static void *magma_ssytrd_sb2st_parallel_section(void *arg)
{
magma_int_t my_core_id = ((magma_sbulge_id_data*)arg) -> id;
magma_sbulge_data* data = ((magma_sbulge_id_data*)arg) -> data;
magma_int_t allcores_num = data -> threads_num;
magma_int_t n = data -> n;
magma_int_t nb = data -> nb;
magma_int_t nbtiles = data -> nbtiles;
magma_int_t grsiz = data -> grsiz;
magma_int_t Vblksiz = data -> Vblksiz;
magma_int_t compT = data -> compT;
float *A = data -> A;
magma_int_t lda = data -> lda;
float *V = data -> V;
magma_int_t ldv = data -> ldv;
float *TAU = data -> TAU;
float *T = data -> T;
magma_int_t ldt = data -> ldt;
volatile magma_int_t* prog = data -> prog;
pthread_barrier_t* barrier = &(data -> barrier);
//magma_int_t sys_corenbr = 1;
#ifdef ENABLE_TIMER
real_Double_t timeB=0.0, timeT=0.0;
#endif
// with MKL and when using omp_set_num_threads instead of mkl_set_num_threads
// it need that all threads setting it to 1.
magma_setlapack_numthreads(1);
#ifdef MAGMA_SETAFFINITY
//#define PRINTAFFINITY
#ifdef PRINTAFFINITY
affinity_set print_set;
print_set.print_affinity(my_core_id, "starting affinity");
#endif
affinity_set original_set;
affinity_set new_set(my_core_id);
int check = 0;
int check2 = 0;
// bind threads
check = original_set.get_affinity();
if (check == 0) {
check2 = new_set.set_affinity();
if (check2 != 0)
printf("Error in sched_setaffinity (single cpu)\n");
}
else
{
printf("Error in sched_getaffinity\n");
}
#ifdef PRINTAFFINITY
print_set.print_affinity(my_core_id, "set affinity");
#endif
#endif
if(compT==1)
{
/* compute the Q1 overlapped with the bulge chasing+T.
* if all_cores_num=1 it call Q1 on GPU and then bulgechasing.
* otherwise the first thread run Q1 on GPU and
* the other threads run the bulgechasing.
* */
if(allcores_num==1)
{
//=========================
// bulge chasing
//=========================
#ifdef ENABLE_TIMER
timeB = magma_wtime();
#endif
magma_stile_bulge_parallel(0, 1, A, lda, V, ldv, TAU, n, nb, nbtiles, grsiz, Vblksiz, prog);
#ifdef ENABLE_TIMER
timeB = magma_wtime()-timeB;
printf(" Finish BULGE timing= %f \n" ,timeB);
#endif
//=========================
// compute the T's to be used when applying Q2
//=========================
#ifdef ENABLE_TIMER
timeT = magma_wtime();
#endif
magma_stile_bulge_computeT_parallel(0, 1, V, ldv, TAU, T, ldt, n, nb, Vblksiz);
#ifdef ENABLE_TIMER
timeT = magma_wtime()-timeT;
printf(" Finish T's timing= %f \n" ,timeT);
#endif
}else{ // allcore_num > 1
magma_int_t id = my_core_id;
magma_int_t tot = allcores_num;
//=========================
// bulge chasing
//=========================
#ifdef ENABLE_TIMER
if(id == 0)
timeB = magma_wtime();
#endif
magma_stile_bulge_parallel(id, tot, A, lda, V, ldv, TAU, n, nb, nbtiles, grsiz, Vblksiz, prog);
pthread_barrier_wait(barrier);
#ifdef ENABLE_TIMER
if(id == 0){
timeB = magma_wtime()-timeB;
printf(" Finish BULGE timing= %f \n" ,timeB);
}
#endif
//=========================
// compute the T's to be used when applying Q2
//=========================
#ifdef ENABLE_TIMER
if(id == 0)
timeT = magma_wtime();
#endif
magma_stile_bulge_computeT_parallel(id, tot, V, ldv, TAU, T, ldt, n, nb, Vblksiz);
pthread_barrier_wait(barrier);
#ifdef ENABLE_TIMER
if (id == 0){
timeT = magma_wtime()-timeT;
printf(" Finish T's timing= %f \n" ,timeT);
}
#endif
} // allcore == 1
}else{ // WANTZ = 0
//=========================
// bulge chasing
//=========================
#ifdef ENABLE_TIMER
if(my_core_id == 0)
timeB = magma_wtime();
#endif
magma_stile_bulge_parallel(my_core_id, allcores_num, A, lda, V, ldv, TAU, n, nb, nbtiles, grsiz, Vblksiz, prog);
