Exemplo n.º 1
0
/**
    Purpose
    -------
    CHETRD reduces a complex Hermitian matrix A to real symmetric
    tridiagonal form T by an orthogonal similarity transformation:
    Q\*\*H * A * Q = T.

    Arguments
    ---------
    @param[in]
    num_gpus INTEGER
             The number of GPUs.  num_gpus > 0.

    @param[in]
    num_streams INTEGER
             The number of GPU streams used for update.  10 >= num_streams > 0.

    @param[in]
    uplo     magma_uplo_t
      -      = MagmaUpper:  Upper triangle of A is stored;
      -      = MagmaLower:  Lower triangle of A is stored.

    @param[in]
    n        INTEGER
             The order of the matrix A.  N >= 0.

    @param[in,out]
    A        COMPLEX array, dimension (LDA,N)
             On entry, the Hermitian matrix A.  If UPLO = MagmaUpper, the leading
             N-by-N upper triangular part of A contains the upper
             triangular part of the matrix A, and the strictly lower
             triangular part of A is not referenced.  If UPLO = MagmaLower, the
             leading N-by-N lower triangular part of A contains the lower
             triangular part of the matrix A, and the strictly upper
             triangular part of A is not referenced.
             On exit, if UPLO = MagmaUpper, the diagonal and first superdiagonal
             of A are overwritten by the corresponding elements of the
             tridiagonal matrix T, and the elements above the first
             superdiagonal, with the array TAU, represent the orthogonal
             matrix Q as a product of elementary reflectors; if UPLO
             = MagmaLower, the diagonal and first subdiagonal of A are over-
             written by the corresponding elements of the tridiagonal
             matrix T, and the elements below the first subdiagonal, with
             the array TAU, represent the orthogonal matrix Q as a product
             of elementary reflectors. See Further Details.

    @param[in]
    lda      INTEGER
             The leading dimension of the array A.  LDA >= max(1,N).

    @param[out]
    d        COMPLEX array, dimension (N)
             The diagonal elements of the tridiagonal matrix T:
             D(i) = A(i,i).
 
    @param[out]
    e        COMPLEX array, dimension (N-1)
             The off-diagonal elements of the tridiagonal matrix T:
             E(i) = A(i,i+1) if UPLO = MagmaUpper, E(i) = A(i+1,i) if UPLO = MagmaLower.

    @param[out]
    tau      COMPLEX array, dimension (N-1)
             The scalar factors of the elementary reflectors (see Further
             Details).

    @param[out]
    work     (workspace) COMPLEX array, dimension (MAX(1,LWORK))
             On exit, if INFO = 0, WORK(1) returns the optimal LWORK.

    @param[in]
    lwork    INTEGER
             The dimension of the array WORK.  LWORK >= 1.
             For optimum performance LWORK >= N*NB, where NB is the
             optimal blocksize.
    \n
             If LWORK = -1, then a workspace query is assumed; the routine
             only calculates the optimal size of the WORK array, returns
             this value as the first entry of the WORK array, and no error
             message related to LWORK is issued by XERBLA.

    @param[out]
    info     INTEGER
      -      = 0:  successful exit
      -      < 0:  if INFO = -i, the i-th argument had an illegal value

    Further Details
    ---------------
    If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary
    reflectors

       Q = H(n-1) . . . H(2) H(1).

    Each H(i) has the form

       H(i) = I - tau * v * v'

    where tau is a complex scalar, and v is a complex vector with
    v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in
    A(1:i-1,i+1), and tau in TAU(i).

    If UPLO = MagmaLower, the matrix Q is represented as a product of elementary
    reflectors

       Q = H(1) H(2) . . . H(n-1).

    Each H(i) has the form

       H(i) = I - tau * v * v'

    where tau is a complex scalar, and v is a complex vector with
    v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i),
    and tau in TAU(i).

    The contents of A on exit are illustrated by the following examples
    with n = 5:

    if UPLO = MagmaUpper:                if UPLO = MagmaLower:

      (  d   e   v2  v3  v4 )              (  d                  )
      (      d   e   v3  v4 )              (  e   d              )
      (          d   e   v4 )              (  v1  e   d          )
      (              d   e  )              (  v1  v2  e   d      )
      (                  d  )              (  v1  v2  v3  e   d  )

    where d and e denote diagonal and off-diagonal elements of T, and vi
    denotes an element of the vector defining H(i).

