FORT_DLL_SPEC void FORT_CALL mpi_alltoallw_ ( void*v1, MPI_Fint v2[], MPI_Fint v3[], MPI_Fint v4[], void*v5, MPI_Fint v6[], MPI_Fint v7[], MPI_Fint v8[], MPI_Fint *v9, MPI_Fint *ierr ){
#ifndef HAVE_MPI_F_INIT_WORKS_WITH_C
if (MPIR_F_NeedInit){ mpirinitf_(); MPIR_F_NeedInit = 0; }
#endif
if (v1 == MPIR_F_MPI_IN_PLACE) v1 = MPI_IN_PLACE;
*ierr = MPI_Alltoallw( v1, v2, v3, (MPI_Datatype *)(v4), v5, v6, v7, (MPI_Datatype *)(v8), (MPI_Comm)(*v9) );
}
FC_FUNC( mpi_alltoallw , MPI_ALLTOALLW )
( void *sendbuf, int *sendcounts, int *sdispls, int *sendtypes,
void *recvbuf, int *recvcounts, int *rdispls, int *recvtypes,
int *comm, int *ierror )
{
*ierror=MPI_Alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
recvbuf, recvcounts, rdispls, recvtypes,
*comm);
}
int main( int argc, char **argv )
{
MPI_Comm comm;
int *sbuf, *rbuf;
int rank, size;
int *sendcounts, *recvcounts, *rdispls, *sdispls;
int i, j, *p, err;
MPI_Datatype *sendtypes, *recvtypes;
MTest_Init( &argc, &argv );
err = 0;
while (MTestGetIntracommGeneral( &comm, 2, 1 )) {
if (comm == MPI_COMM_NULL) continue;
/* Create the buffer */
MPI_Comm_size( comm, &size );
MPI_Comm_rank( comm, &rank );
sbuf = (int *)malloc( size * size * sizeof(int) );
rbuf = (int *)malloc( size * size * sizeof(int) );
if (!sbuf || !rbuf) {
fprintf( stderr, "Could not allocated buffers!\n" );
MPI_Abort( comm, 1 );
}
/* Load up the buffers */
for (i=0; i<size*size; i++) {
sbuf[i] = i + 100*rank;
rbuf[i] = -i;
}
/* Create and load the arguments to alltoallv */
sendcounts = (int *)malloc( size * sizeof(int) );
recvcounts = (int *)malloc( size * sizeof(int) );
rdispls = (int *)malloc( size * sizeof(int) );
sdispls = (int *)malloc( size * sizeof(int) );
sendtypes = (MPI_Datatype *)malloc( size * sizeof(MPI_Datatype) );
recvtypes = (MPI_Datatype *)malloc( size * sizeof(MPI_Datatype) );
if (!sendcounts || !recvcounts || !rdispls || !sdispls || !sendtypes || !recvtypes) {
fprintf( stderr, "Could not allocate arg items!\n" );
MPI_Abort( comm, 1 );
}
/* Note that process 0 sends no data (sendcounts[0] = 0) */
for (i=0; i<size; i++) {
sendcounts[i] = i;
recvcounts[i] = rank;
rdispls[i] = i * rank * sizeof(int);
sdispls[i] = (((i+1) * (i))/2) * sizeof(int);
sendtypes[i] = recvtypes[i] = MPI_INT;
}
MPI_Alltoallw( sbuf, sendcounts, sdispls, sendtypes,
rbuf, recvcounts, rdispls, recvtypes, comm );
/* Check rbuf */
for (i=0; i<size; i++) {
p = rbuf + rdispls[i]/sizeof(int);
for (j=0; j<rank; j++) {
if (p[j] != i * 100 + (rank*(rank+1))/2 + j) {
fprintf( stderr, "[%d] got %d expected %d for %dth\n",
rank, p[j],(i*(i+1))/2 + j, j );
err++;
}
}
}
free(sendtypes);
free(sdispls);
free(sendcounts);
free(sbuf);
#if MTEST_HAVE_MIN_MPI_VERSION(2,2)
/* check MPI_IN_PLACE, added in MPI-2.2 */
free( rbuf );
rbuf = (int *)malloc( size * (2 * size) * sizeof(int) );
if (!rbuf) {
fprintf( stderr, "Could not reallocate rbuf!\n" );
MPI_Abort( comm, 1 );
}
/* Load up the buffers */
for (i = 0; i < size; i++) {
/* alltoallw displs are in bytes, not in type extents */
rdispls[i] = i * (2 * size) * sizeof(int);
recvtypes[i] = MPI_INT;
recvcounts[i] = i + rank;
}
memset(rbuf, -1, size * (2 * size) * sizeof(int));
for (i=0; i < size; i++) {
p = rbuf + (rdispls[i] / sizeof(int));
for (j = 0; j < recvcounts[i]; ++j) {
p[j] = 100 * rank + 10 * i + j;
}
}
MPI_Alltoallw( MPI_IN_PLACE, NULL, NULL, NULL,
rbuf, recvcounts, rdispls, recvtypes, comm );
/* Check rbuf */
