mirror/hdf5
2
0
Fork 0
mirror of https://github.com/HDFGroup/hdf5.git synced 2024-12-09 07:32:32 +08:00
Code Issues Packages Projects Releases Wiki Activity
d06b3afc6c
hdf5/src/H5Smpio.c

580 lines
19 KiB
C
Raw Normal View History

[svn-r597] Functions to convert HDF dataspace selections to MPI datatypes.
1998-08-17 23:13:45 +08:00
/*
* Copyright (C) 1998 NCSA
* All rights reserved.
*
* Programmer: rky 980813
*
* Purpose: Functions to read/write directly between app buffer and file.
*
* Beware of the ifdef'ed print statements.
* I didn't make them portable.
*/
#include <H5private.h>
#include <H5Eprivate.h>
#include <H5Sprivate.h>
[svn-r604] Added #ifdef HAVE_PARALLEL macro so that the code is compiled only in parallel version but an "empty" source in the serial version. Tested in solaris 2.5.
1998-08-19 22:07:14 +08:00
#ifndef HAVE_PARALLEL
/*
* The H5S_mpio_xxxx functions are for parallel I/O only and are
* valid only when HAVE_PARALLEL is #defined. This empty #ifndef
* body is used to allow this source file be included in the serial
* distribution.
* Some compilers/linkers may complain about "empty" object file.
* If that happens, uncomment the following statement to pacify
* them.
*/
/* const hbool_t H5S_mpio_avail = FALSE; */
[svn-r617] Added optimization to coalesce outer dims of hyperslab, and changed error return of H5S_mpio_spaces_xfer to simple FAIL in case file layout is not contiguous (so that the xfer will still be done, but in the old chunk-by-chunk method).
1998-08-27 01:51:19 +08:00
#else /* HAVE_PARALLEL */
[svn-r597] Functions to convert HDF dataspace selections to MPI datatypes.
1998-08-17 23:13:45 +08:00
/* Interface initialization */
#define PABLO_MASK H5S_all_mask
#define INTERFACE_INIT NULL
static intn interface_initialize_g = FALSE;
/*-------------------------------------------------------------------------
* Function: H5S_mpio_all_type
*
* Purpose: Translate an HDF5 "all" selection into an MPI type.
*
* Return: non-negative on success, negative on failure.
*
* Outputs: *new_type the MPI type corresponding to the selection
* *count how many objects of the new_type in selection
* (useful if this is the buffer type for xfer)
* *is_derived_type 0 if MPI primitive type, 1 if derived
*
* Programmer: rky 980813
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_mpio_all_type( const H5S_t *space, const hsize_t elmt_size,
/* out: */
MPI_Datatype *new_type,
hsize_t *count,
hbool_t *is_derived_type )
{
hsize_t total_bytes;
int i;
FUNC_ENTER (H5S_mpio_all_type, FAIL);
/* Check args */
assert (space);
/* Just treat the entire extent as a block of bytes */
total_bytes = elmt_size;
for (i=0; i<space->extent.u.simple.rank; ++i) {
total_bytes *= space->extent.u.simple.size[i];
}
/* fill in the return values */
*new_type = MPI_BYTE;
*count = total_bytes;
*is_derived_type = 0;
#ifdef H5Smpi_DEBUG
fprintf(stdout, "Leave %s total_bytes=%lld\n", FUNC, total_bytes );
#endif
FUNC_LEAVE (SUCCEED);
} /* H5S_mpio_all_type() */
/*-------------------------------------------------------------------------
* Function: H5S_mpio_hyper_type
*
* Purpose: Translate an HDF5 hyperslab selection into an MPI type.
*
* Return: non-negative on success, negative on failure.