    @ingroup magma_cheev_comp
    ********************************************************************/
extern "C" magma_int_t
magma_chetrd_mgpu(
    magma_int_t num_gpus, magma_int_t num_streams, magma_uplo_t uplo, magma_int_t n,
    magmaFloatComplex *A, magma_int_t lda,
    float *d, float *e, magmaFloatComplex *tau,
    magmaFloatComplex *work, magma_int_t lwork,
    magma_int_t *info)
{
#define  A(i, j)     (A           + (j)*lda  + (i))
#define dA(id, i, j) (dA[(id)]    + (j)*ldda + (i))
#define dW(id, i, j) (dwork[(id)] + (j)*ldda + (i))

    const char* uplo_ = lapack_uplo_const( uplo );
    
    magma_int_t ln, ldda;
    magma_int_t nb = magma_get_chetrd_nb(n), ib;

    magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
    magmaFloatComplex c_one = MAGMA_C_ONE;
    float  d_one = MAGMA_D_ONE;
    float mv_time = 0.0;
#ifdef PROFILE_SY2RK
    float up_time = 0.0;
#endif

    magma_int_t kk, nx;
    magma_int_t i = 0, ii, iii, j, did, i_n;
    magma_int_t iinfo;
    magma_int_t ldwork, lddwork, lwkopt, ldwork2;
    magma_int_t lquery;
    magma_queue_t stream[MagmaMaxGPUs][10];
    magmaFloatComplex *dx[MagmaMaxGPUs], *dy[MagmaMaxGPUs], *hwork;
    magmaFloatComplex *dwork2[MagmaMaxGPUs];

    *info = 0;
    int upper = (uplo == MagmaUpper);
    lquery = (lwork == -1);
    if (! upper && uplo != MagmaLower) {
        printf( " uplo = %c\n",uplo );
        *info = -1;
    } else if (n < 0) {
        *info = -2;
    } else if (lda < max(1,n)) {
        *info = -4;
    } else if (lwork < nb*n && ! lquery) {
        *info = -9;
    } else if ( num_streams > 2 ) {
        *info = 2;
    }

    /* Determine the block size. */
    ldwork = lddwork = n;
    lwkopt = n * nb;
    if (*info == 0) {
        work[0] = MAGMA_C_MAKE( lwkopt, 0 );
    }

    if (*info != 0) {
        magma_xerbla( __func__, -(*info) );
        return MAGMA_ERR_ILLEGAL_VALUE;
    }
    else if (lquery)
        return 0;

    /* Quick return if possible */
    if (n == 0) {
        work[0] = c_one;
        return 0;
    }

    magmaFloatComplex *dA[MagmaMaxGPUs];
    magmaFloatComplex *dwork[MagmaMaxGPUs];

    float times[11];
    for( did=0; did < 11; did++ )
        times[did] = 0;
//#define PROFILE_SY2RK
#ifdef PROFILE_SY2RK
    magma_event_t start, stop;
    float etime;
    magma_setdevice(0);
    magma_event_create( &start );
    magma_event_create( &stop  );
#endif
    ldda = lda;
    ln = ((nb*(1+n/(nb*num_gpus))+31)/32)*32;
    ldwork2 = (1+ n / nb + (n % nb != 0)) * ldda;
    for( did=0; did < num_gpus; did++ ) {
        magma_setdevice(did);
        // TODO fix memory leak
        if ( MAGMA_SUCCESS != magma_cmalloc(&dA[did],     ln*ldda+3*lddwork*nb) ||
             MAGMA_SUCCESS != magma_cmalloc(&dx[did],     num_streams*n) ||
             MAGMA_SUCCESS != magma_cmalloc(&dy[did],     num_streams*n) ||
             MAGMA_SUCCESS != magma_cmalloc(&dwork2[did], ldwork2 ) ) {
            for( i=0; i < did; i++ ) {
                magma_setdevice(i);
                magma_free(dA[i]);
                magma_free(dx[i]);
                magma_free(dy[i]);
            }
            *info = MAGMA_ERR_DEVICE_ALLOC;
            return *info;
        }
        dwork[did] = dA[did] + ln*ldda;
        
