Files
hdf5/src/H5Farray.c
Quincey Koziol 990fadfbe5 [svn-r4181] Purpose:
Bug Fix, Code Cleanup, Code Optimization, etc.
Description:
    Fold in the hyperslab speedups, clean up compile warnings and change a
    few things from using 'unsigned' or 'hsize_t' to use 'size_t' instead.
Platforms tested:
    FreeBSD 4.3 (hawkwind), Solaris 2.7 (arabica), Irix64 6.5 (modi4)
2001-07-10 16:19:18 -05:00

641 lines
24 KiB
C
Raw Blame History

This file contains invisible Unicode characters
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
/*
* Copyright (C) 1998-2001 NCSA
* All rights reserved.
*
* Programmer: Robb Matzke <matzke@llnl.gov>
* Thursday, January 15, 1998
*
* Purpose: Provides I/O facilities for multi-dimensional arrays of bytes
* stored with various layout policies. If the caller is
* interested in arrays of elements >1 byte then add an extra
* dimension. For example, a 10x20 array of int would
* translate to a 10x20x4 array of bytes at this level.
*/
#define H5F_PACKAGE /*suppress error about including H5Fpkg */
#include "H5private.h"
#include "H5Dprivate.h"
#include "H5Eprivate.h"
#include "H5Fpkg.h"
#include "H5FDprivate.h" /*file driver */
#include "H5Iprivate.h"
#include "H5MFprivate.h"
#include "H5MMprivate.h" /*memory management */
#include "H5Oprivate.h"
#include "H5Pprivate.h"
#include "H5Vprivate.h"
/* MPIO driver functions are needed for some special checks */
#include "H5FDmpio.h"
/* Interface initialization */
#define PABLO_MASK H5Farray_mask
#define INTERFACE_INIT NULL
static intn interface_initialize_g = 0;
/*-------------------------------------------------------------------------
* Function: H5F_arr_create
*
* Purpose: Creates an array of bytes.
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Robb Matzke
* Friday, January 16, 1998
*
* Modifications:
*
*-------------------------------------------------------------------------
*/
herr_t
H5F_arr_create (H5F_t *f, struct H5O_layout_t *layout/*in,out*/)
{
uintn u;
hsize_t nbytes;
FUNC_ENTER (H5F_arr_create, FAIL);
/* check args */
assert (f);
assert (layout);
layout->addr = HADDR_UNDEF; /*just in case we fail*/
switch (layout->type) {
case H5D_CONTIGUOUS:
/* Reserve space in the file for the entire array */
for (u=0, nbytes=1; u<layout->ndims; u++)
nbytes *= layout->dim[u];
assert (nbytes>0);
if (HADDR_UNDEF==(layout->addr=H5MF_alloc(f, H5FD_MEM_DRAW, nbytes))) {
HRETURN_ERROR (H5E_IO, H5E_NOSPACE, FAIL,
"unable to reserve file space");
}
break;
case H5D_CHUNKED:
/* Create the root of the B-tree that describes chunked storage */
if (H5F_istore_create (f, layout/*out*/)<0) {
HRETURN_ERROR (H5E_IO, H5E_CANTINIT, FAIL,
"unable to initialize chunked storage");
}
break;
default:
assert ("not implemented yet" && 0);
HRETURN_ERROR (H5E_IO, H5E_UNSUPPORTED, FAIL,
"unsupported storage layout");
}
FUNC_LEAVE (SUCCEED);
}
/*-------------------------------------------------------------------------
* Function: H5F_arr_read
*
* Purpose: Reads a hyperslab of a file byte array into a hyperslab of
* a byte array in memory. The data is read from file F and the
* array's size and storage information is in LAYOUT. External
* files are described according to the external file list, EFL.
