extern "C" magma_int_t magma_zmdiagadd( magma_z_sparse_matrix *A, magmaDoubleComplex 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_z_sparse_matrix hA, CSRA; magma_storage_t A_storage = A->storage_type; magma_location_t A_location = A->memory_location; magma_z_mtransfer( *A, &hA, A->memory_location, Magma_CPU, queue ); magma_z_mconvert( hA, &CSRA, hA.storage_type, Magma_CSRCOO, queue ); magma_zmdiagadd( &CSRA, add, queue ); magma_z_mfree( &hA, queue ); magma_z_mfree( A, queue ); magma_z_mconvert( CSRA, &hA, Magma_CSRCOO, A_storage, queue ); magma_z_mtransfer( hA, A, Magma_CPU, A_location, queue ); magma_z_mfree( &hA, queue ); magma_z_mfree( &CSRA, queue ); return MAGMA_SUCCESS; } }
magma_int_t magma_zbaiter( magma_z_sparse_matrix A, magma_z_vector b, magma_z_vector *x, magma_z_solver_par *solver_par ) { // prepare solver feedback solver_par->solver = Magma_BAITER; solver_par->info = 0; magma_z_sparse_matrix A_d, D, R, D_d, R_d; magma_z_mtransfer( A, &A_d, Magma_CPU, Magma_DEV ); // initial residual real_Double_t tempo1, tempo2; double residual; magma_zresidual( A_d, b, *x, &residual ); solver_par->init_res = residual; solver_par->res_vec = NULL; solver_par->timing = NULL; // setup magma_zcsrsplit( 256, A, &D, &R ); magma_z_mtransfer( D, &D_d, Magma_CPU, Magma_DEV ); magma_z_mtransfer( R, &R_d, Magma_CPU, Magma_DEV ); magma_int_t localiter = 1; magma_device_sync(); tempo1=magma_wtime(); // block-asynchronous iteration iterator for( int iter=0; iter<solver_par->maxiter; iter++) magma_zbajac_csr( localiter, D_d, R_d, b, x ); magma_device_sync(); tempo2=magma_wtime(); solver_par->runtime = (real_Double_t) tempo2-tempo1; magma_zresidual( A_d, b, *x, &residual ); solver_par->final_res = residual; solver_par->numiter = solver_par->maxiter; if( solver_par->init_res > solver_par->final_res ) solver_par->info = 0; else solver_par->info = -1; magma_z_mfree(&D); magma_z_mfree(&R); magma_z_mfree(&D_d); magma_z_mfree(&R_d); magma_z_mfree(&A_d); return MAGMA_SUCCESS; } /* magma_zbaiter */
extern "C" magma_int_t magma_zjacobisetup_diagscal( magma_z_sparse_matrix A, magma_z_vector *d, magma_queue_t queue ) { magma_int_t i; magma_z_sparse_matrix A_h1, B; magma_z_vector diag; magma_z_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue ); if ( A.storage_type != Magma_CSR) { magma_z_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue ); magma_z_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue ); } else { magma_z_mtransfer( A, &B, A.memory_location, Magma_CPU, queue ); } for( magma_int_t rowindex=0; rowindex<B.num_rows; rowindex++ ) { magma_int_t start = (B.drow[rowindex]); magma_int_t end = (B.drow[rowindex+1]); for( i=start; i<end; i++ ) { if ( B.dcol[i]==rowindex ) { diag.val[rowindex] = 1.0/B.val[i]; if ( MAGMA_Z_REAL( diag.val[rowindex]) == 0 ) printf(" error: zero diagonal element in row %d!\n", (int) rowindex); } } } magma_z_vtransfer( diag, d, Magma_CPU, A.memory_location, queue ); if ( A.storage_type != Magma_CSR) { magma_z_mfree( &A_h1, queue ); } magma_z_mfree( &B, queue ); magma_z_vfree( &diag, queue ); return MAGMA_SUCCESS; }
extern "C" magma_int_t magma_zcuspaxpy( magmaDoubleComplex *alpha, magma_z_sparse_matrix A, magmaDoubleComplex *beta, magma_z_sparse_matrix B, magma_z_sparse_matrix *AB, magma_queue_t queue ) { if ( A.memory_location == Magma_DEV && B.memory_location == Magma_DEV && ( A.storage_type == Magma_CSR || A.storage_type == Magma_CSRCOO ) && ( B.storage_type == Magma_CSR || B.storage_type == Magma_CSRCOO ) ) { magma_z_sparse_matrix C; C.num_rows = A.num_rows; C.num_cols = A.num_cols; C.storage_type = A.storage_type; C.memory_location = A.memory_location; magma_int_t stat_dev = 0; C.val = NULL; C.col = NULL; C.row = NULL; C.rowidx = NULL; C.blockinfo = NULL; C.diag = NULL; C.dval = NULL; C.dcol = NULL; C.drow = NULL; C.drowidx = NULL; C.ddiag = NULL; // 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; cusparseMatDescr_t descrC; cusparseStatus = cusparseCreateMatDescr(&descrA); cusparseStatus = cusparseCreateMatDescr(&descrB); cusparseStatus = cusparseCreateMatDescr(&descrC); if (cusparseStatus != 0) printf("error in MatrDescr.