/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magma_c_matrix Z={Magma_CSR};
int i=1;
CHECK( magma_cparse_opts( argc, argv, &zopts, &i, queue ));
printf("matrixinfo = [ \n");
printf("%% size (n) || nonzeros (nnz) || nnz/n \n");
printf("%%=============================================================%%\n");
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_cm_5stencil( laplace_size, &Z, queue ));
} else { // file-matrix test
CHECK( magma_c_csr_mtx( &Z, argv[i], queue ));
}
printf(" %10d %10d %10d\n",
int(Z.num_rows), int(Z.nnz), int(Z.nnz/Z.num_rows) );
magma_cmfree(&Z, queue );
i++;
}
printf("%%=============================================================%%\n");
printf("];\n");
cleanup:
magma_cmfree(&Z, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_cparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_c_sparse_matrix Z, Z2, A, A2, AT, AT2, B;
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_cm_5stencil( laplace_size, &Z, queue );
} else { // file-matrix test
magma_c_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 );
// convert to be non-symmetric
magma_c_mconvert( Z, &A, Magma_CSR, Magma_CSRL, queue );
magma_c_mconvert( Z, &B, Magma_CSR, Magma_CSRU, queue );
// transpose
magma_c_mtranspose( A, &AT, queue );
// quite some conversions
//ELL
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_ELL, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_ELL, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//ELLPACKT
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_ELLPACKT, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_ELLPACKT, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//ELLRT
AT2.blocksize = 8;
AT2.alignment = 8;
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_ELLRT, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_ELLRT, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//SELLP
AT2.blocksize = 8;
AT2.alignment = 8;
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_SELLP, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_SELLP, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//ELLD
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_ELLD, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_ELLD, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//CSRCOO
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_CSRCOO, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_CSRCOO, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//CSRD
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_CSRD, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_CSRD, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
//BCSR
magma_c_mconvert( AT, &AT2, Magma_CSR, Magma_BCSR, queue );
magma_c_mfree(&AT, queue );
magma_c_mconvert( AT2, &AT, Magma_BCSR, Magma_CSR, queue );
magma_c_mfree(&AT2, queue );
// transpose
magma_c_mtranspose( AT, &A2, queue );
magma_cmdiff( A, A2, &res, queue);
printf("# ||A-A2||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# conversion tester: ok\n");
else
printf("# conversion tester: failed\n");
magma_cmlumerge( A2, B, &Z2, queue );
magma_cmdiff( Z, Z2, &res, queue);
printf("# ||Z-Z2||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# LUmerge tester: ok\n");
else
printf("# LUmerge tester: failed\n");
magma_c_mfree(&A, queue );
magma_c_mfree(&A2, queue );
magma_c_mfree(&AT, queue );
magma_c_mfree(&AT2, queue );
magma_c_mfree(&B, queue );
magma_c_mfree(&Z2, queue );
magma_c_mfree(&Z, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_cparse_opts( argc, argv, &zopts, &i, queue );
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_sparse_matrix A, B, B_d;
magma_c_vector x, b;
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
if ( zopts.solver_par.solver != Magma_PCG &&
zopts.solver_par.solver != Magma_PGMRES &&
zopts.solver_par.solver != Magma_PBICGSTAB &&
zopts.solver_par.solver != Magma_ITERREF )
zopts.precond_par.solver = Magma_NONE;
magma_csolverinfo_init( &zopts.solver_par, &zopts.precond_par, queue );
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_cm_5stencil( laplace_size, &A, queue );
} else { // file-matrix test
magma_c_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 );
// for the eigensolver case
zopts.solver_par.ev_length = A.num_rows;
