| 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 TEST_BATCHED_DENSE_HELPER_HPP |
| 17 |
#define TEST_BATCHED_DENSE_HELPER_HPP |
| 18 |
|
| 19 |
#include "KokkosBatched_Util.hpp" |
| 20 |
|
| 21 |
namespace KokkosBatched { |
| 22 |
template <typename MatrixViewType, typename VectorViewType> |
| 23 |
void create_tridiagonal_batched_matrices(const MatrixViewType& A, const VectorViewType& B) { |
| 24 |
Kokkos::Random_XorShift64_Pool<typename VectorViewType::device_type::execution_space> random(13718); |
| 25 |
Kokkos::fill_random(B, random, Kokkos::reduction_identity<typename VectorViewType::value_type>::prod()); |
| 26 |
|
| 27 |
auto A_host = Kokkos::create_mirror_view(A); |
| 28 |
|
| 29 |
const int N = A.extent(0); |
| 30 |
const int BlkSize = A.extent(1); |
| 31 |
|
| 32 |
for (int l = 0; l < N; ++l) { |
| 33 |
for (int i = 0; i < BlkSize; ++i) { |
| 34 |
for (int j = i; j < BlkSize; ++j) { |
| 35 |
if (i == j) |
| 36 |
A_host(l, i, j) = typename VectorViewType::value_type(2.0); |
| 37 |
else if (i == j - 1) { |
| 38 |
A_host(l, i, j) = typename VectorViewType::value_type(-1.0); |
| 39 |
A_host(l, j, i) = typename VectorViewType::value_type(-1.0); |
| 40 |
} else { |
| 41 |
A_host(l, i, j) = typename VectorViewType::value_type(0.0); |
| 42 |
A_host(l, j, i) = typename VectorViewType::value_type(0.0); |
| 43 |
} |
| 44 |
} |
| 45 |
} |
| 46 |
} |
| 47 |
|
| 48 |
Kokkos::fence(); |
| 49 |
|
| 50 |
Kokkos::deep_copy(A, A_host); |
| 51 |
|
| 52 |
Kokkos::fence(); |
| 53 |
} |
| 54 |
|
| 55 |
template <typename InViewType, typename OutViewType, typename UploType> |
| 56 |
void create_banded_triangular_matrix(InViewType& in, OutViewType& out, int k = 1, bool band_storage = true) { |
| 57 |
auto h_in = Kokkos::create_mirror_view(in); |
| 58 |
auto h_out = Kokkos::create_mirror_view(out); |
| 59 |
const int N = in.extent(0), BlkSize = in.extent(1); |
| 60 |
|
| 61 |
Kokkos::deep_copy(h_in, in); |
| 62 |
if (band_storage) { |
| 63 |
assert(out.extent(0) == in.extent(0)); |
| 64 |
assert(out.extent(1) == static_cast<std::size_t>(k + 1)); |
| 65 |
assert(out.extent(2) == in.extent(2)); |
| 66 |
if constexpr (std::is_same_v<UploType, KokkosBatched::Uplo::Upper>) { |
| 67 |
for (int i0 = 0; i0 < N; i0++) { |
| 68 |
for (int i1 = 0; i1 < k + 1; i1++) { |
| 69 |
for (int i2 = i1; i2 < BlkSize; i2++) { |
| 70 |
h_out(i0, k - i1, i2) = h_in(i0, i2 - i1, i2); |
| 71 |
} |
| 72 |
} |
| 73 |
} |
| 74 |
} else { |
| 75 |
for (int i0 = 0; i0 < N; i0++) { |
| 76 |
for (int i1 = 0; i1 < k + 1; i1++) { |
| 77 |
for (int i2 = 0; i2 < BlkSize - i1; i2++) { |
| 78 |
h_out(i0, i1, i2) = h_in(i0, i2 + i1, i2); |
| 79 |
} |
| 80 |
} |
| 81 |
} |
| 82 |
} |
| 83 |
} else { |
| 84 |
for (std::size_t i = 0; i < InViewType::rank(); i++) { |
| 85 |
assert(out.extent(i) == in.extent(i)); |
| 86 |
} |
| 87 |
|
| 88 |
if constexpr (std::is_same_v<UploType, KokkosBatched::Uplo::Upper>) { |
| 89 |
for (int i0 = 0; i0 < N; i0++) { |
| 90 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 91 |
for (int i2 = i1; i2 < Kokkos::min(i1 + k + 1, BlkSize); i2++) { |
| 92 |
h_out(i0, i1, i2) = h_in(i0, i1, i2); |
| 93 |
} |
| 94 |
} |
| 95 |
} |
| 96 |
