| 1 |
// @HEADER |
| 2 |
// ***************************************************************************** |
| 3 |
// Stokhos Package |
| 4 |
// |
| 5 |
// Copyright 2009 NTESS and the Stokhos contributors. |
| 6 |
// SPDX-License-Identifier: BSD-3-Clause |
| 7 |
// ***************************************************************************** |
| 8 |
// @HEADER |
| 9 |
|
| 10 |
#include <iostream> |
| 11 |
|
| 12 |
// Kokkos CrsMatrix |
| 13 |
#include "KokkosSparse_CrsMatrix.hpp" |
| 14 |
#include "KokkosSparse_spmv.hpp" |
| 15 |
|
| 16 |
|
| 17 |
// Utilities |
| 18 |
#include "Kokkos_Timer.hpp" |
| 19 |
|
| 20 |
template< typename IntType > |
| 21 |
inline |
| 22 |
IntType map_fem_graph_coord( const IntType & N , |
| 23 |
const IntType & i , |
| 24 |
const IntType & j , |
| 25 |
const IntType & k ) |
| 26 |
{ |
| 27 |
return k + N * ( j + N * i ); |
| 28 |
} |
| 29 |
|
| 30 |
inline |
| 31 |
size_t generate_fem_graph( size_t N , |
| 32 |
std::vector< std::vector<size_t> > & graph ) |
| 33 |
{ |
| 34 |
graph.resize( N * N * N , std::vector<size_t>() ); |
| 35 |
|
| 36 |
size_t total = 0 ; |
| 37 |
|
| 38 |
for ( int i = 0 ; i < (int) N ; ++i ) { |
| 39 |
for ( int j = 0 ; j < (int) N ; ++j ) { |
| 40 |
for ( int k = 0 ; k < (int) N ; ++k ) { |
| 41 |
|
| 42 |
const size_t row = map_fem_graph_coord((int)N,i,j,k); |
| 43 |
|
| 44 |
graph[row].reserve(27); |
| 45 |
|
| 46 |
for ( int ii = -1 ; ii < 2 ; ++ii ) { |
| 47 |
for ( int jj = -1 ; jj < 2 ; ++jj ) { |
| 48 |
for ( int kk = -1 ; kk < 2 ; ++kk ) { |
| 49 |
if ( 0 <= i + ii && i + ii < (int) N && |
| 50 |
0 <= j + jj && j + jj < (int) N && |
| 51 |
0 <= k + kk && k + kk < (int) N ) { |
| 52 |
size_t col = map_fem_graph_coord((int)N,i+ii,j+jj,k+kk); |
| 53 |
|
| 54 |
graph[row].push_back(col); |
| 55 |
} |
| 56 |
}}} |
| 57 |
total += graph[row].size(); |
| 58 |
}}} |
| 59 |
|
| 60 |
return total ; |
| 61 |
} |
| 62 |
|
| 63 |
template <typename ScalarType, typename OrdinalType, typename Device> |
| 64 |
void |
| 65 |
test_spmm(const OrdinalType ensemble_length, |
| 66 |
const OrdinalType nGrid, |
| 67 |
const OrdinalType iterCount, |
| 68 |
std::vector<double>& scalar_perf, |
| 69 |
std::vector<double>& block_left_perf, |
| 70 |
std::vector<double>& block_right_perf) |
| 71 |
{ |
| 72 |
typedef ScalarType value_type; |
| 73 |
typedef OrdinalType ordinal_type; |
| 74 |
typedef Device execution_space; |
| 75 |
typedef Kokkos::View< value_type*, execution_space > vector_type; |
| 76 |
typedef Kokkos::View< value_type**, Kokkos::LayoutLeft, execution_space > left_multivec_type; |
| 77 |
//typedef Kokkos::View< value_type**, Kokkos::LayoutRight, execution_space > right_multivec_type; |
| 78 |
typedef KokkosSparse::CrsMatrix< value_type, ordinal_type, execution_space > matrix_type; |
| 79 |
typedef typename matrix_type::StaticCrsGraphType matrix_graph_type; |
| 80 |
typedef typename matrix_type::values_type matrix_values_type; |
| 81 |
|
| 82 |
//------------------------------ |
| 83 |
// Generate graph for "FEM" box structure: |
| 84 |
|
| 85 |
std::vector< std::vector<size_t> > fem_graph; |
| 86 |
const size_t fem_length = nGrid * nGrid * nGrid; |
| 87 |
const size_t graph_length = generate_fem_graph( nGrid , fem_graph ); |
| 88 |
|
| 89 |
//------------------------------ |
| 90 |
// Generate input vectors: |
| 91 |
|
