| 1 |
//@HEADER |
| 2 |
// ************************************************************************ |
| 3 |
// |
| 4 |
// Kokkos v. 4.0 |
| 5 |
// Copyright (2022) National Technology & Engineering |
| 6 |
// Solutions of Sandia, LLC (NTESS). |
| 7 |
// |
| 8 |
// Under the terms of Contract DE-NA0003525 with NTESS, |
| 9 |
// the U.S. Government retains certain rights in this software. |
| 10 |
// |
| 11 |
// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions. |
| 12 |
// See https://kokkos.org/LICENSE for license information. |
| 13 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 14 |
// |
| 15 |
//@HEADER |
| 16 |
#ifndef KOKKOSBATCHED_GESV_IMPL_HPP |
| 17 |
#define KOKKOSBATCHED_GESV_IMPL_HPP |
| 18 |
|
| 19 |
/// \author Kim Liegeois ([email protected]) |
| 20 |
|
| 21 |
#include "KokkosBatched_Util.hpp" |
| 22 |
#include <KokkosBatched_LU_Decl.hpp> |
| 23 |
#include "KokkosBatched_Trsm_Decl.hpp" |
| 24 |
#include "KokkosBatched_Copy_Decl.hpp" |
| 25 |
|
| 26 |
namespace KokkosBatched { |
| 27 |
|
| 28 |
struct SerialStaticPivoting { |
| 29 |
template <class MatrixType1, class MatrixType2, class VectorType1, class VectorType2> |
| 30 |
KOKKOS_INLINE_FUNCTION static int invoke(const MatrixType1 A, const MatrixType2 PDAD, const VectorType1 Y, |
| 31 |
const VectorType2 PDY, const VectorType2 D2, const VectorType2 tmp_v_1, |
| 32 |
const VectorType2 tmp_v_2); |
| 33 |
}; |
| 34 |
|
| 35 |
template <typename MemberType> |
| 36 |
struct TeamStaticPivoting { |
| 37 |
template <class MatrixType1, class MatrixType2, class VectorType1, class VectorType2> |
| 38 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType1 A, const MatrixType2 PDAD, |
| 39 |
const VectorType1 Y, const VectorType2 PDY, const VectorType2 D2, |
| 40 |
const VectorType2 tmp_v_1, const VectorType2 tmp_v_2); |
| 41 |
}; |
| 42 |
|
| 43 |
template <typename MemberType> |
| 44 |
struct TeamVectorStaticPivoting { |
| 45 |
template <class MatrixType1, class MatrixType2, class VectorType1, class VectorType2> |
| 46 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType1 A, const MatrixType2 PDAD, |
| 47 |
const VectorType1 Y, const VectorType2 PDY, const VectorType2 D2, |
| 48 |
const VectorType2 tmp_v_1, const VectorType2 tmp_v_2); |
| 49 |
}; |
| 50 |
|
| 51 |
template <class MatrixType1, class MatrixType2, class VectorType1, class VectorType2> |
| 52 |
KOKKOS_INLINE_FUNCTION int SerialStaticPivoting::invoke(const MatrixType1 A, const MatrixType2 PDAD, |
| 53 |
const VectorType1 Y, const VectorType2 PDY, |
| 54 |
const VectorType2 D2, const VectorType2 tmp_v_1, |
| 55 |
const VectorType2 tmp_v_2) { |
| 56 |
using value_type = typename MatrixType1::non_const_value_type; |
| 57 |
const size_t n = A.extent(0); |
| 58 |
|
| 59 |
// First, the algorithm loops over the rows and columns and search |
| 60 |
// for the maximal absolute value per row and column. |
| 61 |
for (size_t i = 0; i < n; ++i) { |
| 62 |
D2(i) = Kokkos::ArithTraits<value_type>::zero(); |
| 63 |
tmp_v_1(i) = 0; |
| 64 |
tmp_v_2(i) = 1.; |
| 65 |
for (size_t j = 0; j < n; ++j) { |
| 66 |
if (D2(i) < Kokkos::abs(A(j, i))) D2(i) = Kokkos::abs(A(j, i)); |
| 67 |
if (tmp_v_1(i) < Kokkos::abs(A(i, j))) tmp_v_1(i) = Kokkos::abs(A(i, j)); |
| 68 |
} |
| 69 |
D2(i) = 1. / D2(i); |
| 70 |
} |
| 71 |
|
| 72 |
// Then, the inverse of the maximal value per column is used to scale |
| 73 |
// A by the right. |
| 74 |
for (size_t i = 0; i < n; ++i) { |
| 75 |
for (size_t j = 0; j < n; ++j) { |
| 76 |
A(i, j) *= D2(j); |
| 77 |
} |
| 78 |
} |
| 79 |
|
| 80 |
// Once again, the algorithm loops over the rows and store the maximal |
| 81 |
// absolute value per row but after the right scalling and do a left scalling |
| 82 |
// of A and Y. |
| 83 |
value_type D1_i; |
| 84 |