        for( kk=0; kk < num_streams; kk++ )
            magma_queue_create(&stream[did][kk]);
    }
    magma_setdevice(0);
    // TODO fix memory leak dwork2
    if ( MAGMA_SUCCESS != magma_cmalloc_pinned( &hwork, num_streams*num_gpus*n ) ) {
        for( i=0; i < num_gpus; i++ ) {
            magma_setdevice(i);
            magma_free(dA[i]);
            magma_free(dx[i]);
            magma_free(dy[i]);
        }
        *info = MAGMA_ERR_HOST_ALLOC;
        return *info;
    }

    if (n < 2048)
        nx = n;
    else
        nx = 512;

    if (upper) {
        /* Copy the matrix to the GPU */
        if (1 <= n-nx) {
            magma_chtodhe(num_gpus, uplo, n, nb, A, lda, dA, ldda, stream, &iinfo );
        }

        /*  Reduce the upper triangle of A.
            Columns 1:kk are handled by the unblocked method. */
        for (i = nb*((n-1)/nb); i >= nx; i -= nb) {
            ib = min(nb, n-i);

            ii  = nb*(i/(nb*num_gpus));
            did = (i/nb)%num_gpus;

            /* wait for the next panel */
            if (i != nb*((n-1)/nb)) {
                magma_setdevice(did);
                magma_queue_sync(stream[did][0]);
            }

            magma_clatrd_mgpu(num_gpus, uplo, n, i+ib, ib, nb,
                              A(0, 0), lda, e, tau,
                              work, ldwork,
                              dA, ldda, 0,
                              dwork, i+ib,
                              dwork2, ldwork2,
                              1, dx, dy, hwork,
                              stream, times);

            magma_cher2k_mgpu(num_gpus, MagmaUpper, MagmaNoTrans, nb, i, ib,
                         c_neg_one, dwork, i+ib, 0,
                         d_one,     dA,    ldda, 0,
                         num_streams, stream);

            /* get the next panel */
            if (i-nb >= nx ) {
                ib = min(nb, n-(i-nb));
                
                ii  = nb*((i-nb)/(nb*num_gpus));
                did = ((i-nb)/nb)%num_gpus;
                magma_setdevice(did);
                
                magma_cgetmatrix_async( (i-nb)+ib, ib,
                                        dA(did, 0, ii), ldda,
                                         A(0, i-nb),    lda,
                                        stream[did][0] );
            }

            /* Copy superdiagonal elements back into A, and diagonal
               elements into D */
            for (j = i; j < i+ib; ++j) {
                if ( j > 0 ) {
                    *A(j-1,j) = MAGMA_C_MAKE( e[j - 1], 0 );
                }
                d[j] = MAGMA_C_REAL( *A(j, j) );
            }
        } /* end of for i=... */
      
        if ( nx > 0 ) {
            if (1 <= n-nx) { /* else A is already on CPU */
                for (i=0; i < nx; i += nb) {
                    ib = min(nb, n-i);
                    ii  = nb*(i/(nb*num_gpus));
                    did = (i/nb)%num_gpus;
                
                    magma_setdevice(did);
                    magma_cgetmatrix_async( nx, ib,
                                            dA(did, 0, ii), ldda,
                                            A(0, i),        lda,
                                            stream[did][0] );
                }
            }
            
            for( did=0; did < num_gpus; did++ ) {
                magma_setdevice(did);
                magma_queue_sync(stream[did][0]);
            }
            /*  Use unblocked code to reduce the last or only block */
            lapackf77_chetd2(uplo_, &nx, A(0, 0), &lda, d, e, tau, &iinfo);
        }
    }
    else {
        trace_init( 1, num_gpus, num_streams, (CUstream_st**)stream );
        /* Copy the matrix to the GPU */
        if (1 <= n-nx) {
            magma_chtodhe(num_gpus, uplo, n, nb, A, lda, dA, ldda, stream, &iinfo );
        }

        /* Reduce the lower triangle of A */
        for (i = 0; i < n-nx; i += nb) {
            ib = min(nb, n-i);

            ii  = nb*(i/(nb*num_gpus));
            did = (i/nb)%num_gpus;
            /* Reduce columns i:i+ib-1 to tridiagonal form and form the
               matrix W which is needed to update the unreduced part of
               the matrix */