magma_ceigensolverinfo_init( &zopts.solver_par, queue );
// scale matrix
magma_cmscale( &A, zopts.scaling, queue );
magma_c_mconvert( A, &B, Magma_CSR, zopts.output_format, queue );
magma_c_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
// vectors and initial guess
magma_c_vinit( &b, Magma_DEV, A.num_cols, one, queue );
magma_c_vinit( &x, Magma_DEV, A.num_cols, one, queue );
magma_c_spmv( one, B_d, x, zero, b, queue ); // b = A x
magma_c_vfree(&x, queue );
magma_c_vinit( &x, Magma_DEV, A.num_cols, zero, queue );
magma_c_solver( B_d, b, &x, &zopts, queue );
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
magma_c_mfree(&B_d, queue );
magma_c_mfree(&B, queue );
magma_c_mfree(&A, queue );
magma_c_vfree(&x, queue );
magma_c_vfree(&b, queue );
i++;
}
magma_csolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue=NULL;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
real_Double_t res;
magma_c_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_cparse_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_cm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_c_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_cmscale( &A, zopts.scaling, queue ));
// remove nonzeros in matrix
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_cmcsrcompressor( &A, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA CPU: %.2e seconds.\n", (end-start)/10 );
// transpose
CHECK( magma_cmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_cmconvert( AT, &B, Magma_CSR, Magma_CSR, queue ));
magma_cmfree(&AT, queue );
CHECK( magma_cmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&B, queue );
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_cmcsrcompressor_gpu( &B_d, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA GPU: %.2e seconds.\n", (end-start)/10 );
CHECK( magma_cmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_cmfree(&B_d, queue );
CHECK( magma_cmconvert( B, &AT, Magma_CSR, Magma_CSR, queue ));
magma_cmfree(&B, queue );
// transpose back
CHECK( magma_cmtranspose( AT, &A2, queue ));
magma_cmfree(&AT, queue );
CHECK( magma_cmdiff( 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_cmfree(&A, queue );
magma_cmfree(&A2, queue );
i++;
}
cleanup:
magma_cmfree(&AT, queue );
magma_cmfree(&B, queue );
magma_cmfree(&A, queue );
magma_cmfree(&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_CHECK( magma_init() );
magma_print_environment();
magma_copts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_c_matrix A={Magma_CSR}, A2={Magma_CSR},
A3={Magma_CSR}, A4={Magma_CSR}, A5={Magma_CSR};
int i=1;
TESTING_CHECK( magma_cparse_opts( argc, argv, &zopts, &i, queue ));
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
TESTING_CHECK( magma_cm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_c_csr_mtx( &A, argv[i], queue ));
}
printf("%% matrix info: %lld-by-%lld with %lld nonzeros\n",
(long long) A.num_rows, (long long) A.num_cols, (long long) A.nnz );
// filename for temporary matrix storage
const char *filename = "testmatrix.mtx";
// write to file
TESTING_CHECK( magma_cwrite_csrtomtx( A, filename, queue ));
// read from file
TESTING_CHECK( magma_c_csr_mtx( &A2, filename, queue ));
// delete temporary matrix
unlink( filename );
//visualize
printf("A2:\n");
TESTING_CHECK( magma_cprint_matrix( A2, queue ));
//visualize
TESTING_CHECK( magma_cmconvert(A2, &A4, Magma_CSR, Magma_CSRL, queue ));
printf("A4:\n");
TESTING_CHECK( magma_cprint_matrix( A4, queue ));
TESTING_CHECK( magma_cmconvert(A4, &A5, Magma_CSR, Magma_ELL, queue ));
printf("A5:\n");
TESTING_CHECK( magma_cprint_matrix( A5, queue ));
// pass it to another application and back
magma_int_t m, n;
magma_index_t *row, *col;
magmaFloatComplex *val=NULL;
TESTING_CHECK( magma_ccsrget( A2, &m, &n, &row, &col, &val, queue ));
TESTING_CHECK( magma_ccsrset( m, n, row, col, val, &A3, queue ));
TESTING_CHECK( magma_cmdiff( A, A2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester IO: ok\n");
else
printf("%% tester IO: failed\n");
TESTING_CHECK( magma_cmdiff( A, A3, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix interface: ok\n");
else
printf("%% tester matrix interface: failed\n");
magma_cmfree(&A, queue );
magma_cmfree(&A2, queue );
magma_cmfree(&A4, queue );
