extern "C" magma_int_t
magma_cmdiagadd(
magma_c_sparse_matrix *A,
magmaFloatComplex add,
magma_queue_t queue )
{
if ( A->memory_location == Magma_CPU && A->storage_type == Magma_CSRCOO ) {
for( magma_int_t z=0; z<A->nnz; z++ ) {
if ( A->col[z]== A->rowidx[z] ) {
// add some identity matrix
A->val[z] = A->val[z] + add;
}
}
return MAGMA_SUCCESS;
}
else {
magma_c_sparse_matrix hA, CSRA;
magma_storage_t A_storage = A->storage_type;
magma_location_t A_location = A->memory_location;
magma_c_mtransfer( *A, &hA, A->memory_location, Magma_CPU, queue );
magma_c_mconvert( hA, &CSRA, hA.storage_type, Magma_CSRCOO, queue );
magma_cmdiagadd( &CSRA, add, queue );
magma_c_mfree( &hA, queue );
magma_c_mfree( A, queue );
magma_c_mconvert( CSRA, &hA, Magma_CSRCOO, A_storage, queue );
magma_c_mtransfer( hA, A, Magma_CPU, A_location, queue );
magma_c_mfree( &hA, queue );
magma_c_mfree( &CSRA, queue );
return MAGMA_SUCCESS;
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
magma_int_t
magma_cpastixsetup( magma_c_sparse_matrix A, magma_c_vector b,
magma_c_preconditioner *precond ){
#if defined(HAVE_PASTIX)
#if defined(PRECISION_d)
pastix_data_t *pastix_data = NULL; /* Pointer to a storage structure needed by pastix */
pastix_int_t ncol; /* Size of the matrix */
pastix_int_t *colptr = NULL; /* Indexes of first element of each column in row and values */
pastix_int_t *rows = NULL; /* Row of each element of the matrix */
pastix_float_t *values = NULL; /* Value of each element of the matrix */
pastix_float_t *rhs = NULL; /* right hand side */
pastix_int_t *iparm = NULL; /* integer parameters for pastix */
float *dparm = NULL; /* floating parameters for pastix */
pastix_int_t *perm = NULL; /* Permutation tabular */
pastix_int_t *invp = NULL; /* Reverse permutation tabular */
pastix_int_t mat_type;
magma_c_sparse_matrix A_h1, B;
magma_c_vector diag, c_t, b_h;
magma_c_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_C_ZERO );
magma_c_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_C_ZERO );
magma_c_vtransfer( b, &b_h, A.memory_location, Magma_CPU);
if( A.storage_type != Magma_CSR ){
magma_c_mtransfer( A, &A_h1, A.memory_location, Magma_CPU);
magma_c_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR);
}
else{
magma_c_mtransfer( A, &B, A.memory_location, Magma_CPU);
}
rhs = (pastix_float_t*) b_h.val;
ncol = B.num_rows;
colptr = B.row;
rows = B.col;
values = (pastix_float_t*) B.val;
mat_type = API_SYM_NO;
iparm = (pastix_int_t*)malloc(IPARM_SIZE*sizeof(pastix_int_t));
dparm = (pastix_float_t*)malloc(DPARM_SIZE*sizeof(pastix_float_t));
/*******************************************/
/* Initialize parameters to default values */
/*******************************************/
iparm[IPARM_MODIFY_PARAMETER] = API_NO;
pastix(&pastix_data, MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, rhs, 1, iparm, dparm);
iparm[IPARM_THREAD_NBR] = 16;
iparm[IPARM_SYM] = mat_type;
iparm[IPARM_FACTORIZATION] = API_FACT_LU;
iparm[IPARM_VERBOSE] = API_VERBOSE_YES;
iparm[IPARM_ORDERING] = API_ORDER_SCOTCH;
iparm[IPARM_INCOMPLETE] = API_NO;
iparm[IPARM_RHS_MAKING] = API_RHS_B;
//iparm[IPARM_AMALGAMATION] = 5;
iparm[IPARM_LEVEL_OF_FILL] = 0;
/* if (incomplete == 1)
{
dparm[DPARM_EPSILON_REFINEMENT] = 1e-7;
}
*/
/*
* Matrix needs :
* - to be in fortran numbering
* - to have only the lower triangular part in symmetric case
* - to have a graph with a symmetric structure in unsymmetric case
* If those criteria are not matched, the csc will be reallocated and changed.
