| 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 |
#include "gtest/gtest.h" |
| 17 |
#include "Kokkos_Core.hpp" |
| 18 |
#include "Kokkos_Random.hpp" |
| 19 |
|
| 20 |
#include "KokkosBatched_HostLevel_Gemm.hpp" |
| 21 |
#include "KokkosBatched_HostLevel_Gemm_DblBuf_Impl.hpp" |
| 22 |
|
| 23 |
#include "KokkosKernels_TestUtils.hpp" |
| 24 |
|
| 25 |
using namespace KokkosBatched; |
| 26 |
|
| 27 |
namespace Test { |
| 28 |
template <typename DeviceType, typename ViewType, typename ScalarType, typename ParamTagType> |
| 29 |
void impl_test_batched_gemm_with_handle(BatchedGemmHandle* batchedGemmHandle, const int N, const int matAdim1, |
| 30 |
const int matAdim2, const int matBdim1, const int matBdim2, const int matCdim1, |
| 31 |
const int matCdim2, ScalarType alpha, ScalarType beta) { |
| 32 |
using execution_space = typename DeviceType::execution_space; |
| 33 |
using transA = typename ParamTagType::transA; |
| 34 |
using transB = typename ParamTagType::transB; |
| 35 |
using batchLayout = typename ParamTagType::batchLayout; |
| 36 |
using ats = Kokkos::ArithTraits<ScalarType>; |
| 37 |
|
| 38 |
int ret = 0; |
| 39 |
auto algo_type = batchedGemmHandle->get_kernel_algo_type(); |
| 40 |
ViewType a_expected, a_actual, b_expected, b_actual, c_expected, c_actual; |
| 41 |
std::string fmsg; |
| 42 |
std::string fmsg_rhs = "algo_type:" + batchedGemmHandle->get_kernel_algo_type_str() + ", "; |
| 43 |
fmsg_rhs += ("N:" + std::to_string(N) + ", "); |
| 44 |
fmsg_rhs += ("A:" + std::to_string(matAdim1) + "x" + std::to_string(matAdim2) + ", "); |
| 45 |
fmsg_rhs += ("B:" + std::to_string(matBdim1) + "x" + std::to_string(matBdim2) + ", "); |
| 46 |
fmsg_rhs += ("C:" + std::to_string(matCdim1) + "x" + std::to_string(matCdim2) + "\n"); |
| 47 |
|
| 48 |
if (std::is_same<batchLayout, BatchLayout::Left>::value) { |
| 49 |
a_expected = ViewType("a_expected", N, matAdim1, matAdim2); |
| 50 |
a_actual = ViewType("a_actual", N, matAdim1, matAdim2); |
| 51 |
b_expected = ViewType("b_expected", N, matBdim1, matBdim2); |
| 52 |
b_actual = ViewType("b_actual", N, matBdim1, matBdim2); |
| 53 |
c_expected = ViewType("c_expected", N, matCdim1, matCdim2); |
| 54 |
c_actual = ViewType("c_actual", N, matCdim1, matCdim2); |
| 55 |
} else if (std::is_same<batchLayout, BatchLayout::Right>::value) { |
| 56 |
a_expected = ViewType("a_expected", matAdim1, matAdim2, N); |
| 57 |
a_actual = ViewType("a_actual", matAdim1, matAdim2, N); |
| 58 |
b_expected = ViewType("b_expected", matBdim1, matBdim2, N); |
| 59 |
b_actual = ViewType("b_actual", matBdim1, matBdim2, N); |
| 60 |
c_expected = ViewType("c_expected", matCdim1, matCdim2, N); |
| 61 |
c_actual = ViewType("c_actual", matCdim1, matCdim2, N); |
| 62 |
} |
| 63 |
|
| 64 |
Kokkos::Random_XorShift64_Pool<execution_space> random(13718); |
| 65 |
|
| 66 |
Kokkos::fill_random(a_expected, random, ScalarType(1.0)); |
| 67 |
Kokkos::fill_random(b_expected, random, ScalarType(1.0)); |
| 68 |