for (size_t i = 0; i < n; ++i) { |
| 85 |
D1_i = Kokkos::ArithTraits<value_type>::zero(); |
| 86 |
for (size_t j = 0; j < n; ++j) { |
| 87 |
if (D1_i < Kokkos::abs(A(i, j))) D1_i = Kokkos::abs(A(i, j)); |
| 88 |
} |
| 89 |
D1_i = 1. / D1_i; |
| 90 |
for (size_t j = 0; j < n; ++j) { |
| 91 |
A(i, j) *= D1_i; |
| 92 |
} |
| 93 |
Y(i) *= D1_i; |
| 94 |
} |
| 95 |
|
| 96 |
// Finally, the algorithm starts to loop over the rows in an order such that |
| 97 |
// their initial maximal absolute value decrease (it uses the tmp_v_1 to do |
| 98 |
// so), then for a given row, it finds the available column with the largest |
| 99 |
// absolute value. If this value is zero, the algorithm failed to compute a |
| 100 |
// good pivot, otherwise it puts the current row to the found column index and |
| 101 |
// it labels the row and column index as unavailable and continue the loop |
| 102 |
// over the rows. |
| 103 |
// |
| 104 |
for (size_t i = 0; i < n; ++i) { |
| 105 |
int row_index = 0; |
| 106 |
int col_index = 0; |
| 107 |
value_type tmp_0 = Kokkos::ArithTraits<value_type>::zero(); |
| 108 |
value_type tmp_1 = Kokkos::ArithTraits<value_type>::zero(); |
| 109 |
for (size_t j = 0; j < n; ++j) { |
| 110 |
if (tmp_0 < tmp_v_1(j)) { |
| 111 |
tmp_0 = tmp_v_1(j); |
| 112 |
row_index = j; |
| 113 |
} |
| 114 |
} |
| 115 |
for (size_t j = 0; j < n; ++j) { |
| 116 |
if (tmp_1 < Kokkos::abs(A(row_index, j) * tmp_v_2(j))) { |
| 117 |
tmp_1 = Kokkos::abs(A(row_index, j) * tmp_v_2(j)); |
| 118 |
col_index = j; |
| 119 |
} |
| 120 |
} |
| 121 |
if (tmp_1 == Kokkos::ArithTraits<value_type>::zero()) return 1; |
| 122 |
tmp_v_1(row_index) = Kokkos::ArithTraits<value_type>::zero(); |
| 123 |
tmp_v_2(col_index) = Kokkos::ArithTraits<value_type>::zero(); |
| 124 |
|
| 125 |
for (size_t j = 0; j < n; ++j) { |
| 126 |
PDAD(col_index, j) = A(row_index, j); |
| 127 |
} |
| 128 |
PDY(col_index) = Y(row_index); |
| 129 |
} |
| 130 |
|
| 131 |
return 0; |
| 132 |
} |
| 133 |
|
| 134 |
template <typename MemberType> |
| 135 |
template <class MatrixType1, class MatrixType2, class VectorType1, class VectorType2> |
| 136 |
KOKKOS_INLINE_FUNCTION int TeamStaticPivoting<MemberType>::invoke(const MemberType &member, const MatrixType1 A, |
| 137 |
const MatrixType2 PDAD, const VectorType1 Y, |
| 138 |
const VectorType2 PDY, const VectorType2 D2, |
| 139 |
const VectorType2 tmp_v_1, |
| 140 |
const VectorType2 tmp_v_2) { |
| 141 |
using value_type = typename MatrixType1::non_const_value_type; |
| 142 |
using reducer_value_type = typename Kokkos::MaxLoc<value_type, int>::value_type; |
| 143 |
// This implementation follows the strategy of SerialStaticPivoting but uses |
| 144 |
// an extra level of parallelism. |
| 145 |
|
| 146 |
// Made this non-const in order to WORKAROUND issue #349 (Credit to C. Trott) |
| 147 |
size_t n = A.extent(0); |
| 148 |
|
| 149 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const int &i) { |
| 150 |
D2(i) = Kokkos::ArithTraits<value_type>::zero(); |
| 151 |
tmp_v_1(i) = 0; |
| 152 |
tmp_v_2(i) = 1.; |
| 153 |
for (size_t j = 0; j < n; ++j) { |
| 154 |
if (D2(i) < Kokkos::abs(A(j, i))) D2(i) = Kokkos::abs(A(j, i)); |
| 155 |
if (tmp_v_1(i) < Kokkos::abs(A(i, j))) tmp_v_1(i) = Kokkos::abs(A(i, j)); |
| 156 |
} |
| 157 |
D2(i) = 1. / D2(i); |
| 158 |
}); |
| 159 |
|
| 160 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const int &i) { |
| 161 |
for (size_t j = 0; j < n; ++j) { |
| 162 |
A(i, j) *= D2(j); |
| 163 |
} |
| 164 |
}); |
| 165 |
|
| 166 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const int &i) { |
| 167 |
value_type D1_i = Kokkos::ArithTraits<value_type>::zero(); |
| 168 |
for (size_t j = 0; j < n; ++j) { |
| 169 |
if (D1_i < Kokkos::abs(A(i, j))) D1_i = Kokkos::abs(A(i, j)); |
| 170 |
} |
| 171 |
