Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <gtest/gtest.h> |
| 9 | |
| 10 | #include <algorithm> |
| 11 | #include <array> |
| 12 | #include <cmath> |
| 13 | #include <cstddef> |
| 14 | #include <cstdlib> |
| 15 | #include <functional> |
| 16 | #include <initializer_list> |
| 17 | #include <limits> |
| 18 | #include <random> |
| 19 | #include <vector> |
| 20 | |
| 21 | #include <xnnpack.h> |
| 22 | |
| 23 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 24 | class BinaryElementwiseOperatorTester { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 25 | public: |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 26 | enum class OperationType { |
| 27 | Unknown, |
| 28 | Add, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 29 | Divide, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 30 | Maximum, |
| 31 | Minimum, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 32 | Multiply, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 33 | Subtract, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 34 | }; |
| 35 | |
| 36 | inline BinaryElementwiseOperatorTester& input1_shape(std::initializer_list<size_t> input1_shape) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 37 | assert(input1_shape.size() <= XNN_MAX_TENSOR_DIMS); |
| 38 | this->input1_shape_ = std::vector<size_t>(input1_shape); |
| 39 | return *this; |
| 40 | } |
| 41 | |
| 42 | inline const std::vector<size_t>& input1_shape() const { |
| 43 | return this->input1_shape_; |
| 44 | } |
| 45 | |
| 46 | inline size_t input1_dim(size_t i) const { |
| 47 | return i < num_input1_dims() ? this->input1_shape_[i] : 1; |
| 48 | } |
| 49 | |
| 50 | inline size_t num_input1_dims() const { |
| 51 | return this->input1_shape_.size(); |
| 52 | } |
| 53 | |
| 54 | inline size_t num_input1_elements() const { |
| 55 | return std::accumulate( |
| 56 | this->input1_shape_.begin(), this->input1_shape_.end(), size_t(1), std::multiplies<size_t>()); |
| 57 | } |
| 58 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 59 | inline BinaryElementwiseOperatorTester& input2_shape(std::initializer_list<size_t> input2_shape) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 60 | assert(input2_shape.size() <= XNN_MAX_TENSOR_DIMS); |
| 61 | this->input2_shape_ = std::vector<size_t>(input2_shape); |
| 62 | return *this; |
| 63 | } |
| 64 | |
| 65 | inline const std::vector<size_t>& input2_shape() const { |
| 66 | return this->input2_shape_; |
| 67 | } |
| 68 | |
| 69 | inline size_t input2_dim(size_t i) const { |
| 70 | return i < num_input2_dims() ? this->input2_shape_[i] : 1; |
| 71 | } |
| 72 | |
| 73 | inline size_t num_input2_dims() const { |
| 74 | return this->input2_shape_.size(); |
| 75 | } |
| 76 | |
| 77 | inline size_t num_input2_elements() const { |
| 78 | return std::accumulate( |
| 79 | this->input2_shape_.begin(), this->input2_shape_.end(), size_t(1), std::multiplies<size_t>()); |
| 80 | } |
| 81 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 82 | inline BinaryElementwiseOperatorTester& qmin(uint8_t qmin) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 83 | this->qmin_ = qmin; |
| 84 | return *this; |
| 85 | } |
| 86 | |
| 87 | inline uint8_t qmin() const { |
| 88 | return this->qmin_; |
| 89 | } |
| 90 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 91 | inline BinaryElementwiseOperatorTester& qmax(uint8_t qmax) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 92 | this->qmax_ = qmax; |
| 93 | return *this; |
| 94 | } |
| 95 | |
| 96 | inline uint8_t qmax() const { |
| 97 | return this->qmax_; |
| 98 | } |
| 99 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 100 | inline BinaryElementwiseOperatorTester& operation_type(OperationType operation_type) { |
| 101 | this->operation_type_ = operation_type; |
| 102 | return *this; |
| 103 | } |
| 104 | |
| 105 | inline OperationType operation_type() const { |
| 106 | return this->operation_type_; |
| 107 | } |
| 108 | |
| 109 | inline BinaryElementwiseOperatorTester& iterations(size_t iterations) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 110 | this->iterations_ = iterations; |
