/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_CHECK( magma_init() );
magma_print_environment();
magma_sopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magma_s_matrix Z={Magma_CSR};
int i=1;
TESTING_CHECK( magma_sparse_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] );
TESTING_CHECK( magma_sm_5stencil( laplace_size, &Z, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_s_csr_mtx( &Z, argv[i], queue ));
}
printf(" %10lld %10lld %10lld\n",
(long long) Z.num_rows, (long long) Z.nnz, (long long) (Z.nnz/Z.num_rows) );
magma_smfree(&Z, queue );
i++;
}
printf("%%=============================================================%%\n");
printf("];\n");
magma_queue_destroy( queue );
TESTING_CHECK( magma_finalize() );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
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 );
magma_s_matrix hA={Magma_CSR}, hA_SELLP={Magma_CSR},
dA={Magma_CSR}, dA_SELLP={Magma_CSR};
magma_s_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;
float c_one = MAGMA_S_MAKE(1.0, 0.0);
float c_zero = MAGMA_S_MAKE(0.0, 0.0);
float accuracy = 1e-10;
#define PRECISION_s
#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_sspmm"
" [ --blocksize %lld --alignment %lld (for SELLP) ] matrices\n\n",
(long long) hA_SELLP.blocksize, (long long) 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] );
TESTING_CHECK( magma_sm_5stencil( laplace_size, &hA, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_s_csr_mtx( &hA, argv[i], queue ));
}
printf("%% matrix info: %lld-by-%lld with %lld nonzeros\n",
(long long) hA.num_rows, (long long) hA.num_cols, (long long) 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
TESTING_CHECK( magma_svinit( &hx, Magma_CPU, m, n, c_one, queue ));
TESTING_CHECK( magma_svinit( &hy, Magma_CPU, m, n, c_zero, queue ));
// init DEV vectors
TESTING_CHECK( magma_svinit( &dx, Magma_DEV, m, n, c_one, queue ));
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, m, n, c_zero, queue ));
// calling MKL with CSR
#ifdef MAGMA_WITH_MKL
TESTING_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_CHECK( magma_malloc_cpu( (void**) &col, nnz * sizeof(MKL_INT) ));
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_CHECK( magma_malloc_cpu( (void**) &row, num_rows * sizeof(MKL_INT) ));
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
// === Call MKL with consecutive SpMVs, using mkl_scsrmv ===
// warmp up
mkl_scsrmv( "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_scsrmv( "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_scsrmm ===
char transa = 'n';
MKL_INT ldb = n, ldc=n;
char matdescra[6] = {'g', 'l', 'n', 'c', 'x', 'x'};
// warm up
mkl_scsrmm( &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_scsrmm( &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) );
magma_free_cpu( row );
magma_free_cpu( col );
row = NULL;
col = NULL;
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
TESTING_CHECK( magma_smtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue ));
// SpMV on GPU (CSR)
start = magma_sync_wtime( queue );
for (j=0; j < 10; j++) {
TESTING_CHECK( magma_s_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) );
TESTING_CHECK( magma_smtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue ));
magma_smfree(&dA, queue );
// convert to SELLP and copy to GPU
TESTING_CHECK( magma_smconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue ));
TESTING_CHECK( magma_smtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue ));
magma_smfree(&hA_SELLP, queue );
magma_smfree( &dy, queue );
TESTING_CHECK( magma_svinit( &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++) {
TESTING_CHECK( magma_s_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) );
TESTING_CHECK( magma_smtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k < hA.num_rows; k++ ) {
res=res + MAGMA_S_REAL(hcheck.val[k]) - MAGMA_S_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_smfree( &hcheck, queue );
magma_smfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
magma_smfree( &dy, queue );
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
//#ifdef PRECISION_d
start = magma_sync_wtime( queue );
TESTING_CHECK( cusparseCreate( &cusparseHandle ));
TESTING_CHECK( cusparseSetStream( cusparseHandle, magma_queue_get_cuda_stream(queue) ));
TESTING_CHECK( cusparseCreateMatDescr( &descr ));
TESTING_CHECK( cusparseSetMatType( descr, CUSPARSE_MATRIX_TYPE_GENERAL ));
TESTING_CHECK( cusparseSetMatIndexBase( descr, CUSPARSE_INDEX_BASE_ZERO ));
float alpha = c_one;
float beta = c_zero;
// copy matrix to GPU
TESTING_CHECK( magma_smtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue) );
for (j=0; j < 10; j++) {
cusparseScsrmm(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) );
TESTING_CHECK( magma_smtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k < hA.num_rows; k++ ) {
res = res + MAGMA_S_REAL(hcheck.val[k]) - MAGMA_S_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_smfree( &hcheck, queue );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
descr = NULL;
cusparseHandle = NULL;
//#endif
printf("\n\n");
// free CPU memory
magma_smfree( &hA, queue );
magma_smfree( &hx, queue );
magma_smfree( &hy, queue );
magma_smfree( &hrefvec, queue );