Kokkos::fill_random(c_expected, random, ScalarType(1.0)); |
| 69 |
|
| 70 |
Kokkos::fence(); |
| 71 |
|
| 72 |
Kokkos::deep_copy(a_actual, a_expected); |
| 73 |
Kokkos::deep_copy(b_actual, b_expected); |
| 74 |
Kokkos::deep_copy(c_actual, c_expected); |
| 75 |
|
| 76 |
// Check for expected BatchedDblBufGemm runtime errors |
| 77 |
if (algo_type == GemmKokkosBatchedAlgos::KK_DBLBUF) { |
| 78 |
// Check for DblBuf runtime errors related to team_size |
| 79 |
try { |
| 80 |
fmsg = kk_failure_str(__FILE__, __FUNCTION__, __LINE__); |
| 81 |
Impl::BatchedDblBufGemm<transA, transB, batchLayout, BatchedGemmHandle, ScalarType, decltype(a_actual), |
| 82 |
decltype(b_actual), decltype(c_actual), BoundsCheck::Yes, AlphaTag::No, 65536, 1, 65536>( |
| 83 |
batchedGemmHandle, alpha, a_actual, b_actual, beta, c_actual) |
| 84 |
.invoke(); |
| 85 |
FAIL() << (fmsg + fmsg_rhs); |
| 86 |
} catch (const std::runtime_error& error) { |
| 87 |
; |
| 88 |
} |
| 89 |
|
| 90 |
// Check for DblBuf runtime errors related to vector_len |
| 91 |
try { |
| 92 |
fmsg = kk_failure_str(__FILE__, __FUNCTION__, __LINE__); |
| 93 |
Impl::BatchedDblBufGemm<transA, transB, batchLayout, BatchedGemmHandle, ScalarType, decltype(a_actual), |
| 94 |
decltype(b_actual), decltype(c_actual), BoundsCheck::No, AlphaTag::No, 65536, 65536 * 2, |
| 95 |
65536>(batchedGemmHandle, alpha, a_actual, b_actual, beta, c_actual) |
| 96 |
.invoke(); |
| 97 |
FAIL() << (fmsg + fmsg_rhs); |
| 98 |
} catch (const std::runtime_error& error) { |
| 99 |
; |
| 100 |
} |
| 101 |
} |
| 102 |
|
| 103 |
// Check for expected BatchedGemm runtime errors |
| 104 |
try { |
| 105 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) |
| 106 |
if (algo_type == BaseTplAlgos::ARMPL && N % 2 == 0) { |
| 107 |
auto ninter = batchedGemmHandle->get_tpl_params(); |
| 108 |
// Set ninter parameter for underlying armpl_dgemm_interleave_batch call |
| 109 |
*ninter = N / 2; |
| 110 |
} |
| 111 |
#endif |
| 112 |
|
| 113 |
fmsg = kk_failure_str(__FILE__, __FUNCTION__, __LINE__); |
| 114 |
ret = BatchedGemm<transA, transB, batchLayout>(batchedGemmHandle, alpha, a_actual, b_actual, beta, |
| 115 |
c_actual); // Compute c_actual |
| 116 |
} catch (const std::runtime_error& error) { |
| 117 |
std::string error_msg = error.what(); |
| 118 |
if (algo_type == BaseHeuristicAlgos::SQUARE && matCdim1 != matCdim2) { |
| 119 |
; |
| 120 |
} else if (algo_type == BaseTplAlgos::ARMPL) { |
| 121 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) && ARMPL_BUILD >= 1058 |
| 122 |
auto ninter = batchedGemmHandle->get_tpl_params()[0]; |
| 123 |
// No runtime errors expected since layout is valid, double is a supported |
| 124 |
// type, and ninter != 0 |
| 125 |
if (std::is_same<typename ViewType::value_type, double>::value && ninter != 0) { |
| 126 |
FAIL() << (error_msg + fmsg + fmsg_rhs); |
| 127 |
} |
| 128 |
#else |
| 129 |
; // We expect a runtime error if the ARMPL TPL is not enabled |
| 130 |
#endif |
| 131 |
} else { |
| 132 |
