mirror/hdf5
2
0
Fork 0
mirror of https://github.com/HDFGroup/hdf5.git synced 2025-03-13 16:47:58 +08:00
Code Issues Packages Projects Releases Wiki Activity
hdf5/src/H5Sselect.c

1620 lines
64 KiB
C
Raw Normal View History

[svn-r6430] Purpose: Code cleanup (sorta) Description: Updated to new copyright information as I browsed these files.
2003-02-24 15:13:07 -05:00
/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
* Copyright by the Board of Trustees of the University of Illinois. *
* All rights reserved. *
* *
* This file is part of HDF5. The full HDF5 copyright notice, including *
* terms governing use, modification, and redistribution, is contained in *
* the files COPYING and Copyright.html. COPYING can be found at the root *
* of the source code distribution tree; Copyright.html can be found at the *
* root level of an installed copy of the electronic HDF5 document set and *
* is linked from the top-level documents page. It can also be found at *
* http://hdf.ncsa.uiuc.edu/HDF5/doc/Copyright.html. If you do not have *
* access to either file, you may request a copy from hdfhelp@ncsa.uiuc.edu. *
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
/* Programmer: Quincey Koziol <koziol@ncsa.uiuc.ued>
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
* Friday, May 29, 1998
*
[svn-r6430] Purpose: Code cleanup (sorta) Description: Updated to new copyright information as I browsed these files.
2003-02-24 15:13:07 -05:00
* Purpose: Dataspace selection functions.
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
*/
[svn-r2652] Purpose: Maintainance & performance enhancements Description: Re-arranged header files to protect private symbols better. Changed optimized regular hyperslab I/O to compute the offsets more efficiently from previous method of using matrix operations. Added sequential I/O operations at a more abstract level (at the same level as H5F_arr_read/write), to support the optimized hyperslab I/O. Platforms tested: Solaris 2.6 (baldric) & FreeBSD 4.1.1 (hawkwind)
2000-10-10 02:43:38 -05:00
#define H5S_PACKAGE /*suppress error about including H5Spkg */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
#include "H5private.h" /* Generic Functions */
#include "H5Dprivate.h" /* Datasets (for their properties) */
#include "H5Eprivate.h" /* Error handling */
#include "H5FLprivate.h" /* Free Lists */
#include "H5Iprivate.h" /* ID Functions */
#include "H5Spkg.h" /* Dataspace functions */
#include "H5Vprivate.h" /* Vector functions */
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
/* Interface initialization */
[svn-r1182] Changed name of PABLO Mask to simplify automatic definitions.
1999-04-14 16:48:05 -05:00
#define PABLO_MASK H5Sselect_mask
[svn-r936] Changes since 19981119 ---------------------- ./src/H5.c ./src/H5A.c ./src/H5AC.c ./src/H5B.c ./src/H5D.c ./src/H5E.c ./src/H5F.c ./src/H5Farray.c ./src/H5Fcore.c ./src/H5Ffamily.c ./src/H5Fistore.c ./src/H5Flow.c ./src/H5Fmpio.c ./src/H5Fsec2.c ./src/H5Fsplit.c ./src/H5Fstdio.c ./src/H5G.c ./src/H5Gent.c ./src/H5Gnode.c ./src/H5Gstab.c ./src/H5HG.c ./src/H5HL.c ./src/H5I.c ./src/H5Iprivate.h ./src/H5MF.c ./src/H5MM.c ./src/H5O.c ./src/H5Oattr.c ./src/H5Ocomp.c ./src/H5Ocont.c ./src/H5Odtype.c ./src/H5Oefl.c ./src/H5Ofill.c ./src/H5Olayout.c ./src/H5Omtime.c ./src/H5Oname.c ./src/H5Osdspace.c ./src/H5Oshared.c ./src/H5Ostab.c ./src/H5P.c ./src/H5R.c ./src/H5RA.c ./src/H5S.c ./src/H5Sall.c ./src/H5Shyper.c ./src/H5Smpio.c ./src/H5Snone.c ./src/H5Spoint.c ./src/H5Sselect.c ./src/H5T.c ./src/H5TB.c ./src/H5Tbit.c ./src/H5Tconv.c ./src/H5V.c ./src/H5Z.c ./src/H5detect.c ./src/H5private.h Most of these changes are because the `interface_initialize_g' variable change from hbool_t to int. It's a one line change. Changed the way the library is closed so we have more control over the order the interfaces are shut down. Instead of registering an atexit() function for every interface in some haphazard order we just register one: H5_term_library() which then calls the H5*_term_interface() functions in a well-defined order. If the library is closed and then reopened repeatedly by calling H5close() and H5open() in a loop we only add one copy of the library termination functions with atexit(). Termination is a two-step process in order to help detect programming errors that would cause an infinite loop caused by the termination of one interface waking up some other previously terminated interface. The first step terminates the interface and *marks it as unusable*. After all interfaces are terminated then we mark them all as usable again. The FUNC_ENTER() macro has been modified to return failure or to dump core (depending on whether NDEBUG is defined) if we try to call an interface while it's shutting down. ./src/H5.c The H5dont_atexit() function returns failure if it's called more than once or if it's called too late. However, the error stack is not automatically printed on failure because the library might not be initialized yet ./test/chunk.c ./test/flush1.c ./test/flush2.c ./test/iopipe.c ./test/overhead.c ./test/ragged.c Changed the extra cast for Win32 so we do floating point division again -- it was just confusion about precedence and associativity of casting and the C coercion rules. Removed extra carriage returns inserted by broken operating system. ./src/H5Ffamily.c Fixed an bug where H5F_fam_write() lowered the EOF marker for one of the family members causing H5F_fam_read() to read zeros. ./test/h5test.h [NEW] ./test/h5test.c [NEW] ./test/Makefile.in ./test/bittests.c ./test/cmpd_dset.c ./test/dsets.c ./test/dtypes.c ./test/extend.c ./test/external.c Support library for test files. This isn't done yet but Katie's contractions are ~10 minutes apart so I figured I better back this stuff up just in case I'm not here next week... Eventually all test files will understand HDF5_DRIVER to name the low level file driver and parameters so we can easily test various drivers. They will also understand HDF5_PREFIX to prepend to the beginning of file names which is necessary for testing ROMIO with various drivers. Also, the cleanup function will know how to use the file name prefix and will understand different file driver naming schemes like file families. I'm not sure they'll understand the `gsf:' type prefixes yet. Note, the external test is completely commented out because I'm in the middle of modifying it. It will still compile and run but it doesn't test anything at the moment.
1998-11-20 22:36:51 -05:00
#define INTERFACE_INIT NULL
[svn-r4355] Purpose: Code cleanup (sorta) Description: When the first versions of the HDF5 library were designed, I remembered vividly the difficulties of porting code from a 32-bit platform to a 16-bit platform and asked that people use intn & uintn instead of int & unsigned int, respectively. However, in hindsight, this was overkill and unnecessary since we weren't going to be porting the HDF5 library to 16-bit architectures. Currently, the extra uintn & intn typedefs are causing problems for users who'd like to include both the HDF5 and HDF4 header files in one source module (like Kent's h4toh5 library). Solution: Changed the uintn & intn's to unsigned and int's respectively. Platforms tested: FreeBSD 4.4 (hawkwind)
2001-08-14 17:09:56 -05:00
static int interface_initialize_g = 0;
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
[svn-r2073] Added free-list code to the library and took out the older "temporary buffer" code, since the functionality was superceded. See the followup document for details on the free-list code.
2000-04-04 16:00:31 -05:00
/* Declare external the free list for hssize_t arrays */
H5FL_ARR_EXTERN(hssize_t);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Declare a free list to manage arrays of size_t */
H5FL_ARR_DEFINE_STATIC(size_t,-1);
/* Declare a free list to manage arrays of hsize_t */
H5FL_ARR_DEFINE_STATIC(hsize_t,-1);
/* Declare a free list to manage blocks of single datatype element data */
H5FL_BLK_EXTERN(type_elem);
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
/*--------------------------------------------------------------------------
NAME
H5S_get_vector_size
PURPOSE
Gets the size of the I/O vector
USAGE
ssize_t H5S_get_vector_size(dxpl_id)
hid_t dxpl_id; IN: The dataset transfer property list to query
RETURNS
Non-negative number of entries in I/O vector on success, negative on failure
DESCRIPTION
Retrieves the number of I/O vector entries to use for a given dataset
transfer. If the default dataset property list is used, the default
number of I/O vectors is returned.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
static ssize_t
H5S_get_vector_size(hid_t dxpl_id)
{
ssize_t ret_value; /* return value */
FUNC_ENTER_NOINIT(H5S_get_vector_size);
if(dxpl_id==H5P_DATASET_XFER_DEFAULT) {
ret_value=H5D_XFER_HYPER_VECTOR_SIZE_DEF;
} /* end if */
else {
H5P_genplist_t *dx_plist; /* Dataset transfer property list */
/* Get the hyperslab vector size */
if(NULL == (dx_plist = H5I_object(dxpl_id)))
HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a dataset transfer property list");
if (H5P_get(dx_plist,H5D_XFER_HYPER_VECTOR_SIZE_NAME,&ret_value)<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get value");
} /* end else */
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
} /* H5S_get_vector_size() */
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
/*--------------------------------------------------------------------------
NAME
H5S_select_copy
PURPOSE
Copy a selection from one dataspace to another
USAGE
herr_t H5S_select_copy(dst, src)
H5S_t *dst; OUT: Pointer to the destination dataspace
H5S_t *src; IN: Pointer to the source dataspace
RETURNS
[svn-r797] Changed comments from returning "SUCCEED/FAIL" to "non-negative/negative" and also fixed a few more explicit checks against FAIL.
1998-10-26 16:18:54 -05:00
Non-negative on success/Negative on failure
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
DESCRIPTION
Copies all the selection information (include offset) from the source
dataspace to the destination dataspace.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
[svn-r468] Fixed formatting problems and moved H5S_sel_iter_release calls above dataspace release calls in H5D_read and H5D_write.
1998-07-08 10:05:01 -05:00
herr_t
H5S_select_copy (H5S_t *dst, const H5S_t *src)
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
{
[svn-r530] Added code so H5Scopy copies selections properly.
1998-07-22 17:11:22 -05:00
herr_t ret_value=SUCCEED; /* return value */
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_NOAPI(H5S_select_copy, FAIL);
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
/* Check args */
assert(dst);
assert(src);
[svn-r530] Added code so H5Scopy copies selections properly.
1998-07-22 17:11:22 -05:00
/* Copy regular fields */
dst->select=src->select;
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
/* Need to copy order information still */
/* Copy offset information */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (NULL==(dst->select.offset = H5FL_ARR_CALLOC(hssize_t,src->extent.u.simple.rank)))
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
HGOTO_ERROR (H5E_RESOURCE, H5E_NOSPACE, FAIL, "memory allocation failed");
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
if(src->select.offset!=NULL)
HDmemcpy(dst->select.offset,src->select.offset,(src->extent.u.simple.rank*sizeof(hssize_t)));
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
/* Perform correct type of copy based on the type of selection */
[svn-r530] Added code so H5Scopy copies selections properly.