magma_cmfree(&A5, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_CHECK( magma_finalize() );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
magma_c_sparse_matrix hA, hA_SELLP, hA_ELL, dA, dA_SELLP, dA_ELL;
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
magma_int_t *pntre;
magmaFloatComplex c_one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex c_zero = MAGMA_C_MAKE(0.0, 0.0);
magma_int_t i, j;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--blocksize", argv[i]) == 0 ) {
hA_SELLP.blocksize = atoi( argv[++i] );
} else if ( strcmp("--alignment", argv[i]) == 0 ) {
hA_SELLP.alignment = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_cspmv"
" [ --blocksize %d --alignment %d (for SELLP) ]"
" matrices \n\n", (int) hA_SELLP.blocksize, (int) hA_SELLP.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_cm_5stencil( laplace_size, &hA, queue );
} else { // file-matrix test
magma_c_csr_mtx( &hA, argv[i], queue );
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) hA.num_rows,(int) hA.num_cols,(int) hA.nnz );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
magma_c_vector hx, hy, dx, dy, hrefvec, hcheck;
// init CPU vectors
magma_c_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue );
magma_c_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue );
// init DEV vectors
magma_c_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
#ifdef MAGMA_WITH_MKL
// calling MKL with CSR
pntre = (magma_int_t*)malloc( (hA.num_rows+1)*sizeof(magma_int_t) );
pntre[0] = 0;
for (j=0; j<hA.num_rows; j++ ) {
pntre[j] = hA.row[j+1];
}
MKL_INT num_rows = hA.num_rows;
MKL_INT num_cols = hA.num_cols;
MKL_INT nnz = hA.nnz;
MKL_INT *col;
TESTING_MALLOC_CPU( col, MKL_INT, nnz );
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_MALLOC_CPU( row, MKL_INT, num_rows );
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
start = magma_wtime();
for (j=0; j<10; j++ ) {
mkl_ccsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val),
col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
}
end = magma_wtime();
printf( "\n > MKL : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
TESTING_FREE_CPU( row );
TESTING_FREE_CPU( col );
free(pntre);
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
magma_c_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
// SpMV on GPU (CSR) -- this is the reference!
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_c_spmv( c_one, dA, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_c_mfree(&dA, queue );
magma_c_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue );
// convert to ELL and copy to GPU
magma_c_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue );
magma_c_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue );
magma_c_mfree(&hA_ELL, queue );
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (ELL)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_c_spmv( c_one, dA_ELL, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard ELL).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_c_mfree(&dA_ELL, queue );
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
if ( res < .000001 )
printf("# tester spmv ELL: ok\n");
else
printf("# tester spmv ELL: failed\n");
magma_c_vfree( &hcheck, queue );
// convert to SELLP and copy to GPU
magma_c_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue );
magma_c_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue );
magma_c_mfree(&hA_SELLP, queue );
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_c_spmv( c_one, dA_SELLP, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (SELLP).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv SELL-P: ok\n");
else
printf("# tester spmv SELL-P: failed\n");
magma_c_vfree( &hcheck, queue );
magma_c_mfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
magmaFloatComplex alpha = c_one;
magmaFloatComplex beta = c_zero;
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// copy matrix to GPU
magma_c_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseCcsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
hA.num_rows, hA.num_cols, hA.nnz, &alpha, descr,
dA.dval, dA.drow, dA.dcol, dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE CSR\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
cusparseMatDescr_t descrA;
