Reviewed and revised version information for unix daily tested machines in supported a
d tested platforms sections.
Added description of LD_LIBRARY_PATH change to configure and using scripts to compile
nd run installed example files to Bug Fixes/Configuration section.
Text change only.
Description:
With optimization enabled, gcc 4.8 inserts garbage into the padding bytes of
floating point types when assigning from a literal constant. This caused
problems when H5detect.c scanned the bits in floating point types to determine
their properties.
Modified H5detect.c to scan for padding before further analyzing the type, and
to ignore all information in the padding areas. Also removed code that
temporarily disabled optimization.
Tested: jam, koala, ostrich, platypus (h5committest)
HDFFV-8214 - h5repack failed converting small chunked dataset (size < 1K) to contiguous layout.
Description:
h5repack failed converting small chunked dataset (size < 1K) to contiguous layout.
The first case was when chunk dim is bigger than the dataset dim (at leat one), h5repack failed with displaying error stacks.
The other case is when chunk dim is smaller than the dataset dim, h5repack failed to change layout.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), emu (solaris-BE),fred (mac64-LE), Windows (32-LE cmake), cmake (jam)
HDFFV-7643 - h5diff - incorrect exit-code returns for extra attribute exist
Description:
h5diff: Fixed to return correct exit code 1 when detect unique
attribute. Prior to this fix, h5diff returned exit code 0 indicating
two files are identical.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), fred (mac64-LE), emu (solaris-BE), Windows (32-LE cmake), cmake (jam)
Since this is a testing code error that does not affect the function of the
library or tools, I started a new "Testing" section for this sort of changes.
Description:
Adds 2 new API functions, H5Dscatter and H5Dgather. H5Dscatter retrieves data
from a specified callback function and scatters it into a selection, defined by
a supplied dataspace, within a supplied memory buffer. H5Dgather gathers data
from a selection within a supplied memory buffer and passes it in a contiguous
form to a supplied callback function. Added tests for these functions
Tested: jam, ostrich, koala (h5committest); ummon
HDFFV-8145 h5diff slowness with version >= 1.8.7
Description:
Improved speed when comparing HDF5 files with lots of attributes. Much slower performance was identified with release version from 1.8.7 to 1.8.10 compared to 1.8.6.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), Windows (32-LE cmake), cmake (jam)
HDFFV-8012 - h5repack changes max dims and cause failure if only "-f none" is used without changing layout for chunked dataset when a chunk dim is bigger than a dataset dim
Description:
"h5repack -f <obj>:NONE <file.h5> out.h5" command failed if source file contains chunked dataset and a chunk dim is bigger than a dataset dim.
Another issue is that the command changed max dims if chunk dim is smaller than the dataset dim.
These issue occurred when dataset size is smaller than 64k (compact size limit)
Fixed them.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), cmake (jam)
HDFFV-5919 - GMQS: h5diff - The number of object difference is not consistent between dataset and group/type when attribute(s) have differences
Description:
Object differences are not consistent between dataset and group/datatype when their attribute(s) have differences. This is because attribute(s) differences is not accumulated to group or datatype object’s difference, but accumulated to dataset difference.
To fix, do not accumulate attribute difference to dataset difference. This is referred to h5diff’s default behavior and also past report from users that users were confused by the accumulated behavior. (also can’t figure out only for dataset difference , also hard to spot dataset difference when it has lots of attributes or differences)
This also lead to fix inconsistent format indicating difference between dataset and group/datatype object.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), cmake (jam)
HDFFV-8143 Provide a routine(s) for telling the user why the library broke collective data access
Description:
Added H5Pget_mpio_no_collective_cause() function that retrive reasons why the collective I/O was broken during Read/Write IO access.
Reasons to break collective I/O:
- SET_INDEPENDENT
- DATATYPE_CONVERSION
- DATA_TRANSFORMS
- MPIPOSIX
- NOT_SIMPLE_OR_SCALAR_DATASPACES (NULL Space)
- POINT_SELECTIONS
- NOT_CONTIGUOUS_OR_CHUNKED_DATASET (Compact or External-Storage)
- FILTERS
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE)
Address HDFFV-7942 - h5diff: incorrect result for comparing attribute data with different type size in same class
Description:
When comparing attribute data values with same type class but different size, the result was incorrect. It was due to the size difference and got truncated.
Fixed to match up the smaller type size to big type size like what dataset does.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), cmake (jam)
Purpose:
HDFFV-8003 - ph5diff (parallel h5diff): daily test failure on ember intermittently during non comparable test file comparison
HDFFV-7755 - parallel h5diff : hanging on koala intermittently during non comparable test file comparison
Description:
non-comparable test intermittently hung on koala and ember, but not on jam. it didn't occur until -np reaches 4 or bigger. it occurred once out of many repeated attempts of the same test.
There was a incorrectly (mistakenly?) duplicated code in MPI section which caused such hang in a certain condition. The test used more processes than other tests, which increased chance to trigger more undone processes, and such process could enter the incorrect code section and wait for wrong pair of send. it explains why it occurred intermittently according to machine condition and using a certain feature.
Removed incorrect code which blocked correct code.
Tested: some manually repeated test performed
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE)
Fix for HDFFV-8107 testh5diff will fail if build/test in HDF5 source tree
Description:
This is sub-task for "HDFFV-8105 testh5diff.sh uses the wrong operator (-a) in an if statement."
From the HDFFV-8105's update, h5diff test failed if build&test is performed in HDF5 source tree because 'cp' try to copy test files to self dir.
It's addressed by skipping if cp's src dir and dest dir is same.
Also this applied for all other tools under src/tools dir.