* The hyperslab offset is FILE_OFFSET[] in the file and
* MEM_OFFSET[] in memory (offsets are relative to the origin of
* the array) and the size of the hyperslab is HSLAB_SIZE[]. The
* total size of the file array is implied in the LAYOUT
* argument and the total size of the memory array is
* MEM_SIZE[]. The dimensionality of these vectors is implied by
* the LAYOUT argument.
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Robb Matzke
* Friday, January 16, 1998
*
* Modifications:
* Albert Cheng, 1998-06-02
* Added xfer_mode argument
*
* Robb Matzke, 1998-09-28
* Added the `xfer' argument and removed the `xfer_mode'
* argument since it's a field of `xfer'.
*
* Robb Matzke, 1999-08-02
* Data transfer properties are passed by ID since that's how
* the virtual file layer wants them.
*-------------------------------------------------------------------------
*/
herr_t
H5F_arr_read(H5F_t *f, hid_t dxpl_id, const struct H5O_layout_t *layout,
const struct H5O_pline_t *pline, const H5O_fill_t *fill,
const struct H5O_efl_t *efl, const hsize_t _hslab_size[],
const hsize_t mem_size[], const hssize_t mem_offset[],
const hssize_t file_offset[], void *_buf/*out*/)
{
uint8_t *buf = (uint8_t*)_buf; /*cast for arithmetic */
hssize_t file_stride[H5O_LAYOUT_NDIMS]; /*strides through file */
hssize_t mem_stride[H5O_LAYOUT_NDIMS]; /*strides through memory*/
hsize_t hslab_size[H5O_LAYOUT_NDIMS]; /*hyperslab size */
hsize_t idx[H5O_LAYOUT_NDIMS]; /*multi-dim counter */
size_t mem_start; /*byte offset to start */
hsize_t file_start; /*byte offset to start */
hsize_t max_data = 0; /*bytes in dataset */
hsize_t elmt_size = 1; /*bytes per element */
size_t nelmts, z; /*number of elements */
uintn ndims; /*stride dimensionality */
haddr_t addr; /*address in file */
intn j; /*counters */
uintn u; /*counters */
hbool_t carray; /*carry for subtraction */
#ifdef H5_HAVE_PARALLEL
H5FD_mpio_xfer_t xfer_mode=H5FD_MPIO_INDEPENDENT;
#endif
#ifdef COALESCE_READS
H5D_xfer_t *xfer_parms; /*transfer property list*/
#endif
FUNC_ENTER(H5F_arr_read, FAIL);
/* Check args */
assert(f);
assert(layout);
assert(_hslab_size);
assert(file_offset);
assert(mem_offset);
assert(mem_size);
assert(buf);
/* Make a local copy of size so we can modify it */
H5V_vector_cpy(layout->ndims, hslab_size, _hslab_size);
#ifdef H5_HAVE_PARALLEL
{
/* Get the transfer mode */
H5D_xfer_t *dxpl;
H5FD_mpio_dxpl_t *dx;
if (H5P_DEFAULT!=dxpl_id && (dxpl=H5I_object(dxpl_id)) &&
H5FD_MPIO==dxpl->driver_id && (dx=dxpl->driver_info) &&
H5FD_MPIO_INDEPENDENT!=dx->xfer_mode) {
xfer_mode = dx->xfer_mode;
}
}
#endif
#ifdef H5_HAVE_PARALLEL
/* Collective MPIO access is unsupported for non-contiguous datasets */
if (H5D_CONTIGUOUS!=layout->type && H5FD_MPIO_COLLECTIVE==xfer_mode) {
HRETURN_ERROR (H5E_DATASET, H5E_READERROR, FAIL,
"collective access on non-contiguous datasets not "
"supported yet");
}
#endif
#ifdef QAK
{
extern int qak_debug;
if(qak_debug) {
printf("%s: layout->ndims=%d\n",FUNC,(int)layout->ndims);
for(u=0; u<layout->ndims; u++)
printf("%s: %u: hslab_size=%d, mem_size=%d, mem_offset=%d, file_offset=%d\n",FUNC,u,(int)_hslab_size[u],(int)mem_size[u],(int)mem_offset[u],(int)file_offset[u]);
printf("%s: *buf=%d, *(buf+1)=%d\n", FUNC,(int)*(const uint16_t *)buf,(int)*((const uint16 *)buf+1));
}
}
#endif /* QAK */
switch (layout->type) {
case H5D_CONTIGUOUS:
ndims = layout->ndims;
/*
* Offsets must not be negative for this type of storage.
