int P_zunmqr(
const char *side,
const char *trans,
int M,
int N,
int K,
void *A,
int LDA,
void *T,
void *B,
int LDB
) {
PLASMA_enum s, t;
int info;
if (*side == 'L') {
s = PlasmaLeft;
} else {
s = PlasmaRight;
}
if (*trans == 'C') {
t = PlasmaConjTrans;
} else {
t = PlasmaNoTrans;
}
#if CHECK_VERSION_BEQ(2,4,5)
info = PLASMA_zunmqr(s, t, M, N, K, A, LDA, T, B, LDB);
#else
PLASMA_desc *descT;
int NB, IB;
int MT, NT;
/* Get autotuned or set tile size; T matrix allocated with R */
PLASMA_Alloc_Workspace_zgeqrf(1, 1, &descT);
PLASMA_Get(PLASMA_TILE_SIZE, &NB);
PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &IB);
PLASMA_Dealloc_Handle_Tile(&descT);
MT = (M%NB==0) ? (M/NB) : (M/NB+1);
NT = (N%NB==0) ? (N/NB) : (N/NB+1);
// possibly allocate space for descT in R and keep it in qr object instead
info = PLASMA_Desc_Create(&descT, T, PlasmaComplexDouble,
IB, NB, IB*NB, MT*IB, NT*NB, 0, 0, MT*IB, NT*NB);
info = PLASMA_zunmqr(s, t, M, N, K, A, LDA, descT, B, LDB);
PLASMA_Desc_Destroy(&descT);
#endif
return(info);
}
int P_dgeqrf(
int M,
int N,
double *A,
double *T
) {
int info;
#if CHECK_VERSION_BEQ(2,4,5)
info = PLASMA_dgeqrf(M, N, A, M, T);
#else
PLASMA_desc *descT;
int NB, IB;
int MT, NT;
/* Get autotuned or set tile size; T matrix allocated with R */
PLASMA_Alloc_Workspace_dgeqrf(1, 1, &descT);
PLASMA_Get(PLASMA_TILE_SIZE, &NB);
PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &IB);
PLASMA_Dealloc_Handle_Tile(&descT);
MT = (M%NB==0) ? (M/NB) : (M/NB+1);
NT = (N%NB==0) ? (N/NB) : (N/NB+1);
// possibly allocate space for descT in R and keep it in qr object instead
info = PLASMA_Desc_Create(&descT, T, PlasmaRealDouble,
IB, NB, IB*NB, MT*IB, NT*NB, 0, 0, MT*IB, NT*NB);
/*
printf("MB=%d NB=%d BSIZ=%d LM=%d LN=%d M=%d N=%d MT=%d NT=%d\n",
descT->mb, descT->nb, descT->bsiz, descT->lm, descT->ln,
descT->m, descT->n, descT->mt, descT->nt);
*/
info = PLASMA_dgeqrf(M, N, A, M, descT);
PLASMA_Desc_Destroy(&descT);
#endif
return(info);
}
int P_zungqr(
int M,
int N,
int K,
void *A,
int LDA,
void *T,
void *Q,
int LDQ
) {
int info;
#if CHECK_VERSION_BEQ(2,4,5)
info = PLASMA_zungqr(M, N, K, A, LDA, T, Q, LDQ);
#else
PLASMA_desc *descT;
int NB, IB;
int MT, NT;
/* Get autotuned or set tile size; T matrix allocated with R */
PLASMA_Alloc_Workspace_dgeqrf(1, 1, &descT);
PLASMA_Get(PLASMA_TILE_SIZE, &NB);
PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &IB);
PLASMA_Dealloc_Handle_Tile(&descT);
MT = (M%NB==0) ? (M/NB) : (M/NB+1);
NT = (N%NB==0) ? (N/NB) : (N/NB+1);
// possibly allocate space for descT in R and keep it in qr object instead
info = PLASMA_Desc_Create(&descT, T, PlasmaComplexDouble,
IB, NB, IB*NB, MT*IB, NT*NB, 0, 0, MT*IB, NT*NB);
info = PLASMA_zungqr(M, N, K, A, LDA, descT, Q, LDQ);
PLASMA_Desc_Destroy(&descT);
#endif
return(info);
}
static int
RunTest(int *iparam, double *dparam, real_Double_t *t_)
{
double *AT;
real_Double_t t;
PLASMA_desc *descA;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int check = iparam[TIMING_CHECK];
PLASMA_enum uplo = PlasmaLower;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* } */
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
/* Allocate Data */
AT = (double *)malloc(nt*nt*nb2*sizeof(double));
/* Check if unable to allocate memory */
if ( !AT ){
printf("Out of Memory \n ");
exit(0);
}
/*
* Initialize Data
* It's done in static to avoid having the same sequence than one
* the function we want to trace
*/
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
PLASMA_Desc_Create(&descA, AT, PlasmaRealDouble, nb, nb, nb*nb, n, n, 0, 0, n, n);
