| 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 |
/// \author Yuuichi Asahi ([email protected]) |
| 17 |
#include <gtest/gtest.h> |
| 18 |
#include <Kokkos_Core.hpp> |
| 19 |
#include <Kokkos_Random.hpp> |
| 20 |
|
| 21 |
#include "KokkosBatched_Util.hpp" |
| 22 |
#include "KokkosBatched_Pbtrf.hpp" |
| 23 |
#include "Test_Batched_DenseUtils.hpp" |
| 24 |
|
| 25 |
using namespace KokkosBatched; |
| 26 |
|
| 27 |
namespace Test { |
| 28 |
namespace Pbtrf { |
| 29 |
|
| 30 |
template <typename U> |
| 31 |
struct ParamTag { |
| 32 |
using uplo = U; |
| 33 |
}; |
| 34 |
|
| 35 |
template <typename DeviceType, typename ABViewType, typename ParamTagType, typename AlgoTagType> |
| 36 |
struct Functor_BatchedSerialPbtrf { |
| 37 |
using execution_space = typename DeviceType::execution_space; |
| 38 |
ABViewType _ab; |
| 39 |
|
| 40 |
KOKKOS_INLINE_FUNCTION |
| 41 |
Functor_BatchedSerialPbtrf(const ABViewType &ab) : _ab(ab) {} |
| 42 |
|
| 43 |
KOKKOS_INLINE_FUNCTION |
| 44 |
void operator()(const ParamTagType &, const int k, int &info) const { |
| 45 |
auto sub_ab = Kokkos::subview(_ab, k, Kokkos::ALL(), Kokkos::ALL()); |
| 46 |
|
| 47 |
info += KokkosBatched::SerialPbtrf<typename ParamTagType::uplo, AlgoTagType>::invoke(sub_ab); |
| 48 |
} |
| 49 |
|
| 50 |
inline int run() { |
| 51 |
using value_type = typename ABViewType::non_const_value_type; |
| 52 |
std::string name_region("KokkosBatched::Test::SerialPbtrf"); |
| 53 |
const std::string name_value_type = Test::value_type_name<value_type>(); |
| 54 |
std::string name = name_region + name_value_type; |
| 55 |
int info_sum = 0; |
| 56 |
Kokkos::Profiling::pushRegion(name.c_str()); |
| 57 |
Kokkos::RangePolicy<execution_space, ParamTagType> policy(0, _ab.extent(0)); |
| 58 |
Kokkos::parallel_reduce(name.c_str(), policy, *this, info_sum); |
| 59 |
Kokkos::Profiling::popRegion(); |
| 60 |
return info_sum; |
| 61 |
} |
| 62 |
}; |
| 63 |
|
| 64 |
template <typename DeviceType, typename ScalarType, typename AViewType, typename BViewType, typename CViewType, |
| 65 |
typename ArgTransA, typename ArgTransB> |
| 66 |
struct Functor_BatchedSerialGemm { |
| 67 |
using execution_space = typename DeviceType::execution_space; |
| 68 |
AViewType _a; |
| 69 |
BViewType _b; |
| 70 |
CViewType _c; |
| 71 |
ScalarType _alpha, _beta; |
| 72 |
|
| 73 |
KOKKOS_INLINE_FUNCTION |
| 74 |
Functor_BatchedSerialGemm(const ScalarType alpha, const AViewType &a, const BViewType &b, const ScalarType beta, |
| 75 |
const CViewType &c) |
| 76 |
: _a(a), _b(b), _c(c), _alpha(alpha), _beta(beta) {} |
| 77 |
|
| 78 |
KOKKOS_INLINE_FUNCTION |
| 79 |
void operator()(const int k) const { |
| 80 |
auto aa = Kokkos::subview(_a, k, Kokkos::ALL(), Kokkos::ALL()); |
| 81 |
