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> |
Marat Dukhan | c5ee9ff | 2020-04-13 01:32:59 -0700 | [diff] [blame] | 18 | #include <numeric> |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 19 | #include <random> |
| 20 | #include <vector> |
| 21 | |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 22 | #include <fp16.h> |
| 23 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 24 | #include <xnnpack.h> |
| 25 | |
| 26 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 27 | class BinaryElementwiseOperatorTester { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 28 | public: |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 29 | enum class OperationType { |
| 30 | Unknown, |
| 31 | Add, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 32 | Divide, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 33 | Maximum, |
| 34 | Minimum, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 35 | Multiply, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 36 | Subtract, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 37 | SquaredDifference, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 38 | }; |
| 39 | |
| 40 | inline BinaryElementwiseOperatorTester& input1_shape(std::initializer_list<size_t> input1_shape) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 41 | assert(input1_shape.size() <= XNN_MAX_TENSOR_DIMS); |
| 42 | this->input1_shape_ = std::vector<size_t>(input1_shape); |
| 43 | return *this; |
| 44 | } |
| 45 | |
| 46 | inline const std::vector<size_t>& input1_shape() const { |
| 47 | return this->input1_shape_; |
| 48 | } |
| 49 | |
| 50 | inline size_t input1_dim(size_t i) const { |
| 51 | return i < num_input1_dims() ? this->input1_shape_[i] : 1; |
| 52 | } |
| 53 | |
| 54 | inline size_t num_input1_dims() const { |
| 55 | return this->input1_shape_.size(); |
| 56 | } |
| 57 | |
| 58 | inline size_t num_input1_elements() const { |
| 59 | return std::accumulate( |
| 60 | this->input1_shape_.begin(), this->input1_shape_.end(), size_t(1), std::multiplies<size_t>()); |
| 61 | } |
| 62 | |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 63 | inline BinaryElementwiseOperatorTester& input1_zero_point(int16_t input1_zero_point) { |
| 64 | this->input1_zero_point_ = input1_zero_point; |
| 65 | return *this; |
| 66 | } |
| 67 | |
| 68 | inline int16_t input1_zero_point() const { |
| 69 | return this->input1_zero_point_; |
| 70 | } |
| 71 | |
| 72 | inline BinaryElementwiseOperatorTester& input1_scale(float input1_scale) { |
| 73 | assert(isfinite(input1_scale)); |
| 74 | this->input1_scale_ = input1_scale; |
| 75 | return *this; |
| 76 | } |
| 77 | |
| 78 | inline float input1_scale() const { |
| 79 | return this->input1_scale_; |
| 80 | } |
| 81 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 82 | inline BinaryElementwiseOperatorTester& input2_shape(std::initializer_list<size_t> input2_shape) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 83 | assert(input2_shape.size() <= XNN_MAX_TENSOR_DIMS); |
| 84 | this->input2_shape_ = std::vector<size_t>(input2_shape); |
| 85 | return *this; |
| 86 | } |
| 87 | |
| 88 | inline const std::vector<size_t>& input2_shape() const { |
| 89 | return this->input2_shape_; |
| 90 | } |
| 91 | |
| 92 | inline size_t input2_dim(size_t i) const { |
| 93 | return i < num_input2_dims() ? this->input2_shape_[i] : 1; |
| 94 | } |
| 95 | |
| 96 | inline size_t num_input2_dims() const { |
| 97 | return this->input2_shape_.size(); |
| 98 | } |
| 99 | |
| 100 | inline size_t num_input2_elements() const { |
| 101 | return std::accumulate( |
| 102 | this->input2_shape_.begin(), this->input2_shape_.end(), size_t(1), std::multiplies<size_t>()); |
| 103 | } |
| 104 | |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 105 | inline BinaryElementwiseOperatorTester& input2_zero_point(int16_t input2_zero_point) { |
| 106 | this->input2_zero_point_ = input2_zero_point; |
| 107 | return *this; |
| 108 | } |
| 109 | |
| 110 | inline int16_t input2_zero_point() const { |
| 111 | return this->input2_zero_point_; |
| 112 | } |
| 113 | |
| 114 | inline BinaryElementwiseOperatorTester& input2_scale(float input2_scale) { |
| 115 | assert(isfinite(input2_scale)); |
| 116 | this->input2_scale_ = input2_scale; |
| 117 | return *this; |
| 118 | } |
| 119 | |