PLASMA_dplgsy_Tile( (double)n, descA, 51 );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/* Save AT in lapack layout for check */
if ( check ) {
}
/* PLASMA DPOTRF / DTRTRI / DLAUUM */
/*
* Example of the different way to combine several asynchonous calls
*/
{
#if defined(TRACE_BY_SEQUENCE)
PLASMA_sequence *sequence[3];
PLASMA_request request[3] = { PLASMA_REQUEST_INITIALIZER,
PLASMA_REQUEST_INITIALIZER,
PLASMA_REQUEST_INITIALIZER };
PLASMA_Sequence_Create(&sequence[0]);
PLASMA_Sequence_Create(&sequence[1]);
PLASMA_Sequence_Create(&sequence[2]);
t = -cWtime();
#if defined(POTRI_SYNC)
PLASMA_dpotrf_Tile_Async(uplo, descA, sequence[0], &request[0]);
PLASMA_Sequence_Wait(sequence[0]);
PLASMA_dtrtri_Tile_Async(uplo, PlasmaNonUnit, descA, sequence[1], &request[1]);
PLASMA_Sequence_Wait(sequence[1]);
PLASMA_dlauum_Tile_Async(uplo, descA, sequence[2], &request[2]);
PLASMA_Sequence_Wait(sequence[2]);
#else
PLASMA_dpotrf_Tile_Async(uplo, descA, sequence[0], &request[0]);
PLASMA_dtrtri_Tile_Async(uplo, PlasmaNonUnit, descA, sequence[1], &request[1]);
PLASMA_dlauum_Tile_Async(uplo, descA, sequence[2], &request[2]);
PLASMA_Sequence_Wait(sequence[0]);
PLASMA_Sequence_Wait(sequence[1]);
PLASMA_Sequence_Wait(sequence[2]);
#endif
t += cWtime();
PLASMA_Sequence_Destroy(sequence[0]);
PLASMA_Sequence_Destroy(sequence[1]);
PLASMA_Sequence_Destroy(sequence[2]);
#else
#if defined(POTRI_SYNC)
t = -cWtime();
PLASMA_dpotrf_Tile(uplo, descA);
PLASMA_dtrtri_Tile(uplo, PlasmaNonUnit, descA);
PLASMA_dlauum_Tile(uplo, descA);
t += cWtime();
#else
/* Default: we use Asynchonous call with only one sequence */
PLASMA_sequence *sequence;
PLASMA_request request[2] = { PLASMA_REQUEST_INITIALIZER,
PLASMA_REQUEST_INITIALIZER };
t = -cWtime();
PLASMA_Sequence_Create(&sequence);
PLASMA_dpotrf_Tile_Async(uplo, descA, sequence, &request[0]);
PLASMA_dpotri_Tile_Async(uplo, descA, sequence, &request[1]);
PLASMA_Sequence_Wait(sequence);
t += cWtime();
PLASMA_Sequence_Destroy(sequence);
#endif
#endif
*t_ = t;
}
/* Check the solution */
if ( check )
{
dparam[TIMING_ANORM] = 0.0;
dparam[TIMING_XNORM] = 0.0;
dparam[TIMING_BNORM] = 0.0;
dparam[TIMING_RES] = 0.0;
}
PLASMA_Desc_Destroy(&descA);
PLASMA_Finalize();
free(AT);
return 0;
}
static int
RunTest(int *iparam, double *dparam, real_Double_t *t_)
{
PLASMA_Complex64_t *A = NULL, *AT, *b, *bT, *x;
PLASMA_desc *descA, *descB, *descL;
real_Double_t t;
int *piv;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int nrhs = iparam[TIMING_NRHS];
int check = iparam[TIMING_CHECK];
int lda = n;
int ldb = n;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &iparam[TIMING_IB] ); */
/* } */
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
/* Allocate Data */
AT = (PLASMA_Complex64_t *)malloc(nt*nt*nb2*sizeof(PLASMA_Complex64_t));
/* Check if unable to allocate memory */
if ( !AT ){
printf("Out of Memory \n ");
exit(0);
}
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaComplexDouble, nb, nb, nb*nb, n, n, 0, 0, n, n);
LAPACKE_zlarnv_work(1, ISEED, nt*nt*nb2, AT);
/* Allocate Workspace */
PLASMA_Alloc_Workspace_zgesv_incpiv_Tile(n, &descL, &piv);
/* Save AT in lapack layout for check */
if ( check ) {
A = (PLASMA_Complex64_t *)malloc(lda*n *sizeof(PLASMA_Complex64_t));
PLASMA_Tile_to_Lapack(descA, (void*)A, n);
}
t = -cWtime();
PLASMA_zgetrf_incpiv_Tile( descA, descL, piv );
t += cWtime();
*t_ = t;
/* Check the solution */
if ( check )
{
b = (PLASMA_Complex64_t *)malloc(ldb*nrhs *sizeof(PLASMA_Complex64_t));
