//##################################################################################################
static void *magma_capplyQ_parallel_section(void *arg)
{
magma_int_t my_core_id = ((magma_capplyQ_id_data*)arg) -> id;
magma_capplyQ_data* data = ((magma_capplyQ_id_data*)arg) -> data;
magma_int_t allcores_num = data -> threads_num;
magma_int_t n = data -> n;
magma_int_t ne = data -> ne;
magma_int_t n_gpu = data -> n_gpu;
magma_int_t nb = data -> nb;
magma_int_t Vblksiz = data -> Vblksiz;
magmaFloatComplex *E = data -> E;
magma_int_t lde = data -> lde;
magmaFloatComplex *V = data -> V;
magma_int_t ldv = data -> ldv;
magmaFloatComplex *TAU = data -> TAU;
magmaFloatComplex *T = data -> T;
magma_int_t ldt = data -> ldt;
magmaFloatComplex *dE = data -> dE;
magma_int_t ldde = data -> ldde;
pthread_barrier_t* barrier = &(data -> barrier);
magma_int_t info;
#ifdef ENABLE_TIMER
real_Double_t timeQcpu=0.0, timeQgpu=0.0;
#endif
magma_int_t n_cpu = ne - n_gpu;
// 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(my_core_id==0)
{
//=============================================
// on GPU on thread 0:
// - apply V2*Z(:,1:N_GPU)
//=============================================
#ifdef ENABLE_TIMER
timeQgpu = magma_wtime();
#endif
magma_csetmatrix(n, n_gpu, E, lde, dE, ldde);
magma_cbulge_applyQ_v2('L', n_gpu, n, nb, Vblksiz, dE, ldde, V, ldv, T, ldt, &info);
magma_device_sync();
#ifdef ENABLE_TIMER
timeQgpu = magma_wtime()-timeQgpu;
printf(" Finish Q2_GPU GGG timing= %f \n" ,timeQgpu);
#endif
} else {
//=============================================
// on CPU on threads 1:allcores_num-1:
// - apply V2*Z(:,N_GPU+1:NE)
//=============================================
#ifdef ENABLE_TIMER
if(my_core_id == 1)
timeQcpu = magma_wtime();
#endif
magma_int_t n_loc = magma_ceildiv(n_cpu, allcores_num-1);
magmaFloatComplex* E_loc = E + (n_gpu+ n_loc * (my_core_id-1))*lde;
n_loc = min(n_loc,n_cpu - n_loc * (my_core_id-1));
magma_ctile_bulge_applyQ(my_core_id, 'L', n_loc, n, nb, Vblksiz, E_loc, lde, V, ldv, TAU, T, ldt);
pthread_barrier_wait(barrier);
#ifdef ENABLE_TIMER
if(my_core_id == 1) {
timeQcpu = magma_wtime()-timeQcpu;
printf(" Finish Q2_CPU CCC timing= %f \n" ,timeQcpu);
}
#endif
} // END if my_core_id
#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;
}
extern "C" magma_int_t
magma_cbulge_back(magma_int_t threads, char uplo,
magma_int_t n, magma_int_t nb,
magma_int_t ne, magma_int_t Vblksiz,
magmaFloatComplex *Z, magma_int_t ldz,
magmaFloatComplex *dZ, magma_int_t lddz,
magmaFloatComplex *V, magma_int_t ldv,
magmaFloatComplex *TAU,
magmaFloatComplex *T, magma_int_t ldt,
magma_int_t* info)
{
magma_setlapack_numthreads(1);
float timeaplQ2=0.0;
float f= 1.;
magma_int_t n_gpu = ne;
//#if defined(PRECISION_s) || defined(PRECISION_d)
//float gpu_cpu_perf = 50; // gpu over cpu performance //100% ev // SandyB. - Kepler (K20c)
//float gpu_cpu_perf = 16; // gpu over cpu performance //100% ev // SandyB. - Fermi (M2090)
//#else
// float gpu_cpu_perf = 27.5; // gpu over cpu performance //100% ev // Westmere - Fermi (M2090)
//float gpu_cpu_perf = 37; // gpu over cpu performance //100% ev // SandyB. - Kepler (K20c)
// float gpu_cpu_perf = 130; // gpu over cpu performance //100% ev // Bulldozer - Kepler (K20X)
//#endif
magma_int_t gpu_cpu_perf = magma_get_cbulge_gcperf();
if(threads>1) {
f = 1. / (1. + (float)(threads-1)/ ((float)gpu_cpu_perf) );
n_gpu = (magma_int_t)(f*ne);
}
/****************************************************