| 120 | inline float input2_scale() const { |
| 121 | return this->input2_scale_; |
| 122 | } |
| 123 | |
| 124 | inline BinaryElementwiseOperatorTester& output_zero_point(int16_t output_zero_point) { |
| 125 | this->output_zero_point_ = output_zero_point; |
| 126 | return *this; |
| 127 | } |
| 128 | |
| 129 | inline int16_t output_zero_point() const { |
| 130 | return this->output_zero_point_; |
| 131 | } |
| 132 | |
| 133 | inline BinaryElementwiseOperatorTester& output_scale(float output_scale) { |
| 134 | assert(isfinite(output_scale)); |
| 135 | this->output_scale_ = output_scale; |
| 136 | return *this; |
| 137 | } |
| 138 | |
| 139 | inline float output_scale() const { |
| 140 | return this->output_scale_; |
| 141 | } |
| 142 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 143 | inline BinaryElementwiseOperatorTester& qmin(uint8_t qmin) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 144 | this->qmin_ = qmin; |
| 145 | return *this; |
| 146 | } |
| 147 | |
| 148 | inline uint8_t qmin() const { |
| 149 | return this->qmin_; |
| 150 | } |
| 151 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 152 | inline BinaryElementwiseOperatorTester& qmax(uint8_t qmax) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 153 | this->qmax_ = qmax; |
| 154 | return *this; |
| 155 | } |
| 156 | |
| 157 | inline uint8_t qmax() const { |
| 158 | return this->qmax_; |
| 159 | } |
| 160 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 161 | inline BinaryElementwiseOperatorTester& operation_type(OperationType operation_type) { |
| 162 | this->operation_type_ = operation_type; |
| 163 | return *this; |
| 164 | } |
| 165 | |
| 166 | inline OperationType operation_type() const { |
| 167 | return this->operation_type_; |
| 168 | } |
| 169 | |
| 170 | inline BinaryElementwiseOperatorTester& iterations(size_t iterations) { |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 171 | this->iterations_ = iterations; |
| 172 | return *this; |
| 173 | } |
| 174 | |
| 175 | inline size_t iterations() const { |
| 176 | return this->iterations_; |
| 177 | } |
| 178 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 179 | float Compute(float a, float b) const { |
| 180 | switch (operation_type()) { |
| 181 | case OperationType::Add: |
| 182 | return a + b; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 183 | case OperationType::Divide: |
| 184 | return a / b; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 185 | case OperationType::Maximum: |
| 186 | return std::max<float>(a, b); |
| 187 | case OperationType::Minimum: |
| 188 | return std::min<float>(a, b); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 189 | case OperationType::Multiply: |
| 190 | return a * b; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 191 | case OperationType::Subtract: |
| 192 | return a - b; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 193 | case OperationType::SquaredDifference: |
| 194 | return (a - b) * (a - b); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 195 | default: |
| 196 | return std::nanf(""); |
| 197 | } |
| 198 | } |
| 199 | |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 200 | void TestQS8() const { |
| 201 | ASSERT_NE(operation_type(), OperationType::Unknown); |
| 202 | ASSERT_GE(input1_zero_point(), std::numeric_limits<int8_t>::min()); |
| 203 | ASSERT_LE(input1_zero_point(), std::numeric_limits<int8_t>::max()); |
| 204 | ASSERT_GE(input2_zero_point(), std::numeric_limits<int8_t>::min()); |
| 205 | ASSERT_LE(input2_zero_point(), std::numeric_limits<int8_t>::max()); |
| 206 | ASSERT_GE(output_zero_point(), std::numeric_limits<int8_t>::min()); |
| 207 | ASSERT_LE(output_zero_point(), std::numeric_limits<int8_t>::max()); |
| 208 | |
| 209 | std::random_device random_device; |
| 210 | auto rng = std::mt19937(random_device()); |
| 211 | auto i8rng = std::bind( |
| 212 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), std::ref(rng)); |
| 213 | |
| 214 | // Compute generalized shapes. |
| 215 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_dims; |
| 216 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_dims; |
| 217 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_dims; |
| 218 | std::fill(input1_dims.begin(), input1_dims.end(), 1); |
| 219 | std::fill(input2_dims.begin(), input2_dims.end(), 1); |
