|
| 1 | +/* |
| 2 | + * Copyright (c) Meta Platforms, Inc. and affiliates. |
| 3 | + * All rights reserved. |
| 4 | + * |
| 5 | + * This source code is licensed under the BSD-style license found in the |
| 6 | + * LICENSE file in the root directory of this source tree. |
| 7 | + */ |
| 8 | + |
| 9 | +#include <cstdint> |
| 10 | +#include <map> |
| 11 | +#include <typeindex> |
| 12 | +#include <variant> |
| 13 | + |
| 14 | +#include <executorch/kernels/test/FunctionHeaderWrapper.h> // Declares the operator. |
| 15 | +#include <executorch/kernels/test/TestUtil.h> |
| 16 | +#include <executorch/kernels/test/supported_features.h> |
| 17 | +#include <executorch/runtime/core/exec_aten/exec_aten.h> |
| 18 | +#include <executorch/runtime/core/exec_aten/testing_util/tensor_factory.h> |
| 19 | +#include <executorch/runtime/core/exec_aten/testing_util/tensor_util.h> |
| 20 | + |
| 21 | +#include <gtest/gtest.h> |
| 22 | + |
| 23 | +using namespace ::testing; |
| 24 | +using executorch::aten::ArrayRef; |
| 25 | +using executorch::aten::ScalarType; |
| 26 | +using executorch::aten::Tensor; |
| 27 | +using std::optional; |
| 28 | +using torch::executor::testing::TensorFactory; |
| 29 | + |
| 30 | +class OpDimOrderCloneTest : public OperatorTest { |
| 31 | + protected: |
| 32 | + Tensor& op__clone_dim_order_out( |
| 33 | + const Tensor& self, |
| 34 | + bool non_blocking, |
| 35 | + std::optional<ArrayRef<int64_t>> dim_order, |
| 36 | + Tensor& out) { |
| 37 | + return torch::executor::dim_order_ops::_clone_dim_order_outf( |
| 38 | + context_, self, non_blocking, dim_order, out); |
| 39 | + } |
| 40 | + |
| 41 | + template <typename INPUT_CTYPE, typename OUTPUT_CTYPE> |
| 42 | + std::vector<OUTPUT_CTYPE> vector_type_cast(std::vector<INPUT_CTYPE> input) { |
| 43 | + std::vector<OUTPUT_CTYPE> output(input.size()); |
| 44 | + std::transform( |
| 45 | + input.begin(), input.end(), output.begin(), [](INPUT_CTYPE x) { |
| 46 | + return static_cast<OUTPUT_CTYPE>(x); |
| 47 | + }); |
| 48 | + return output; |
| 49 | + } |
| 50 | + |
| 51 | + template <typename INPUT_CTYPE, typename OUTPUT_CTYPE> |
| 52 | + struct ToTestCase { |
| 53 | + const std::vector<int32_t> sizes; |
| 54 | + const std::vector<INPUT_CTYPE> data_in; |
| 55 | + const std::vector<OUTPUT_CTYPE> data_out; |
| 56 | + }; |
| 57 | + |
| 58 | + template <typename CTYPE, ScalarType DTYPE> |
| 59 | + void test_runner_clone(std::vector<ToTestCase<double, double>> test_cases) { |
| 60 | + TensorFactory<DTYPE> tf_in; |
| 61 | + TensorFactory<DTYPE> tf_out; |
| 62 | + |
| 63 | + for (const auto& test_case : test_cases) { |
| 64 | + auto data_in = vector_type_cast<double, CTYPE>(test_case.data_in); |
| 65 | + |
| 66 | + Tensor input = tf_in.make(test_case.sizes, data_in); |