*/
iparm[IPARM_MATRIX_VERIFICATION] = API_YES;
perm = (pastix_int_t*)malloc(ncol*sizeof(pastix_int_t));
invp = (pastix_int_t*)malloc(ncol*sizeof(pastix_int_t));
/*******************************************/
/* Step 1 - Ordering / Scotch */
/* Perform it only when the pattern of */
/* matrix change. */
/* eg: mesh refinement */
/* In many cases users can simply go from */
/* API_TASK_ORDERING to API_TASK_ANALYSE */
/* in one call. */
/*******************************************/
/*******************************************/
/* Step 2 - Symbolic factorization */
/* Perform it only when the pattern of */
/* matrix change. */
/*******************************************/
/*******************************************/
/* Step 3 - Mapping and Compute scheduling */
/* Perform it only when the pattern of */
/* matrix change. */
/*******************************************/
/*******************************************/
/* Step 4 - Numerical Factorisation */
/* Perform it each time the values of the */
/* matrix changed. */
/*******************************************/
iparm[IPARM_START_TASK] = API_TASK_ORDERING;
iparm[IPARM_END_TASK] = API_TASK_NUMFACT;
pastix(&pastix_data, MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, NULL, 1, iparm, dparm);
precond->int_array_1 = (magma_int_t*) perm;
precond->int_array_2 = (magma_int_t*) invp;
precond->M.val = (magmaFloatComplex*) values;
precond->M.col = (magma_int_t*) colptr;
precond->M.row = (magma_int_t*) rows;
precond->M.num_rows = A.num_rows;
precond->M.num_cols = A.num_cols;
precond->M.memory_location = Magma_CPU;
precond->pastix_data = pastix_data;
precond->iparm = iparm;
precond->dparm = dparm;
if( A.storage_type != Magma_CSR){
magma_c_mfree( &A_h1 );
}
magma_c_vfree( &b_h);
magma_c_mfree( &B );
#else
printf( "error: only real supported yet.\n");
#endif
#else
printf( "error: pastix not available.\n");
#endif
return MAGMA_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
extern "C" magma_int_t
magma_c_cucsrtranspose(
magma_c_sparse_matrix A,
magma_c_sparse_matrix *B,
magma_queue_t queue )
{
// for symmetric matrices: convert to csc using cusparse
if( A.storage_type == Magma_CSR && A.memory_location == Magma_DEV ) {
magma_c_sparse_matrix C;
magma_c_mtransfer( A, &C, Magma_DEV, Magma_DEV, queue );
// CUSPARSE context //
cusparseHandle_t handle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&handle);
cusparseSetStream( handle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseMatDescr_t descrB;
cusparseStatus = cusparseCreateMatDescr(&descrA);
cusparseStatus = cusparseCreateMatDescr(&descrB);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrB,CUSPARSE_MATRIX_TYPE_GENERAL);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrB,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
cusparseStatus =
cusparseCcsr2csc( handle, A.num_rows, A.num_rows, A.nnz,