| 220 | std::fill(output_dims.begin(), output_dims.end(), 1); |
| 221 | std::copy(input1_shape().cbegin(), input1_shape().cend(), input1_dims.end() - num_input1_dims()); |
| 222 | std::copy(input2_shape().cbegin(), input2_shape().cend(), input2_dims.end() - num_input2_dims()); |
| 223 | for (size_t i = 0; i < XNN_MAX_TENSOR_DIMS; i++) { |
| 224 | if (input1_dims[i] != 1 && input2_dims[i] != 1) { |
| 225 | ASSERT_EQ(input1_dims[i], input2_dims[i]); |
| 226 | } |
| 227 | output_dims[i] = std::max(input1_dims[i], input2_dims[i]); |
| 228 | } |
| 229 | const size_t num_output_elements = |
| 230 | std::accumulate(output_dims.begin(), output_dims.end(), size_t(1), std::multiplies<size_t>()); |
| 231 | |
| 232 | // Compute generalized strides. |
| 233 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_strides; |
| 234 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_strides; |
| 235 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_strides; |
| 236 | size_t input1_stride = 1, input2_stride = 1, output_stride = 1; |
| 237 | for (size_t i = XNN_MAX_TENSOR_DIMS; i != 0; i--) { |
| 238 | input1_strides[i - 1] = input1_dims[i - 1] == 1 ? 0 : input1_stride; |
| 239 | input2_strides[i - 1] = input2_dims[i - 1] == 1 ? 0 : input2_stride; |
| 240 | output_strides[i - 1] = output_stride; |
| 241 | input1_stride *= input1_dims[i - 1]; |
| 242 | input2_stride *= input2_dims[i - 1]; |
| 243 | output_stride *= output_dims[i - 1]; |
| 244 | } |
| 245 | |
| 246 | std::vector<int8_t> input1(XNN_EXTRA_BYTES / sizeof(uint16_t) + num_input1_elements()); |
| 247 | std::vector<int8_t> input2(XNN_EXTRA_BYTES / sizeof(uint16_t) + num_input2_elements()); |
| 248 | std::vector<int8_t> output(num_output_elements); |
| 249 | std::vector<float> output_ref(num_output_elements); |
| 250 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 251 | std::generate(input1.begin(), input1.end(), std::ref(i8rng)); |
| 252 | std::generate(input2.begin(), input2.end(), std::ref(i8rng)); |
| 253 | std::fill(output.begin(), output.end(), 0xAA); |
| 254 | |
| 255 | // Compute reference results. |
| 256 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 257 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 258 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 259 | for (size_t l = 0; l < output_dims[3]; l++) { |
| 260 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 261 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 262 | 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( |
| 263 | input1_scale() * (int32_t(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]]) - input1_zero_point()), |
| 264 | input2_scale() * (int32_t(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]]) - input2_zero_point())) / |
| 265 | output_scale() + float(output_zero_point()); |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | } |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | for (float& output_value : output_ref) { |
| 274 | output_value = std::min(std::max(output_value, float(int8_t(qmin() - 0x80))), float(int8_t(qmax() - 0x80))); |
| 275 | } |
| 276 | |
| 277 | // Create, setup, run, and destroy a binary elementwise operator. |
| 278 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 279 | xnn_operator_t binary_elementwise_op = nullptr; |
| 280 | xnn_status status = xnn_status_unsupported_parameter; |
| 281 | switch (operation_type()) { |
| 282 | case OperationType::Add: |
| 283 | status = xnn_create_add_nd_qs8( |
| 284 | input1_zero_point(), input1_scale(), |
| 285 | input2_zero_point(), input2_scale(), |
| 286 | output_zero_point(), output_scale(), |
| 287 | int8_t(qmin() - 0x80), int8_t(qmax() - 0x80), |
| 288 | 0, &binary_elementwise_op); |
| 289 | break; |
| 290 | default: |
| 291 | FAIL() << "Unsupported operation type"; |
| 292 | } |
| 293 | if (status == xnn_status_unsupported_hardware) { |
| 294 | GTEST_SKIP(); |
| 295 | } |
| 296 | ASSERT_EQ(xnn_status_success, status); |
| 297 | ASSERT_NE(nullptr, binary_elementwise_op); |
| 298 | |
| 299 | // Smart pointer to automatically delete binary_elementwise_op. |
| 300 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_binary_elementwise_op(binary_elementwise_op, xnn_delete_operator); |
| 301 | |
| 302 | switch (operation_type()) { |
| 303 | case OperationType::Add: |