\n"); cusparseStatus = cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL); cusparseSetMatType(descrB,CUSPARSE_MATRIX_TYPE_GENERAL); cusparseSetMatType(descrC,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); cusparseSetMatIndexBase(descrC,CUSPARSE_INDEX_BASE_ZERO); if (cusparseStatus != 0) printf("error in IndexBase.\n"); // multiply A and B on the device magma_int_t baseC; // nnzTotalDevHostPtr points to host memory magma_index_t *nnzTotalDevHostPtr = (magma_index_t*) &C.nnz; cusparseSetPointerMode(handle, CUSPARSE_POINTER_MODE_HOST); stat_dev += magma_index_malloc( &C.drow, (A.num_rows + 1) ); cusparseXcsrgeamNnz(handle,A.num_rows, A.num_cols, descrA, A.nnz, A.drow, A.dcol, descrB, B.nnz, B.drow, B.dcol, descrC, C.row, nnzTotalDevHostPtr); if (NULL != nnzTotalDevHostPtr) { C.nnz = *nnzTotalDevHostPtr; } else { // workaround as nnz and base C are magma_int_t magma_index_t base_t, nnz_t; magma_index_getvector( 1, C.drow+C.num_rows, 1, &nnz_t, 1 ); magma_index_getvector( 1, C.drow, 1, &base_t, 1 ); C.nnz = (magma_int_t) nnz_t; baseC = (magma_int_t) base_t; C.nnz -= baseC; } stat_dev += magma_index_malloc( &C.dcol, C.nnz ); stat_dev += magma_zmalloc( &C.dval, C.nnz ); if( stat_dev != 0 ) { magma_z_mfree( &C, queue ); return MAGMA_ERR_DEVICE_ALLOC; } cusparseZcsrgeam(handle, A.num_rows, A.num_cols, alpha, descrA, A.nnz, A.dval, A.drow, A.dcol, beta, descrB, B.nnz, B.dval, B.drow, B.dcol, descrC, C.dval, C.drow, C.dcol); cusparseDestroyMatDescr( descrA ); cusparseDestroyMatDescr( descrB ); cusparseDestroyMatDescr( descrC ); cusparseDestroy( handle ); // end CUSPARSE context // magma_z_mtransfer( C, AB, Magma_DEV, Magma_DEV, queue ); magma_z_mfree( &C, queue ); return MAGMA_SUCCESS; } else { printf("error: CSRSPAXPY only supported on device and CSR format.\n"); return MAGMA_SUCCESS; } }
extern "C" magma_int_t magma_zpastixsetup( magma_z_sparse_matrix A, magma_z_vector b, magma_z_preconditioner *precond, magma_queue_t queue ) { #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 */ double *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_z_sparse_matrix A_h1, B; magma_z_vector diag, c_t, b_h; magma_z_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue ); magma_z_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue ); magma_z_vtransfer( b, &b_h, A.memory_location, Magma_CPU, queue ); if ( A.storage_type != Magma_CSR ) { magma_z_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue ); magma_z_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue ); } else { magma_z_mtransfer( A, &B, A.memory_location, Magma_CPU, queue ); } rhs = (pastix_float_t*) b_h.dval; ncol = B.num_rows; colptr = B.drow; rows = B.dcol; values = (pastix_float_t*) B.dval; 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.dval = (magmaDoubleComplex*) values; precond->M.dcol = (magma_int_t*) colptr; precond->M.drow = (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_z_mfree( &A_h1, queue ); } magma_z_vfree( &b_h, queue ); magma_z_mfree( &B, queue ); #else printf( "error: only double precision supported yet.\n"); #endif #else printf( "error: pastix not available.\n"); #endif return MAGMA_SUCCESS; }
extern "C" magma_int_t magma_zjacobisetup( magma_z_sparse_matrix A, magma_z_vector b, magma_z_sparse_matrix *M, magma_z_vector *c, magma_queue_t queue ) { magma_int_t i; magma_z_sparse_matrix A_h1, A_h2, B, C; magma_z_vector diag, c_t, b_h; magma_z_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue ); magma_z_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue ); magma_z_vtransfer( b, &b_h, A.memory_location, Magma_CPU, queue ); if ( A.storage_type != Magma_CSR ) { magma_z_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue ); magma_z_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue ); } else { magma_z_mtransfer( A, &B, A.memory_location, Magma_CPU, queue ); } for( magma_int_t rowindex=0; rowindex<B.num_rows; rowindex++ ) { magma_int_t start = (B.drow[rowindex]); magma_int_t end = (B.drow[rowindex+1]); for( i=start; i<end; i++ ) { if ( B.dcol[i]==rowindex ) { diag.val[rowindex] = B.val[i]; if ( MAGMA_Z_REAL( diag.val[rowindex]) == 0 ) printf(" error: zero diagonal element in row %d!