| 111 | return *this; |
| 112 | } |
| 113 | |
| 114 | inline size_t iterations() const { |
| 115 | return this->iterations_; |
| 116 | } |
| 117 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 118 | float Compute(float a, float b) const { |
| 119 | switch (operation_type()) { |
| 120 | case OperationType::Add: |
| 121 | return a + b; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 122 | case OperationType::Divide: |
| 123 | return a / b; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 124 | case OperationType::Maximum: |
| 125 | return std::max<float>(a, b); |
| 126 | case OperationType::Minimum: |
| 127 | return std::min<float>(a, b); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 128 | case OperationType::Multiply: |
| 129 | return a * b; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 130 | case OperationType::Subtract: |
| 131 | return a - b; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 132 | default: |
| 133 | return std::nanf(""); |
| 134 | } |
| 135 | } |
| 136 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 137 | void TestF32() const { |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 138 | ASSERT_NE(operation_type(), OperationType::Unknown); |
| 139 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 140 | std::random_device random_device; |
| 141 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 142 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.01f, 1.0f), rng); |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 143 | |
| 144 | // Compute generalized shapes. |
| 145 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_dims; |
| 146 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_dims; |
| 147 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_dims; |
| 148 | std::fill(input1_dims.begin(), input1_dims.end(), 1); |
| 149 | std::fill(input2_dims.begin(), input2_dims.end(), 1); |
| 150 | std::fill(output_dims.begin(), output_dims.end(), 1); |
| 151 | std::copy(input1_shape().cbegin(), input1_shape().cend(), input1_dims.end() - num_input1_dims()); |
| 152 | std::copy(input2_shape().cbegin(), input2_shape().cend(), input2_dims.end() - num_input2_dims()); |
| 153 | for (size_t i = 0; i < XNN_MAX_TENSOR_DIMS; i++) { |
| 154 | if (input1_dims[i] != 1 && input2_dims[i] != 1) { |
| 155 | ASSERT_EQ(input1_dims[i], input2_dims[i]); |
| 156 | } |
| 157 | output_dims[i] = std::max(input1_dims[i], input2_dims[i]); |
| 158 | } |
| 159 | const size_t num_output_elements = |
| 160 | std::accumulate(output_dims.begin(), output_dims.end(), size_t(1), std::multiplies<size_t>()); |
| 161 | |
| 162 | // Compute generalized strides. |
| 163 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_strides; |
| 164 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_strides; |
| 165 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_strides; |
| 166 | size_t input1_stride = 1, input2_stride = 1, output_stride = 1; |
| 167 | for (size_t i = XNN_MAX_TENSOR_DIMS; i != 0; i--) { |
| 168 | input1_strides[i - 1] = input1_dims[i - 1] == 1 ? 0 : input1_stride; |
| 169 | input2_strides[i - 1] = input2_dims[i - 1] == 1 ? 0 : input2_stride; |
| 170 | output_strides[i - 1] = output_stride; |
| 171 | input1_stride *= input1_dims[i - 1]; |
| 172 | input2_stride *= input2_dims[i - 1]; |
| 173 | output_stride *= output_dims[i - 1]; |
| 174 | } |
| 175 | |
| 176 | std::vector<float> input1(XNN_EXTRA_BYTES / sizeof(float) + num_input1_elements()); |
| 177 | std::vector<float> input2(XNN_EXTRA_BYTES / sizeof(float) + num_input2_elements()); |
| 178 | std::vector<float> output(num_output_elements); |
| 179 | std::vector<float> output_ref(num_output_elements); |
| 180 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 181 | std::generate(input1.begin(), input1.end(), std::ref(f32rng)); |
| 182 | std::generate(input2.begin(), input2.end(), std::ref(f32rng)); |
| 183 | std::fill(output.begin(), output.end(), nanf("")); |
| 184 | |
| 185 | // Compute reference results. |
