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;
int *types = NULL;
MPI_Comm comm;
MPI_Request req;
/* 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);
/* 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(int));
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;
}
MPI_Ibarrier(comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ibcast(sbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Igather(sbuf, NUM_INTS, MPI_INT, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
if (0 == rank)
MPI_Igather(MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
else
MPI_Igather(sbuf, NUM_INTS, MPI_INT, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Igatherv(sbuf, NUM_INTS, MPI_INT, rbuf, rcounts, rdispls, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
if (0 == rank)
MPI_Igatherv(MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, rbuf, rcounts, rdispls, MPI_INT, 0, comm,
&req);
else
MPI_Igatherv(sbuf, NUM_INTS, MPI_INT, rbuf, rcounts, rdispls, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iscatter(sbuf, NUM_INTS, MPI_INT, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
if (0 == rank)
MPI_Iscatter(sbuf, NUM_INTS, MPI_INT, MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, 0, comm, &req);
else
MPI_Iscatter(sbuf, NUM_INTS, MPI_INT, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iscatterv(sbuf, scounts, sdispls, MPI_INT, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
if (0 == rank)
MPI_Iscatterv(sbuf, scounts, sdispls, MPI_INT, MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, 0, comm,
&req);
else
MPI_Iscatterv(sbuf, scounts, sdispls, MPI_INT, rbuf, NUM_INTS, MPI_INT, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iallgather(sbuf, NUM_INTS, MPI_INT, rbuf, NUM_INTS, MPI_INT, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iallgather(MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, rbuf, NUM_INTS, MPI_INT, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iallgatherv(sbuf, NUM_INTS, MPI_INT, rbuf, rcounts, rdispls, MPI_INT, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iallgatherv(MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, rbuf, rcounts, rdispls, MPI_INT, comm,
&req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ialltoall(sbuf, NUM_INTS, MPI_INT, rbuf, NUM_INTS, MPI_INT, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ialltoall(MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, rbuf, NUM_INTS, MPI_INT, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ialltoallv(sbuf, scounts, sdispls, MPI_INT, rbuf, rcounts, rdispls, MPI_INT, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ialltoallv(MPI_IN_PLACE, NULL, NULL, MPI_DATATYPE_NULL, rbuf, rcounts, rdispls, MPI_INT,
comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ialltoallw(sbuf, scounts, sdispls, types, rbuf, rcounts, rdispls, types, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ialltoallw(MPI_IN_PLACE, NULL, NULL, NULL, rbuf, rcounts, rdispls, types, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ireduce(sbuf, rbuf, NUM_INTS, MPI_INT, MPI_SUM, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
if (0 == rank)
MPI_Ireduce(MPI_IN_PLACE, rbuf, NUM_INTS, MPI_INT, MPI_SUM, 0, comm, &req);
else
MPI_Ireduce(sbuf, rbuf, NUM_INTS, MPI_INT, MPI_SUM, 0, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iallreduce(sbuf, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iallreduce(MPI_IN_PLACE, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ireduce_scatter(sbuf, rbuf, rcounts, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ireduce_scatter(MPI_IN_PLACE, rbuf, rcounts, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ireduce_scatter_block(sbuf, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Ireduce_scatter_block(MPI_IN_PLACE, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iscan(sbuf, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iscan(MPI_IN_PLACE, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iexscan(sbuf, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Iexscan(MPI_IN_PLACE, rbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
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 (rank == 0) {
if (errs)
