zgemm.c

Go to the documentation of this file.

#include "common.h"

/***************************************************************************/
int PLASMA_zgemm(PLASMA_enum transA, PLASMA_enum transB, int M, int N, int K,
                 PLASMA_Complex64_t alpha, PLASMA_Complex64_t *A, int LDA,
                                           PLASMA_Complex64_t *B, int LDB,
                 PLASMA_Complex64_t beta,  PLASMA_Complex64_t *C, int LDC)
{
    int NB;
    int Am, An, Bm, Bn;
    int status;
    PLASMA_desc descA, descB, descC;
    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_zgemm", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }

    /* Check input arguments */
    if ((transA != PlasmaNoTrans) && (transA != PlasmaTrans) && (transA != PlasmaConjTrans)) {
        plasma_error("PLASMA_zgemm", "illegal value of transA");
        return -1;
    }
    if ((transB != PlasmaNoTrans) && (transB != PlasmaTrans) && (transB != PlasmaConjTrans)) {
        plasma_error("PLASMA_zgemm", "illegal value of transB");
        return -2;
    }
    if ( transA == PlasmaNoTrans ) { 
        Am = M; An = K;
    } else {
        Am = K; An = M;
    }
    if ( transB == PlasmaNoTrans ) { 
        Bm = K; Bn = N;
    } else {
        Bm = N; Bn = K;
    }
    if (M < 0) {
        plasma_error("PLASMA_zgemm", "illegal value of M");
        return -3;
    }
    if (N < 0) {
        plasma_error("PLASMA_zgemm", "illegal value of N");
        return -4;
    }
    if (K < 0) {
        plasma_error("PLASMA_zgemm", "illegal value of N");
        return -5;
    }
    if (LDA < max(1, Am)) {
        plasma_error("PLASMA_zgemm", "illegal value of LDA");
        return -8;
    }
    if (LDB < max(1, Bm)) {
        plasma_error("PLASMA_zgemm", "illegal value of LDB");
        return -10;
    }
    if (LDC < max(1, M)) {
        plasma_error("PLASMA_zgemm", "illegal value of LDC");
        return -13;
    }

    /* Quick return */
    if (M == 0 || N == 0 ||
        ((alpha == (PLASMA_Complex64_t)0.0 || K == 0) && beta == (PLASMA_Complex64_t)1.0))
        return PLASMA_SUCCESS;

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

    /* Set MT & NT & KT */
    NB = PLASMA_NB;

    plasma_sequence_create(plasma, &sequence);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, plasma_desc_mat_free(&(descA)) );
        plasma_zooplap2tile( descB, B, NB, NB, LDB, Bn, 0, 0, Bm, Bn, plasma_desc_mat_free(&(descA)); plasma_desc_mat_free(&(descB)));
        plasma_zooplap2tile( descC, C, NB, NB, LDC, N,  0, 0, M,  N,  plasma_desc_mat_free(&(descA)); plasma_desc_mat_free(&(descB)); plasma_desc_mat_free(&(descC)));
    } else {
        plasma_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An );
        plasma_ziplap2tile( descB, B, NB, NB, LDB, Bn, 0, 0, Bm, Bn );
        plasma_ziplap2tile( descC, C, NB, NB, LDC, N,  0, 0, M,  N  );
    }

    /* Call the tile interface */
    PLASMA_zgemm_Tile_Async(
        transA, transB, alpha, &descA, &descB, beta, &descC, sequence, &request);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_zooptile2lap( descC, C, NB, NB, LDC, N );
        plasma_dynamic_sync();
        plasma_desc_mat_free(&descA);
        plasma_desc_mat_free(&descB);
        plasma_desc_mat_free(&descC);
    } else {
        plasma_ziptile2lap( descA, A, NB, NB, LDA, An );
        plasma_ziptile2lap( descB, B, NB, NB, LDB, Bn );
        plasma_ziptile2lap( descC, C, NB, NB, LDC, N  );
        plasma_dynamic_sync();
    }

    status = sequence->status;
    plasma_sequence_destroy(plasma, sequence);
    return status;
}

/***************************************************************************/
int PLASMA_zgemm_Tile(PLASMA_enum transA, PLASMA_enum transB,
                      PLASMA_Complex64_t alpha, PLASMA_desc *A, PLASMA_desc *B,
                      PLASMA_Complex64_t beta,  PLASMA_desc *C)
{
    plasma_context_t *plasma;
    PLASMA_sequence *sequence = NULL;
    PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
    int status;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_zgemm_Tile", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    plasma_sequence_create(plasma, &sequence);
    PLASMA_zgemm_Tile_Async(transA, transB, alpha, A, B, beta, C, sequence, &request);
    plasma_dynamic_sync();
    status = sequence->status;
    plasma_sequence_destroy(plasma, sequence);
    return status;
}