*
* Outputs: *new_type the MPI type corresponding to the selection
* *count how many objects of the new_type in selection
* (useful if this is the buffer type for xfer)
* *is_derived_type 0 if MPI primitive type, 1 if derived
*
* Programmer: rky 980813
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_mpio_hyper_type( const H5S_t *space, const hsize_t elmt_size,
/* out: */
MPI_Datatype *new_type,
hsize_t *count,
hbool_t *is_derived_type )
{
struct dim { /* less hassle than malloc/free & ilk */
hssize_t start;
hsize_t strid;
hsize_t block;
hsize_t xtent;
hsize_t count;
} d[32];
int i, err, new_rank, num_to_collapse;
H5S_hyper_dim_t *diminfo; /* [rank] */
intn rank;
MPI_Datatype inner_type, outer_type;
MPI_Aint s[2]; /* array of displacements for struct type */
MPI_Aint extent_len, start_Aint; /* for calculating d[1] */
FUNC_ENTER (H5S_mpio_hyper_type, FAIL);
/* Check and abbreviate args */
assert (space);
diminfo = space->select.sel_info.hyper.diminfo;
assert (diminfo);
rank = space->extent.u.simple.rank;
assert (rank >= 0);
/* make a local copy of the dimension info so we can transform them */
#ifdef H5Smpi_DEBUG
fprintf(stdout, "rank=%d ", rank );
#endif
assert(rank<=32); /* within array bounds */
for ( i=0; i<rank; ++i) {
d[i].start = diminfo[i].start;
d[i].strid = diminfo[i].stride;
d[i].block = diminfo[i].block;
d[i].count = diminfo[i].count;
d[i].xtent = space->extent.u.simple.size[i];
#ifdef H5Smpi_DEBUG
fprintf(stdout,
"hyper_type: start=%lld count=%lld stride=%lld block=%lld xtent=%lld\n",
d[i].start, d[i].count, d[i].strid, d[i].block, d[i].xtent );
#endif
}
/* Create a type covering the selected hyperslab.
* Multidimensional dataspaces are stored in row-major order.
* The type is built from the inside out, going from the
* fastest-changing (i.e., inner) dimension * to the slowest (outer). */
/* Optimization: check for contiguous inner dimensions.
* Supposing the dimensions were numbered from 1 to rank, we find that
*
* dim d=rank is contiguous if:
* stride[d] = block[d]
* and count[d] * block[d] = extent.u.simple.size[d]
*
* (i.e., there's no overlap or gaps and the entire extent is filled.)
*
* dim d (1<=d<rank) is contiguous if:
* dim d+1 is contiguous
* and stride[d] = block[d]
* and count[d] * block[d] = extent.u.simple.size[d]
*
* There is also a weak sense in which the first noncollapsible dim
* is contiguous if it consists of a single unbroken range,
* and we also take advantage of that.
*/
/* figure out how many dimensions we can eliminate */
/* This loop examines contiguity from the inside out. */
for ( i=0; i<rank; ++i) {
if ((d[rank-i].strid != d[rank-i].block)
||
(d[rank-i].count*d[rank-i].block) != space->extent.u.simple.size[rank-i]) {
break;
}
} /* end for */
num_to_collapse = (i)? i-1: 0;
assert(0<=num_to_collapse && num_to_collapse<rank);
new_rank = rank - num_to_collapse;
#ifdef H5Smpi_DEBUG
fprintf(stdout, "hyper_type: new_rank=%d\n", new_rank );
#endif
/* To collapse dims, we only need to transform the dimension info */
for (i=0; i<num_to_collapse; ++i) {
d[rank-i-1].block *= d[rank-i].strid;
d[rank-i-1].strid *= d[rank-i].strid;
d[rank-i-1].xtent *= d[rank-i].strid;
assert( d[rank-i].start == 0 );
/* d[rank-i-1].start stays unchanged */
/* d[rank-i-1].count stays unchanged */
}
[svn-r617] Added optimization to coalesce outer dims of hyperslab, and changed error return of H5S_mpio_spaces_xfer to simple FAIL in case file layout is not contiguous (so that the xfer will still be done, but in the old chunk-by-chunk method).
1998-08-27 01:51:19 +08:00
/* check for possibility to coalesce blocks of the uncoalesced dimensions */
for (i=0; i<new_rank; ++i) {
if (d[i].strid == d[i].block) {
/* transform smaller blocks to 1 larger block of combined size */
d[i].block *= d[i].count;
d[i].count = 1;
}
[svn-r597] Functions to convert HDF dataspace selections to MPI datatypes.