for (i=0; i<size; i++) {
p = rbuf + (rdispls[i] / sizeof(int));
for (j=0; j<recvcounts[i]; j++) {
int expected = 100 * i + 10 * rank + j;
if (p[j] != expected) {
fprintf(stderr, "[%d] got %d expected %d for block=%d, element=%dth\n",
rank, p[j], expected, i, j);
++err;
}
}
}
#endif
free(recvtypes);
free(rdispls);
free(recvcounts);
free(rbuf);
MTestFreeComm( &comm );
}
MTest_Finalize( err );
MPI_Finalize();
return 0;
}
int main(int argc, char **argv)
{
int errs = 0;
int i;
int rank, size;
int *sbuf = NULL;
int *rbuf = NULL;
int *scounts = NULL;
int *rcounts = NULL;
int *sdispls = NULL;
int *rdispls = NULL;
MPI_Datatype *types = NULL;
MPI_Comm comm;
/* intentionally not using MTest_Init/MTest_Finalize in order to make it
* easy to take this test and use it as an NBC sanity test outside of the
* MPICH test suite */
MPI_Init(&argc, &argv);
comm = MPI_COMM_WORLD;
MPI_Comm_size(comm, &size);
MPI_Comm_rank(comm, &rank);
MPI_Comm_set_errhandler(MPI_COMM_WORLD, MPI_ERRORS_RETURN);
/* enough space for every process to contribute at least NUM_INTS ints to any
* collective operation */
sbuf = malloc(NUM_INTS * size * sizeof(int));
my_assert(sbuf);
rbuf = malloc(NUM_INTS * size * sizeof(int));
my_assert(rbuf);
scounts = malloc(size * sizeof(int));
my_assert(scounts);
rcounts = malloc(size * sizeof(int));
my_assert(rcounts);
sdispls = malloc(size * sizeof(int));
my_assert(sdispls);
rdispls = malloc(size * sizeof(int));
my_assert(rdispls);
types = malloc(size * sizeof(MPI_Datatype));
my_assert(types);
for (i = 0; i < size; ++i) {
sbuf[2 * i] = i;
sbuf[2 * i + 1] = i;
rbuf[2 * i] = i;
rbuf[2 * i + 1] = i;
scounts[i] = NUM_INTS;
rcounts[i] = NUM_INTS;
sdispls[i] = i * NUM_INTS;
rdispls[i] = i * NUM_INTS;
types[i] = MPI_INT;
}
if (rank == 0 && MPI_SUCCESS ==
MPI_Gather(sbuf, NUM_INTS, MPI_INT, sbuf, NUM_INTS, MPI_INT, 0, comm))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Gatherv(sbuf, NUM_INTS, MPI_INT, sbuf, rcounts, rdispls, MPI_INT, 0, comm))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Scatter(sbuf, NUM_INTS, MPI_INT, sbuf, NUM_INTS, MPI_INT, 0, comm))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Scatterv(sbuf, scounts, sdispls, MPI_INT, sbuf, NUM_INTS, MPI_INT, 0, comm))
errs++;
if (MPI_SUCCESS == MPI_Allgather(&sbuf[rank], 1, MPI_INT, sbuf, 1, MPI_INT, comm))
errs++;
if (MPI_SUCCESS ==
MPI_Allgatherv(&sbuf[rank * rcounts[rank]], rcounts[rank], MPI_INT, sbuf, rcounts, rdispls,
MPI_INT, comm))
errs++;
if (MPI_SUCCESS == MPI_Alltoall(sbuf, NUM_INTS, MPI_INT, sbuf, NUM_INTS, MPI_INT, comm))
errs++;
if (MPI_SUCCESS ==
MPI_Alltoallv(sbuf, scounts, sdispls, MPI_INT, sbuf, scounts, sdispls, MPI_INT, comm))
errs++;
if (MPI_SUCCESS ==
MPI_Alltoallw(sbuf, scounts, sdispls, types, sbuf, scounts, sdispls, types, comm))
errs++;
if (rank == 0 && MPI_SUCCESS == MPI_Reduce(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, 0, comm))
errs++;
if (MPI_SUCCESS == MPI_Allreduce(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm))
errs++;
if (MPI_SUCCESS == MPI_Reduce_scatter(sbuf, sbuf, rcounts, MPI_INT, MPI_SUM, comm))
errs++;
if (MPI_SUCCESS == MPI_Reduce_scatter_block(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm))
errs++;
if (MPI_SUCCESS == MPI_Scan(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm))
errs++;
if (MPI_SUCCESS == MPI_Exscan(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm))
errs++;
if (sbuf)
free(sbuf);
if (rbuf)
free(rbuf);