| 186 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 187 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 188 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 189 | for (size_t l = 0; l < output_dims[3]; l++) { |
Marat Dukhan | fc2b96e | 2019-12-03 12:04:04 -0800 | [diff] [blame] | 190 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 191 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 192 | output_ref[i * output_strides[0] + j * output_strides[1] + k * output_strides[2] + l * output_strides[3] + m * output_strides[4] + n * output_strides[5]] = Compute( |
| 193 | input1[i * input1_strides[0] + j * input1_strides[1] + k * input1_strides[2] + l * input1_strides[3] + m * input1_strides[4] + n * input1_strides[5]], |
| 194 | input2[i * input2_strides[0] + j * input2_strides[1] + k * input2_strides[2] + l * input2_strides[3] + m * input2_strides[4] + n * input2_strides[5]]); |
| 195 | } |
| 196 | } |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 197 | } |
| 198 | } |
| 199 | } |
| 200 | } |
| 201 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 202 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 203 | const float accumulated_range = accumulated_max - accumulated_min; |
| 204 | const float output_min = num_output_elements == 1 ? |
| 205 | -std::numeric_limits<float>::infinity() : accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 206 | const float output_max = num_output_elements == 1 ? |
| 207 | +std::numeric_limits<float>::infinity() : accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 208 | for (float& output_value : output_ref) { |
| 209 | output_value = std::min(std::max(output_value, output_min), output_max); |
| 210 | } |
| 211 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 212 | // Create, setup, run, and destroy a binary elementwise operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 213 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 214 | xnn_operator_t binary_elementwise_op = nullptr; |
| 215 | |
| 216 | switch (operation_type()) { |
| 217 | case OperationType::Add: |
| 218 | ASSERT_EQ(xnn_status_success, |
| 219 | xnn_create_add_nd_f32( |
| 220 | output_min, output_max, |
| 221 | 0, &binary_elementwise_op)); |
| 222 | break; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 223 | case OperationType::Divide: |
| 224 | ASSERT_EQ(xnn_status_success, |
| 225 | xnn_create_divide_nd_f32( |
| 226 | output_min, output_max, |
| 227 | 0, &binary_elementwise_op)); |
| 228 | break; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 229 | case OperationType::Maximum: |
| 230 | ASSERT_EQ(xnn_status_success, |
| 231 | xnn_create_maximum_nd_f32( |
| 232 | 0, &binary_elementwise_op)); |
| 233 | break; |
| 234 | case OperationType::Minimum: |
| 235 | ASSERT_EQ(xnn_status_success, |
| 236 | xnn_create_minimum_nd_f32( |
| 237 | 0, &binary_elementwise_op)); |
| 238 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 239 | case OperationType::Multiply: |
| 240 | ASSERT_EQ(xnn_status_success, |
| 241 | xnn_create_multiply_nd_f32( |
| 242 | output_min, output_max, |
| 243 | 0, &binary_elementwise_op)); |
| 244 | break; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 245 | case OperationType::Subtract: |
| 246 | ASSERT_EQ(xnn_status_success, |
| 247 | xnn_create_subtract_nd_f32( |
| 248 | output_min, output_max, |
| 249 | 0, &binary_elementwise_op)); |
| 250 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 251 | default: |
| 252 | FAIL() << "Unsupported operation type"; |
| 253 | } |
| 254 | ASSERT_NE(nullptr, binary_elementwise_op); |
| 255 | |
| 256 | // Smart pointer to automatically delete binary_elementwise_op. |
| 257 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_binary_elementwise_op(binary_elementwise_op, xnn_delete_operator); |
| 258 | |
| 259 | switch (operation_type()) { |
| 260 | case OperationType::Add: |
| 261 | ASSERT_EQ(xnn_status_success, |
| 262 | xnn_setup_add_nd_f32( |
| 263 | binary_elementwise_op, |
| 264 | num_input1_dims(), |
| 265 | input1_shape().data(), |
| 266 | num_input2_dims(), |
| 267 | input2_shape().data(), |