1998-07-22 17:11:22 -05:00
switch (src->extent.type) {
case H5S_SCALAR:
/*nothing needed */
break;
case H5S_SIMPLE:
/* Deep copy extra stuff */
switch(src->select.type) {
case H5S_SEL_NONE:
case H5S_SEL_ALL:
/*nothing needed */
break;
case H5S_SEL_POINTS:
ret_value=H5S_point_copy(dst,src);
break;
case H5S_SEL_HYPERSLABS:
ret_value=H5S_hyper_copy(dst,src);
break;
default:
assert("unknown selection type" && 0);
break;
} /* end switch */
break;
case H5S_COMPLEX:
/*void */
break;
default:
assert("unknown data space type" && 0);
break;
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
} /* end switch */
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
} /* H5S_select_copy() */
[svn-r539] Added H5Sselect_all code
1998-07-23 18:29:44 -05:00
[svn-r544] Added code for H5Sselect_none call.
1998-07-24 15:46:19 -05:00
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
/*--------------------------------------------------------------------------
NAME
[svn-r641] Changed function names for beta release.
1998-08-31 22:35:23 -05:00
H5Sget_select_npoints
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
PURPOSE
Get the number of elements in current selection
USAGE
[svn-r793] Changed return value for H5Sget_select_npoints to hssize_t to enable negative return values.
1998-10-26 12:42:48 -05:00
hssize_t H5Sget_select_npoints(dsid)
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
hid_t dsid; IN: Dataspace ID of selection to query
RETURNS
The number of elements in selection on success, 0 on failure
DESCRIPTION
Returns the number of elements in current selection for dataspace.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
[svn-r793] Changed return value for H5Sget_select_npoints to hssize_t to enable negative return values.
1998-10-26 12:42:48 -05:00
hssize_t
[svn-r641] Changed function names for beta release.
1998-08-31 22:35:23 -05:00
H5Sget_select_npoints(hid_t spaceid)
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
{
H5S_t *space = NULL; /* Dataspace to modify selection of */
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
hssize_t ret_value; /* return value */
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_API(H5Sget_select_npoints, 0);
[svn-r795] Changes since 19981026 ---------------------- ./MANIFEST ./test/th5s.h5 [NEW] ./test/th5s.c Added a test to make sure that creating a data space with too large a rank fails. Added a test to make sure that reading a file that has a dataset with a space with too large a rank fails. Actually, this one is a little weird: the code that reads the data space message assumes the space is scalar if the message cannot be read. Fortunately the layout message fails also, preventing the dataset from being opened. However, since the data type message is still visible h5ls will report that the object is a named data type. ./test/space_overflow.c [NEW] This is the little program that makes the th5s.h5 file. ./src/H5A.c ./src/H5R.c ./src/H5Sselect.c Updated trace info. ./src/H5Olayout.c ./src/H5Osdspace.c Added code to fail if the dimensionality is too large when decoding a layout or simple data space message. ./src/H5Oprivate.h Redefined H5O_LAYOUT_NDIMS in terms of H5S_MAX_RANK. ./src/H5P.c ./src/H5S.c Check for ndims>H5S_MAX_RANK in API function calls, added assert to internal functions. ./src/H5V.c Changed a `<' to an `<=' in an assert. ./test/flush2.c Includes stdlib.h for getenv(). ./tools/h5tools.c Able to handle up to H5S_MAX_RANK dimensions during output.
1998-10-26 14:55:54 -05:00
H5TRACE1("Hs","i",spaceid);
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
/* Check args */
[svn-r5840] Purpose: Code cleanup Description: Created a new H5I function which combined the some of the functionality of H5I_get_type and H5I_object: H5I_object_verify. Using this new function in the library trims another ~200 lines of code off the library and makes the resulting binaries smaller and faster also. Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 14:17:12 -05:00
if (NULL == (space=H5I_object_verify(spaceid, H5I_DATASPACE)))
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, 0, "not a data space");
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value = (*space->select.get_npoints)(space);
[svn-r451] Completely tore out existing dataspace API and replaced with code to match API defined in the html/Dataspaces.html document. This code does not include support for strides, merging selections, or permutations of coordinates yet, but it's a drop-in replacement for the existing API with the same features.
1998-07-06 16:01:13 -05:00
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_API(ret_value);
[svn-r641] Changed function names for beta release.
1998-08-31 22:35:23 -05:00
} /* H5Sget_select_npoints() */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
/*--------------------------------------------------------------------------
NAME
H5Sselect_valid
PURPOSE
Check whether the selection fits within the extent, with the current
offset defined.
USAGE
[svn-r800] Made a "htri_t" as a return value from "boolean" functions returning TRUE/FALSE/FAIL, hbool_t is now strictly for true boolean values.
1998-10-26 17:44:13 -05:00
htri_t H5Sselect_void(dsid)
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
hid_t dsid; IN: Dataspace ID to query
RETURNS
TRUE if the selection fits within the extent, FALSE if it does not and
[svn-r797] Changed comments from returning "SUCCEED/FAIL" to "non-negative/negative" and also fixed a few more explicit checks against FAIL.
1998-10-26 16:18:54 -05:00
Negative on an error.
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
DESCRIPTION
Determines if the current selection at the current offet fits within the
extent for the dataspace.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
[svn-r800] Made a "htri_t" as a return value from "boolean" functions returning TRUE/FALSE/FAIL, hbool_t is now strictly for true boolean values.
1998-10-26 17:44:13 -05:00
htri_t
[svn-r620] Changes since 19980825 ---------------------- ./MANIFEST ./src/H5R.c [NEW] ./src/H5Rprivate.h [NEW] ./src/H5Rpublic.h [NEW] ./src/Makefile.in ./src/hdf5.h ./test/ragged.c [NEW] Preliminary support for 2d ragged arrays for Mark Miller and Jim Reus. Not fully implemented yet. The test is not actually part of `make test' because we still have some memory problems. ./src/H5E.c ./src/H5Epublic.h Added H5E_RAGGED as a major error number. ./bin/release Checks the MANIFEST file against `svf ls' on systems that have it. ./bin/trace Fixed a bug that caused arguments of type `void *x[]' to not be handled. ./src/H5.c Removed unused variables and changed a couple types to fix compiler warnings. Added tracing support for ragged array object ID's and arrays of pointers. ./src/H5D.c H5Dcreate() will complain if either of the property lists are invalid (instead of using the default). ./src/H5D.c ./src/H5Dprivate.h Split H5Dget_space() into an API and internal function so it can be called from the new ragged array layer. ./src/H5Fistore.c Fixed warnings about unsigned vs. signed comparisons. ./src/H5Flow.c Fixed a warning about a variable being shadowed in the MPI-IO stuff. ./src/H5Iprivate.h ./src/H5Ipublic.h Added the H5_RAGGED atom group. ./src/H5Shyper.c Fixed some freeing-free-memory errors that resulted when certain arrays were freed but the pointers were left in the data structures. I simply set the pointers to null after they were freed. ./src/H5Sprivate.h ./src/H5Sselect.c Split the H5Sselect_hyperslab() function into an API and a private function so it could be called from the ragged array layer. Added H5S_SEL_ERROR and H5S_SEL_N to the switch statements to get rid or compiler warnings. ./src/H5Tconv.c Removed a misleading comment. ./test/bittests.c Fixed a warning about a printf(). ./test/cmpd_dset.c Fixed warnings about unused variables because of test #11 being commented out. ./bin/trace Shortened the right margin for the output to allow room for the `);' at the end of the TRACE() macros.
1998-08-27 11:48:50 -05:00
H5Sselect_valid(hid_t spaceid)
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
{
H5S_t *space = NULL; /* Dataspace to modify selection of */
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
htri_t ret_value; /* return value */
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_API(H5Sselect_valid, 0);
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
H5TRACE1("b","i",spaceid);
/* Check args */
[svn-r5840] Purpose: Code cleanup Description: Created a new H5I function which combined the some of the functionality of H5I_get_type and H5I_object: H5I_object_verify. Using this new function in the library trims another ~200 lines of code off the library and makes the resulting binaries smaller and faster also. Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 14:17:12 -05:00
if (NULL == (space=H5I_object_verify(spaceid, H5I_DATASPACE)))
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, 0, "not a data space");
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value = (*space->select.is_valid)(space);
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_API(ret_value);
[svn-r566] Fixed offset code [finally :-/ ] for selections.
1998-08-03 19:30:35 -05:00
} /* H5Sselect_valid() */
[svn-r4586] Purpose: Code speedups, etc. Description: Bring in new algorithms and data structures for dealing with hyperslabs. This speeds up the hyperslab I/O for non-regular hyperslabs by a huge amount. Currently, the new API functions are ifdef'ed out, pending discussion and consensus approval. Platforms tested: FreeBSD 4.4 (hawkwind)
2001-11-02 15:31:35 -05:00
[svn-r947] Lots of patches to enable dataset region references. The testing for these need to be finished and checked in still.
1998-11-24 19:29:09 -05:00
/*--------------------------------------------------------------------------
NAME
H5S_select_deserialize
PURPOSE
Deserialize the current selection from a user-provided buffer into a real
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
selection in the dataspace.
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
USAGE
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
herr_t H5S_select_deserialize(space, buf)
H5S_t *space; IN/OUT: Dataspace pointer to place selection into
uint8 *buf; IN: Buffer to retrieve serialized selection from
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
RETURNS
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
Non-negative on success/Negative on failure
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
DESCRIPTION
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
Deserializes the current selection into a buffer. (Primarily for retrieving
from disk). This routine just hands off to the appropriate routine for each
type of selection. The format of the serialized information is shown in
the H5S_select_serialize() header.
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
herr_t
H5S_select_deserialize (H5S_t *space, const uint8_t *buf)
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
{
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
const uint8_t *tbuf; /* Temporary pointer to the selection type */
uint32_t sel_type; /* Pointer to the selection type */
herr_t ret_value=FAIL; /* return value */
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
FUNC_ENTER_NOAPI(H5S_select_deserialize, FAIL);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
assert(space);
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
tbuf=buf;
UINT32DECODE(tbuf, sel_type);
switch(sel_type) {
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
case H5S_SEL_POINTS: /* Sequence of points selected */
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value=H5S_point_deserialize(space,buf);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
break;
case H5S_SEL_HYPERSLABS: /* Hyperslab selection defined */
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value=H5S_hyper_deserialize(space,buf);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
break;
case H5S_SEL_ALL: /* Entire extent selected */
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value=H5S_all_deserialize(space,buf);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
break;
case H5S_SEL_NONE: /* Nothing selected */
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value=H5S_none_deserialize(space,buf);
break;
default:
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
break;
}
[svn-r5867] Purpose: Code cleanup Description: Changed the last HRETURN* statements in the FUNC_ENTER macros into HGOTO* macros, which reduces the size of the library binary in certain configurations by another 10% Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel, IRIX64 6.5 (modi4) serial & parallel
2002-08-09 15:48:23 -05:00
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
} /* H5S_select_deserialize() */
[svn-r4586] Purpose: Code speedups, etc. Description: Bring in new algorithms and data structures for dealing with hyperslabs. This speeds up the hyperslab I/O for non-regular hyperslabs by a huge amount. Currently, the new API functions are ifdef'ed out, pending discussion and consensus approval. Platforms tested: FreeBSD 4.4 (hawkwind)
2001-11-02 15:31:35 -05:00
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
/*--------------------------------------------------------------------------
NAME
H5Sget_select_bounds
PURPOSE
Gets the bounding box containing the selection.