pthread_barrier_wait(barrier);
#ifdef ENABLE_TIMER
if(my_core_id == 0){
timeB = magma_wtime()-timeB;
printf(" Finish BULGE timing= %f \n" ,timeB);
}
#endif
} // WANTZ > 0
#ifdef MAGMA_SETAFFINITY
// unbind threads
if (check == 0){
check2 = original_set.set_affinity();
if (check2 != 0)
printf("Error in sched_setaffinity (restore cpu list)\n");
}
#ifdef PRINTAFFINITY
print_set.print_affinity(my_core_id, "restored_affinity");
#endif
#endif
return 0;
}
static void *magma_ssytrd_hb2st_parallel_section(void *arg)
{
magma_int_t my_core_id = ((magma_sbulge_id_data*)arg) -> id;
magma_sbulge_data* data = ((magma_sbulge_id_data*)arg) -> data;
magma_int_t allcores_num = data -> threads_num;
magma_int_t n = data -> n;
magma_int_t nb = data -> nb;
magma_int_t nbtiles = data -> nbtiles;
magma_int_t grsiz = data -> grsiz;
magma_int_t Vblksiz = data -> Vblksiz;
magma_int_t compT = data -> compT;
float *A = data -> A;
magma_int_t lda = data -> lda;
float *V = data -> V;
magma_int_t ldv = data -> ldv;
float *TAU = data -> TAU;
float *T = data -> T;
magma_int_t ldt = data -> ldt;
volatile magma_int_t* prog = data -> prog;
pthread_barrier_t* barrier = &(data -> barrier);
magma_int_t sys_corenbr = 1;
float timeB=0.0, timeT=0.0;
#if defined(SETAFFINITY)
// bind threads
cpu_set_t set;
// bind threads
CPU_ZERO( &set );
CPU_SET( my_core_id, &set );
sched_setaffinity( 0, sizeof(set), &set) ;
#endif
if(compT==1)
{
/* compute the Q1 overlapped with the bulge chasing+T.
* if all_cores_num=1 it call Q1 on GPU and then bulgechasing.
* otherwise the first thread run Q1 on GPU and
* the other threads run the bulgechasing.
* */
if(allcores_num==1)
{
//=========================
// bulge chasing
//=========================
timeB = magma_wtime();
magma_stile_bulge_parallel(0, 1, A, lda, V, ldv, TAU, n, nb, nbtiles, grsiz, Vblksiz, prog);
timeB = magma_wtime()-timeB;
printf(" Finish BULGE timing= %f \n" ,timeB);
//=========================
// compute the T's to be used when applying Q2
//=========================
timeT = magma_wtime();
magma_stile_bulge_computeT_parallel(0, 1, V, ldv, TAU, T, ldt, n, nb, Vblksiz);
timeT = magma_wtime()-timeT;
printf(" Finish T's timing= %f \n" ,timeT);
}else{ // allcore_num > 1
magma_int_t id = my_core_id;
magma_int_t tot = allcores_num;
//=========================
// bulge chasing
//=========================
if(id == 0)timeB = magma_wtime();
magma_stile_bulge_parallel(id, tot, A, lda, V, ldv, TAU, n, nb, nbtiles, grsiz, Vblksiz, prog);
pthread_barrier_wait(barrier);
if(id == 0){
timeB = magma_wtime()-timeB;
printf(" Finish BULGE timing= %f \n" ,timeB);
}
//=========================
// compute the T's to be used when applying Q2
//=========================
if(id == 0)timeT = magma_wtime();
magma_stile_bulge_computeT_parallel(id, tot, V, ldv, TAU, T, ldt, n, nb, Vblksiz);
pthread_barrier_wait(barrier);
if (id == 0){
timeT = magma_wtime()-timeT;
printf(" Finish T's timing= %f \n" ,timeT);
}
} // allcore == 1
}else{ // WANTZ = 0
//=========================
// bulge chasing
//=========================
if(my_core_id == 0)
timeB = magma_wtime();
magma_stile_bulge_parallel(my_core_id, allcores_num, A, lda, V, ldv, TAU, n, nb, nbtiles, grsiz, Vblksiz, prog);
pthread_barrier_wait(barrier);
if(my_core_id == 0){
timeB = magma_wtime()-timeB;
printf(" Finish BULGE timing= %f \n" ,timeB);
}
} // WANTZ > 0
#if defined(SETAFFINITY)
// unbind threads
sys_corenbr = sysconf(_SC_NPROCESSORS_ONLN);
CPU_ZERO( &set );
for(magma_int_t i=0; i<sys_corenbr; i++)
CPU_SET( i, &set );
sched_setaffinity( 0, sizeof(set), &set) ;
#endif
return 0;
}