No change to the CMakeLists.txt files because CMake cautions/demands that in-source builds be avoided.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), some manual tests as well
Description:
When jumping out from between H5_BEGIN_TAG and H5_END_TAG macros using
HGOTO_ERROR or HGOTO_DONE, the previous metadata tag is not reset on the dxpl.
This could cause problems when, for example, calling H5Ocopy within an
H5Literate callback.
Added new HGOTO_ERROR_TAG and HGOTO_DONE_TAG macros which must be used in place
of the above between H5_BEGIN_TAG and H5_END_TAG.
Tested:
jam, koala, ostrich (h5committest), durandal
Description:
The test command in COPY_TESTFILES_TO_TESTDIR used the wrong operator -a which
means and. It is not sufficient to test mere existence as $tstfile could have
been a directory or other non-file things. If the file is not readable, the
$CP command would fail but its return code was not checked.
Solution:
Get rid of the test command and go ahead to $CP it and then check its returned
code and report errors if detected.
Tested; committested. (note that ostrick still failed on errors committed by
other staff).
- Fixed a typo in return value of the nh5dread_f_c function ( was 1
instead of 0 on success); fixed the return value to make it consistent
with other Fortran functions; cleaned the code from debug statements.
Platforms tested: jam with GNU and Intel compilers, fortran 2003 feature.
Fix for HDFFV-7998-h5diff: incorrect behavior and output for comparing symbolic dangling links without following-symlinks option
Description:
Fix not to check and display dangling link status without --follow-symlinks option.
Berfor the fix, when comparing symbolic link to symbolic link without follow-symlinks option, h5diff followed to check if those links are dangling or not. It caused to display output incorrectly (dangling link instead of symbolic link).
The fix also improved performance when comparing lots of symbolic links without the --follow-symlinks option.
Test cases were added and tagged with jira#.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), Cmake (jam)
Fix for HDFFV-8001 - h5unjam: segfault when used -V (show version) option
Description:
h5unjam has an option '-V' to display version information.
"h5unjam -V" caused segfulat for 1.8.9 pre-release.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), Cmake (jam)
Fix for HDFFV-7993 - h5repack fails with error "chunk size must be <= maximum dimension size for fixed-sized dimensions"
Description:
Fixed a failure when change the chunk size of a specified chunked dataset with unlimited max dims.
Also took care of converting to contiguous and compact from the dataset.
Test cases were added and tagged with jira#.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), Cmake (jam)
could waste memory if any dataset size is smaller than the sieve buffer
size. Now the library picks the smaller one between the dataset size
and the sieve buffer size from the file access property.
Tested on jam, koala, and ostrich.
Fix for HDFFV-7644 GMQS: h5diff - incorrect behaviors when comparing HDF5 files with different type of objects (dataset, group, type) as common objects
Description:
Fixed failure for comparing same named object with different object types in comparing groups. (common object comparison)
Prior to the fix, h5diff resulted in error. After the fix, h5diff detects such case as non-comparable and display messages accordingly.
Test cases were added and tagged with jira#.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), Cmake (jam)
Description:
When copying an object with attribute creation order tracked, the attribute
creation order was not copied correctly to the destination file, causing an
error if the creation order was also indexed (due to attempting to insert
duplicate keys) or incorrect creation orders otherwise. Fixed to copy the
creation order correctly.
Also fixed the attribute character set not being copied, and fixed an issue
where an attribute opened with H5Aopen (or similar, but not by_idx), from an
object using the latest format but without creation order being tracked, would
always report the creation order as 0 (and marked as valid).
Tested: jam, koala, ostrich (h5committest), durandal
Bring "merge committed datatypes during H5Ocopy" feature from branch to
trunk. (Also has some minor bugfixes with it)
Tested on:
Mac OSX/64 10.7.3 (amazon) w/debug
(h5committest coming up)
HDFFV-944: Fortran test problem with HDF5 182, intel 11.0, open-mpi-1.2.8
Description:
Moved libtool post-processing 'ed' snippet that replaces wl=""
with wl="-Wl," when intel compilers are used to the end of the
configure script, after libtool is actually generated. (I imagine
that at some point in time during a libtool version upgrade the
libtool script ceased being generated immediately after the libtool
initialization macro and is generated at the end of configure with
all other files).
Additionally, added a leading 'g' to the replacement invocation to
change all instances of wl="" to wl="-Wl," instead of just the
first that it comes across. (There's a separate one for each
compiler used (C, Fortran, C++)).
Tested:
Reproduced the reported make failure and tested resolution on
NCSA's Forge with OpenMPI built with Intel 12.0.4.
Also, h5committested.
Fix for HDFFV-7837 - h5diff: incorrect behavior with exclude-path option when unique object exist only in one file
Description:
If unique objects exists only in one file and try to exclude
the unique objects with --exclude-path option, h5diff missed
excluding some objects.
Fixed to exclude objects correctly in such case.
Test cases were added and tagged with jira#.
Tested:
jam (linux32-LE), koala (linux64-LE), ostrich (linuxppc64-BE), tejeda (mac32-LE), linew (solaris-BE), Windows (32-LE cmake), Cmake (jam)
The parser mistaked "-" as substraction. I fixed it and also fixed another problem
with some special cases like 100-x and 2/x.
Tested on jam, koala, and ostrich.
Description:
When using the new object header format, it was possible for corruption to occur
if the first object header chunk changed size such that the lenght of the "chunk
0 size" field changed. This only occurred if there were messages that had not
been decoded. The original algorithm that changed the object header chunk size
marked all messages as dirty, causing those that had not been decoded to have
both the raw and native form invalidated. Changed the algorithm to avoid
marking messages dirty and added assertions to catch the case where messages
are dirtied without being decoded (or recently created) first.
Tested: jam, koala, ostrich (h5committest), durandal