bT = (PLASMA_Complex64_t *)malloc(nt*nb2 *sizeof(PLASMA_Complex64_t));
x = (PLASMA_Complex64_t *)malloc(ldb*nrhs *sizeof(PLASMA_Complex64_t));
LAPACKE_zlarnv_work(1, ISEED, n*nrhs, b);
PLASMA_Desc_Create(&descB, bT, PlasmaComplexDouble, nb, nb, nb*nb, n, nrhs, 0, 0, n, nrhs);
PLASMA_Lapack_to_Tile((void*)b, n, descB);
PLASMA_zgetrs_incpiv_Tile( descA, descL, piv, descB );
PLASMA_Tile_to_Lapack(descB, (void*)x, n);
dparam[TIMING_RES] = z_check_solution(n, n, nrhs, A, lda, b, x, ldb,
&(dparam[TIMING_ANORM]), &(dparam[TIMING_BNORM]),
&(dparam[TIMING_XNORM]));
PLASMA_Desc_Destroy(&descB);
free( A ); free( b ); free( bT ); free( x );
}
/* Deallocate Workspace */
PLASMA_Dealloc_Handle_Tile(&descL);
PLASMA_Desc_Destroy(&descA);
free( AT );
free( piv );
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, float *dparam, real_Double_t *t_)
{
float *AT, *BT, *CT;
float *A = NULL, *B = NULL, *C1 = NULL, *C2 = NULL;
float alpha, beta;
PLASMA_desc *descA, *descB, *descC;
real_Double_t t;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int check = iparam[TIMING_CHECK];
int lda = n;
/* Allocate Data */
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
/* } */
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* } */
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
AT = (float *)malloc(nt*nt*nb2*sizeof(float));
BT = (float *)malloc(nt*nt*nb2*sizeof(float));
CT = (float *)malloc(nt*nt*nb2*sizeof(float));
/* Check if unable to allocate memory */
if ( (!AT) || (!BT) || (!CT) ) {
printf("Out of Memory \n ");
exit(0);
}
#if defined(PLASMA_CUDA)
cudaHostRegister(AT, nt*nt*nb2*sizeof(float), cudaHostRegisterPortable);
cudaHostRegister(BT, nt*nt*nb2*sizeof(float), cudaHostRegisterPortable);
cudaHostRegister(CT, nt*nt*nb2*sizeof(float), cudaHostRegisterPortable);
#endif
/* Initialiaze Data */
LAPACKE_slarnv_work(1, ISEED, 1, &alpha);
LAPACKE_slarnv_work(1, ISEED, 1, &beta);
LAPACKE_slarnv_work(1, ISEED, nt*nt*nb2, AT);
LAPACKE_slarnv_work(1, ISEED, nt*nt*nb2, BT);
LAPACKE_slarnv_work(1, ISEED, nt*nt*nb2, CT);
/* Initialize AT and bT for Symmetric Positif Matrix */
PLASMA_Desc_Create(&descA, AT, PlasmaRealFloat, nb, nb, nb*nb, n, n, 0, 0, n, n);
PLASMA_Desc_Create(&descB, BT, PlasmaRealFloat, nb, nb, nb*nb, n, n, 0, 0, n, n);
PLASMA_Desc_Create(&descC, CT, PlasmaRealFloat, nb, nb, nb*nb, n, n, 0, 0, n, n);
if (check)
{
C2 = (float *)malloc(n*lda*sizeof(float));
PLASMA_Tile_to_Lapack(descC, (void*)C2, n);
}
#if defined(PLASMA_CUDA)
core_cublas_init();
#endif
t = -cWtime();
PLASMA_sgemm_Tile( PlasmaNoTrans, PlasmaNoTrans, alpha, descA, descB, beta, descC );
t += cWtime();
*t_ = t;
/* Check the solution */
if (check)
{
A = (float *)malloc(n*lda*sizeof(float));
PLASMA_Tile_to_Lapack(descA, (void*)A, n);
free(AT);
B = (float *)malloc(n*lda*sizeof(float));
PLASMA_Tile_to_Lapack(descB, (void*)B, n);
free(BT);
C1 = (float *)malloc(n*lda*sizeof(float));
PLASMA_Tile_to_Lapack(descC, (void*)C1, n);
free(CT);
dparam[TIMING_RES] = s_check_gemm( PlasmaNoTrans, PlasmaNoTrans, n, n, n,
alpha, A, lda, B, lda, beta, C1, C2, lda,
&(dparam[TIMING_ANORM]), &(dparam[TIMING_BNORM]),
&(dparam[TIMING_XNORM]));
free(C2);
}
else {
free( AT );
free( BT );
free( CT );
}
PLASMA_Desc_Destroy(&descA);
PLASMA_Desc_Destroy(&descB);
PLASMA_Desc_Destroy(&descC);
PLASMA_Finalize();
return 0;
}
static double
RunTest(real_Double_t *t_, struct user_parameters* params)
{
double t;
PLASMA_desc *descT;
int64_t N = params->matrix_size;
int64_t IB = params->iblocksize;
int64_t NB = params->blocksize;
int check = params->check;
