extern "C" magma_int_t
magma_dmshrink(
magma_d_matrix A,
magma_d_matrix *B,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_d_matrix hA={Magma_CSR}, hACSR={Magma_CSR}, hB={Magma_CSR}, hBCSR={Magma_CSR};
if( A.num_rows<=A.num_cols){
if( A.memory_location == Magma_CPU && A.storage_type == Magma_CSR ){
CHECK( magma_dmconvert( A, B, Magma_CSR, Magma_CSR, queue ));
for(magma_int_t i=0; i<A.nnz; i++){
if( B->col[i] >= A.num_rows ){
B->val[i] = MAGMA_D_ZERO;
}
}
CHECK( magma_dmcsrcompressor( B, queue ) );
B->num_cols = B->num_rows;
} else {
CHECK( magma_dmtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
CHECK( magma_dmconvert( hA, &hACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_dmshrink( hACSR, &hBCSR, queue ));
CHECK( magma_dmconvert( hBCSR, &hB, Magma_CSR, A.storage_type, queue ));
CHECK( magma_dmtransfer( hB, B, Magma_CPU, A.memory_location, queue ));
}
} else {
printf("%% error: A has too many rows: m > n.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
cleanup:
magma_dmfree( &hA, queue );
magma_dmfree( &hB, queue );
magma_dmfree( &hACSR, queue );
magma_dmfree( &hBCSR, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_dparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_d_sparse_matrix Z, A, AT, A2, B, B_d;
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_dm_5stencil( laplace_size, &Z, queue );
} else { // file-matrix test
magma_d_csr_mtx( &Z, argv[i], queue );
}
printf( "# matrix info: %d-by-%d with %d nonzeros\n",
(int) Z.num_rows,(int) Z.num_cols,(int) Z.nnz );
// scale matrix
magma_dmscale( &Z, zopts.scaling, queue );
// remove nonzeros in matrix
magma_dmcsrcompressor( &Z, queue );
// convert to be non-symmetric
magma_d_mconvert( Z, &A, Magma_CSR, Magma_CSRL, queue );
// transpose
magma_d_mtranspose( A, &AT, queue );
// convert, copy back and forth to check everything works
magma_d_mconvert( AT, &B, Magma_CSR, zopts.output_format, queue );
magma_d_mfree(&AT, queue );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&B, queue );
magma_dmcsrcompressor_gpu( &B_d, queue );
magma_d_mtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue );
magma_d_mfree(&B_d, queue );
magma_d_mconvert( B, &AT, zopts.output_format,Magma_CSR, queue );
magma_d_mfree(&B, queue );
// transpose back
magma_d_mtranspose( AT, &A2, queue );
magma_d_mfree(&AT, queue );
magma_dmdiff( A, A2, &res, queue);
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester: ok\n");
else
printf("# tester: failed\n");
magma_d_mfree(&A, queue );
magma_d_mfree(&A2, queue );
magma_d_mfree(&Z, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
magma_int_t
magma_dm_5stencil(
magma_int_t n,
magma_d_matrix *A,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i,j,k;
magma_d_matrix hA={Magma_CSR};
// generate matrix of desired structure and size (2d 5-point stencil)
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset=NULL;
double *diag_vals=NULL;
CHECK( magma_dmalloc_cpu( &diag_vals, offdiags+1 ));
CHECK( magma_index_malloc_cpu( &diag_offset, offdiags+1 ));
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
#define REAL
#ifdef COMPLEX
// real case
diag_vals[0] = MAGMA_D_MAKE( 4.0, 4.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, -1.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, -1.0 );
#else
// real case
diag_vals[0] = MAGMA_D_MAKE( 4.0, 0.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, 0.0 );
#endif
CHECK( magma_dmgenerator( nn, offdiags, diag_offset, diag_vals, &hA, queue ));
// now set some entries to zero (boundary...)