| 92 |
std::vector<vector_type> x(ensemble_length); |
| 93 |
std::vector<vector_type> y(ensemble_length); |
| 94 |
for (ordinal_type e=0; e<ensemble_length; ++e) { |
| 95 |
x[e] = vector_type(Kokkos::ViewAllocateWithoutInitializing("x"), fem_length); |
| 96 |
y[e] = vector_type(Kokkos::ViewAllocateWithoutInitializing("y"), fem_length); |
| 97 |
|
| 98 |
Kokkos::deep_copy( x[e] , value_type(1.0) ); |
| 99 |
Kokkos::deep_copy( y[e] , value_type(0.0) ); |
| 100 |
} |
| 101 |
left_multivec_type xl(Kokkos::ViewAllocateWithoutInitializing("xl"), fem_length, ensemble_length); |
| 102 |
left_multivec_type yl(Kokkos::ViewAllocateWithoutInitializing("yl"), fem_length, ensemble_length); |
| 103 |
// right_multivec_type xr(Kokkos::ViewAllocateWithoutInitializing("xr"), fem_length, ensemble_length); |
| 104 |
// right_multivec_type yr(Kokkos::ViewAllocateWithoutInitializing("yr"), fem_length, ensemble_length); |
| 105 |
Kokkos::deep_copy(xl, value_type(1.0)); |
| 106 |
//Kokkos::deep_copy(xr, value_type(1.0)); |
| 107 |
Kokkos::deep_copy(yl, value_type(0.0)); |
| 108 |
//Kokkos::deep_copy(yr, value_type(0.0)); |
| 109 |
|
| 110 |
//------------------------------ |
| 111 |
// Generate matrix |
| 112 |
|
| 113 |
matrix_graph_type matrix_graph = |
| 114 |
Kokkos::create_staticcrsgraph<matrix_graph_type>( |
| 115 |
std::string("test crs graph"), fem_graph); |
| 116 |
matrix_values_type matrix_values = |
| 117 |
matrix_values_type(Kokkos::ViewAllocateWithoutInitializing("matrix"), graph_length); |
| 118 |
matrix_type matrix("matrix", fem_length, matrix_values, matrix_graph); |
| 119 |
Kokkos::deep_copy( matrix_values , value_type(1.0) ); |
| 120 |
|
| 121 |
//------------------------------ |
| 122 |
// Scalar multiply |
| 123 |
|
| 124 |
{ |
| 125 |
// warm up |
| 126 |
for (ordinal_type iter = 0; iter < iterCount; ++iter) { |
| 127 |
for (ordinal_type e=0; e<ensemble_length; ++e) { |
| 128 |
KokkosSparse::spmv( "N", value_type(1.0), matrix, x[e] , value_type(0.0) , y[e]); |
| 129 |
} |
| 130 |
} |
| 131 |
|
| 132 |
execution_space().fence(); |
| 133 |
Kokkos::Timer clock ; |
| 134 |
for (ordinal_type iter = 0; iter < iterCount; ++iter) { |
| 135 |
for (ordinal_type e=0; e<ensemble_length; ++e) { |
| 136 |
KokkosSparse::spmv( "N", value_type(1.0), matrix, x[e] , value_type(0.0) , y[e]); |
| 137 |
} |
| 138 |
} |
| 139 |
execution_space().fence(); |
| 140 |
|
| 141 |
const double seconds_per_iter = clock.seconds() / ((double) iterCount ); |
| 142 |
const double flops = 1.0e-9 * 2.0 * graph_length * ensemble_length; |
| 143 |
|
| 144 |
scalar_perf.resize(5); |
| 145 |
scalar_perf[0] = fem_length; |
| 146 |
scalar_perf[1] = ensemble_length; |
| 147 |
scalar_perf[2] = graph_length; |
| 148 |
scalar_perf[3] = seconds_per_iter; |
| 149 |
scalar_perf[4] = flops / seconds_per_iter; |
| 150 |
} |
| 151 |
|
| 152 |
//------------------------------ |
| 153 |
// Block-left multiply |
| 154 |
|
| 155 |
{ |
| 156 |
// warm up |
| 157 |
for (ordinal_type iter = 0; iter < iterCount; ++iter) { |
| 158 |
KokkosSparse::spmv( "N", value_type(1.0), matrix, xl , value_type(0.0) , yl); |
| 159 |
} |
| 160 |
|
| 161 |
execution_space().fence(); |
| 162 |
Kokkos::Timer clock ; |
| 163 |
for (ordinal_type iter = 0; iter < iterCount; ++iter) { |
| 164 |
KokkosSparse::spmv( "N", value_type(1.0), matrix, xl , value_type(0.0) , yl); |
