extern "C" magma_int_t
magma_smshrink(
magma_s_matrix A,
magma_s_matrix *B,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_s_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_smconvert( 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_S_ZERO;
}
}
CHECK( magma_smcsrcompressor( B, queue ) );
B->num_cols = B->num_rows;
} else {
CHECK( magma_smtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smconvert( hA, &hACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_smshrink( hACSR, &hBCSR, queue ));
CHECK( magma_smconvert( hBCSR, &hB, Magma_CSR, A.storage_type, queue ));
CHECK( magma_smtransfer( 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_smfree( &hA, queue );
magma_smfree( &hB, queue );
magma_smfree( &hACSR, queue );
magma_smfree( &hBCSR, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_sopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
real_Double_t res;
magma_s_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_sparse_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_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_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_smscale( &A, zopts.scaling, queue ));
// remove nonzeros in matrix
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_smcsrcompressor( &A, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA CPU: %.2e seconds.\n", (end-start)/10 );
// transpose
CHECK( magma_smtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_smconvert( AT, &B, Magma_CSR, Magma_CSR, queue ));
magma_smfree(&AT, queue );
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_smfree(&B, queue );
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_smcsrcompressor_gpu( &B_d, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA GPU: %.2e seconds.\n", (end-start)/10 );
CHECK( magma_smtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_smfree(&B_d, queue );
CHECK( magma_smconvert( B, &AT, Magma_CSR, Magma_CSR, queue ));
magma_smfree(&B, queue );
// transpose back
CHECK( magma_smtranspose( AT, &A2, queue ));
magma_smfree(&AT, queue );
CHECK( magma_smdiff( 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_smfree(&A, queue );
magma_smfree(&A2, queue );
i++;
}
cleanup:
magma_smfree(&AT, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&A2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_sopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_sparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_s_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_sm_5stencil( laplace_size, &Z, queue );
} else { // file-matrix test
magma_s_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_smscale( &Z, zopts.scaling, queue );
// remove nonzeros in matrix
magma_smcsrcompressor( &Z, queue );
// convert to be non-symmetric
magma_s_mconvert( Z, &A, Magma_CSR, Magma_CSRL, queue );
// transpose
magma_s_mtranspose( A, &AT, queue );
// convert, copy back and forth to check everything works
printf("here0\n");
magma_s_mconvert( AT, &B, Magma_CSR, zopts.output_format, queue );
magma_s_mfree(&AT, queue );
magma_s_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
magma_s_mfree(&B, queue );
magma_smcsrcompressor_gpu( &B_d, queue );
magma_s_mtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue );
magma_s_mfree(&B_d, queue );
magma_s_mconvert( B, &AT, zopts.output_format,Magma_CSR, queue );
magma_s_mfree(&B, queue );
// transpose back
magma_s_mtranspose( AT, &A2, queue );
magma_s_mfree(&AT, queue );
magma_smdiff( 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_s_mfree(&A, queue );
magma_s_mfree(&A2, queue );
magma_s_mfree(&Z, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