auto bb = Kokkos::subview(_b, k, Kokkos::ALL(), Kokkos::ALL()); |
| 82 |
auto cc = Kokkos::subview(_c, k, Kokkos::ALL(), Kokkos::ALL()); |
| 83 |
|
| 84 |
KokkosBatched::SerialGemm<ArgTransA, ArgTransB, Algo::Gemm::Unblocked>::invoke(_alpha, aa, bb, _beta, cc); |
| 85 |
} |
| 86 |
|
| 87 |
inline void run() { |
| 88 |
using value_type = typename AViewType::non_const_value_type; |
| 89 |
std::string name_region("KokkosBatched::Test::SerialPbtrf"); |
| 90 |
const std::string name_value_type = Test::value_type_name<value_type>(); |
| 91 |
std::string name = name_region + name_value_type; |
| 92 |
Kokkos::RangePolicy<execution_space> policy(0, _a.extent(0)); |
| 93 |
Kokkos::parallel_for(name.c_str(), policy, *this); |
| 94 |
} |
| 95 |
}; |
| 96 |
|
| 97 |
template <typename DeviceType, typename ScalarType, typename LayoutType, typename ParamTagType, typename AlgoTagType> |
| 98 |
/// \brief Implementation details of batched pbtrf test |
| 99 |
/// Confirm A = U**H * U or L * L**H, where |
| 100 |
/// For full storage, |
| 101 |
/// A: [[4, 1], |
| 102 |
/// [1, 4]] |
| 103 |
/// L: [[sqrt(4), 0], |
| 104 |
/// [1/sqrt(4), sqrt(4 - (1/sqrt(4))**2)] |
| 105 |
/// U: [[sqrt(4), 1/sqrt(4)], |
| 106 |
/// [0, sqrt(4 - (1/sqrt(4))**2)] |
| 107 |
/// |
| 108 |
/// For lower banded storage, Ab = Lb * Lb**H |
| 109 |
/// Ab: [[4, 4], |
| 110 |
/// [1, 0]] |
| 111 |
/// Lb: [[sqrt(4), sqrt(4 - (1/sqrt(4))**2)], |
| 112 |
/// [1/sqrt(4), 0]] |
| 113 |
/// |
| 114 |
/// For upper banded storage, Ab = Ub**H * Ub |
| 115 |
/// Ab: [[0, 1], |
| 116 |
/// [4, 4]] |
| 117 |
/// Ub: [[0, 1/sqrt(4)], |
| 118 |
/// [sqrt(4), sqrt(4 - (1/sqrt(4))**2)]] |
| 119 |
/// \param N [in] Batch size of AB |
| 120 |
void impl_test_batched_pbtrf_analytical(const int N) { |
| 121 |
using ats = typename Kokkos::ArithTraits<ScalarType>; |
| 122 |
using RealType = typename ats::mag_type; |
| 123 |
using View3DType = Kokkos::View<ScalarType ***, LayoutType, DeviceType>; |
| 124 |
|
| 125 |
constexpr int BlkSize = 2, k = 1; |
| 126 |
View3DType A("A", N, BlkSize, BlkSize), Ab("Ab", N, k + 1, BlkSize), |
| 127 |
Ab_ref("Ab_ref", N, k + 1, BlkSize); // Banded matrix |
| 128 |
|
| 129 |
auto h_A = Kokkos::create_mirror_view(A); |
| 130 |
for (int ib = 0; ib < N; ib++) { |
| 131 |
for (int i = 0; i < BlkSize; i++) { |
| 132 |
for (int j = 0; j < BlkSize; j++) { |
| 133 |
h_A(ib, i, j) = i == j ? 4.0 : 1.0; |
| 134 |
} |
| 135 |
} |
| 136 |
} |
| 137 |
|
| 138 |
Kokkos::deep_copy(A, h_A); |
| 139 |
|
| 140 |
// Create banded triangluar matrix in normal and banded storage |
| 141 |
using ArgUplo = typename ParamTagType::uplo; |
| 142 |
create_banded_triangular_matrix<View3DType, View3DType, ArgUplo>(A, Ab, k, true); |
| 143 |
|
