void IMB_iallreduce_pure(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
{
int i = 0;
Type_Size s_size;
int s_num = 0;
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->red_data_type, &s_size);
if (s_size != 0)
{
s_num = size / s_size;
}
if(c_info->rank != -1)
{
IMB_do_n_barriers (c_info->communicator, N_BARR);
for(i = 0; i < ITERATIONS->n_sample; i++)
{
t_pure -= MPI_Wtime();
ierr = MPI_Iallreduce((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
s_num,
c_info->red_data_type,
c_info->op_type,
c_info->communicator,
&request);
MPI_ERRHAND(ierr);
MPI_Wait(&request, &status);
t_pure += MPI_Wtime();
CHK_DIFF("Iallreduce_pure", c_info,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
0, size, size, asize, put, 0, ITERATIONS->n_sample, i, -1, &defect);
IMB_do_n_barriers (c_info->communicator, c_info->sync);
}
t_pure /= ITERATIONS->n_sample;
}
time[0] = t_pure;
}
void IMB_reduce_scatter(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Reduce_scatter
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
*/
{
double t1, t2;
int i;
size_t pos1,pos2;
#ifdef CHECK
size_t pos;
int Locsize;
#endif
Type_Size s_size;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->red_data_type,&s_size);
for (i=0;i<c_info->num_procs ;i++)
{
if( size > 0)
{
IMB_get_rank_portion(i, c_info->num_procs, size, s_size,
&pos1, &pos2);
c_info->reccnt[i] = (pos2-pos1+1)/s_size;
#ifdef CHECK
if( i==c_info->rank ) {pos=pos1; Locsize= s_size*c_info->reccnt[i];}
#endif
} else
{
c_info->reccnt[i] = 0;
#ifdef CHECK
if( i==c_info->rank ) {pos=0; Locsize= 0;}
#endif
}
}
if(c_info->rank!=-1)
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
ierr = MPI_Reduce_scatter ((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
c_info->reccnt,
c_info->red_data_type,c_info->op_type,
c_info->communicator);
MPI_ERRHAND(ierr);
CHK_DIFF("Reduce_scatter",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
pos,
Locsize, size, asize,
put, 0, ITERATIONS->n_sample, i,
-1, &defect);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
} else /*if(c_info->rank==-1)*/
{
*time = 0.;
}
}
void IMB_igatherv_pure(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
/*
MPI-NBC benchmark kernel
Benchmarks IMB_Igatherv_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;
}
/* INITIALIZATION OF DISPLACEMENT and RECEIVE COUNTS */
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_Igatherv((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,
i % c_info->num_procs, // root = round robin
c_info->communicator,
&request);
MPI_ERRHAND(ierr);
MPI_Wait(&request, &status);
#ifdef CHECK
if (c_info->rank == i % c_info->num_procs) {
CHK_DIFF("Igatherv_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);
}
#endif // CHECK
}
t_pure = (MPI_Wtime() - t_pure) / ITERATIONS->n_sample;
}
time[0] = t_pure;
}
void IMB_allreduce(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Allreduce
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
*/
{
double t1, t2;
int i;
Type_Size s_size;
int s_num;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
*time = 0.;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->red_data_type,&s_size);
if (s_size!=0) s_num=size/s_size;
if(c_info->rank!=-1)
{
IMB_do_n_barriers (c_info->communicator, N_BARR);
for(i=0;i< ITERATIONS->n_sample;i++)
{
t1 = MPI_Wtime();
ierr = MPI_Allreduce((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
s_num,
c_info->red_data_type,c_info->op_type,
c_info->communicator);
MPI_ERRHAND(ierr);
t2 = MPI_Wtime();
*time += (t2 - t1);
CHK_DIFF("Allreduce",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs, 0,
size, size, asize,
put, 0, ITERATIONS->n_sample, i,
-1, &defect);
IMB_do_n_barriers (c_info->communicator, c_info->sync);
}
*time /= ITERATIONS->n_sample;
}
}
void IMB_sendrecv(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Sendrecv
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
*/
{
double t1,t2;
int i;
Type_Size s_size, r_size;
int s_num,r_num;
int s_tag, r_tag;
int dest, source;