} else { |
| 97 |
for (int i0 = 0; i0 < N; i0++) { |
| 98 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 99 |
for (int i2 = Kokkos::max(0, i1 - k); i2 <= i1; i2++) { |
| 100 |
h_out(i0, i1, i2) = h_in(i0, i1, i2); |
| 101 |
} |
| 102 |
} |
| 103 |
} |
| 104 |
} |
| 105 |
} |
| 106 |
Kokkos::deep_copy(out, h_out); |
| 107 |
} |
| 108 |
|
| 109 |
/// \brief Create a diagonal matrix from an input vector: |
| 110 |
/// Copies the input vector into the diagonal of the output matrix specified |
| 111 |
/// by the parameter k. k > 0 means that the matrix is upper-diagonal and |
| 112 |
/// k < 0 means the lower-diagonal. k = 0 means the diagonal. |
| 113 |
/// |
| 114 |
/// \tparam InViewType: Input type for the vector, needs to be a 2D view |
| 115 |
/// \tparam OutViewType: Output type for the matrix, needs to be a 3D view |
| 116 |
/// |
| 117 |
/// \param in [in]: Input batched vector, a rank 2 view |
| 118 |
/// \param out [out]: Output batched matrix, where the diagonal compnent |
| 119 |
/// specified by k is filled with the input vector, a rank 3 view |
| 120 |
/// \param k [in]: The diagonal offset to be filled (default is 0). |
| 121 |
/// |
| 122 |
template <typename InViewType, typename OutViewType> |
| 123 |
void create_diagonal_matrix(InViewType& in, OutViewType& out, int k = 0) { |
| 124 |
auto h_in = Kokkos::create_mirror_view(in); |
| 125 |
auto h_out = Kokkos::create_mirror_view(out); |
| 126 |
const int N = in.extent(0), BlkSize = in.extent(1); |
| 127 |
|
| 128 |
assert(out.extent(0) == in.extent(0)); |
| 129 |
assert(out.extent(1) == in.extent(1) + abs(k)); |
| 130 |
|
| 131 |
int i1_start = k >= 0 ? 0 : -k; |
| 132 |
int i2_start = k >= 0 ? k : 0; |
| 133 |
|
| 134 |
// Zero clear the output matrix |
| 135 |
using ScalarType = typename OutViewType::non_const_value_type; |
| 136 |
Kokkos::deep_copy(h_out, ScalarType(0.0)); |
| 137 |
|
| 138 |
Kokkos::deep_copy(h_in, in); |
| 139 |
for (int i0 = 0; i0 < N; i0++) { |
| 140 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 141 |
h_out(i0, i1 + i1_start, i1 + i2_start) = h_in(i0, i1); |
| 142 |
} |
| 143 |
} |
| 144 |
|
| 145 |
Kokkos::deep_copy(out, h_out); |
| 146 |
} |
| 147 |
|
| 148 |
/// \brief Creates a positive definite symmetric (PDS) matrix. |
| 149 |
/// Takes a full random matrix and converts it to a full pds matrix. |
| 150 |
/// |
| 151 |
/// \tparam InViewType: Input type for the matrix, needs to be a 3D view |
| 152 |
/// \tparam OutViewType: Output type for the matrix, needs to be a 3D view |
| 153 |
/// |
| 154 |
/// \param in [in]: Input batched banded matrix, a rank 3 view |
| 155 |
/// \param out [out]: Output batched full matrix, a rank 3 view |
| 156 |
/// |
| 157 |
template <typename InViewType, typename OutViewType> |
| 158 |
void random_to_pds(InViewType& in, OutViewType& out) { |
| 159 |
auto h_in = Kokkos::create_mirror_view(in); |
| 160 |
auto h_out = Kokkos::create_mirror_view(out); |
| 161 |
const int N = in.extent(0), BlkSize = in.extent(1); |
| 162 |
using value_type = typename InViewType::non_const_value_type; |
| 163 |
|
| 164 |
for (std::size_t i = 0; i < InViewType::rank(); i++) { |
| 165 |