cusparseStatus = cusparseCreateMatDescr(&descrA);
if (cusparseStatus != 0) printf("error\n");
cusparseHybMat_t hybA;
cusparseStatus = cusparseCreateHybMat( &hybA );
if (cusparseStatus != 0) printf("error\n");
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE CSR: ok\n");
else
printf("# tester spmv cuSPARSE CSR: failed\n");
magma_c_vfree( &hcheck, queue );
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
cusparseCcsr2hyb(cusparseHandle, hA.num_rows, hA.num_cols,
descrA, dA.dval, dA.drow, dA.dcol,
hybA, 0, CUSPARSE_HYB_PARTITION_AUTO);
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseChybmv( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
&alpha, descrA, hybA,
dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE HYB\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (HYB).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE HYB: ok\n");
else
printf("# tester spmv cuSPARSE HYB: failed\n");
magma_c_vfree( &hcheck, queue );
cusparseDestroyMatDescr( descrA );
cusparseDestroyHybMat( hybA );
cusparseDestroy( cusparseHandle );
magma_c_mfree(&dA, queue );
printf("\n\n");
// free CPU memory
magma_c_mfree(&hA, queue );
magma_c_vfree(&hx, queue );
magma_c_vfree(&hy, queue );
magma_c_vfree(&hrefvec, queue );
// free GPU memory
magma_c_vfree(&dx, queue );
magma_c_vfree(&dy, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_CHECK( magma_init() );
magma_print_environment();
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_matrix A={Magma_CSR}, B_d={Magma_CSR};
magma_c_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
TESTING_CHECK( magma_cm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_c_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %lld-by-%lld with %lld nonzeros\n\n",
(long long) A.num_rows, (long long) A.num_cols, (long long) A.nnz );
magma_int_t n = A.num_rows;
TESTING_CHECK( magma_cmtransfer( A, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
TESTING_CHECK( magma_cvinit( &b, Magma_DEV, A.num_cols, 1, zero, queue ));
TESTING_CHECK( magma_cvinit( &x, Magma_DEV, A.num_cols, 1, one, queue ));
TESTING_CHECK( magma_cprint_vector( b, 90, 10, queue ));
TESTING_CHECK( magma_cprint_matrix( A, queue ));
printf("\n\n\n");
TESTING_CHECK( magma_cprint_matrix( B_d, queue ));
float res;
res = magma_scnrm2( n, b.dval, 1, queue );
printf("norm0: %f\n", res);
TESTING_CHECK( magma_c_spmv( one, B_d, x, zero, b, queue )); // b = A x
TESTING_CHECK( magma_cprint_vector( b, 0, 100, queue ));
TESTING_CHECK( magma_cprint_vector( b, b.num_rows-10, 10, queue ));
res = magma_scnrm2( n, b.dval, 1, queue );
printf("norm: %f\n", res);
TESTING_CHECK( magma_cresidual( B_d, x, b, &res, queue ));
printf("res: %f\n", res);
magma_cmfree(&B_d, queue );
magma_cmfree(&A, queue );
magma_cmfree(&x, queue );
magma_cmfree(&b, queue );
i++;
}
magma_queue_destroy( queue );
magma_finalize();
return info;
}
magma_int_t
magma_ccustomicsetup(
magma_c_matrix A,
magma_c_matrix b,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_c_matrix hA={Magma_CSR};
char preconditionermatrix[255];
snprintf( preconditionermatrix, sizeof(preconditionermatrix),
"/Users/hanzt0114cl306/work/matrices/ani/ani7_crop_ichol.mtx" );
CHECK( magma_c_csr_mtx( &hA, preconditionermatrix , queue) );
// for CUSPARSE
CHECK( magma_cmtransfer( hA, &precond->M, Magma_CPU, Magma_DEV , queue ));
// copy the matrix to precond->L and (transposed) to precond->U
CHECK( magma_cmtransfer(precond->M, &(precond->L), Magma_DEV, Magma_DEV, queue ));
CHECK( magma_cmtranspose( precond->L, &(precond->U), queue ));
// extract the diagonal of L into precond->d
CHECK( magma_cjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_cvinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_cjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_cvinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_C_ZERO, queue ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCcsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrL,
precond->M.val, precond->M.row, precond->M.col, precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseCcsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrU,
precond->M.val, precond->M.row, precond->M.col, precond->cuinfoU ));
cleanup:
cusparseDestroy( cusparseHandle );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseHandle=NULL;