*/
for (u=0; u<ndims; u++) {
if (mem_offset[u]<0 || file_offset[u]<0) {
HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL,
"negative offsets are not valid");
}
}
/*
* Filters cannot be used for contiguous data.
*/
if (pline && pline->nfilters>0) {
HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL,
"filters are not allowed for contiguous data");
}
/*
* Calculate the strides needed to walk through the array on disk
* and memory. Optimize the strides to result in the fewest number of
* I/O requests.
*/
mem_start = H5V_hyper_stride(ndims, hslab_size, mem_size,
mem_offset, mem_stride/*out*/);
file_start = H5V_hyper_stride(ndims, hslab_size, layout->dim,
file_offset, file_stride/*out*/);
H5V_stride_optimize2(&ndims, &elmt_size, hslab_size,
mem_stride, file_stride);
/*
* Initialize loop variables. The loop is a multi-dimensional loop
* that counts from SIZE down to zero and IDX is the counter. Each
* element of IDX is treated as a digit with IDX[0] being the least
* significant digit.
*/
H5V_vector_cpy(ndims, idx, hslab_size);
nelmts = H5V_vector_reduce_product(ndims, hslab_size);
if (efl && efl->nused>0) {
addr = 0;
} else {
addr = layout->addr;
/* Compute the size of the dataset in bytes */
for(u=0, max_data=1; u<layout->ndims; u++)
max_data *= layout->dim[u];
/* Adjust the maximum size of the data by the offset into it */
max_data -= file_start;
}
addr += file_start;
buf += mem_start;
/*
* Now begin to walk through the array, copying data from disk to
* memory.
*/
#ifdef H5_HAVE_PARALLEL
if (H5FD_MPIO_COLLECTIVE==xfer_mode){
/*
* Currently supports same number of collective access. Need to
* be changed LATER to combine all reads into one collective MPIO
* call.
*/
unsigned long max, min, temp;
temp = nelmts;
assert(temp==nelmts); /* verify no overflow */
MPI_Allreduce(&temp, &max, 1, MPI_UNSIGNED_LONG, MPI_MAX,
H5FD_mpio_communicator(f->shared->lf));
MPI_Allreduce(&temp, &min, 1, MPI_UNSIGNED_LONG, MPI_MIN,
H5FD_mpio_communicator(f->shared->lf));
#ifdef AKC
printf("nelmts=%lu, min=%lu, max=%lu\n", temp, min, max);
#endif
if (max != min)
HRETURN_ERROR(H5E_DATASET, H5E_READERROR, FAIL,
"collective access with unequal number of "
"blocks not supported yet");
}
#endif
#ifdef COALESCE_READS
/* Get the dataset transfer property list */
if (H5P_DEFAULT == dxpl_id) {
xfer_parms = &H5D_xfer_dflt;
} else if (H5P_DATASET_XFER != H5P_get_class (dxpl_id) ||
NULL == (xfer_parms = H5I_object (dxpl_id))) {
HRETURN_ERROR (H5E_ARGS, H5E_BADTYPE, FAIL, "not xfer parms");
}
for (z=0, xfer_parms->gather_reads = nelmts - 1;
z<nelmts;
z++, xfer_parms->gather_reads--) {
#else
#ifdef QAK
printf("%s: nelmts=%d, addr=%lu, elmt_size=%lu\n",FUNC,(int)nelmts,(unsigned long)addr,(unsigned long)elmt_size);
printf("%s: sieve_buf=%p, sieve_loc=%lu, sieve_size=%lu, sieve_buf_size=%lu, sieve_dirty=%u\n",FUNC,f->shared->sieve_buf,(unsigned long)f->shared->sieve_loc,(unsigned long)f->shared->sieve_size,(unsigned long)f->shared->sieve_buf_size,(unsigned)f->shared->sieve_dirty);
printf("%s: feature_flags=%lx\n",FUNC,(unsigned long)f->shared->lf->feature_flags);
#endif /* QAK */