if (scounts)
free(scounts);
if (rcounts)
free(rcounts);
if (sdispls)
free(sdispls);
if (rdispls)
free(rdispls);
if (types)
free(types);
if (rank == 0) {
if (errs)
fprintf(stderr, "Found %d errors\n", errs);
else
printf(" No errors\n");
}
MPI_Finalize();
return 0;
}
void Transpose(float *localA, float *localB, int M, int N, MPI_Comm comm)
/* transpose MxN matrix A that is block distributed (1-D) on
processes of comm onto block distributed matrix B */
{
int i, j, extent, myrank, p, n[2], m[2];
int lasti, lastj;
int *sendcounts, *recvcounts;
int *sdispls, *rdispls;
MPI_Datatype xtype[2][2], stype[2][2], *sendtypes, *recvtypes;
MTestPrintfMsg( 2, "M = %d, N = %d\n", M, N );
/* compute parameters */
MPI_Comm_size(comm, &p);
MPI_Comm_rank(comm, &myrank);
extent = sizeof(float);
/* allocate arrays */
sendcounts = (int *)malloc(p*sizeof(int));
recvcounts = (int *)malloc(p*sizeof(int));
sdispls = (int *)malloc(p*sizeof(int));
rdispls = (int *)malloc(p*sizeof(int));
sendtypes = (MPI_Datatype *)malloc(p*sizeof(MPI_Datatype));
recvtypes = (MPI_Datatype *)malloc(p*sizeof(MPI_Datatype));
/* compute block sizes */
m[0] = M/p;
m[1] = M - (p-1)*(M/p);
n[0] = N/p;
n[1] = N - (p-1)*(N/p);
/* compute types */
for (i=0; i <= 1; i++)
for (j=0; j <= 1; j++) {
xtype[i][j] = transpose_type(N, m[i], n[j], MPI_FLOAT);
stype[i][j] = submatrix_type(M, m[i], n[j], MPI_FLOAT);
}
/* prepare collective operation arguments */
lasti = myrank == p-1;
for (j=0; j < p; j++) {
lastj = j == p-1;
sendcounts[j] = 1;
sdispls[j] = j*n[0]*extent;
sendtypes[j] = xtype[lasti][lastj];
recvcounts[j] = 1;
rdispls[j] = j*m[0]*extent;
recvtypes[j] = stype[lastj][lasti];
}
/* communicate */
MTestPrintfMsg( 2, "Begin Alltoallw...\n" );
/* -- Note that the book incorrectly uses &localA and &localB
as arguments to MPI_Alltoallw */
MPI_Alltoallw(localA, sendcounts, sdispls, sendtypes,
localB, recvcounts, rdispls, recvtypes, comm);
MTestPrintfMsg( 2, "Done with Alltoallw\n" );
/* Free buffers */
free( sendcounts );
free( recvcounts );
free( sdispls );
free( rdispls );
free( sendtypes );
free( recvtypes );
/* Free datatypes */
for (i=0; i <= 1; i++)
for (j=0; j <= 1; j++) {
MPI_Type_free( &xtype[i][j] );
MPI_Type_free( &stype[i][j] );
}
}
/* Avery Ching and Kenin Columa's reworked two-phase algorithm. Key features
* - persistent file domains
* - an option to use alltoall instead of point-to-point
*/
void ADIOI_IOStridedColl(ADIO_File fd, void *buf, int count, int rdwr,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status * status, int *error_code)
{
ADIO_Offset min_st_offset = 0, max_end_offset = 0;
ADIO_Offset st_end_offset[2];
ADIO_Offset *all_st_end_offsets = NULL;
int filetype_is_contig, buftype_is_contig, is_contig;
ADIO_Offset off;
int interleave_count = 0, i, nprocs, myrank, nprocs_for_coll;
int cb_enable;
ADIO_Offset bufsize;
MPI_Aint extent;
#ifdef DEBUG2
MPI_Aint bufextent;
#endif
MPI_Count size;
int agg_rank;
ADIO_Offset agg_disp; /* aggregated file offset */
MPI_Datatype agg_dtype; /* aggregated file datatype */
int aggregators_done = 0;
ADIO_Offset buffered_io_size = 0;
int *alltoallw_disps;
int *alltoallw_counts;
int *client_alltoallw_counts;
int *agg_alltoallw_counts;
char *cb_buf = NULL;
MPI_Datatype *client_comm_dtype_arr; /* aggregator perspective */
MPI_Datatype *agg_comm_dtype_arr; /* client perspective */
ADIO_Offset *client_comm_sz_arr; /* aggregator perspective */