A.dval, A.drow, A.dcol, C.dval, C.dcol, C.drow,
CUSPARSE_ACTION_NUMERIC,
CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0)
printf("error in transpose: %d.\n", cusparseStatus);
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroy( handle );
magma_c_mtransfer( C, B, Magma_DEV, Magma_DEV, queue );
if( A.fill_mode == Magma_FULL ){
B->fill_mode = Magma_FULL;
}
else if( A.fill_mode == Magma_LOWER ){
B->fill_mode = Magma_UPPER;
}
else if ( A.fill_mode == Magma_UPPER ){
B->fill_mode = Magma_LOWER;
}
// end CUSPARSE context //
return MAGMA_SUCCESS;
}else if( A.storage_type == Magma_CSR && A.memory_location == Magma_CPU ){
magma_c_sparse_matrix A_d, B_d;
magma_c_mtransfer( A, &A_d, A.memory_location, Magma_DEV, queue );
magma_c_cucsrtranspose( A_d, &B_d, queue );
magma_c_mtransfer( B_d, B, Magma_DEV, A.memory_location, queue );
magma_c_mfree( &A_d, queue );
magma_c_mfree( &B_d, queue );
return MAGMA_SUCCESS;
}else {
magma_c_sparse_matrix ACSR, BCSR;
magma_c_mconvert( A, &ACSR, A.storage_type, Magma_CSR, queue );
magma_c_cucsrtranspose( ACSR, &BCSR, queue );
magma_c_mconvert( BCSR, B, Magma_CSR, A.storage_type, queue );
magma_c_mfree( &ACSR, queue );
magma_c_mfree( &BCSR, queue );
return MAGMA_SUCCESS;
}
}
extern "C" magma_int_t
magma_cmscale(
magma_c_sparse_matrix *A,
magma_scale_t scaling,
magma_queue_t queue )
{
if ( A->memory_location == Magma_CPU && A->storage_type == Magma_CSRCOO ) {
if ( scaling == Magma_NOSCALE ) {
// no scale
;
}
else if ( scaling == Magma_UNITROW ) {
// scale to unit rownorm
magmaFloatComplex *tmp;
magma_cmalloc_cpu( &tmp, A->num_rows );
for( magma_int_t z=0; z<A->num_rows; z++ ) {
magmaFloatComplex s = MAGMA_C_MAKE( 0.0, 0.0 );
for( magma_int_t f=A->row[z]; f<A->row[z+1]; f++ )
s+= MAGMA_C_REAL(A->val[f])*MAGMA_C_REAL(A->val[f]);
tmp[z] = MAGMA_C_MAKE( 1.0/sqrt( MAGMA_C_REAL( s ) ), 0.0 );
}
for( magma_int_t z=0; z<A->nnz; z++ ) {
A->val[z] = A->val[z] * tmp[A->col[z]] * tmp[A->rowidx[z]];
}
magma_free_cpu( tmp );
}
else if (scaling == Magma_UNITDIAG ) {
// scale to unit diagonal
magmaFloatComplex *tmp;
magma_cmalloc_cpu( &tmp, A->num_rows );
for( magma_int_t z=0; z<A->num_rows; z++ ) {
magmaFloatComplex s = MAGMA_C_MAKE( 0.0, 0.0 );
for( magma_int_t f=A->row[z]; f<A->row[z+1]; f++ ) {
if ( A->col[f]== z ) {
// add some identity matrix
//A->val[f] = A->val[f] + MAGMA_C_MAKE( 100000.0, 0.0 );
s = A->val[f];
}
}
if ( s == MAGMA_C_MAKE( 0.0, 0.0 ) )
printf("error: zero diagonal element.\n");
tmp[z] = MAGMA_C_MAKE( 1.0/sqrt( MAGMA_C_REAL( s ) ), 0.0 );
}
for( magma_int_t z=0; z<A->nnz; z++ ) {
A->val[z] = A->val[z] * tmp[A->col[z]] * tmp[A->rowidx[z]];
}
magma_free_cpu( tmp );
}
else
printf( "error: scaling not supported\n" );
return MAGMA_SUCCESS;
}
else {
magma_c_sparse_matrix hA, CSRA;
magma_storage_t A_storage = A->storage_type;