descrL=NULL;
descrU=NULL;
magma_cmfree( &hA, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_copts zopts;
int i=1;
magma_cparse_opts( argc, argv, &zopts, &i);
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_sparse_matrix A, B, B_d;
magma_c_vector x, b;
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
if ( zopts.solver_par.solver != Magma_PCG &&
zopts.solver_par.solver != Magma_PGMRES &&
zopts.solver_par.solver != Magma_PBICGSTAB &&
zopts.solver_par.solver != Magma_ITERREF )
zopts.precond_par.solver = Magma_NONE;
magma_csolverinfo_init( &zopts.solver_par, &zopts.precond_par );
while( i < argc ){
magma_c_csr_mtx( &A, argv[i] );
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
magma_cmscale( &A, zopts.scaling );
magma_c_mconvert( A, &B, Magma_CSR, zopts.output_format );
magma_c_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_c_vinit( &b, Magma_DEV, A.num_cols, one );
magma_c_vinit( &x, Magma_DEV, A.num_cols, one );
magma_c_spmv( one, B_d, x, zero, b ); // b = A x
magma_c_vfree(&x);
magma_c_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_c_solver( B_d, b, &x, &zopts );
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par );
magma_c_mfree(&B_d);
magma_c_mfree(&B);
magma_c_mfree(&A);
magma_c_vfree(&x);
magma_c_vfree(&b);
i++;
}
magma_csolverinfo_free( &zopts.solver_par, &zopts.precond_par );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( &queue );
magma_c_matrix hA={Magma_CSR}, hA_SELLP={Magma_CSR}, hA_ELL={Magma_CSR},
dA={Magma_CSR}, dA_SELLP={Magma_CSR}, dA_ELL={Magma_CSR};
magma_c_matrix hx={Magma_CSR}, hy={Magma_CSR}, dx={Magma_CSR},
dy={Magma_CSR}, hrefvec={Magma_CSR}, hcheck={Magma_CSR};
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
#ifdef MAGMA_WITH_MKL
magma_int_t *pntre=NULL;
#endif
cusparseHandle_t cusparseHandle = NULL;
cusparseMatDescr_t descr = NULL;
magmaFloatComplex c_one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex c_zero = MAGMA_C_MAKE(0.0, 0.0);
float accuracy = 1e-10;
#define PRECISION_c
#if defined(PRECISION_c)
accuracy = 1e-4;
#endif
#if defined(PRECISION_s)
accuracy = 1e-4;
#endif
magma_int_t i, j;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--blocksize", argv[i]) == 0 ) {
hA_SELLP.blocksize = atoi( argv[++i] );
} else if ( strcmp("--alignment", argv[i]) == 0 ) {
hA_SELLP.alignment = atoi( argv[++i] );
} else
break;
}
printf("\n# usage: ./run_cspmm"
" [ --blocksize %d --alignment %d (for SELLP) ]"
" matrices \n\n", int(hA_SELLP.blocksize), int(hA_SELLP.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_cm_5stencil( laplace_size, &hA, queue ));
} else { // file-matrix test
CHECK( magma_c_csr_mtx( &hA, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(hA.num_rows), int(hA.num_cols), int(hA.nnz) );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
// m - number of rows for the sparse matrix
// n - number of vectors to be multiplied in the SpMM product
magma_int_t m, n;
m = hA.num_rows;
n = 48;
// init CPU vectors
CHECK( magma_cvinit( &hx, Magma_CPU, m, n, c_one, queue ));
CHECK( magma_cvinit( &hy, Magma_CPU, m, n, c_zero, queue ));
// init DEV vectors
CHECK( magma_cvinit( &dx, Magma_DEV, m, n, c_one, queue ));
CHECK( magma_cvinit( &dy, Magma_DEV, m, n, c_zero, queue ));
// calling MKL with CSR
#ifdef MAGMA_WITH_MKL
CHECK( magma_imalloc_cpu( &pntre, m + 1 ) );
pntre[0] = 0;
for (j=0; j < m; j++ ) {
pntre[j] = hA.row[j+1];
}
MKL_INT num_rows = hA.num_rows;
MKL_INT num_cols = hA.num_cols;
MKL_INT nnz = hA.nnz;
MKL_INT num_vecs = n;
MKL_INT *col;
TESTING_MALLOC_CPU( col, MKL_INT, nnz );
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_MALLOC_CPU( row, MKL_INT, num_rows );
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
// === Call MKL with consecutive SpMVs, using mkl_ccsrmv ===
// warmp up
mkl_ccsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
start = magma_wtime();
for (j=0; j<10; j++ )
mkl_ccsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
end = magma_wtime();
printf( "\n > MKL SpMVs : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
// === Call MKL with blocked SpMVs, using mkl_ccsrmm ===
char transa = 'n';
MKL_INT ldb = n, ldc=n;
char matdescra[6] = {'g', 'l', 'n', 'c', 'x', 'x'};
// warm up