USAGE
herr_t H5S_get_select_bounds(space, start, end)
hid_t dsid; IN: Dataspace ID of selection to query
hsize_t *start; OUT: Starting coordinate of bounding box
hsize_t *end; OUT: Opposite coordinate of bounding box
RETURNS
Non-negative on success, negative on failure
DESCRIPTION
Retrieves the bounding box containing the current selection and places
it into the user's buffers. The start and end buffers must be large
enough to hold the dataspace rank number of coordinates. The bounding box
exactly contains the selection, ie. if a 2-D element selection is currently
defined with the following points: (4,5), (6,8) (10,7), the bounding box
with be (4, 5), (10, 8). Calling this function on a "none" selection
returns fail.
The bounding box calculations _does_ include the current offset of the
selection within the dataspace extent.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
herr_t
H5Sget_select_bounds(hid_t spaceid, hsize_t *start, hsize_t *end)
{
H5S_t *space = NULL; /* Dataspace to modify selection of */
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
herr_t ret_value; /* return value */
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_API(H5Sget_select_bounds, FAIL);
[svn-r1138] Changes since 19990304 ---------------------- ./config/*-aix4.* [REMOVED] ./config/*-aix4.x [NEW] ./config/solaris2.5 [REMOVED] ./config/solaris2.x [NEW] ./configure.in ./configure [REGENERATED] ./MANIFEST Changed the names of the IBM-SP2 config files by replacing the minor version numbers with an `x'. The solaris config warns about old versions of gcc. The RUNSERIAL value is set to the empty string by default. The config/* file (or user) can override it by setting it to some other value. The `--enable-parallel' with no flags turns on the HAVE_PARALLEL constant in the source code but doesn't add any MPI or MPI-IO libraries to the link line. If we are compiling for parallel then configure checks that we can actually link a very simple program. This has the benefit of detecting config errors before we waste time compiling the entire library. Configure recognizes `mpcc_r' as a parallel compiler. ./src/H5A.c ./src/H5Apublic.c Fixed some compiler warnings. Changed the return type of H5Aget_name() from hssize_t to ssize_t because the name can never be larger then memory.
1999-03-12 13:35:04 -05:00
H5TRACE3("e","i*h*h",spaceid,start,end);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
/* Check args */
if(start==NULL || end==NULL)
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "invalid pointer");
[svn-r5840] Purpose: Code cleanup Description: Created a new H5I function which combined the some of the functionality of H5I_get_type and H5I_object: H5I_object_verify. Using this new function in the library trims another ~200 lines of code off the library and makes the resulting binaries smaller and faster also. Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 14:17:12 -05:00
if (NULL == (space=H5I_object_verify(spaceid, H5I_DATASPACE)))
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a data space");
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
ret_value = (*space->select.bounds)(space,start,end);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_API(ret_value);
[svn-r1132] Dataset region references are now finished and working correctly. Also, there are five new API functions for querying selections: H5Sget_select_hyper_nblocks - retrieves the number of hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_npoints - retrieves the number of element points in current element selection for a dataspace H5Sget_select_hyper_blocklist - retrieves a list of the hyperslab blocks in current hyperslab selection for a dataspace H5Sget_select_elem_pointlist - retrieves a list of the element points in current element selection for a dataspace H5Sget_select_bounds - retrieves a n-dimensional bounding box containing current selection.
1999-03-10 18:50:03 -05:00
} /* H5Sget_select_bounds() */
[svn-r4586] Purpose: Code speedups, etc. Description: Bring in new algorithms and data structures for dealing with hyperslabs. This speeds up the hyperslab I/O for non-regular hyperslabs by a huge amount. Currently, the new API functions are ifdef'ed out, pending discussion and consensus approval. Platforms tested: FreeBSD 4.4 (hawkwind)
2001-11-02 15:31:35 -05:00
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
/*--------------------------------------------------------------------------
NAME
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
H5S_select_iterate
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
PURPOSE
Iterate over the selected elements in a memory buffer.
USAGE
herr_t H5S_select_iterate(buf, type_id, space, operator, operator_data)
void *buf; IN/OUT: Buffer containing elements to iterate over
hid_t type_id; IN: Datatype ID of BUF array.
H5S_t *space; IN: Dataspace object containing selection to iterate over
[svn-r1566] Changed C++ "operator" keyword to "op"
1999-08-10 13:54:06 -05:00
H5D_operator_t op; IN: Function pointer to the routine to be
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
called for each element in BUF iterated over.
void *operator_data; IN/OUT: Pointer to any user-defined data
associated with the operation.
RETURNS
Returns the return value of the last operator if it was non-zero, or zero
if all elements were processed. Otherwise returns a negative value.
DESCRIPTION
Iterates over the selected elements in a memory buffer, calling the user's
callback function for each element. The selection in the dataspace is
modified so that any elements already iterated over are removed from the
selection if the iteration is interrupted (by the H5D_operator_t function
returning non-zero) in the "middle" of the iteration and may be re-started
by the user where it left off.
NOTE: Until "subtracting" elements from a selection is implemented,
the selection is not modified.
--------------------------------------------------------------------------*/
herr_t
[svn-r1566] Changed C++ "operator" keyword to "op"
1999-08-10 13:54:06 -05:00
H5S_select_iterate(void *buf, hid_t type_id, H5S_t *space, H5D_operator_t op,
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
void *operator_data)
{
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
H5T_t *dt; /* Datatype structure */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
H5S_sel_iter_t iter; /* Selection iteration info */
hbool_t iter_init=0; /* Selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
uint8_t *loc; /* Current element location in buffer */
hssize_t coords[H5O_LAYOUT_NDIMS]; /* Coordinates of element in dataspace */
hssize_t nelmts; /* Number of elements in selection */
hsize_t space_size[H5O_LAYOUT_NDIMS]; /* Dataspace size */
hsize_t *off=NULL; /* Array to store sequence offsets */
hsize_t curr_off; /* Current offset within sequence */
hsize_t tmp_off; /* Temporary offset within sequence */
size_t *len=NULL; /* Array to store sequence lengths */
size_t curr_len; /* Length of bytes left to process in sequence */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t nseq; /* Number of sequences generated */
size_t curr_seq; /* Current sequnce being worked on */
size_t nbytes; /* Number of bytes used in sequences */
size_t max_bytes; /* Maximum number of bytes allowed in sequences */
size_t elmt_size; /* Datatype size */
int ndims; /* Number of dimensions in dataspace */
int i; /* Local Index variable */
herr_t user_ret=0; /* User's return value */
herr_t ret_value=SUCCEED; /* Return value */
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_NOAPI(H5S_select_iterate, FAIL);
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
/* Check args */
assert(buf);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
assert(H5I_DATATYPE == H5I_get_type(type_id));
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
assert(space);
[svn-r1566] Changed C++ "operator" keyword to "op"
1999-08-10 13:54:06 -05:00
assert(op);
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the hyperslab vector size */
/* (from the default data transfer property list, for now) */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(H5P_DATASET_XFER_DEFAULT))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
/* Get the datatype size */
[svn-r5840] Purpose: Code cleanup Description: Created a new H5I function which combined the some of the functionality of H5I_get_type and H5I_object: H5I_object_verify. Using this new function in the library trims another ~200 lines of code off the library and makes the resulting binaries smaller and faster also. Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 14:17:12 -05:00
if (NULL==(dt=H5I_object_verify(type_id,H5I_DATATYPE)))
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not an valid base datatype");
if((elmt_size=H5T_get_size(dt))==0)
HGOTO_ERROR(H5E_DATATYPE, H5E_BADSIZE, FAIL, "datatype size invalid");
/* Initialize iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if ((*space->select.iter_init)(space, elmt_size, &iter)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
iter_init=1; /* Selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the number of elements in selection */
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
if((nelmts = (*space->select.get_npoints)(space))<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTCOUNT, FAIL, "can't get number of elements selected");
/* Get the rank of the dataspace */
ndims=space->extent.u.simple.rank;
[svn-r6006] Purpose: Bug fix (ID 841) Description: h5dumptst (generating h5dump test files) failed with an assertion error in H5S_select_iterate. H5S_select_iterate did not account for scalar type that has rank=0, thus no dimension sizes. Solution: Put the assertion of valid sizes only after verify rank is larger than 0. (fix is coached by Quincey.) Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)}? YES. Misc. update: Update release_docs/RELEASE for bug fixes, new features, etc. YES.
2002-10-15 16:12:48 -05:00
if (ndims > 0){
/* Copy the size of the space */
assert(space->extent.u.simple.size);
HDmemcpy(space_size, space->extent.u.simple.size, ndims*sizeof(hsize_t));
}
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
space_size[ndims]=elmt_size;
/* Compute the maximum number of bytes required */
H5_ASSIGN_OVERFLOW(max_bytes,nelmts*elmt_size,hsize_t,size_t);
/* Loop, while elements left in selection */
while(max_bytes>0 && user_ret==0) {
/* Get the sequences of bytes */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*space->select.get_seq_list)(space,0,&iter,elmt_size,(size_t)vector_size,max_bytes,&nseq,&nbytes,off,len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Loop, while sequences left to process */
for(curr_seq=0; curr_seq<nseq && user_ret==0; curr_seq++) {
/* Get the current offset */
curr_off=off[curr_seq];
/* Get the number of bytes in sequence */
curr_len=len[curr_seq];
/* Loop, while bytes left in sequence */
while(curr_len>0 && user_ret==0) {
/* Compute the coordinate from the offset */
for(i=ndims, tmp_off=curr_off; i>=0; i--) {
coords[i]=tmp_off%space_size[i];
tmp_off/=space_size[i];
} /* end for */
/* Get the location within the user's buffer */
loc=(unsigned char *)buf+curr_off;
/* Call user's callback routine */
user_ret=(*op)(loc,type_id,(hsize_t)ndims,coords,operator_data);
/* Increment offset in dataspace */
curr_off+=elmt_size;
/* Decrement number of bytes left in sequence */
curr_len-=elmt_size;
} /* end while */
} /* end for */
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Decrement number of elements left to process */
assert((nbytes%elmt_size)==0);
max_bytes-=nbytes;
} /* end while */
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Set return value */
ret_value=user_ret;
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
done:
/* Release selection iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if(iter_init) {
if ((*space->select.iter_release)(&iter)<0)
HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end if */
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Release length & offset vectors */
if(len!=NULL)
H5FL_ARR_FREE(size_t,len);
if(off!=NULL)
H5FL_ARR_FREE(hsize_t,off);
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r1374] Added in code to support the H5Diterate function, which I've got to add tests for now. Also, I revised some of the code for hyperslab I/O, which should provide a modest speedup in situations with lots of hyperslabs.
1999-06-23 21:16:13 -05:00
} /* end H5S_select_iterate() */
[svn-r4914] Purpose: Bug fix & feature add Description: Added new API function H5Sget_select_type to determine type of selection in a dataspace. Return values are defined by the H5S_sel_type enumerated type in H5Spublic.h Also, hyperslab operations involving a "all" or "none" selection are not generating the correct resulting selections. Solution: Added more code to make hyperslab operations against an "all" or "none" selection generate the correct results. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-02-07 11:21:24 -05:00
/*--------------------------------------------------------------------------
NAME
H5Sget_select_type
PURPOSE
Retrieve the type of selection in a dataspace
USAGE
H5S_sel_type H5Sget_select_type(space_id)
hid_t space_id; IN: Dataspace object to reset
RETURNS
Non-negative on success/Negative on failure. Return value is from the
set of values in the H5S_sel_type enumerated type.
