void magmaf_dpotri_gpu(
magma_uplo_t *uplo, magma_int_t *n,
devptr_t *dA, magma_int_t *ldda,
magma_int_t *info )
{
magma_dpotri_gpu(
*uplo, *n,
magma_ddevptr(dA), *ldda,
info );
}
double *cholesky_gpu(double *ml, int m) {
magma_int_t mm = m*m;
magma_int_t info;
double *a;
double *d_a ;
magma_err_t err;
err = magma_dmalloc_cpu ( &a , mm );
err = magma_dmalloc ( &d_a , mm );
magma_dsetmatrix ( m, m, ml, m, d_a , m );
magma_dpotrf_gpu('L',m,d_a,m,&info);
magma_dpotri_gpu('L',m,d_a,m,&info);
magma_dgetmatrix ( m, m, d_a , m, a, m );
magma_free (d_a );
return a;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dpotrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double *h_A, *h_R;
double *d_A;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t N, n2, lda, ldda, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], error;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n");
printf("=================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
gflops = FLOPS_DPOTRI( N ) / 1e9;
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_PIN( h_R, double, n2 );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
magma_dmake_hpd( N, h_A, lda );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* factorize matrix */
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda );
magma_dpotrf_gpu( opts.uplo, N, d_A, ldda, &info );
// check for exact singularity
//magma_dgetmatrix( N, N, d_A, ldda, h_R, lda );
//h_R[ 10 + 10*lda ] = MAGMA_D_MAKE( 0.0, 0.0 );
//magma_dsetmatrix( N, N, h_R, lda, d_A, ldda );
gpu_time = magma_wtime();
magma_dpotri_gpu( opts.uplo, N, d_A, ldda, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dpotri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
lapackf77_dpotrf( &opts.uplo, &N, h_A, &lda, &info );
cpu_time = magma_wtime();
lapackf77_dpotri( &opts.uplo, &N, h_A, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dpotri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_dgetmatrix( N, N, d_A, ldda, h_R, lda );
error = lapackf77_dlange("f", &N, &N, h_A, &lda, work);
blasf77_daxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_dlange("f", &N, &N, h_R, &lda, work) / error;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e%s\n",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "" : " failed") );
status |= ! (error < tol);
}
else {
printf("%5d --- ( --- ) %7.2f (%7.2f) ---\n",
(int) N, gpu_perf, gpu_time );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
SEXP magma_dpoMatrix_solve(SEXP x)
{
#ifdef HIPLAR_WITH_MAGMA
SEXP Chol = magma_dpoMatrix_chol(x);
SEXP val = PROTECT(NEW_OBJECT(MAKE_CLASS("dpoMatrix")));
int *dims = INTEGER(GET_SLOT(x, Matrix_DimSym)), info;
SET_SLOT(val, Matrix_factorSym, allocVector(VECSXP, 0));
slot_dup(val, Chol, Matrix_uploSym);
slot_dup(val, Chol, Matrix_xSym);
slot_dup(val, Chol, Matrix_DimSym);
SET_SLOT(val, Matrix_DimNamesSym,
duplicate(GET_SLOT(x, Matrix_DimNamesSym)));
double *A = REAL(GET_SLOT(val, Matrix_xSym));
int N = *dims;
int lda = N;
const char *uplo = uplo_P(val);
if(GPUFlag == 0) {
F77_CALL(dpotri)(uplo_P(val), dims, A, dims, &info);
}
else if(GPUFlag == 1 && Interface == 0) {
#ifdef HIPLAR_DBG
R_ShowMessage("DBG: Solving using magma_dpotri");
#endif
magma_dpotri(uplo[0], N, A, lda, &info);
}
else if(GPUFlag == 1 && Interface == 1){
double *d_A;
cublasStatus retStatus;
cublasAlloc( N * lda , sizeof(double), (void**)&d_A);
#ifdef HIPLAR_DBG
R_ShowMessage("DBG: Solving using magma_dpotri_gpu");
#endif
/* Error Checking */
retStatus = cublasGetError ();
if (retStatus != CUBLAS_STATUS_SUCCESS)
error(_("CUBLAS: Error in Memory Allocation"));
/********************************************/
cublasSetVector( N * lda, sizeof(double), A, 1, d_A, 1);