MPI_Status stat;
#ifdef CHECK
defect=0;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE's in s_size and r_size */
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;
}
s_tag = 1;
r_tag = MPI_ANY_TAG;
if(c_info->rank!=-1)
{
/* CALCULATE SOURCE AND DESTINATION */
dest = (c_info->rank + 1) % (c_info->num_procs);
source = (c_info->rank + c_info->num_procs-1) % (c_info->num_procs);
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
ierr= MPI_Sendrecv((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num,c_info->s_data_type, dest,s_tag,
(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
r_num,c_info->r_data_type,source,r_tag,
c_info->communicator,&stat);
MPI_ERRHAND(ierr);
CHK_DIFF("Sendrecv",c_info,(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
0, size, size, asize,
put, 0, ITERATIONS->n_sample, i,
source, &defect);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
}
else
{
*time = 0.;
}
}
void IMB_bcast(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Bcast
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
*/
{
double t1, t2;
int i,i1;
Type_Size s_size;
int s_num;
void* bc_buf;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->s_data_type,&s_size);
if (s_size!=0) s_num=size/s_size;
if(c_info->rank!=-1)
{
i1=0;
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
/* Provide that s_buffer is not overwritten */
bc_buf = (i1 == c_info->rank) ? c_info->s_buffer : c_info->r_buffer;
ierr= MPI_Bcast((char*)bc_buf+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num,c_info->s_data_type,
i1,c_info->communicator);
MPI_ERRHAND(ierr);
CHK_DIFF("Bcast", c_info,
(char*)bc_buf + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
0, size, size, 1, put, 0, ITERATIONS->n_sample, i, i1, &defect);
/* CHANGE THE ROOT NODE */
i1=(++i1)%c_info->num_procs;
}
t2 = MPI_Wtime();
*time=(t2 - t1)/(ITERATIONS->n_sample);
}
else
{
*time = 0.;
}
}
void IMB_ireduce_scatter(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
{
int i = 0;
Type_Size s_size;
MPI_Request request;
MPI_Status status;
double t_pure = 0.,
t_comp = 0.,
t_ovrlp = 0.;
#ifdef CHECK
size_t pos = 0,
pos1 = 0,
pos2 = 0;
int Locsize = 0;
defect = 0.;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->red_data_type, &s_size);
#ifdef CHECK
if(size > 0) {
for (i = 0; i < c_info->num_procs; i++) {
IMB_get_rank_portion(i, c_info->num_procs, size, s_size, &pos1, &pos2);
if (i == c_info->rank) {
pos = pos1;
Locsize = s_size * c_info->reccnt[i];
}
}
}
#endif // CHECK
if(c_info->rank != -1) {
IMB_ireduce_scatter_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_Ireduce_scatter((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
c_info->reccnt,
c_info->red_data_type,
c_info->op_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("Ireduce_scatter", c_info,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
pos, Locsize, size, asize, put, 0, ITERATIONS->n_sample, i, -1, &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_pingping(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
2 process exchange; MPI_Isend + MPI_Recv
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
*/
{
double t1,t2;
int i;
Type_Size s_size, r_size;
int s_num, r_num;
int s_tag, r_tag;
int dest, source;
MPI_Status stat;
MPI_Request request;
#ifdef CHECK
defect=0;
#endif
ierr = 0;
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;
}
s_tag = 1;
r_tag = c_info->select_tag ? s_tag : MPI_ANY_TAG;
dest = -1;
if (c_info->rank == c_info->pair0)
dest = c_info->pair1;
else if (c_info->rank == c_info->pair1)
dest = c_info->pair0;
source = c_info->select_source ? dest : MPI_ANY_SOURCE;
if( dest != -1 )
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
ierr= MPI_Isend((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num,
c_info->s_data_type,dest,s_tag,
c_info->communicator,&request);
MPI_ERRHAND(ierr);
ierr = MPI_Recv((char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
r_num,c_info->r_data_type,source,
r_tag,c_info->communicator,&stat);