mkl_ccsrmm( &transa, &num_rows, &num_vecs, &num_cols, MKL_ADDR(&c_one), matdescra,
MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val), &ldb,
MKL_ADDR(&c_zero),
MKL_ADDR(hy.val), &ldc );
start = magma_wtime();
for (j=0; j<10; j++ )
mkl_ccsrmm( &transa, &num_rows, &num_vecs, &num_cols, MKL_ADDR(&c_one), matdescra,
MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val), &ldb,
MKL_ADDR(&c_zero),
MKL_ADDR(hy.val), &ldc );
end = magma_wtime();
printf( "\n > MKL SpMM : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
TESTING_FREE_CPU( row );
TESTING_FREE_CPU( col );
row = NULL;
col = NULL;
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
CHECK( magma_cmtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue ));
// SpMV on GPU (CSR)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
CHECK( magma_c_spmv( c_one, dA, dx, c_zero, dy, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard CSR).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
CHECK( magma_cmtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue ));
magma_cmfree(&dA, queue );
// convert to SELLP and copy to GPU
CHECK( magma_cmconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue ));
CHECK( magma_cmtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&hA_SELLP, queue );
magma_cmfree( &dy, queue );
CHECK( magma_cvinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
CHECK( magma_c_spmv( c_one, dA_SELLP, dx, c_zero, dy, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (SELLP).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
CHECK( magma_cmtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("%% |x-y|_F = %8.2e\n", res);
if ( res < accuracy )
printf("%% tester spmm SELL-P: ok\n");
else
printf("%% tester spmm SELL-P: failed\n");
magma_cmfree( &hcheck, queue );
magma_cmfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
magma_cmfree( &dy, queue );
CHECK( magma_cvinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
//#ifdef PRECISION_d
start = magma_sync_wtime( queue );
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseSetStream( cusparseHandle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descr ));
CHECK_CUSPARSE( cusparseSetMatType( descr, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descr, CUSPARSE_INDEX_BASE_ZERO ));
magmaFloatComplex alpha = c_one;
magmaFloatComplex beta = c_zero;
// copy matrix to GPU
CHECK( magma_cmtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue) );
for (j=0; j<10; j++)
cusparseCcsrmm(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
dA.num_rows, n, dA.num_cols, dA.nnz,
&alpha, descr, dA.dval, dA.drow, dA.dcol,
dx.dval, dA.num_cols, &beta, dy.dval, dA.num_cols);
end = magma_sync_wtime( queue );
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10*n/(end-start) );
CHECK( magma_cmtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("%% |x-y|_F = %8.2e\n", res);
if ( res < accuracy )
printf("%% tester spmm cuSPARSE: ok\n");
else
printf("%% tester spmm cuSPARSE: failed\n");
magma_cmfree( &hcheck, queue );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
descr = NULL;
cusparseHandle = NULL;
//#endif
printf("\n\n");
// free CPU memory
magma_cmfree(&hA, queue );
magma_cmfree(&hx, queue );
magma_cmfree(&hy, queue );
magma_cmfree(&hrefvec, queue );
// free GPU memory
magma_cmfree(&dx, queue );
magma_cmfree(&dy, queue );
magma_cmfree(&dA, queue);
i++;
}
cleanup:
#ifdef MAGMA_WITH_MKL
magma_free_cpu(pntre);
#endif
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
magma_cmfree(&hA, queue );
magma_cmfree(&dA, queue );
magma_cmfree(&hA_ELL, queue );
magma_cmfree(&dA_ELL, queue );
magma_cmfree(&hA_SELLP, queue );
magma_cmfree(&dA_SELLP, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_CHECK( magma_init() );
magma_print_environment();
magma_copts zopts;
magma_queue_t queue;
magma_queue_create( 0, &queue );
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_c_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
TESTING_CHECK( magma_cparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
TESTING_CHECK( magma_csolverinfo_init( &zopts.solver_par, &zopts.precond_par, queue ));
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
TESTING_CHECK( magma_cm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_c_csr_mtx( &A, argv[i], queue ));