assert(out.extent(i) == in.extent(i)); |
| 166 |
} |
| 167 |
|
| 168 |
Kokkos::deep_copy(h_in, in); |
| 169 |
Kokkos::deep_copy(h_out, 0.0); |
| 170 |
|
| 171 |
for (int i0 = 0; i0 < N; i0++) { |
| 172 |
// Make a hermitian matrix |
| 173 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 174 |
for (int i2 = i1; i2 < BlkSize; i2++) { |
| 175 |
if (i1 == i2) { |
| 176 |
// Diagonal elements must be real |
| 177 |
h_out(i0, i1, i2) = Kokkos::ArithTraits<value_type>::real(h_in(i0, i1, i2)); |
| 178 |
} else { |
| 179 |
// Off-diagonal elements are complex and Hermitian |
| 180 |
h_out(i0, i1, i2) = h_in(i0, i1, i2); |
| 181 |
h_out(i0, i2, i1) = Kokkos::ArithTraits<value_type>::conj(h_in(i0, i1, i2)); |
| 182 |
} |
| 183 |
} |
| 184 |
} |
| 185 |
|
| 186 |
// Make matrix diagonal dominant |
| 187 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 188 |
value_type sum = 0; |
| 189 |
for (int i2 = 0; i2 < BlkSize; i2++) { |
| 190 |
if (i1 != i2) { |
| 191 |
sum += Kokkos::abs(h_out(i0, i1, i2)); |
| 192 |
} |
| 193 |
} |
| 194 |
h_out(i0, i1, i1) = sum + 1.0; |
| 195 |
} |
| 196 |
} |
| 197 |
Kokkos::deep_copy(out, h_out); |
| 198 |
} |
| 199 |
|
| 200 |
/// \brief Creates a banded positive definite symmetric (PDS) matrix. |
| 201 |
/// Takes a full diagonal dominant matrix and converts it to a banded pds matrix either |
| 202 |
/// in banded or full storage. |
| 203 |
/// |
| 204 |
/// \tparam InViewType: Input type for the matrix, needs to be a 3D view |
| 205 |
/// \tparam OutViewType: Output type for the matrix, needs to be a 3D view |
| 206 |
/// \tparam UploType: Type indicating whether the matrix is upper or lower triangular |
| 207 |
/// |
| 208 |
/// \param in [in]: Input batched banded matrix, a rank 3 view |
| 209 |
/// \param out [out]: Output batched full matrix, a rank 3 view |
| 210 |
/// \param k [in]: Number of sub/super-diagonals for lower/upper triangular (default is 1) |
| 211 |
/// \param band_storage [in]: Boolean flag indicating whether the output should be in banded storage format (default is |
| 212 |
/// true) |
| 213 |
template <typename InViewType, typename OutViewType, typename UploType> |
| 214 |
void create_banded_pds_matrix(InViewType& in, OutViewType& out, int k = 1, bool band_storage = true) { |
| 215 |
auto h_in = Kokkos::create_mirror_view(in); |
| 216 |
auto h_out = Kokkos::create_mirror_view(out); |
| 217 |
using value_type = typename InViewType::non_const_value_type; |
| 218 |
const int N = in.extent(0), BlkSize = in.extent(1); |
| 219 |
|
| 220 |
Kokkos::deep_copy(h_in, in); |
| 221 |
|
| 222 |
if (band_storage) { |
| 223 |
assert(out.extent(0) == in.extent(0)); |
| 224 |
assert(out.extent(1) == static_cast<std::size_t>(k + 1)); |
| 225 |
assert(out.extent(2) == in.extent(2)); |
| 226 |
if constexpr (std::is_same_v<UploType, KokkosBatched::Uplo::Upper>) { |
| 227 |
for (int i0 = 0; i0 < N; i0++) { |
| 228 |
for (int i1 = 0; i1 < k + 1; i1++) { |
| 229 |
for (int i2 = i1; i2 < BlkSize; i2++) { |
| 230 |
h_out(i0, k - i1, i2) = h_in(i0, i2 - i1, i2); |
| 231 |
} |
| 232 |
} |
| 233 |
} |
| 234 |
} else { |
| 235 |
for (int i0 = 0; i0 < N; i0++) { |
| 236 |
for (int i1 = 0; i1 < k + 1; i1++) { |