| 67 | + Tensor output = tf_out.zeros_like(input); |
| 68 | + |
| 69 | + std::vector<int64_t> dim_order_vec; |
| 70 | + for (int64_t i = 0; i < input.dim(); i++) { |
| 71 | + dim_order_vec.push_back(i); |
| 72 | + } |
| 73 | + ArrayRef<int64_t> dim_order(dim_order_vec.data(), dim_order_vec.size()); |
| 74 | + |
| 75 | + Tensor ret = op__clone_dim_order_out( |
| 76 | + /*self=*/input, |
| 77 | + /*non_blocking=*/false, |
| 78 | + dim_order, |
| 79 | + output); |
| 80 | + |
| 81 | + Tensor expected = tf_out.make(test_case.sizes, data_in); |
| 82 | + |
| 83 | + // Verifies that the returned and output tensor from _clone_dim_order both |
| 84 | + // match the original input (expected). |
| 85 | + EXPECT_TENSOR_EQ(ret, output); |
| 86 | + EXPECT_TENSOR_EQ(ret, expected); |
| 87 | + } |
| 88 | + } |
| 89 | + |
| 90 | + /* %python |
| 91 | + import torch |
| 92 | + torch.manual_seed(0) |
| 93 | + x = torch.rand(2, 3) |
| 94 | + res = x.clone(memory_format = torch.preserve_format) |
| 95 | + op = "op__clone_dim_order_out" |
| 96 | + opt_setup_params = """ |
| 97 | + bool non_blocking = false; |
| 98 | + optional<MemoryFormat> memory_format; |
| 99 | + """ |
| 100 | + opt_extra_params = "non_blocking, memory_format," |
| 101 | + out_args = "out_shape, dynamism" |
| 102 | + dtype = "ScalarType::Float" |
| 103 | + check = "EXPECT_TENSOR_EQ" */ |
| 104 | + |
| 105 | + // Helper for testing dynamic shape outputs. |
| 106 | + void test_dynamic_shape( |
| 107 | + const std::vector<int32_t>& out_shape, |
| 108 | + enum torch::executor::TensorShapeDynamism dynamism) { |
| 109 | + /* %python |
| 110 | + %rewrite(unary_op) */ |
| 111 | + |
| 112 | + TensorFactory<ScalarType::Float> tf; |
| 113 | + |
| 114 | + Tensor x = tf.make( |
| 115 | + {2, 3}, |
| 116 | + {0.49625658988952637, |
| 117 | + 0.7682217955589294, |
| 118 | + 0.08847743272781372, |
| 119 | + 0.13203048706054688, |
| 120 | + 0.30742281675338745, |
| 121 | + 0.6340786814689636}); |
| 122 | + Tensor expected = tf.make( |
| 123 | + {2, 3}, |
| 124 | + {0.49625658988952637, |
| 125 | + 0.7682217955589294, |
| 126 | + 0.08847743272781372, |
| 127 | + 0.13203048706054688, |
| 128 | + 0.30742281675338745, |
| 129 | + 0.6340786814689636}); |
| 130 | + |
| 131 | + bool non_blocking = false; |
| 132 | + |
| 133 | + Tensor out = tf.zeros(out_shape, dynamism); |
| 134 | + |
| 135 | + std::vector<int64_t> dim_order_vec; |
| 136 | + for (int64_t i = 0; i < x.dim(); i++) { |
| 137 | + dim_order_vec.push_back(i); |
| 138 | + } |
| 139 | + ArrayRef<int64_t> dim_order(dim_order_vec.data(), dim_order_vec.size()); |
| 140 | + |
| 141 | + Tensor ret = op__clone_dim_order_out( |
| 142 | + /*self=*/x, non_blocking, dim_order, out); |
| 143 | + |