DESCRIPTION
This function retrieves the type of selection currently defined for
a dataspace.
--------------------------------------------------------------------------*/
H5S_sel_type
H5Sget_select_type(hid_t space_id)
{
H5S_t *space = NULL; /* dataspace to modify */
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
H5S_sel_type ret_value; /* Return value */
[svn-r4914] Purpose: Bug fix & feature add Description: Added new API function H5Sget_select_type to determine type of selection in a dataspace. Return values are defined by the H5S_sel_type enumerated type in H5Spublic.h Also, hyperslab operations involving a "all" or "none" selection are not generating the correct resulting selections. Solution: Added more code to make hyperslab operations against an "all" or "none" selection generate the correct results. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-02-07 11:21:24 -05:00
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_API(H5Sget_select_type, H5S_SEL_ERROR);
[svn-r4914] Purpose: Bug fix & feature add Description: Added new API function H5Sget_select_type to determine type of selection in a dataspace. Return values are defined by the H5S_sel_type enumerated type in H5Spublic.h Also, hyperslab operations involving a "all" or "none" selection are not generating the correct resulting selections. Solution: Added more code to make hyperslab operations against an "all" or "none" selection generate the correct results. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-02-07 11:21:24 -05:00
H5TRACE1("St","i",space_id);
/* Check args */
[svn-r5840] Purpose: Code cleanup Description: Created a new H5I function which combined the some of the functionality of H5I_get_type and H5I_object: H5I_object_verify. Using this new function in the library trims another ~200 lines of code off the library and makes the resulting binaries smaller and faster also. Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 14:17:12 -05:00
if (NULL == (space = H5I_object_verify(space_id, H5I_DATASPACE)))
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
HGOTO_ERROR(H5E_ATOM, H5E_BADATOM, H5S_SEL_ERROR, "not a data space");
[svn-r4914] Purpose: Bug fix & feature add Description: Added new API function H5Sget_select_type to determine type of selection in a dataspace. Return values are defined by the H5S_sel_type enumerated type in H5Spublic.h Also, hyperslab operations involving a "all" or "none" selection are not generating the correct resulting selections. Solution: Added more code to make hyperslab operations against an "all" or "none" selection generate the correct results. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-02-07 11:21:24 -05:00
[svn-r5842] Purpose: Code cleanup Description: Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with HRETURN macros to HGOTO_DONE macros. This unifies the error return path from functions and reduces the size of the library by up to 10% on some platforms. Additionally, I improved a lot of the error cleanup code in many routines. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial & parallel.
2002-08-08 11:52:55 -05:00
/* Set return value */
ret_value=space->select.type;
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_API(ret_value);
[svn-r4914] Purpose: Bug fix & feature add Description: Added new API function H5Sget_select_type to determine type of selection in a dataspace. Return values are defined by the H5S_sel_type enumerated type in H5Spublic.h Also, hyperslab operations involving a "all" or "none" selection are not generating the correct resulting selections. Solution: Added more code to make hyperslab operations against an "all" or "none" selection generate the correct results. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-02-07 11:21:24 -05:00
} /* end H5Sget_select_type() */
[svn-r5138] Purpose: Bug Fix & Code Cleanup Description: The MPI-IO optimized transfer routines (H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all the cases where they could be used. Additionally, the code for determining if an optimized transfer is wrapped into the actual I/O transfer routine in a very confusing way. Solution: Re-enabled MPI-IO optimized transfer routines in all the cases where they should work. Extracted all the pre-conditions for optimized transfers into separate routines from the transfer routines. Platforms tested: FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
2002-04-03 12:07:14 -05:00
/*--------------------------------------------------------------------------
NAME
H5S_select_shape_same
PURPOSE
Check if two selections are the same shape
USAGE
htri_t H5S_select_shape_same(space1, space2)
const H5S_t *space1; IN: 1st Dataspace pointer to compare
const H5S_t *space2; IN: 2nd Dataspace pointer to compare
RETURNS
TRUE/FALSE/FAIL
DESCRIPTION
Checks to see if the current selection in the dataspaces are the same
dimensionality and shape.
This is primarily used for reading the entire selection in one swoop.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
Assumes that there is only a single "block" for hyperslab selections.
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
htri_t
H5S_select_shape_same(const H5S_t *space1, const H5S_t *space2)
{
H5S_hyper_span_t *span1=NULL,*span2=NULL; /* Hyperslab span node */
[svn-r5843] Purpose: Code cleanup Description: Clean up a few warnings which were showing up with --enable-production turned on. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel
2002-08-08 12:52:17 -05:00
hsize_t elmts1=0,elmts2=0; /* Number of elements in each dimension of selection */
[svn-r5138] Purpose: Bug Fix & Code Cleanup Description: The MPI-IO optimized transfer routines (H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all the cases where they could be used. Additionally, the code for determining if an optimized transfer is wrapped into the actual I/O transfer routine in a very confusing way. Solution: Re-enabled MPI-IO optimized transfer routines in all the cases where they should work. Extracted all the pre-conditions for optimized transfers into separate routines from the transfer routines. Platforms tested: FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
2002-04-03 12:07:14 -05:00
unsigned u; /* Index variable */
htri_t ret_value=TRUE; /* return value */
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_NOAPI(H5S_select_shape_same, FAIL);
[svn-r5138] Purpose: Bug Fix & Code Cleanup Description: The MPI-IO optimized transfer routines (H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all the cases where they could be used. Additionally, the code for determining if an optimized transfer is wrapped into the actual I/O transfer routine in a very confusing way. Solution: Re-enabled MPI-IO optimized transfer routines in all the cases where they should work. Extracted all the pre-conditions for optimized transfers into separate routines from the transfer routines. Platforms tested: FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
2002-04-03 12:07:14 -05:00
/* Check args */
assert(space1);
assert(space2);
if (space1->extent.u.simple.rank!=space2->extent.u.simple.rank)
HGOTO_DONE(FALSE);
/* Get information about memory and file */
for (u=0; u<space1->extent.u.simple.rank; u++) {
switch(space1->select.type) {
case H5S_SEL_HYPERSLABS:
/* Check size hyperslab selection in this dimension */
if(space1->select.sel_info.hslab.diminfo != NULL) {
elmts1=space1->select.sel_info.hslab.diminfo[u].block;
} /* end if */
else {
/* Check for the first dimension */
if(span1==NULL)
span1=space1->select.sel_info.hslab.span_lst->head;
/* Get the number of elements in the span */
elmts1=(span1->high-span1->low)+1;
/* Advance to the next dimension */
span1=span1->down->head;
} /* end else */
break;
case H5S_SEL_ALL:
elmts1=space1->extent.u.simple.size[u];
break;
case H5S_SEL_POINTS:
elmts1=1;
break;
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
case H5S_SEL_NONE:
elmts1=0;
break;
[svn-r5138] Purpose: Bug Fix & Code Cleanup Description: The MPI-IO optimized transfer routines (H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all the cases where they could be used. Additionally, the code for determining if an optimized transfer is wrapped into the actual I/O transfer routine in a very confusing way. Solution: Re-enabled MPI-IO optimized transfer routines in all the cases where they should work. Extracted all the pre-conditions for optimized transfers into separate routines from the transfer routines. Platforms tested: FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
2002-04-03 12:07:14 -05:00
default:
assert(0 && "Invalid selection type!");
} /* end switch */
switch(space2->select.type) {
case H5S_SEL_HYPERSLABS:
/* Check size hyperslab selection in this dimension */
if(space2->select.sel_info.hslab.diminfo != NULL) {
elmts2=space2->select.sel_info.hslab.diminfo[u].block;
} /* end if */
else {
/* Check for the first dimension */
if(span2==NULL)
span2=space2->select.sel_info.hslab.span_lst->head;
/* Get the number of elements in the span */
elmts2=(span2->high-span2->low)+1;
/* Advance to the next dimension */
span2=span2->down->head;
} /* end else */
break;
case H5S_SEL_ALL:
elmts2=space2->extent.u.simple.size[u];
break;
case H5S_SEL_POINTS:
elmts2=1;
break;
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
case H5S_SEL_NONE:
elmts2=0;
break;
[svn-r5138] Purpose: Bug Fix & Code Cleanup Description: The MPI-IO optimized transfer routines (H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all the cases where they could be used. Additionally, the code for determining if an optimized transfer is wrapped into the actual I/O transfer routine in a very confusing way. Solution: Re-enabled MPI-IO optimized transfer routines in all the cases where they should work. Extracted all the pre-conditions for optimized transfers into separate routines from the transfer routines. Platforms tested: FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
2002-04-03 12:07:14 -05:00
default:
assert(0 && "Invalid selection type!");
} /* end switch */
/* Make certaint the selections have the same number of elements in this dimension */
if (elmts1!=elmts2)
HGOTO_DONE(FALSE);
} /* end for */
done:
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5138] Purpose: Bug Fix & Code Cleanup Description: The MPI-IO optimized transfer routines (H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all the cases where they could be used. Additionally, the code for determining if an optimized transfer is wrapped into the actual I/O transfer routine in a very confusing way. Solution: Re-enabled MPI-IO optimized transfer routines in all the cases where they should work. Extracted all the pre-conditions for optimized transfers into separate routines from the transfer routines. Platforms tested: FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
2002-04-03 12:07:14 -05:00
} /* H5S_select_shape_same() */
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
/*--------------------------------------------------------------------------
NAME
H5S_select_fill
PURPOSE
Fill a selection in memory with a value
USAGE
herr_t H5S_select_fill(fill,fill_size,space,buf)
const void *fill; IN: Pointer to fill value to use
size_t fill_size; IN: Size of elements in memory buffer & size of
fill value
H5S_t *space; IN: Dataspace describing memory buffer &
containing selection to use.
void *buf; IN/OUT: Memory buffer to fill selection in
RETURNS
Non-negative on success/Negative on failure.
DESCRIPTION
Use the selection in the dataspace to fill elements in a memory buffer.
GLOBAL VARIABLES
COMMENTS, BUGS, ASSUMPTIONS
The memory buffer elements are assumed to have the same datatype as the
fill value being placed into them.
EXAMPLES
REVISION LOG
--------------------------------------------------------------------------*/
herr_t
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
H5S_select_fill(void *_fill, size_t fill_size, const H5S_t *space, void *_buf)
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
{
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
H5S_sel_iter_t iter; /* Selection iteration info */
hbool_t iter_init=0; /* Selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
uint8_t *buf; /* Current location in buffer */
[svn-r5259] Purpose: Code cleanup Description: Previously, the I/O pipeline (pline), external file list (efl) and fill- value (fill) structs were passed down the raw data function call chain, even into and/or through functions which didn't use them. Since all three of these pieces of information are available from the dataset creation property list, just pass the dataset creation property list down the function call chain and query for the information needed in a particular function. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-04-25 12:56:56 -05:00
void *fill=_fill; /* Alias for fill-value buffer */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
hssize_t nelmts; /* Number of elements in selection */
hsize_t *off=NULL; /* Array to store sequence offsets */
size_t *len=NULL; /* Array to store sequence lengths */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t nseq; /* Number of sequences generated */
size_t curr_seq; /* Current sequnce being worked on */
size_t nbytes; /* Number of bytes used in sequences */
size_t max_bytes; /* Total number of bytes in selection */
herr_t ret_value=SUCCEED; /* return value */
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
[svn-r5471] Purpose: Code cleanup Description: Broke the FUNC_ENTER macro into several macros, with more specialized uses (which followup mail will describe). This was designed to move most/all of the checks which could be done at compile time to that point, instead of needlessly performing them (over & over :-) at run-time. This reduces the library's size (and thus staticly linked binaries) and has a minor speedup effect also. Platforms tested: IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing on FreeBSD and Solaris immediately after the checkin.
