void IMB_ibarrier(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
{
int i = 0;
MPI_Request request;
MPI_Status status;
double t_pure = 0.,
t_comp = 0.,
t_ovrlp = 0.;
#ifdef CHECK
defect=0.;
#endif
ierr = 0;
if(c_info->rank != -1) {
IMB_ibarrier_pure(c_info, size, ITERATIONS, RUN_MODE, &t_pure);
/* INITIALIZATION CALL */
IMB_cpu_exploit(t_pure, 1);
IMB_do_n_barriers (c_info->communicator, N_BARR);
t_ovrlp = MPI_Wtime();
for(i=0; i < ITERATIONS->n_sample; i++) {
ierr = MPI_Ibarrier(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);
}
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_ialltoall(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)
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) {
IMB_ialltoall_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_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);
t_comp -= MPI_Wtime();
IMB_cpu_exploit(t_pure, 0);
t_comp += MPI_Wtime();
MPI_Wait(&request, &status);
CHK_DIFF("Ialltoall", 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_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_iallreduce(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.,
t_comp = 0.,
t_ovrlp = 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_iallreduce_pure(c_info, size, ITERATIONS, RUN_MODE, &t_pure);
/* INITIALIZATION CALL */
IMB_cpu_exploit(t_pure, 1);
IMB_do_n_barriers (c_info->communicator, N_BARR);
for(i = 0; i < ITERATIONS->n_sample; i++)
{
t_ovrlp -= 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);
t_comp -= MPI_Wtime();
IMB_cpu_exploit(t_pure, 0);
t_comp += MPI_Wtime();
MPI_Wait(&request, &status);
t_ovrlp += MPI_Wtime();
CHK_DIFF("Iallreduce", 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_ovrlp /= ITERATIONS->n_sample;
t_comp /= ITERATIONS->n_sample;
}
time[0] = t_pure;
time[1] = t_ovrlp;
time[2] = t_comp;
}
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_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_igatherv(struct comm_info* c_info,
int size,
struct iter_schedule* ITERATIONS,
MODES RUN_MODE,
double* time)
/*
MPI-NBC benchmark kernel
Benchmarks MPI_Igatherv
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) {
int root = 0;
/* GET PURE TIME. DISPLACEMENT AND RECEIVE COUNT WILL BE INITIALIZED HERE */
IMB_igatherv_pure(c_info, size, ITERATIONS, RUN_MODE, &t_pure);
/* INITIALIZATION CALL */
IMB_cpu_exploit(t_pure, 1);
IMB_do_n_barriers(c_info->communicator, N_BARR);
for(i=0; i < ITERATIONS->n_sample; i++)
{
t_ovrlp -= MPI_Wtime();
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,
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);
t_ovrlp += MPI_Wtime();
#ifdef CHECK
if (c_info->rank == root) {
CHK_DIFF("Igatherv", 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
root = (root + c_info->root_shift) % c_info->num_procs;
IMB_do_n_barriers(c_info->communicator, c_info->sync);
}
t_ovrlp /= ITERATIONS->n_sample;
t_comp /= ITERATIONS->n_sample;
}
time[0] = t_pure;
time[1] = t_ovrlp;
time[2] = t_comp;
}
