cheev.c

Go to the documentation of this file.

#include <lapacke.h>
#include "common.h"

/***************************************************************************/
int PLASMA_cheev(PLASMA_enum jobz, PLASMA_enum uplo, int N,
                 PLASMA_Complex32_t *A, int LDA,
                 float *W,
                 PLASMA_desc *descT,
                 PLASMA_Complex32_t *Q, int LDQ)
{
    int NB, IB, NT;
    int status;
    plasma_context_t *plasma;
    PLASMA_sequence *sequence = NULL;
    PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
    PLASMA_desc descA, descQ;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_error("PLASMA_cheev", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }

    /* Tune NB & IB depending on N; Set NBNB */
    status = plasma_tune(PLASMA_FUNC_CHEEV, N, N, 0);
    if (status != PLASMA_SUCCESS) {
        plasma_error("PLASMA_cheev", "plasma_tune() failed");
        return status;
    }

    /* Set NT */
    NB = PLASMA_NB;
    IB = PLASMA_IB;
    NT = (N%NB==0) ? (N/NB) : (N/NB+1);

    /* Check input arguments */
    if (jobz != PlasmaNoVec && jobz != PlasmaVec) {
        plasma_error("PLASMA_cheev", "illegal value of jobz");
        return -1;
    }
    if (uplo != PlasmaLower && uplo != PlasmaUpper) {
        plasma_error("PLASMA_cheev", "illegal value of uplo");
        return -2;
    }
    if (N < 0) {
        plasma_error("PLASMA_cheev", "illegal value of N");
        return -3;
    }
    if (LDA < max(1, N)) {
        plasma_error("PLASMA_cheev", "illegal value of LDA");
        return -5;
    }
    if ( (plasma_desc_check(descT) != PLASMA_SUCCESS) || 
         ( descT->m != NT*IB ) || (descT->n != NT*NB) ) {
        plasma_error("PLASMA_cheev", "invalid T descriptor");
        return -7;
    }
    if (LDQ < max(1, N)) {
        plasma_error("PLASMA_cheev", "illegal value of LDQ");
        return -9;
    }
    /* Quick return */
    if (N == 0)
        return PLASMA_SUCCESS;

    if (jobz == PlasmaVec) {
        plasma_error("PLASMA_cheev", "computing the eigenvectors is not supported in this version");
        return -1;
    }

    plasma_sequence_create(plasma, &sequence);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_cooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, plasma_desc_mat_free(&(descA)) );
        if (jobz == PlasmaVec) {
            /* No need for conversion, it's just output */
            plasma_cdesc_alloc( descQ, NB, NB, LDQ, N, 0, 0, N, N, 
                                plasma_desc_mat_free(&(descA)); plasma_desc_mat_free(&(descQ)) );
        }
    } else {
        plasma_ciplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N );
        if (jobz == PlasmaVec) {
            /* No need for conversion, it's just output */
            descQ = plasma_desc_init(
                PlasmaComplexFloat, NB, NB, NB*NB,
                LDQ, N, 0, 0, N, N);
            descQ.mat = Q;
        }
    }

    /* Call the tile interface */
    PLASMA_cheev_Tile_Async(jobz, uplo, &descA, W, descT, &descQ, sequence, &request);

    if ( PLASMA_TRANSLATION == PLASMA_OUTOFPLACE ) {
        plasma_cooptile2lap( descA, A, NB, NB, LDA, N );
        if (jobz == PlasmaVec) {
           plasma_cooptile2lap( descQ, Q, NB, NB, LDQ, N );
        }
        plasma_dynamic_sync();
        plasma_desc_mat_free(&descA);
        if (jobz == PlasmaVec)
           plasma_desc_mat_free(&descQ);
    } else {
        plasma_ciptile2lap( descA, A, NB, NB, LDA, N );
        if (jobz == PlasmaVec)
           plasma_ciptile2lap( descQ, Q, NB, NB, LDQ, N );
        plasma_dynamic_sync();
    }

    status = sequence->status;
    plasma_sequence_destroy(plasma, sequence);
    return status;
}
/***************************************************************************/
int PLASMA_cheev_Tile(PLASMA_enum jobz, PLASMA_enum uplo,
                      PLASMA_desc *A, float *W, 
                      PLASMA_desc *T, PLASMA_desc *Q)
{
    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_cheev_Tile", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    plasma_sequence_create(plasma, &sequence);
    PLASMA_cheev_Tile_Async(jobz, uplo, A, W, T, Q, sequence, &request);
    plasma_dynamic_sync();
    status = sequence->status;
    plasma_sequence_destroy(plasma, sequence);
    return status;
}