2002-05-29 10:07:55 -05:00
FUNC_ENTER_NOAPI(H5S_select_fill, FAIL);
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
/* Check args */
assert(fill_size>0);
assert(space);
[svn-r5835] Purpose: Bug fix. Description: Correct erroneous assertion. Platforms tested: IRIX64 6.5 (modi4) w/parallel
2002-07-24 14:09:55 -05:00
assert(_buf);
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
[svn-r5254] Purpose: Code tweak. Description: Allow the 'fill' parameter of H5S_select_fill to be NULL and allocate a temporary buffer for it, if so. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-04-25 08:15:22 -05:00
/* Check if we need a temporary fill value buffer */
if(fill==NULL) {
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (NULL==(fill = H5FL_BLK_CALLOC(type_elem,fill_size)))
[svn-r5254] Purpose: Code tweak. Description: Allow the 'fill' parameter of H5S_select_fill to be NULL and allocate a temporary buffer for it, if so. Platforms tested: FreeBSD 4.5 (sleipnir)
2002-04-25 08:15:22 -05:00
HGOTO_ERROR (H5E_RESOURCE, H5E_NOSPACE, FAIL, "fill value buffer allocation failed");
} /* end if */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the hyperslab vector size */
/* (from the default data transfer property list, for now) */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(H5P_DATASET_XFER_DEFAULT))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
/* Initialize iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if ((*space->select.iter_init)(space, fill_size, &iter)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
iter_init=1; /* Selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the number of elements in selection */
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
if((nelmts = (*space->select.get_npoints)(space))<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTCOUNT, FAIL, "can't get number of elements selected");
/* Compute the number of bytes to process */
H5_CHECK_OVERFLOW(nelmts,hssize_t,size_t);
max_bytes=(size_t)nelmts*fill_size;
/* Loop, while elements left in selection */
while(max_bytes>0) {
/* Get the sequences of bytes */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*space->select.get_seq_list)(space,0,&iter,fill_size,(size_t)vector_size,max_bytes,&nseq,&nbytes,off,len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Loop over sequences */
for(curr_seq=0; curr_seq<nseq; curr_seq++) {
/* Get offset in memory buffer */
buf=(uint8_t *)_buf+off[curr_seq];
/* Fill each sequence in memory with fill value */
assert((len[curr_seq]%fill_size)==0);
H5V_array_fill(buf, fill, fill_size, (len[curr_seq]/fill_size));
} /* end for */
/* Decrement number of bytes left to process */
max_bytes-=nbytes;
} /* end while */
done:
/* Release selection iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if(iter_init) {
if ((*space->select.iter_release)(&iter)<0)
HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end if */
/* Release length & offset vectors */
if(len!=NULL)
H5FL_ARR_FREE(size_t,len);
if(off!=NULL)
H5FL_ARR_FREE(hsize_t,off);
/* Release fill value, if allocated */
if(_fill==NULL && fill)
H5FL_BLK_FREE(type_elem,fill);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* H5S_select_fill() */
/*-------------------------------------------------------------------------
* Function: H5S_select_fscat
*
* Purpose: Scatters dataset elements from the type conversion buffer BUF
* to the file F where the data points are arranged according to
* the file data space FILE_SPACE and stored according to
* LAYOUT and EFL. Each element is ELMT_SIZE bytes.
* The caller is requesting that NELMTS elements are copied.
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Quincey Koziol
* Thursday, June 20, 2002
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
[svn-r5894] Purpose: Bug fix/Code cleanup/New Feature Description: Correct problems with writing fill-values to external storage and allocate the data storage at the correct times. Also, mostly straighten out the strange code which allocates and fills raw data storage for datasets. Things are still a bit odd in that the fill-values for chunked datasets are written when the space is allocated, instead of in a separate routine, but there are two reasons for this: it's inefficient (especially in parallel) to iterate through all the chunks twice, and (more importantly) the space needed to store compressed chunks isn't known until we've got a buffer of compressed fill-values ready to write to the chunk. Additionally, add in the H5D_SPACE_ALLOC_INCR and H5D_SPACE_ALLOC_DEFAULT setting for the "space time", which incorporate the previous behavior of the space allocation for chunked datasets. The default settings for the different types of dataset storage are now as follows: Contiguous - Late Chunked - Incremental Compact - Early This checkin also incorporates a change to the behavior of external data storage in two ways - fill-values are _never_ written to external storage (under the assumption that writing fill-values is triggered by allocating space in an HDF5 file, and since space is not allocated in the file, the fill-values should not be written) and external data files are now created if they don't exist when data is written to them. The fill-value will probably need to be revisited at some time in the future, this just seemed like the safer course currently. I think I cleaned up some compiler errors also, before getting bogged down in the fixes for the space allocation and fill-values. Platforms tested: FreeBSD 4.6 (sleipnir) w/serial & parallel. Will be testing on IRIX64 6.5 (modi4) in serial & parallel shortly.
2002-08-27 08:41:32 -05:00
H5S_select_fscat (H5F_t *f, struct H5O_layout_t *layout,
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
H5P_genplist_t *dc_plist, const H5O_efl_t *efl, size_t elmt_size,
const H5S_t *space, H5S_sel_iter_t *iter, hsize_t nelmts,
hid_t dxpl_id, const void *_buf)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
{
const uint8_t *buf=_buf; /* Alias for pointer arithmetic */
hsize_t *off=NULL; /* Array to store sequence offsets */
size_t *len=NULL; /* Array to store sequence lengths */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t maxbytes; /* Number of bytes in the buffer */
size_t nseq; /* Number of sequences generated */
size_t nbytes; /* Number of bytes used in sequences */
herr_t ret_value=SUCCEED; /* Return value */
FUNC_ENTER_NOAPI(H5S_select_fscat, FAIL);
/* Check args */
assert (f);
assert (layout);
assert (elmt_size>0);
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
assert (efl);
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
assert (space);
assert (iter);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
assert (nelmts>0);
assert (_buf);
[svn-r6047] Purpose: Bug fix & code cleanup Description: Hyperslab code for collapsing dimensions was incorrectly collapsing selections inappropriately when the fastest changing dimension couldn't be collapsed. Also add some more assertions which will make similar bugs easier to find. Solution: Break out of loop earlier. Platforms tested: modi4 (parallel), too small to triple check.
2002-11-01 13:39:20 -05:00
assert(TRUE==H5P_isa_class(dxpl_id,H5P_DATASET_XFER));
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the hyperslab vector size */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(dxpl_id))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
/* Compute the number of bytes available in buffer */
H5_ASSIGN_OVERFLOW(maxbytes,nelmts*elmt_size,hsize_t,size_t);
/* Loop until all elements are written */
while(maxbytes>0) {
/* Get list of sequences for selection to write */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*space->select.get_seq_list)(space,H5S_GET_SEQ_LIST_SORTED,iter,elmt_size,(size_t)vector_size,maxbytes,&nseq,&nbytes,off,len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Write sequence list out */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (H5F_seq_writev(f, dxpl_id, layout, dc_plist, efl, space, elmt_size, nseq, len, off, buf)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error");
/* Update buffer */
buf += nbytes;
/* Decrement number of elements left to process */
assert(nbytes%elmt_size==0);
maxbytes -= nbytes;
} /* end while */
done:
if(len!=NULL)
H5FL_ARR_FREE(size_t,len);
if(off!=NULL)
H5FL_ARR_FREE(hsize_t,off);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* H5S_select_fscat() */
/*-------------------------------------------------------------------------
* Function: H5S_select_fgath
*
* Purpose: Gathers data points from file F and accumulates them in the
* type conversion buffer BUF. The LAYOUT argument describes
* how the data is stored on disk and EFL describes how the data
* is organized in external files. ELMT_SIZE is the size in
* bytes of a datum which this function treats as opaque.
* FILE_SPACE describes the data space of the dataset on disk
* and the elements that have been selected for reading (via
* hyperslab, etc). This function will copy at most NELMTS
* elements.
*
* Return: Success: Number of elements copied.
* Failure: 0
*
* Programmer: Quincey Koziol
* Monday, June 24, 2002
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
hsize_t
H5S_select_fgath (H5F_t *f, const struct H5O_layout_t *layout,
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
H5P_genplist_t *dc_plist, const H5O_efl_t *efl, size_t elmt_size,
const H5S_t *space, H5S_sel_iter_t *iter, hsize_t nelmts,
hid_t dxpl_id, void *_buf/*out*/)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
{
uint8_t *buf=_buf; /* Alias for pointer arithmetic */
hsize_t *off=NULL; /* Array to store sequence offsets */
size_t *len=NULL; /* Array to store sequence lengths */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t maxbytes; /* Number of bytes in the buffer */
size_t nseq; /* Number of sequences generated */
size_t nbytes; /* Number of bytes used in sequences */
hsize_t ret_value=nelmts; /* Return value */
FUNC_ENTER_NOAPI(H5S_select_fgath, 0);
/* Check args */
assert (f);
assert (layout);
assert (elmt_size>0);
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
assert (efl);
[svn-r5839] Purpose: Code cleanup Description: Move operations on each type of selection into the source code file for each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.) Remove central H5S_select_<foo> operations, instead calling the operations through function pointers in each selection (a much more object-oriented approach). Platforms tested: FreeBSD 4.6 (sleipnir)
2002-07-31 10:27:07 -05:00
assert (space);
assert (iter);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
assert (nelmts>0);
assert (_buf);
/* Get the hyperslab vector size */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(dxpl_id))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, 0, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, 0, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, 0, "can't allocate I/O offset vector array");
/* Compute the number of bytes available in buffer */
H5_ASSIGN_OVERFLOW(maxbytes,nelmts*elmt_size,hsize_t,size_t);
/* Loop until all elements are written */
while(maxbytes>0) {
/* Get list of sequences for selection to write */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*space->select.get_seq_list)(space,H5S_GET_SEQ_LIST_SORTED,iter,elmt_size,(size_t)vector_size,maxbytes,&nseq,&nbytes,off,len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, 0, "sequence length generation failed");
/* Read sequence list in */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (H5F_seq_readv(f, dxpl_id, layout, dc_plist, efl, space, elmt_size, nseq, len, off, buf)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, 0, "read error");
/* Update buffer */
buf += nbytes;
/* Decrement number of elements left to process */
assert(nbytes%elmt_size==0);
maxbytes -= nbytes;
} /* end while */
done:
if(len!=NULL)
H5FL_ARR_FREE(size_t,len);
if(off!=NULL)
H5FL_ARR_FREE(hsize_t,off);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* H5S_select_fgath() */
/*-------------------------------------------------------------------------
* Function: H5S_select_mscat
*
* Purpose: Scatters NELMTS data points from the scatter buffer
* TSCAT_BUF to the application buffer BUF. Each element is
* ELMT_SIZE bytes and they are organized in application memory
* according to SPACE.
