| 1 |
/* |
| 2 |
* (C) Copyright 2013 ECMWF. |
| 3 |
* |
| 4 |
* This software is licensed under the terms of the Apache Licence Version 2.0 |
| 5 |
* which can be obtained at http://www.apache.org/licenses/LICENSE-2.0. |
| 6 |
* In applying this licence, ECMWF does not waive the privileges and immunities |
| 7 |
* granted to it by virtue of its status as an intergovernmental organisation |
| 8 |
* nor does it submit to any jurisdiction. |
| 9 |
*/ |
| 10 |
|
| 11 |
#include "atlas/linalg/sparse/SparseMatrixMultiply_OpenMP.h" |
| 12 |
|
| 13 |
#include "atlas/parallel/omp/omp.h" |
| 14 |
#include "atlas/runtime/Exception.h" |
| 15 |
|
| 16 |
namespace atlas { |
| 17 |
namespace linalg { |
| 18 |
namespace sparse { |
| 19 |
|
| 20 |
template <bool SetZero, typename MatrixValue, typename SourceValue, typename TargetValue> |
| 21 |
void spmv_layout_left(const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 1>& src, View<TargetValue, 1>& tgt) { |
| 22 |
using Value = TargetValue; |
| 23 |
const auto outer = W.outer(); |
| 24 |
const auto index = W.inner(); |
| 25 |
const auto weight = W.value(); |
| 26 |
const idx_t rows = static_cast<idx_t>(W.rows()); |
| 27 |
|
| 28 |
ATLAS_ASSERT(src.shape(0) >= W.cols()); |
| 29 |
ATLAS_ASSERT(tgt.shape(0) >= W.rows()); |
| 30 |
|
| 31 |
atlas_omp_parallel_for(idx_t r = 0; r < rows; ++r) { |
| 32 |
if constexpr (SetZero) { |
| 33 |
tgt[r] = 0.; |
| 34 |
} |
| 35 |
for (idx_t c = outer[r]; c < outer[r + 1]; ++c) { |
| 36 |
idx_t n = index[c]; |
| 37 |
Value w = static_cast<Value>(weight[c]); |
| 38 |
tgt[r] += w * src[n]; |
| 39 |
} |
| 40 |
} |
| 41 |
} |
| 42 |
|
| 43 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 44 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 1, MatrixValue, SourceValue, TargetValue>::multiply( |
| 45 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 1>& src, View<TargetValue, 1>& tgt, const Configuration&) { |
| 46 |
spmv_layout_left<true>(W, src, tgt); |
| 47 |
} |
| 48 |
|
| 49 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 50 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 1, MatrixValue, SourceValue, TargetValue>::multiply_add( |
| 51 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 1>& src, View<TargetValue, 1>& tgt, const Configuration&) { |
| 52 |
spmv_layout_left<false>(W, src, tgt); |
| 53 |
} |
| 54 |
|
| 55 |
template <bool SetZero, typename MatrixValue, typename SourceValue, typename TargetValue> |
| 56 |
void spmm_layout_left(const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 2>& src, View<TargetValue, 2>& tgt) { |
| 57 |
using Value = TargetValue; |
| 58 |
const auto outer = W.outer(); |
| 59 |
const auto index = W.inner(); |
| 60 |
const auto weight = W.value(); |
| 61 |
const idx_t rows = static_cast<idx_t>(W.rows()); |
| 62 |
const idx_t Nk = src.shape(1); |
| 63 |
|
| 64 |
ATLAS_ASSERT(src.shape(0) >= W.cols()); |
| 65 |
ATLAS_ASSERT(tgt.shape(0) >= W.rows()); |
| 66 |
|
| 67 |
atlas_omp_parallel_for(idx_t r = 0; r < rows; ++r) { |
| 68 |
if constexpr (SetZero) { |
| 69 |
for (idx_t k = 0; k < Nk; ++k) { |
| 70 |
tgt(r, k) = 0.; |
| 71 |
} |
| 72 |
} |
| 73 |
for (idx_t c = outer[r]; c < outer[r + 1]; ++c) { |
| 74 |
idx_t n = index[c]; |
| 75 |
Value w = static_cast<Value>(weight[c]); |
| 76 |
for (idx_t k = 0; k < Nk; ++k) { |
| 77 |
tgt(r, k) += w * src(n, k); |
| 78 |
} |
| 79 |
} |
| 80 |
} |
| 81 |
} |
| 82 |
|
| 83 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 84 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 2, MatrixValue, SourceValue, TargetValue>::multiply( |
| 85 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 2>& src, View<TargetValue, 2>& tgt, const Configuration&) { |
| 86 |
spmm_layout_left<true>(W, src, tgt); |
| 87 |
} |
| 88 |
|
| 89 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 90 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 2, MatrixValue, SourceValue, TargetValue>::multiply_add( |
| 91 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 2>& src, View<TargetValue, 2>& tgt, const Configuration&) { |
| 92 |
spmm_layout_left<false>(W, src, tgt); |
| 93 |
} |
| 94 |
|
| 95 |
template <bool SetZero,typename MatrixValue, typename SourceValue, typename TargetValue> |
| 96 |
void spmt_layout_left(const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 3>& src, View<TargetValue, 3>& tgt) { |