| 144 | + EXPECT_TENSOR_EQ(out, expected); |
| 145 | + EXPECT_TENSOR_EQ(ret, expected); |
| 146 | + } |
| 147 | +}; |
| 148 | + |
| 149 | +// Clones tensors of all real dtypes. |
| 150 | +TEST_F(OpDimOrderCloneTest, AllDtypesSupported) { |
| 151 | + std::vector<ToTestCase<double, double>> test_cases = { |
| 152 | + { |
| 153 | + /*sizes=*/{2, 4}, |
| 154 | + /*data_in=*/{2.11, 3.2, 2.3, 4.0, 1.1, 5.2, 1.1, 6.3}, |
| 155 | + /*data_out=*/{}, // data_out shouldn't be used in test_runner_clone |
| 156 | + }, |
| 157 | + { |
| 158 | + /*sizes=*/{3, 4, 0, 5}, |
| 159 | + /*data_in=*/{}, |
| 160 | + /*data_out=*/{}, |
| 161 | + }, |
| 162 | + { |
| 163 | + /*sizes=*/{}, |
| 164 | + /*data_in=*/{10.0}, |
| 165 | + /*data_out=*/{}, // data_out shouldn't be used in test_runner_clone |
| 166 | + }, |
| 167 | + }; |
| 168 | + |
| 169 | +#define TEST_KERNEL(CTYPE, DTYPE) \ |
| 170 | + test_runner_clone<CTYPE, ScalarType::DTYPE>(test_cases); |
| 171 | + |
| 172 | + ET_FORALL_REAL_TYPES(TEST_KERNEL); |
| 173 | + |
| 174 | +#undef TEST_KERNEL |
| 175 | +} |
| 176 | + |
| 177 | +// Cloning with mismatched input and output tensor shapes should fail. |
| 178 | +TEST_F(OpDimOrderCloneTest, MismatchedSizesDie) { |
| 179 | + if (torch::executor::testing::SupportedFeatures::get()->is_aten) { |
| 180 | + GTEST_SKIP() << "Skipping: ATen kernel supports mismatched sizes."; |
| 181 | + } |
| 182 | + TensorFactory<ScalarType::Int> tf; |
| 183 | + Tensor input = tf.make(/*sizes=*/{3, 1, 1, 2}, /*data=*/{1, 2, 3, 4, 5, 6}); |
| 184 | + Tensor out = tf.zeros({3, 2, 1, 1}); |
| 185 | + std::vector<int64_t> dim_order_vec; |
| 186 | + for (int64_t i = 0; i < input.dim(); i++) { |
| 187 | + dim_order_vec.push_back(i); |
| 188 | + } |
| 189 | + ArrayRef<int64_t> dim_order(dim_order_vec.data(), dim_order_vec.size()); |
| 190 | + |
| 191 | + ET_EXPECT_KERNEL_FAILURE( |
| 192 | + context_, |
| 193 | + op__clone_dim_order_out( |
| 194 | + /*self=*/input, |
| 195 | + /*non_blocking=*/false, |
| 196 | + dim_order, |
| 197 | + out)); |
| 198 | +} |
| 199 | + |
| 200 | +// Cloning with a non-contiguous memory format should fail. |
| 201 | +TEST_F(OpDimOrderCloneTest, MismatchedMemoryFormatDies) { |
| 202 | + if (torch::executor::testing::SupportedFeatures::get()->is_aten) { |
| 203 | + GTEST_SKIP() |
| 204 | + << "Skipping: ATen kernel supports non-contiguous memory formats."; |
| 205 | + } |
| 206 | + TensorFactory<ScalarType::Float> tf_in; |
| 207 | + TensorFactory<ScalarType::Float> tf_out; |
| 208 | + Tensor input = |
| 209 | + tf_in.make(/*sizes=*/{3, 1, 1, 2}, /*data=*/{1, 2, 3, 4, 5, 6}); |
| 210 | + Tensor out = tf_out.zeros({3, 1, 1, 2}); |
| 211 | + |
| 212 | + std::vector<int64_t> dim_order_vec; |
| 213 | + for (int64_t i = 0; i < input.dim(); i++) { |