| 165 |
} |
| 166 |
execution_space().fence(); |
| 167 |
|
| 168 |
const double seconds_per_iter = clock.seconds() / ((double) iterCount ); |
| 169 |
const double flops = 1.0e-9 * 2.0 * graph_length * ensemble_length; |
| 170 |
|
| 171 |
block_left_perf.resize(5); |
| 172 |
block_left_perf[0] = fem_length; |
| 173 |
block_left_perf[1] = ensemble_length; |
| 174 |
block_left_perf[2] = graph_length; |
| 175 |
block_left_perf[3] = seconds_per_iter; |
| 176 |
block_left_perf[4] = flops / seconds_per_iter; |
| 177 |
} |
| 178 |
|
| 179 |
#if 0 |
| 180 |
//------------------------------ |
| 181 |
// Block-right multiply |
| 182 |
|
| 183 |
{ |
| 184 |
// warm up |
| 185 |
for (ordinal_type iter = 0; iter < iterCount; ++iter) { |
| 186 |
KokkosSparse::spmv( "N", value_type(1.0), matrix, xr , value_type(0.0) , yr); |
| 187 |
} |
| 188 |
|
| 189 |
execution_space().fence(); |
| 190 |
Kokkos::Timer clock ; |
| 191 |
for (ordinal_type iter = 0; iter < iterCount; ++iter) { |
| 192 |
KokkosSparse::spmv( "N", value_type(1.0), matrix, xr , value_type(0.0) , yr); |
| 193 |
} |
| 194 |
execution_space().fence(); |
| 195 |
|
| 196 |
const double seconds_per_iter = clock.seconds() / ((double) iterCount ); |
| 197 |
const double flops = 1.0e-9 * 2.0 * graph_length * ensemble_length; |
| 198 |
|
| 199 |
block_right_perf.resize(5); |
| 200 |
block_right_perf[0] = fem_length; |
| 201 |
block_right_perf[1] = ensemble_length; |
| 202 |
block_right_perf[2] = graph_length; |
| 203 |
block_right_perf[3] = seconds_per_iter; |
| 204 |
block_right_perf[4] = flops / seconds_per_iter; |
| 205 |
} |
| 206 |
#endif |
| 207 |
|
| 208 |
} |
| 209 |
|
| 210 |
template <typename Scalar, typename Ordinal, typename Device> |
| 211 |
void performance_test_driver( const Ordinal nGrid, |
| 212 |
const Ordinal nIter, |
| 213 |
const Ordinal ensemble_min, |
| 214 |
const Ordinal ensemble_max, |
| 215 |
const Ordinal ensemble_step ) |
| 216 |
{ |
| 217 |
std::cout.precision(8); |
| 218 |
std::cout << std::endl |
| 219 |
<< "\"Grid Size\" , " |
| 220 |
<< "\"FEM Size\" , " |
| 221 |
<< "\"FEM Graph Size\" , " |
| 222 |
<< "\"Ensemble Size\" , " |
| 223 |
<< "\"Scalar SpMM Time\" , " |
| 224 |
<< "\"Scalar SpMM Speedup\" , " |
| 225 |
<< "\"Scalar SpMM GFLOPS\" , " |
| 226 |
<< "\"Block-Left SpMM Speedup\" , " |
| 227 |
<< "\"Block-Left SpMM GFLOPS\" , " |
| 228 |
//<< "\"Block_Right SpMM Speedup\" , " |
| 229 |
//<< "\"Block_Right SpMM GFLOPS\" , " |
| 230 |
<< std::endl; |
| 231 |
|
| 232 |
std::vector<double> perf_scalar, perf_block_left, perf_block_right; |
| 233 |
for (Ordinal e=ensemble_min; e<=ensemble_max; e+=ensemble_step) { |
| 234 |
|
| 235 |
test_spmm<Scalar,Ordinal,Device>( |
| 236 |
e, nGrid, nIter, perf_scalar, perf_block_left, perf_block_right ); |
| 237 |
|
| 238 |
std::cout << nGrid << " , " |
| 239 |
<< perf_scalar[0] << " , " |
| 240 |
<< perf_scalar[2] << " , " |
| 241 |
<< perf_scalar[1] << " , " |
| 242 |
<< perf_scalar[3] << " , " |
| 243 |
<< perf_scalar[4] / perf_scalar[4] << " , " |
| 244 |
<< perf_scalar[4] << " , " |
| 245 |
<< perf_block_left[4]/ perf_scalar[4] << " , " |
| 246 |
<< perf_block_left[4] << " , " |
| 247 |
//<< perf_block_right[4]/ perf_scalar[4] << " , " |
| 248 |
//<< perf_block_right[4] << " , " |
| 249 |
<< std::endl; |
| 250 |
|
| 251 |
} |
| 252 |
} |