ADIO_Offset *agg_comm_sz_arr; /* client perspective */
/* file views for each client and aggregator */
view_state *client_file_view_state_arr = NULL;
view_state *agg_file_view_state_arr = NULL;
/* mem views for local process */
view_state *my_mem_view_state_arr = NULL;
MPI_Status *agg_comm_statuses = NULL;
MPI_Request *agg_comm_requests = NULL;
MPI_Status *client_comm_statuses = NULL;
MPI_Request *client_comm_requests = NULL;
int aggs_client_count = 0;
int clients_agg_count = 0;
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
#ifdef DEBUG
fprintf(stderr, "p%d: entering ADIOI_IOStridedColl\n", myrank);
#endif
#ifdef AGGREGATION_PROFILE
if (rdwr == ADIOI_READ)
MPE_Log_event(5010, 0, NULL);
else
MPE_Log_event(5012, 0, NULL);
#endif
/* I need to check if there are any outstanding nonblocking writes
* to the file, which could potentially interfere with the writes
* taking place in this collective write call. Since this is not
* likely to be common, let me do the simplest thing possible here:
* Each process completes all pending nonblocking operations before
* completing. */
nprocs_for_coll = fd->hints->cb_nodes;
if (rdwr == ADIOI_READ)
cb_enable = fd->hints->cb_read;
else
cb_enable = fd->hints->cb_write;
/* only check for interleaving if cb_read isn't disabled */
if (cb_enable != ADIOI_HINT_DISABLE) {
/* find the starting and ending byte of my I/O access */
ADIOI_Calc_bounds(fd, count, datatype, file_ptr_type, offset,
&st_end_offset[0], &st_end_offset[1]);
/* allocate an array of start/end pairs */
all_st_end_offsets = (ADIO_Offset *)
ADIOI_Malloc(2 * nprocs * sizeof(ADIO_Offset));
MPI_Allgather(st_end_offset, 2, ADIO_OFFSET, all_st_end_offsets, 2, ADIO_OFFSET, fd->comm);
min_st_offset = all_st_end_offsets[0];
max_end_offset = all_st_end_offsets[1];
for (i = 1; i < nprocs; i++) {
/* are the accesses of different processes interleaved? */
if ((all_st_end_offsets[i * 2] < all_st_end_offsets[i * 2 - 1]) &&
(all_st_end_offsets[i * 2] <= all_st_end_offsets[i * 2 + 1]))
interleave_count++;
/* This is a rudimentary check for interleaving, but should
* suffice for the moment. */
min_st_offset = MPL_MIN(all_st_end_offsets[i * 2], min_st_offset);
max_end_offset = MPL_MAX(all_st_end_offsets[i * 2 + 1], max_end_offset);
}
}
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
if ((cb_enable == ADIOI_HINT_DISABLE || (!interleave_count && (cb_enable == ADIOI_HINT_AUTO)))
&& (fd->hints->cb_pfr != ADIOI_HINT_ENABLE)) {
if (cb_enable != ADIOI_HINT_DISABLE) {
ADIOI_Free(all_st_end_offsets);
}
if (buftype_is_contig && filetype_is_contig) {
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
off = fd->disp + (fd->etype_size) * offset;
if (rdwr == ADIOI_READ)
ADIO_ReadContig(fd, buf, count, datatype,
ADIO_EXPLICIT_OFFSET, off, status, error_code);
else
ADIO_WriteContig(fd, buf, count, datatype,
ADIO_EXPLICIT_OFFSET, off, status, error_code);
} else {
if (rdwr == ADIOI_READ)
ADIO_ReadContig(fd, buf, count, datatype, ADIO_INDIVIDUAL,
0, status, error_code);
else
ADIO_WriteContig(fd, buf, count, datatype, ADIO_INDIVIDUAL,
0, status, error_code);
}
} else {
if (rdwr == ADIOI_READ)
ADIO_ReadStrided(fd, buf, count, datatype, file_ptr_type,
offset, status, error_code);
else
ADIO_WriteStrided(fd, buf, count, datatype, file_ptr_type,
offset, status, error_code);
}
return;
}
MPI_Type_extent(datatype, &extent);
#ifdef DEBUG2
bufextent = extent * count;
#endif
MPI_Type_size_x(datatype, &size);