1998-08-17 23:13:45 +08:00
}
/* initialize induction variables */
s[0] = 0; /* stays constant */
/* create contig type for inner contig dims */
#ifdef H5Smpi_DEBUG
fprintf(stdout, "hyper_type: Making contig type %lld MPI_BYTEs\n", elmt_size );
#endif
err = MPI_Type_contiguous( elmt_size, MPI_BYTE, &inner_type );
if (err) {
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,"couldn't create MPI contiguous type");
}
/* construct the type by walking the hyperslab dims from the inside out */
for ( i=new_rank-1; i>=0; --i) {
#ifdef H5Smpi_DEBUG
fprintf(stdout, "hyper_type: i=%d Making vector type\n count=%lld block=%lld stride=%lld\n", i, d[i].count, d[i].block, d[i].strid );
#endif
err = MPI_Type_vector( d[i].count, /* count */
d[i].block, /* blocklength */
d[i].strid, /* stride */
inner_type, /* old type */
&outer_type ); /* new type */
if (err) {
MPI_Type_free( &inner_type ); /* free before abort */
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,"couldn't create MPI vector type");
}
/* from here to end of loop, inner_type actually will get the value
* of the outermost type: it will be inner for the next iteration */
if (0 == d[i].start) {
/* don't need to compensate for the start displacement */
MPI_Type_free( &inner_type ); /* old inner no longer needed */
inner_type = outer_type; /* prepare for next iter */
} else {
/* need to compensate for the start displacement */
int b[2]; /* array of rep counts */
MPI_Datatype t[2]; /* array of MPI types */
/* fill in the b, d, and t arrays, length is 2 */
/* b gives rep count for each type in t */
b[0] = 1;
b[1] = 1;
/* s gives the byte displacement for each "field" by induction:
* for 0<=i<rank-1 s[1]_i = start[i]*extent_length[i+1];
* with base case s[1]_(rank-1) = elmt_size (i decreasing).
* (Assuming dimension index increases as we go deeper in.)
* Note that in this loop, extent_length[i+1] is the extent length
* of the inner type (i.e., the type constructed in previous trip).
*/
err = MPI_Type_extent( inner_type, &extent_len );
if (err) {
MPI_Type_free( &inner_type ); /* free before abort */
MPI_Type_free( &outer_type ); /* free before abort */
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,
"couldn't get extent of MPI type");
}
start_Aint = (MPI_Aint)(d[i].start);
if (start_Aint != d[i].start) {
MPI_Type_free( &inner_type ); /* free before abort */
MPI_Type_free( &outer_type ); /* free before abort */
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,
"start value overflows MPI_Aint");
}
s[1] = start_Aint * extent_len;
/* t gives the MPI types for the "fields" */
/* I think we could do without the LB type and just have
* one "field" consisting of the vector type with displacement,
* but I guess there's no harm in doing it this way. */
t[0] = MPI_LB;
t[1] = outer_type; /* the just-created vector type */
/* Create new struct type to compensate for start displacement.
* The struct's first "field" is the displacement,
* and its second "field" is the just-created vector type */
#ifdef H5Smpi_DEBUG
fprintf(stdout, "hyper_type: i=%d Making struct type\n b[1]=%d d[1]=%lld\n", i, b[1], (long long)s[1] );
#endif
err = MPI_Type_struct( 2, b, s, t, &inner_type/*becomes outer*/ );
MPI_Type_free( &outer_type ); /* no longer needed */
if (err) {
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,"couldn't create MPI struct type");
}
} /* end else */
/* at this point, inner_type is actually the outermost type */
} /* end for */
/* here inner_type is actually the outermost type, even for 0-trip loop */
*new_type = inner_type; /* return the just-constructed type */
err = MPI_Type_commit( new_type );
if (err) {
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,"couldn't commit MPI vector type");
}
/* fill in the remaining return values */
*count = 1; /* only have to move one of these suckers! */
*is_derived_type = 1;
#ifdef H5Smpi_DEBUG
fprintf(stdout, "Leave %s\n", FUNC );
#endif
FUNC_LEAVE (SUCCEED);
} /* H5S_mpio_hyper_type() */
/*-------------------------------------------------------------------------
* Function: H5S_mpio_space_type
*
* Purpose: Translate an HDF5 dataspace selection into an MPI type.
* Currently handle only hyperslab and "all" selections.
*
* Return: non-negative on success, negative on failure.