            /*   Get the current panel (no need for the 1st iteration) */
            if (i != 0) {
                magma_setdevice(did);
                trace_gpu_start( did, 0, "comm", "get" );
                magma_cgetmatrix_async( n-i, ib,
                                        dA(did, i, ii), ldda,
                                         A(i,i),        lda,
                                        stream[did][0] );
                trace_gpu_end( did, 0 );
                magma_queue_sync(stream[did][0]);
                magma_setdevice(0);
            }
            
            
            mv_time +=
            magma_clatrd_mgpu(num_gpus, uplo, n, n-i, ib, nb,
                              A(i, i), lda, &e[i],
                              &tau[i], work, ldwork,
                              dA, ldda, i,
                              dwork,  (n-i),
                              dwork2, ldwork2,
                              1, dx, dy, hwork,
                              stream, times );

#ifdef PROFILE_SY2RK
            magma_setdevice(0);
            if ( i > 0 ) {
                cudaEventElapsedTime(&etime, start, stop);
                up_time += (etime/1000.0);
            }
            magma_event_record(start, 0);
#endif
            magma_cher2k_mgpu(num_gpus, MagmaLower, MagmaNoTrans, nb, n-i-ib, ib,
                         c_neg_one, dwork, n-i, ib,
                         d_one, dA, ldda, i+ib, num_streams, stream);
#ifdef PROFILE_SY2RK
            magma_setdevice(0);
            magma_event_record(stop, 0);
#endif

            /* Copy subdiagonal elements back into A, and diagonal
               elements into D */
            for (j = i; j < i+ib; ++j) {
                if ( j+1 < n ) {
                    *A(j+1,j) = MAGMA_C_MAKE( e[j], 0 );
                }
                d[j] = MAGMA_C_REAL( *A(j, j) );
            }
        } /* for i=... */

        /* Use unblocked code to reduce the last or only block */
        if ( i < n ) {
            iii = i;
            i_n = n-i;
            if ( i > 0 ) {
                for (; i < n; i += nb) {
                    ib = min(nb, n-i);
                    ii  = nb*(i/(nb*num_gpus));
                    did = (i/nb)%num_gpus;
                
                    magma_setdevice(did);
                    magma_cgetmatrix_async( i_n, ib,
                                            dA(did, iii, ii), ldda,
                                             A(iii, i),       lda,
                                            stream[did][0] );
                }
                for( did=0; did < num_gpus; did++ ) {
                    magma_setdevice(did);
                    magma_queue_sync(stream[did][0]);
                }
            }
            lapackf77_chetrd(uplo_, &i_n, A(iii, iii), &lda, &d[iii], &e[iii],
                             &tau[iii], work, &lwork, &iinfo);
        }
    }
#ifdef PROFILE_SY2RK
    magma_setdevice(0);
    if ( n > nx ) {
        cudaEventElapsedTime(&etime, start, stop);
        up_time += (etime/1000.0);
    }
    magma_event_destroy( start );
    magma_event_destroy( stop  );
#endif

    trace_finalize( "chetrd.svg", "trace.css" );
    for( did=0; did < num_gpus; did++ ) {
        magma_setdevice(did);
        for( kk=0; kk < num_streams; kk++ )
            magma_queue_sync(stream[did][kk]);
        for( kk=0; kk < num_streams; kk++ )
            magma_queue_destroy(stream[did][kk]);
        magma_free(dA[did]);
        magma_free(dx[did]);
        magma_free(dy[did]);
        magma_free(dwork2[did]);
    }
    magma_setdevice(0);
    magma_free_pinned(hwork);
    work[0] = MAGMA_C_MAKE( lwkopt, 0 );

#ifdef PROFILE_SY2RK
    printf( " n=%d nb=%d\n",n,nb );
    printf( " Time in CLARFG: %.2e seconds\n",times[0] );
    printf( " Time in CHEMV : %.2e seconds\n",mv_time );
    printf( " Time in CHER2K: %.2e seconds\n",up_time );
#endif
    return MAGMA_SUCCESS;
} /* magma_chetrd */
Exemplo n.º 2
0
/**
    Purpose
    -------
    CHETRD reduces a complex Hermitian matrix A to real symmetric
    tridiagonal form T by an orthogonal similarity transformation:
    Q**H * A * Q = T.