fprintf(stderr, "Found %d errors\n", errs);
else
printf(" No errors\n");
}
MPI_Finalize();
return 0;
}
/* Starts a "random" operation on "comm" corresponding to "rndnum" and returns
* in (*req) a request handle corresonding to that operation. This call should
* be considered collective over comm (with a consistent value for "rndnum"),
* even though the operation may only be a point-to-point request. */
static void start_random_nonblocking(MPI_Comm comm, unsigned int rndnum, MPI_Request *req, struct laundry *l)
{
int i, j;
int rank, size;
int *buf = NULL;
int *recvbuf = NULL;
int *sendcounts = NULL;
int *recvcounts = NULL;
int *sdispls = NULL;
int *rdispls = NULL;
int *sendtypes = NULL;
int *recvtypes = NULL;
signed char *buf_alias = NULL;
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &size);
*req = MPI_REQUEST_NULL;
l->case_num = -1;
l->comm = comm;
l->buf = buf = malloc(COUNT*size*sizeof(int));
l->recvbuf = recvbuf = malloc(COUNT*size*sizeof(int));
l->sendcounts = sendcounts = malloc(size*sizeof(int));
l->recvcounts = recvcounts = malloc(size*sizeof(int));
l->sdispls = sdispls = malloc(size*sizeof(int));
l->rdispls = rdispls = malloc(size*sizeof(int));
l->sendtypes = sendtypes = malloc(size*sizeof(MPI_Datatype));
l->recvtypes = recvtypes = malloc(size*sizeof(MPI_Datatype));
#define NUM_CASES (21)
l->case_num = rand_range(rndnum, 0, NUM_CASES);
switch (l->case_num) {
case 0: /* MPI_Ibcast */
for (i = 0; i < COUNT; ++i) {
if (rank == 0) {
buf[i] = i;
}
else {
buf[i] = 0xdeadbeef;
}
}
MPI_Ibcast(buf, COUNT, MPI_INT, 0, comm, req);
break;
case 1: /* MPI_Ibcast (again, but designed to stress scatter/allgather impls) */
/* FIXME fiddle with PRIME and buffer allocation s.t. PRIME is much larger (1021?) */
buf_alias = (signed char *)buf;
my_assert(COUNT*size*sizeof(int) > PRIME); /* sanity */
for (i = 0; i < PRIME; ++i) {
if (rank == 0)
buf_alias[i] = i;
else
buf_alias[i] = 0xdb;
}
for (i = PRIME; i < COUNT * size * sizeof(int); ++i) {
buf_alias[i] = 0xbf;
}
MPI_Ibcast(buf_alias, PRIME, MPI_SIGNED_CHAR, 0, comm, req);
break;
case 2: /* MPI_Ibarrier */
MPI_Ibarrier(comm, req);
break;
case 3: /* MPI_Ireduce */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Ireduce(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, 0, comm, req);
break;
case 4: /* same again, use a user op and free it before the wait */
{
MPI_Op op = MPI_OP_NULL;
MPI_Op_create(sum_fn, /*commute=*/1, &op);
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Ireduce(buf, recvbuf, COUNT, MPI_INT, op, 0, comm, req);
MPI_Op_free(&op);
}
break;
case 5: /* MPI_Iallreduce */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iallreduce(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 6: /* MPI_Ialltoallv (a weak test, neither irregular nor sparse) */
for (i = 0; i < size; ++i) {
sendcounts[i] = COUNT;
recvcounts[i] = COUNT;
sdispls[i] = COUNT * i;
rdispls[i] = COUNT * i;
for (j = 0; j < COUNT; ++j) {
buf[i*COUNT+j] = rank + (i * j);
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Ialltoallv(buf, sendcounts, sdispls, MPI_INT, recvbuf, recvcounts, rdispls, MPI_INT, comm, req);
break;
case 7: /* MPI_Igather */
for (i = 0; i < size*COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Igather(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, 0, comm, req);
break;
case 8: /* same test again, just use a dup'ed datatype and free it before the wait */
{
MPI_Datatype type = MPI_DATATYPE_NULL;
MPI_Type_dup(MPI_INT, &type);
for (i = 0; i < size*COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Igather(buf, COUNT, MPI_INT, recvbuf, COUNT, type, 0, comm, req);
MPI_Type_free(&type); /* should cause implementations that don't refcount
correctly to blow up or hang in the wait */
}
break;
case 9: /* MPI_Iscatter */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
if (rank == 0)
buf[i*COUNT+j] = i + j;
else
buf[i*COUNT+j] = 0xdeadbeef;
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Iscatter(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, 0, comm, req);
break;
case 10: /* MPI_Iscatterv */
for (i = 0; i < size; ++i) {
/* weak test, just test the regular case where all counts are equal */
sendcounts[i] = COUNT;
sdispls[i] = i * COUNT;
for (j = 0; j < COUNT; ++j) {
if (rank == 0)
