extern "C" magma_int_t
magma_cbaiter_overlap(
magma_c_matrix A,
magma_c_matrix b,
magma_c_matrix *x,
magma_c_solver_par *solver_par,
magma_c_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_BAITERO;
// some useful variables
magmaFloatComplex c_zero = MAGMA_C_ZERO;
// initial residual
real_Double_t tempo1, tempo2, runtime=0;
float residual;
magma_int_t localiter = precond_par->maxiter;
magma_c_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_c_matrix D_d[ 256 ];
struct magma_c_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_cmtransfer( A, &Ah, A.memory_location, Magma_CPU, queue ));
CHECK( magma_cmconvert( Ah, &ACSR, Ah.storage_type, Magma_CSR, queue ));
CHECK( magma_cmtransfer( ACSR, &A_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_cresidualvec( 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_ccsrsplit( i*overlap, 256, ACSR, &D, &R, queue ));
CHECK( magma_cmtransfer( D, &D_d[i], Magma_CPU, Magma_DEV, queue ));
CHECK( magma_cmtransfer( R, &R_d[i], Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&D, queue );
magma_cmfree(&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_cbajac_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_cresidualvec( 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_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 ){
info = MAGMA_SUCCESS;
}
else {
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_cmfree(&r, queue );
magma_cmfree(&D, queue );
magma_cmfree(&R, queue );
for( int i=0; i<matrices; i++ ){
magma_cmfree(&D_d[i], queue );
magma_cmfree(&R_d[i], queue );
}
magma_cmfree(&A_d, queue );
magma_cmfree(&ACSR, queue );
magma_cmfree(&Ah, queue );
solver_par->info = info;
return info;
} /* magma_cbaiter_overlap */
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue=NULL;
magma_queue_create( &queue );
real_Double_t res;
magma_c_matrix A={Magma_CSR}, AT={Magma_CSR}, A2={Magma_CSR},
B={Magma_CSR}, B_d={Magma_CSR};
int i=1;
real_Double_t start, end;
CHECK( magma_cparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_cm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_c_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
// scale matrix
CHECK( magma_cmscale( &A, zopts.scaling, queue ));
// remove nonzeros in matrix
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_cmcsrcompressor( &A, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA CPU: %.2e seconds.\n", (end-start)/10 );
// transpose
CHECK( magma_cmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_cmconvert( AT, &B, Magma_CSR, Magma_CSR, queue ));
magma_cmfree(&AT, queue );
CHECK( magma_cmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_cmfree(&B, queue );
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_cmcsrcompressor_gpu( &B_d, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA GPU: %.2e seconds.\n", (end-start)/10 );
CHECK( magma_cmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_cmfree(&B_d, queue );
CHECK( magma_cmconvert( B, &AT, Magma_CSR, Magma_CSR, queue ));
magma_cmfree(&B, queue );
// transpose back
CHECK( magma_cmtranspose( AT, &A2, queue ));
magma_cmfree(&AT, queue );
CHECK( magma_cmdiff( A, A2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix compressor: ok\n");
else
printf("%% tester matrix compressor: failed\n");
magma_cmfree(&A, queue );
magma_cmfree(&A2, queue );
i++;
}
cleanup:
magma_cmfree(&AT, queue );
magma_cmfree(&B, queue );
magma_cmfree(&A, queue );
magma_cmfree(&A2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}