| 97 |
if (src.contiguous() && tgt.contiguous()) { |
| 98 |
// We can take a more optimized route by reducing rank |
| 99 |
auto src_v = View<SourceValue, 2>(src.data(), array::make_shape(src.shape(0), src.stride(0))); |
| 100 |
auto tgt_v = View<TargetValue, 2>(tgt.data(), array::make_shape(tgt.shape(0), tgt.stride(0))); |
| 101 |
spmm_layout_left<SetZero>(W, src_v, tgt_v); |
| 102 |
return; |
| 103 |
} |
| 104 |
using Value = TargetValue; |
| 105 |
const auto outer = W.outer(); |
| 106 |
const auto index = W.inner(); |
| 107 |
const auto weight = W.value(); |
| 108 |
const idx_t rows = static_cast<idx_t>(W.rows()); |
| 109 |
const idx_t Nk = src.shape(1); |
| 110 |
const idx_t Nl = src.shape(2); |
| 111 |
|
| 112 |
atlas_omp_parallel_for(idx_t r = 0; r < rows; ++r) { |
| 113 |
if constexpr (SetZero) { |
| 114 |
for (idx_t k = 0; k < Nk; ++k) { |
| 115 |
for (idx_t l = 0; l < Nl; ++l) { |
| 116 |
tgt(r, k, l) = 0.; |
| 117 |
} |
| 118 |
} |
| 119 |
} |
| 120 |
for (idx_t c = outer[r]; c < outer[r + 1]; ++c) { |
| 121 |
idx_t n = index[c]; |
| 122 |
const Value w = static_cast<Value>(weight[c]); |
| 123 |
for (idx_t k = 0; k < Nk; ++k) { |
| 124 |
for (idx_t l = 0; l < Nl; ++l) { |
| 125 |
tgt(r, k, l) += w * src(n, k, l); |
| 126 |
} |
| 127 |
} |
| 128 |
} |
| 129 |
} |
| 130 |
} |
| 131 |
|
| 132 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 133 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 3, MatrixValue, SourceValue, TargetValue>::multiply( |
| 134 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 3>& src, View<TargetValue, 3>& tgt, const Configuration& config) { |
| 135 |
spmt_layout_left<true>(W, src, tgt); |
| 136 |
} |
| 137 |
|
| 138 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 139 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 3, MatrixValue, SourceValue, TargetValue>::multiply_add( |
| 140 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 3>& src, View<TargetValue, 3>& tgt, const Configuration& config) { |
| 141 |
spmt_layout_left<false>(W, src, tgt); |
| 142 |
} |
| 143 |
|
| 144 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 145 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 1, MatrixValue, SourceValue, TargetValue>::multiply( |
| 146 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 1>& src, View<TargetValue, 1>& tgt, const Configuration& config) { |
| 147 |
return SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 1, MatrixValue, SourceValue, TargetValue>::multiply(W, src, tgt, config); |
| 148 |
} |
| 149 |
|
| 150 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 151 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 1, MatrixValue, SourceValue, TargetValue>::multiply_add( |
| 152 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 1>& src, View<TargetValue, 1>& tgt, const Configuration& config) { |
| 153 |
return SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 1, MatrixValue, SourceValue, TargetValue>::multiply_add(W, src, tgt, config); |
| 154 |
} |
| 155 |
|
| 156 |
template <bool SetZero, typename MatrixValue, typename SourceValue, typename TargetValue> |
| 157 |
void spmm_layout_right(const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 2>& src, View<TargetValue, 2>& tgt) { |
| 158 |
using Value = TargetValue; |
| 159 |
const auto outer = W.outer(); |
| 160 |
const auto index = W.inner(); |
| 161 |
const auto weight = W.value(); |
| 162 |
const idx_t rows = static_cast<idx_t>(W.rows()); |
| 163 |
const idx_t Nk = src.shape(0); |
| 164 |
|
| 165 |
ATLAS_ASSERT(src.shape(1) >= W.cols()); |
| 166 |
ATLAS_ASSERT(tgt.shape(1) >= W.rows()); |
| 167 |
|
| 168 |
atlas_omp_parallel_for(idx_t r = 0; r < rows; ++r) { |
| 169 |
if constexpr (SetZero) { |
| 170 |
for (idx_t k = 0; k < Nk; ++k) { |
| 171 |
tgt(k, r) = 0.; |
| 172 |
} |
| 173 |
} |
| 174 |
for (idx_t c = outer[r]; c < outer[r + 1]; ++c) { |
| 175 |
idx_t n = index[c]; |
| 176 |
Value w = static_cast<Value>(weight[c]); |
| 177 |
for (idx_t k = 0; k < Nk; ++k) { |
| 178 |
tgt(k, r) += w * src(k, n); |
| 179 |
} |
| 180 |
} |
| 181 |
} |
| 182 |
} |
| 183 |
|
| 184 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 185 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 2, MatrixValue, SourceValue, TargetValue>::multiply( |