*
* Outputs: *new_type the MPI type corresponding to the selection
* *count how many objects of the new_type in selection
* (useful if this is the buffer type for xfer)
* *is_derived_type 0 if MPI primitive type, 1 if derived
*
* Programmer: rky 980813
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_mpio_space_type( const H5S_t *space, const size_t elmt_size,
/* out: */
MPI_Datatype *new_type,
hsize_t *count,
hbool_t *is_derived_type )
{
int err;
herr_t ret_value = SUCCEED;
FUNC_ENTER (H5S_mpio_space_type, FAIL);
/* Check args */
assert (space);
/* Creat MPI type based on the kind of selection */
switch (space->extent.type) {
case H5S_SCALAR:
/* not yet implemented */
ret_value = FAIL;
break;
case H5S_SIMPLE:
switch(space->select.type) {
case H5S_SEL_NONE:
case H5S_SEL_ALL:
err = H5S_mpio_all_type( space, elmt_size,
/* out: */ new_type, count, is_derived_type );
if (err<0) {
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't convert \"all\" selection to MPI type");
}
break;
case H5S_SEL_POINTS:
/* not yet implemented */
ret_value = FAIL;
break;
case H5S_SEL_HYPERSLABS:
err = H5S_mpio_hyper_type( space, elmt_size,
/* out: */ new_type, count, is_derived_type );
if (err) {
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't convert \"all\" selection to MPI type");
}
break;
default:
assert("unknown selection type" && 0);
break;
} /* end switch */
break;
case H5S_COMPLEX:
/* not yet implemented */
ret_value = FAIL;
break;
default:
assert("unknown data space type" && 0);
break;
}
FUNC_LEAVE (ret_value);
} /* H5S_mpio_space_type() */
/*-------------------------------------------------------------------------
* Function: H5S_mpio_spaces_xfer
*
* Purpose: Use MPI-IO to transfer data efficiently
* directly between app buffer and file.
*
* Return: non-negative on success, negative on failure.
*
* Programmer: rky 980813
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_mpio_spaces_xfer (H5F_t *f, const struct H5O_layout_t *layout,
const struct H5O_pline_t *pline,
const struct H5O_efl_t *efl, size_t elmt_size,
const H5S_t *file_space, const H5S_t *mem_space,
const H5D_transfer_t xfer_mode, void *buf /*out*/,
const hbool_t do_write )
{
[svn-r617] Added optimization to coalesce outer dims of hyperslab, and changed error return of H5S_mpio_spaces_xfer to simple FAIL in case file layout is not contiguous (so that the xfer will still be done, but in the old chunk-by-chunk method).
1998-08-27 01:51:19 +08:00
herr_t ret_value = SUCCEED;
[svn-r597] Functions to convert HDF dataspace selections to MPI datatypes.
1998-08-17 23:13:45 +08:00
int err;
haddr_t disp, addr;
size_t mpi_count;
hsize_t mpi_buf_count, mpi_unused_count;
hsize_t elmt_hsize;
MPI_Datatype mpi_buf_type, mpi_file_type;
hbool_t mbt_is_derived, mft_is_derived;
FUNC_ENTER (H5S_mpio_spaces_xfer, FAIL);
/* Check args */
assert (f);
assert (layout);
assert (file_space);
assert (mem_space);
assert (buf);
assert (f->shared->access_parms->driver == H5F_LOW_MPIO);
/* INCOMPLETE!!! rky 980816 */
/* Currently can only handle H5D_CONTIGUOUS layout */
if (layout->type != H5D_CONTIGUOUS) {
[svn-r617] Added optimization to coalesce outer dims of hyperslab, and changed error return of H5S_mpio_spaces_xfer to simple FAIL in case file layout is not contiguous (so that the xfer will still be done, but in the old chunk-by-chunk method).
1998-08-27 01:51:19 +08:00
#ifdef H5S_DEBUG
fprintf (stderr, "H5S: can only create MPI datatype for hyperslab selection when layout is contiguous.\n" );
#endif
ret_value = FAIL;
goto done;
[svn-r597] Functions to convert HDF dataspace selections to MPI datatypes.