FAIL() << (error_msg + fmsg + fmsg_rhs); |
| 133 |
} |
| 134 |
return; |
| 135 |
} |
| 136 |
ASSERT_EQ(ret, 0) << (fmsg + fmsg_rhs); |
| 137 |
|
| 138 |
Functor_BatchedVanillaGEMM<ViewType, ViewType, ViewType, execution_space> vgemm; |
| 139 |
vgemm.A_t = std::is_same<transA, Trans::Transpose>::value; |
| 140 |
vgemm.B_t = std::is_same<transB, Trans::Transpose>::value; |
| 141 |
vgemm.batch_size_last_dim = std::is_same<batchLayout, BatchLayout::Right>::value; |
| 142 |
vgemm.A_c = vgemm.B_c = false; |
| 143 |
vgemm.A = a_expected; |
| 144 |
vgemm.B = b_expected; |
| 145 |
vgemm.C = c_expected; |
| 146 |
vgemm.alpha = alpha; |
| 147 |
vgemm.beta = beta; |
| 148 |
vgemm.run(); // Compute c_expected |
| 149 |
|
| 150 |
Kokkos::fence(); |
| 151 |
|
| 152 |
typename ViewType::HostMirror c_expected_host = Kokkos::create_mirror_view(c_expected); |
| 153 |
typename ViewType::HostMirror c_actual_host = Kokkos::create_mirror_view(c_actual); |
| 154 |
|
| 155 |
// Copy to host |
| 156 |
Kokkos::deep_copy(c_expected_host, c_expected); |
| 157 |
Kokkos::deep_copy(c_actual_host, c_actual); |
| 158 |
|
| 159 |
Kokkos::fence(); |
| 160 |
|
| 161 |
// check c_expected = c_actual ; this eps is about 2^-9 |
| 162 |
// Set mag_type to host_value_type, we may not have half precision on host |
| 163 |
using mag_type = float; |
| 164 |
mag_type sum(1), diff(0); |
| 165 |
|
| 166 |
auto eps = static_cast<mag_type>(ats::epsilon()); |
| 167 |
|
| 168 |
eps *= std::is_same<ScalarType, Kokkos::Experimental::half_t>::value || |
| 169 |
std::is_same<ScalarType, Kokkos::Experimental::bhalf_t>::value |
| 170 |
? 4 |
| 171 |
: 1e3; |
| 172 |
|
| 173 |
for (int k = 0; k < N; ++k) { |
| 174 |
for (int i = 0; i < matCdim1; ++i) { |
| 175 |
for (int j = 0; j < matCdim2; ++j) { |
| 176 |
if (std::is_same<batchLayout, BatchLayout::Right>::value) { |
| 177 |
sum += ats::abs(c_expected_host(i, j, k)); |
| 178 |
diff += ats::abs(c_expected_host(i, j, k) - c_actual_host(i, j, k)); |
| 179 |
} else { |
| 180 |
sum += ats::abs(c_expected_host(k, i, j)); |
| 181 |
diff += ats::abs(c_expected_host(k, i, j) - c_actual_host(k, i, j)); |
| 182 |
} |
| 183 |
} |
| 184 |
} |
| 185 |
} |
| 186 |
|
| 187 |
EXPECT_NEAR_KK(diff / sum, 0, eps, fmsg + fmsg_rhs); |
| 188 |
} |
| 189 |
|
| 190 |
template <typename DeviceType, typename ViewType, typename ScalarType, typename ParamTagType> |
| 191 |
void impl_test_batched_gemm(const int N, const int matAdim1, const int matAdim2, const int matBdim1, const int matBdim2, |
| 192 |
const int matCdim1, const int matCdim2) { |
| 193 |
{ |
| 194 |
BatchedGemmHandle batchedGemmHandle; |
| 195 |
|
| 196 |
ASSERT_EQ(batchedGemmHandle.get_kernel_algo_type(), BaseHeuristicAlgos::SQUARE); |
| 197 |
ASSERT_EQ(batchedGemmHandle.teamSz, 0); |
| 198 |
ASSERT_EQ(batchedGemmHandle.vecLen, 0); |
| 199 |
|
| 200 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 201 |
cublasHandle_t cublas_handle; |
| 202 |
BatchedGemmHandle batchedGemmHandleCublas(cublas_handle, GemmTplAlgos::CUBLAS, 0, 0); |