| 304 | ASSERT_EQ(xnn_status_success, |
| 305 | xnn_setup_add_nd_qs8( |
| 306 | binary_elementwise_op, |
| 307 | num_input1_dims(), |
| 308 | input1_shape().data(), |
| 309 | num_input2_dims(), |
| 310 | input2_shape().data(), |
| 311 | input1.data(), input2.data(), output.data(), |
| 312 | nullptr /* thread pool */)); |
| 313 | break; |
| 314 | default: |
| 315 | FAIL() << "Unsupported operation type"; |
| 316 | } |
| 317 | |
| 318 | ASSERT_EQ(xnn_status_success, |
| 319 | xnn_run_operator(binary_elementwise_op, nullptr /* thread pool */)); |
| 320 | |
| 321 | // Verify results. |
| 322 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 323 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 324 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 325 | for (size_t l = 0; l < output_dims[3]; l++) { |
| 326 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 327 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 328 | const size_t index = |
| 329 | i * output_strides[0] + j * output_strides[1] + k * output_strides[2] + l * output_strides[3] + m * output_strides[4] + n * output_strides[5]; |
| 330 | ASSERT_NEAR(float(output[index]), output_ref[index], 0.6f) |
| 331 | << "(i, j, k, l, m, n) = (" << i << ", " << j << ", " << k << ", " << l << ", " << m << ", " << n << ")" |
| 332 | << ", input1 zero point = " << input1_zero_point() << ", input1 scale = " << input1_scale() |
| 333 | << ", input2 zero point = " << input2_zero_point() << ", input2 scale = " << input2_scale() |
| 334 | << ", output zero point = " << output_zero_point() << ", output scale = " << output_scale(); |
| 335 | } |
| 336 | } |
| 337 | } |
| 338 | } |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | } |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 343 | |
| 344 | void TestF16() const { |
| 345 | ASSERT_NE(operation_type(), OperationType::Unknown); |
| 346 | |
| 347 | std::random_device random_device; |
| 348 | auto rng = std::mt19937(random_device()); |
Frank Barchard | 7d2c1f2 | 2020-09-14 16:43:53 -0700 | [diff] [blame] | 349 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 350 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 351 | |
| 352 | // Compute generalized shapes. |
| 353 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_dims; |
| 354 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_dims; |
| 355 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_dims; |
| 356 | std::fill(input1_dims.begin(), input1_dims.end(), 1); |
| 357 | std::fill(input2_dims.begin(), input2_dims.end(), 1); |
| 358 | std::fill(output_dims.begin(), output_dims.end(), 1); |
| 359 | std::copy(input1_shape().cbegin(), input1_shape().cend(), input1_dims.end() - num_input1_dims()); |
| 360 | std::copy(input2_shape().cbegin(), input2_shape().cend(), input2_dims.end() - num_input2_dims()); |
| 361 | for (size_t i = 0; i < XNN_MAX_TENSOR_DIMS; i++) { |
| 362 | if (input1_dims[i] != 1 && input2_dims[i] != 1) { |
| 363 | ASSERT_EQ(input1_dims[i], input2_dims[i]); |
| 364 | } |
| 365 | output_dims[i] = std::max(input1_dims[i], input2_dims[i]); |
| 366 | } |
| 367 | const size_t num_output_elements = |
| 368 | std::accumulate(output_dims.begin(), output_dims.end(), size_t(1), std::multiplies<size_t>()); |
| 369 | |
| 370 | // Compute generalized strides. |
| 371 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_strides; |
| 372 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_strides; |
| 373 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_strides; |
| 374 | size_t input1_stride = 1, input2_stride = 1, output_stride = 1; |
| 375 | for (size_t i = XNN_MAX_TENSOR_DIMS; i != 0; i--) { |
| 376 | input1_strides[i - 1] = input1_dims[i - 1] == 1 ? 0 : input1_stride; |
| 377 | input2_strides[i - 1] = input2_dims[i - 1] == 1 ? 0 : input2_stride; |
| 378 | output_strides[i - 1] = output_stride; |
| 379 | input1_stride *= input1_dims[i - 1]; |
| 380 | input2_stride *= input2_dims[i - 1]; |
| 381 | output_stride *= output_dims[i - 1]; |
| 382 | } |
| 383 | |
| 384 | std::vector<uint16_t> input1(XNN_EXTRA_BYTES / sizeof(uint16_t) + num_input1_elements()); |
| 385 | std::vector<uint16_t> input2(XNN_EXTRA_BYTES / sizeof(uint16_t) + num_input2_elements()); |
| 386 | std::vector<uint16_t> output(num_output_elements); |