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Quincey Koziol
* Monday, July 8, 2002
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_select_mscat (const void *_tscat_buf, size_t elmt_size, const H5S_t *space,
H5S_sel_iter_t *iter, hsize_t nelmts, hid_t dxpl_id, void *_buf/*out*/)
{
uint8_t *buf=(uint8_t *)_buf; /* Get local copies for address arithmetic */
const uint8_t *tscat_buf=(const uint8_t *)_tscat_buf;
hsize_t *off=NULL; /* Array to store sequence offsets */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t *len=NULL; /* Array to store sequence lengths */
size_t curr_len; /* Length of bytes left to process in sequence */
size_t maxbytes; /* Number of bytes in the buffer */
size_t nseq; /* Number of sequences generated */
size_t curr_seq; /* Current sequence being processed */
size_t nbytes; /* Number of bytes used in sequences */
herr_t ret_value=SUCCEED; /* Number of elements scattered */
FUNC_ENTER_NOAPI(H5S_select_mscat, FAIL);
/* Check args */
assert (tscat_buf);
assert (elmt_size>0);
assert (space);
assert (iter);
assert (nelmts>0);
assert (buf);
/* Get the hyperslab vector size */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(dxpl_id))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
/* Compute the number of bytes available in buffer */
H5_ASSIGN_OVERFLOW(maxbytes,nelmts*elmt_size,hsize_t,size_t);
/* Loop until all elements are written */
while(maxbytes>0) {
/* Get list of sequences for selection to write */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*space->select.get_seq_list)(space,0,iter,elmt_size,(size_t)vector_size,maxbytes,&nseq,&nbytes,off,len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, 0, "sequence length generation failed");
/* Loop, while sequences left to process */
for(curr_seq=0; curr_seq<nseq; curr_seq++) {
/* Get the number of bytes in sequence */
curr_len=len[curr_seq];
HDmemcpy(buf+off[curr_seq],tscat_buf,curr_len);
/* Advance offset in destination buffer */
tscat_buf+=curr_len;
} /* end for */
/* Decrement number of elements left to process */
assert(nbytes%elmt_size==0);
maxbytes -= nbytes;
} /* end while */
done:
if(len!=NULL)
H5FL_ARR_FREE(size_t,len);
if(off!=NULL)
H5FL_ARR_FREE(hsize_t,off);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* H5S_select_mscat() */
/*-------------------------------------------------------------------------
* Function: H5S_select_mgath
*
* Purpose: Gathers dataset elements from application memory BUF and
* copies them into the gather buffer TGATH_BUF.
* Each element is ELMT_SIZE bytes and arranged in application
* memory according to SPACE.
* The caller is requesting that at most NELMTS be gathered.
*
* Return: Success: Number of elements copied.
* Failure: 0
*
* Programmer: Quincey Koziol
* Monday, June 24, 2002
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
hsize_t
H5S_select_mgath (const void *_buf, size_t elmt_size, const H5S_t *space,
H5S_sel_iter_t *iter, hsize_t nelmts, hid_t dxpl_id, void *_tgath_buf/*out*/)
{
const uint8_t *buf=(const uint8_t *)_buf; /* Get local copies for address arithmetic */
uint8_t *tgath_buf=(uint8_t *)_tgath_buf;
hsize_t *off=NULL; /* Array to store sequence offsets */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t *len=NULL; /* Array to store sequence lengths */
size_t curr_len; /* Length of bytes left to process in sequence */
size_t maxbytes; /* Number of bytes in the buffer */
size_t nseq; /* Number of sequences generated */
size_t curr_seq; /* Current sequence being processed */
size_t nbytes; /* Number of bytes used in sequences */
hsize_t ret_value=nelmts; /* Number of elements gathered */
FUNC_ENTER_NOAPI(H5S_select_mgath, 0);
/* Check args */
assert (buf);
assert (elmt_size>0);
assert (space);
assert (iter);
assert (nelmts>0);
assert (tgath_buf);
/* Get the hyperslab vector size */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(dxpl_id))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, 0, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, 0, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, 0, "can't allocate I/O offset vector array");
/* Compute the number of bytes available in buffer */
H5_ASSIGN_OVERFLOW(maxbytes,nelmts*elmt_size,hsize_t,size_t);
/* Loop until all elements are written */
while(maxbytes>0) {
/* Get list of sequences for selection to write */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*space->select.get_seq_list)(space,0,iter,elmt_size,(size_t)vector_size,maxbytes,&nseq,&nbytes,off,len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, 0, "sequence length generation failed");
/* Loop, while sequences left to process */
for(curr_seq=0; curr_seq<nseq; curr_seq++) {
/* Get the number of bytes in sequence */
curr_len=len[curr_seq];
HDmemcpy(tgath_buf,buf+off[curr_seq],curr_len);
/* Advance offset in gather buffer */
tgath_buf+=curr_len;
} /* end for */
/* Decrement number of elements left to process */
assert(nbytes%elmt_size==0);
maxbytes -= nbytes;
} /* end while */
done:
if(len!=NULL)
H5FL_ARR_FREE(size_t,len);
if(off!=NULL)
H5FL_ARR_FREE(hsize_t,off);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* H5S_select_mgath() */
/*-------------------------------------------------------------------------
* Function: H5S_select_read
*
* Purpose: Reads directly from file into application memory.
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Quincey Koziol
* Tuesday, July 23, 2002
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_select_read(H5F_t *f, const H5O_layout_t *layout, H5P_genplist_t *dc_plist,
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
const H5O_efl_t *efl, size_t elmt_size, const H5S_t *file_space,
const H5S_t *mem_space, hid_t dxpl_id, void *_buf/*out*/)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
{
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
H5S_sel_iter_t mem_iter; /* Memory selection iteration info */
hbool_t mem_iter_init=0; /* Memory selection iteration info has been initialized */
H5S_sel_iter_t file_iter; /* File selection iteration info */
hbool_t file_iter_init=0; /* File selection iteration info has been initialized */
[svn-r5843] Purpose: Code cleanup Description: Clean up a few warnings which were showing up with --enable-production turned on. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel
2002-08-08 12:52:17 -05:00
uint8_t *buf=NULL; /* Local buffer pointer, for address arithmetic */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
hsize_t *mem_off=NULL; /* Array to store sequence offsets in memory */
hsize_t *file_off=NULL; /* Array to store sequence offsets in the file */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t *mem_len=NULL; /* Array to store sequence lengths in memory */
size_t *file_len=NULL; /* Array to store sequence lengths in the file */
size_t maxbytes; /* Number of bytes in selection */
size_t mem_nseq; /* Number of sequences generated in the file */
size_t file_nseq; /* Number of sequences generated in memory */
size_t mem_nbytes; /* Number of bytes used in memory sequences */
size_t file_nbytes; /* Number of bytes used in file sequences */
size_t curr_mem_seq; /* Current memory sequence to operate on */
size_t curr_file_seq; /* Current file sequence to operate on */
size_t tmp_file_len; /* Temporary number of bytes in file sequence */
unsigned partial_file; /* Whether a partial file sequence was accessed */
[svn-r5843] Purpose: Code cleanup Description: Clean up a few warnings which were showing up with --enable-production turned on. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel
2002-08-08 12:52:17 -05:00
size_t orig_file_len=0; /* Original file sequence length for partial file access */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t orig_file_seq; /* Original file sequence to operate on */
size_t tot_file_seq; /* Number of file sequences to access */
herr_t ret_value=SUCCEED; /* Return value */
FUNC_ENTER_NOAPI(H5S_select_read, FAIL);
[svn-r6047] Purpose: Bug fix & code cleanup Description: Hyperslab code for collapsing dimensions was incorrectly collapsing selections inappropriately when the fastest changing dimension couldn't be collapsed. Also add some more assertions which will make similar bugs easier to find. Solution: Break out of loop earlier. Platforms tested: modi4 (parallel), too small to triple check.
2002-11-01 13:39:20 -05:00
/* Check args */
assert(f);
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
assert(efl);
[svn-r6047] Purpose: Bug fix & code cleanup Description: Hyperslab code for collapsing dimensions was incorrectly collapsing selections inappropriately when the fastest changing dimension couldn't be collapsed. Also add some more assertions which will make similar bugs easier to find. Solution: Break out of loop earlier. Platforms tested: modi4 (parallel), too small to triple check.