| 203 |
ASSERT_EQ(&cublas_handle, batchedGemmHandleCublas.get_tpl_params()); |
| 204 |
ASSERT_EQ(batchedGemmHandleCublas.get_kernel_algo_type(), (int)GemmTplAlgos::CUBLAS); |
| 205 |
ASSERT_EQ(batchedGemmHandleCublas.teamSz, 0); |
| 206 |
ASSERT_EQ(batchedGemmHandleCublas.vecLen, 0); |
| 207 |
#endif |
| 208 |
|
| 209 |
// FIXME temporary workaround to run this magma test only if cublas is not |
| 210 |
// enabled the design of the BatchedGemmHandle currently does not allow |
| 211 |
// simultanous testing in this way. See issue #2177 |
| 212 |
#if defined(KOKKOSKERNELS_ENABLE_TPL_MAGMA) && !defined(KOKKOSKERNELS_ENABLE_TPL_CUBLAS) |
| 213 |
magma_queue_t magma_queue; |
| 214 |
BatchedGemmHandle batchedGemmHandleMagma(magma_queue, GemmTplAlgos::MAGMA, 0, 0); |
| 215 |
ASSERT_EQ(&magma_queue, batchedGemmHandleMagma.get_tpl_params()); |
| 216 |
ASSERT_EQ(batchedGemmHandleMagma.get_kernel_algo_type(), (int)GemmTplAlgos::MAGMA); |
| 217 |
ASSERT_EQ(batchedGemmHandleMagma.teamSz, 0); |
| 218 |
ASSERT_EQ(batchedGemmHandleMagma.vecLen, 0); |
| 219 |
#endif |
| 220 |
} |
| 221 |
|
| 222 |
for (int algo_type = BaseHeuristicAlgos::SQUARE; algo_type < GemmKokkosBatchedAlgos::N; ++algo_type) { |
| 223 |
{ |
| 224 |
try { |
| 225 |
BatchedGemmHandle batchedGemmHandle(algo_type); |
| 226 |
|
| 227 |
ASSERT_EQ(batchedGemmHandle.get_kernel_algo_type(), algo_type); |
| 228 |
|
| 229 |
if (algo_type == BaseTplAlgos::ARMPL || algo_type == BaseKokkosBatchedAlgos::KK_SERIAL || |
| 230 |
algo_type == GemmKokkosBatchedAlgos::KK_SERIAL_RANK0 || algo_type == GemmKokkosBatchedAlgos::KK_DBLBUF) { |
| 231 |
impl_test_batched_gemm_with_handle<DeviceType, ViewType, ScalarType, ParamTagType>( |
| 232 |
&batchedGemmHandle, N, matAdim1, matAdim2, matBdim1, matBdim2, matCdim1, matCdim2, 1.5, 3.0); |
| 233 |
} else if (algo_type == BaseHeuristicAlgos::SQUARE) { |
| 234 |
// Invoke 4 times to ensure we cover all paths for alpha and beta |
| 235 |
impl_test_batched_gemm_with_handle<DeviceType, ViewType, ScalarType, ParamTagType>( |
| 236 |
&batchedGemmHandle, N, matAdim1, matAdim2, matBdim1, matBdim2, matCdim1, matCdim2, 0.0, 0.0); |
| 237 |
impl_test_batched_gemm_with_handle<DeviceType, ViewType, ScalarType, ParamTagType>( |
| 238 |
&batchedGemmHandle, N, matAdim1, matAdim2, matBdim1, matBdim2, matCdim1, matCdim2, 1.0, 0.0); |
| 239 |
impl_test_batched_gemm_with_handle<DeviceType, ViewType, ScalarType, ParamTagType>( |
| 240 |
&batchedGemmHandle, N, matAdim1, matAdim2, matBdim1, matBdim2, matCdim1, matCdim2, 0.0, 1.0); |
| 241 |
impl_test_batched_gemm_with_handle<DeviceType, ViewType, ScalarType, ParamTagType>( |
| 242 |
&batchedGemmHandle, N, matAdim1, matAdim2, matBdim1, matBdim2, matCdim1, matCdim2, 1.5, 3.0); |
| 243 |
} else { |
| 244 |
try { |
| 245 |
// Allocate these views to invoke BatchedGemm with an unsupported |
| 246 |
// algo type |
| 247 |
ViewType a_actual("a_actual", N, matAdim1, matAdim2); |
| 248 |
ViewType b_actual("b_actual", N, matBdim1, matBdim2); |