bufsize = size * (MPI_Count) count;
/* Calculate file realms */
if ((fd->hints->cb_pfr != ADIOI_HINT_ENABLE) || (fd->file_realm_types == NULL))
ADIOI_Calc_file_realms(fd, min_st_offset, max_end_offset);
my_mem_view_state_arr = (view_state *)
ADIOI_Calloc(1, nprocs * sizeof(view_state));
agg_file_view_state_arr = (view_state *)
ADIOI_Calloc(1, nprocs * sizeof(view_state));
client_comm_sz_arr = (ADIO_Offset *)
ADIOI_Calloc(1, nprocs * sizeof(ADIO_Offset));
if (fd->is_agg) {
client_file_view_state_arr = (view_state *)
ADIOI_Calloc(1, nprocs * sizeof(view_state));
} else {
client_file_view_state_arr = NULL;
}
/* Alltoallw doesn't like a null array even if the counts are
* zero. If you do not include this code, it will fail. */
client_comm_dtype_arr = (MPI_Datatype *)
ADIOI_Calloc(1, nprocs * sizeof(MPI_Datatype));
if (!fd->is_agg)
for (i = 0; i < nprocs; i++)
client_comm_dtype_arr[i] = MPI_BYTE;
ADIOI_Exch_file_views(myrank, nprocs, file_ptr_type, fd, count,
datatype, offset, my_mem_view_state_arr,
agg_file_view_state_arr, client_file_view_state_arr);
agg_comm_sz_arr = (ADIO_Offset *)
ADIOI_Calloc(1, nprocs * sizeof(ADIO_Offset));
agg_comm_dtype_arr = (MPI_Datatype *)
ADIOI_Malloc(nprocs * sizeof(MPI_Datatype));
if (fd->is_agg) {
ADIOI_Build_agg_reqs(fd, rdwr, nprocs,
client_file_view_state_arr,
client_comm_dtype_arr, client_comm_sz_arr, &agg_disp, &agg_dtype);
buffered_io_size = 0;
for (i = 0; i < nprocs; i++) {
if (client_comm_sz_arr[i] > 0)
buffered_io_size += client_comm_sz_arr[i];
}
}
#ifdef USE_PRE_REQ
else {
/* Example use of ADIOI_Build_client_pre_req. to an
* appropriate section */
for (i = 0; i < fd->hints->cb_nodes; i++) {
agg_rank = fd->hints->ranklist[(i + myrank) % fd->hints->cb_nodes];
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5040, 0, NULL);
#endif
ADIOI_Build_client_pre_req(fd, agg_rank, (i + myrank) % fd->hints->cb_nodes,
&(my_mem_view_state_arr[agg_rank]),
&(agg_file_view_state_arr[agg_rank]),
2 * 1024 * 1024, 64 * 1024);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5041, 0, NULL);
#endif
}
}
#endif
if (fd->is_agg)
cb_buf = (char *) ADIOI_Malloc(fd->hints->cb_buffer_size);
alltoallw_disps = (int *) ADIOI_Calloc(nprocs, sizeof(int));
alltoallw_counts = client_alltoallw_counts = (int *)
ADIOI_Calloc(2 * nprocs, sizeof(int));
agg_alltoallw_counts = &alltoallw_counts[nprocs];
if (fd->hints->cb_alltoall == ADIOI_HINT_DISABLE) {
/* aggregators pre-post all Irecv's for incoming data from clients */
if ((fd->is_agg) && (rdwr == ADIOI_WRITE))
post_aggregator_comm(fd->comm, rdwr, nprocs, cb_buf,
client_comm_dtype_arr,
client_comm_sz_arr, &agg_comm_requests, &aggs_client_count);
}
/* Aggregators send amounts for data requested to clients */
Exch_data_amounts(fd, nprocs, client_comm_sz_arr, agg_comm_sz_arr,
client_alltoallw_counts, agg_alltoallw_counts, &aggregators_done);
#ifdef DEBUG
fprintf(stderr, "client_alltoallw_counts[ ");
for (i = 0; i < nprocs; i++) {
fprintf(stderr, "%d ", client_alltoallw_counts[i]);
}
fprintf(stderr, "]\n");
fprintf(stderr, "agg_alltoallw_counts[ ");
for (i = 0; i < nprocs; i++) {
fprintf(stderr, "%d ", agg_alltoallw_counts[i]);
}
fprintf(stderr, "]\n");
#endif
/* keep looping while aggregators still have I/O to do */
while (aggregators_done != nprocs_for_coll) {
if (fd->hints->cb_alltoall == ADIOI_HINT_DISABLE) {
/* clients should build datatypes for local memory locations
* for data communication with aggregators and post