| 214 | + dim_order_vec.push_back(i); |
| 215 | + } |
| 216 | + |
| 217 | + // Mutate dim_order_vec to create an illegal dim_order. |
| 218 | + dim_order_vec[1] = 3; |
| 219 | + dim_order_vec[3] = 1; |
| 220 | + ArrayRef<int64_t> dim_order(dim_order_vec.data(), dim_order_vec.size()); |
| 221 | + |
| 222 | + ET_EXPECT_KERNEL_FAILURE( |
| 223 | + context_, |
| 224 | + op__clone_dim_order_out( |
| 225 | + /*self=*/input, |
| 226 | + /*non_blocking=*/false, |
| 227 | + dim_order, |
| 228 | + out)); |
| 229 | +} |
| 230 | + |
| 231 | +// Cloning with non‑blocking=true should fail because portable kernels only |
| 232 | +// support blocking. |
| 233 | +TEST_F(OpDimOrderCloneTest, MismatchedBlockingDie) { |
| 234 | + if (torch::executor::testing::SupportedFeatures::get()->is_aten) { |
| 235 | + GTEST_SKIP() |
| 236 | + << "Skipping: ATen kernel supports non-blocking data transfer."; |
| 237 | + } |
| 238 | + TensorFactory<ScalarType::Int> tf; |
| 239 | + Tensor input = tf.make(/*sizes=*/{3, 1, 1, 2}, /*data=*/{1, 2, 3, 4, 5, 6}); |
| 240 | + Tensor out = tf.zeros(/*sizes=*/{3, 1, 1, 2}); |
| 241 | + |
| 242 | + std::vector<int64_t> dim_order_vec; |
| 243 | + for (int64_t i = 0; i < input.dim(); i++) { |
| 244 | + dim_order_vec.push_back(i); |
| 245 | + } |
| 246 | + ArrayRef<int64_t> dim_order(dim_order_vec.data(), dim_order_vec.size()); |
| 247 | + |
| 248 | + ET_EXPECT_KERNEL_FAILURE( |
| 249 | + context_, |
| 250 | + op__clone_dim_order_out( |
| 251 | + /*self=*/input, |
| 252 | + /*non_blocking=*/true, |
| 253 | + dim_order, |
| 254 | + out)); |
| 255 | +} |
| 256 | + |
| 257 | +TEST_F(OpDimOrderCloneTest, DynamicShapeUpperBoundSameAsExpected) { |
| 258 | + test_dynamic_shape( |
| 259 | + {2, 3}, torch::executor::TensorShapeDynamism::DYNAMIC_BOUND); |
| 260 | +} |
| 261 | + |
| 262 | +TEST_F(OpDimOrderCloneTest, DynamicShapeUpperBoundLargerThanExpected) { |
| 263 | + test_dynamic_shape( |
| 264 | + {10, 10}, torch::executor::TensorShapeDynamism::DYNAMIC_BOUND); |
| 265 | +} |
| 266 | + |
| 267 | +TEST_F(OpDimOrderCloneTest, DynamicShapeUnbound) { |
| 268 | + if (!torch::executor::testing::SupportedFeatures::get()->output_resize) { |
| 269 | + GTEST_SKIP() << "Dynamic shape unbound not supported."; |
| 270 | + } |
| 271 | + test_dynamic_shape( |
| 272 | + {1, 1}, torch::executor::TensorShapeDynamism::DYNAMIC_UNBOUND); |
| 273 | +} |
| 274 | + |
| 275 | +TEST_F(OpDimOrderCloneTest, ContiguousToChannelsLast) { |
| 276 | + TensorFactory<ScalarType::Float> tf; |
| 277 | + |
| 278 | + Tensor x = tf.make_with_dimorder( |
| 279 | + {3, 5, 2, 2}, |
| 280 | + {0.2432, 0.5248, 0.5361, 0.8513, 0.8184, 0.8206, 0.7357, 0.9655, 0.6138, |
| 281 | + 0.1112, 0.2799, 0.1079, 0.9680, 0.2548, 0.0393, 0.6002, 0.2257, 0.8766, |