* apply V2 from left to the eigenvectors Z. dZ = (I-V2*T2*V2')*Z
* **************************************************/
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
//n_gpu=ne;
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
//$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
timeaplQ2 = magma_wtime();
/*============================
* use GPU+CPU's
*==========================*/
if(n_gpu < ne)
{
// define the size of Q to be done on CPU's and the size on GPU's
// note that GPU use Q(1:N_GPU) and CPU use Q(N_GPU+1:N)
#ifdef ENABLE_DEBUG
printf("---> calling GPU + CPU(if N_CPU>0) to apply V2 to Z with NE %d N_GPU %d N_CPU %d\n",ne, n_gpu, ne-n_gpu);
#endif
magma_capplyQ_data data_applyQ(threads, n, ne, n_gpu, nb, Vblksiz, Z, ldz, V, ldv, TAU, T, ldt, dZ, lddz);
magma_capplyQ_id_data* arg;
magma_malloc_cpu((void**) &arg, threads*sizeof(magma_capplyQ_id_data));
pthread_t* thread_id;
magma_malloc_cpu((void**) &thread_id, threads*sizeof(pthread_t));
pthread_attr_t thread_attr;
// ===============================
// relaunch thread to apply Q
// ===============================
// Set one thread per core
pthread_attr_init(&thread_attr);
pthread_attr_setscope(&thread_attr, PTHREAD_SCOPE_SYSTEM);
pthread_setconcurrency(threads);
// Launch threads
for (magma_int_t thread = 1; thread < threads; thread++)
{
arg[thread] = magma_capplyQ_id_data(thread, &data_applyQ);
pthread_create(&thread_id[thread], &thread_attr, magma_capplyQ_parallel_section, &arg[thread]);
}
arg[0] = magma_capplyQ_id_data(0, &data_applyQ);
magma_capplyQ_parallel_section(&arg[0]);
// Wait for completion
for (magma_int_t thread = 1; thread < threads; thread++)
{
void *exitcodep;
pthread_join(thread_id[thread], &exitcodep);
}
magma_free_cpu(thread_id);
magma_free_cpu(arg);
magma_csetmatrix(n, ne-n_gpu, Z + n_gpu*ldz, ldz, dZ + n_gpu*ldz, lddz);
/*============================
* use only GPU
*==========================*/
} else {
magma_csetmatrix(n, ne, Z, ldz, dZ, lddz);
magma_cbulge_applyQ_v2('L', ne, n, nb, Vblksiz, dZ, lddz, V, ldv, T, ldt, info);
magma_device_sync();
}
timeaplQ2 = magma_wtime()-timeaplQ2;
magma_setlapack_numthreads(threads);
return MAGMA_SUCCESS;
}
extern "C" magma_int_t magma_cbulge_back(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t ne, magma_int_t Vblksiz,
cuFloatComplex *Z, magma_int_t ldz, cuFloatComplex *dZ, magma_int_t lddz,
cuFloatComplex *V, magma_int_t ldv, cuFloatComplex *TAU, cuFloatComplex *T, magma_int_t ldt, magma_int_t* info)
{
magma_int_t mklth = threads;
float timeaplQ2=0.0;
#if defined(USEMKL)
mkl_set_num_threads(1);
#endif
#if defined(USEACML)
omp_set_num_threads(1);
#endif
float f= 1.;
magma_int_t n_gpu = ne;
if(threads>40){
f = 0.5;
n_gpu = (magma_int_t)(f*ne)/64*64;
}
else if(threads>10){
#if (defined(PRECISION_s) || defined(PRECISION_d))
f = 0.68;
#else
f = 0.72;
#endif
n_gpu = (magma_int_t)(f*ne)/64*64;
}
else if(threads>5){
#if (defined(PRECISION_s) || defined(PRECISION_d))
f = 0.82;
#else
f = 0.86;
#endif
n_gpu = (magma_int_t)(f*ne)/64*64;
}
else if(threads>1){
#if (defined(PRECISION_s) || defined(PRECISION_d))
f = 0.96;
#else
f = 0.96;
#endif
n_gpu = (magma_int_t)(f*ne)/64*64;
}
/****************************************************
* apply V2 from left to the eigenvectors Z. dZ = (I-V2*T2*V2')*Z
* **************************************************/
timeaplQ2 = magma_wtime();
/*============================
* use GPU+CPU's
*==========================*/
if(n_gpu < ne)
{
// define the size of Q to be done on CPU's and the size on GPU's