| 249 |
ViewType c_actual("c_actual", N, matCdim1, matCdim2); |
| 250 |
using ta = typename ParamTagType::transA; |
| 251 |
using tb = typename ParamTagType::transB; |
| 252 |
using bl = typename ParamTagType::batchLayout; |
| 253 |
ScalarType alpha = 0.34; |
| 254 |
ScalarType beta = 0.43; |
| 255 |
BatchedGemm<ta, tb, bl>(&batchedGemmHandle, alpha, a_actual, b_actual, beta, c_actual); |
| 256 |
std::string fmsg = kk_failure_str(__FILE__, __FUNCTION__, __LINE__); |
| 257 |
FAIL() << fmsg; |
| 258 |
} catch (const std::runtime_error& error) { |
| 259 |
; |
| 260 |
} |
| 261 |
} |
| 262 |
} catch (const std::runtime_error& error) { |
| 263 |
#if !defined(KOKKOSKERNELS_ENABLE_TPL_ARMPL) || (ARMPL_BUILD < 1058) |
| 264 |
if (algo_type == BaseTplAlgos::ARMPL) { |
| 265 |
; |
| 266 |
} else { |
| 267 |
std::string fmsg = kk_failure_str(__FILE__, __FUNCTION__, __LINE__); |
| 268 |
FAIL() << fmsg; |
| 269 |
} |
| 270 |
#endif // KOKKOSKERNELS_ENABLE_TPL_ARMPL |
| 271 |
} |
| 272 |
} |
| 273 |
} |
| 274 |
} |
| 275 |
} // namespace Test |
| 276 |
|
| 277 |
template <typename ViewType, typename DeviceType, typename ValueType, typename ScalarType, typename ParamTagType> |
| 278 |
void test_batched_gemm_with_layout(int N) { |
| 279 |
// Square cases |
| 280 |
{ |
| 281 |
int i = 0; |
| 282 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, i, i, i, i, i, i); |
| 283 |
|
| 284 |
i = 10; |
| 285 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, i, i, i, i, i, i); |
| 286 |
|
| 287 |
i = 25; |
| 288 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, i, i, i, i, i, i); |
| 289 |
|
| 290 |
i = 32; |
| 291 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, i, i, i, i, i, i); |
| 292 |
} |
| 293 |
|
| 294 |
// Non-square cases |
| 295 |
for (int i = 1; i < 5; ++i) { |
| 296 |
int dimM = 1 * i; |
| 297 |
int dimN = 2 * i; |
| 298 |
int dimK = 3 * i; |
| 299 |
if ((std::is_same<typename ParamTagType::transA, KokkosBatched::Trans::NoTranspose>::value) && |
| 300 |
(std::is_same<typename ParamTagType::transB, KokkosBatched::Trans::NoTranspose>::value)) { |
| 301 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, dimM, dimK, dimK, dimN, dimM, |
| 302 |
dimN); |
| 303 |
} |
| 304 |
if ((std::is_same<typename ParamTagType::transA, KokkosBatched::Trans::NoTranspose>::value) && |
| 305 |
(std::is_same<typename ParamTagType::transB, KokkosBatched::Trans::Transpose>::value)) { |
| 306 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, dimM, dimK, dimN, dimK, dimM, |
| 307 |
dimN); |
| 308 |
} |
| 309 |
if ((std::is_same<typename ParamTagType::transA, KokkosBatched::Trans::Transpose>::value) && |
| 310 |
(std::is_same<typename ParamTagType::transB, KokkosBatched::Trans::NoTranspose>::value)) { |
| 311 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, dimK, dimM, dimK, dimN, dimM, |
| 312 |
dimN); |
| 313 |
} |
| 314 |
if ((std::is_same<typename ParamTagType::transA, KokkosBatched::Trans::Transpose>::value) && |