| 282 | + 0.2715, 0.1595, 0.2029, 0.7026, 0.6982, 0.8529, 0.4405, 0.6560, 0.9217, |
| 283 | + 0.6372, 0.2446, 0.6590, 0.3866, 0.7185, 0.4439, 0.5346, 0.3179, 0.4492, |
| 284 | + 0.3491, 0.6970, 0.8456, 0.2516, 0.2345, 0.2924, 0.7695, 0.0911, 0.8530, |
| 285 | + 0.8560, 0.6909, 0.7719, 0.8923, 0.5546, 0.6978, 0.8151, 0.3007, 0.3961, |
| 286 | + 0.8416, 0.4296, 0.7203, 0.8963, 0.3597, 0.5552}); |
| 287 | + |
| 288 | + Tensor out = tf.full_channels_last({3, 5, 2, 2}, 0.0); |
| 289 | + Tensor expected = tf.make_with_dimorder( |
| 290 | + {3, 5, 2, 2}, |
| 291 | + {0.2432, 0.8184, 0.6138, 0.9680, 0.2257, 0.5248, 0.8206, 0.1112, 0.2548, |
| 292 | + 0.8766, 0.5361, 0.7357, 0.2799, 0.0393, 0.2715, 0.8513, 0.9655, 0.1079, |
| 293 | + 0.6002, 0.1595, 0.2029, 0.4405, 0.2446, 0.4439, 0.3491, 0.7026, 0.6560, |
| 294 | + 0.6590, 0.5346, 0.6970, 0.6982, 0.9217, 0.3866, 0.3179, 0.8456, 0.8529, |
| 295 | + 0.6372, 0.7185, 0.4492, 0.2516, 0.2345, 0.8530, 0.8923, 0.3007, 0.7203, |
| 296 | + 0.2924, 0.8560, 0.5546, 0.3961, 0.8963, 0.7695, 0.6909, 0.6978, 0.8416, |
| 297 | + 0.3597, 0.0911, 0.7719, 0.8151, 0.4296, 0.5552}, |
| 298 | + /*dim_order=*/{0, 2, 3, 1}); |
| 299 | + |
| 300 | + std::vector<int64_t> dim_order_vec = {0, 2, 3, 1}; |
| 301 | + executorch::aten::ArrayRef<int64_t> dim_order( |
| 302 | + dim_order_vec.data(), dim_order_vec.size()); |
| 303 | + Tensor ret = op__clone_dim_order_out( |
| 304 | + /*self*/ x, /*non_blocking*/ false, /*dim_order*/ dim_order, out); |
| 305 | + |
| 306 | + EXPECT_TENSOR_EQ(out, expected); |
| 307 | + EXPECT_TENSOR_EQ(ret, expected); |
| 308 | +} |
| 309 | + |
| 310 | +TEST_F(OpDimOrderCloneTest, ChannelsLastToContiguous) { |
| 311 | + TensorFactory<ScalarType::Float> tf; |
| 312 | + |
| 313 | + Tensor out = tf.full({3, 5, 2, 2}, 0.0); |
| 314 | + Tensor x = tf.make_with_dimorder( |
| 315 | + {3, 5, 2, 2}, |
| 316 | + {0.2432, 0.8184, 0.6138, 0.9680, 0.2257, 0.5248, 0.8206, 0.1112, 0.2548, |
| 317 | + 0.8766, 0.5361, 0.7357, 0.2799, 0.0393, 0.2715, 0.8513, 0.9655, 0.1079, |
| 318 | + 0.6002, 0.1595, 0.2029, 0.4405, 0.2446, 0.4439, 0.3491, 0.7026, 0.6560, |
| 319 | + 0.6590, 0.5346, 0.6970, 0.6982, 0.9217, 0.3866, 0.3179, 0.8456, 0.8529, |
| 320 | + 0.6372, 0.7185, 0.4492, 0.2516, 0.2345, 0.8530, 0.8923, 0.3007, 0.7203, |
| 321 | + 0.2924, 0.8560, 0.5546, 0.3961, 0.8963, 0.7695, 0.6909, 0.6978, 0.8416, |
| 322 | + 0.3597, 0.0911, 0.7719, 0.8151, 0.4296, 0.5552}, |
| 323 | + /*dim_order=*/{0, 2, 3, 1}); |
| 324 | + |
| 325 | + Tensor expected = tf.make_with_dimorder( |
| 326 | + {3, 5, 2, 2}, |
| 327 | + {0.2432, 0.5248, 0.5361, 0.8513, 0.8184, 0.8206, 0.7357, 0.9655, 0.6138, |
| 328 | + 0.1112, 0.2799, 0.1079, 0.9680, 0.2548, 0.0393, 0.6002, 0.2257, 0.8766, |
