void PLASMA_DPOTRF_TILE(PLASMA_enum *uplo, intptr_t *A, int *INFO)
{ *INFO = PLASMA_dpotrf_Tile(*uplo, (PLASMA_desc *)(*A)); }
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_)
{
double *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] );
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_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] );
/* } 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 */
#if defined(PLASMA_CUDA)
cudaHostAlloc((void**)&AT, nt*nt*nb2*sizeof(double), cudaHostAllocPortable);
#else
AT = (double *)malloc(nt*nt*nb2*sizeof(double));
#endif
/* Check if unable to allocate memory */
if ( !AT ){
printf("Out of Memory \n ");
exit(0);
}
// cudaHostRegister(AT, nt*nt*nb2*sizeof(double), cudaHostRegisterPortable);
/* Initialiaze Data */
PLASMA_Desc_Create(&descA, AT, PlasmaRealDouble, nb, nb, nb*nb, n, n, 0, 0, n, n);
PLASMA_dplgsy_Tile((double)n, descA, 51 );
/* Save AT in lapack layout for check */
if ( check ) {
A = (double *)malloc(lda*n *sizeof(double));
PLASMA_Tile_to_Lapack(descA, (void*)A, n);
}
/* PLASMA DPOSV */
t = -cWtime();
PLASMA_dpotrf_Tile(PlasmaUpper, descA);
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_dpotrs_Tile( PlasmaUpper, descA, 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 );
}
PLASMA_Desc_Destroy(&descA);
PLASMA_Finalize();
#if defined(PLASMA_CUDA)
cudaFreeHost(AT);
#else
free(AT);
#endif
// cudaHostUnregister(AT);
return 0;
}