| 387 | std::vector<float> output_ref(num_output_elements); |
| 388 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 389 | std::generate(input1.begin(), input1.end(), std::ref(f16rng)); |
| 390 | std::generate(input2.begin(), input2.end(), std::ref(f16rng)); |
| 391 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 392 | |
| 393 | // Compute reference results. |
| 394 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 395 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 396 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 397 | for (size_t l = 0; l < output_dims[3]; l++) { |
| 398 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 399 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 400 | 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( |
| 401 | fp16_ieee_to_fp32_value(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]]), |
| 402 | fp16_ieee_to_fp32_value(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]])); |
| 403 | } |
| 404 | } |
| 405 | } |
| 406 | } |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | // Compute clamping parameters. |
| 411 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 412 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 413 | const float accumulated_range = accumulated_max - accumulated_min; |
| 414 | const float scaled_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_min + accumulated_range / 255.0f * float(qmin()))); |
| 415 | const float scaled_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_max - accumulated_range / 255.0f * float(255 - qmax()))); |
| 416 | const float output_min = scaled_min == scaled_max ? -std::numeric_limits<float>::infinity() : scaled_min; |
| 417 | const float output_max = scaled_min == scaled_max ? +std::numeric_limits<float>::infinity() : scaled_max; |
| 418 | |
| 419 | for (float& output_value : output_ref) { |
| 420 | output_value = std::min(std::max(output_value, output_min), output_max); |
| 421 | } |
| 422 | |
| 423 | // Create, setup, run, and destroy a binary elementwise operator. |
| 424 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 425 | xnn_operator_t binary_elementwise_op = nullptr; |
| 426 | xnn_status status = xnn_status_unsupported_parameter; |
| 427 | switch (operation_type()) { |
| 428 | case OperationType::Add: |
| 429 | status = xnn_create_add_nd_f16(output_min, output_max, 0, &binary_elementwise_op); |
| 430 | break; |
Frank Barchard | 0ea6a77 | 2020-09-09 15:26:31 -0700 | [diff] [blame] | 431 | case OperationType::Multiply: |
| 432 | status = xnn_create_multiply_nd_f16(output_min, output_max, 0, &binary_elementwise_op); |
| 433 | break; |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 434 | default: |
| 435 | FAIL() << "Unsupported operation type"; |
| 436 | } |
| 437 | if (status == xnn_status_unsupported_hardware) { |
| 438 | GTEST_SKIP(); |
| 439 | } |
| 440 | ASSERT_EQ(xnn_status_success, status); |
| 441 | ASSERT_NE(nullptr, binary_elementwise_op); |
| 442 | |
| 443 | // Smart pointer to automatically delete binary_elementwise_op. |
| 444 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_binary_elementwise_op(binary_elementwise_op, xnn_delete_operator); |
| 445 | |
| 446 | switch (operation_type()) { |
| 447 | case OperationType::Add: |
| 448 | ASSERT_EQ(xnn_status_success, |
| 449 | xnn_setup_add_nd_f16( |
| 450 | binary_elementwise_op, |
| 451 | num_input1_dims(), |
| 452 | input1_shape().data(), |
| 453 | num_input2_dims(), |
| 454 | input2_shape().data(), |
| 455 | input1.data(), input2.data(), output.data(), |
| 456 | nullptr /* thread pool */)); |
| 457 | break; |
Frank Barchard | 0ea6a77 | 2020-09-09 15:26:31 -0700 | [diff] [blame] | 458 | case OperationType::Multiply: |
| 459 | ASSERT_EQ(xnn_status_success, |
| 460 | xnn_setup_multiply_nd_f16( |
| 461 | binary_elementwise_op, |
| 462 | num_input1_dims(), |
| 463 | input1_shape().data(), |
| 464 | num_input2_dims(), |
| 465 | input2_shape().data(), |
| 466 | input1.data(), input2.data(), output.data(), |
| 467 | nullptr /* thread pool */)); |
| 468 | break; |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 469 | default: |
| 470 | FAIL() << "Unsupported operation type"; |
| 471 | } |
| 472 | |
| 473 | ASSERT_EQ(xnn_status_success, |
| 474 | xnn_run_operator(binary_elementwise_op, nullptr /* thread pool */)); |