| 237 |
for (int i2 = 0; i2 < BlkSize - i1; i2++) { |
| 238 |
h_out(i0, i1, i2) = h_in(i0, i2 + i1, i2); |
| 239 |
} |
| 240 |
} |
| 241 |
} |
| 242 |
} |
| 243 |
} else { |
| 244 |
for (std::size_t i = 0; i < InViewType::rank(); i++) { |
| 245 |
assert(out.extent(i) == in.extent(i)); |
| 246 |
} |
| 247 |
|
| 248 |
if constexpr (std::is_same_v<UploType, KokkosBatched::Uplo::Upper>) { |
| 249 |
for (int i0 = 0; i0 < N; i0++) { |
| 250 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 251 |
for (int i2 = i1; i2 < Kokkos::min(i1 + k + 1, BlkSize); i2++) { |
| 252 |
h_out(i0, i1, i2) = h_in(i0, i1, i2); |
| 253 |
h_out(i0, i2, i1) = Kokkos::ArithTraits<value_type>::conj(h_in(i0, i1, i2)); |
| 254 |
} |
| 255 |
} |
| 256 |
} |
| 257 |
} else { |
| 258 |
for (int i0 = 0; i0 < N; i0++) { |
| 259 |
for (int i1 = 0; i1 < BlkSize; i1++) { |
| 260 |
for (int i2 = Kokkos::max(0, i1 - k); i2 <= i1; i2++) { |
| 261 |
h_out(i0, i1, i2) = h_in(i0, i1, i2); |
| 262 |
h_out(i0, i2, i1) = Kokkos::ArithTraits<value_type>::conj(h_in(i0, i1, i2)); |
| 263 |
} |
| 264 |
} |
| 265 |
} |
| 266 |
} |
| 267 |
} |
| 268 |
Kokkos::deep_copy(out, h_out); |
| 269 |
} |
| 270 |
|
| 271 |
/// \brief Converts a banded matrix to a full matrix. |
| 272 |
/// Takes a banded matrix in banded storage and converts it to a full matrix. |
| 273 |
/// |
| 274 |
/// \tparam InViewType: Input type for the matrix, needs to be a 3D view |
| 275 |
/// \tparam OutViewType: Output type for the matrix, needs to be a 3D view |
| 276 |
/// \tparam UploType: Type indicating whether the matrix is upper or lower triangular |
| 277 |
/// |
| 278 |
/// \param in [in]: Input batched banded matrix, a rank 3 view |
| 279 |
/// \param out [out]: Output batched full matrix, a rank 3 view |
| 280 |
/// \param k [in]: Number of sub/super-diagonals for lower/upper triangular (default is 1) |
| 281 |
/// |
| 282 |
template <typename InViewType, typename OutViewType, typename UploType> |
| 283 |
void banded_to_full(InViewType& in, OutViewType& out, int k = 1) { |
| 284 |
auto h_in = Kokkos::create_mirror_view(in); |
| 285 |
auto h_out = Kokkos::create_mirror_view(out); |
| 286 |
const int N = in.extent(0), BlkSize = in.extent(2); |
| 287 |
|
| 288 |
Kokkos::deep_copy(h_in, in); |
| 289 |
assert(in.extent(0) == out.extent(0)); |
| 290 |
assert(in.extent(1) == static_cast<std::size_t>(k + 1)); |
| 291 |
assert(in.extent(2) == out.extent(2)); |
| 292 |
if constexpr (std::is_same_v<UploType, KokkosBatched::Uplo::Upper>) { |
| 293 |
for (int i0 = 0; i0 < N; i0++) { |
| 294 |
for (int i1 = 0; i1 < k + 1; i1++) { |
| 295 |
for (int i2 = i1; i2 < BlkSize; i2++) { |
| 296 |
h_out(i0, i2 - i1, i2) = h_in(i0, k - i1, i2); |
| 297 |
} |
| 298 |
} |
| 299 |
} |
| 300 |
} else { |
| 301 |
for (int i0 = 0; i0 < N; i0++) { |
| 302 |
for (int i1 = 0; i1 < k + 1; i1++) { |
| 303 |
for (int i2 = 0; i2 < BlkSize - i1; i2++) { |
| 304 |
h_out(i0, i2 + i1, i2) = h_in(i0, i1, i2); |
| 305 |
} |
| 306 |
} |
| 307 |
} |
| 308 |
} |
| 309 |
Kokkos::deep_copy(out, h_out); |
| 310 |
} |
| 311 |
|
| 312 |
} // namespace KokkosBatched |
| 313 |
|
| 314 |
#endif // TEST_BATCHED_DENSE_HELPER_HPP |