| 315 |
(std::is_same<typename ParamTagType::transB, KokkosBatched::Trans::Transpose>::value)) { |
| 316 |
Test::impl_test_batched_gemm<DeviceType, ViewType, ScalarType, ParamTagType>(N, dimK, dimM, dimN, dimK, dimM, |
| 317 |
dimN); |
| 318 |
} |
| 319 |
} |
| 320 |
} |
| 321 |
|
| 322 |
template <typename DeviceType, typename ValueType, typename ScalarType, typename ParamTagType> |
| 323 |
int test_batched_gemm() { |
| 324 |
#if defined(KOKKOSKERNELS_INST_LAYOUTLEFT) || \ |
| 325 |
(!defined(KOKKOSKERNELS_ETI_ONLY) && !defined(KOKKOSKERNELS_IMPL_CHECK_ETI_CALLS)) |
| 326 |
if constexpr (std::is_same_v<typename ParamTagType::batchLayout, typename BatchLayout::Right>) { |
| 327 |
using param_tag_type = |
| 328 |
::Test::SharedParamTag<typename ParamTagType::transA, typename ParamTagType::transB, BatchLayout::Right>; |
| 329 |
typedef Kokkos::View<ValueType***, Kokkos::LayoutLeft, DeviceType> llVt; |
| 330 |
test_batched_gemm_with_layout<llVt, DeviceType, ValueType, ScalarType, param_tag_type>(0); |
| 331 |
test_batched_gemm_with_layout<llVt, DeviceType, ValueType, ScalarType, param_tag_type>(1); |
| 332 |
test_batched_gemm_with_layout<llVt, DeviceType, ValueType, ScalarType, param_tag_type>(4); |
| 333 |
test_batched_gemm_with_layout<llVt, DeviceType, ValueType, ScalarType, param_tag_type>(8); |
| 334 |
test_batched_gemm_with_layout<llVt, DeviceType, ValueType, ScalarType, param_tag_type>(16); |
| 335 |
} else { |
| 336 |
std::cerr << "TEST SKIPPED since BatchLayout is not Right." << std::endl; |
| 337 |
} |
| 338 |
#else |
| 339 |
std::cerr << "TEST SKIPPED since LayoutLeft is not ETI'd." << std::endl; |
| 340 |
#endif // KOKKOSKERNELS_INST_LAYOUTLEFT |
| 341 |
|
| 342 |
#if defined(KOKKOSKERNELS_INST_LAYOUTRIGHT) || \ |
| 343 |
(!defined(KOKKOSKERNELS_ETI_ONLY) && !defined(KOKKOSKERNELS_IMPL_CHECK_ETI_CALLS)) |
| 344 |
if constexpr (std::is_same_v<typename ParamTagType::batchLayout, typename BatchLayout::Left>) { |
| 345 |
using param_tag_type = |
| 346 |
::Test::SharedParamTag<typename ParamTagType::transA, typename ParamTagType::transB, BatchLayout::Left>; |
| 347 |
typedef Kokkos::View<ValueType***, Kokkos::LayoutRight, DeviceType> lrVt; |
| 348 |
test_batched_gemm_with_layout<lrVt, DeviceType, ValueType, ScalarType, param_tag_type>(0); |
| 349 |
test_batched_gemm_with_layout<lrVt, DeviceType, ValueType, ScalarType, param_tag_type>(1); |
| 350 |
test_batched_gemm_with_layout<lrVt, DeviceType, ValueType, ScalarType, param_tag_type>(4); |
| 351 |
test_batched_gemm_with_layout<lrVt, DeviceType, ValueType, ScalarType, param_tag_type>(8); |
| 352 |
test_batched_gemm_with_layout<lrVt, DeviceType, ValueType, ScalarType, param_tag_type>(16); |
| 353 |
} else { |
| 354 |
std::cerr << "TEST SKIPPED since BatchLayout is not Left." << std::endl; |
| 355 |
} |
| 356 |
#else |
| 357 |
std::cerr << "TEST SKIPPED since LayoutRight is not ETI'd." << std::endl; |
| 358 |
#endif // KOKKOSKERNELS_INST_LAYOUTRIGHT |
| 359 |
return 0; |
| 360 |
} |