| 475 | |
| 476 | // Verify results. |
| 477 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 478 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 479 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 480 | for (size_t l = 0; l < output_dims[3]; l++) { |
| 481 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 482 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 483 | const size_t index = |
| 484 | i * output_strides[0] + j * output_strides[1] + k * output_strides[2] + l * output_strides[3] + m * output_strides[4] + n * output_strides[5]; |
Frank Barchard | 2b9d29b | 2020-09-17 12:03:39 -0700 | [diff] [blame^] | 485 | ASSERT_NEAR(fp16_ieee_to_fp32_value(output[index]), output_ref[index], std::max(1.0e-4f, std::abs(output_ref[index]) * 1.0e-2f)) |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 486 | << "(i, j, k, l, m, n) = (" << i << ", " << j << ", " << k << ", " << l << ", " << m << ", " << n << ")"; |
| 487 | } |
| 488 | } |
| 489 | } |
| 490 | } |
| 491 | } |
| 492 | } |
| 493 | } |
| 494 | } |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 495 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 496 | void TestF32() const { |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 497 | ASSERT_NE(operation_type(), OperationType::Unknown); |
| 498 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 499 | std::random_device random_device; |
| 500 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 501 | 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] | 502 | |
| 503 | // Compute generalized shapes. |
| 504 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_dims; |
| 505 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_dims; |
| 506 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_dims; |
| 507 | std::fill(input1_dims.begin(), input1_dims.end(), 1); |
| 508 | std::fill(input2_dims.begin(), input2_dims.end(), 1); |
| 509 | std::fill(output_dims.begin(), output_dims.end(), 1); |
| 510 | std::copy(input1_shape().cbegin(), input1_shape().cend(), input1_dims.end() - num_input1_dims()); |
| 511 | std::copy(input2_shape().cbegin(), input2_shape().cend(), input2_dims.end() - num_input2_dims()); |
| 512 | for (size_t i = 0; i < XNN_MAX_TENSOR_DIMS; i++) { |
| 513 | if (input1_dims[i] != 1 && input2_dims[i] != 1) { |
| 514 | ASSERT_EQ(input1_dims[i], input2_dims[i]); |
| 515 | } |
| 516 | output_dims[i] = std::max(input1_dims[i], input2_dims[i]); |
| 517 | } |
| 518 | const size_t num_output_elements = |
| 519 | std::accumulate(output_dims.begin(), output_dims.end(), size_t(1), std::multiplies<size_t>()); |
| 520 | |
| 521 | // Compute generalized strides. |
| 522 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input1_strides; |
| 523 | std::array<size_t, XNN_MAX_TENSOR_DIMS> input2_strides; |
| 524 | std::array<size_t, XNN_MAX_TENSOR_DIMS> output_strides; |
| 525 | size_t input1_stride = 1, input2_stride = 1, output_stride = 1; |
| 526 | for (size_t i = XNN_MAX_TENSOR_DIMS; i != 0; i--) { |
| 527 | input1_strides[i - 1] = input1_dims[i - 1] == 1 ? 0 : input1_stride; |
| 528 | input2_strides[i - 1] = input2_dims[i - 1] == 1 ? 0 : input2_stride; |
| 529 | output_strides[i - 1] = output_stride; |
| 530 | input1_stride *= input1_dims[i - 1]; |
| 531 | input2_stride *= input2_dims[i - 1]; |
| 532 | output_stride *= output_dims[i - 1]; |
| 533 | } |
| 534 | |
| 535 | std::vector<float> input1(XNN_EXTRA_BYTES / sizeof(float) + num_input1_elements()); |
| 536 | std::vector<float> input2(XNN_EXTRA_BYTES / sizeof(float) + num_input2_elements()); |
| 537 | std::vector<float> output(num_output_elements); |
| 538 | std::vector<float> output_ref(num_output_elements); |
| 539 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 540 | std::generate(input1.begin(), input1.end(), std::ref(f32rng)); |
| 541 | std::generate(input2.begin(), input2.end(), std::ref(f32rng)); |
| 542 | std::fill(output.begin(), output.end(), nanf("")); |
| 543 | |
| 544 | // Compute reference results. |
| 545 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 546 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 547 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 548 | for (size_t l = 0; l < output_dims[3]; l++) { |