| 329 | + 0.2715, 0.1595, 0.2029, 0.7026, 0.6982, 0.8529, 0.4405, 0.6560, 0.9217, |
| 330 | + 0.6372, 0.2446, 0.6590, 0.3866, 0.7185, 0.4439, 0.5346, 0.3179, 0.4492, |
| 331 | + 0.3491, 0.6970, 0.8456, 0.2516, 0.2345, 0.2924, 0.7695, 0.0911, 0.8530, |
| 332 | + 0.8560, 0.6909, 0.7719, 0.8923, 0.5546, 0.6978, 0.8151, 0.3007, 0.3961, |
| 333 | + 0.8416, 0.4296, 0.7203, 0.8963, 0.3597, 0.5552}); |
| 334 | + |
| 335 | + std::vector<int64_t> dim_order_vec = {0, 1, 2, 3}; |
| 336 | + executorch::aten::ArrayRef<int64_t> dim_order( |
| 337 | + dim_order_vec.data(), dim_order_vec.size()); |
| 338 | + Tensor ret = op__clone_dim_order_out( |
| 339 | + /*self*/ x, /*non_blocking*/ false, /*dim_order*/ dim_order, out); |
| 340 | + |
| 341 | + EXPECT_TENSOR_EQ(out, expected); |
| 342 | + EXPECT_TENSOR_EQ(ret, expected); |
| 343 | +} |
| 344 | + |
| 345 | +TEST_F(OpDimOrderCloneTest, PreserveChannelsLast) { |
| 346 | + TensorFactory<ScalarType::Float> tf; |
| 347 | + |
| 348 | + Tensor out = tf.full_channels_last({3, 5, 2, 2}, 0.0); |
| 349 | + Tensor x = tf.make_with_dimorder( |
| 350 | + {3, 5, 2, 2}, |
| 351 | + {0.2432, 0.8184, 0.6138, 0.9680, 0.2257, 0.5248, 0.8206, 0.1112, 0.2548, |
| 352 | + 0.8766, 0.5361, 0.7357, 0.2799, 0.0393, 0.2715, 0.8513, 0.9655, 0.1079, |
| 353 | + 0.6002, 0.1595, 0.2029, 0.4405, 0.2446, 0.4439, 0.3491, 0.7026, 0.6560, |
| 354 | + 0.6590, 0.5346, 0.6970, 0.6982, 0.9217, 0.3866, 0.3179, 0.8456, 0.8529, |
| 355 | + 0.6372, 0.7185, 0.4492, 0.2516, 0.2345, 0.8530, 0.8923, 0.3007, 0.7203, |
| 356 | + 0.2924, 0.8560, 0.5546, 0.3961, 0.8963, 0.7695, 0.6909, 0.6978, 0.8416, |
| 357 | + 0.3597, 0.0911, 0.7719, 0.8151, 0.4296, 0.5552}, |
| 358 | + /*dim_order=*/{0, 2, 3, 1}); |
| 359 | + |
| 360 | + Tensor expected = tf.make_with_dimorder( |
| 361 | + {3, 5, 2, 2}, |
| 362 | + {0.2432, 0.8184, 0.6138, 0.9680, 0.2257, 0.5248, 0.8206, 0.1112, 0.2548, |
| 363 | + 0.8766, 0.5361, 0.7357, 0.2799, 0.0393, 0.2715, 0.8513, 0.9655, 0.1079, |
| 364 | + 0.6002, 0.1595, 0.2029, 0.4405, 0.2446, 0.4439, 0.3491, 0.7026, 0.6560, |
| 365 | + 0.6590, 0.5346, 0.6970, 0.6982, 0.9217, 0.3866, 0.3179, 0.8456, 0.8529, |
| 366 | + 0.6372, 0.7185, 0.4492, 0.2516, 0.2345, 0.8530, 0.8923, 0.3007, 0.7203, |
| 367 | + 0.2924, 0.8560, 0.5546, 0.3961, 0.8963, 0.7695, 0.6909, 0.6978, 0.8416, |
| 368 | + 0.3597, 0.0911, 0.7719, 0.8151, 0.4296, 0.5552}, |
| 369 | + /*dim_order=*/{0, 2, 3, 1}); |
| 370 | + |
| 371 | + Tensor ret = op__clone_dim_order_out( |
| 372 | + /*self*/ x, |
| 373 | + /*non_blocking*/ false, |
| 374 | + /*dim_order*/ executorch::aten::nullopt, |
| 375 | + out); |
| 376 | + |
| 377 | + EXPECT_TENSOR_EQ(out, expected); |
| 378 | + EXPECT_TENSOR_EQ(ret, expected); |
| 379 | +} |
0 commit comments