MPI_ERRHAND(ierr);
ierr = MPI_Wait(&request, &stat);
MPI_ERRHAND(ierr);
CHK_DIFF("PingPing",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
0, size, size, asize,
put, 0, ITERATIONS->n_sample, i,
dest, &defect);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
}
else
{
*time = 0.;
}
}
void IMB_shr_spmd_swapm(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks shr_spmd_swapm
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
*/
{
double t1, t2;
int i;
Type_Size s_size,r_size;
int s_num, r_num;
MPI_Datatype s_data_types[c_info->num_procs];
MPI_Datatype r_data_types[c_info->num_procs];
static bool handshake, isend, firstpass=true;
static int maxreqs;
int flow_cntl;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
if(firstpass){
firstpass = false;
shr_swapm_getenv(&handshake, &isend, &maxreqs);
}
if(maxreqs >= 0){
flow_cntl = min(c_info->num_procs, maxreqs);
}else{
flow_cntl = max(2,-1* c_info->num_procs/maxreqs);
}
if(c_info->rank==0){
printf("SPMD_SWAPM: handshake %d isend %d flow_cntl = %d\n",handshake, isend,flow_cntl );
}
/* 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;
}
/* INITIALIZATION OF DISPLACEMENT and SEND/RECEIVE COUNTS */
for (i=0;i<c_info->num_procs ;i++)
{
s_data_types[i] = c_info->s_data_type;
r_data_types[i] = c_info->r_data_type;
c_info->sdispl[i] = s_num*i;
c_info->sndcnt[i] = s_num;
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);
t1 = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
ierr = shr_spmd_swapm((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
c_info->sndcnt,c_info->sdispl,
s_data_types,
(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
c_info->reccnt,c_info->rdispl,
r_data_types,
c_info->communicator, handshake, isend, flow_cntl);
MPI_ERRHAND(ierr);
CHK_DIFF("Alltoallw",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
(size_t) c_info->rank* (size_t) size, 0,
(size_t) c_info->num_procs* (size_t) size, 1,
put, 0, ITERATIONS->n_sample, i,
-2, &defect);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
}
else
{
*time = 0.;
}
}
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_gatherv(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Gatherv
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
*/
{
double t1, t2;
int i;
Type_Size s_size,r_size;
int s_num, r_num;
#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;
}
/* INITIALIZATION OF DISPLACEMENT and RECEIVE COUNTS */
for (i=0;i<c_info->num_procs ;i++)
{
c_info->rdispl[i] = r_num*i;
c_info->reccnt[i] = r_num;
}
*time = 0.;
if(c_info->rank!=-1)
{
int root = 0;
IMB_do_n_barriers(c_info->communicator, N_BARR);
for(i=0;i<ITERATIONS->n_sample;i++)
{
t1 = MPI_Wtime();
ierr = MPI_Gatherv((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,
root,
c_info->communicator);
MPI_ERRHAND(ierr);
t2 = MPI_Wtime();
*time += (t2 - t1);
#ifdef CHECK
if( c_info->rank == root )
{
CHK_DIFF("Gatherv",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);
}
#endif
root = (root + c_info->root_shift) % c_info->num_procs;
IMB_do_n_barriers(c_info->communicator, c_info->sync);
}
*time /= ITERATIONS->n_sample;
}
}
void IMB_accumulate (struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-2 benchmark kernel
Benchmarks MPI_Accumulate
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)
Mode (aggregate/non aggregate; blocking/nonblocking);
see "IMB_benchmark.h" for definition
Output variables:
-time (type double*)
Timing result per sample
*/
{
double t1, t2;
Type_Size s_size,r_size;
int s_num, r_num;
/* IMB 3.1 << */
int r_off;
/* >> IMB 3.1 */
int s_tag, r_tag;
int dest, source, root;
int i;
MPI_Status stat;
#ifdef CHECK
defect=0;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->red_data_type,&s_size);
/* IMB 3.1 << */
s_num=size/s_size;
r_size=s_size;
r_num=s_num;
r_off=ITERATIONS->r_offs/r_size;
/* >> IMB 3.1 */
root = (c_info-> rank == 0);
if( c_info-> rank < 0 )
*time = 0.;
else
{
if( !RUN_MODE->AGGREGATE )
{
*time = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