Marat Dukhan | fc2b96e | 2019-12-03 12:04:04 -0800 | [diff] [blame] | 549 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 550 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 551 | 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( |
| 552 | 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]], |
| 553 | 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]]); |
| 554 | } |
| 555 | } |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 556 | } |
| 557 | } |
| 558 | } |
| 559 | } |
| 560 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 561 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 562 | const float accumulated_range = accumulated_max - accumulated_min; |
| 563 | const float output_min = num_output_elements == 1 ? |
| 564 | -std::numeric_limits<float>::infinity() : accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 565 | const float output_max = num_output_elements == 1 ? |
| 566 | +std::numeric_limits<float>::infinity() : accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 567 | for (float& output_value : output_ref) { |
| 568 | output_value = std::min(std::max(output_value, output_min), output_max); |
| 569 | } |
| 570 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 571 | // Create, setup, run, and destroy a binary elementwise operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 572 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 573 | xnn_operator_t binary_elementwise_op = nullptr; |
| 574 | |
| 575 | switch (operation_type()) { |
| 576 | case OperationType::Add: |
| 577 | ASSERT_EQ(xnn_status_success, |
| 578 | xnn_create_add_nd_f32( |
| 579 | output_min, output_max, |
| 580 | 0, &binary_elementwise_op)); |
| 581 | break; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 582 | case OperationType::Divide: |
| 583 | ASSERT_EQ(xnn_status_success, |
| 584 | xnn_create_divide_nd_f32( |
| 585 | output_min, output_max, |
| 586 | 0, &binary_elementwise_op)); |
| 587 | break; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 588 | case OperationType::Maximum: |
| 589 | ASSERT_EQ(xnn_status_success, |
| 590 | xnn_create_maximum_nd_f32( |
| 591 | 0, &binary_elementwise_op)); |
| 592 | break; |
| 593 | case OperationType::Minimum: |
| 594 | ASSERT_EQ(xnn_status_success, |
| 595 | xnn_create_minimum_nd_f32( |
| 596 | 0, &binary_elementwise_op)); |
| 597 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 598 | case OperationType::Multiply: |
| 599 | ASSERT_EQ(xnn_status_success, |
| 600 | xnn_create_multiply_nd_f32( |
| 601 | output_min, output_max, |
| 602 | 0, &binary_elementwise_op)); |
| 603 | break; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 604 | case OperationType::Subtract: |
| 605 | ASSERT_EQ(xnn_status_success, |
| 606 | xnn_create_subtract_nd_f32( |
| 607 | output_min, output_max, |
| 608 | 0, &binary_elementwise_op)); |
| 609 | break; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 610 | case OperationType::SquaredDifference: |
| 611 | ASSERT_EQ(xnn_status_success, |
| 612 | xnn_create_squared_difference_nd_f32( |
| 613 | 0, &binary_elementwise_op)); |
| 614 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 615 | default: |
| 616 | FAIL() << "Unsupported operation type"; |
| 617 | } |
| 618 | ASSERT_NE(nullptr, binary_elementwise_op); |
| 619 | |
| 620 | // Smart pointer to automatically delete binary_elementwise_op. |
| 621 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_binary_elementwise_op(binary_elementwise_op, xnn_delete_operator); |
| 622 | |
| 623 | switch (operation_type()) { |
| 624 | case OperationType::Add: |
| 625 | ASSERT_EQ(xnn_status_success, |
| 626 | xnn_setup_add_nd_f32( |
| 627 | binary_elementwise_op, |
| 628 | num_input1_dims(), |
| 629 | input1_shape().data(), |
| 630 | num_input2_dims(), |
| 631 | input2_shape().data(), |
| 632 | input1.data(), input2.data(), output.data(), |
| 633 | nullptr /* thread pool */)); |
| 634 | break; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 635 | case OperationType::Divide: |
| 636 | ASSERT_EQ(xnn_status_success, |
| 637 | xnn_setup_divide_nd_f32( |
| 638 | binary_elementwise_op, |
| 639 | num_input1_dims(), |
| 640 | input1_shape().data(), |
| 641 | num_input2_dims(), |
| 642 | input2_shape().data(), |
| 643 | input1.data(), input2.data(), output.data(), |
| 644 | nullptr /* thread pool */)); |