\n", (int) rowindex); } } for( i=start; i<end; i++ ) { B.val[i] = B.val[i] / diag.val[rowindex]; if ( B.dcol[i]==rowindex ) { B.val[i] = MAGMA_Z_MAKE( 0., 0. ); } } c_t.val[rowindex] = b_h.val[rowindex] / diag.val[rowindex]; } magma_z_csr_compressor(&B.val, &B.drow, &B.dcol, &C.val, &C.drow, &C.dcol, &B.num_rows, queue ); C.num_rows = B.num_rows; C.num_cols = B.num_cols; C.memory_location = B.memory_location; C.nnz = C.drow[B.num_rows]; C.storage_type = B.storage_type; C.memory_location = B.memory_location; if ( A.storage_type != Magma_CSR) { A_h2.alignment = A.alignment; A_h2.blocksize = A.blocksize; magma_z_mconvert( C, &A_h2, Magma_CSR, A_h1.storage_type, queue ); magma_z_mtransfer( A_h2, M, Magma_CPU, A.memory_location, queue ); } else { magma_z_mtransfer( C, M, Magma_CPU, A.memory_location, queue ); } magma_z_vtransfer( c_t, c, Magma_CPU, A.memory_location, queue ); if ( A.storage_type != Magma_CSR) { magma_z_mfree( &A_h1, queue ); magma_z_mfree( &A_h2, queue ); } magma_z_mfree( &B, queue ); magma_z_mfree( &C, queue ); magma_z_vfree( &diag, queue ); magma_z_vfree( &c_t, queue ); magma_z_vfree( &b_h, queue ); return MAGMA_SUCCESS; }
/* //////////////////////////////////////////////////////////////////////////// -- testing csr matrix add */ int main( int argc, char** argv ) { TESTING_INIT(); magma_queue_t queue; magma_queue_create( /*devices[ opts->device ],*/ &queue ); real_Double_t res; magma_z_sparse_matrix A, B, B2, C, A_d, B_d, C_d; magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0); magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0); magmaDoubleComplex mone = MAGMA_Z_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] ); magma_zm_5stencil( laplace_size, &A, queue ); } else { // file-matrix test magma_z_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] ); magma_zm_5stencil( laplace_size, &B, queue ); } else { // file-matrix test magma_z_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 ); magma_z_mtransfer( A, &A_d, Magma_CPU, Magma_DEV, queue ); magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ); magma_zcuspaxpy( &one, A_d, &one, B_d, &C_d, queue ); magma_z_mfree(&B_d, queue ); magma_zcuspaxpy( &mone, A_d, &one, C_d, &B_d, queue ); magma_z_mtransfer( B_d, &B2, Magma_DEV, Magma_CPU, queue ); magma_z_mfree(&A_d, queue ); magma_z_mfree(&B_d, queue ); magma_z_mfree(&C_d, queue ); // check difference magma_zmdiff( 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_z_mfree(&A, queue ); magma_z_mfree(&B, queue ); magma_z_mfree(&B2, queue ); magma_queue_destroy( queue ); TESTING_FINALIZE(); return 0; }
/* //////////////////////////////////////////////////////////////////////////// -- running magma_zgmres */ int main( int argc, char** argv) { TESTING_INIT(); magma_z_solver_par solver_par; magma_z_preconditioner precond_par; solver_par.epsilon = 10e-16; solver_par.maxiter = 1000; solver_par.restart = 30; solver_par.num_eigenvalues = 0; solver_par.ortho = Magma_CGS; solver_par.verbose = 0; int format = 0; int ortho = 0; int scale = 0; magma_scale_t scaling = Magma_NOSCALE; magma_z_sparse_matrix A, B, B_d; magma_z_vector x, b; B.blocksize = 8; B.alignment = 8; magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0); magmaDoubleComplex zero = MAGMA_Z_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("--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("--ortho", argv[i]) == 0 ) { ortho = atoi( argv[++i] ); switch( ortho ) { case 0: solver_par.ortho = Magma_CGS; break; case 1: solver_par.ortho = Magma_MGS; break; case 2: solver_par.ortho = Magma_FUSED_CGS; break; } } else if ( strcmp("--restart", argv[i]) == 0 ) { solver_par.restart = 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], "%lf", &solver_par.epsilon ); } else break; } printf( "\n# usage: ./run_zgmres" " [ --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)" " --restart %d --maxiter %d --tol %.2e" " --ortho %d (0=CGS, 1=MGS, 2=FUSED_CGS) ]" " matrices \n\n", format, (int) B.blocksize, (int) B.alignment, (int) scale, (int) solver_par.verbose, (int) solver_par.restart, (int) solver_par.maxiter, solver_par.epsilon, ortho ); magma_zsolverinfo_init( &solver_par, &precond_par ); while( i < argc ){ magma_z_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_zmscale( &A, scaling ); magma_z_mconvert( A, &B, Magma_CSR, B.storage_type ); magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV ); // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zgmres( B_d, b, &x, &solver_par ); magma_zsolverinfo( &solver_par, &precond_par ); magma_z_mfree(&B_d); magma_z_mfree(&B); magma_z_mfree(&A); magma_z_vfree(&x); magma_z_vfree(&b); i++; } magma_zsolverinfo_free( &solver_par, &precond_par ); TESTING_FINALIZE(); return 0; }
extern "C" magma_int_t magma_zmscale( magma_z_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 magmaDoubleComplex *tmp; magma_zmalloc_cpu( &tmp, A->num_rows ); for( magma_int_t z=0; z<A->num_rows; z++ ) { magmaDoubleComplex s = MAGMA_Z_MAKE( 0.0, 0.0 ); for( magma_int_t f=A->row[z]; f<A->row[z+1]; f++ ) s+= MAGMA_Z_REAL(A->val[f])*MAGMA_Z_REAL(A->val[f]); tmp[z] = MAGMA_Z_MAKE( 1.0/sqrt( MAGMA_Z_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 magmaDoubleComplex *tmp; magma_zmalloc_cpu( &tmp, A->num_rows ); for( magma_int_t z=0; z<A->num_rows; z++ ) { magmaDoubleComplex s = MAGMA_Z_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_Z_MAKE( 100000.0, 0.0 ); s = A->val[f]; } } if ( s == MAGMA_Z_MAKE( 0.0, 0.0 ) ) printf("error: zero diagonal element.\n"); tmp[z] = MAGMA_Z_MAKE( 1.0/sqrt( MAGMA_Z_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_z_sparse_matrix hA, CSRA; magma_storage_t A_storage = A->storage_type; magma_location_t A_location = A->memory_location; magma_z_mtransfer( *A, &hA, A->memory_location, Magma_CPU, queue ); magma_z_mconvert( hA, &CSRA, hA.storage_type, Magma_CSRCOO, queue ); magma_zmscale( &CSRA, scaling, queue ); magma_z_mfree( &hA, queue ); magma_z_mfree( A, queue ); magma_z_mconvert( CSRA, &hA, Magma_CSRCOO, A_storage, queue ); magma_z_mtransfer( hA, A, Magma_CPU, A_location, queue ); magma_z_mfree( &hA, queue ); magma_z_mfree( &CSRA, queue ); return MAGMA_SUCCESS; } }
extern "C" magma_int_t magma_z_cucsrtranspose( magma_z_sparse_matrix A, magma_z_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_z_sparse_matrix C; magma_z_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 = cusparseZcsr2csc( 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_z_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_z_sparse_matrix A_d, B_d; magma_z_mtransfer( A, &A_d, A.memory_location, Magma_DEV, queue ); magma_z_cucsrtranspose( A_d, &B_d, queue ); magma_z_mtransfer( B_d, B, Magma_DEV, A.memory_location, queue ); magma_z_mfree( &A_d, queue ); magma_z_mfree( &B_d, queue ); return MAGMA_SUCCESS; }else { magma_z_sparse_matrix ACSR, BCSR; magma_z_mconvert( A, &ACSR, A.storage_type, Magma_CSR, queue ); magma_z_cucsrtranspose( ACSR, &BCSR, queue ); magma_z_mconvert( BCSR, B, Magma_CSR, A.storage_type, queue ); magma_z_mfree( &ACSR, queue ); magma_z_mfree( &BCSR, queue ); return MAGMA_SUCCESS; } }
magma_int_t magma_zailusetup( magma_z_sparse_matrix A, magma_z_preconditioner *precond ){ magma_z_sparse_matrix hAh, hA, hAL, hALCOO, hAU, hAUT, hAUCOO, dAL, dAU, hL, hU, dL, dU, DL, RL, DU, RU; // copy original matrix as CSRCOO to device magma_z_mtransfer(A, &hAh, A.memory_location, Magma_CPU); magma_z_mconvert( hAh, &hA, hAh.storage_type, Magma_CSR ); magma_z_mfree(&hAh); // in case using fill-in magma_zilustruct( &hA, precond->levels); // need only lower triangular hAL.diagorder_type == Magma_UNITY; magma_z_mconvert( hA, &hAL, Magma_CSR, Magma_CSRL ); magma_z_mconvert( hAL, &hALCOO, Magma_CSR, Magma_CSRCOO ); magma_z_mtransfer( hALCOO, &dAL, Magma_CPU, Magma_DEV ); magma_z_mtransfer( hALCOO, &dAU, Magma_CPU, Magma_DEV ); // need only upper triangular magma_z_mconvert( hA, &hAU, Magma_CSR, Magma_CSRU ); magma_z_cucsrtranspose( hAU, &hAUT ); magma_z_mconvert( hAUT, &hAUCOO, Magma_CSR, Magma_CSRCOO ); magma_z_mtransfer( hAUCOO, &dL, Magma_CPU, Magma_DEV ); magma_z_mtransfer( hAUCOO, &dU, Magma_CPU, Magma_DEV ); magma_z_mfree(&hALCOO); magma_z_mfree(&hAL); magma_z_mfree(&hAUCOO); magma_z_mfree(&hAUT); magma_z_mfree(&hAU); for(int i=0; i<precond->sweeps; i++){ magma_zailu_csr_s( dAL, dAU, dL, dU ); } magma_z_mtransfer( dL, &hL, Magma_DEV, Magma_CPU ); magma_z_mtransfer( dU, &hU, Magma_DEV, Magma_CPU ); magma_z_LUmergein( hL, hU, &hA); magma_z_mtransfer( hA, &precond->M, Magma_CPU, Magma_DEV ); magma_z_mfree(&dL); magma_z_mfree(&dU); magma_z_mfree(&dAL); magma_z_mfree(&dAU); hAL.diagorder_type = Magma_UNITY; magma_z_mconvert(hA, &hAL, Magma_CSR, Magma_CSRL); hAL.storage_type = Magma_CSR; magma_z_mconvert(hA, &hAU, Magma_CSR, Magma_CSRU); hAU.storage_type = Magma_CSR; magma_z_mfree(&hA); magma_z_mfree(&hL); magma_z_mfree(&hU); magma_zcsrsplit( 256, hAL, &DL, &RL ); magma_zcsrsplit( 256, hAU, &DU, &RU ); magma_z_mtransfer( DL, &precond->LD, Magma_CPU, Magma_DEV ); magma_z_mtransfer( DU, &precond->UD, Magma_CPU, Magma_DEV ); // for cusparse uncomment this magma_z_mtransfer( hAL, &precond->L, Magma_CPU, Magma_DEV ); magma_z_mtransfer( hAU, &precond->U, Magma_CPU, Magma_DEV ); // for ba-solve uncomment this /* if( RL.nnz != 0 ) magma_z_mtransfer( RL, &precond->L, Magma_CPU, Magma_DEV ); else{ precond->L.nnz = 0; precond->L.val = NULL; precond->L.col = NULL; precond->L.row = NULL; precond->L.blockinfo = NULL; } if( RU.nnz != 0 ) magma_z_mtransfer( RU, &precond->U, Magma_CPU, Magma_DEV ); else{ precond->U.nnz = 0; precond->L.val = NULL; precond->L.col = NULL; precond->L.row = NULL; precond->L.blockinfo = NULL; } */ magma_z_mfree(&hAL); magma_z_mfree(&hAU); magma_z_mfree(&DL); magma_z_mfree(&RL); magma_z_mfree(&DU); magma_z_mfree(&RU); // CUSPARSE context // cusparseHandle_t cusparseHandle; cusparseStatus_t cusparseStatus; cusparseStatus = cusparseCreate(&cusparseHandle); if(cusparseStatus != 0) printf("error in Handle.\n"); cusparseMatDescr_t descrL; cusparseStatus = cusparseCreateMatDescr(&descrL); if(cusparseStatus != 0) printf("error in MatrDescr.\n"); cusparseStatus = cusparseSetMatType(descrL,CUSPARSE_MATRIX_TYPE_TRIANGULAR); if(cusparseStatus != 0) printf("error in MatrType.\n"); cusparseStatus = cusparseSetMatDiagType (descrL, CUSPARSE_DIAG_TYPE_UNIT); if(cusparseStatus != 0) printf("error in DiagType.\n"); cusparseStatus = cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO); if(cusparseStatus != 0) printf("error in IndexBase.\n"); cusparseStatus = cusparseSetMatFillMode(descrL,CUSPARSE_FILL_MODE_LOWER); if(cusparseStatus != 0) printf("error in fillmode.\n"); cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoL); if(cusparseStatus != 0) printf("error in info.\n"); cusparseStatus = cusparseZcsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, precond->L.num_rows, precond->L.nnz, descrL, precond->L.val, precond->L.row, precond->L.col, precond->cuinfoL ); if(cusparseStatus != 0) printf("error in analysis.\n"); cusparseDestroyMatDescr( descrL ); cusparseMatDescr_t descrU; cusparseStatus = cusparseCreateMatDescr(&descrU); if(cusparseStatus != 0) printf("error in MatrDescr.\n"); cusparseStatus = cusparseSetMatType(descrU,CUSPARSE_MATRIX_TYPE_TRIANGULAR); if(cusparseStatus != 0) printf("error in MatrType.\n"); cusparseStatus = cusparseSetMatDiagType (descrU, CUSPARSE_DIAG_TYPE_NON_UNIT); if(cusparseStatus != 0) printf("error in DiagType.\n"); cusparseStatus = cusparseSetMatIndexBase(descrU,CUSPARSE_INDEX_BASE_ZERO); if(cusparseStatus != 0) printf("error in IndexBase.\n"); cusparseStatus = cusparseSetMatFillMode(descrU,CUSPARSE_FILL_MODE_UPPER); if(cusparseStatus != 0) printf("error in fillmode.\n"); cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoU); if(cusparseStatus != 0) printf("error in info.\n"); cusparseStatus = cusparseZcsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, precond->U.num_rows, precond->U.nnz, descrU, precond->U.val, precond->U.row, precond->U.col, precond->cuinfoU ); if(cusparseStatus != 0) printf("error in analysis.\n"); cusparseDestroyMatDescr( descrU ); cusparseDestroy( cusparseHandle ); return MAGMA_SUCCESS; }