// free GPU memory
magma_smfree( &dx, queue );
magma_smfree( &dy, queue );
magma_smfree( &dA, queue);
#ifdef MAGMA_WITH_MKL
magma_free_cpu( pntre );
#endif
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_sopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
magma_s_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_s_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
CHECK( magma_sparse_opts( argc, argv, &zopts, &i, queue ));
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.solver_par.solver != Magma_LOBPCG )
zopts.precond_par.solver = Magma_NONE;
CHECK( magma_ssolverinfo_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_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 );
// for the eigensolver case
zopts.solver_par.ev_length = A.num_rows;
CHECK( magma_seigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
CHECK( magma_smscale( &A, zopts.scaling, queue ));
CHECK( magma_smconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_svinit( &b, Magma_DEV, A.num_cols, 1, one, queue ));
//magma_svinit( &x, Magma_DEV, A.num_cols, 1, one, queue );
//magma_s_spmv( one, B_d, x, zero, b, queue ); // b = A x
//magma_smfree(&x, queue );
CHECK( magma_svinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
info = magma_s_solver( B_d, b, &x, &zopts, queue );
if( info != 0 ){
printf("error: solver returned: %s (%d).\n",
magma_strerror( info ), info );
}
magma_ssolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
i++;
}
cleanup:
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
magma_ssolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing csr matrix add
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( &queue );
real_Double_t res;
magma_s_matrix A={Magma_CSR}, B={Magma_CSR}, B2={Magma_CSR},
A_d={Magma_CSR}, B_d={Magma_CSR}, C_d={Magma_CSR};
float one = MAGMA_S_MAKE(1.0, 0.0);
float mone = MAGMA_S_MAKE(-1.0, 0.0);
magma_int_t i=1;
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("%% matrix info: %d-by-%d with %d nonzeros\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
i++;
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, &B, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &B, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(B.num_rows), int(B.num_cols), int(B.nnz) );
CHECK( magma_smtransfer( A, &A_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_scuspaxpy( &one, A_d, &one, B_d, &C_d, queue ));
magma_smfree(&B_d, queue );
CHECK( magma_scuspaxpy( &mone, A_d, &one, C_d, &B_d, queue ));
CHECK( magma_smtransfer( B_d, &B2, Magma_DEV, Magma_CPU, queue ));
magma_smfree(&A_d, queue );
magma_smfree(&B_d, queue );
magma_smfree(&C_d, queue );
// check difference
CHECK( magma_smdiff( B, B2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix add: ok\n");
else
printf("%% tester matrix add: failed\n");
magma_smfree(&A, queue );
magma_smfree(&B, queue );
magma_smfree(&B2, queue );
cleanup:
magma_smfree(&A_d, queue );
magma_smfree(&B_d, queue );
magma_smfree(&C_d, queue );
magma_smfree(&A, queue );
magma_smfree(&B, queue );
magma_smfree(&B2, 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_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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_scg magma_scg_merge
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_s_solver_par solver_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.num_eigenvalues = 0;
magma_s_preconditioner precond_par;
precond_par.solver = Magma_JACOBI;
precond_par.levels = 0;
precond_par.sweeps = 10;
int precond = 0;
int format = 0;
int version = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_s_sparse_matrix A, B, B_d;
magma_s_vector x, b;
B.blocksize = 8;
B.alignment = 8;
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
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 ) {
precond = atoi( argv[++i] );
switch( precond ) {
case 0: precond_par.solver = Magma_JACOBI; break;
case 1: precond_par.solver = Magma_ICC; break;
case 2: precond_par.solver = Magma_AICC; break;
}
}else if ( strcmp("--version", argv[i]) == 0 ) {
version = atoi( argv[++i] );
}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("--levels", argv[i]) == 0 ) {
precond_par.levels = atoi( argv[++i] );
}else if ( strcmp("--sweeps", argv[i]) == 0 ) {
precond_par.sweeps = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_spcg"
" [ --format %d (0=CSR, 1=ELL 2=ELLRT, 3=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"
" --precond %d (0=Jacobi, 1=IC, 2=AIC [ --levels %d --sweeps %d]) ]"
" matrices \n\n", format, (int)B.blocksize, (int)B.alignment,
scale,
(int)solver_par.verbose,
(int)solver_par.maxiter, solver_par.epsilon,
precond, (int) precond_par.levels, (int) precond_par.sweeps );
magma_ssolverinfo_init( &solver_par, &precond_par );
while( i < argc ){
magma_s_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_smscale( &A, scaling );
magma_s_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_s_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_s_vinit( &b, Magma_DEV, A.num_cols, one );
magma_s_vinit( &x, Magma_DEV, A.num_cols, one );
magma_s_spmv( one, B_d, x, zero, b ); // b = A x
magma_s_vfree(&x);
magma_s_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_s_precondsetup( B_d, b, &precond_par );