buf[i*COUNT+j] = i + j;
else
buf[i*COUNT+j] = 0xdeadbeef;
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Iscatterv(buf, sendcounts, sdispls, MPI_INT, recvbuf, COUNT, MPI_INT, 0, comm, req);
break;
case 11: /* MPI_Ireduce_scatter */
for (i = 0; i < size; ++i) {
recvcounts[i] = COUNT;
for (j = 0; j < COUNT; ++j) {
buf[i*COUNT+j] = rank + i;
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Ireduce_scatter(buf, recvbuf, recvcounts, MPI_INT, MPI_SUM, comm, req);
break;
case 12: /* MPI_Ireduce_scatter_block */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
buf[i*COUNT+j] = rank + i;
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Ireduce_scatter_block(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 13: /* MPI_Igatherv */
for (i = 0; i < size*COUNT; ++i) {
buf[i] = 0xdeadbeef;
recvbuf[i] = 0xdeadbeef;
}
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
}
for (i = 0; i < size; ++i) {
recvcounts[i] = COUNT;
rdispls[i] = i * COUNT;
}
MPI_Igatherv(buf, COUNT, MPI_INT, recvbuf, recvcounts, rdispls, MPI_INT, 0, comm, req);
break;
case 14: /* MPI_Ialltoall */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
buf[i*COUNT+j] = rank + (i * j);
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Ialltoall(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, comm, req);
break;
case 15: /* MPI_Iallgather */
for (i = 0; i < size*COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iallgather(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, comm, req);
break;
case 16: /* MPI_Iallgatherv */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
recvcounts[i] = COUNT;
rdispls[i] = i * COUNT;
}
for (i = 0; i < COUNT; ++i)
buf[i] = rank + i;
MPI_Iallgatherv(buf, COUNT, MPI_INT, recvbuf, recvcounts, rdispls, MPI_INT, comm, req);
break;
case 17: /* MPI_Iscan */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iscan(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 18: /* MPI_Iexscan */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iexscan(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 19: /* MPI_Ialltoallw (a weak test, neither irregular nor sparse) */
for (i = 0; i < size; ++i) {
sendcounts[i] = COUNT;
recvcounts[i] = COUNT;
sdispls[i] = COUNT * i * sizeof(int);
rdispls[i] = COUNT * i * sizeof(int);
sendtypes[i] = MPI_INT;
recvtypes[i] = MPI_INT;
for (j = 0; j < COUNT; ++j) {
buf[i*COUNT+j] = rank + (i * j);
recvbuf[i*COUNT+j] = 0xdeadbeef;
}
}
MPI_Ialltoallw(buf, sendcounts, sdispls, sendtypes, recvbuf, recvcounts, rdispls, recvtypes, comm, req);
break;
case 20: /* basic pt2pt MPI_Isend/MPI_Irecv pairing */
/* even ranks send to odd ranks, but only if we have a full pair */
if ((rank % 2 != 0) || (rank != size-1)) {
for (j = 0; j < COUNT; ++j) {
buf[j] = j;
recvbuf[j] = 0xdeadbeef;
}
if (rank % 2 == 0)
MPI_Isend(buf, COUNT, MPI_INT, rank+1, 5, comm, req);
else
MPI_Irecv(recvbuf, COUNT, MPI_INT, rank-1, 5, comm, req);
}
break;
default:
fprintf(stderr, "unexpected value for l->case_num=%d)\n", (l->case_num));
MPI_Abort(comm, 1);
exit(1);
break;
}
}
void IMB_iallgatherv(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
/*
MPI-NBC benchmark kernel
Benchmarks MPI_Iallgatherv
Input variables:
-c_info (type struct comm_info*)
Collection of all base data for MPI;
see [1] for more information
-size (type int)
Basic message size in bytes
-ITERATIONS (type struct iter_schedule *)
Repetition scheduling
-RUN_MODE (type MODES)
(only MPI-2 case: see [1])
Output variables:
-time (type double*)
Timing result per sample
*/
{
int i = 0;
Type_Size s_size,
r_size;
int s_num = 0,
r_num = 0;
MPI_Request request;
MPI_Status status;
double t_pure = 0.,
t_comp = 0.,
t_ovrlp = 0.;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->s_data_type, &s_size);
MPI_Type_size(c_info->r_data_type, &r_size);
if ((s_size != 0) && (r_size != 0)) {
s_num = size / s_size;
r_num = size / r_size;
}
if(c_info->rank != -1) {
/* GET PURE TIME. DISPLACEMENT AND RECEIVE COUNT WILL BE INITIALIZED HERE */
IMB_iallgatherv_pure(c_info, size, ITERATIONS, RUN_MODE, &t_pure);
/* INITIALIZATION CALL */
IMB_cpu_exploit(t_pure, 1);
for(i=0; i<N_BARR; i++) {
MPI_Barrier(c_info->communicator);
}
t_ovrlp = MPI_Wtime();
for(i=0; i < ITERATIONS->n_sample; i++)