D1_i = 1. / D1_i; |
| 172 |
for (size_t j = 0; j < n; ++j) { |
| 173 |
A(i, j) *= D1_i; |
| 174 |
} |
| 175 |
Y(i) *= D1_i; |
| 176 |
}); |
| 177 |
|
| 178 |
for (size_t i = 0; i < n; ++i) { |
| 179 |
int row_index, col_index; |
| 180 |
reducer_value_type value; |
| 181 |
Kokkos::MaxLoc<value_type, int> reducer_value(value); |
| 182 |
Kokkos::parallel_reduce( |
| 183 |
Kokkos::TeamThreadRange(member, n), |
| 184 |
[&](const int &j, reducer_value_type &update) { |
| 185 |
if (tmp_v_1(j) > update.val) { |
| 186 |
update.val = tmp_v_1(j); |
| 187 |
update.loc = j; |
| 188 |
} |
| 189 |
}, |
| 190 |
reducer_value); |
| 191 |
row_index = value.loc; |
| 192 |
value.loc = 0; |
| 193 |
value.val = Kokkos::ArithTraits<value_type>::zero(); |
| 194 |
Kokkos::parallel_reduce( |
| 195 |
Kokkos::TeamThreadRange(member, n), |
| 196 |
[&](const int &j, reducer_value_type &update) { |
| 197 |
if (Kokkos::abs(A(row_index, j) * tmp_v_2(j)) > update.val) { |
| 198 |
update.val = Kokkos::abs(A(row_index, j) * tmp_v_2(j)); |
| 199 |
update.loc = j; |
| 200 |
} |
| 201 |
}, |
| 202 |
reducer_value); |
| 203 |
col_index = value.loc; |
| 204 |
if (value.val == Kokkos::ArithTraits<value_type>::zero()) return 1; |
| 205 |
tmp_v_1(row_index) = Kokkos::ArithTraits<value_type>::zero(); |
| 206 |
tmp_v_2(col_index) = Kokkos::ArithTraits<value_type>::zero(); |
| 207 |
|
| 208 |
for (size_t j = 0; j < n; ++j) { |
| 209 |
PDAD(col_index, j) = A(row_index, j); |
| 210 |
} |
| 211 |
PDY(col_index) = Y(row_index); |
| 212 |
} |
| 213 |
return 0; |
| 214 |
} |
| 215 |
|
| 216 |
template <typename MemberType> |
| 217 |
template <class MatrixType1, class MatrixType2, class VectorType1, class VectorType2> |
| 218 |
KOKKOS_INLINE_FUNCTION int TeamVectorStaticPivoting<MemberType>::invoke(const MemberType &member, const MatrixType1 A, |
| 219 |
const MatrixType2 PDAD, const VectorType1 Y, |
| 220 |
const VectorType2 PDY, const VectorType2 D2, |
| 221 |
const VectorType2 tmp_v_1, |
| 222 |
const VectorType2 tmp_v_2) { |
| 223 |
using value_type = typename MatrixType1::non_const_value_type; |
| 224 |
using reducer_value_type = typename Kokkos::MaxLoc<value_type, int>::value_type; |
| 225 |
// This implementation follows the strategy of SerialStaticPivoting but uses |
| 226 |
// two extra levels of parallelism. |
| 227 |
|
| 228 |
const size_t n = A.extent(0); |
| 229 |
|
| 230 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const int &i) { |
| 231 |
D2(i) = Kokkos::ArithTraits<value_type>::zero(); |
| 232 |
tmp_v_1(i) = 0; |
| 233 |
tmp_v_2(i) = 1.; |
| 234 |
reducer_value_type value; |
| 235 |
Kokkos::MaxLoc<value_type, int> reducer_value(value); |
| 236 |
Kokkos::parallel_reduce( |
| 237 |
Kokkos::ThreadVectorRange(member, n), |
| 238 |
[&](const int &j, reducer_value_type &update) { |
| 239 |
if (Kokkos::abs(A(j, i)) > update.val) { |
| 240 |
update.val = Kokkos::abs(A(j, i)); |
| 241 |
update.loc = j; |
| 242 |
} |
| 243 |
}, |
| 244 |
reducer_value); |
| 245 |
D2(i) = 1. / value.val; |
| 246 |
Kokkos::parallel_reduce( |
| 247 |
Kokkos::ThreadVectorRange(member, n), |
| 248 |
[&](const int &j, reducer_value_type &update) { |
| 249 |
if (Kokkos::abs(A(i, j)) > update.val) { |
| 250 |
update.val = Kokkos::abs(A(i, j)); |
| 251 |
update.loc = j; |
| 252 |
} |
| 253 |
}, |
| 254 |
reducer_value); |
| 255 |
tmp_v_1(i) = value.val; |
| 256 |
}); |
| 257 |
|
| 258 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const int &i) { |
| 259 |
Kokkos::parallel_for(Kokkos::ThreadVectorRange(member, n), [&](const int &j) { A(i, j) *= D2(j); }); |
| 260 |
}); |
| 261 |
|
| 262 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const int &i) { |
| 263 |
value_type D1_i = Kokkos::ArithTraits<value_type>::zero(); |
| 264 |
reducer_value_type value; |
| 265 |