* communication as the datatypes are built */
client_comm_requests = (MPI_Request *)
ADIOI_Calloc(fd->hints->cb_nodes, sizeof(MPI_Request));
for (i = 0; i < fd->hints->cb_nodes; i++) {
clients_agg_count = 0;
agg_rank = fd->hints->ranklist[(i + myrank) % fd->hints->cb_nodes];
if (agg_comm_sz_arr[agg_rank] > 0) {
ADIOI_Build_client_req(fd, agg_rank,
(i + myrank) % fd->hints->cb_nodes,
&(my_mem_view_state_arr[agg_rank]),
&(agg_file_view_state_arr[agg_rank]),
agg_comm_sz_arr[agg_rank],
&(agg_comm_dtype_arr[agg_rank]));
#ifdef AGGREGATION_PROFILE
if (i == 0)
MPE_Log_event(5038, 0, NULL);
#endif
post_client_comm(fd, rdwr, agg_rank, buf,
agg_comm_dtype_arr[agg_rank],
agg_alltoallw_counts[agg_rank],
&client_comm_requests[clients_agg_count]);
clients_agg_count++;
}
}
#ifdef AGGREGATION_PROFILE
if (!clients_agg_count)
MPE_Log_event(5039, 0, NULL);
#endif
if (rdwr == ADIOI_READ) {
if (fd->is_agg && buffered_io_size) {
ADIOI_IOFiletype(fd, cb_buf, buffered_io_size, MPI_BYTE,
ADIO_EXPLICIT_OFFSET, agg_disp, agg_dtype,
ADIOI_READ, status, error_code);
if (*error_code != MPI_SUCCESS)
return;
MPI_Type_free(&agg_dtype);
}
#ifdef DEBUG
fprintf(stderr, "expecting from [agg](disp,size,cnt)=");
for (i = 0; i < nprocs; i++) {
MPI_Type_size_x(agg_comm_dtype_arr[i], &size);
fprintf(stderr, "[%d](%d,%d,%d)", i, alltoallw_disps[i],
size, agg_alltoallw_counts[i]);
if (i != nprocs - 1)
fprintf(stderr, ",");
}
fprintf(stderr, "]\n");
if (fd->is_agg) {
fprintf(stderr, "sending to [client](disp,size,cnt)=");
for (i = 0; i < nprocs; i++) {
if (fd->is_agg)
MPI_Type_size_x(client_comm_dtype_arr[i], &size);
else
size = -1;
fprintf(stderr, "[%d](%d,%d,%d)", i, alltoallw_disps[i],
size, client_alltoallw_counts[i]);
if (i != nprocs - 1)
fprintf(stderr, ",");
}
fprintf(stderr, "\n");
}
fflush(NULL);
#endif
/* aggregators post all Isends for outgoing data to clients */
if (fd->is_agg)
post_aggregator_comm(fd->comm, rdwr, nprocs, cb_buf,
client_comm_dtype_arr,
client_comm_sz_arr,
&agg_comm_requests, &aggs_client_count);
if (fd->is_agg && aggs_client_count) {
#ifdef MPI_STATUSES_IGNORE
agg_comm_statuses = MPI_STATUSES_IGNORE;
#else
agg_comm_statuses = ADIOI_Malloc(aggs_client_count * sizeof(MPI_Status));
#endif
MPI_Waitall(aggs_client_count, agg_comm_requests, agg_comm_statuses);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5033, 0, NULL);
#endif
ADIOI_Free(agg_comm_requests);
#ifndef MPI_STATUSES_IGNORE
ADIOI_Free(agg_comm_statuses);
#endif
}
if (clients_agg_count) {
#ifdef MPI_STATUSES_IGNORE
client_comm_statuses = MPI_STATUSES_IGNORE;
#else
client_comm_statuses = ADIOI_Malloc(clients_agg_count * sizeof(MPI_Status));
#endif
MPI_Waitall(clients_agg_count, client_comm_requests, client_comm_statuses);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5039, 0, NULL);
#endif
ADIOI_Free(client_comm_requests);
#ifndef MPI_STATUSES_IGNORE
ADIOI_Free(client_comm_statuses);
#endif
}
#ifdef DEBUG2
fprintf(stderr, "buffered_io_size = %lld\n", buffered_io_size);
if (fd->is_agg && buffered_io_size) {
fprintf(stderr, "buf = [");
for (i = 0; i < bufextent; i++)
fprintf(stderr, "%c", ((char *) buf)[i]);
fprintf(stderr, "]\n");
fprintf(stderr, "cb_buf = [");
for (i = 0; i < buffered_io_size; i++)
fprintf(stderr, "%c", cb_buf[i]);
fprintf(stderr, "]\n");
fflush(NULL);
}
#endif
} else { /* Write Case */
#ifdef DEBUG
fprintf(stderr, "sending to [agg](disp,size,cnt)=");
for (i = 0; i < nprocs; i++) {