for (z=0; z<nelmts; z++) {
#endif
/* Read directly from file if the dataset is in an external file */
/* Note: We can't use data sieve buffers for datasets in external files
* because the 'addr' of all external files is set to 0 (above) and
* all datasets in external files would alias to the same set of
* file offsets, totally mixing up the data sieve buffer information. -QAK
*/
if (efl && efl->nused>0) {
H5_CHECK_OVERFLOW(elmt_size,hsize_t,size_t);
if (H5O_efl_read(f, efl, addr, (size_t)elmt_size, buf)<0) {
HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL,
"external data read failed");
}
} else {
H5_CHECK_OVERFLOW(elmt_size,hsize_t,size_t);
if (H5F_contig_read(f, max_data, H5FD_MEM_DRAW, addr, (size_t)elmt_size, dxpl_id, buf)<0) {
HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL,
"block read failed");
}
} /* end else */
/* Decrement indices and advance pointers */
for (j=ndims-1, carray=TRUE; j>=0 && carray; --j) {
addr += file_stride[j];
buf += mem_stride[j];
/* Adjust the maximum size of the data by the offset into it */
max_data -= file_stride[j];
if (--idx[j])
carray = FALSE;
else
idx[j] = hslab_size[j];
}
}
break;
case H5D_CHUNKED:
/*
* This method is unable to access external raw data files or to copy
* into a proper hyperslab.
*/
if (efl && efl->nused>0) {
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL,
"chunking and external files are mutually exclusive");
}
for (u=0; u<layout->ndims; u++) {
if (0!=mem_offset[u] || hslab_size[u]!=mem_size[u]) {
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL,
"unable to copy into a proper hyperslab");
}
}
if (H5F_istore_read(f, dxpl_id, layout, pline, fill, file_offset,
hslab_size, buf)<0) {
HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL, "chunked read failed");
}
break;
default:
assert("not implemented yet" && 0);
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL,
"unsupported storage layout");
}
FUNC_LEAVE(SUCCEED);
}
/*-------------------------------------------------------------------------
* Function: H5F_arr_write
*
* Purpose: Copies a hyperslab of a memory array to a hyperslab of a
* file array. The data is written to file F and the file
* array's size and storage information is implied by LAYOUT.
* The data is stored in external files according to the
* external file list, EFL. The hyperslab offset is
* FILE_OFFSET[] in the file and MEM_OFFSET[] in memory (offsets
* are relative to the origin of the array) and the size of the
* hyperslab is HSLAB_SIZE[]. The total size of the file array
* is implied by the LAYOUT argument and the total size of the
* memory array is MEM_SIZE[]. The dimensionality of these
* vectors is implied by the LAYOUT argument.
*
* Return: Non-negative on success/Negative on failure
*
* Programmer: Robb Matzke
* Friday, January 16, 1998
*
* Modifications:
* Albert Cheng, 1998-06-02
* Added xfer_mode argument
*
* Robb Matzke, 1998-09-28
* Added `xfer' argument, removed `xfer_mode' argument since it
* is a member of H5D_xfer_t.
*
* Robb Matzke, 1999-08-02
* Data transfer properties are passed by ID since that's how
* the virtual file layer wants them.