Kokkos::MaxLoc<value_type, int> reducer_value(value); |
| 266 |
Kokkos::parallel_reduce( |
| 267 |
Kokkos::ThreadVectorRange(member, n), |
| 268 |
[&](const int &j, reducer_value_type &update) { |
| 269 |
if (Kokkos::abs(A(i, j)) > update.val) { |
| 270 |
update.val = Kokkos::abs(A(i, j)); |
| 271 |
update.loc = j; |
| 272 |
} |
| 273 |
}, |
| 274 |
reducer_value); |
| 275 |
D1_i = 1. / value.val; |
| 276 |
Kokkos::parallel_for(Kokkos::ThreadVectorRange(member, n), [&](const int &j) { A(i, j) *= D1_i; }); |
| 277 |
Y(i) *= D1_i; |
| 278 |
}); |
| 279 |
|
| 280 |
for (size_t i = 0; i < n; ++i) { |
| 281 |
int row_index, col_index; |
| 282 |
reducer_value_type value{}; |
| 283 |
Kokkos::MaxLoc<value_type, int> reducer_value(value); |
| 284 |
Kokkos::parallel_reduce( |
| 285 |
Kokkos::TeamVectorRange(member, n), |
| 286 |
[&](const int &j, reducer_value_type &update) { |
| 287 |
if (tmp_v_1(j) > update.val) { |
| 288 |
update.val = tmp_v_1(j); |
| 289 |
update.loc = j; |
| 290 |
} |
| 291 |
}, |
| 292 |
reducer_value); |
| 293 |
row_index = value.loc; |
| 294 |
value.loc = 0; |
| 295 |
value.val = Kokkos::ArithTraits<value_type>::zero(); |
| 296 |
Kokkos::parallel_reduce( |
| 297 |
Kokkos::TeamVectorRange(member, n), |
| 298 |
[&](const int &j, reducer_value_type &update) { |
| 299 |
if (Kokkos::abs(A(row_index, j) * tmp_v_2(j)) > update.val) { |
| 300 |
update.val = Kokkos::abs(A(row_index, j) * tmp_v_2(j)); |
| 301 |
update.loc = j; |
| 302 |
} |
| 303 |
}, |
| 304 |
reducer_value); |
| 305 |
col_index = value.loc; |
| 306 |
if (value.val == Kokkos::ArithTraits<value_type>::zero()) return 1; |
| 307 |
tmp_v_1(row_index) = Kokkos::ArithTraits<value_type>::zero(); |
| 308 |
tmp_v_2(col_index) = Kokkos::ArithTraits<value_type>::zero(); |
| 309 |
|
| 310 |
Kokkos::parallel_for(Kokkos::TeamVectorRange(member, n), |
| 311 |
[&](const int &j) { PDAD(col_index, j) = A(row_index, j); }); |
| 312 |
PDY(col_index) = Y(row_index); |
| 313 |
} |
| 314 |
return 0; |
| 315 |
} |
| 316 |
|
| 317 |
template <class VectorType1, class VectorType2, class VectorType3> |
| 318 |
KOKKOS_INLINE_FUNCTION void SerialHadamard1D(const VectorType1 X, const VectorType2 D, const VectorType3 DX) { |
| 319 |
const size_t n = X.extent(0); |
| 320 |
|
| 321 |
for (size_t i = 0; i < n; ++i) { |
| 322 |
DX(i) = D(i) * X(i); |
| 323 |
} |
| 324 |
} |
| 325 |
|
| 326 |
template <typename MemberType, class VectorType1, class VectorType2, class VectorType3> |
| 327 |
KOKKOS_INLINE_FUNCTION void TeamHadamard1D(const MemberType &member, const VectorType1 X, const VectorType2 D, |
| 328 |
const VectorType3 DX) { |
| 329 |
const size_t n = X.extent(0); |
| 330 |
|
| 331 |
Kokkos::parallel_for(Kokkos::TeamThreadRange(member, n), [&](const size_t &i) { DX(i) = D(i) * X(i); }); |
| 332 |
} |
| 333 |
|
| 334 |
template <typename MemberType, class VectorType1, class VectorType2, class VectorType3> |
| 335 |
KOKKOS_INLINE_FUNCTION void TeamVectorHadamard1D(const MemberType &member, const VectorType1 X, const VectorType2 D, |
| 336 |
const VectorType3 DX) { |
| 337 |
const size_t n = X.extent(0); |
| 338 |
|
| 339 |
Kokkos::parallel_for(Kokkos::TeamVectorRange(member, n), [&](const size_t &i) { DX(i) = D(i) * X(i); }); |
| 340 |
} |
| 341 |
|
| 342 |
/// |
| 343 |
/// Serial Impl |
| 344 |
/// =========== |
| 345 |
template <> |
| 346 |
struct SerialGesv<Gesv::StaticPivoting> { |
| 347 |
template <typename MatrixType, typename XVectorType, typename YVectorType> |
| 348 |
KOKKOS_INLINE_FUNCTION static int invoke(const MatrixType A, const XVectorType X, const YVectorType Y, |
| 349 |
const MatrixType tmp) { |
| 350 |
#if (KOKKOSKERNELS_DEBUG_LEVEL > 0) |
| 351 |
static_assert(Kokkos::is_view<MatrixType>::value, "KokkosBatched::gesv: MatrixType is not a Kokkos::View."); |