    Arguments
    ---------
    @param[in]
    ngpu    INTEGER
            Number of GPUs to use. ngpu > 0.

    @param[in]
    nqueue  INTEGER
            The number of GPU queues used for update.  10 >= nqueue > 0.

    @param[in]
    uplo    magma_uplo_t
      -     = MagmaUpper:  Upper triangle of A is stored;
      -     = MagmaLower:  Lower triangle of A is stored.

    @param[in]
    n       INTEGER
            The order of the matrix A.  N >= 0.

    @param[in,out]
    A       COMPLEX array, dimension (LDA,N)
            On entry, the Hermitian matrix A.  If UPLO = MagmaUpper, the leading
            N-by-N upper triangular part of A contains the upper
            triangular part of the matrix A, and the strictly lower
            triangular part of A is not referenced.  If UPLO = MagmaLower, the
            leading N-by-N lower triangular part of A contains the lower
            triangular part of the matrix A, and the strictly upper
            triangular part of A is not referenced.
            On exit, if UPLO = MagmaUpper, the diagonal and first superdiagonal
            of A are overwritten by the corresponding elements of the
            tridiagonal matrix T, and the elements above the first
            superdiagonal, with the array TAU, represent the orthogonal
            matrix Q as a product of elementary reflectors; if UPLO
            = MagmaLower, the diagonal and first subdiagonal of A are over-
            written by the corresponding elements of the tridiagonal
            matrix T, and the elements below the first subdiagonal, with
            the array TAU, represent the orthogonal matrix Q as a product
            of elementary reflectors. See Further Details.

    @param[in]
    lda     INTEGER
            The leading dimension of the array A.  LDA >= max(1,N).

    @param[out]
    d       COMPLEX array, dimension (N)
            The diagonal elements of the tridiagonal matrix T:
            D(i) = A(i,i).
 
    @param[out]
    e       COMPLEX array, dimension (N-1)
            The off-diagonal elements of the tridiagonal matrix T:
            E(i) = A(i,i+1) if UPLO = MagmaUpper, E(i) = A(i+1,i) if UPLO = MagmaLower.

    @param[out]
    tau     COMPLEX array, dimension (N-1)
            The scalar factors of the elementary reflectors (see Further
            Details).

    @param[out]
    work    (workspace) COMPLEX array, dimension (MAX(1,LWORK))
            On exit, if INFO = 0, WORK[0] returns the optimal LWORK.

    @param[in]
    lwork   INTEGER
            The dimension of the array WORK.  LWORK >= N*NB, where NB is the
            optimal blocksize given by magma_get_chetrd_nb().
    \n
            If LWORK = -1, then a workspace query is assumed; the routine
            only calculates the optimal size of the WORK array, returns
            this value as the first entry of the WORK array, and no error
            message related to LWORK is issued by XERBLA.

    @param[out]
    info    INTEGER
      -     = 0:  successful exit
      -     < 0:  if INFO = -i, the i-th argument had an illegal value

    Further Details
    ---------------
    If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary
    reflectors

        Q = H(n-1) . . . H(2) H(1).

    Each H(i) has the form

        H(i) = I - tau * v * v'

    where tau is a complex scalar, and v is a complex vector with
    v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in
    A(1:i-1,i+1), and tau in TAU(i).

    If UPLO = MagmaLower, the matrix Q is represented as a product of elementary
    reflectors

        Q = H(1) H(2) . . . H(n-1).

    Each H(i) has the form

        H(i) = I - tau * v * v'

    where tau is a complex scalar, and v is a complex vector with
    v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i),
    and tau in TAU(i).

    The contents of A on exit are illustrated by the following examples
    with n = 5:

    if UPLO = MagmaUpper:                if UPLO = MagmaLower:

        (  d   e   v2  v3  v4 )              (  d                  )
        (      d   e   v3  v4 )              (  e   d              )
        (          d   e   v4 )              (  v1  e   d          )
        (              d   e  )              (  v1  v2  e   d      )
        (                  d  )              (  v1  v2  v3  e   d  )

    where d and e denote diagonal and off-diagonal elements of T, and vi
    denotes an element of the vector defining H(i).