magma_location_t A_location = A->memory_location;
magma_c_mtransfer( *A, &hA, A->memory_location, Magma_CPU, queue );
magma_c_mconvert( hA, &CSRA, hA.storage_type, Magma_CSRCOO, queue );
magma_cmscale( &CSRA, scaling, queue );
magma_c_mfree( &hA, queue );
magma_c_mfree( A, queue );
magma_c_mconvert( CSRA, &hA, Magma_CSRCOO, A_storage, queue );
magma_c_mtransfer( hA, A, Magma_CPU, A_location, queue );
magma_c_mfree( &hA, queue );
magma_c_mfree( &CSRA, queue );
return MAGMA_SUCCESS;
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
extern "C" magma_int_t
magma_cbaiter(
magma_c_sparse_matrix A,
magma_c_vector b,
magma_c_vector *x,
magma_c_solver_par *solver_par,
magma_queue_t queue )
{
// prepare solver feedback
solver_par->solver = Magma_BAITER;
solver_par->info = MAGMA_SUCCESS;
magma_c_sparse_matrix Ah, ACSR, A_d, D, R, D_d, R_d;
magma_c_mtransfer( A, &Ah, A.memory_location, Magma_CPU, queue );
magma_c_mconvert( Ah, &ACSR, Ah.storage_type, Magma_CSR, queue );
magma_c_mtransfer( ACSR, &A_d, Magma_CPU, Magma_DEV, queue );
// initial residual
real_Double_t tempo1, tempo2;
float residual;
magma_cresidual( A_d, b, *x, &residual, queue );
solver_par->init_res = residual;
solver_par->res_vec = NULL;
solver_par->timing = NULL;
// setup
magma_ccsrsplit( 256, ACSR, &D, &R, queue );
magma_c_mtransfer( D, &D_d, Magma_CPU, Magma_DEV, queue );
magma_c_mtransfer( R, &R_d, Magma_CPU, Magma_DEV, queue );
magma_int_t localiter = 1;
tempo1 = magma_sync_wtime( queue );
// block-asynchronous iteration iterator
for( int iter=0; iter<solver_par->maxiter; iter++)
magma_cbajac_csr( localiter, D_d, R_d, b, x, queue );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
magma_cresidual( A_d, b, *x, &residual, queue );
solver_par->final_res = residual;
solver_par->numiter = solver_par->maxiter;
if ( solver_par->init_res > solver_par->final_res )
solver_par->info = MAGMA_SUCCESS;
else
solver_par->info = MAGMA_DIVERGENCE;
magma_c_mfree(&D, queue );
magma_c_mfree(&R, queue );
magma_c_mfree(&D_d, queue );
magma_c_mfree(&R_d, queue );
magma_c_mfree(&A_d, queue );
magma_c_mfree(&ACSR, queue );
magma_c_mfree(&Ah, queue );
return MAGMA_SUCCESS;
} /* magma_cbaiter */
extern "C" magma_int_t
magma_ccsrsplit(
magma_int_t bsize,
magma_c_sparse_matrix A,
magma_c_sparse_matrix *D,
magma_c_sparse_matrix *R,
magma_queue_t queue )
{
if ( A.memory_location == Magma_CPU &&
( A.storage_type == Magma_CSR ||
A.storage_type == Magma_CSRCOO ) ) {
magma_int_t i, k, j, nnz_diag, nnz_offd;
magma_int_t stat_cpu = 0, stat_dev = 0;
D->val = NULL;
D->col = NULL;
D->row = NULL;
D->rowidx = NULL;
D->blockinfo = NULL;
D->diag = NULL;
D->dval = NULL;
D->dcol = NULL;
D->drow = NULL;
D->drowidx = NULL;
D->ddiag = NULL;
R->val = NULL;
R->col = NULL;
R->row = NULL;
R->rowidx = NULL;
R->blockinfo = NULL;
R->diag = NULL;
R->dval = NULL;
R->dcol = NULL;
R->drow = NULL;