| 352 |
static_assert(Kokkos::is_view<XVectorType>::value, "KokkosBatched::gesv: XVectorType is not a Kokkos::View."); |
| 353 |
static_assert(Kokkos::is_view<YVectorType>::value, "KokkosBatched::gesv: YVectorType is not a Kokkos::View."); |
| 354 |
static_assert(MatrixType::rank == 2, "KokkosBatched::gesv: MatrixType must have rank 2."); |
| 355 |
static_assert(XVectorType::rank == 1, "KokkosBatched::gesv: XVectorType must have rank 1."); |
| 356 |
static_assert(YVectorType::rank == 1, "KokkosBatched::gesv: YVectorType must have rank 1."); |
| 357 |
|
| 358 |
// Check compatibility of dimensions at run time. |
| 359 |
|
| 360 |
if (A.extent(0) != tmp.extent(0) || A.extent(1) + 4 != tmp.extent(1)) { |
| 361 |
Kokkos::printf( |
| 362 |
"KokkosBatched::gesv: dimensions of A and tmp do not match: A: " |
| 363 |
"%d x %d, tmp (note: its second dimension should be the second " |
| 364 |
"dimension of A + 4): %d x %d\n", |
| 365 |
(int)A.extent(0), (int)A.extent(1), (int)tmp.extent(0), (int)tmp.extent(1)); |
| 366 |
return 1; |
| 367 |
} |
| 368 |
|
| 369 |
if (A.extent(0) != X.extent(0) || A.extent(1) != X.extent(0) || A.extent(0) != Y.extent(0)) { |
| 370 |
Kokkos::printf( |
| 371 |
"KokkosBatched::gesv: dimensions of A and X and Y do not match: A: " |
| 372 |
"%d x %d, X: %d, Y: %d\n", |
| 373 |
(int)A.extent(0), (int)A.extent(1), (int)X.extent(0), (int)Y.extent(0)); |
| 374 |
return 1; |
| 375 |
} |
| 376 |
#endif |
| 377 |
|
| 378 |
const int n = A.extent(0); |
| 379 |
|
| 380 |
auto PDAD = Kokkos::subview(tmp, Kokkos::ALL, Kokkos::make_pair(0, n)); |
| 381 |
auto PDY = Kokkos::subview(tmp, Kokkos::ALL, n); |
| 382 |
auto D2 = Kokkos::subview(tmp, Kokkos::ALL, n + 1); |
| 383 |
auto tmp_v_1 = Kokkos::subview(tmp, Kokkos::ALL, n + 2); |
| 384 |
auto tmp_v_2 = Kokkos::subview(tmp, Kokkos::ALL, n + 3); |
| 385 |
|
| 386 |
if (SerialStaticPivoting::invoke(A, PDAD, Y, PDY, D2, tmp_v_1, tmp_v_2) == 1) { |
| 387 |
Kokkos::printf( |
| 388 |
"KokkosBatched::gesv: the currently implemented static pivoting " |
| 389 |
"failed.\n"); |
| 390 |
return 1; |
| 391 |
} |
| 392 |
|
| 393 |
int r_val = SerialLU<Algo::Level3::Unblocked>::invoke(PDAD); |
| 394 |
|
| 395 |
if (r_val == 0) |
| 396 |
r_val = SerialTrsm<Side::Left, Uplo::Lower, Trans::NoTranspose, Diag::Unit, Algo::Level3::Unblocked>::invoke( |
| 397 |
1.0, PDAD, PDY); |
| 398 |
|
| 399 |
if (r_val == 0) |
| 400 |
r_val = SerialTrsm<Side::Left, Uplo::Upper, Trans::NoTranspose, Diag::NonUnit, Algo::Level3::Unblocked>::invoke( |
| 401 |
1.0, PDAD, PDY); |
| 402 |
|
| 403 |
if (r_val == 0) SerialHadamard1D(PDY, D2, X); |
| 404 |
return r_val; |
| 405 |
} |
| 406 |
}; |
| 407 |
|
| 408 |
template <> |
| 409 |
struct SerialGesv<Gesv::NoPivoting> { |
| 410 |
template <typename MatrixType, typename XVectorType, typename YVectorType> |
| 411 |
KOKKOS_INLINE_FUNCTION static int invoke(const MatrixType A, const XVectorType X, const YVectorType Y, |
| 412 |
const MatrixType /*tmp*/) { |
| 413 |
#if (KOKKOSKERNELS_DEBUG_LEVEL > 0) |
| 414 |
static_assert(Kokkos::is_view<MatrixType>::value, "KokkosBatched::gesv: MatrixType is not a Kokkos::View."); |
| 415 |
static_assert(Kokkos::is_view<XVectorType>::value, "KokkosBatched::gesv: XVectorType is not a Kokkos::View."); |
| 416 |
static_assert(Kokkos::is_view<YVectorType>::value, "KokkosBatched::gesv: YVectorType is not a Kokkos::View."); |
| 417 |
static_assert(MatrixType::rank == 2, "KokkosBatched::gesv: MatrixType must have rank 2."); |
| 418 |
static_assert(XVectorType::rank == 1, "KokkosBatched::gesv: XVectorType must have rank 1."); |
| 419 |
static_assert(YVectorType::rank == 1, "KokkosBatched::gesv: YVectorType must have rank 1."); |
| 420 |
|
| 421 |
// Check compatibility of dimensions at run time. |
| 422 |