ierr = MPI_Accumulate(
(char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num, c_info->red_data_type,
0, i%ITERATIONS->r_cache_iter*r_off,
r_num, c_info->red_data_type, c_info->op_type,
c_info->WIN );
MPI_ERRHAND(ierr);
ierr = MPI_Win_fence(0, c_info->WIN);
MPI_ERRHAND(ierr);
#ifdef CHECK
if( root )
{
CHK_DIFF("Accumulate",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
0, size, size, asize,
put, 0, ITERATIONS->n_sample, i,
-1, &defect);
IMB_ass_buf((char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs, 0, 0, size-1, 0);
}
MPI_Barrier(c_info->communicator);
#endif
}
*time=(MPI_Wtime()-*time)/ITERATIONS->n_sample;
}
if( RUN_MODE->AGGREGATE )
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
*time = MPI_Wtime();
#ifdef CHECK
for(i=0;i< ITERATIONS->r_cache_iter; i++)
#else
for(i=0;i< ITERATIONS->n_sample;i++)
#endif
{
ierr = MPI_Accumulate(
(char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num, c_info->red_data_type,
0, i%ITERATIONS->r_cache_iter*r_off,
r_num, c_info->red_data_type, c_info->op_type,
c_info->WIN );
MPI_ERRHAND(ierr);
}
ierr = MPI_Win_fence(0, c_info->WIN);
MPI_ERRHAND(ierr);
*time=(MPI_Wtime()-*time)/ITERATIONS->n_sample;
#ifdef CHECK
if( root )
{
for(i=0;i< ITERATIONS->r_cache_iter; i++)
{
CHK_DIFF("Accumulate", c_info, (char*)c_info->r_buffer+i*ITERATIONS->r_offs,
0, size, size, asize,
put, 0, ITERATIONS->n_sample, i,
-1, &defect);
}
}
#endif
}
}
}
void IMB_exchange(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Chainwise exchange; MPI_Isend (left+right) + MPI_Recv (right+left)
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
*/
{
double t1, t2;
int i;
Type_Size s_size, r_size;
int s_num, r_num;
int s_tag, r_tag;
int left, right;
MPI_Status stat[2];
MPI_Request request[2];
#ifdef CHECK
defect=0;
#endif
ierr = 0;
/*GET SIZE OF DATA TYPE's in s_size and r_size*/
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;
}
s_tag = 1;
r_tag = c_info->select_tag ? s_tag : MPI_ANY_TAG;
if(c_info->rank != -1)
{
if(c_info->rank < c_info->num_procs-1) right = c_info->rank+1;
if(c_info->rank > 0) left = c_info->rank-1;
if(c_info->rank == c_info->num_procs-1) right = 0;
if(c_info->rank == 0) left = c_info->num_procs-1 ;
if((c_info->rank >= 0) && (c_info->rank <= c_info->num_procs-1))
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0; i< ITERATIONS->n_sample; i++)
{
ierr= MPI_Isend((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num,c_info->s_data_type,
right,s_tag,c_info->communicator,&request[0]);
MPI_ERRHAND(ierr);
ierr= MPI_Isend((char*)c_info->s_buffer+size+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
s_num,c_info->s_data_type,
left ,s_tag,c_info->communicator,&request[1]);
MPI_ERRHAND(ierr);
ierr= MPI_Recv( (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
r_num,c_info->r_data_type,
left ,r_tag,c_info->communicator,stat);
MPI_ERRHAND(ierr);
CHK_DIFF("Exchange",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
0, size, size, 1,
put, 0, ITERATIONS->n_sample, i,
left, &defect);
ierr= MPI_Recv( (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
r_num,c_info->r_data_type,
right,r_tag,c_info->communicator,stat);
MPI_ERRHAND(ierr);
CHK_DIFF("Exchange",c_info, (char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
s_num, size, size, 1,
put, 0, ITERATIONS->n_sample, i,
right, &defect);
ierr= MPI_Waitall(2,request,stat);
MPI_ERRHAND(ierr);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
}
}
else
{
*time = 0.;
}
}
void Accumulate (struct comm_info* c_info,
int size,int n_sample,MODES RUN_MODE,double* time)
/*************************************************************************/
/*------------------------------------------------------------
VARIABLE | TYPE | MEANING
------------------------------------------------------------
Input : c_info | struct comm_info* | see comm_info.h
size | int | message length in byte
n_sample | int | repetition count
RUN_MODE | MODES (typedef, | Distinction aggregate/
| see Benchmark.h) | non aggr., see docu.