2002-11-01 13:39:20 -05:00
assert(_buf);
assert(TRUE==H5P_isa_class(dxpl_id,H5P_DATASET_XFER));
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the hyperslab vector size */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(dxpl_id))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((mem_len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((mem_off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((file_len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((file_off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
/* Initialize file iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if ((*file_space->select.iter_init)(file_space, elmt_size, &file_iter)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
file_iter_init=1; /* File selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Initialize memory iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if ((*mem_space->select.iter_init)(mem_space, elmt_size, &mem_iter)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
mem_iter_init=1; /* Memory selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get number of bytes in selection */
[svn-r6296] Purpose: Code cleanup Description: Reduce warnings on Windows Platforms tested: FreeBSD 4.7 (sleipnir)
2003-01-17 15:34:14 -05:00
#ifndef NDEBUG
{
hsize_t tmp_maxbytes=(*file_space->select.get_npoints)(file_space)*elmt_size;
H5_ASSIGN_OVERFLOW(maxbytes,tmp_maxbytes,hsize_t,size_t);
}
#else /* NDEBUG */
maxbytes=(size_t)((*file_space->select.get_npoints)(file_space)*elmt_size);
#endif /* NDEBUG */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Initialize sequence counts */
curr_mem_seq=curr_file_seq=0;
mem_nseq=file_nseq=0;
/* Loop, until all bytes are processed */
while(maxbytes>0) {
/* Check if more file sequences are needed */
if(curr_file_seq>=file_nseq) {
/* Get sequences for file selection */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*file_space->select.get_seq_list)(file_space,H5S_GET_SEQ_LIST_SORTED,&file_iter,elmt_size,(size_t)vector_size,maxbytes,&file_nseq,&file_nbytes,file_off,file_len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Start at the beginning of the sequences again */
curr_file_seq=0;
} /* end if */
/* Check if more memory sequences are needed */
if(curr_mem_seq>=mem_nseq) {
/* Get sequences for memory selection */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*mem_space->select.get_seq_list)(mem_space,0,&mem_iter,elmt_size,(size_t)vector_size,maxbytes,&mem_nseq,&mem_nbytes,mem_off,mem_len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Start at the beginning of the sequences again */
curr_mem_seq=0;
/* Set the buffer pointer using the first sequence */
H5_CHECK_OVERFLOW(mem_off[0],hsize_t,size_t);
buf=(uint8_t *)_buf+(size_t)mem_off[0];
} /* end if */
/* Check if current file sequence will fit into current memory sequence */
if(mem_len[curr_mem_seq]>=file_len[curr_file_seq]) {
/* Save the current number file sequence */
orig_file_seq=curr_file_seq;
/* Determine how many file sequences will fit into current memory sequence */
tmp_file_len=0;
tot_file_seq=0;
[svn-r5924] Purpose: Code cleanup Description: Switch order of test on array bounds to avoid reading data outside of array Located by purify. Platforms tested: FreeBSD 4.6 (sleipnir) w/C++ Solaris 2.7 (arabica) w/C++ IRIX64 6.5 (modi4) w/parallel & fortran
2002-09-13 11:27:09 -05:00
while( curr_file_seq<file_nseq && (tmp_file_len+file_len[curr_file_seq])<=mem_len[curr_mem_seq] ) {
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
tmp_file_len+=file_len[curr_file_seq];
curr_file_seq++;
tot_file_seq++;
} /* end while */
/* Check for partial file sequence */
if(tmp_file_len<mem_len[curr_mem_seq] && curr_file_seq<file_nseq) {
/* Get the original file sequence length */
orig_file_len=file_len[curr_file_seq];
/* Make the last file sequence a partial access */
file_len[curr_file_seq]=mem_len[curr_mem_seq]-tmp_file_len;
/* Increase the number of bytes to access */
tmp_file_len=mem_len[curr_mem_seq];
/* Indicate that there is an extra sequence to include in the file access */
tot_file_seq++;
/* Indicate a partial file sequence */
partial_file=1;
} /* end if */
else
partial_file=0;
/* Read file sequences into current memory sequence */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (H5F_seq_readv(f, dxpl_id, layout, dc_plist, efl, file_space, elmt_size, tot_file_seq, &file_len[orig_file_seq], &file_off[orig_file_seq], buf)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, FAIL, "read error");
/* Update last file sequence, if it was partially accessed */
if(partial_file) {
file_off[curr_file_seq]+=orig_file_len-file_len[curr_file_seq];
file_len[curr_file_seq]=orig_file_len-file_len[curr_file_seq];
} /* end if */
/* Check if the current memory sequence was only partially accessed */
if(tmp_file_len<mem_len[curr_mem_seq]) {
/* Adjust current memory sequence */
mem_off[curr_mem_seq]+=tmp_file_len;
mem_len[curr_mem_seq]-=tmp_file_len;
/* Adjust memory buffer pointer */
buf+=tmp_file_len;
} /* end if */
else {
/* Must have used entire memory sequence, advance to next one */
curr_mem_seq++;
/* Check if it is valid to adjust buffer pointer */
if(curr_mem_seq<mem_nseq) {
H5_CHECK_OVERFLOW(mem_off[curr_mem_seq],hsize_t,size_t);
buf=(uint8_t *)_buf+(size_t)mem_off[curr_mem_seq];
} /* end if */
} /* end else */
/* Decrement number of bytes left to process */
maxbytes-=tmp_file_len;
} /* end if */
else {
/* Save number of bytes to access */
tmp_file_len=mem_len[curr_mem_seq];
/* Read part of current file sequence into current memory sequence */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (H5F_seq_read(f, dxpl_id, layout, dc_plist, efl, file_space, elmt_size, tmp_file_len, file_off[curr_file_seq], buf)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, FAIL, "read error");
/* Update current file sequence information */
file_off[curr_file_seq]+=tmp_file_len;
file_len[curr_file_seq]-=tmp_file_len;
/* Increment memory sequence */
curr_mem_seq++;
/* Check if it is valid to adjust buffer pointer */
if(curr_mem_seq<mem_nseq) {
H5_CHECK_OVERFLOW(mem_off[curr_mem_seq],hsize_t,size_t);
buf=(uint8_t *)_buf+(size_t)mem_off[curr_mem_seq];
} /* end if */
/* Decrement number of bytes left to process */
maxbytes-=tmp_file_len;
} /* end else */
} /* end while */
done:
/* Release file selection iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if(file_iter_init) {
if ((*file_space->select.iter_release)(&file_iter)<0)
HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end if */
/* Release memory selection iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if(mem_iter_init) {
if ((*mem_space->select.iter_release)(&mem_iter)<0)
HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end if */
/* Free vector arrays */
if(file_len!=NULL)
H5FL_ARR_FREE(size_t,file_len);
if(file_off!=NULL)
H5FL_ARR_FREE(hsize_t,file_off);
if(mem_len!=NULL)
H5FL_ARR_FREE(size_t,mem_len);
if(mem_off!=NULL)
H5FL_ARR_FREE(hsize_t,mem_off);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end H5S_select_read() */
/*-------------------------------------------------------------------------
* Function: H5S_select_write
*
* Purpose: Writes directly from application memory into a file
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Quincey Koziol
* Tuesday, July 23, 2002
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
[svn-r5894] Purpose: Bug fix/Code cleanup/New Feature Description: Correct problems with writing fill-values to external storage and allocate the data storage at the correct times. Also, mostly straighten out the strange code which allocates and fills raw data storage for datasets. Things are still a bit odd in that the fill-values for chunked datasets are written when the space is allocated, instead of in a separate routine, but there are two reasons for this: it's inefficient (especially in parallel) to iterate through all the chunks twice, and (more importantly) the space needed to store compressed chunks isn't known until we've got a buffer of compressed fill-values ready to write to the chunk. Additionally, add in the H5D_SPACE_ALLOC_INCR and H5D_SPACE_ALLOC_DEFAULT setting for the "space time", which incorporate the previous behavior of the space allocation for chunked datasets. The default settings for the different types of dataset storage are now as follows: Contiguous - Late Chunked - Incremental Compact - Early This checkin also incorporates a change to the behavior of external data storage in two ways - fill-values are _never_ written to external storage (under the assumption that writing fill-values is triggered by allocating space in an HDF5 file, and since space is not allocated in the file, the fill-values should not be written) and external data files are now created if they don't exist when data is written to them. The fill-value will probably need to be revisited at some time in the future, this just seemed like the safer course currently. I think I cleaned up some compiler errors also, before getting bogged down in the fixes for the space allocation and fill-values. Platforms tested: FreeBSD 4.6 (sleipnir) w/serial & parallel. Will be testing on IRIX64 6.5 (modi4) in serial & parallel shortly.
2002-08-27 08:41:32 -05:00
H5S_select_write(H5F_t *f, H5O_layout_t *layout, H5P_genplist_t *dc_plist,
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
const H5O_efl_t *efl, size_t elmt_size, const H5S_t *file_space,
const H5S_t *mem_space, hid_t dxpl_id, const void *_buf/*out*/)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
{
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
H5S_sel_iter_t mem_iter; /* Memory selection iteration info */
hbool_t mem_iter_init=0; /* Memory selection iteration info has been initialized */
H5S_sel_iter_t file_iter; /* File selection iteration info */
hbool_t file_iter_init=0; /* File selection iteration info has been initialized */
[svn-r5843] Purpose: Code cleanup Description: Clean up a few warnings which were showing up with --enable-production turned on. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel
2002-08-08 12:52:17 -05:00
const uint8_t *buf=NULL; /* Local buffer pointer, for address arithmetic */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
hsize_t *mem_off=NULL; /* Array to store sequence offsets in memory */
hsize_t *file_off=NULL; /* Array to store sequence offsets in the file */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
ssize_t vector_size; /* Value for vector size */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t *mem_len=NULL; /* Array to store sequence lengths in memory */
size_t *file_len=NULL; /* Array to store sequence lengths in the file */
size_t maxbytes; /* Number of bytes in selection */
size_t mem_nseq; /* Number of sequences generated in the file */
size_t file_nseq; /* Number of sequences generated in memory */
size_t mem_nbytes; /* Number of bytes used in memory sequences */
size_t file_nbytes; /* Number of bytes used in file sequences */
size_t curr_mem_seq; /* Current memory sequence to operate on */
size_t curr_file_seq; /* Current file sequence to operate on */
size_t tmp_file_len; /* Temporary number of bytes in file sequence */
unsigned partial_file; /* Whether a partial file sequence was accessed */
[svn-r5843] Purpose: Code cleanup Description: Clean up a few warnings which were showing up with --enable-production turned on. Platforms tested: FreeBSD 4.6 (sleipnir) serial & parallel
2002-08-08 12:52:17 -05:00
size_t orig_file_len=0; /* Original file sequence length for partial file access */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
size_t orig_file_seq; /* Original file sequence to operate on */
size_t tot_file_seq; /* Number of file sequences to access */
herr_t ret_value=SUCCEED; /* Return value */
FUNC_ENTER_NOAPI(H5S_select_write, FAIL);
[svn-r6047] Purpose: Bug fix & code cleanup Description: Hyperslab code for collapsing dimensions was incorrectly collapsing selections inappropriately when the fastest changing dimension couldn't be collapsed. Also add some more assertions which will make similar bugs easier to find. Solution: Break out of loop earlier. Platforms tested: modi4 (parallel), too small to triple check.
2002-11-01 13:39:20 -05:00
/* Check args */
assert(f);
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
assert(efl);
[svn-r6047] Purpose: Bug fix & code cleanup Description: Hyperslab code for collapsing dimensions was incorrectly collapsing selections inappropriately when the fastest changing dimension couldn't be collapsed. Also add some more assertions which will make similar bugs easier to find. Solution: Break out of loop earlier. Platforms tested: modi4 (parallel), too small to triple check.