R->drowidx = NULL;
R->ddiag = NULL;
nnz_diag = nnz_offd = 0;
// Count the new number of nonzeroes in the two matrices
for( i=0; i<A.num_rows; i+=bsize )
for( k=i; k<min(A.num_rows,i+bsize); k++ )
for( j=A.row[k]; j<A.row[k+1]; j++ )
if ( A.col[j] < i )
nnz_offd++;
else if ( A.col[j] < i+bsize )
nnz_diag++;
else
nnz_offd++;
// Allocate memory for the new matrices
D->storage_type = Magma_CSRD;
D->memory_location = A.memory_location;
D->num_rows = A.num_rows;
D->num_cols = A.num_cols;
D->nnz = nnz_diag;
R->storage_type = Magma_CSR;
R->memory_location = A.memory_location;
R->num_rows = A.num_rows;
R->num_cols = A.num_cols;
R->nnz = nnz_offd;
stat_cpu += magma_cmalloc_cpu( &D->val, nnz_diag );
stat_cpu += magma_index_malloc_cpu( &D->row, A.num_rows+1 );
stat_cpu += magma_index_malloc_cpu( &D->col, nnz_diag );
stat_cpu += magma_cmalloc_cpu( &R->val, nnz_offd );
stat_cpu += magma_index_malloc_cpu( &R->row, A.num_rows+1 );
stat_cpu += magma_index_malloc_cpu( &R->col, nnz_offd );
if( stat_cpu != 0 ){
magma_c_mfree( D, queue );
magma_c_mfree( R, queue );
return MAGMA_ERR_HOST_ALLOC;
}
// Fill up the new sparse matrices
D->row[0] = 0;
R->row[0] = 0;
nnz_offd = nnz_diag = 0;
for( i=0; i<A.num_rows; i+=bsize) {
for( k=i; k<min(A.num_rows,i+bsize); k++ ) {
D->row[k+1] = D->row[k];
R->row[k+1] = R->row[k];
for( j=A.row[k]; j<A.row[k+1]; j++ ) {
if ( A.col[j] < i ) {
R->val[nnz_offd] = A.val[j];
R->col[nnz_offd] = A.col[j];
R->row[k+1]++;
nnz_offd++;
}
else if ( A.col[j] < i+bsize ) {
// larger than diagonal remain as before
if ( A.col[j]>k ) {
D->val[nnz_diag] = A.val[ j ];
D->col[nnz_diag] = A.col[ j ];
D->row[k+1]++;
}
// diagonal is written first
else if ( A.col[j]==k ) {
D->val[D->row[k]] = A.val[ j ];
D->col[D->row[k]] = A.col[ j ];
D->row[k+1]++;
}
// smaller than diagonal are shifted one to the right
// to have room for the diagonal
else {
D->val[nnz_diag+1] = A.val[ j ];
D->col[nnz_diag+1] = A.col[ j ];
D->row[k+1]++;
}
nnz_diag++;
}
else {
R->val[nnz_offd] = A.val[j];
R->col[nnz_offd] = A.col[j];
R->row[k+1]++;
nnz_offd++;
}
}
}
}
return MAGMA_SUCCESS;
}
else {
magma_c_sparse_matrix Ah, ACSR, DCSR, RCSR, Dh, Rh;
magma_c_mtransfer( A, &Ah, A.memory_location, Magma_CPU, queue );
magma_c_mconvert( Ah, &ACSR, A.storage_type, Magma_CSR, queue );
magma_ccsrsplit( bsize, ACSR, &DCSR, &RCSR, queue );
magma_c_mconvert( DCSR, &Dh, Magma_CSR, A.storage_type, queue );
magma_c_mconvert( RCSR, &Rh, Magma_CSR, A.storage_type, queue );
magma_c_mtransfer( Dh, D, Magma_CPU, A.memory_location, queue );
magma_c_mtransfer( Rh, R, Magma_CPU, A.memory_location, queue );
magma_c_mfree( &Ah, queue );
magma_c_mfree( &ACSR, queue );
magma_c_mfree( &Dh, queue );
magma_c_mfree( &DCSR, queue );
magma_c_mfree( &Rh, queue );
magma_c_mfree( &RCSR, queue );
return MAGMA_SUCCESS;
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
