int ColorSYMGS( const SparseMatrix & A, const Vector & r, Vector & x){
assert(x.localLength == A.localNumberOfColumns); // Make sure x contains space for halo values
#ifndef HPCG_NO_MPI
ExchangeHalo(A,x);
#endif
Optimatrix* A_Optimized = (Optimatrix*)A.optimizationData;
local_matrix_type localMatrix = A_Optimized->localMatrix;
local_int_1d_type matrixDiagonal = A_Optimized->matrixDiagonal;
local_int_1d_type colors_ind = A_Optimized->colors_ind;
host_local_int_1d_type host_colors_ind = A_Optimized->host_colors_ind;
local_int_1d_type colors_map = A_Optimized->colors_map;
host_local_int_1d_type host_colors_map = A_Optimized->host_colors_map;
const int numColors = A_Optimized->numColors;
Optivector * r_Optimized = (Optivector*)r.optimizationData;
double_1d_type r_values = r_Optimized->values;
Optivector * x_Optimized = (Optivector*)x.optimizationData;
double_1d_type x_values = x_Optimized->values;
// Forward Sweep!
#ifdef KOKKOS_TEAM
int vector_size = 32;
int teamSizeMax = 8;
for(int i = 0; i < numColors; i++){
int color_index_begin = host_colors_map(i);
int color_index_end = host_colors_map(i + 1);
int numberOfTeams = color_index_end - color_index_begin;
Kokkos::parallel_for(team_policy(numberOfTeams / teamSizeMax + 1, teamSizeMax, vector_size),
ColouredSweep(color_index_begin, color_index_end, localMatrix, colors_ind, r_values, x_values));
execution_space::fence();
}
for(int i = numColors - 1; i >= 0; i--){
int color_index_begin = host_colors_map(i);
int color_index_end = host_colors_map(i+1);
int numberOfTeams = color_index_end - color_index_begin;
Kokkos::parallel_for(team_policy(numberOfTeams / teamSizeMax + 1, teamSizeMax, vector_size),
ColouredSweep(color_index_begin, color_index_end, localMatrix, colors_ind, r_values, x_values));
execution_space::fence();
}
#else
local_int_t dummy = 0;
for(int i = 0; i < numColors; i++){
int start = host_colors_map(i); // Colors start at 1, i starts at 0
int end = host_colors_map(i+1);
dummy += end - start;
Kokkos::parallel_for(end - start, colouredForwardSweep(start, colors_ind, localMatrix, r_values, x_values, matrixDiagonal));
}
assert(dummy == A.localNumberOfRows);
// Back Sweep!
for(int i = numColors -1; i >= 0; --i){
int start = host_colors_map(i); // Colors start at 1, i starts at 0
int end = host_colors_map(i+1);
Kokkos::parallel_for(end - start, colouredBackSweep(start, colors_ind, localMatrix, r_values, x_values, matrixDiagonal));
}
#endif
return(0);
}
hpx::future<void> ComputeSPMV_async( const SparseMatrix & A, /*const*/ Vector & x, Vector & y) {
assert(x.localLength>=A.localNumberOfColumns); // Test vector lengths
assert(y.localLength>=A.localNumberOfRows);
#ifndef HPCG_NOMPI
ExchangeHalo(A,x);
#endif
const double * const xv = x.values;
double * const yv = y.values;
const local_int_t nrow = A.localNumberOfRows;
typedef boost::counting_iterator<local_int_t> iterator;
return hpx::parallel::for_each(
hpx::parallel::par(hpx::parallel::task), iterator(0), iterator(nrow),
[xv, yv, &A](local_int_t i) {
double sum = 0.0;
const double * const cur_vals = A.matrixValues[i];
const local_int_t * const cur_inds = A.mtxIndL[i];
const int cur_nnz = A.nonzerosInRow[i];
for (int j=0; j< cur_nnz; j++)
sum += cur_vals[j]*xv[cur_inds[j]];
yv[i] = sum;
});
}
/*!
Routine to compute matrix vector product y = Ax where:
Precondition: First call exchange_externals to get off-processor values of x
This is the reference SPMV implementation. It CANNOT be modified for the
purposes of this benchmark.
@param[in] A the known system matrix
@param[in] x the known vector
@param[out] y the On exit contains the result: Ax.
@return returns 0 upon success and non-zero otherwise
@see ComputeSPMV
*/
int ComputeSPMV_ref( const SparseMatrix & A, Vector & x, Vector & y) {
assert(x.localLength>=A.localNumberOfColumns); // Test vector lengths
assert(y.localLength>=A.localNumberOfRows);
#ifndef HPCG_NOMPI
ExchangeHalo(A,x);
#endif
const double * const xv = x.values;
double * const yv = y.values;
const local_int_t nrow = A.localNumberOfRows;
#ifndef HPCG_NOOPENMP
#pragma omp parallel for
#endif
for (local_int_t i=0; i< nrow; i++) {
double sum = 0.0;
const double * const cur_vals = A.matrixValues[i];
const local_int_t * const cur_inds = A.mtxIndL[i];
const int cur_nnz = A.nonzerosInRow[i];
for (int j=0; j< cur_nnz; j++)
sum += cur_vals[j]*xv[cur_inds[j]];
yv[i] = sum;
}
return(0);
}