MPI_Type_size_x(agg_comm_dtype_arr[i], &size);
fprintf(stderr, "[%d](%d,%d,%d)", i, alltoallw_disps[i],
size, agg_alltoallw_counts[i]);
if (i != nprocs - 1)
fprintf(stderr, ",");
}
fprintf(stderr, "]\n");
fprintf(stderr, "expecting from [client](disp,size,cnt)=");
for (i = 0; i < nprocs; i++) {
if (fd->is_agg)
MPI_Type_size_x(client_comm_dtype_arr[i], &size);
else
size = -1;
fprintf(stderr, "[%d](%d,%d,%d)", i, alltoallw_disps[i],
size, client_alltoallw_counts[i]);
if (i != nprocs - 1)
fprintf(stderr, ",");
}
fprintf(stderr, "\n");
fflush(NULL);
#endif
#ifdef DEBUG
fprintf(stderr, "buffered_io_size = %lld\n", buffered_io_size);
#endif
if (clients_agg_count) {
#ifdef MPI_STATUSES_IGNORE
client_comm_statuses = MPI_STATUSES_IGNORE;
#else
client_comm_statuses = ADIOI_Malloc(clients_agg_count * sizeof(MPI_Status));
#endif
MPI_Waitall(clients_agg_count, client_comm_requests, client_comm_statuses);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5039, 0, NULL);
#endif
ADIOI_Free(client_comm_requests);
#ifndef MPI_STATUSES_IGNORE
ADIOI_Free(client_comm_statuses);
#endif
}
#ifdef DEBUG2
if (bufextent) {
fprintf(stderr, "buf = [");
for (i = 0; i < bufextent; i++)
fprintf(stderr, "%c", ((char *) buf)[i]);
fprintf(stderr, "]\n");
}
#endif
if (fd->is_agg && buffered_io_size) {
ADIOI_Assert(aggs_client_count != 0);
/* make sure we actually have the data to write out */
#ifdef MPI_STATUSES_IGNORE
agg_comm_statuses = MPI_STATUSES_IGNORE;
#else
agg_comm_statuses = (MPI_Status *)
ADIOI_Malloc(aggs_client_count * sizeof(MPI_Status));
#endif
MPI_Waitall(aggs_client_count, agg_comm_requests, agg_comm_statuses);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5033, 0, NULL);
#endif
ADIOI_Free(agg_comm_requests);
#ifndef MPI_STATUSES_IGNORE
ADIOI_Free(agg_comm_statuses);
#endif
#ifdef DEBUG2
fprintf(stderr, "cb_buf = [");
for (i = 0; i < buffered_io_size; i++)
fprintf(stderr, "%c", cb_buf[i]);
fprintf(stderr, "]\n");
fflush(NULL);
#endif
ADIOI_IOFiletype(fd, cb_buf, buffered_io_size, MPI_BYTE,
ADIO_EXPLICIT_OFFSET, agg_disp, agg_dtype,
ADIOI_WRITE, status, error_code);
if (*error_code != MPI_SUCCESS)
return;
MPI_Type_free(&agg_dtype);
}
}
} else {
/* Alltoallw version of everything */
ADIOI_Build_client_reqs(fd, nprocs, my_mem_view_state_arr,
agg_file_view_state_arr, agg_comm_sz_arr, agg_comm_dtype_arr);
if (rdwr == ADIOI_READ) {
if (fd->is_agg && buffered_io_size) {
ADIOI_IOFiletype(fd, cb_buf, buffered_io_size, MPI_BYTE,
ADIO_EXPLICIT_OFFSET, agg_disp, agg_dtype,
ADIOI_READ, status, error_code);
if (*error_code != MPI_SUCCESS)
return;
MPI_Type_free(&agg_dtype);
}
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5032, 0, NULL);
#endif
MPI_Alltoallw(cb_buf, client_alltoallw_counts, alltoallw_disps,
client_comm_dtype_arr,
buf, agg_alltoallw_counts, alltoallw_disps,
agg_comm_dtype_arr, fd->comm);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5033, 0, NULL);
#endif
} else { /* Write Case */
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5032, 0, NULL);
#endif
MPI_Alltoallw(buf, agg_alltoallw_counts, alltoallw_disps,
agg_comm_dtype_arr,
cb_buf, client_alltoallw_counts, alltoallw_disps,
client_comm_dtype_arr, fd->comm);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5033, 0, NULL);
#endif
if (fd->is_agg && buffered_io_size) {
ADIOI_IOFiletype(fd, cb_buf, buffered_io_size, MPI_BYTE,
ADIO_EXPLICIT_OFFSET, agg_disp, agg_dtype,
ADIOI_WRITE, status, error_code);