*-------------------------------------------------------------------------
*/
herr_t
H5F_arr_write(H5F_t *f, hid_t dxpl_id, const struct H5O_layout_t *layout,
const struct H5O_pline_t *pline,
const struct H5O_fill_t *fill, const struct H5O_efl_t *efl,
const hsize_t _hslab_size[], const hsize_t mem_size[],
const hssize_t mem_offset[], const hssize_t file_offset[],
const void *_buf)
{
const uint8_t *buf = (const uint8_t *)_buf; /*cast for arithmetic */
hssize_t file_stride[H5O_LAYOUT_NDIMS]; /*strides through file */
hssize_t mem_stride[H5O_LAYOUT_NDIMS]; /*strides through memory*/
hsize_t hslab_size[H5O_LAYOUT_NDIMS]; /*hyperslab size */
hsize_t idx[H5O_LAYOUT_NDIMS]; /*multi-dim counter */
hsize_t mem_start; /*byte offset to start */
hsize_t file_start; /*byte offset to start */
hsize_t max_data = 0; /*bytes in dataset */
hsize_t elmt_size = 1; /*bytes per element */
size_t nelmts, z; /*number of elements */
uintn ndims; /*dimensionality */
haddr_t addr; /*address in file */
intn j; /*counters */
uintn u; /*counters */
hbool_t carray; /*carry for subtraction */
#ifdef H5_HAVE_PARALLEL
H5FD_mpio_xfer_t xfer_mode=H5FD_MPIO_INDEPENDENT;
#endif
FUNC_ENTER(H5F_arr_write, FAIL);
/* Check args */
assert(f);
assert(layout);
assert(_hslab_size);
assert(file_offset);
assert(mem_offset);
assert(mem_size);
assert(buf);
/* Make a local copy of _size so we can modify it */
H5V_vector_cpy(layout->ndims, hslab_size, _hslab_size);
#ifdef H5_HAVE_PARALLEL
{
/* Get the transfer mode */
H5D_xfer_t *dxpl;
H5FD_mpio_dxpl_t *dx;
if (H5P_DEFAULT!=dxpl_id && (dxpl=H5I_object(dxpl_id)) &&
H5FD_MPIO==dxpl->driver_id && (dx=dxpl->driver_info) &&
H5FD_MPIO_INDEPENDENT!=dx->xfer_mode) {
xfer_mode = dx->xfer_mode;
}
}
#endif
#ifdef H5_HAVE_PARALLEL
if (H5D_CONTIGUOUS!=layout->type && H5FD_MPIO_COLLECTIVE==xfer_mode) {
HRETURN_ERROR (H5E_DATASET, H5E_WRITEERROR, FAIL,
"collective access on non-contiguous datasets not "
"supported yet");
}
#endif
#ifdef QAK
{
extern int qak_debug;
printf("%s: layout->ndims=%d\n",FUNC,(int)layout->ndims);
for(i=0; i<layout->ndims; i++)
printf("%s: %d: hslab_size=%d, mem_size=%d, mem_offset=%d, "
"file_offset=%d\n", FUNC, i, (int)_hslab_size[i],
(int)mem_size[i],(int)mem_offset[i],(int)file_offset[i]);
if(qak_debug) {
printf("%s: *buf=%d, *(buf+1)=%d\n", FUNC,
(int)*(const uint16_t *)buf, (int)*((const uint16_t *)buf+1));
}
}
#endif /* QAK */
switch (layout->type) {
case H5D_CONTIGUOUS:
ndims = layout->ndims;
/*
* Offsets must not be negative for this type of storage.
*/
for (u=0; u<ndims; u++) {
if (mem_offset[u]<0 || file_offset[u]<0) {
HRETURN_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,
"negative offsets are not valid");
}
}
/*
* Filters cannot be used for contiguous data
*/
if (pline && pline->nfilters>0) {
HRETURN_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,
"filters are not allowed for contiguous data");
}
/*
* Calculate the strides needed to walk through the array on disk.
* Optimize the strides to result in the fewest number of I/O
* requests.