| 144 |
// Make a reference using the naive Cholesky decomposition |
| 145 |
// Cholesky decomposition for full storage |
| 146 |
// l_kk = np.sqrt( a_kk - sum_{i=1}^{k-1}( l_ik^2 ) ) |
| 147 |
// l_ik = 1/l_kk * ( a_ik - sum_{j=1}^{k-1}( l_ij * l_kj ) ) |
| 148 |
auto h_Ab_ref = Kokkos::create_mirror_view(Ab_ref); |
| 149 |
auto h_Ab = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), Ab); |
| 150 |
if (std::is_same_v<typename ParamTagType::uplo, KokkosBatched::Uplo::Upper>) { |
| 151 |
// A = U**H * U |
| 152 |
for (int ib = 0; ib < N; ib++) { |
| 153 |
h_Ab_ref(ib, 1, 0) = Kokkos::sqrt(h_Ab(ib, 1, 0)); |
| 154 |
h_Ab_ref(ib, 0, 1) = 1.0 / h_Ab_ref(ib, 1, 0); |
| 155 |
h_Ab_ref(ib, 1, 1) = Kokkos::sqrt(h_Ab(ib, 1, 1) - h_Ab_ref(ib, 0, 1) * h_Ab_ref(ib, 0, 1)); |
| 156 |
} |
| 157 |
} else { |
| 158 |
// A = L * L**H |
| 159 |
for (int ib = 0; ib < N; ib++) { |
| 160 |
h_Ab_ref(ib, 0, 0) = Kokkos::sqrt(h_Ab(ib, 0, 0)); |
| 161 |
h_Ab_ref(ib, 1, 0) = 1.0 / h_Ab_ref(ib, 0, 0); |
| 162 |
h_Ab_ref(ib, 0, 1) = Kokkos::sqrt(h_Ab(ib, 0, 1) - h_Ab_ref(ib, 1, 0) * h_Ab_ref(ib, 1, 0)); |
| 163 |
} |
| 164 |
} |
| 165 |
|
| 166 |
// Factorize with Pbtrf: A = U**H * U or A = L * L**H |
| 167 |
auto info = Functor_BatchedSerialPbtrf<DeviceType, View3DType, ParamTagType, AlgoTagType>(Ab).run(); |
| 168 |
Kokkos::fence(); |
| 169 |
EXPECT_EQ(info, 0); |
| 170 |
|
| 171 |
// this eps is about 10^-14 |
| 172 |
RealType eps = 1.0e3 * ats::epsilon(); |
| 173 |
|
| 174 |
// Check if Ab == Ub or Lb |
| 175 |
Kokkos::deep_copy(h_Ab, Ab); |
| 176 |
for (int ib = 0; ib < N; ib++) { |
| 177 |
for (int i = 0; i < k + 1; i++) { |
| 178 |
for (int j = 0; j < BlkSize; j++) { |
| 179 |
EXPECT_NEAR_KK(h_Ab(ib, i, j), h_Ab_ref(ib, i, j), eps); |
| 180 |
} |
| 181 |
} |
| 182 |
} |
| 183 |
} |
| 184 |
|
| 185 |
template <typename DeviceType, typename ScalarType, typename LayoutType, typename ParamTagType, typename AlgoTagType> |
| 186 |
/// \brief Implementation details of batched pbtrs test |
| 187 |
/// |
| 188 |
/// \param N [in] Batch size of RHS (banded matrix can also be batched matrix) |
| 189 |
/// \param k [in] Number of superdiagonals or subdiagonals of matrix A |
| 190 |
/// \param BlkSize [in] Block size of matrix A |
| 191 |
void impl_test_batched_pbtrf(const int N, const int k, const int BlkSize) { |
| 192 |
using View3DType = Kokkos::View<ScalarType ***, LayoutType, DeviceType>; |
| 193 |
View3DType A("A", N, BlkSize, BlkSize), A_reconst("A_reconst", N, BlkSize, BlkSize), |
| 194 |
Ab("Ab", N, k + 1, BlkSize); // Banded matrix |
| 195 |
|
| 196 |
using execution_space = typename DeviceType::execution_space; |
| 197 |
Kokkos::Random_XorShift64_Pool<execution_space> rand_pool(13718); |
| 198 |