    @ingroup magma_cheev_comp
    ********************************************************************/
extern "C" magma_int_t
magma_chetrd_mgpu(
    magma_int_t ngpu,
    magma_int_t nqueue, magma_uplo_t uplo, magma_int_t n,
    magmaFloatComplex *A, magma_int_t lda,
    float *d, float *e, magmaFloatComplex *tau,
    magmaFloatComplex *work, magma_int_t lwork,
    magma_int_t *info)
{
#define  A(i, j)     (A           + (j)*lda  + (i))
#define dA(id, i, j) (dA[(id)]    + (j)*ldda + (i))
#define dW(id, i, j) (dW[(id)] + (j)*ldda + (i))

    /* Constants */
    const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
    const magmaFloatComplex c_one     = MAGMA_C_ONE;
    const float             d_one     = MAGMA_D_ONE;
    
    /* Local variables */
    const char* uplo_ = lapack_uplo_const( uplo );
    
    magma_int_t nlocal, ldda;
    magma_int_t nb = magma_get_chetrd_nb(n), ib, ib2;

    #ifdef PROFILE_SY2RK
    float mv_time = 0.0;
    float up_time = 0.0;
    #endif

    magma_int_t kk, nx;
    magma_int_t i, ii, iii, j, dev, i_n;
    magma_int_t iinfo;
    magma_int_t ldwork, lddw, lwkopt, ldwork2, lhwork;
    
    // set pointers to NULL so it is safe to goto CLEANUP if any malloc fails.
    magma_queue_t queues[MagmaMaxGPUs][10] = { { NULL, NULL } };
    magma_queue_t queues0[MagmaMaxGPUs]    = { NULL };
    magmaFloatComplex *hwork = NULL;
    magmaFloatComplex_ptr dwork2[MagmaMaxGPUs] = { NULL };
    magmaFloatComplex_ptr dA[MagmaMaxGPUs]     = { NULL };
    magmaFloatComplex_ptr dW[MagmaMaxGPUs]     = { NULL };

    *info = 0;
    bool upper = (uplo == MagmaUpper);
    bool lquery = (lwork == -1);
    if (! upper && uplo != MagmaLower) {
        *info = -1;
    } else if (n < 0) {
        *info = -2;
    } else if (lda < max(1,n)) {
        *info = -4;
    } else if (lwork < nb*n && ! lquery) {
        *info = -9;
    } else if ( nqueue > 2 ) {
        *info = 2;  // TODO fix
    }

    /* Determine the block size. */
    ldwork = n;
    lwkopt = n * nb;
    if (*info == 0) {
        work[0] = magma_cmake_lwork( lwkopt );
    }

    if (*info != 0) {
        magma_xerbla( __func__, -(*info) );
        return *info;
    }
    else if (lquery) {
        return *info;
    }

    /* Quick return if possible */
    if (n == 0) {
        work[0] = c_one;
        return *info;
    }

    magma_device_t orig_dev;
    magma_getdevice( &orig_dev );

    //#define PROFILE_SY2RK
    #ifdef PROFILE_SY2RK
    float times[11] = { 0 };
    magma_event_t start, stop;
    float etime;
    magma_setdevice( 0 );
    magma_event_create( &start );
    magma_event_create( &stop  );
    #endif

    ldda = magma_roundup( lda, 32 );
    lddw = ldda;
    nlocal = nb*(1 + n/(nb*ngpu));
    ldwork2 = ldda*( magma_ceildiv( n, nb ) + 1);  // i.e., ldda*(blocks + 1)
    for( dev=0; dev < ngpu; dev++ ) {
        magma_setdevice( dev );
        // TODO fix memory leak
        if ( MAGMA_SUCCESS != magma_cmalloc( &dA[dev],     nlocal*ldda + 3*lddw*nb ) ||
             MAGMA_SUCCESS != magma_cmalloc( &dwork2[dev], ldwork2 ) ) {
            *info = MAGMA_ERR_DEVICE_ALLOC;
            goto CLEANUP;
        }
        dW[dev] = dA[dev] + nlocal*ldda;
        
        for( kk=0; kk < nqueue; kk++ ) {
            magma_device_t cdev;
            magma_getdevice( &cdev );
            magma_queue_create( cdev, &queues[dev][kk] );
        }
        queues0[dev] = queues[dev][0];
    }
    
    lhwork = nqueue*ngpu*n;
    if ( MAGMA_SUCCESS != magma_cmalloc_pinned( &hwork, lhwork ) ) {
        *info = MAGMA_ERR_HOST_ALLOC;
        goto CLEANUP;
    }