| 186 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 2>& src, View<TargetValue, 2>& tgt, const Configuration&) { |
| 187 |
spmm_layout_right<true>(W, src, tgt); |
| 188 |
} |
| 189 |
|
| 190 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 191 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 2, MatrixValue, SourceValue, TargetValue>::multiply_add( |
| 192 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 2>& src, View<TargetValue, 2>& tgt, const Configuration&) { |
| 193 |
spmm_layout_right<false>(W, src, tgt); |
| 194 |
} |
| 195 |
|
| 196 |
template <bool SetZero, typename MatrixValue, typename SourceValue, typename TargetValue> |
| 197 |
void spmt_layout_right(const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 3>& src, View<TargetValue, 3>& tgt) { |
| 198 |
if (src.contiguous() && tgt.contiguous()) { |
| 199 |
// We can take a more optimized route by reducing rank |
| 200 |
auto src_v = View<SourceValue, 2>(src.data(), array::make_shape(src.shape(0), src.stride(0))); |
| 201 |
auto tgt_v = View<TargetValue, 2>(tgt.data(), array::make_shape(tgt.shape(0), tgt.stride(0))); |
| 202 |
spmm_layout_right<SetZero>(W, src_v, tgt_v); |
| 203 |
return; |
| 204 |
} |
| 205 |
using Value = TargetValue; |
| 206 |
const auto outer = W.outer(); |
| 207 |
const auto index = W.inner(); |
| 208 |
const auto weight = W.value(); |
| 209 |
const idx_t rows = static_cast<idx_t>(W.rows()); |
| 210 |
const idx_t Nk = src.shape(1); |
| 211 |
const idx_t Nl = src.shape(0); |
| 212 |
|
| 213 |
atlas_omp_parallel_for(idx_t r = 0; r < rows; ++r) { |
| 214 |
if constexpr (SetZero){ |
| 215 |
for (idx_t k = 0; k < Nk; ++k) { |
| 216 |
for (idx_t l = 0; l < Nl; ++l) { |
| 217 |
tgt(l, k, r) = 0.; |
| 218 |
} |
| 219 |
} |
| 220 |
} |
| 221 |
for (idx_t c = outer[r]; c < outer[r + 1]; ++c) { |
| 222 |
idx_t n = index[c]; |
| 223 |
const Value w = static_cast<Value>(weight[c]); |
| 224 |
for (idx_t k = 0; k < Nk; ++k) { |
| 225 |
for (idx_t l = 0; l < Nl; ++l) { |
| 226 |
tgt(l, k, r) += w * src(l, k, n); |
| 227 |
} |
| 228 |
} |
| 229 |
} |
| 230 |
} |
| 231 |
} |
| 232 |
|
| 233 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 234 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 3, MatrixValue, SourceValue, TargetValue>::multiply( |
| 235 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 3>& src, View<TargetValue, 3>& tgt, const Configuration& config) { |
| 236 |
spmt_layout_right<true>(W, src, tgt); |
| 237 |
} |
| 238 |
|
| 239 |
template <typename MatrixValue, typename SourceValue, typename TargetValue> |
| 240 |
void SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 3, MatrixValue, SourceValue, TargetValue>::multiply_add( |
| 241 |
const SparseMatrixView<MatrixValue>& W, const View<SourceValue, 3>& src, View<TargetValue, 3>& tgt, const Configuration& config) { |
| 242 |
spmt_layout_right<false>(W, src, tgt); |
| 243 |
} |
| 244 |
|
| 245 |
#define EXPLICIT_TEMPLATE_INSTANTIATION(TYPE) \ |
| 246 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 1, double, TYPE const, TYPE>; \ |
| 247 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 2, double, TYPE const, TYPE>; \ |
| 248 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 3, double, TYPE const, TYPE>; \ |
| 249 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 1, double, TYPE const, TYPE>; \ |
| 250 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 2, double, TYPE const, TYPE>; \ |
| 251 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 3, double, TYPE const, TYPE>; \ |
| 252 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 1, float, TYPE const, TYPE>; \ |
| 253 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 2, float, TYPE const, TYPE>; \ |
| 254 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_left, 3, float, TYPE const, TYPE>; \ |
| 255 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 1, float, TYPE const, TYPE>; \ |
| 256 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 2, float, TYPE const, TYPE>; \ |
| 257 |
template struct SparseMatrixMultiply<backend::openmp, Indexing::layout_right, 3, float, TYPE const, TYPE>; |
| 258 |
|
| 259 |
EXPLICIT_TEMPLATE_INSTANTIATION(double); |
| 260 |
EXPLICIT_TEMPLATE_INSTANTIATION(float); |
| 261 |
|
| 262 |
} // namespace sparse |
| 263 |
} // namespace linalg |
| 264 |
} // namespace atlas |