/***************************************************************************/
int PLASMA_cheev_Tile_Async(PLASMA_enum jobz, PLASMA_enum uplo,
                            PLASMA_desc *A,
                            float *W,
                            PLASMA_desc *T,
                            PLASMA_desc *Q,
                            PLASMA_sequence *sequence, PLASMA_request *request)
{
    PLASMA_desc descA = *A;
    PLASMA_desc descT = *T;
    float *E;
    plasma_context_t *plasma;

    plasma = plasma_context_self();
    if (plasma == NULL) {
        plasma_fatal_error("PLASMA_cheev_Tile_Async", "PLASMA not initialized");
        return PLASMA_ERR_NOT_INITIALIZED;
    }
    if (sequence == NULL) {
        plasma_fatal_error("PLASMA_cheev_Tile_Async", "NULL sequence");
        return PLASMA_ERR_UNALLOCATED;
    }
    if (request == NULL) {
        plasma_fatal_error("PLASMA_cheev_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_cheev_Tile_Async", "invalid descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (plasma_desc_check(&descT) != PLASMA_SUCCESS) {
        plasma_error("PLASMA_cheev_Tile_Async", "invalid descriptor");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (jobz == PlasmaVec){
       if (plasma_desc_check(Q) != PLASMA_SUCCESS) {
           plasma_error("PLASMA_cheev_Tile_Async", "invalid descriptor");
           return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
       }
    }
    /* Check input arguments */
    if (jobz != PlasmaNoVec && jobz != PlasmaVec) {
        plasma_error("PLASMA_cheev_Tile_Async", "illegal value of jobz");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (uplo != PlasmaLower && uplo != PlasmaUpper) {
        plasma_error("PLASMA_cheev_Tile_Async", "illegal value of uplo");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (descA.m != descA.n) {
        plasma_error("PLASMA_cheev_Tile_Async", "matrix need to be square");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (descA.nb != descA.mb) {
        plasma_error("PLASMA_cheev_Tile_Async", "only square tiles supported");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (jobz == PlasmaVec) {
        plasma_error("PLASMA_cheev_Tile_Async", "computing the eigenvectors is not supported in this version");
        return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
    }
    if (jobz == PlasmaVec){
       if (Q->nb != Q->mb) {
           plasma_error("PLASMA_cheev_Tile_Async", "only square tiles supported");
           return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
       }
    }

    E = (float *)plasma_shared_alloc(plasma, descA.n-1, PlasmaRealDouble);

    /* Currently NOT equivalent to LAPACK's
     */

    /* Reduction to tridiagonal form
     * with a two-stage approach.
     */


    /* Reduction to BAND tridiagonal form
     */
    plasma_dynamic_call_5(plasma_pcherbt,
        PLASMA_enum, uplo,
        PLASMA_desc, descA,
        PLASMA_desc, descT,
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);

    /*
     * Build the Q of the first stage
     */
    /* if (jobz == PlasmaVec){ */
    /*     /\* Initialize Q to Identity *\/ */
    /*     plasma_dynamic_call_6(plasma_pclaset, */
    /*                           PLASMA_enum, PlasmaUpperLower, */
    /*                           PLASMA_Complex32_t, 0.0, */
    /*                           PLASMA_Complex32_t, 1.0, */
    /*                           PLASMA_desc, descQ, */
    /*                           PLASMA_sequence*, sequence, */
    /*                           PLASMA_request*, request); */

    /*     /\* Accumulate the transformations from the first stage *\/ */
    /*     plasma_dynamic_call_6(plasma_pcungtr, */
    /*                           PLASMA_enum, uplo, */
    /*                           PLASMA_desc, descA, */
    /*                           PLASMA_desc, descQ, */
    /*                           PLASMA_desc, descT, */
    /*                           PLASMA_sequence*, sequence, */
    /*                           PLASMA_request*, request); */
    /* } */
    
    /* Set the V's to zero before the 2nd stage (bulge chasing) */
    plasma_dynamic_call_5(
        plasma_pclaset2,
        PLASMA_enum, uplo,
        PLASMA_Complex32_t, 0.0,
        PLASMA_desc, uplo == PlasmaLower ? plasma_desc_submatrix(descA, descA.mb, 0, descA.m-descA.mb, descA.n-descA.nb) : 
                                           plasma_desc_submatrix(descA, 0, descA.nb, descA.m-descA.mb, descA.n-descA.nb),
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);
    /* Reduction from BAND tridiagonal to the final condensed form
     */
    plasma_dynamic_call_7(plasma_pchbrdt,
        PLASMA_enum, uplo,
        PLASMA_desc, descA,
        float*, W,
        float*, E,
        PLASMA_desc, descT,
        PLASMA_sequence*, sequence,
        PLASMA_request*, request);

    /* 
     * For eigenvalues only, call DSTERF.
     * For eigenvectors, first call CUNGTR to generate the unitary matrix,
     * then call ZSTEQR.
     */
    plasma_dynamic_sync();
    if (jobz == PlasmaNoVec){
        LAPACKE_ssterf(descA.n, W, E);
    }else {
        LAPACKE_ssterf(descA.n, W, E);
        /* Accumulate the transformations from the second stage */
        /*
        plasma_dynamic_call_6(plasma_pcungtr,
            PLASMA_enum, uplo,
            PLASMA_desc, descA,
            PLASMA_desc, descQ,
            PLASMA_desc, descT,
            PLASMA_sequence*, sequence,
            PLASMA_request*, request);
        LAPACKE_csteqr(jobz, descA.n, W, E, Q->mat, Q->lm);
        */
    }
    /* If matrix was scaled, then rescale eigenvalues appropriately.
    */

    plasma_shared_free(plasma, E);

    return PLASMA_SUCCESS;
}