// note that GPU use Q(1:N_GPU) and CPU use Q(N_GPU+1:N)
printf("---> calling GPU + CPU(if N_CPU>0) to apply V2 to Z with NE %d N_GPU %d N_CPU %d\n",ne, n_gpu, ne-n_gpu);
magma_capplyQ_data data_applyQ(threads, n, ne, n_gpu, nb, Vblksiz, Z, ldz, V, ldv, TAU, T, ldt, dZ, lddz);
magma_capplyQ_id_data* arg = new magma_capplyQ_id_data[threads];
pthread_t* thread_id = new pthread_t[threads];
pthread_attr_t thread_attr;
// ===============================
// relaunch thread to apply Q
// ===============================
// Set one thread per core
pthread_attr_init(&thread_attr);
pthread_attr_setscope(&thread_attr, PTHREAD_SCOPE_SYSTEM);
pthread_setconcurrency(threads);
// Launch threads
for (magma_int_t thread = 1; thread < threads; thread++)
{
arg[thread] = magma_capplyQ_id_data(thread, &data_applyQ);
pthread_create(&thread_id[thread], &thread_attr, magma_capplyQ_parallel_section, &arg[thread]);
}
arg[0] = magma_capplyQ_id_data(0, &data_applyQ);
magma_capplyQ_parallel_section(&arg[0]);
// Wait for completion
for (magma_int_t thread = 1; thread < threads; thread++)
{
void *exitcodep;
pthread_join(thread_id[thread], &exitcodep);
}
delete[] thread_id;
delete[] arg;
magma_csetmatrix(n, ne-n_gpu, Z + n_gpu*ldz, ldz, dZ + n_gpu*ldz, lddz);
/*============================
* use only GPU
*==========================*/
}else{
magma_csetmatrix(n, ne, Z, ldz, dZ, lddz);
magma_cbulge_applyQ_v2('L', ne, n, nb, Vblksiz, dZ, lddz, V, ldv, T, ldt, info);
magma_device_sync();
}
timeaplQ2 = magma_wtime()-timeaplQ2;
#if defined(USEMKL)
mkl_set_num_threads(mklth);
#endif
#if defined(USEACML)
omp_set_num_threads(mklth);
#endif
return MAGMA_SUCCESS;
}
//##################################################################################################
static void *magma_capplyQ_parallel_section(void *arg)
{
magma_int_t my_core_id = ((magma_capplyQ_id_data*)arg) -> id;
magma_capplyQ_data* data = ((magma_capplyQ_id_data*)arg) -> data;
magma_int_t allcores_num = data -> threads_num;
magma_int_t n = data -> n;
magma_int_t ne = data -> ne;
magma_int_t n_gpu = data -> n_gpu;
magma_int_t nb = data -> nb;
magma_int_t Vblksiz = data -> Vblksiz;
cuFloatComplex *E = data -> E;
magma_int_t lde = data -> lde;
cuFloatComplex *V = data -> V;
magma_int_t ldv = data -> ldv;
cuFloatComplex *TAU = data -> TAU;
cuFloatComplex *T = data -> T;
magma_int_t ldt = data -> ldt;
cuFloatComplex *dE = data -> dE;
magma_int_t ldde = data -> ldde;
pthread_barrier_t* barrier = &(data -> barrier);
magma_int_t info;
real_Double_t timeQcpu=0.0, timeQgpu=0.0;
magma_int_t n_cpu = ne - n_gpu;
#if defined(SETAFFINITY)
cpu_set_t set;
CPU_ZERO( &set );
CPU_SET( my_core_id, &set );
sched_setaffinity( 0, sizeof(set), &set) ;
#endif
if(my_core_id==0)
{
//=============================================
// on GPU on thread 0:
// - apply V2*Z(:,1:N_GPU)
//=============================================
timeQgpu = magma_wtime();
magma_csetmatrix(n, n_gpu, E, lde, dE, ldde);
magma_cbulge_applyQ_v2('L', n_gpu, n, nb, Vblksiz, dE, ldde, V, ldv, T, ldt, &info);
magma_device_sync();
timeQgpu = magma_wtime()-timeQgpu;
printf(" Finish Q2_GPU GGG timing= %f \n" ,timeQgpu);
}else{
//=============================================
// on CPU on threads 1:allcores_num-1:
// - apply V2*Z(:,N_GPU+1:NE)
//=============================================
if(my_core_id == 1)
timeQcpu = magma_wtime();
magma_int_t n_loc = magma_ceildiv(n_cpu, allcores_num-1);
cuFloatComplex* E_loc = E + (n_gpu+ n_loc * (my_core_id-1))*lde;