| |
Output : time | double* | *time: time/sample in usec
| |
In/Out : - | - | -
| |
------------------------------------------------------------
------------------------------------------------------------
Description: see the accompanying document
-------------------------------------------------------------*/
{
double t1, t2;
Type_Size s_size,r_size;
int s_num, r_num;
int s_tag, r_tag;
int dest, source, root;
int i;
MPI_Status stat;
#ifdef CHECK
defect=0;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->red_data_type,&s_size);
if (s_size!=0) s_num=size/s_size;
root = (c_info-> rank == 0);
if( c_info-> rank < 0 )
*time = 0.;
else
{
if( !RUN_MODE->AGGREGATE )
{
*time = MPI_Wtime();
for(i=0;i< n_sample;i++)
{
ierr = MPI_Accumulate
(c_info->s_buffer, s_num, c_info->red_data_type,
0, i*s_num, s_num, c_info->red_data_type, c_info->op_type,
c_info->WIN );
MPI_ERRHAND(ierr);
ierr = MPI_Win_fence(0, c_info->WIN);
MPI_ERRHAND(ierr);
#ifdef CHECK
if( root )
{
CHK_DIFF("Accumulate",c_info, (void*)(c_info->r_data+i*s_num), 0,
size, size, asize,
put, 0, n_sample, i,
-1, &defect);
ass_buf(c_info->r_buffer, 0, 0, size-1, 0);
}
MPI_Barrier(c_info->communicator);
#endif
}
*time=(MPI_Wtime()-*time)/n_sample;
}
if( RUN_MODE->AGGREGATE )
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
*time = MPI_Wtime();
for(i=0;i< n_sample;i++)
{
ierr = MPI_Accumulate
((void*)(c_info->s_data+i*s_num), s_num, c_info->red_data_type,
0, i*s_num, s_num, c_info->red_data_type, c_info->op_type,
c_info->WIN );
MPI_ERRHAND(ierr);
}
ierr = MPI_Win_fence(0, c_info->WIN);
MPI_ERRHAND(ierr);
*time=(MPI_Wtime()-*time)/n_sample;
#ifdef CHECK
if( root )
{
CHK_DIFF("Accumulate",c_info, c_info->r_buffer, 0,
n_sample*size, n_sample*size, asize,
put, 0, n_sample, -1,
-1, &defect);
}
#endif
}
}
}
void IMB_ireduce_scatter_pure(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
{
int i = 0;
Type_Size s_size;
size_t pos1 = 0,
pos2 = 0;
MPI_Request request;
MPI_Status status;
double t_pure = 0.;
#ifdef CHECK
size_t pos = 0;
int Locsize = 0;
defect = 0.;
#endif
ierr = 0;
/* GET SIZE OF DATA TYPE */
MPI_Type_size(c_info->red_data_type, &s_size);
for (i = 0; i < c_info->num_procs; i++) {
if( size > 0) {
IMB_get_rank_portion(i, c_info->num_procs, size, s_size, &pos1, &pos2);
c_info->reccnt[i] = (pos2 - pos1 + 1) / s_size;
#ifdef CHECK
if (i == c_info->rank) {
pos = pos1;
Locsize = s_size * c_info->reccnt[i];
}
#endif
} else {
c_info->reccnt[i] = 0;
#ifdef CHECK
if (i == c_info->rank) {
pos=0;
Locsize = 0;
}
#endif
}
}
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_Ireduce_scatter((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
c_info->reccnt,
c_info->red_data_type,
c_info->op_type,
c_info->communicator,
&request);
MPI_ERRHAND(ierr);
MPI_Wait(&request, &status);
CHK_DIFF("Ireduce_scatter_pure", c_info,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
pos, Locsize, size, asize, put, 0, ITERATIONS->n_sample, i, -1, &defect);
}
t_pure = (MPI_Wtime() - t_pure) / ITERATIONS->n_sample;
}
time[0] = t_pure;
}
void IMB_ialltoall_pure(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
/*
MPI-NBC benchmark kernel
Benchmarks MPI_Ialltoall.