}
// for the eigensolver case
zopts.solver_par.ev_length = A.num_cols;
TESTING_CHECK( magma_ceigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
TESTING_CHECK( magma_cmscale( &A, zopts.scaling, queue ));
// preconditioner
if ( zopts.solver_par.solver != Magma_ITERREF ) {
TESTING_CHECK( magma_c_precondsetup( A, b, &zopts.solver_par, &zopts.precond_par, queue ) );
}
TESTING_CHECK( magma_cmconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
printf( "\n%% matrix info: %lld-by-%lld with %lld nonzeros\n\n",
(long long) A.num_rows, (long long) A.num_cols, (long long) A.nnz );
printf("matrixinfo = [\n");
printf("%% size (m x n) || nonzeros (nnz) || nnz/m || stored nnz\n");
printf("%%============================================================================%%\n");
printf(" %8lld %8lld %10lld %4lld %10lld\n",
(long long) B.num_rows, (long long) B.num_cols, (long long) B.true_nnz,
(long long) (B.true_nnz/B.num_rows), (long long) B.nnz );
printf("%%============================================================================%%\n");
printf("];\n");
TESTING_CHECK( magma_cmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
TESTING_CHECK( magma_cvinit( &b, Magma_DEV, A.num_rows, 1, one, queue ));
//magma_cvinit( &x, Magma_DEV, A.num_cols, 1, one, queue );
//magma_c_spmv( one, B_d, x, zero, b, queue ); // b = A x
//magma_cmfree(&x, queue );
TESTING_CHECK( magma_cvinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
info = magma_c_solver( B_d, b, &x, &zopts, queue );
if( info != 0 ) {
printf("%%error: solver returned: %s (%lld).\n",
magma_strerror( info ), (long long) info );
}
printf("convergence = [\n");
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
printf("];\n\n");
zopts.solver_par.verbose = 0;
printf("solverinfo = [\n");
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
printf("];\n\n");
printf("precondinfo = [\n");
printf("%% setup runtime\n");
printf(" %.6f %.6f\n",
zopts.precond_par.setuptime, zopts.precond_par.runtime );
printf("];\n\n");
magma_cmfree(&B_d, queue );
magma_cmfree(&B, queue );
magma_cmfree(&A, queue );
magma_cmfree(&x, queue );
magma_cmfree(&b, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_CHECK( magma_finalize() );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_c_matrix x={Magma_CSR}, b={Magma_CSR}, t={Magma_CSR};
magma_c_matrix x1={Magma_CSR}, x2={Magma_CSR};
//Chronometry
real_Double_t tempo1, tempo2;
int i=1;
CHECK( magma_cparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
CHECK( magma_csolverinfo_init( &zopts.solver_par, &zopts.precond_par, queue ));
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_cm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_c_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) );
// for the eigensolver case
zopts.solver_par.ev_length = A.num_rows;
CHECK( magma_ceigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
CHECK( magma_cmscale( &A, zopts.scaling, queue ));
CHECK( magma_cmconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
CHECK( magma_cmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_cvinit( &b, Magma_DEV, A.num_cols, 1, one, queue ));
CHECK( magma_cvinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_cvinit( &t, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_cvinit( &x1, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_cvinit( &x2, Magma_DEV, A.num_cols, 1, zero, queue ));
//preconditioner
CHECK( magma_c_precondsetup( B_d, b, &zopts.solver_par, &zopts.precond_par, queue ) );
float residual;
CHECK( magma_cresidual( B_d, b, x, &residual, queue ));
zopts.solver_par.init_res = residual;
printf("data = [\n");
printf("%%runtime left preconditioner:\n");
tempo1 = magma_sync_wtime( queue );
info = magma_c_applyprecond_left( MagmaNoTrans, B_d, b, &x1, &zopts.precond_par, queue );
tempo2 = magma_sync_wtime( queue );
if( info != 0 ){
printf("error: preconditioner returned: %s (%d).\n",
magma_strerror( info ), int(info) );
}
CHECK( magma_cresidual( B_d, b, x1, &residual, queue ));
printf("%.8e %.8e\n", tempo2-tempo1, residual );
printf("%%runtime right preconditioner:\n");
tempo1 = magma_sync_wtime( queue );
info = magma_c_applyprecond_right( MagmaNoTrans, B_d, b, &x2, &zopts.precond_par, queue );
tempo2 = magma_sync_wtime( queue );
if( info != 0 ){