/* Error Checking */
retStatus = cublasGetError ();
if (retStatus != CUBLAS_STATUS_SUCCESS)
error(_("CUBLAS: Error in Data Transfer to Device"));
/********************************************/
magma_dpotri_gpu(uplo[0], N, d_A, lda, &info);
cublasGetVector(N * lda, sizeof(double), d_A, 1, val, 1);
/* Error Checking */
retStatus = cublasGetError ();
if (retStatus != CUBLAS_STATUS_SUCCESS)
error(_("CUBLAS: Error in Data Transfer from Device"));
/********************************************/
cublasFree(d_A);
}
else
error(_("MAGMA/LAPACK/Interface Flag not defined correctly"));
if (info) {
if(info > 0)
error(_("the leading minor of order %d is not positive definite"),
info);
else /* should never happen! */
error(_("Lapack routine %s returned error code %d"), "dpotrf", info);
}
UNPROTECT(1);
return val;
#endif
return R_NilValue;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dpotrf
*/
int main( int argc, char** argv)
{
TESTING_CUDA_INIT();
magma_timestr_t start, end;
double flops, gpu_perf, cpu_perf;
double *h_A, *h_R;
double *d_A;
magma_int_t N = 0, n2, lda, ldda;
magma_int_t size[10] = {1024,2048,3072,4032,5184,6016,7040,8064,9088,10112};
magma_int_t i, info;
const char *uplo = MagmaUpperStr;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], matnorm;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
}
if (N>0) size[0] = size[9] = N;
else exit(1);
}
else {
printf("\nUsage: \n");
printf(" testing_dpotri_gpu -N %d\n\n", 1024);
}
/* Allocate host memory for the matrix */
n2 = size[9] * size[9];
ldda = ((size[9]+31)/32) * 32;
TESTING_MALLOC( h_A, double, n2);
TESTING_HOSTALLOC( h_R, double, n2);
TESTING_DEVALLOC( d_A, double, ldda*size[9] );
printf(" N CPU GFlop/s GPU GFlop/s ||R||_F / ||A||_F\n");
printf("========================================================\n");
for(i=0; i<10; i++){
N = size[i];
lda = N;
n2 = lda*N;
flops = FLOPS_DPOTRI( (double)N ) / 1000000;
ldda = ((N+31)/32)*32;
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
/* Symmetrize and increase the diagonal */
{
magma_int_t i, j;
for(i=0; i<N; i++) {
MAGMA_D_SET2REAL( h_A[i*lda+i], ( MAGMA_D_REAL(h_A[i*lda+i]) + 1.*N ) );
for(j=0; j<i; j++)
h_A[i*lda+j] = (h_A[j*lda+i]);
}
}
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
//cublasSetMatrix( N, N, sizeof(double), h_A, lda, d_A, ldda);
//magma_dpotrf_gpu(uplo[0], N, d_A, ldda, &info);
/* factorize matrix */
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda );
magma_dpotrf_gpu(uplo[0], N, d_A, ldda, &info);
// check for exact singularity
//magma_dgetmatrix( N, N, d_A, ldda, h_R, lda );
//h_R[ 10 + 10*lda ] = MAGMA_D_MAKE( 0.0, 0.0 );
//magma_dsetmatrix( N, N, h_R, lda, d_A, ldda );
start = get_current_time();
magma_dpotri_gpu(uplo[0], N, d_A, ldda, &info);
end = get_current_time();
if (info != 0)
printf("magma_dpotri_gpu returned error %d\n", (int) info);
gpu_perf = flops / GetTimerValue(start, end);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_dpotrf(uplo, &N, h_A, &lda, &info);
start = get_current_time();
lapackf77_dpotri(uplo, &N, h_A, &lda, &info);
end = get_current_time();
if (info != 0)
printf("lapackf77_dpotri returned error %d\n", (int) info);
cpu_perf = flops / GetTimerValue(start, end);
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
magma_dgetmatrix( N, N, d_A, ldda, h_R, lda );
matnorm = lapackf77_dlange("f", &N, &N, h_A, &lda, work);
blasf77_daxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
printf("%5d %6.2f %6.2f %e\n",
(int) size[i], cpu_perf, gpu_perf,
lapackf77_dlange("f", &N, &N, h_R, &lda, work) / matnorm);
if (argc != 1)
break;
}
/* Memory clean up */
TESTING_FREE( h_A );
TESTING_HOSTFREE( h_R );
TESTING_DEVFREE( d_A );
/* Shutdown */
TESTING_CUDA_FINALIZE();
}