|
| 423 |
if (A.extent(0) != X.extent(0) || A.extent(1) != X.extent(0) || A.extent(0) != Y.extent(0)) { |
| 424 |
Kokkos::printf( |
| 425 |
"KokkosBatched::gesv: dimensions of A and X and Y do not match: A: " |
| 426 |
"%d x %d, X: %d, Y: %d\n", |
| 427 |
(int)A.extent(0), (int)A.extent(1), (int)X.extent(0), (int)Y.extent(0)); |
| 428 |
return 1; |
| 429 |
} |
| 430 |
#endif |
| 431 |
|
| 432 |
int r_val = SerialLU<Algo::Level3::Unblocked>::invoke(A); |
| 433 |
|
| 434 |
if (r_val == 0) r_val = SerialCopy<Trans::NoTranspose, 1>::invoke(Y, X); |
| 435 |
|
| 436 |
if (r_val == 0) |
| 437 |
r_val = SerialTrsm<Side::Left, Uplo::Lower, Trans::NoTranspose, Diag::Unit, Algo::Level3::Unblocked>::invoke( |
| 438 |
1.0, A, X); |
| 439 |
|
| 440 |
if (r_val == 0) |
| 441 |
r_val = SerialTrsm<Side::Left, Uplo::Upper, Trans::NoTranspose, Diag::NonUnit, Algo::Level3::Unblocked>::invoke( |
| 442 |
1.0, A, X); |
| 443 |
|
| 444 |
return r_val; |
| 445 |
} |
| 446 |
}; |
| 447 |
|
| 448 |
/// |
| 449 |
/// Team Impl |
| 450 |
/// ========= |
| 451 |
|
| 452 |
template <typename MemberType> |
| 453 |
struct TeamGesv<MemberType, Gesv::StaticPivoting> { |
| 454 |
template <typename MatrixType, typename VectorType> |
| 455 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType A, const VectorType X, |
| 456 |
const VectorType Y) { |
| 457 |
#if (KOKKOSKERNELS_DEBUG_LEVEL > 0) |
| 458 |
static_assert(Kokkos::is_view<MatrixType>::value, "KokkosBatched::gesv: MatrixType is not a Kokkos::View."); |
| 459 |
static_assert(Kokkos::is_view<VectorType>::value, "KokkosBatched::gesv: VectorType is not a Kokkos::View."); |
| 460 |
static_assert(MatrixType::rank == 2, "KokkosBatched::gesv: MatrixType must have rank 2."); |
| 461 |
static_assert(VectorType::rank == 1, "KokkosBatched::gesv: VectorType must have rank 1."); |
| 462 |
|
| 463 |
// Check compatibility of dimensions at run time. |
| 464 |
if (A.extent(0) != X.extent(0) || A.extent(1) != X.extent(0) || A.extent(0) != Y.extent(0)) { |
| 465 |
Kokkos::printf( |
| 466 |
"KokkosBatched::gesv: dimensions of A and X and Y do not match: A: " |
| 467 |
"%d x %d, X: %d, Y: %d\n", |
| 468 |
(int)A.extent(0), (int)A.extent(1), (int)X.extent(0), (int)Y.extent(0)); |
| 469 |
return 1; |
| 470 |
} |
| 471 |
#endif |
| 472 |
using ScratchPadMatrixViewType = Kokkos::View<typename MatrixType::non_const_value_type **, |
| 473 |
typename MatrixType::execution_space::scratch_memory_space>; |
| 474 |
|
| 475 |
const int n = A.extent(0); |
| 476 |
|
| 477 |
ScratchPadMatrixViewType tmp(member.team_scratch(0), n, n + 4); |
| 478 |
auto PDAD = Kokkos::subview(tmp, Kokkos::ALL, Kokkos::make_pair(0, n)); |
| 479 |
auto PDY = Kokkos::subview(tmp, Kokkos::ALL, n); |
| 480 |
auto D2 = Kokkos::subview(tmp, Kokkos::ALL, n + 1); |
| 481 |
auto tmp_v_1 = Kokkos::subview(tmp, Kokkos::ALL, n + 2); |
| 482 |
auto tmp_v_2 = Kokkos::subview(tmp, Kokkos::ALL, n + 3); |
| 483 |
|
| 484 |
if (TeamStaticPivoting<MemberType>::invoke(member, A, PDAD, Y, PDY, D2, tmp_v_1, tmp_v_2) == 1) { |
| 485 |
Kokkos::printf( |
| 486 |
"KokkosBatched::gesv: the currently implemented static pivoting " |
| 487 |
"failed.\n"); |
| 488 |
return 1; |
| 489 |
} |
| 490 |
member.team_barrier(); |
| 491 |
|
| 492 |
int r_val = TeamLU<MemberType, Algo::Level3::Unblocked>::invoke(member, PDAD); |
| 493 |
member.team_barrier(); |
| 494 |
|
| 495 |
if (r_val == 0) { |
| 496 |
r_val = TeamTrsm<MemberType, Side::Left, Uplo::Lower, Trans::NoTranspose, Diag::Unit, |
| 497 |
Algo::Level3::Unblocked>::invoke(member, 1.0, PDAD, PDY); |
| 498 |
member.team_barrier(); |
| 499 |
} |
| 500 |
|
| 501 |
if (r_val == 0) { |
| 502 |
r_val = TeamTrsm<MemberType, Side::Left, Uplo::Upper, Trans::NoTranspose, Diag::NonUnit, |
| 503 |
Algo::Level3::Unblocked>::invoke(member, 1.0, PDAD, PDY); |