    // nx <= n is required
    // use LAPACK for n < 3000, otherwise switch at 512
    if (n < 3000)
        nx = n;
    else
        nx = 512;

    if (upper) {
        /* Copy the matrix to the GPU */
        if (1 <= n-nx) {
            magma_chtodhe( ngpu, uplo, n, nb, A, lda, dA, ldda, queues, &iinfo );
        }

        /*  Reduce the upper triangle of A.
            Columns 1:kk are handled by the unblocked method. */
        for (i = nb*((n-1)/nb); i >= nx; i -= nb) {
            ib = min(nb, n-i);

            ii  = nb*(i/(nb*ngpu));
            dev = (i/nb)%ngpu;

            /* wait for the next panel */
            if (i != nb*((n-1)/nb)) {
                magma_setdevice( dev );
                magma_queue_sync( queues[dev][0] );
            }

            magma_clatrd_mgpu( ngpu, uplo, i+ib, ib, nb,
                               A(0, 0), lda, e, tau,
                               work, ldwork,
                               dA, ldda, 0,
                               dW, i+ib,
                               hwork,  lhwork,
                               dwork2, ldwork2,
                               queues0 );

            magma_cher2k_mgpu( ngpu, MagmaUpper, MagmaNoTrans, nb, i, ib,
                               c_neg_one, dW, i+ib, 0,
                               d_one,     dA, ldda, 0,
                               nqueue, queues );

            /* get the next panel */
            if (i-nb >= nx ) {
                ib2 = min(nb, n-(i-nb));
                
                ii  = nb*((i-nb)/(nb*ngpu));
                dev = ((i-nb)/nb)%ngpu;
                magma_setdevice( dev );
                
                magma_cgetmatrix_async( (i-nb)+ib2, ib2,
                                        dA(dev, 0, ii), ldda,
                                        A(0, i-nb),     lda,
                                        queues[dev][0] );
            }

            /* Copy superdiagonal elements back into A, and diagonal
               elements into D */
            for (j = i; j < i+ib; ++j) {
                if ( j > 0 ) {
                    *A(j-1,j) = MAGMA_C_MAKE( e[j - 1], 0 );
                }
                d[j] = MAGMA_C_REAL( *A(j, j) );
            }
        } /* end of for i=... */
      
        if ( nx > 0 ) {
            if (1 <= n-nx) { /* else A is already on CPU */
                for (i=0; i < nx; i += nb) {
                    ib = min(nb, n-i);
                    ii  = nb*(i/(nb*ngpu));
                    dev = (i/nb)%ngpu;
                
                    magma_setdevice( dev );
                    magma_cgetmatrix_async( nx, ib,
                                            dA(dev, 0, ii), ldda,
                                            A(0, i),        lda,
                                            queues[dev][0] );
                }
            }
            
            for( dev=0; dev < ngpu; dev++ ) {
                magma_setdevice( dev );
                magma_queue_sync( queues[dev][0] );
            }
            /* Use CPU code to reduce the last or only block */
            lapackf77_chetrd( uplo_, &nx, A(0, 0), &lda, d, e, tau,
                              work, &lwork, &iinfo );
        }
    }
    else {
        trace_init( 1, ngpu, nqueue, queues );
        /* Copy the matrix to the GPU */
        if (1 <= n-nx) {
            magma_chtodhe( ngpu, uplo, n, nb, A, lda, dA, ldda, queues, &iinfo );
        }

        /* Reduce the lower triangle of A */
        for (i = 0; i < n-nx; i += nb) {
            ib = min(nb, n-i);

            ii  = nb*(i/(nb*ngpu));
            dev = (i/nb)%ngpu;
            /* Reduce columns i:i+ib-1 to tridiagonal form and form the
               matrix W which is needed to update the unreduced part of
               the matrix */