magma_int_t magma_zaiccsetup( magma_z_sparse_matrix A, magma_z_preconditioner *precond ){ magma_z_sparse_matrix hAh, hA, hAL, hALCOO, dAL, hL, dL, DL, RL; // copy original matrix as CSRCOO to device magma_z_mtransfer(A, &hAh, A.memory_location, Magma_CPU); magma_z_mconvert( hAh, &hA, hAh.storage_type, Magma_CSR ); magma_z_mfree(&hAh); // in case using fill-in magma_zilustruct( &hA, precond->levels); magma_z_mconvert( hA, &hAL, Magma_CSR, Magma_CSRL ); magma_z_mconvert( hAL, &hALCOO, Magma_CSR, Magma_CSRCOO ); magma_z_mtransfer( hALCOO, &dAL, Magma_CPU, Magma_DEV ); magma_z_mtransfer( hALCOO, &dL, Magma_CPU, Magma_DEV ); magma_z_mfree(&hALCOO); magma_z_mfree(&hAL); magma_z_mfree(&hA); for(int i=0; i<precond->sweeps; i++){ magma_zaic_csr_s( dAL, dL ); } magma_z_mtransfer( dL, &hL, Magma_DEV, Magma_CPU ); magma_z_mfree(&dL); magma_z_mfree(&dAL); magma_z_mconvert(hL, &hAL, hL.storage_type, Magma_CSR); // for CUSPARSE magma_z_mtransfer( hAL, &precond->M, Magma_CPU, Magma_DEV ); magma_zcsrsplit( 256, hAL, &DL, &RL ); magma_z_mtransfer( DL, &precond->LD, Magma_CPU, Magma_DEV ); magma_z_mtransfer( RL, &precond->L, Magma_CPU, Magma_DEV ); magma_z_mfree(&hL); magma_z_cucsrtranspose( hAL, &hL ); magma_zcsrsplit( 256, hL, &DL, &RL ); magma_z_mtransfer( DL, &precond->UD, Magma_CPU, Magma_DEV ); magma_z_mtransfer( RL, &precond->U, Magma_CPU, Magma_DEV ); magma_z_mfree(&hAL); magma_z_mfree(&hL); magma_z_mfree(&DL); magma_z_mfree(&RL); // CUSPARSE context // cusparseHandle_t cusparseHandle; cusparseStatus_t cusparseStatus; cusparseStatus = cusparseCreate(&cusparseHandle); if(cusparseStatus != 0) printf("error in Handle.\n"); cusparseMatDescr_t descrL; cusparseStatus = cusparseCreateMatDescr(&descrL); if(cusparseStatus != 0) printf("error in MatrDescr.\n"); cusparseStatus = cusparseSetMatType(descrL,CUSPARSE_MATRIX_TYPE_TRIANGULAR); if(cusparseStatus != 0) printf("error in MatrType.\n"); cusparseStatus = cusparseSetMatDiagType (descrL, CUSPARSE_DIAG_TYPE_NON_UNIT); if(cusparseStatus != 0) printf("error in DiagType.\n"); cusparseStatus = cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO); if(cusparseStatus != 0) printf("error in IndexBase.\n"); cusparseStatus = cusparseSetMatFillMode(descrL,CUSPARSE_FILL_MODE_LOWER); if(cusparseStatus != 0) printf("error in fillmode.\n"); cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoL); if(cusparseStatus != 0) printf("error in info.\n"); cusparseStatus = cusparseZcsrsv_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 ); if(cusparseStatus != 0) printf("error in analysis L.\n"); cusparseDestroyMatDescr( descrL ); cusparseMatDescr_t descrU; cusparseStatus = cusparseCreateMatDescr(&descrU); if(cusparseStatus != 0) printf("error in MatrDescr.\n"); cusparseStatus = cusparseSetMatType(descrU,CUSPARSE_MATRIX_TYPE_TRIANGULAR); if(cusparseStatus != 0) printf("error in MatrType.\n"); cusparseStatus = cusparseSetMatDiagType (descrU, CUSPARSE_DIAG_TYPE_NON_UNIT); if(cusparseStatus != 0) printf("error in DiagType.\n"); cusparseStatus = cusparseSetMatIndexBase(descrU,CUSPARSE_INDEX_BASE_ZERO); if(cusparseStatus != 0) printf("error in IndexBase.\n"); cusparseStatus = cusparseSetMatFillMode(descrU,CUSPARSE_FILL_MODE_LOWER); if(cusparseStatus != 0) printf("error in fillmode.\n"); cusparseStatus = cusparseCreateSolveAnalysisInfo(&precond->cuinfoU); if(cusparseStatus != 0) printf("error in info.\n"); cusparseStatus = cusparseZcsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows, precond->M.nnz, descrU, precond->M.val, precond->M.row, precond->M.col, precond->cuinfoU ); if(cusparseStatus != 0) printf("error in analysis U.\n"); cusparseDestroyMatDescr( descrU ); cusparseDestroy( cusparseHandle ); return MAGMA_SUCCESS; }