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;
}
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_Ialltoall((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,
r_num,
c_info->r_data_type,
c_info->communicator,
&request);
MPI_ERRHAND(ierr);
MPI_Wait(&request, &status);
CHK_DIFF("Ialltoall_pure", c_info,
(char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
((size_t)c_info->rank * (size_t) size), 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;
}
void IMB_ibcast(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
/*
MPI-NBC benchmark kernel
Benchmarks MPI_Ibcast
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,
root = 0;
Type_Size s_size;
int s_num = 0;
void* bc_buf = NULL;
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);
if (s_size != 0) {
s_num = size / s_size;
}
if(c_info->rank != -1) {
IMB_ibcast_pure(c_info, size, ITERATIONS, RUN_MODE, &t_pure);
/* INITIALIZATION CALL */
IMB_cpu_exploit(t_pure, 1);
root = 0;
for(i=0; i<N_BARR; i++) {
MPI_Barrier(c_info->communicator);
}
t_ovrlp = MPI_Wtime();
for(i=0; i < ITERATIONS->n_sample; i++)
{
bc_buf = (root == c_info->rank)
? c_info->s_buffer
: c_info->r_buffer;
ierr = MPI_Ibcast((char*)bc_buf + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
s_num,
c_info->s_data_type,
root,
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("Ibcast", c_info,
(char*)bc_buf + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
0, size, size, 1, put, 0, ITERATIONS->n_sample, i, root, &defect);
/* CHANGE THE ROOT NODE */
root = (++root) % c_info->num_procs;
}
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_allgather(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Allgather
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
*/
{
double t1, t2;
int i;
Type_Size s_size,r_size;
int s_num, r_num;
#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)
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0;i< ITERATIONS->n_sample;i++)
{
ierr = MPI_Allgather((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,
r_num,c_info->r_data_type,
c_info->communicator);
MPI_ERRHAND(ierr);
CHK_DIFF("Allgather",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);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
}
else
{
*time = 0.;
}
}
void IMB_scatterv(struct comm_info* c_info, int size, struct iter_schedule* ITERATIONS,
MODES RUN_MODE, double* time)
/*
MPI-1 benchmark kernel
Benchmarks MPI_Scatterv
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
*/
{
double t1, t2;
int i;
Type_Size s_size,r_size;
int s_num, r_num;
#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;
}
/* INITIALIZATION OF DISPLACEMENT and RECEIVE COUNTS */
for (i=0;i<c_info->num_procs ;i++)
{
c_info->sdispl[i] = s_num*i;
c_info->sndcnt[i] = s_num;
}
if(c_info->rank!=-1)
{
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
t1 = MPI_Wtime();
for(i=0;i<ITERATIONS->n_sample;i++)
{
ierr = MPI_Scatterv((char*)c_info->s_buffer+i%ITERATIONS->s_cache_iter*ITERATIONS->s_offs,
c_info->sndcnt,c_info->sdispl, c_info->s_data_type,
(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
// root = round robin
r_num, c_info->r_data_type, i%c_info->num_procs,
c_info->communicator);
MPI_ERRHAND(ierr);
CHK_DIFF("Scatterv",c_info,
(char*)c_info->r_buffer+i%ITERATIONS->r_cache_iter*ITERATIONS->r_offs,
c_info->sdispl[c_info->rank], size, size, 1,
put, 0, ITERATIONS->n_sample, i,
i%c_info->num_procs, &defect);
}
t2 = MPI_Wtime();
*time=(t2 - t1)/ITERATIONS->n_sample;
}
else
{
*time = 0.;
}
}