| 268 | input1.data(), input2.data(), output.data(), |
| 269 | nullptr /* thread pool */)); |
| 270 | break; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 271 | case OperationType::Divide: |
| 272 | ASSERT_EQ(xnn_status_success, |
| 273 | xnn_setup_divide_nd_f32( |
| 274 | binary_elementwise_op, |
| 275 | num_input1_dims(), |
| 276 | input1_shape().data(), |
| 277 | num_input2_dims(), |
| 278 | input2_shape().data(), |
| 279 | input1.data(), input2.data(), output.data(), |
| 280 | nullptr /* thread pool */)); |
| 281 | break; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 282 | case OperationType::Maximum: |
| 283 | ASSERT_EQ(xnn_status_success, |
| 284 | xnn_setup_maximum_nd_f32( |
| 285 | binary_elementwise_op, |
| 286 | num_input1_dims(), |
| 287 | input1_shape().data(), |
| 288 | num_input2_dims(), |
| 289 | input2_shape().data(), |
| 290 | input1.data(), input2.data(), output.data(), |
| 291 | nullptr /* thread pool */)); |
| 292 | break; |
| 293 | case OperationType::Minimum: |
| 294 | ASSERT_EQ(xnn_status_success, |
| 295 | xnn_setup_minimum_nd_f32( |
| 296 | binary_elementwise_op, |
| 297 | num_input1_dims(), |
| 298 | input1_shape().data(), |
| 299 | num_input2_dims(), |
| 300 | input2_shape().data(), |
| 301 | input1.data(), input2.data(), output.data(), |
| 302 | nullptr /* thread pool */)); |
| 303 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 304 | case OperationType::Multiply: |
| 305 | ASSERT_EQ(xnn_status_success, |
| 306 | xnn_setup_multiply_nd_f32( |
| 307 | binary_elementwise_op, |
| 308 | num_input1_dims(), |
| 309 | input1_shape().data(), |
| 310 | num_input2_dims(), |
| 311 | input2_shape().data(), |
| 312 | input1.data(), input2.data(), output.data(), |
| 313 | nullptr /* thread pool */)); |
| 314 | break; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 315 | case OperationType::Subtract: |
| 316 | ASSERT_EQ(xnn_status_success, |
| 317 | xnn_setup_subtract_nd_f32( |
| 318 | binary_elementwise_op, |
| 319 | num_input1_dims(), |
| 320 | input1_shape().data(), |
| 321 | num_input2_dims(), |
| 322 | input2_shape().data(), |
| 323 | input1.data(), input2.data(), output.data(), |
| 324 | nullptr /* thread pool */)); |
| 325 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 326 | default: |
| 327 | FAIL() << "Unsupported operation type"; |
| 328 | } |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 329 | |
| 330 | ASSERT_EQ(xnn_status_success, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 331 | xnn_run_operator(binary_elementwise_op, nullptr /* thread pool */)); |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 332 | |
| 333 | // Verify results. |
| 334 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 335 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 336 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 337 | for (size_t l = 0; l < output_dims[3]; l++) { |
Marat Dukhan | fc2b96e | 2019-12-03 12:04:04 -0800 | [diff] [blame] | 338 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 339 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 340 | const size_t index = |
| 341 | i * output_strides[0] + j * output_strides[1] + k * output_strides[2] + l * output_strides[3] + m * output_strides[4] + n * output_strides[5]; |
| 342 | ASSERT_NEAR(output[index], output_ref[index], 1.0e-6f * std::abs(output_ref[index])) |
| 343 | << "(i, j, k, l, m, n) = (" << i << ", " << j << ", " << k << ", " << l << ", " << m << ", " << n << ")"; |
| 344 | } |
| 345 | } |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 346 | } |
| 347 | } |
| 348 | } |
| 349 | } |
| 350 | } |
| 351 | } |
| 352 | |
| 353 | private: |
| 354 | std::vector<size_t> input1_shape_; |
| 355 | std::vector<size_t> input2_shape_; |
| 356 | uint8_t qmin_{0}; |
| 357 | uint8_t qmax_{255}; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 358 | OperationType operation_type_{OperationType::Unknown}; |
Marat Dukhan | ab4af57 | 2019-12-03 11:11:18 -0800 | [diff] [blame] | 359 | size_t iterations_{3}; |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 360 | }; |