zherk.c

Go to the documentation of this file.

#include "common.h"

/***************************************************************************/
int PLASMA_zherk(PLASMA_enum uplo, PLASMA_enum trans, int N, int K,
                 double alpha, PLASMA_Complex64_t *A, int LDA,
                 double beta,  PLASMA_Complex64_t *C, int LDC)
{
    int NB;
    int Am, An;
    int status;
    PLASMA_desc descA, 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_zherk", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }

    /* Check input arguments */
    if ((uplo != PlasmaUpper) && (uplo != PlasmaLower)) {
        plasma_error("PLASMA_zherk", "illegal value of uplo");
        return -1;
    }
    if ((trans != PlasmaNoTrans) && (trans != PlasmaConjTrans)) {
        plasma_error("PLASMA_zherk", "illegal value of trans");
        return -2;
    }
    if ( trans == PlasmaNoTrans ) { 
        Am = N; An = K;
    } else {
        Am = K; An = N;
    }
    if (N < 0) {
        plasma_error("PLASMA_zherk", "illegal value of N");
        return -3;
    }
    if (K < 0) {
        plasma_error("PLASMA_zherk", "illegal value of K");
        return -4;
    }
    if (LDA < max(1, Am)) {
        plasma_error("PLASMA_zherk", "illegal value of LDA");
        return -7;
    }
    if (LDC < max(1, N)) {
        plasma_error("PLASMA_zherk", "illegal value of LDC");
        return -10;
    }

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

    /* Tune NB depending on M, N & NRHS; Set NBNB */
    status = plasma_tune(PLASMA_FUNC_ZHERK, N, K, 0);
    if (status != PLASMA_SUCCESS) {
        plasma_error("PLASMA_zherk", "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( descC, C, NB, NB, LDC, N,  0, 0, N,  N,  plasma_desc_mat_free(&(descA)); plasma_desc_mat_free(&(descC)));
    } else {
        plasma_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An );
        plasma_ziplap2tile( descC, C, NB, NB, LDC, N,  0, 0, N,  N );
    }

    /* Call the tile interface */
    PLASMA_zherk_Tile_Async(uplo, trans, alpha, &descA, 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(&descC);
    } else {
        plasma_ziptile2lap( descA, A, NB, NB, LDA, An );
        plasma_ziptile2lap( descC, C, NB, NB, LDC, N );
        plasma_dynamic_sync();
    }

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

/***************************************************************************/
int PLASMA_zherk_Tile(PLASMA_enum uplo, PLASMA_enum trans,
                      double alpha, PLASMA_desc *A,
                      double 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_zherk_Tile", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    plasma_sequence_create(plasma, &sequence);
    PLASMA_zherk_Tile_Async(uplo, trans, alpha, A, beta, C, sequence, &request);
    plasma_dynamic_sync();
    status = sequence->status;
    plasma_sequence_destroy(plasma, sequence);
    return status;
}

/***************************************************************************/
int PLASMA_zherk_Tile_Async(PLASMA_enum uplo, PLASMA_enum trans,
                            double alpha, PLASMA_desc *A,
                            double beta,  PLASMA_desc *C,
                            PLASMA_sequence *sequence, PLASMA_request *request)
{
    plasma_context_t *plasma;
    PLASMA_desc descA = *A;
    PLASMA_desc descC = *C;
    int N, K;
    int Am, An, Amb;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_zherk_Tile_Async", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    if (sequence == NULL) {
        plasma_fatal_error("PLASMA_zherk_Tile_Async", "NULL sequence");
        return PLASMA_ERR_UNALLOCATED;
    }
    if (request == NULL) {
        plasma_fatal_error("PLASMA_zherk_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_zherk_Tile_Async", "invalid first descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (plasma_desc_check(&descC) != PLASMA_SUCCESS) {
        plasma_error("PLASMA_zherk_Tile_Async", "invalid third descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    /* Check input arguments */
    if ((uplo != PlasmaUpper) && (uplo != PlasmaLower)) {
        plasma_error("PLASMA_zherk", "illegal value of uplo");
        return plasma_request_fail(sequence, request, -1);
    }
    if ((trans != PlasmaNoTrans) && (trans != PlasmaConjTrans)) {
        plasma_error("PLASMA_zherk", "illegal value of transA");
        return plasma_request_fail(sequence, request, -2);
    }

    if ( trans == PlasmaNoTrans ) {
        Am  = descA.m;
        An  = descA.n;
        Amb = descA.mb;
    } else {
        Am  = descA.n;
        An  = descA.m;
        Amb = descA.nb;
    }

    if (descC.mb != descC.nb) {
        plasma_error("PLASMA_zherk_Tile_Async", "only square tiles are supported");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (Amb != descC.mb) {
        plasma_error("PLASMA_zherk_Tile_Async", "tile sizes have to match");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (descC.m != descC.n) {
        plasma_error("PLASMA_zherk_Tile_Async", "only square matrix C is supported");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (Am != descC.m) {
        plasma_error("PLASMA_zherk_Tile_Async", "sizes of matrices have to match");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }

    N = descC.m;
    K = An;

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

    plasma_parallel_call_8(plasma_pzherk,
        PLASMA_enum, uplo,
        PLASMA_enum, trans,
        double, alpha,
        PLASMA_desc, descA,
        double, beta,
        PLASMA_desc, descC,
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);

    return PLASMA_SUCCESS;
}