2002-11-01 13:39:20 -05:00
assert(_buf);
assert(TRUE==H5P_isa_class(dxpl_id,H5P_DATASET_XFER));
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get the hyperslab vector size */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((vector_size=H5S_get_vector_size(dxpl_id))<0)
HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "unable to get I/O vector size");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Allocate the vector I/O arrays */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((mem_len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((mem_off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((file_len = H5FL_ARR_MALLOC(size_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O length vector array");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((file_off = H5FL_ARR_MALLOC(hsize_t,(size_t)vector_size))==NULL)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "can't allocate I/O offset vector array");
/* Initialize file iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if ((*file_space->select.iter_init)(file_space, elmt_size, &file_iter)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
file_iter_init=1; /* File selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Initialize memory iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if ((*mem_space->select.iter_init)(mem_space, elmt_size, &mem_iter)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator");
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
mem_iter_init=1; /* Memory selection iteration info has been initialized */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Get number of bytes in selection */
[svn-r6296] Purpose: Code cleanup Description: Reduce warnings on Windows Platforms tested: FreeBSD 4.7 (sleipnir)
2003-01-17 15:34:14 -05:00
#ifndef NDEBUG
{
hsize_t tmp_maxbytes=(*file_space->select.get_npoints)(file_space)*elmt_size;
H5_ASSIGN_OVERFLOW(maxbytes,tmp_maxbytes,hsize_t,size_t);
}
#else /* NDEBUG */
maxbytes=(size_t)((*file_space->select.get_npoints)(file_space)*elmt_size);
#endif /* NDEBUG */
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
/* Initialize sequence counts */
curr_mem_seq=curr_file_seq=0;
mem_nseq=file_nseq=0;
/* Loop, until all bytes are processed */
while(maxbytes>0) {
/* Check if more file sequences are needed */
if(curr_file_seq>=file_nseq) {
/* Get sequences for file selection */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*file_space->select.get_seq_list)(file_space,H5S_GET_SEQ_LIST_SORTED,&file_iter,elmt_size,(size_t)vector_size,maxbytes,&file_nseq,&file_nbytes,file_off,file_len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Start at the beginning of the sequences again */
curr_file_seq=0;
} /* end if */
/* Check if more memory sequences are needed */
if(curr_mem_seq>=mem_nseq) {
/* Get sequences for memory selection */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if((*mem_space->select.get_seq_list)(mem_space,0,&mem_iter,elmt_size,(size_t)vector_size,maxbytes,&mem_nseq,&mem_nbytes,mem_off,mem_len)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR (H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed");
/* Start at the beginning of the sequences again */
curr_mem_seq=0;
/* Set the buffer pointer using the first sequence */
H5_CHECK_OVERFLOW(mem_off[0],hsize_t,size_t);
buf=(const uint8_t *)_buf+(size_t)mem_off[0];
} /* end if */
/* Check if current file sequence will fit into current memory sequence */
if(mem_len[curr_mem_seq]>=file_len[curr_file_seq]) {
/* Save the current number file sequence */
orig_file_seq=curr_file_seq;
/* Determine how many file sequences will fit into current memory sequence */
tmp_file_len=0;
tot_file_seq=0;
[svn-r5924] Purpose: Code cleanup Description: Switch order of test on array bounds to avoid reading data outside of array Located by purify. Platforms tested: FreeBSD 4.6 (sleipnir) w/C++ Solaris 2.7 (arabica) w/C++ IRIX64 6.5 (modi4) w/parallel & fortran
2002-09-13 11:27:09 -05:00
while( curr_file_seq<file_nseq && (tmp_file_len+file_len[curr_file_seq])<=mem_len[curr_mem_seq] ) {
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
tmp_file_len+=file_len[curr_file_seq];
curr_file_seq++;
tot_file_seq++;
} /* end while */
/* Check for partial file sequence */
if(tmp_file_len<mem_len[curr_mem_seq] && curr_file_seq<file_nseq) {
/* Get the original file sequence length */
orig_file_len=file_len[curr_file_seq];
/* Make the last file sequence a partial access */
file_len[curr_file_seq]=mem_len[curr_mem_seq]-tmp_file_len;
/* Increase the number of bytes to access */
tmp_file_len=mem_len[curr_mem_seq];
/* Indicate that there is an extra sequence to include in the file access */
tot_file_seq++;
/* Indicate a partial file sequence */
partial_file=1;
} /* end if */
else
partial_file=0;
/* Write current memory sequence into file sequences */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (H5F_seq_writev(f, dxpl_id, layout, dc_plist, efl, file_space, elmt_size, tot_file_seq, &file_len[orig_file_seq], &file_off[orig_file_seq], buf)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error");
/* Update last file sequence, if it was partially accessed */
if(partial_file) {
file_off[curr_file_seq]+=orig_file_len-file_len[curr_file_seq];
file_len[curr_file_seq]=orig_file_len-file_len[curr_file_seq];
} /* end if */
/* Check if the current memory sequence was only partially accessed */
if(tmp_file_len<mem_len[curr_mem_seq]) {
/* Adjust current memory sequence */
mem_off[curr_mem_seq]+=tmp_file_len;
mem_len[curr_mem_seq]-=tmp_file_len;
/* Adjust memory buffer pointer */
buf+=tmp_file_len;
} /* end if */
else {
/* Must have used entire memory sequence, advance to next one */
curr_mem_seq++;
/* Check if it is valid to adjust buffer pointer */
if(curr_mem_seq<mem_nseq) {
H5_CHECK_OVERFLOW(mem_off[curr_mem_seq],hsize_t,size_t);
buf=(const uint8_t *)_buf+(size_t)mem_off[curr_mem_seq];
} /* end if */
} /* end else */
/* Decrement number of bytes left to process */
maxbytes-=tmp_file_len;
} /* end if */
else {
/* Save number of bytes to access */
tmp_file_len=mem_len[curr_mem_seq];
/* Write part of current memory sequence to current file sequence */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if (H5F_seq_write(f, dxpl_id, layout, dc_plist, efl, file_space, elmt_size, tmp_file_len, file_off[curr_file_seq], buf)<0)
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error");
/* Update current file sequence information */
file_off[curr_file_seq]+=tmp_file_len;
file_len[curr_file_seq]-=tmp_file_len;
/* Increment memory sequence */
curr_mem_seq++;
/* Check if it is valid to adjust buffer pointer */
if(curr_mem_seq<mem_nseq) {
H5_CHECK_OVERFLOW(mem_off[curr_mem_seq],hsize_t,size_t);
buf=(const uint8_t *)_buf+(size_t)mem_off[curr_mem_seq];
} /* end if */
/* Decrement number of bytes left to process */
maxbytes-=tmp_file_len;
} /* end else */
} /* end while */
done:
/* Release file selection iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if(file_iter_init) {
if ((*file_space->select.iter_release)(&file_iter)<0)
HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end if */
/* Release memory selection iterator */
[svn-r6252] Purpose: Lots of performance improvements & a couple new internal API interfaces. Description: Performance Improvements: - Cached file offset & length sizes in shared file struct, to avoid constantly looking them up in the FCPL. - Generic property improvements: - Added "revision" number to generic property classes to speed up comparisons. - Changed method of storing properties from using a hash-table to the TBBT routines in the library. - Share the propery names between classes and the lists derived from them. - Removed redundant 'def_value' buffer from each property. - Switching code to use a "copy on write" strategy for properties in each list, where the properties in each list are shared with the properties in the class, until a property's value is changed in a list. - Fixed error in layout code which was allocating too many buffers. - Redefined public macros of the form (H5open()/H5check, <variable>) internally to only be (<variable>), avoiding innumerable useless calls to H5open() and H5check_version(). - Reuse already zeroed buffers in H5F_contig_fill instead of constantly re-zeroing them. - Don't write fill values if writing entire dataset. - Use gettimeofday() system call instead of time() system when checking the modification time of a dataset. - Added reference counted string API and use it for tracking the names of objects opening in a file (for the ID->name code). - Removed redundant H5P_get() calls in B-tree routines. - Redefine H5T datatype macros internally to the library, to avoid calling H5check redundantly. - Keep dataspace information for dataset locally instead of reading from disk each time. Added new module to track open objects in a file, to allow this (which will be useful eventually for some FPH5 metadata caching issues). - Remove H5AC_find macro which was inlining metadata cache lookups, and call function instead. - Remove redundant memset() calls from H5G_namei() routine. - Remove redundant checking of object type when locating objects in metadata cache and rely on the address only. - Create default dataset object to use when default dataset creation property list is used to create datasets, bypassing querying for all the property list values. - Use default I/O vector size when performing raw data with the default dataset transfer property list, instead of querying for I/O vector size. - Remove H5P_DEFAULT internally to the library, replacing it with more specific default property list based on the type of property list needed. - Remove redundant memset() calls in object header message (H5O*) routines. - Remove redunant memset() calls in data I/O routines. - Split free-list allocation routines into malloc() and calloc()- like routines, instead of one combined routine. - Remove lots of indirection in H5O*() routines. - Simplify metadata cache entry comparison routine (used when flushing entire cache out). - Only enable metadata cache statistics when H5AC_DEBUG is turned on, instead of always tracking them. - Simplify address comparison macro (H5F_addr_eq). - Remove redundant metadata cache entry protections during dataset creation by protecting the object header once and making all the modifications necessary for the dataset creation before unprotecting it. - Reduce # of "number of element in extent" computations performed by computing and storing the value during dataspace creation. - Simplify checking for group location's file information, when file has not been involving in file-mounting operations. - Use binary encoding for modification time, instead of ASCII. - Hoist H5HL_peek calls (to get information in a local heap) out of loops in many group routine. - Use static variable for iterators of selections, instead of dynamically allocation them each time. - Lookup & insert new entries in one step, avoiding traversing group's B-tree twice. - Fixed memory leak in H5Gget_objname_idx() routine (tangential to performance improvements, but fixed along the way). - Use free-list for reference counted strings. - Don't bother copying object names into cached group entries, since they are re-created when an object is opened. The benchmark I used to measure these results created several thousand small (2K) datasets in a file and wrote out the data for them. This is Elena's "regular.c" benchmark. These changes resulted in approximately ~4.3x speedup of the development branch when compared to the previous code in the development branch and ~1.4x speedup compared to the release branch. Additionally, these changes reduce the total memory used (code and data) by the development branch by ~800KB, bringing the development branch back into the same ballpark as the release branch. I'll send out a more detailed description of the benchmark results as a followup note. New internal API routines: Added "reference counted strings" API for tracking strings that get used by multiple owners without duplicating the strings. Added "ternary search tree" API for text->object mappings. Platforms tested: Tested h5committest {arabica (fortran), eirene (fortran, C++) modi4 (parallel, fortran)} Other platforms/configurations tested? FreeBSD 4.7 (sleipnir) serial & parallel Solaris 2.6 (baldric) serial
2003-01-09 12:20:03 -05:00
if(mem_iter_init) {
if ((*mem_space->select.iter_release)(&mem_iter)<0)
HDONE_ERROR (H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator");
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end if */
/* Free vector arrays */
if(file_len!=NULL)
H5FL_ARR_FREE(size_t,file_len);
if(file_off!=NULL)
H5FL_ARR_FREE(hsize_t,file_off);
if(mem_len!=NULL)
H5FL_ARR_FREE(size_t,mem_len);
if(mem_off!=NULL)
H5FL_ARR_FREE(hsize_t,mem_off);
[svn-r6266] Purpose: Code cleanup/new feature. Description: Split FUNC_LEAVE into API and non-API specific versions. This allows a solution to compiling this branch with C++, as well as reducing the size of the binaries produced. Platforms tested: FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
2003-01-10 15:26:02 -05:00
FUNC_LEAVE_NOAPI(ret_value);
[svn-r5834] Purpose: Large code cleanup/re-write Description: This is phase 1 of the data I/O re-architecture, with the following changes: - Changed the selection drivers to not actually do any I/O, they only generate the sequences of offset/length pairs needed for the I/O (or memory access, in the case of iterating or filling a selection in a memory buffer) - Wrote more abstract I/O routines which get the sequence of offset/ length pairs for each selection and access perform the I/O or memory access. Benefits of this change include: - Removed ~3400 lines of quite redundant code, with corresponding reduction in the size of library binary. - Any selection can now directly access memory when performing I/O, if no type conversions are required, instead of just "regular" hyperslab and 'all' selections, which speeds up I/O. - Sped up I/O for hyperslab selections which have contiguous lower dimensions by "flattening" them out into lesser dimensional objects for the I/O. No file format or API changes were necessary for this change. The next phase will be to create a "selection driver" for each type of selection, allowing each type of selection to directly call certain methods that only apply to that type of selection, instead of passing through dozens of functions which have switch statements to call the appropriate method for each selection type. This will also reduce the amount of code in the library and speed things up a bit more. Phase 3 will involve generating an MPI datatype for all types of selections, instead of only "regular" hyperslab and 'all' selections. This will allow collective parallel I/O for all I/O operations which don't require type conversions. It will also open up the door for allowing collective I/O on datasets which require type conversion. Phase 4 will involve changing the access pattern to deal with chunked datasets in a more optimal way (in serial). Phase 5 will deal with accessing chunked datasets more optimally for collective parallel I/O operations. Platforms tested: FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
2002-07-24 13:56:48 -05:00
} /* end H5S_select_write() */
[svn-r5152] Purpose: New Feature Description: Added new H5Dfill() routine to fill the elements in a selection for a memory buffer with a fill value. This is a user API wrapper around some internal routines which were needed for the fill-value modifications from Raymond as well as Pedro's code for reducing the size of a chunked dataset. Platforms tested: FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few minutes]
2002-04-09 07:47:34 -05:00
Reference in New Issue Copy Permalink