| 504 |
member.team_barrier(); |
| 505 |
} |
| 506 |
|
| 507 |
if (r_val == 0) { |
| 508 |
TeamHadamard1D(member, PDY, D2, X); |
| 509 |
member.team_barrier(); |
| 510 |
} |
| 511 |
|
| 512 |
return r_val; |
| 513 |
} |
| 514 |
}; |
| 515 |
|
| 516 |
template <typename MemberType> |
| 517 |
struct TeamGesv<MemberType, Gesv::NoPivoting> { |
| 518 |
template <typename MatrixType, typename VectorType> |
| 519 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType A, const VectorType X, |
| 520 |
const VectorType Y) { |
| 521 |
#if (KOKKOSKERNELS_DEBUG_LEVEL > 0) |
| 522 |
static_assert(Kokkos::is_view<MatrixType>::value, "KokkosBatched::gesv: MatrixType is not a Kokkos::View."); |
| 523 |
static_assert(Kokkos::is_view<VectorType>::value, "KokkosBatched::gesv: VectorType is not a Kokkos::View."); |
| 524 |
static_assert(MatrixType::rank == 2, "KokkosBatched::gesv: MatrixType must have rank 2."); |
| 525 |
static_assert(VectorType::rank == 1, "KokkosBatched::gesv: VectorType must have rank 1."); |
| 526 |
|
| 527 |
// Check compatibility of dimensions at run time. |
| 528 |
if (A.extent(0) != X.extent(0) || A.extent(1) != X.extent(0) || A.extent(0) != Y.extent(0)) { |
| 529 |
Kokkos::printf( |
| 530 |
"KokkosBatched::gesv: dimensions of A and X and Y do not match: A: " |
| 531 |
"%d x %d, X: %d, Y: %d\n", |
| 532 |
(int)A.extent(0), (int)A.extent(1), (int)X.extent(0), (int)Y.extent(0)); |
| 533 |
return 1; |
| 534 |
} |
| 535 |
#endif |
| 536 |
|
| 537 |
int r_val = TeamLU<MemberType, Algo::Level3::Unblocked>::invoke(member, A); |
| 538 |
member.team_barrier(); |
| 539 |
|
| 540 |
if (r_val == 0) { |
| 541 |
TeamCopy<MemberType, Trans::NoTranspose, 1>::invoke(member, Y, X); |
| 542 |
member.team_barrier(); |
| 543 |
} |
| 544 |
|
| 545 |
if (r_val == 0) { |
| 546 |
TeamTrsm<MemberType, Side::Left, Uplo::Lower, Trans::NoTranspose, Diag::Unit, Algo::Level3::Unblocked>::invoke( |
| 547 |
member, 1.0, A, X); |
| 548 |
member.team_barrier(); |
| 549 |
} |
| 550 |
|
| 551 |
if (r_val == 0) { |
| 552 |
TeamTrsm<MemberType, Side::Left, Uplo::Upper, Trans::NoTranspose, Diag::NonUnit, Algo::Level3::Unblocked>::invoke( |
| 553 |
member, 1.0, A, X); |
| 554 |
member.team_barrier(); |
| 555 |
} |
| 556 |
|
| 557 |
return r_val; |
| 558 |
} |
| 559 |
}; |
| 560 |
|
| 561 |
/// |
| 562 |
/// TeamVector Impl |
| 563 |
/// ========= |
| 564 |
|
| 565 |
template <typename MemberType> |
| 566 |
struct TeamVectorGesv<MemberType, Gesv::StaticPivoting> { |
| 567 |
template <typename MatrixType, typename VectorType> |
| 568 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType A, const VectorType X, |
| 569 |
const VectorType Y) { |
| 570 |
#if (KOKKOSKERNELS_DEBUG_LEVEL > 0) |
| 571 |
static_assert(Kokkos::is_view<MatrixType>::value, "KokkosBatched::gesv: MatrixType is not a Kokkos::View."); |
| 572 |
static_assert(Kokkos::is_view<VectorType>::value, "KokkosBatched::gesv: VectorType is not a Kokkos::View."); |
| 573 |
static_assert(MatrixType::rank == 2, "KokkosBatched::gesv: MatrixType must have rank 2."); |
| 574 |
static_assert(VectorType::rank == 1, "KokkosBatched::gesv: VectorType must have rank 1."); |
| 575 |
|
| 576 |
// Check compatibility of dimensions at run time. |
| 577 |
if (A.extent(0) != X.extent(0) || A.extent(1) != X.extent(0) || A.extent(0) != Y.extent(0)) { |
| 578 |
Kokkos::printf( |
| 579 |
"KokkosBatched::gesv: dimensions of A and X and Y do not match: A: " |
| 580 |
"%d x %d, X: %d, Y: %d\n", |
| 581 |
(int)A.extent(0), (int)A.extent(1), (int)X.extent(0), (int)Y.extent(0)); |
| 582 |
return 1; |
| 583 |
} |
| 584 |
#endif |
| 585 |
using ScratchPadMatrixViewType = Kokkos::View<typename MatrixType::non_const_value_type **, |
| 586 |
typename MatrixType::execution_space::scratch_memory_space>; |