double check_res = 0;
/* Allocate Data */
PLASMA_desc *descA = NULL;
double *ptr = (double*)malloc(N * N * sizeof(double));
PLASMA_Desc_Create(&descA, ptr, PlasmaRealDouble, NB, NB, NB*NB, N, N, 0, 0, N, N);
#pragma omp parallel
{
#pragma omp single
{
plasma_pdpltmg_quark(*descA, 5373 );
}
}
/* Save A for check */
double *A = NULL;
if ( check ) {
A = (double*)malloc(N * N * sizeof(double));
plasma_pdtile_to_lapack_quark(*descA, (void*)A, N);
}
/* Allocate Workspace */
plasma_alloc_ibnb_tile(N, N, PlasmaRealDouble, &descT, IB, NB);
/* Do the computations */
START_TIMING();
#pragma omp parallel
{
#pragma omp single
{
plasma_pdgeqrf_quark( *descA, *descT , IB);
}
}
STOP_TIMING();
/* Check the solution */
if ( check )
{
/* Allocate B for check */
PLASMA_desc *descB = NULL;
double* ptr = (double*)malloc(N * sizeof(double));
PLASMA_Desc_Create(&descB, ptr, PlasmaRealDouble, NB, NB, NB*NB, N, 1, 0, 0, N, 1);
/* Initialize and save B */
plasma_pdpltmg_seq(*descB, 2264 );
double *B = (double*)malloc(N * sizeof(double));
plasma_pdtile_to_lapack_quark(*descB, (void*)B, N);
/* Compute the solution */
PLASMA_dgeqrs_Tile( descA, descT, descB , IB);
/* Copy solution to X */
double *X = (double*)malloc(N * sizeof(double));
plasma_pdtile_to_lapack_quark(*descB, (void*)X, N);
check_res = d_check_solution(N, N, 1, A, N, B, X, N);
/* Free checking structures */
PASTE_CODE_FREE_MATRIX( descB );
free( A );
free( B );
free( X );
}
/* Free data */
PLASMA_Dealloc_Handle_Tile(&descT);
PASTE_CODE_FREE_MATRIX( descA );
return check_res;
}
static int
RunTest(int *iparam, double *dparam, real_Double_t *t_)
{
double *A = NULL, *AT, *b = NULL, *bT, *x;
PLASMA_desc *descA, *descB, *descT;
real_Double_t t;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int nrhs = iparam[TIMING_NRHS];
int check = iparam[TIMING_CHECK];
int lda = n;
int ldb = n;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
#if defined(PLASMA_CUDA)
core_cublas_init();
#endif
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &iparam[TIMING_IB] ); */
/* } */
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
/* Allocate Data */
AT = (double *)malloc(nt*nt*nb2*sizeof(double));
/* Check if unable to allocate memory */
if ( !AT ){
printf("Out of Memory \n ");
exit(0);
}
#if defined(PLASMA_CUDA)
cudaHostRegister((void*)AT, nt*nt*nb2*sizeof(double), cudaHostRegisterPortable);
#endif
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaRealDouble, nb, nb, nb*nb, n, n, 0, 0, n, n);
LAPACKE_dlarnv_work(1, ISEED, nt*nt*nb2, AT);
/* Allocate Workspace */
PLASMA_Alloc_Workspace_dgels_Tile(n, n, &descT);
#if defined(PLASMA_CUDA)
cudaHostRegister((void*)descT->mat, descT->lm*descT->ln*sizeof(double), cudaHostRegisterPortable);
#endif
/* Save AT in lapack layout for check */
if ( check ) {
A = (double *)malloc(lda*n *sizeof(double));
PLASMA_Tile_to_Lapack(descA, (void*)A, n);
}
t = -cWtime();
PLASMA_dgeqrf_Tile( descA, descT );
t += cWtime();
*t_ = t;
/* Check the solution */
if ( check )
{
b = (double *)malloc(ldb*nrhs *sizeof(double));
bT = (double *)malloc(nt*nb2 *sizeof(double));
x = (double *)malloc(ldb*nrhs *sizeof(double));
LAPACKE_dlarnv_work(1, ISEED, nt*nb2, bT);
PLASMA_Desc_Create(&descB, bT, PlasmaRealDouble, nb, nb, nb*nb, n, nrhs, 0, 0, n, nrhs);
PLASMA_Tile_to_Lapack(descB, (void*)b, n);
PLASMA_dgeqrs_Tile( descA, descT, descB );
PLASMA_Tile_to_Lapack(descB, (void*)x, n);
dparam[TIMING_RES] = d_check_solution(n, n, nrhs, A, lda, b, x, ldb,
&(dparam[TIMING_ANORM]), &(dparam[TIMING_BNORM]),
