slange.c

Go to the documentation of this file.

#include "common.h"

/***************************************************************************/
float PLASMA_slange(PLASMA_enum norm, int M, int N,
                     float *A, int LDA, float *work)
{
    int NB;
    int status;
    float value;
    plasma_context_t *plasma;
    PLASMA_sequence *sequence = NULL;
    PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
    PLASMA_desc descA;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_slange", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    /* Check input arguments */
    if ( (norm != PlasmaMaxNorm) && (norm != PlasmaOneNorm)
        && (norm != PlasmaInfNorm) && (norm != PlasmaFrobeniusNorm) ) {
        plasma_error("PLASMA_slange", "illegal value of norm");
        return -1;
    }
    if (M < 0) {
        plasma_error("PLASMA_slange", "illegal value of M");
        return -2;
    }
    if (N < 0) {
        plasma_error("PLASMA_slange", "illegal value of N");
        return -3;
    }
    if (LDA < max(1, M)) {
        plasma_error("PLASMA_slange", "illegal value of LDA");
        return -5;
    }

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

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

    /* Set NT */
    NB   = PLASMA_NB;

    plasma_sequence_create(plasma, &sequence);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_sooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, plasma_desc_mat_free(&(descA)) );
    } else {
        plasma_siplap2tile(  descA, A, NB, NB, LDA, N, 0, 0, M, N);
    }

    /* Call the tile interface */
    PLASMA_slange_Tile_Async(norm, &descA, work, &value, sequence, &request);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_dynamic_sync();
        plasma_desc_mat_free(&descA);
    } else {
        plasma_siptile2lap( descA, A, NB, NB, LDA, N );
        plasma_dynamic_sync();
    }

    plasma_sequence_destroy(plasma, sequence);
    return value;
}

/***************************************************************************/
float PLASMA_slange_Tile(PLASMA_enum norm, PLASMA_desc *A, float *work)
{
    plasma_context_t *plasma;
    PLASMA_sequence *sequence = NULL;
    PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
    float value;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_slange_Tile", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    plasma_sequence_create(plasma, &sequence);
    PLASMA_slange_Tile_Async(norm, A, work, &value, sequence, &request);
    plasma_dynamic_sync();
    plasma_sequence_destroy(plasma, sequence);
    return value;
}

/***************************************************************************/
int PLASMA_slange_Tile_Async(PLASMA_enum norm, PLASMA_desc *A, float *work, float *value,
                             PLASMA_sequence *sequence, PLASMA_request *request)
{
    PLASMA_desc descA = *A;
    plasma_context_t *plasma;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_slange_Tile", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    if (sequence == NULL) {
        plasma_fatal_error("PLASMA_slange_Tile", "NULL sequence");
        return PLASMA_ERR_UNALLOCATED;
    }
    if (request == NULL) {
        plasma_fatal_error("PLASMA_slange_Tile", "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_slange_Tile", "invalid descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    /* Check input arguments */
    if (descA.nb != descA.mb) {
        plasma_error("PLASMA_slange_Tile", "only square tiles supported");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if ( (norm != PlasmaMaxNorm) && (norm != PlasmaOneNorm)
         && (norm != PlasmaInfNorm) && (norm != PlasmaFrobeniusNorm) ) {
        plasma_error("PLASMA_slange", "illegal value of norm");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    /* Quick return */
    if (min(descA.m, descA.n) == 0) {
        *value = 0.0;
        return PLASMA_SUCCESS;
    }

    plasma_parallel_call_6(plasma_pslange,
        PLASMA_enum, norm,
        PLASMA_desc, descA,
        float*, work,
        float*, value,
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);

    return PLASMA_SUCCESS;
}