| 587 |
|
| 588 |
const int n = A.extent(0); |
| 589 |
|
| 590 |
ScratchPadMatrixViewType tmp(member.team_scratch(0), n, n + 4); |
| 591 |
auto PDAD = Kokkos::subview(tmp, Kokkos::ALL, Kokkos::make_pair(0, n)); |
| 592 |
auto PDY = Kokkos::subview(tmp, Kokkos::ALL, n); |
| 593 |
auto D2 = Kokkos::subview(tmp, Kokkos::ALL, n + 1); |
| 594 |
auto tmp_v_1 = Kokkos::subview(tmp, Kokkos::ALL, n + 2); |
| 595 |
auto tmp_v_2 = Kokkos::subview(tmp, Kokkos::ALL, n + 3); |
| 596 |
|
| 597 |
if (TeamVectorStaticPivoting<MemberType>::invoke(member, A, PDAD, Y, PDY, D2, tmp_v_1, tmp_v_2) == 1) { |
| 598 |
Kokkos::printf( |
| 599 |
"KokkosBatched::gesv: the currently implemented static pivoting " |
| 600 |
"failed.\n"); |
| 601 |
return 1; |
| 602 |
} |
| 603 |
|
| 604 |
member.team_barrier(); |
| 605 |
|
| 606 |
int r_val = TeamLU<MemberType, Algo::Level3::Unblocked>::invoke(member, PDAD); |
| 607 |
member.team_barrier(); |
| 608 |
|
| 609 |
if (r_val == 0) { |
| 610 |
TeamVectorTrsm<MemberType, Side::Left, Uplo::Lower, Trans::NoTranspose, Diag::Unit, |
| 611 |
Algo::Level3::Unblocked>::invoke(member, 1.0, PDAD, PDY); |
| 612 |
member.team_barrier(); |
| 613 |
} |
| 614 |
|
| 615 |
if (r_val == 0) { |
| 616 |
TeamVectorTrsm<MemberType, Side::Left, Uplo::Upper, Trans::NoTranspose, Diag::NonUnit, |
| 617 |
Algo::Level3::Unblocked>::invoke(member, 1.0, PDAD, PDY); |
| 618 |
member.team_barrier(); |
| 619 |
} |
| 620 |
|
| 621 |
if (r_val == 0) { |
| 622 |
TeamVectorHadamard1D(member, PDY, D2, X); |
| 623 |
member.team_barrier(); |
| 624 |
} |
| 625 |
|
| 626 |
return r_val; |
| 627 |
} |
| 628 |
}; |
| 629 |
|
| 630 |
template <typename MemberType> |
| 631 |
struct TeamVectorGesv<MemberType, Gesv::NoPivoting> { |
| 632 |
template <typename MatrixType, typename VectorType> |
| 633 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType A, const VectorType X, |
| 634 |
const VectorType Y) { |
| 635 |
#if (KOKKOSKERNELS_DEBUG_LEVEL > 0) |
| 636 |
static_assert(Kokkos::is_view<MatrixType>::value, "KokkosBatched::gesv: MatrixType is not a Kokkos::View."); |
| 637 |
static_assert(Kokkos::is_view<VectorType>::value, "KokkosBatched::gesv: VectorType is not a Kokkos::View."); |
| 638 |
static_assert(MatrixType::rank == 2, "KokkosBatched::gesv: MatrixType must have rank 2."); |
| 639 |
static_assert(VectorType::rank == 1, "KokkosBatched::gesv: VectorType must have rank 1."); |
| 640 |
|
| 641 |
// Check compatibility of dimensions at run time. |
| 642 |
if (A.extent(0) != X.extent(0) || A.extent(1) != X.extent(0) || A.extent(0) != Y.extent(0)) { |
| 643 |
Kokkos::printf( |
| 644 |
"KokkosBatched::gesv: dimensions of A and X and Y do not match: A: " |
| 645 |
"%d x %d, X: %d, Y: %d\n", |
| 646 |
(int)A.extent(0), (int)A.extent(1), (int)X.extent(0), (int)Y.extent(0)); |
| 647 |
return 1; |
| 648 |
} |
| 649 |
#endif |
| 650 |
|
| 651 |
int r_val = TeamLU<MemberType, Algo::Level3::Unblocked>::invoke(member, A); |
| 652 |
member.team_barrier(); |
| 653 |
|
| 654 |
if (r_val == 0) { |
| 655 |
TeamVectorCopy<MemberType, Trans::NoTranspose, 1>::invoke(member, Y, X); |
| 656 |
member.team_barrier(); |
| 657 |
} |
| 658 |
|
| 659 |
if (r_val == 0) { |
| 660 |
TeamVectorTrsm<MemberType, Side::Left, Uplo::Lower, Trans::NoTranspose, Diag::Unit, |
| 661 |
Algo::Level3::Unblocked>::invoke(member, 1.0, A, X); |
| 662 |
member.team_barrier(); |
| 663 |
} |
| 664 |
|
| 665 |
if (r_val == 0) { |
| 666 |
TeamVectorTrsm<MemberType, Side::Left, Uplo::Upper, Trans::NoTranspose, Diag::NonUnit, |
| 667 |
Algo::Level3::Unblocked>::invoke(member, 1.0, A, X); |
| 668 |
member.team_barrier(); |
| 669 |
} |
| 670 |
|
| 671 |
return r_val; |
| 672 |
} |
| 673 |
}; |
| 674 |
|
| 675 |
} // namespace KokkosBatched |
| 676 |
|
| 677 |
#endif |