for( i=0; i<n; i++ ) {
for( j=0; j<n; j++ ) {
magma_index_t row = i*n+j;
for( k=hA.row[row]; k<hA.row[row+1]; k++) {
if ((hA.col[k] == row-1 ) && (row+1)%n == 1 )
hA.val[k] = MAGMA_D_MAKE( 0.0, 0.0 );
if ((hA.col[k] == row+1 ) && (row)%n ==n-1 )
hA.val[k] = MAGMA_D_MAKE( 0.0, 0.0 );
}
}
}
CHECK( magma_dmconvert( hA, A, Magma_CSR, Magma_CSR, queue ));
magma_dmcsrcompressor( A, queue );
A->true_nnz = A->nnz;
cleanup:
magma_free_cpu( diag_vals );
magma_free_cpu( diag_offset );
magma_dmfree( &hA, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_d_matrix A={Magma_CSR}, AT={Magma_CSR}, A2={Magma_CSR},
B={Magma_CSR}, B_d={Magma_CSR};
int i=1;
real_Double_t start, end;
CHECK( magma_dparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_dm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_d_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
// scale matrix
CHECK( magma_dmscale( &A, zopts.scaling, queue ));
// remove nonzeros in matrix
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_dmcsrcompressor( &A, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA CPU: %.2e seconds.\n", (end-start)/10 );
// transpose
CHECK( magma_dmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_dmconvert( AT, &B, Magma_CSR, Magma_CSR, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&B, queue );
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_dmcsrcompressor_gpu( &B_d, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA GPU: %.2e seconds.\n", (end-start)/10 );
CHECK( magma_dmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_dmfree(&B_d, queue );
CHECK( magma_dmconvert( B, &AT, Magma_CSR, Magma_CSR, queue ));
magma_dmfree(&B, queue );
// transpose back
CHECK( magma_dmtranspose( AT, &A2, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmdiff( A, A2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix compressor: ok\n");
else
printf("%% tester matrix compressor: failed\n");
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
i++;
}
cleanup:
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_d_matrix Z={Magma_CSR}, A={Magma_CSR}, AT={Magma_CSR},
A2={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_index_t *comm_i=NULL;
double *comm_v=NULL;
magma_int_t start, end;
int i=1;
CHECK( magma_dparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_dm_5stencil( laplace_size, &Z, queue ));
} else { // file-matrix test
CHECK( magma_d_csr_mtx( &Z, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(Z.num_rows), int(Z.num_cols), int(Z.nnz) );
// slice matrix
CHECK( magma_index_malloc_cpu( &comm_i, Z.num_rows ) );
CHECK( magma_dmalloc_cpu( &comm_v, Z.num_rows ) );
CHECK( magma_dmslice( 1, 0, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 0, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 1, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 8, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
// scale matrix
CHECK( magma_dmscale( &Z, zopts.scaling, queue ));
// remove nonzeros in matrix
CHECK( magma_dmcsrcompressor( &Z, queue ));
// convert to be non-symmetric
CHECK( magma_dmconvert( Z, &A, Magma_CSR, Magma_CSRL, queue ));
// transpose
CHECK( magma_dmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_dmconvert( AT, &B, Magma_CSR, zopts.output_format, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&B, queue );
CHECK( magma_dmcsrcompressor_gpu( &B_d, queue ));
CHECK( magma_dmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_dmfree(&B_d, queue );
CHECK( magma_dmconvert( B, &AT, zopts.output_format,Magma_CSR, queue ));
magma_dmfree(&B, queue );
// transpose back
CHECK( magma_dmtranspose( AT, &A2, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmdiff( A, A2, &res, queue));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester: ok\n");
else
printf("%% tester: failed\n");
magma_free_cpu( comm_i );
magma_free_cpu( comm_v );
comm_i=NULL;
comm_v=NULL;
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&Z, queue );
i++;
}
cleanup:
magma_free_cpu( comm_i );
magma_free_cpu( comm_v );
magma_dmfree(&AT, queue );
magma_dmfree(&A, queue );
magma_dmfree(&B, queue );
magma_dmfree(&B_d, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&Z, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}