/* //////////////////////////////////////////////////////////////////////////// -- Debugging file */ int main( int argc, char** argv) { TESTING_INIT(); magma_z_solver_par solver_par; magma_z_preconditioner precond_par; solver_par.epsilon = 10e-16; solver_par.maxiter = 1000; solver_par.verbose = 0; solver_par.restart = 30; solver_par.num_eigenvalues = 0; solver_par.ortho = Magma_CGS; magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0); magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0); magma_z_sparse_matrix A, B, B_d; magma_z_vector x, b; // generate matrix of desired structure and size magma_int_t n=100; // size is n*n magma_int_t nn = n*n; magma_int_t offdiags = 2; magma_index_t *diag_offset; magmaDoubleComplex *diag_vals; magma_zmalloc_cpu( &diag_vals, offdiags+1 ); magma_index_malloc_cpu( &diag_offset, offdiags+1 ); diag_offset[0] = 0; diag_offset[1] = 1; diag_offset[2] = n; diag_vals[0] = MAGMA_Z_MAKE( 4.1, 0.0 ); diag_vals[1] = MAGMA_Z_MAKE( -1.0, 0.0 ); diag_vals[2] = MAGMA_Z_MAKE( -1.0, 0.0 ); magma_zmgenerator( nn, offdiags, diag_offset, diag_vals, &A ); // convert marix into desired format B.storage_type = Magma_SELLC; B.blocksize = 8; B.alignment = 8; // scale matrix magma_zmscale( &A, Magma_UNITDIAG ); magma_z_mconvert( A, &B, Magma_CSR, B.storage_type ); magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV ); // test CG #################################### // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // solver magma_zcg_res( B_d, b, &x, &solver_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test PCG Jacobi ############################ // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_JACOBI; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpcg( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test PCG IC ################################ // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_ICC; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpcg( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test PCG IC ################################ // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_ICC; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpcg( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test BICGSTAB #################################### // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // solver magma_zbicgstab( B_d, b, &x, &solver_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test PBICGSTAB Jacobi ############################ // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_JACOBI; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); /* // test PBICGSTAB ILU ############################### // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_ILU; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test PBICGSTAB ILU ############################### // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x);printf("here\n"); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_ILU; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test GMRES #################################### // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // solver magma_zgmres( B_d, b, &x, &solver_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); // test PGMRES Jacobi ############################ // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_JACOBI; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpgmres( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b);*/ // test PGMRES ILU ############################### // vectors and initial guess magma_z_vinit( &b, Magma_DEV, A.num_cols, one ); magma_z_vinit( &x, Magma_DEV, A.num_cols, one ); magma_z_spmv( one, B_d, x, zero, b ); // b = A x magma_z_vfree(&x); magma_z_vinit( &x, Magma_DEV, A.num_cols, zero ); magma_zsolverinfo_init( &solver_par, &precond_par ); // Preconditioner precond_par.solver = Magma_ILU; magma_z_precondsetup( B_d, b, &precond_par ); // solver magma_zpgmres( B_d, b, &x, &solver_par, &precond_par ); // solverinfo magma_zsolverinfo( &solver_par, &precond_par ); if( solver_par.numiter > 150 ){ printf("error: test not passed!\n"); exit(-1); } magma_zsolverinfo_free( &solver_par, &precond_par ); magma_z_vfree(&x); magma_z_vfree(&b); printf("all tests passed.\n"); magma_z_mfree(&B_d); magma_z_mfree(&B); magma_z_mfree(&A); TESTING_FINALIZE(); return 0; }