printf("error: preconditioner returned: %s (%d).\n",
magma_strerror( info ), int(info) );
}
CHECK( magma_cresidual( B_d, b, x2, &residual, queue ));
printf("%.8e %.8e\n", tempo2-tempo1, residual );
printf("];\n");
info = magma_c_applyprecond_left( MagmaNoTrans, B_d, b, &t, &zopts.precond_par, queue );
info = magma_c_applyprecond_right( MagmaNoTrans, B_d, t, &x, &zopts.precond_par, queue );
CHECK( magma_cresidual( B_d, b, x, &residual, queue ));
zopts.solver_par.final_res = residual;
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
magma_cmfree(&B_d, queue );
magma_cmfree(&B, queue );
magma_cmfree(&A, queue );
magma_cmfree(&x, queue );
magma_cmfree(&x1, queue );
magma_cmfree(&x2, queue );
magma_cmfree(&b, queue );
magma_cmfree(&t, queue );
i++;
}
cleanup:
magma_cmfree(&B_d, queue );
magma_cmfree(&B, queue );
magma_cmfree(&A, queue );
magma_cmfree(&x, queue );
magma_cmfree(&x1, queue );
magma_cmfree(&x2, queue );
magma_cmfree(&b, queue );
magma_cmfree(&t, queue );
magma_csolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_clobpcg
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_c_solver_par solver_par;
solver_par.epsilon = 1e-5;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.num_eigenvalues = 32;
solver_par.solver = Magma_LOBPCG;
magma_c_preconditioner precond_par;
precond_par.solver = Magma_JACOBI;
int precond = 0;
int format = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_c_sparse_matrix A, B, dA;
B.blocksize = 8;
B.alignment = 8;
B.storage_type = Magma_CSR;
int i;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--format", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( format ) {
case 0: B.storage_type = Magma_CSR; break;
case 1: B.storage_type = Magma_ELL; break;
case 2: B.storage_type = Magma_ELLRT; break;
case 3: B.storage_type = Magma_SELLP; break;
}
}else if ( strcmp("--mscale", argv[i]) == 0 ) {
scale = atoi( argv[++i] );
switch( scale ) {
case 0: scaling = Magma_NOSCALE; break;
case 1: scaling = Magma_UNITDIAG; break;
case 2: scaling = Magma_UNITROW; break;
}
}else if ( strcmp("--precond", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( precond ) {
case 0: precond_par.solver = Magma_JACOBI; break;
}
}else if ( strcmp("--blocksize", argv[i]) == 0 ) {
B.blocksize = atoi( argv[++i] );
}else if ( strcmp("--alignment", argv[i]) == 0 ) {
B.alignment = atoi( argv[++i] );
}else if ( strcmp("--verbose", argv[i]) == 0 ) {
solver_par.verbose = atoi( argv[++i] );
} else if ( strcmp("--maxiter", argv[i]) == 0 ) {
solver_par.maxiter = atoi( argv[++i] );
} else if ( strcmp("--tol", argv[i]) == 0 ) {
sscanf( argv[++i], "%f", &solver_par.epsilon );
} else if ( strcmp("--eigenvalues", argv[i]) == 0 ) {
solver_par.num_eigenvalues = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_clobpcg"
" [ --format %d (0=CSR, 1=ELL, 2=ELLRT, 4=SELLP)"
" [ --blocksize %d --alignment %d ]"
" --mscale %d (0=no, 1=unitdiag, 2=unitrownrm)"
" --verbose %d (0=summary, k=details every k iterations)"
" --maxiter %d --tol %.2e"
" --preconditioner %d (0=Jacobi) "
" --eigenvalues %d ]"
" matrices \n\n", format, (int) B.blocksize, (int) B.alignment,
(int) scale,
(int) solver_par.verbose,
(int) solver_par.maxiter, solver_par.epsilon, precond,
(int) solver_par.num_eigenvalues);
while( i < argc ){
magma_c_csr_mtx( &A, argv[i] );
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// scale initial guess
magma_cmscale( &A, scaling );
solver_par.ev_length = A.num_cols;
magma_c_sparse_matrix A2;
A2.storage_type = Magma_SELLC;
A2.blocksize = 8;
A2.alignment = 4;
magma_c_mconvert( A, &A2, Magma_CSR, A2.storage_type );
// copy matrix to GPU
magma_c_mtransfer( A2, &dA, Magma_CPU, Magma_DEV);
magma_csolverinfo_init( &solver_par, &precond_par ); // inside the loop!
// as the matrix size has influence on the EV-length
real_Double_t gpu_time;
// Find the blockSize smallest eigenvalues and corresponding eigen-vectors
gpu_time = magma_wtime();
magma_clobpcg( dA, &solver_par );
gpu_time = magma_wtime() - gpu_time;
printf("Time (sec) = %7.2f\n", gpu_time);
printf("solver runtime (sec) = %7.2f\n", solver_par.runtime );
magma_csolverinfo_free( &solver_par, &precond_par );
magma_c_mfree( &dA );
magma_c_mfree( &A2 );
magma_c_mfree( &A );
i++;
}
TESTING_FINALIZE();
return 0;
}