&(dparam[TIMING_XNORM]));
PLASMA_Desc_Destroy(&descB);
free( A );
free( b );
free( bT );
free( x );
}
/* Allocate Workspace */
PLASMA_Dealloc_Handle_Tile(&descT);
PLASMA_Desc_Destroy(&descA);
free( AT );
PLASMA_Finalize();
#if defined(PLASMA_CUDA)
#endif
return 0;
}
static double
RunTest(real_Double_t *t_, struct user_parameters* params)
{
double t;
int64_t N = params->matrix_size;
int64_t NB = params->blocksize;
int check = params->check;
int uplo = PlasmaUpper;
double check_res = 0;
/* Allocate Data */
PLASMA_desc *descA = NULL;
double* ptr = malloc(N * N * sizeof(double));
PLASMA_Desc_Create(&descA, ptr, PlasmaRealDouble, NB, NB, NB*NB, N, N, 0, 0, N, N);
#pragma omp parallel
{
#pragma omp single
{
plasma_pdplgsy_quark( (double)N, *descA, 51 );
}
}
/* Save A for check */
double *A = NULL;
if(check) {
A = (double*)malloc(N * N * sizeof(double));
plasma_pdtile_to_lapack_quark(*descA, (void*)A, N);
}
/* PLASMA DPOSV */
START_TIMING();
#pragma omp parallel
{
#pragma omp single
{
plasma_pdpotrf_quark(uplo, *descA);
}
}
STOP_TIMING();
/* Check the solution */
if ( check )
{
PLASMA_desc *descB = NULL;
double* ptr = (double*)malloc(N * sizeof(double));
PLASMA_Desc_Create(&descB, ptr, PlasmaRealDouble, NB, NB, NB*NB, N, 1, 0, 0, N, 1);
plasma_pdpltmg_seq(* descB, 7672 );
double* B = (double*)malloc(N * sizeof(double));
plasma_pdtile_to_lapack_quark(*descB, (void*)B, N);
PLASMA_dpotrs_Tile( uplo, descA, descB );
double* X = (double*)malloc(N * sizeof(double));
plasma_pdtile_to_lapack_quark(*descB, (void*)X, N);
check_res = d_check_solution(N, N, 1, A, N, B, X, N);
PASTE_CODE_FREE_MATRIX( descB );
free( A );
free( B );
free( X );
}
PASTE_CODE_FREE_MATRIX( descA );
return check_res;
}
static int
RunTest(int *iparam, float *dparam, real_Double_t *t_)
{
float *AT, *Q = NULL;
float *W;
PLASMA_desc *descA = NULL;
PLASMA_desc *descQ = NULL;
PLASMA_desc *descT = NULL;
real_Double_t t;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int check = iparam[TIMING_CHECK];
int lda = n;
int uplo = PlasmaUpper;
int vec = PlasmaNoVec;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
/* Allocate Data */
AT = (float *)malloc(lda*n*sizeof(float));
W = (float *)malloc(n*sizeof(float));
if (vec == PlasmaVec){
Q = (float *)malloc(lda*n*sizeof(float));
if ( (!Q) ) {
printf("Out of Memory -Q-\n ");
return -2;
}
}
/* Check if unable to allocate memory */
if ( (!AT) || (!W) ) {
printf("Out of Memory -\n ");
return -2;
}
/* Initialize Data */
PLASMA_Desc_Create(&descA, AT, PlasmaRealFloat, nb, nb, nb*nb, lda, n, 0, 0, n, n);
PLASMA_splgsy_Tile((float)0.0, descA, 51 );
if (vec == PlasmaVec)
PLASMA_Desc_Create(&descQ, Q, PlasmaRealFloat, nb, nb, nb*nb, lda, n, 0, 0, n, n);
/* Save AT and bT in lapack layout for check */
if ( check ) {
}
/* Allocate Workspace */
PLASMA_Alloc_Workspace_ssyev(n, n, &descT);
t = -cWtime();
PLASMA_ssyev_Tile( vec, uplo, descA, W, descT, descQ );
t += cWtime();
*t_ = t;
/* Check the solution */
if ( check )
{
}
/* DeAllocate Workspace */
PLASMA_Dealloc_Handle_Tile(&descT);
PLASMA_Desc_Destroy(&descA);
if (vec == PlasmaVec) {
PLASMA_Desc_Destroy(&descQ);
free( Q );
}
free( AT );
free( W );
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, float *dparam, real_Double_t *t_)
{
float *AT, *bT, *x;
float *A = NULL;
float *b = NULL;
PLASMA_desc *descA, *descB;
real_Double_t t;
int *piv;
int n = iparam[TIMING_N];
int nb = iparam[TIMING_NB];
int nrhs = iparam[TIMING_NRHS];
int check = iparam[TIMING_CHECK];
int lda = n;
int ldb = n;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &iparam[TIMING_IB] ); */