/***************************************************************************/
int PLASMA_zgemm_Tile_Async(PLASMA_enum transA, PLASMA_enum transB,
                            PLASMA_Complex64_t alpha, PLASMA_desc *A, PLASMA_desc *B,
                            PLASMA_Complex64_t beta,  PLASMA_desc *C,
                            PLASMA_sequence *sequence, PLASMA_request *request)
{
    plasma_context_t *plasma;
    PLASMA_desc descA = *A;
    PLASMA_desc descB = *B;
    PLASMA_desc descC = *C;
    int M, N, K;
    int Am, An, Ai, Aj, Amb, Anb;
    int Bm, Bn, Bi, Bj, Bmb, Bnb;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_zgemm_Tile_Async", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    if (sequence == NULL) {
        plasma_fatal_error("PLASMA_zgemm_Tile_Async", "NULL sequence");
        return PLASMA_ERR_UNALLOCATED;
    }
    if (request == NULL) {
        plasma_fatal_error("PLASMA_zgemm_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_zgemm_Tile_Async", "invalid first descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (plasma_desc_check(&descB) != PLASMA_SUCCESS) {
        plasma_error("PLASMA_zgemm_Tile_Async", "invalid second descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (plasma_desc_check(&descC) != PLASMA_SUCCESS) {
        plasma_error("PLASMA_zgemm_Tile_Async", "invalid third descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    /* Check input arguments */
    if ((transA != PlasmaNoTrans) && (transA != PlasmaTrans) && (transA != PlasmaConjTrans)) {
        plasma_error("PLASMA_zgemm_Tile_Async", "illegal value of transA");
        return plasma_request_fail(sequence, request, -1);
    }
    if ((transB != PlasmaNoTrans) && (transB != PlasmaTrans) && (transB != PlasmaConjTrans)) {
        plasma_error("PLASMA_zgemm_Tile_Async", "illegal value of transB");
        return plasma_request_fail(sequence, request, -2);
    }

    if ( transA == PlasmaNoTrans ) {
        Am  = descA.m;
        An  = descA.n;
        Amb = descA.mb;
        Anb = descA.nb;
        Ai  = descA.i;
        Aj  = descA.j;
    } else {
        Am  = descA.n;
        An  = descA.m;
        Amb = descA.nb;
        Anb = descA.mb;
        Ai  = descA.j;
        Aj  = descA.i;
    }

    if ( transB == PlasmaNoTrans ) {
        Bm  = descB.m;
        Bn  = descB.n;
        Bmb = descB.mb;
        Bnb = descB.nb;
        Bi  = descB.i;
        Bj  = descB.j;
    } else {
        Bm  = descB.n;
        Bn  = descB.m;
        Bmb = descB.nb;
        Bnb = descB.mb;
        Bi  = descB.j;
        Bj  = descB.i;
    }

    if ( (Amb != descC.mb) || (Anb != Bmb) || (Bnb != descC.nb) ) {
        plasma_error("PLASMA_zgemm_Tile_Async", "tile sizes have to match");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if ( (Am != descC.m) || (An != Bm) || (Bn != descC.n) ) {
        plasma_error("PLASMA_zgemm_Tile_Async", "sizes of matrices have to match");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if ( (Ai != descC.i) || (Aj != Bi) || (Bj != descC.j) ) {
        plasma_error("PLASMA_zgemm_Tile_Async", "start indexes have to match");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }

    M = descC.m;
    N = descC.n;
    K = An;

    /* Quick return */
    if (M == 0 || N == 0 ||
        ((alpha == (PLASMA_Complex64_t)0.0 || K == 0) && beta == (PLASMA_Complex64_t)1.0))
        return PLASMA_SUCCESS;

    plasma_parallel_call_9(plasma_pzgemm,
        PLASMA_enum, transA,
        PLASMA_enum, transB,
        PLASMA_Complex64_t, alpha,
        PLASMA_desc, descA,
        PLASMA_desc, descB,
        PLASMA_Complex64_t, beta,
        PLASMA_desc, descC,
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);

    return PLASMA_SUCCESS;
}