// ------------------------------------------------------------
// Solve dA * dX = dB, where dA and dX are stored in GPU device memory.
// Internally, MAGMA uses a hybrid CPU + GPU algorithm.
void gpu_interface( magma_int_t n, magma_int_t nrhs )
{
magmaDoubleComplex *dA=NULL, *dX=NULL;
magma_int_t *ipiv=NULL;
magma_int_t ldda = magma_roundup( n, 32 ); // round up to multiple of 32 for best GPU performance
magma_int_t lddx = ldda;
magma_int_t info = 0;
magma_queue_t queue=NULL;
// magma_*malloc routines for GPU memory are type-safe,
// but you can use cudaMalloc if you prefer.
magma_zmalloc( &dA, ldda*n );
magma_zmalloc( &dX, lddx*nrhs );
magma_imalloc_cpu( &ipiv, n ); // ipiv always on CPU
if ( dA == NULL || dX == NULL || ipiv == NULL ) {
fprintf( stderr, "malloc failed\n" );
goto cleanup;
}
magma_int_t dev = 0;
magma_queue_create( dev, &queue );
// Replace these with your code to initialize A and X
zfill_matrix_gpu( n, n, dA, ldda, queue );
zfill_rhs_gpu( n, nrhs, dX, lddx, queue );
magma_zgesv_gpu( n, 1, dA, ldda, ipiv, dX, ldda, &info );
if ( info != 0 ) {
fprintf( stderr, "magma_zgesv_gpu failed with info=%d\n", info );
}
// TODO: use result in dX
cleanup:
magma_queue_destroy( queue );
magma_free( dA );
magma_free( dX );
magma_free_cpu( ipiv );
}
// ------------------------------------------------------------
// Replace with your code to initialize the dX rhs on the GPU device.
void zfill_rhs_gpu(
magma_int_t m, magma_int_t nrhs, magmaDoubleComplex *dX, magma_int_t lddx )
{
zfill_matrix_gpu( m, nrhs, dX, lddx );
}