/* } */
/* Allocate Data */
AT = (float *)malloc(lda*n *sizeof(float));
bT = (float *)malloc(ldb*nrhs*sizeof(float));
piv = (int *)malloc( n*sizeof(int));
/* Check if unable to allocate memory */
if ( (!AT) || (!bT) || (!piv) ) {
printf("Out of Memory \n ");
return -1;
}
/* Initialize AT and bT for Symmetric Positif Matrix */
PLASMA_Desc_Create(&descA, AT, PlasmaRealFloat, nb, nb, nb*nb, lda, n, 0, 0, n, n);
PLASMA_Desc_Create(&descB, bT, PlasmaRealFloat, nb, nb, nb*nb, ldb, nrhs, 0, 0, n, nrhs);
LAPACKE_slarnv_work(1, ISEED, lda*n, AT);
LAPACKE_slarnv_work(1, ISEED, ldb*nrhs, bT);
/* Save AT and bT in lapack layout for check */
if ( check ) {
A = (float *)malloc(lda*n *sizeof(float));
b = (float *)malloc(ldb*nrhs*sizeof(float));
PLASMA_sTile_to_Lapack(descA, (void*)A, lda);
PLASMA_sTile_to_Lapack(descB, (void*)b, ldb);
}
t = -cWtime();
PLASMA_sgesv_Tile( descA, piv, descB );
t += cWtime();
*t_ = t;
/* Check the solution */
if ( check )
{
x = (float *)malloc(ldb*nrhs *sizeof(float));
PLASMA_sTile_to_Lapack(descB, (void*)x, n);
dparam[TIMING_RES] = s_check_solution(n, n, nrhs, A, lda, b, x, ldb,
&(dparam[TIMING_ANORM]), &(dparam[TIMING_BNORM]),
&(dparam[TIMING_XNORM]));
free(A); free(b); free(x);
}
PLASMA_Desc_Destroy(&descA);
PLASMA_Desc_Destroy(&descB);
free( AT ); free( bT );
free( piv );
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, float *dparam, real_Double_t *t_)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
PLASMA_Complex32_t *A, *AT, *A2 = NULL;
PLASMA_desc *descA;
real_Double_t t;
int *ipiv, *ipiv2 = NULL;
int i;
int nb = iparam[TIMING_NB];
int m = iparam[TIMING_N];
int n = iparam[TIMING_NRHS];
int check = iparam[TIMING_CHECK];
int lda = m;
PLASMA_sequence *sequence = NULL;
PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* Allocate Data */
A = (PLASMA_Complex32_t *)malloc(lda*n*sizeof(PLASMA_Complex32_t));
AT = (PLASMA_Complex32_t *)malloc(lda*n*sizeof(PLASMA_Complex32_t));
/* Check if unable to allocate memory */
if ( ( !AT ) || (! A) ) {
printf("Out of Memory \n ");
return -1;
}
/* Initialiaze Data */
LAPACKE_clarnv_work(1, ISEED, lda*n, A);
/* for(i=0; i<n; i++) { */
/* A[i*lda+i] += (float)m; */
/* } */
PLASMA_Desc_Create(&descA, AT, PlasmaComplexFloat, nb, nb, nb*nb, lda, n, 0, 0, m, n);
PLASMA_cLapack_to_Tile((void*)A, lda, descA);
/* Allocate Workspace */
ipiv = (int *)malloc( n*sizeof(int) );
/* Save AT in lapack layout for check */
if ( check ) {
A2 = (PLASMA_Complex32_t *)malloc(lda*n*sizeof(PLASMA_Complex32_t));
ipiv2 = (int *)malloc( n*sizeof(int) );
LAPACKE_clacpy_work(LAPACK_COL_MAJOR,' ', m, n, A, lda, A2, lda);
LAPACKE_cgetrf_work(LAPACK_COL_MAJOR, m, n, A2, lda, ipiv2 );
}
plasma = plasma_context_self();
PLASMA_Sequence_Create(&sequence);
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
QUARK_Task_Flag_Set(&task_flags, TASK_THREAD_COUNT, iparam[TIMING_THRDNBR] );
plasma_dynamic_spawn();
CORE_cgetrf_rectil_init();
t = -cWtime();
QUARK_CORE_cgetrf_rectil(plasma->quark, &task_flags,
*descA, AT, descA->mb*descA->nb, ipiv,
sequence, &request,
0, 0,
iparam[TIMING_THRDNBR]);
PLASMA_Sequence_Wait(sequence);
t += cWtime();
*t_ = t;
PLASMA_Sequence_Destroy(sequence);
/* Check the solution */
if ( check )
{
float *work = (float *)malloc(max(m,n)*sizeof(float));
PLASMA_cTile_to_Lapack(descA, (void*)A, lda);
/* Check ipiv */
for(i=0; i<n; i++)
{
if( ipiv[i] != ipiv2[i] ) {
fprintf(stderr, "\nPLASMA (ipiv[%d] = %d, A[%d] = %e) / LAPACK (ipiv[%d] = %d, A[%d] = [%e])\n",