ScalarType randStart, randEnd; |
| 199 |
|
| 200 |
// Initialize A_reconst with random matrix |
| 201 |
KokkosKernels::Impl::getRandomBounds(1.0, randStart, randEnd); |
| 202 |
Kokkos::fill_random(A, rand_pool, randStart, randEnd); |
| 203 |
|
| 204 |
// Make the matrix Positive Definite Symmetric and Diagonal dominant |
| 205 |
random_to_pds(A, A_reconst); |
| 206 |
Kokkos::deep_copy(A, 0.0); |
| 207 |
|
| 208 |
// Create banded triangluar matrix in normal and banded storage |
| 209 |
using ArgUplo = typename ParamTagType::uplo; |
| 210 |
create_banded_pds_matrix<View3DType, View3DType, ArgUplo>(A_reconst, A, k, false); |
| 211 |
|
| 212 |
create_banded_triangular_matrix<View3DType, View3DType, ArgUplo>(A_reconst, Ab, k, true); |
| 213 |
|
| 214 |
// Clear matrix |
| 215 |
Kokkos::deep_copy(A_reconst, 0.0); |
| 216 |
|
| 217 |
// Factorize with Pbtrf: A = U**H * U or A = L * L**H |
| 218 |
auto info = Functor_BatchedSerialPbtrf<DeviceType, View3DType, ParamTagType, AlgoTagType>(Ab).run(); |
| 219 |
|
| 220 |
Kokkos::fence(); |
| 221 |
EXPECT_EQ(info, 0); |
| 222 |
|
| 223 |
if (std::is_same_v<typename ParamTagType::uplo, KokkosBatched::Uplo::Upper>) { |
| 224 |
// A = U**H * U |
| 225 |
View3DType U("U", N, BlkSize, BlkSize), Uc("Uc", N, BlkSize, BlkSize); |
| 226 |
banded_to_full<View3DType, View3DType, ArgUplo>(Ab, U, k); |
| 227 |
|
| 228 |
// Compute the complex conjugate of U |
| 229 |
// U -> conj(U) |
| 230 |
auto h_U = Kokkos::create_mirror_view(U); |
| 231 |
auto h_Uc = Kokkos::create_mirror_view(Uc); |
| 232 |
Kokkos::deep_copy(h_U, U); |
| 233 |
Kokkos::deep_copy(h_Uc, Uc); |
| 234 |
for (int ib = 0; ib < N; ib++) { |
| 235 |
for (int i = 0; i < BlkSize; i++) { |
| 236 |
for (int j = 0; j < BlkSize; j++) { |
| 237 |
h_Uc(ib, i, j) = Kokkos::ArithTraits<ScalarType>::conj(h_U(ib, i, j)); |
| 238 |
} |
| 239 |
} |
| 240 |
} |
| 241 |
Kokkos::deep_copy(Uc, h_Uc); |
| 242 |
|
| 243 |
// Create conjugate of U |
| 244 |
Functor_BatchedSerialGemm<DeviceType, ScalarType, View3DType, View3DType, View3DType, Trans::Transpose, |
| 245 |
Trans::NoTranspose>(1.0, Uc, U, 0.0, A_reconst) |
| 246 |
.run(); |
| 247 |
} else { |
| 248 |
// A = L * L**H |
| 249 |
View3DType L("L", N, BlkSize, BlkSize), Lc("Lc", N, BlkSize, BlkSize); |
| 250 |
banded_to_full<View3DType, View3DType, ArgUplo>(Ab, L, k); |
| 251 |
|
| 252 |
// Compute the complex conjugate of L |
| 253 |
// L -> conj(L) |
| 254 |
auto h_L = Kokkos::create_mirror_view(L); |
| 255 |
auto h_Lc = Kokkos::create_mirror_view(Lc); |
| 256 |
Kokkos::deep_copy(h_L, L); |
| 257 |
Kokkos::deep_copy(h_Lc, Lc); |
| 258 |
for (int ib = 0; ib < N; ib++) { |
| 259 |
for (int i = 0; i < BlkSize; i++) { |
| 260 |
for (int j = 0; j < BlkSize; j++) { |