| 645 | break; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 646 | case OperationType::Maximum: |
| 647 | ASSERT_EQ(xnn_status_success, |
| 648 | xnn_setup_maximum_nd_f32( |
| 649 | binary_elementwise_op, |
| 650 | num_input1_dims(), |
| 651 | input1_shape().data(), |
| 652 | num_input2_dims(), |
| 653 | input2_shape().data(), |
| 654 | input1.data(), input2.data(), output.data(), |
| 655 | nullptr /* thread pool */)); |
| 656 | break; |
| 657 | case OperationType::Minimum: |
| 658 | ASSERT_EQ(xnn_status_success, |
| 659 | xnn_setup_minimum_nd_f32( |
| 660 | binary_elementwise_op, |
| 661 | num_input1_dims(), |
| 662 | input1_shape().data(), |
| 663 | num_input2_dims(), |
| 664 | input2_shape().data(), |
| 665 | input1.data(), input2.data(), output.data(), |
| 666 | nullptr /* thread pool */)); |
| 667 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 668 | case OperationType::Multiply: |
| 669 | ASSERT_EQ(xnn_status_success, |
| 670 | xnn_setup_multiply_nd_f32( |
| 671 | binary_elementwise_op, |
| 672 | num_input1_dims(), |
| 673 | input1_shape().data(), |
| 674 | num_input2_dims(), |
| 675 | input2_shape().data(), |
| 676 | input1.data(), input2.data(), output.data(), |
| 677 | nullptr /* thread pool */)); |
| 678 | break; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 679 | case OperationType::Subtract: |
| 680 | ASSERT_EQ(xnn_status_success, |
| 681 | xnn_setup_subtract_nd_f32( |
| 682 | binary_elementwise_op, |
| 683 | num_input1_dims(), |
| 684 | input1_shape().data(), |
| 685 | num_input2_dims(), |
| 686 | input2_shape().data(), |
| 687 | input1.data(), input2.data(), output.data(), |
| 688 | nullptr /* thread pool */)); |
| 689 | break; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 690 | case OperationType::SquaredDifference: |
| 691 | ASSERT_EQ(xnn_status_success, |
| 692 | xnn_setup_squared_difference_nd_f32( |
| 693 | binary_elementwise_op, |
| 694 | num_input1_dims(), |
| 695 | input1_shape().data(), |
| 696 | num_input2_dims(), |
| 697 | input2_shape().data(), |
| 698 | input1.data(), input2.data(), output.data(), |
| 699 | nullptr /* thread pool */)); |
| 700 | break; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 701 | default: |
| 702 | FAIL() << "Unsupported operation type"; |
| 703 | } |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 704 | |
| 705 | ASSERT_EQ(xnn_status_success, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 706 | xnn_run_operator(binary_elementwise_op, nullptr /* thread pool */)); |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 707 | |
| 708 | // Verify results. |
| 709 | for (size_t i = 0; i < output_dims[0]; i++) { |
| 710 | for (size_t j = 0; j < output_dims[1]; j++) { |
| 711 | for (size_t k = 0; k < output_dims[2]; k++) { |
| 712 | for (size_t l = 0; l < output_dims[3]; l++) { |
Marat Dukhan | fc2b96e | 2019-12-03 12:04:04 -0800 | [diff] [blame] | 713 | for (size_t m = 0; m < output_dims[4]; m++) { |
| 714 | for (size_t n = 0; n < output_dims[5]; n++) { |
| 715 | const size_t index = |
| 716 | i * output_strides[0] + j * output_strides[1] + k * output_strides[2] + l * output_strides[3] + m * output_strides[4] + n * output_strides[5]; |
| 717 | ASSERT_NEAR(output[index], output_ref[index], 1.0e-6f * std::abs(output_ref[index])) |
| 718 | << "(i, j, k, l, m, n) = (" << i << ", " << j << ", " << k << ", " << l << ", " << m << ", " << n << ")"; |
| 719 | } |
| 720 | } |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 721 | } |
| 722 | } |
| 723 | } |
| 724 | } |
| 725 | } |
| 726 | } |
| 727 | |
| 728 | private: |
| 729 | std::vector<size_t> input1_shape_; |
| 730 | std::vector<size_t> input2_shape_; |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 731 | int16_t input1_zero_point_{0}; |
| 732 | float input1_scale_{1.0f}; |
| 733 | int16_t input2_zero_point_{0}; |
| 734 | float input2_scale_{1.0f}; |
| 735 | int16_t output_zero_point_{0}; |
| 736 | float output_scale_{1.0f}; |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 737 | uint8_t qmin_{0}; |
| 738 | uint8_t qmax_{255}; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 739 | OperationType operation_type_{OperationType::Unknown}; |
Marat Dukhan | ab4af57 | 2019-12-03 11:11:18 -0800 | [diff] [blame] | 740 | size_t iterations_{3}; |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 741 | }; |