            /*   Get the current panel (no need for the 1st iteration) */
            if (i != 0) {
                magma_setdevice( dev );
                trace_gpu_start( dev, 0, "comm", "get" );
                magma_cgetmatrix_async( n-i, ib,
                                        dA(dev, i, ii), ldda,
                                        A(i,i),         lda,
                                        queues[dev][0] );
                trace_gpu_end( dev, 0 );
                magma_queue_sync( queues[dev][0] );
                magma_setdevice( 0 );
            }
            
            magma_clatrd_mgpu( ngpu, uplo, n-i, ib, nb,
                               A(i, i), lda, &e[i], &tau[i],
                               work, ldwork,
                               dA, ldda, i,
                               dW, n-i,
                               hwork,  lhwork,
                               dwork2, ldwork2,
                               queues0 );

            #ifdef PROFILE_SY2RK
            magma_setdevice( 0 );
            if ( i > 0 ) {
                cudaEventElapsedTime( &etime, start, stop );
                up_time += (etime/1000.0);
            }
            magma_event_record( start, 0 );
            #endif
            
            magma_cher2k_mgpu( ngpu, MagmaLower, MagmaNoTrans, nb, n-i-ib, ib,
                               c_neg_one, dW, n-i, ib,
                               d_one, dA, ldda, i+ib, nqueue, queues );
            
            #ifdef PROFILE_SY2RK
            magma_setdevice( 0 );
            magma_event_record( stop, 0 );
            #endif

            /* Copy subdiagonal elements back into A, and diagonal
               elements into D */
            for (j = i; j < i+ib; ++j) {
                if ( j+1 < n ) {
                    *A(j+1,j) = MAGMA_C_MAKE( e[j], 0 );
                }
                d[j] = MAGMA_C_REAL( *A(j, j) );
            }
        } /* for i=... */

        /* Use CPU code to reduce the last or only block */
        if ( i < n ) {
            iii = i;
            i_n = n-i;
            if ( i > 0 ) {
                for (; i < n; i += nb) {
                    ib = min(nb, n-i);
                    ii  = nb*(i/(nb*ngpu));
                    dev = (i/nb)%ngpu;
                
                    magma_setdevice( dev );
                    magma_cgetmatrix_async( i_n, ib,
                                            dA(dev, iii, ii), ldda,
                                            A(iii, i),        lda,
                                            queues[dev][0] );
                }
                for( dev=0; dev < ngpu; dev++ ) {
                    magma_setdevice( dev );
                    magma_queue_sync( queues[dev][0] );
                }
            }
            lapackf77_chetrd( uplo_, &i_n, A(iii, iii), &lda, &d[iii], &e[iii],
                              &tau[iii], work, &lwork, &iinfo );
        }
    }
    
    for( dev=0; dev < ngpu; dev++ ) {
        magma_setdevice( dev );
        for( kk=0; kk < nqueue; kk++ ) {
            magma_queue_sync( queues[dev][kk] );
        }
    }
    
    #ifdef PROFILE_SY2RK
    magma_setdevice( 0 );
    if ( n > nx ) {
        cudaEventElapsedTime( &etime, start, stop );
        up_time += (etime/1000.0);
    }
    magma_event_destroy( start );
    magma_event_destroy( stop  );
    #endif

    trace_finalize( "chetrd.svg", "trace.css" );
    
    #ifdef PROFILE_SY2RK
    printf( " n=%d nb=%d\n", n, nb );
    printf( " Time in CLARFG: %.2e seconds\n", times[0] );
    //printf( " Time in CHEMV : %.2e seconds\n", mv_time );
    printf( " Time in CHER2K: %.2e seconds\n", up_time );
    #endif
    
CLEANUP:
    for( dev=0; dev < ngpu; dev++ ) {
        magma_setdevice( dev );
        for( kk=0; kk < nqueue; kk++ ) {
            magma_queue_destroy( queues[dev][kk] );
        }
        magma_free( dA[dev] );
        magma_free( dwork2[dev] );
    }
    magma_free_pinned( hwork );
    
    magma_setdevice( orig_dev );
    
    work[0] = magma_cmake_lwork( lwkopt );
    
    return *info;
} /* magma_chetrd */