*/
mem_start = H5V_hyper_stride(ndims, hslab_size, mem_size,
mem_offset, mem_stride/*out*/);
file_start = H5V_hyper_stride(ndims, hslab_size, layout->dim,
file_offset, file_stride/*out*/);
H5V_stride_optimize2(&ndims, &elmt_size, hslab_size,
mem_stride, file_stride);
/*
* Initialize loop variables. The loop is a multi-dimensional loop
* that counts from SIZE down to zero and IDX is the counter. Each
* element of IDX is treated as a digit with IDX[0] being the least
* significant digit.
*/
H5V_vector_cpy(ndims, idx, hslab_size);
nelmts = H5V_vector_reduce_product(ndims, hslab_size);
if (efl && efl->nused>0) {
addr = 0;
} else {
addr = layout->addr;
/* Compute the size of the dataset in bytes */
for(u=0, max_data=1; u<layout->ndims; u++)
max_data *= layout->dim[u];
/* Adjust the maximum size of the data by the offset into it */
max_data -= file_start;
}
addr += file_start;
buf += mem_start;
/*
* Now begin to walk through the array, copying data from memory to
* disk.
*/
#ifdef H5_HAVE_PARALLEL
if (H5FD_MPIO_COLLECTIVE==xfer_mode){
/*
* Currently supports same number of collective access. Need to
* be changed LATER to combine all writes into one collective
* MPIO call.
*/
unsigned long max, min, temp;
temp = nelmts;
assert(temp==nelmts); /* verify no overflow */
MPI_Allreduce(&temp, &max, 1, MPI_UNSIGNED_LONG, MPI_MAX,
H5FD_mpio_communicator(f->shared->lf));
MPI_Allreduce(&temp, &min, 1, MPI_UNSIGNED_LONG, MPI_MIN,
H5FD_mpio_communicator(f->shared->lf));
#ifdef AKC
printf("nelmts=%lu, min=%lu, max=%lu\n", temp, min, max);
#endif
if (max != min) {
HRETURN_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
"collective access with unequal number of "
"blocks not supported yet");
}
}
#endif
for (z=0; z<nelmts; z++) {
/* Write to file */
if (efl && efl->nused>0) {
H5_CHECK_OVERFLOW(elmt_size,hsize_t,size_t);
if (H5O_efl_write(f, efl, addr, (size_t)elmt_size, buf)<0) {
HRETURN_ERROR(H5E_IO, H5E_READERROR, FAIL,
"external data write failed");
}
} else {
H5_CHECK_OVERFLOW(elmt_size,hsize_t,size_t);
if (H5F_contig_write(f, max_data, H5FD_MEM_DRAW, addr, (size_t)elmt_size, dxpl_id, buf)<0) {
HRETURN_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,
"block write failed");
}
} /* end else */
/* Decrement indices and advance pointers */
for (j=ndims-1, carray=TRUE; j>=0 && carray; --j) {
addr += file_stride[j];
buf += mem_stride[j];
/* Adjust the maximum size of the data by the offset into it */
max_data -= file_stride[j];
if (--idx[j])
carray = FALSE;
else
idx[j] = hslab_size[j];
}
}
break;
case H5D_CHUNKED:
/*
* This method is unable to access external raw data files or to copy
* from a proper hyperslab.
*/
if (efl && efl->nused>0) {
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL,
"chunking and external files are mutually "
"exclusive");
}
for (u=0; u<layout->ndims; u++) {
if (0!=mem_offset[u] || hslab_size[u]!=mem_size[u]) {
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL,
"unable to copy from a proper hyperslab");
}
}
if (H5F_istore_write(f, dxpl_id, layout, pline, fill, file_offset,
hslab_size, buf)<0) {
HRETURN_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,
"chunked write failed");
}
break;
default:
assert("not implemented yet" && 0);
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL,
"unsupported storage layout");
}
FUNC_LEAVE (SUCCEED);
}