if (*error_code != MPI_SUCCESS)
return;
MPI_Type_free(&agg_dtype);
}
}
}
/* Free (uncommit) datatypes for reuse */
if (fd->is_agg) {
if (buffered_io_size > 0) {
for (i = 0; i < nprocs; i++) {
if (client_comm_sz_arr[i] > 0)
MPI_Type_free(&client_comm_dtype_arr[i]);
}
}
}
for (i = 0; i < nprocs; i++) {
if (agg_comm_sz_arr[i] > 0)
MPI_Type_free(&agg_comm_dtype_arr[i]);
}
/* figure out next set up requests */
if (fd->is_agg) {
ADIOI_Build_agg_reqs(fd, rdwr, nprocs,
client_file_view_state_arr,
client_comm_dtype_arr, client_comm_sz_arr, &agg_disp, &agg_dtype);
buffered_io_size = 0;
for (i = 0; i < nprocs; i++) {
if (client_comm_sz_arr[i] > 0)
buffered_io_size += client_comm_sz_arr[i];
}
}
#ifdef USE_PRE_REQ
else {
/* Example use of ADIOI_Build_client_pre_req. to an
* appropriate section */
for (i = 0; i < fd->hints->cb_nodes; i++) {
agg_rank = fd->hints->ranklist[(i + myrank) % fd->hints->cb_nodes];
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5040, 0, NULL);
#endif
ADIOI_Build_client_pre_req(fd, agg_rank, (i + myrank) % fd->hints->cb_nodes,
&(my_mem_view_state_arr[agg_rank]),
&(agg_file_view_state_arr[agg_rank]),
2 * 1024 * 1024, 64 * 1024);
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5041, 0, NULL);
#endif
}
}
#endif
/* aggregators pre-post all Irecv's for incoming data from
* clients. if nothing is needed, agg_comm_requests is not
* allocated */
if (fd->hints->cb_alltoall == ADIOI_HINT_DISABLE) {
if ((fd->is_agg) && (rdwr == ADIOI_WRITE))
post_aggregator_comm(fd->comm, rdwr, nprocs, cb_buf,
client_comm_dtype_arr,
client_comm_sz_arr, &agg_comm_requests, &aggs_client_count);
}
/* Aggregators send amounts for data requested to clients */
Exch_data_amounts(fd, nprocs, client_comm_sz_arr, agg_comm_sz_arr,
client_alltoallw_counts, agg_alltoallw_counts, &aggregators_done);
}
/* Clean up */
if (fd->hints->cb_pfr != ADIOI_HINT_ENABLE) {
/* AAR, FSIZE, and User provided uniform File realms */
if (1) {
MPI_Type_free(&fd->file_realm_types[0]);
} else {
for (i = 0; i < fd->hints->cb_nodes; i++) {
ADIOI_Datatype_iscontig(fd->file_realm_types[i], &is_contig);
MPI_Type_free(&fd->file_realm_types[i]);
}
}
ADIOI_Free(fd->file_realm_types);
ADIOI_Free(fd->file_realm_st_offs);
}
if (fd->is_agg) {
if (buffered_io_size > 0)
MPI_Type_free(&agg_dtype);
for (i = 0; i < nprocs; i++) {
MPI_Type_free(&client_comm_dtype_arr[i]);
ADIOI_Free(client_file_view_state_arr[i].flat_type_p->indices);
ADIOI_Free(client_file_view_state_arr[i].flat_type_p->blocklens);
ADIOI_Free(client_file_view_state_arr[i].flat_type_p);
}
ADIOI_Free(client_file_view_state_arr);
ADIOI_Free(cb_buf);
}
for (i = 0; i < nprocs; i++)
if (agg_comm_sz_arr[i] > 0)
MPI_Type_free(&agg_comm_dtype_arr[i]);
ADIOI_Free(client_comm_sz_arr);
ADIOI_Free(client_comm_dtype_arr);
ADIOI_Free(my_mem_view_state_arr);
ADIOI_Free(agg_file_view_state_arr);
ADIOI_Free(agg_comm_sz_arr);
ADIOI_Free(agg_comm_dtype_arr);
ADIOI_Free(alltoallw_disps);
ADIOI_Free(alltoallw_counts);
ADIOI_Free(all_st_end_offsets);
#ifdef HAVE_STATUS_SET_BYTES
MPIR_Status_set_bytes(status, datatype, bufsize);
/* This is a temporary way of filling in status. The right way is
* to keep track of how much data was actually read and placed in
* buf during collective I/O. */
#endif
fd->fp_sys_posn = -1; /* set it to null. */
#ifdef AGGREGATION_PROFILE
if (rdwr == ADIOI_READ)
MPE_Log_event(5011, 0, NULL);
else
MPE_Log_event(5013, 0, NULL);
#endif
}