{
ierr = MPI_Iallgatherv((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
s_num,
c_info->s_data_type,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
c_info->reccnt,
c_info->rdispl,
c_info->r_data_type,
c_info->communicator,
&request);
MPI_ERRHAND(ierr);
t_comp -= MPI_Wtime();
IMB_cpu_exploit(t_pure, 0);
t_comp += MPI_Wtime();
MPI_Wait(&request, &status);
CHK_DIFF("Iallgatherv", c_info,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
0, 0, ((size_t)c_info->num_procs * (size_t)size),
1, put, 0, ITERATIONS->n_sample, i, -2, &defect);
}
t_ovrlp = (MPI_Wtime() - t_ovrlp) / ITERATIONS->n_sample;
t_comp /= ITERATIONS->n_sample;
}
time[0] = t_pure;
time[1] = t_ovrlp;
time[2] = t_comp;
}
void IMB_iallgatherv_pure(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
/*
MPI-NBC benchmark kernel
Benchmarks IMB_Iallgatherv_pure
Input variables:
-c_info (type struct comm_info*)
Collection of all base data for MPI;
see [1] for more information
-size (type int)
Basic message size in bytes
-ITERATIONS (type struct iter_schedule *)
Repetition scheduling
-RUN_MODE (type MODES)
(only MPI-2 case: see [1])
Output variables:
-time (type double*)
Timing result per sample
*/
{
int i = 0;
Type_Size s_size,
r_size;
int s_num = 0,
r_num;
MPI_Request request;
MPI_Status status;
double t_pure = 0.;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->s_data_type, &s_size);
MPI_Type_size(c_info->s_data_type, &r_size);
if ((s_size != 0) && (r_size != 0)) {
s_num = size / s_size;
r_num = size / r_size;
}
for (i=0 ; i < c_info->num_procs; ++i) {
c_info->rdispl[i] = r_num * i;
c_info->reccnt[i] = r_num;
}
if(c_info->rank != -1) {
for (i = 0; i < N_BARR; i++) {
MPI_Barrier(c_info->communicator);
}
t_pure = MPI_Wtime();
for(i = 0; i < ITERATIONS->n_sample; i++)
{
ierr = MPI_Iallgatherv((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
s_num,
c_info->s_data_type,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
c_info->reccnt,
c_info->rdispl,
c_info->r_data_type,
c_info->communicator,
&request);
MPI_ERRHAND(ierr);
MPI_Wait(&request, &status);
CHK_DIFF("Iallgatherv_pure", c_info,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
0, 0, ((size_t)c_info->num_procs * (size_t)size),
1, put, 0, ITERATIONS->n_sample, i, -2, &defect);
}
t_pure = (MPI_Wtime() - t_pure) / ITERATIONS->n_sample;
}
time[0] = t_pure;
}
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;
MPI_Request req;
/* 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_Igather(sbuf, NUM_INTS, MPI_INT, sbuf, NUM_INTS, MPI_INT, 0, comm, &req))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Igatherv(sbuf, NUM_INTS, MPI_INT, sbuf, rcounts, rdispls, MPI_INT, 0, comm, &req))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Iscatter(sbuf, NUM_INTS, MPI_INT, sbuf, NUM_INTS, MPI_INT, 0, comm, &req))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Iscatterv(sbuf, scounts, sdispls, MPI_INT, sbuf, NUM_INTS, MPI_INT, 0, comm, &req))
errs++;
if (MPI_SUCCESS == MPI_Iallgather(&sbuf[rank], 1, MPI_INT, sbuf, 1, MPI_INT, comm, &req))
errs++;
if (MPI_SUCCESS ==
MPI_Iallgatherv(&sbuf[rank * rcounts[rank]], rcounts[rank], MPI_INT, sbuf, rcounts, rdispls,
MPI_INT, comm, &req))
errs++;
if (MPI_SUCCESS == MPI_Ialltoall(sbuf, NUM_INTS, MPI_INT, sbuf, NUM_INTS, MPI_INT, comm, &req))
errs++;
if (MPI_SUCCESS ==
MPI_Ialltoallv(sbuf, scounts, sdispls, MPI_INT, sbuf, scounts, sdispls, MPI_INT, comm,
&req))
errs++;
if (MPI_SUCCESS ==
MPI_Ialltoallw(sbuf, scounts, sdispls, types, sbuf, scounts, sdispls, types, comm, &req))
errs++;
if (rank == 0 && MPI_SUCCESS ==
MPI_Ireduce(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, 0, comm, &req))
errs++;
if (MPI_SUCCESS == MPI_Iallreduce(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req))
errs++;
if (MPI_SUCCESS == MPI_Ireduce_scatter(sbuf, sbuf, rcounts, MPI_INT, MPI_SUM, comm, &req))
errs++;
if (MPI_SUCCESS ==
MPI_Ireduce_scatter_block(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req))
errs++;
if (MPI_SUCCESS == MPI_Iscan(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req))
errs++;
if (MPI_SUCCESS == MPI_Iexscan(sbuf, sbuf, NUM_INTS, MPI_INT, MPI_SUM, comm, &req))
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;
}