i, ipiv[i], i, crealf(A[ i * lda + i ]),
i, ipiv2[i], i, crealf(A2[ i * lda + i ]));
break;
}
}
dparam[TIMING_ANORM] = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaMaxNorm),
m, n, A, lda, work);
dparam[TIMING_XNORM] = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaMaxNorm),
m, n, A2, lda, work);
dparam[TIMING_BNORM] = 0.0;
CORE_caxpy( m, n, -1.0, A, lda, A2, lda);
dparam[TIMING_RES] = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaMaxNorm),
m, n, A2, lda, work);
free( A2 );
free( ipiv2 );
free( work );
}
/* Deallocate Workspace */
PLASMA_Desc_Destroy(&descA);
free( A );
free( AT );
free( ipiv );
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, _PREC *dparam, real_Double_t *t_)
{
PLASMA_Complex32_t *A = NULL, *AT;
PLASMA_desc *descA;
real_Double_t t;
int n = iparam[TIMING_N];
int nb = iparam[TIMING_NB];
int check = iparam[TIMING_CHECK];
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
/* } */
n = ((n % nb) == 0) ? (n / nb) * nb : ((n / nb) + 1) * nb ;
dparam[TIMING_ANORM] = (_PREC)n;
/* Allocate Data */
AT = (PLASMA_Complex32_t *)malloc(n*n*sizeof(PLASMA_Complex32_t));
/* Check if unable to allocate memory */
if ( (!AT) ) {
printf("Out of Memory \n ");
exit(0);
}
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaComplexFloat, nb, nb, nb*nb, n, n, 0, 0, n, n);
LAPACKE_clarnv_work(1, ISEED, n*n, AT);
/* Save A and b */
if (check) {
A = (PLASMA_Complex32_t *)malloc(n*n*sizeof(PLASMA_Complex32_t));
LAPACKE_clacpy_work(LAPACK_COL_MAJOR, lapack_const(PlasmaUpperLower), n, n, AT, n, A, n);
}
t = -cWtime();
PLASMA_Lapack_to_Tile( (void *)A, n, descA);
t += cWtime();
*t_ = t;
/* Check the solution */
if (check)
{
dparam[TIMING_RES] = (_PREC)c_check_conversion(n, n, n, 1, nb, nb, A, AT, map_CM, map_CCRB);
free(A);
}
PLASMA_Desc_Destroy(&descA);
free( AT );
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, double *dparam, real_Double_t *t_)
{
PLASMA_Complex64_t *AT;
PLASMA_desc *descA, *descT;
real_Double_t t;
int nb;
int M = iparam[TIMING_N];
int N = iparam[TIMING_M];
//int N = M/nrhs; //RUN WITH NRHS = 10 or 20 (ALSO USED IN TIMING.C)
int lda = M;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &iparam[TIMING_IB] ); */
/* } */
nb = iparam[TIMING_NB];
/* Householder mode */
//PLASMA_Set(PLASMA_HOUSEHOLDER_MODE, PLASMA_FLAT_HOUSEHOLDER);
PLASMA_Set(PLASMA_HOUSEHOLDER_MODE, PLASMA_TREE_HOUSEHOLDER);
PLASMA_Set(PLASMA_HOUSEHOLDER_SIZE, 4);
/* Allocate Data */
AT = (PLASMA_Complex64_t *)malloc(lda*N*sizeof(PLASMA_Complex64_t));
/* Check if unable to allocate memory */
if ( !AT ){
printf("Out of Memory \n ");
exit(0);
}
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaComplexDouble, nb, nb, nb*nb, M, N, 0, 0, M, N);
LAPACKE_zlarnv_work(1, ISEED, lda*N, AT);
/* Allocate Workspace */
PLASMA_Alloc_Workspace_zgels_Tile(M, N, &descT);
t = -cWtime();
PLASMA_zgeqrf_Tile( descA, descT );
t += cWtime();
*t_ = t;
/* Allocate Workspace */
PLASMA_Dealloc_Handle_Tile(&descT);
PLASMA_Desc_Destroy(&descA);
free( AT );
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, float *dparam, real_Double_t *t_)
{
float *A = NULL, *AT, *b = NULL, *bT, *x;
real_Double_t t;
PLASMA_desc *descA, *descB;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int nrhs = iparam[TIMING_NRHS];
int check = iparam[TIMING_CHECK];
int lda = n;
int ldb = n;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
else
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING );
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* } */
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
