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;
}