1998-08-17 23:13:45 +08:00
}
/* create the MPI buffer type */
elmt_hsize = (hsize_t)elmt_size;
if (elmt_hsize != elmt_size)
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"element size overflows hsize_t");
err = H5S_mpio_space_type( mem_space, elmt_size,
/* out: */
&mpi_buf_type,
&mpi_buf_count,
&mbt_is_derived );
if (MPI_SUCCESS != err)
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI buf type");
/* pass the buf type to low-level write via access_parms */
f->shared->access_parms->u.mpio.btype = mpi_buf_type;
/* create the MPI file type */
err = H5S_mpio_space_type( file_space, elmt_size,
/* out: */
&mpi_file_type,
&mpi_unused_count,
&mft_is_derived );
if (MPI_SUCCESS != err)
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI file type");
/* pass the file type to low-level write via access_parms */
f->shared->access_parms->u.mpio.ftype = mpi_file_type;
/* calculate the absolute base addr (i.e., the file view disp) */
disp = f->shared->base_addr;
H5F_addr_add( &disp, &(layout->addr) );
f->shared->access_parms->u.mpio.disp = disp;
#ifdef H5Smpi_DEBUG
fprintf(stdout, "spaces_xfer: disp=%lld\n", disp.offset );
#endif
/* Effective address determined by base addr and the MPI file type */
H5F_addr_reset( &addr ); /* set to 0 */
/* request a dataspace xfer (instead of an elementary byteblock xfer) */
f->shared->access_parms->u.mpio.use_types = 1;
/* transfer the data */
mpi_count = (size_t)mpi_buf_count;
if (mpi_count != mpi_buf_count)
HRETURN_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"transfer size overflows size_t");
if (do_write) {
err = H5F_low_write( f->shared->lf, f->shared->access_parms,
xfer_mode, &addr, mpi_count, buf );
if (err) HRETURN_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,"MPI write failed");
} else {
err = H5F_low_read ( f->shared->lf, f->shared->access_parms,
xfer_mode, &addr, mpi_count, buf );
if (err) HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL,"MPI read failed");
}
/* free the MPI buf and file types */
if (mbt_is_derived) {
err = MPI_Type_free( &mpi_buf_type );
if (MPI_SUCCESS != err) {
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,"couldn't free MPI file type");
}
}
if (mft_is_derived) {
err = MPI_Type_free( &mpi_file_type );
if (MPI_SUCCESS != err) {
HRETURN_ERROR(H5E_DATASPACE, H5E_MPI, FAIL,"couldn't free MPI file type");
}
}
[svn-r617] Added optimization to coalesce outer dims of hyperslab, and changed error return of H5S_mpio_spaces_xfer to simple FAIL in case file layout is not contiguous (so that the xfer will still be done, but in the old chunk-by-chunk method).
1998-08-27 01:51:19 +08:00
done:
FUNC_LEAVE (ret_value);
[svn-r597] Functions to convert HDF dataspace selections to MPI datatypes.
1998-08-17 23:13:45 +08:00
} /* H5S_mpio_spaces_xfer() */
/*-------------------------------------------------------------------------
* Function: H5S_mpio_spaces_read
*
* Purpose: MPI-IO function to read directly from app buffer to file.
*
* Return: non-negative on success, negative on failure.
*
* Programmer: rky 980813
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_mpio_spaces_read (H5F_t *f, const struct H5O_layout_t *layout,
const struct H5O_pline_t *pline,
const struct H5O_efl_t *efl, size_t elmt_size,
const H5S_t *file_space, const H5S_t *mem_space,
const H5D_transfer_t xfer_mode, void *buf /*out*/ )
{
herr_t ret_value = FAIL;
FUNC_ENTER (H5S_mpio_spaces_read, FAIL);
ret_value = H5S_mpio_spaces_xfer( f, layout, pline, efl, elmt_size,
file_space, mem_space, xfer_mode, (void*)buf,
0 /*read*/ );
FUNC_LEAVE (ret_value);
} /* H5S_mpio_spaces_read() */
/*-------------------------------------------------------------------------
* Function: H5S_mpio_spaces_write
*
* Purpose: MPI-IO function to write directly from app buffer to file.
*
* Return: non-negative on success, negative on failure.
*
* Programmer: rky 980813
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5S_mpio_spaces_write(H5F_t *f, const struct H5O_layout_t *layout,
const struct H5O_pline_t *pline,
const struct H5O_efl_t *efl, size_t elmt_size,
const H5S_t *file_space, const H5S_t *mem_space,
const H5D_transfer_t xfer_mode, const void *buf )
{
herr_t ret_value = FAIL;
FUNC_ENTER (H5S_mpio_spaces_write, FAIL);
ret_value = H5S_mpio_spaces_xfer( f, layout, pline, efl, elmt_size,
file_space, mem_space, xfer_mode, (void*)buf,
1 /*write*/ );
FUNC_LEAVE (ret_value);
} /* H5S_mpio_spaces_write() */
[svn-r604] Added #ifdef HAVE_PARALLEL macro so that the code is compiled only in parallel version but an "empty" source in the serial version. Tested in solaris 2.5.
1998-08-19 22:07:14 +08:00
[svn-r617] Added optimization to coalesce outer dims of hyperslab, and changed error return of H5S_mpio_spaces_xfer to simple FAIL in case file layout is not contiguous (so that the xfer will still be done, but in the old chunk-by-chunk method).
1998-08-27 01:51:19 +08:00
#endif /* HAVE_PARALLEL */
Reference in New Issue Copy Permalink