magma_spcg( B_d, b, &x, &solver_par, &precond_par );
magma_ssolverinfo( &solver_par, &precond_par );
magma_s_mfree(&B_d);
magma_s_mfree(&B);
magma_s_mfree(&A);
magma_s_vfree(&x);
magma_s_vfree(&b);
i++;
}
magma_ssolverinfo_free( &solver_par, &precond_par );
TESTING_FINALIZE();
return 0;
}
magma_int_t
magma_scustomicsetup(
magma_s_matrix A,
magma_s_matrix b,
magma_s_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_s_matrix hA={Magma_CSR};
char preconditionermatrix[255];
snprintf( preconditionermatrix, sizeof(preconditionermatrix),
"precondL.mtx" );
CHECK( magma_s_csr_mtx( &hA, preconditionermatrix , queue) );
// for CUSPARSE
CHECK( magma_smtransfer( hA, &precond->M, Magma_CPU, Magma_DEV , queue ));
// copy the matrix to precond->L and (transposed) to precond->U
CHECK( magma_smtransfer(precond->M, &(precond->L), Magma_DEV, Magma_DEV, queue ));
CHECK( magma_smtranspose( precond->L, &(precond->U), queue ));
// extract the diagonal of L into precond->d
CHECK( magma_sjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_svinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_S_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_sjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_svinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_S_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( cusparseScsrsv_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( cusparseScsrsv_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_smfree( &hA, 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( &queue );
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
magma_s_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_s_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
CHECK( magma_sparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
CHECK( magma_ssolverinfo_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_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &A, argv[i], queue ));
}
// for the eigensolver case
zopts.solver_par.ev_length = A.num_cols;
CHECK( magma_seigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
CHECK( magma_smscale( &A, zopts.scaling, queue ));
// preconditioner
if ( zopts.solver_par.solver != Magma_ITERREF ) {
CHECK( magma_s_precondsetup( A, b, &zopts.solver_par, &zopts.precond_par, queue ) );
}
CHECK( magma_smconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
printf( "\n%% matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
printf("matrixinfo = [ \n");
printf("%% size (m x n) || nonzeros (nnz) || nnz/m || stored nnz\n");
printf("%%======================================================================"
"======%%\n");
printf(" %8d %8d %10d %4d %10d\n",
int(B.num_rows), int(B.num_cols), int(B.true_nnz), int(B.true_nnz/B.num_rows), int(B.nnz) );
printf("%%======================================================================"
"======%%\n");
printf("];\n");
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_svinit( &b, Magma_DEV, A.num_rows, 1, one, queue ));
//magma_svinit( &x, Magma_DEV, A.num_cols, 1, one, queue );
//magma_s_spmv( one, B_d, x, zero, b, queue ); // b = A x
//magma_smfree(&x, queue );
CHECK( magma_svinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
info = magma_s_solver( B_d, b, &x, &zopts, queue );
if( info != 0 ) {
printf("%%error: solver returned: %s (%d).\n",
magma_strerror( info ), int(info) );
}
printf("data = [\n");
magma_ssolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
printf("];\n\n");
printf("precond_info = [\n");
printf("%% setup runtime\n");
printf(" %.6f %.6f\n",
zopts.precond_par.setuptime, zopts.precond_par.runtime );
printf("];\n\n");
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
i++;
}
cleanup:
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
magma_ssolverinfo_free( &zopts.solver_par, &zopts.precond_par, 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_queue_t queue=NULL;
magma_queue_create( 0, &queue );
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
magma_s_matrix A={Magma_CSR}, B_d={Magma_CSR};
magma_s_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] );
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) );
magma_int_t n = A.num_rows;
CHECK( magma_smtransfer( A, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_svinit( &b, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_svinit( &x, Magma_DEV, A.num_cols, 1, one, queue ));
CHECK( magma_sprint_vector( b, 90, 10, queue ));
CHECK( magma_sprint_matrix( A, queue ));
printf("\n\n\n");
CHECK( magma_sprint_matrix( B_d, queue ));
float res;
res = magma_snrm2(n, b.dval, 1, queue );
printf("norm0: %f\n", res);
CHECK( magma_s_spmv( one, B_d, x, zero, b, queue )); // b = A x
CHECK( magma_sprint_vector( b, 0, 100, queue ));
CHECK( magma_sprint_vector( b, b.num_rows-10, 10, queue ));
res = magma_snrm2( n, b.dval, 1, queue );
printf("norm: %f\n", res);
CHECK( magma_sresidual( B_d, x, b, &res, queue));
printf("res: %f\n", res);
magma_smfree(&B_d, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
i++;
}
cleanup:
magma_smfree(&A, queue );
magma_smfree(&B_d, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
magma_queue_destroy( queue );
magma_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;
}