n_loc = min(n_loc,n_cpu - n_loc * (my_core_id-1));
magma_ctile_bulge_applyQ('L', n_loc, n, nb, Vblksiz, E_loc, lde, V, ldv, TAU, T, ldt);
pthread_barrier_wait(barrier);
if(my_core_id == 1){
timeQcpu = magma_wtime()-timeQcpu;
printf(" Finish Q2_CPU CCC timing= %f \n" ,timeQcpu);
}
} // END if my_core_id
#if defined(SETAFFINITY)
// unbind threads
magma_int_t sys_corenbr = 1;
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;
}
//##################################################################################################
static void *magma_capplyQ_parallel_section(void *arg)
{
magma_int_t my_core_id = ((magma_capplyQ_id_data*)arg) -> id;
magma_capplyQ_data* data = ((magma_capplyQ_id_data*)arg) -> data;
magma_int_t allcores_num = data -> threads_num;
magma_int_t n = data -> n;
magma_int_t ne = data -> ne;
magma_int_t n_gpu = data -> n_gpu;
magma_int_t nb = data -> nb;
magma_int_t Vblksiz = data -> Vblksiz;
magmaFloatComplex *E = data -> E;
magma_int_t lde = data -> lde;
magmaFloatComplex *V = data -> V;
magma_int_t ldv = data -> ldv;
magmaFloatComplex *TAU = data -> TAU;
magmaFloatComplex *T = data -> T;
magma_int_t ldt = data -> ldt;
magmaFloatComplex *dE = data -> dE;
magma_int_t ldde = data -> ldde;
pthread_barrier_t* barrier = &(data -> barrier);
magma_int_t info;
#ifdef ENABLE_TIMER
real_Double_t timeQcpu=0.0, timeQgpu=0.0;
#endif
magma_int_t n_cpu = ne - n_gpu;
// 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_set_lapack_numthreads(1);
#ifndef MAGMA_NOAFFINITY
//#define PRINTAFFINITY
#ifdef PRINTAFFINITY
affinity_set print_set;
print_set.print_affinity(my_core_id, "starting affinity");
#endif
cpu_set_t old_set, new_set;
//store current affinity
CPU_ZERO(&old_set);
sched_getaffinity( 0, sizeof(old_set), &old_set);
//set new affinity
// bind threads
CPU_ZERO(&new_set);
CPU_SET(my_core_id, &new_set);
sched_setaffinity( 0, sizeof(new_set), &new_set);
#ifdef PRINTAFFINITY
print_set.print_affinity(my_core_id, "set affinity");
#endif
#endif
if (my_core_id == 0) {
//=============================================
// on GPU on thread 0:
// - apply V2*Z(:,1:N_GPU)
//=============================================
#ifdef ENABLE_TIMER
timeQgpu = magma_wtime();
#endif
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
magma_csetmatrix( n, n_gpu, E, lde, dE, ldde, queue );
magma_cbulge_applyQ_v2(MagmaLeft, n_gpu, n, nb, Vblksiz, dE, ldde, V, ldv, T, ldt, &info);
magma_queue_destroy( queue );
#ifdef ENABLE_TIMER
timeQgpu = magma_wtime()-timeQgpu;
printf(" Finish Q2_GPU GGG timing= %f\n", timeQgpu);
#endif
} else {
//=============================================
// on CPU on threads 1:allcores_num-1:
// - apply V2*Z(:,N_GPU+1:NE)
//=============================================
#ifdef ENABLE_TIMER
if (my_core_id == 1)
timeQcpu = magma_wtime();
#endif
magma_int_t n_loc = magma_ceildiv(n_cpu, allcores_num-1);
magmaFloatComplex* E_loc = E + (n_gpu+ n_loc * (my_core_id-1))*lde;
n_loc = min(n_loc,n_cpu - n_loc * (my_core_id-1));
magma_ctile_bulge_applyQ(my_core_id, MagmaLeft, n_loc, n, nb, Vblksiz, E_loc, lde, V, ldv, TAU, T, ldt);
pthread_barrier_wait(barrier);
#ifdef ENABLE_TIMER
if (my_core_id == 1) {
timeQcpu = magma_wtime()-timeQcpu;
printf(" Finish Q2_CPU CCC timing= %f\n", timeQcpu);
}
#endif
} // END if my_core_id
#ifndef MAGMA_NOAFFINITY
//restore old affinity
sched_setaffinity(0, sizeof(old_set), &old_set);
#ifdef PRINTAFFINITY
print_set.print_affinity(my_core_id, "restored_affinity");
#endif
#endif
return 0;
}
