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_zbaiter_overlap(
magma_z_matrix A,
magma_z_matrix b,
magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_BAITERO;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
// initial residual
real_Double_t tempo1, tempo2, runtime=0;
double residual;
magma_int_t localiter = precond_par->maxiter;
magma_z_matrix Ah={Magma_CSR}, ACSR={Magma_CSR}, A_d={Magma_CSR}, r={Magma_CSR},
D={Magma_CSR}, R={Magma_CSR};
// setup
magma_int_t matrices;
matrices = precond_par->levels;
struct magma_z_matrix D_d[ 256 ];
struct magma_z_matrix R_d[ 256 ];
magma_int_t overlap;
magma_int_t blocksize = 256;
if( matrices==2 ||
matrices==4 ||
matrices==8 ||
matrices==16 ||
matrices==32 ||
matrices==64 ||
matrices==128 ){
overlap = blocksize/matrices;
}else if( matrices == 1){
overlap = 0;
}else{
printf("error: overlap ratio not supported.\n");
goto cleanup;
}
CHECK( magma_zmtransfer( A, &Ah, A.memory_location, Magma_CPU, queue ));
CHECK( magma_zmconvert( Ah, &ACSR, Ah.storage_type, Magma_CSR, queue ));
CHECK( magma_zmtransfer( ACSR, &A_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zresidualvec( A_d, b, *x, &r, &residual, queue));
solver_par->init_res = residual;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t) residual;
}
// setup
for( int i=0; i<matrices; i++ ){
CHECK( magma_zcsrsplit( i*overlap, 256, ACSR, &D, &R, queue ));
CHECK( magma_zmtransfer( D, &D_d[i], Magma_CPU, Magma_DEV, queue ));
CHECK( magma_zmtransfer( R, &R_d[i], Magma_CPU, Magma_DEV, queue ));
magma_zmfree(&D, queue );
magma_zmfree(&R, queue );
}
magma_int_t iterinc;
if( solver_par->verbose == 0 ){
iterinc = solver_par->maxiter;
}
else{
iterinc = solver_par->verbose;
}
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// block-asynchronous iteration iterator
do
{
tempo1 = magma_sync_wtime( queue );
solver_par->numiter+= iterinc;
for( int z=0; z<iterinc; z++){
CHECK( magma_zbajac_csr_overlap( localiter, matrices, overlap, D_d, R_d, b, x, queue ));
}
tempo2 = magma_sync_wtime( queue );
runtime += tempo2-tempo1;
if ( solver_par->verbose > 0 ) {
CHECK( magma_zresidualvec( A_d, b, *x, &r, &residual, queue));
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) residual;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) runtime;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
solver_par->runtime = runtime;
CHECK( magma_zresidual( 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 ){
info = MAGMA_SUCCESS;
}
else {
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&D, queue );
magma_zmfree(&R, queue );
for( int i=0; i<matrices; i++ ){
magma_zmfree(&D_d[i], queue );
magma_zmfree(&R_d[i], queue );
}
magma_zmfree(&A_d, queue );
magma_zmfree(&ACSR, queue );
magma_zmfree(&Ah, queue );
solver_par->info = info;
return info;
} /* magma_zbaiter_overlap */
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;
}
