Ejemplo n.º 1
0
//
//      m, n - dimensions in the source (input) matrix.
//             This routine copies the hA matrix from the CPU
//             to dAT on the GPU. In addition, the output matrix
//             is transposed. The routine uses a buffer of size
//             2*lddwork*nb pointed to by dwork (lddwork > m) on the GPU. 
//             Note that lda >= m and lddat >= n.
//
extern "C" void 
magmablas_zsetmatrix_transpose_q(
    magma_int_t m, magma_int_t n,
    const magmaDoubleComplex     *hA, magma_int_t lda, 
    magmaDoubleComplex_ptr       dAT, magma_int_t ldda,
    magmaDoubleComplex_ptr     dwork, magma_int_t lddwork, magma_int_t nb,
    magma_queue_t queues[2] )
{
#define    hA(i_, j_)    (hA + (i_) + (j_)*lda)
#define   dAT(i_, j_)   (dAT + (i_) + (j_)*ldda)
#define dwork(i_, j_) (dwork + (i_) + (j_)*lddwork)

    magma_int_t i = 0, j = 0, ib;

    /* Quick return */
    if ( (m == 0) || (n == 0) )
        return;

    // TODO standard check arguments
    if (lda < m || ldda < n || lddwork < m) {
        fprintf( stderr, "%s: wrong arguments.\n", __func__ );
        return;
    }

    /* Move data from CPU to GPU in the first panel in the dwork buffer */
    ib = min(n-i, nb);
    magma_zsetmatrix_async( m, ib,
                            hA(0,i), lda,
                            dwork(0,(j%2)*nb), lddwork, queues[j%2] );
    j++;

    for (i=nb; i < n; i += nb) {
        /* Move data from CPU to GPU in the second panel in the dwork buffer */
        ib = min(n-i, nb);
        magma_zsetmatrix_async( m, ib,
                                hA(0,i), lda,
                                dwork(0,(j%2)*nb), lddwork, queues[j%2] );
        j++;
        
        /* Note that the previous panel (i.e., j%2) comes through the queue
           for the kernel so there is no need to synchronize.             */
        // TODO should this be ib not nb?
        magmablas_ztranspose_q( m, nb, dwork(0,(j%2)*nb), lddwork, dAT(i-nb,0), ldda, queues[j%2] );
    }

    /* Transpose the last part of the matrix.                            */
    j++;
    magmablas_ztranspose_q( m, ib, dwork(0,(j%2)*nb), lddwork, dAT(i-nb,0), ldda, queues[j%2] );
}
Ejemplo n.º 2
0
/**
    @see magmablas_ztranspose_q
    @ingroup magma_zaux2
    ********************************************************************/
extern "C" void
magmablas_ztranspose(
    magma_int_t m, magma_int_t n,
    magmaDoubleComplex_const_ptr dA,  magma_int_t ldda,
    magmaDoubleComplex_ptr       dAT, magma_int_t lddat )
{
    magmablas_ztranspose_q( m, n, dA, ldda, dAT, lddat, magmablasGetQueue() );
}