/* Allocate Data */
AT = (float *)malloc(nt*nt*nb2*sizeof(float));
/* Check if unable to allocate memory */
if ( !AT ){
printf("Out of Memory \n ");
exit(0);
}
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaRealFloat, nb, nb, nb*nb, n, n, 0, 0, n, n);
PLASMA_splgsy_Tile((float)n, descA, 51 );
/* Save AT in lapack layout for check */
if ( check ) {
A = (float *)malloc(lda*n *sizeof(float));
PLASMA_Tile_to_Lapack(descA, (void*)A, n);
}
/* PLASMA SPOSV */
t = -cWtime();
PLASMA_spotrf_Tile(PlasmaUpper, descA);
t += cWtime();
*t_ = t;
/* Check the solution */
if ( check )
{
b = (float *)malloc(ldb*nrhs *sizeof(float));
bT = (float *)malloc(nt*nb2 *sizeof(float));
x = (float *)malloc(ldb*nrhs *sizeof(float));
LAPACKE_slarnv_work(1, ISEED, nt*nb2, bT);
PLASMA_Desc_Create(&descB, bT, PlasmaRealFloat, nb, nb, nb*nb, n, nrhs, 0, 0, n, nrhs);
PLASMA_Tile_to_Lapack(descB, (void*)b, n);
PLASMA_spotrs_Tile( PlasmaUpper, descA, descB );
PLASMA_Tile_to_Lapack(descB, (void*)x, n);
dparam[TIMING_RES] = s_check_solution(n, n, nrhs, A, lda, b, x, ldb,
&(dparam[TIMING_ANORM]), &(dparam[TIMING_BNORM]),
&(dparam[TIMING_XNORM]));
PLASMA_Desc_Destroy(&descB);
free( A );
free( b );
free( bT );
free( x );
}
PLASMA_Desc_Destroy(&descA);
free(AT);
PLASMA_Finalize();
return 0;
}
static int
RunTest(int *iparam, float *dparam, real_Double_t *t_)
{
PLASMA_Complex32_t *AT, *BT, *Q = NULL;
float *W;
PLASMA_desc *descA, *descB, *descQ, *descT;
real_Double_t t;
int nb, nb2, nt;
int n = iparam[TIMING_N];
int check = iparam[TIMING_CHECK];
int lda = n;
int itype = 1;
int vec = PlasmaNoVec;
int uplo = PlasmaUpper;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
// if ( iparam[TIMING_SCHEDULER] )
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
/* else */
/* PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_STATIC_SCHEDULING ); */
/*if ( !iparam[TIMING_AUTOTUNING] ) {*/
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* } else { */
/* PLASMA_Get(PLASMA_TILE_SIZE, &iparam[TIMING_NB] ); */
/* PLASMA_Get(PLASMA_INNER_BLOCK_SIZE, &iparam[TIMING_IB] ); */
/* } */
nb = iparam[TIMING_NB];
nb2 = nb * nb;
nt = n / nb + ((n % nb == 0) ? 0 : 1);
/* Allocate Data */
AT = (PLASMA_Complex32_t *)malloc(nt*nt*nb2*sizeof(PLASMA_Complex32_t));
BT = (PLASMA_Complex32_t *)malloc(nt*nt*nb2*sizeof(PLASMA_Complex32_t));
W = (float *)malloc(n*sizeof(float));
if (vec == PlasmaVec){
Q = (PLASMA_Complex32_t *)malloc(nt*nt*nb2*sizeof(PLASMA_Complex32_t));
if ( (!Q) ) {
printf("Out of Memory -Q-\n ");
exit(0);
}
}
/* Check if unable to allocate memory */
if ( (!AT) || (!BT) || (!W) ) {
printf("Out of Memory -\n ");
exit(0);
}
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaComplexFloat, nb, nb, nb*nb, lda, n, 0, 0, n, n);
PLASMA_cplghe_Tile((float)0.0, descA, 51 );
PLASMA_Desc_Create(&descB, BT, PlasmaComplexFloat, nb, nb, nb*nb, lda, n, 0, 0, n, n);
PLASMA_cplghe_Tile((float)n, descB, 51 );
PLASMA_Desc_Create(&descQ, Q, PlasmaComplexFloat, nb, nb, nb*nb, lda, n, 0, 0, n, n);
/* Save AT and bT in lapack layout for check */
if ( check ) {
}
/* Allocate Workspace */
PLASMA_Alloc_Workspace_chegv(n, n, &descT);
t = -cWtime();
PLASMA_chegv_Tile( itype, vec, uplo, descA, descB, W, descT, descQ );
t += cWtime();
*t_ = t;
/* Check the solution */
if ( check )
{
}
/* DeAllocate Workspace */
PLASMA_Dealloc_Handle_Tile(&descT);
PLASMA_Desc_Destroy(&descA);
PLASMA_Desc_Destroy(&descB);
PLASMA_Desc_Destroy(&descQ);
if (vec == PlasmaVec)
free( Q );
free( AT );
free( W );
PLASMA_Finalize();
return 0;
}