| 261 |
h_Lc(ib, i, j) = Kokkos::ArithTraits<ScalarType>::conj(h_L(ib, i, j)); |
| 262 |
} |
| 263 |
} |
| 264 |
} |
| 265 |
Kokkos::deep_copy(Lc, h_Lc); |
| 266 |
|
| 267 |
// Create conjugate of L |
| 268 |
Functor_BatchedSerialGemm<DeviceType, ScalarType, View3DType, View3DType, View3DType, Trans::NoTranspose, |
| 269 |
Trans::Transpose>(1.0, L, Lc, 0.0, A_reconst) |
| 270 |
.run(); |
| 271 |
} |
| 272 |
|
| 273 |
// this eps is about 10^-14 |
| 274 |
using ats = typename Kokkos::ArithTraits<ScalarType>; |
| 275 |
using RealType = typename ats::mag_type; |
| 276 |
RealType eps = 1.0e3 * ats::epsilon(); |
| 277 |
|
| 278 |
auto h_A = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A); |
| 279 |
auto h_A_reconst = Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), A_reconst); |
| 280 |
|
| 281 |
// Check if A = U**H * U or A = L * L**H |
| 282 |
for (int ib = 0; ib < N; ib++) { |
| 283 |
for (int i = 0; i < BlkSize; i++) { |
| 284 |
for (int j = 0; j < BlkSize; j++) { |
| 285 |
EXPECT_NEAR_KK(h_A_reconst(ib, i, j), h_A(ib, i, j), eps); |
| 286 |
} |
| 287 |
} |
| 288 |
} |
| 289 |
} |
| 290 |
|
| 291 |
} // namespace Pbtrf |
| 292 |
} // namespace Test |
| 293 |
|
| 294 |
template <typename DeviceType, typename ScalarType, typename ParamTagType, typename AlgoTagType> |
| 295 |
int test_batched_pbtrf() { |
| 296 |
#if defined(KOKKOSKERNELS_INST_LAYOUTLEFT) |
| 297 |
{ |
| 298 |
using LayoutType = Kokkos::LayoutLeft; |
| 299 |
Test::Pbtrf::impl_test_batched_pbtrf_analytical<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(1); |
| 300 |
Test::Pbtrf::impl_test_batched_pbtrf_analytical<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(2); |
| 301 |
for (int i = 0; i < 10; i++) { |
| 302 |
int k = 1; |
| 303 |
Test::Pbtrf::impl_test_batched_pbtrf<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(1, k, i); |
| 304 |
Test::Pbtrf::impl_test_batched_pbtrf<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(2, k, i); |
| 305 |
} |
| 306 |
} |
| 307 |
#endif |
| 308 |
#if defined(KOKKOSKERNELS_INST_LAYOUTRIGHT) |
| 309 |
{ |
| 310 |
using LayoutType = Kokkos::LayoutRight; |
| 311 |
Test::Pbtrf::impl_test_batched_pbtrf_analytical<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(1); |
| 312 |
Test::Pbtrf::impl_test_batched_pbtrf_analytical<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(2); |
| 313 |
for (int i = 0; i < 10; i++) { |
| 314 |
int k = 1; |
| 315 |
Test::Pbtrf::impl_test_batched_pbtrf<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(1, k, i); |
| 316 |
Test::Pbtrf::impl_test_batched_pbtrf<DeviceType, ScalarType, LayoutType, ParamTagType, AlgoTagType>(2, k, i); |
| 317 |
} |
| 318 |
} |
| 319 |
#endif |
| 320 |
|
| 321 |
return 0; |
| 322 |
} |