/* //////////////////////////////////////////////////////////////////////////// -- 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_z_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; magmaDoubleComplex c_one = MAGMA_Z_MAKE(1.0, 0.0); magmaDoubleComplex c_zero = MAGMA_Z_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_zspmv" " [ --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_zm_5stencil( laplace_size, &hA, queue ); } else { // file-matrix test magma_z_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_z_vector hx, hy, dx, dy, hrefvec, hcheck; // init CPU vectors magma_z_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue ); magma_z_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue ); // init DEV vectors magma_z_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue ); magma_z_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_zcsrmv( "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_z_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_z_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_z_mfree(&dA, queue ); magma_z_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue ); // convert to ELL and copy to GPU magma_z_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue ); magma_z_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue ); magma_z_mfree(&hA_ELL, queue ); magma_z_vfree( &dy, queue ); magma_z_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_z_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_z_mfree(&dA_ELL, queue ); magma_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_REAL(hrefvec.val[k]); if ( res < .000001 ) printf("# tester spmv ELL: ok\n"); else printf("# tester spmv ELL: failed\n"); magma_z_vfree( &hcheck, queue ); // convert to SELLP and copy to GPU magma_z_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue ); magma_z_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue ); magma_z_mfree(&hA_SELLP, queue ); magma_z_vfree( &dy, queue ); magma_z_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_z_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_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_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_z_vfree( &hcheck, queue ); magma_z_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); magmaDoubleComplex alpha = c_one; magmaDoubleComplex beta = c_zero; magma_z_vfree( &dy, queue ); magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue ); // copy matrix to GPU magma_z_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue ); start = magma_sync_wtime( queue ); for (j=0; j<10; j++) cusparseStatus = cusparseZcsrmv(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_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_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_z_vfree( &hcheck, queue ); magma_z_vfree( &dy, queue ); magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue ); cusparseZcsr2hyb(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 = cusparseZhybmv( 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_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_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_z_vfree( &hcheck, queue ); cusparseDestroyMatDescr( descrA ); cusparseDestroyHybMat( hybA ); cusparseDestroy( cusparseHandle ); magma_z_mfree(&dA, queue ); printf("\n\n"); // free CPU memory magma_z_mfree(&hA, queue ); magma_z_vfree(&hx, queue ); magma_z_vfree(&hy, queue ); magma_z_vfree(&hrefvec, queue ); // free GPU memory magma_z_vfree(&dx, queue ); magma_z_vfree(&dy, queue ); i++; } magma_queue_destroy( queue ); TESTING_FINALIZE(); return 0; }