dlacpy.c

Go to the documentation of this file.

#include "common.h"

/***************************************************************************/
int PLASMA_dlacpy(PLASMA_enum uplo, int M, int N,
                  double *A, int LDA,
                  double *B, int LDB)
{
    int NB;
    int status;
    plasma_context_t *plasma;
    PLASMA_sequence *sequence = NULL;
    PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
    PLASMA_desc descA, descB;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_dlacpy", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    /* Check input arguments */
    if ( (uplo != PlasmaUpperLower) && 
         (uplo != PlasmaUpper) &&
         (uplo != PlasmaLower) ) {
        plasma_error("PLASMA_dlacpy", "illegal value of uplo");
        return -1;
    }
    if (M < 0) {
        plasma_error("PLASMA_dlacpy", "illegal value of M");
        return -2;
    }
    if (N < 0) {
        plasma_error("PLASMA_dlacpy", "illegal value of N");
        return -3;
    }
    if (LDA < max(1, M)) {
        plasma_error("PLASMA_dlacpy", "illegal value of LDA");
        return -5;
    }
    if (LDB < max(1, M)) {
        plasma_error("PLASMA_dlacpy", "illegal value of LDB");
        return -7;
    }

    /* Quick return */
    if (min(N, M) == 0)
      return (double)0.0;

    /* Tune NB depending on M, N & NRHS; Set NBNB */
    status = plasma_tune(PLASMA_FUNC_DGEMM, M, N, 0);
    if (status != PLASMA_SUCCESS) {
        plasma_error("PLASMA_dlacpy", "plasma_tune() failed");
        return status;
    }

    /* Set NT */
    NB   = PLASMA_NB;

    plasma_sequence_create(plasma, &sequence);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_dooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, plasma_desc_mat_free(&(descA)) );
        plasma_dooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, plasma_desc_mat_free(&(descA)); plasma_desc_mat_free(&(descB)) );
    } else {
        plasma_diplap2tile(  descA, A, NB, NB, LDA, N, 0, 0, M, N);
        plasma_diplap2tile(  descB, B, NB, NB, LDA, N, 0, 0, M, N);
    }

    /* Call the tile interface */
    PLASMA_dlacpy_Tile_Async(uplo, &descA, &descB, sequence, &request);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_dooptile2lap( descB, B, NB, NB, LDB, N );
        plasma_dynamic_sync();
        plasma_desc_mat_free(&descA);
        plasma_desc_mat_free(&descB);
    } else {
        plasma_diptile2lap( descB, B, NB, NB, LDB, N );
        plasma_dynamic_sync();
    }

    plasma_sequence_destroy(plasma, sequence);
    return PLASMA_SUCCESS;
}

/***************************************************************************/
int PLASMA_dlacpy_Tile(PLASMA_enum uplo, PLASMA_desc *A, PLASMA_desc *B)
{
    plasma_context_t *plasma;
    PLASMA_sequence *sequence = NULL;
    PLASMA_request request = PLASMA_REQUEST_INITIALIZER;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_dlacpy_Tile", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    plasma_sequence_create(plasma, &sequence);
    PLASMA_dlacpy_Tile_Async(uplo, A, B, sequence, &request);
    plasma_dynamic_sync();
    plasma_sequence_destroy(plasma, sequence);
    return PLASMA_SUCCESS;
}

/***************************************************************************/
int PLASMA_dlacpy_Tile_Async(PLASMA_enum uplo, PLASMA_desc *A, PLASMA_desc *B,
                             PLASMA_sequence *sequence, PLASMA_request *request)
{
    PLASMA_desc descA = *A;
    PLASMA_desc descB = *B;
    plasma_context_t *plasma;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_dlacpy_Tile_Async", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    if (sequence == NULL) {
        plasma_fatal_error("PLASMA_dlacpy_Tile_Async", "NULL sequence");
        return PLASMA_ERR_UNALLOCATED;
    }
    if (request == NULL) {
        plasma_fatal_error("PLASMA_dlacpy_Tile_Async", "NULL request");
        return PLASMA_ERR_UNALLOCATED;
    }
    /* Check sequence status */
    if (sequence->status == PLASMA_SUCCESS)
        request->status = PLASMA_SUCCESS;
    else
        return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);

    /* Check descriptors for correctness */
    if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
        plasma_error("PLASMA_dlacpy_Tile_Async", "invalid descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    /* Check input arguments */
    if (descA.nb != descA.mb) {
        plasma_error("PLASMA_dlacpy_Tile_Async", "only square tiles supported");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    /* Check input arguments */
    if ( (uplo != PlasmaUpperLower) && 
         (uplo != PlasmaUpper) &&
         (uplo != PlasmaLower) ) {
        plasma_error("PLASMA_dlacpy_Tile_Async", "illegal value of uplo");
        return -1;
    }
    /* Quick return */
    if (min(descA.m, descA.n) == 0) {
        return PLASMA_SUCCESS;
    }

    plasma_parallel_call_5(plasma_pdlacpy,
        PLASMA_enum, uplo,
        PLASMA_desc, descA,
        PLASMA_desc, descB,
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);

    return PLASMA_SUCCESS;
}