XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [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 | |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 10 | #include <fp16.h> |
| 11 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 12 | #include <algorithm> |
| 13 | #include <cmath> |
| 14 | #include <cstddef> |
| 15 | #include <cstdlib> |
| 16 | #include <functional> |
| 17 | #include <random> |
| 18 | #include <vector> |
| 19 | |
| 20 | #include <xnnpack.h> |
| 21 | |
| 22 | |
| 23 | class PReLUOperatorTester { |
| 24 | public: |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 25 | enum class WeightsType { |
| 26 | Default, |
| 27 | FP32, |
| 28 | }; |
| 29 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 30 | inline PReLUOperatorTester& batch_size(size_t batch_size) { |
| 31 | assert(batch_size != 0); |
| 32 | this->batch_size_ = batch_size; |
| 33 | return *this; |
| 34 | } |
| 35 | |
| 36 | inline size_t batch_size() const { |
| 37 | return this->batch_size_; |
| 38 | } |
| 39 | |
| 40 | inline PReLUOperatorTester& channels(size_t channels) { |
| 41 | assert(channels != 0); |
| 42 | this->channels_ = channels; |
| 43 | return *this; |
| 44 | } |
| 45 | |
| 46 | inline size_t channels() const { |
| 47 | return this->channels_; |
| 48 | } |
| 49 | |
| 50 | inline PReLUOperatorTester& x_stride(size_t x_stride) { |
| 51 | assert(x_stride != 0); |
| 52 | this->x_stride_ = x_stride; |
| 53 | return *this; |
| 54 | } |
| 55 | |
| 56 | inline size_t x_stride() const { |
| 57 | if (this->x_stride_ == 0) { |
| 58 | return this->channels_; |
| 59 | } else { |
| 60 | assert(this->x_stride_ >= this->channels_); |
| 61 | return this->x_stride_; |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | inline PReLUOperatorTester& y_stride(size_t y_stride) { |
| 66 | assert(y_stride != 0); |
| 67 | this->y_stride_ = y_stride; |
| 68 | return *this; |
| 69 | } |
| 70 | |
| 71 | inline size_t y_stride() const { |
| 72 | if (this->y_stride_ == 0) { |
| 73 | return this->channels_; |
| 74 | } else { |
| 75 | assert(this->y_stride_ >= this->channels_); |
| 76 | return this->y_stride_; |
| 77 | } |
| 78 | } |
| 79 | |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 80 | inline PReLUOperatorTester& weights_type(WeightsType weights_type) { |
| 81 | this->weights_type_ = weights_type; |
| 82 | return *this; |
| 83 | } |
| 84 | |
| 85 | inline WeightsType weights_type() const { |
| 86 | return this->weights_type_; |
| 87 | } |
| 88 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 89 | inline PReLUOperatorTester& iterations(size_t iterations) { |
| 90 | this->iterations_ = iterations; |
| 91 | return *this; |
| 92 | } |
| 93 | |
| 94 | inline size_t iterations() const { |
| 95 | return this->iterations_; |
| 96 | } |
| 97 | |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 98 | void TestF16() const { |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 99 | switch (weights_type()) { |
| 100 | case WeightsType::Default: |
| 101 | break; |
| 102 | case WeightsType::FP32: |
| 103 | break; |
| 104 | default: |
| 105 | GTEST_FAIL() << "unexpected weights type"; |
| 106 | } |
| 107 | |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 108 | std::random_device random_device; |
| 109 | auto rng = std::mt19937(random_device()); |
| 110 | auto f32irng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), rng); |
| 111 | auto f16irng = std::bind(fp16_ieee_from_fp32_value, f32irng); |
| 112 | auto f32wrng = std::bind(std::uniform_real_distribution<float>(0.25f, 0.75f), rng); |
| 113 | auto f16wrng = std::bind(fp16_ieee_from_fp32_value, f32wrng); |
| 114 | |
| 115 | std::vector<uint16_t> x((batch_size() - 1) * x_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 116 | std::vector<uint16_t> w(channels()); |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 117 | std::vector<float> w_as_float(channels()); |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 118 | std::vector<uint16_t> y((batch_size() - 1) * y_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 119 | std::vector<float> y_ref(batch_size() * channels()); |
| 120 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 121 | std::generate(x.begin(), x.end(), std::ref(f16irng)); |
| 122 | std::generate(w.begin(), w.end(), std::ref(f16wrng)); |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 123 | std::transform(w.cbegin(), w.cend(), w_as_float.begin(), fp16_ieee_to_fp32_value); |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 124 | std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */); |
| 125 | |
| 126 | // Compute reference results, without clamping. |
| 127 | for (size_t i = 0; i < batch_size(); i++) { |
| 128 | for (size_t c = 0; c < channels(); c++) { |
| 129 | const float x_value = fp16_ieee_to_fp32_value(x[i * x_stride() + c]); |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 130 | const float w_value = w_as_float[c]; |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 131 | y_ref[i * channels() + c] = signbit(x_value) ? x_value * w_value : x_value; |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | // Create, setup, run, and destroy PReLU operator. |
| 136 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 137 | xnn_operator_t prelu_op = nullptr; |
| 138 | |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 139 | const void* negative_slope_data = w.data(); |
| 140 | if (weights_type() == WeightsType::FP32) { |
| 141 | negative_slope_data = w_as_float.data(); |
| 142 | } |
| 143 | uint32_t flags = 0; |
| 144 | if (weights_type() == WeightsType::FP32) { |
| 145 | flags |= XNN_FLAG_FP32_STATIC_WEIGHTS; |
| 146 | } |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 147 | ASSERT_EQ(xnn_status_success, |
| 148 | xnn_create_prelu_nc_f16( |
| 149 | channels(), x_stride(), y_stride(), |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 150 | negative_slope_data, |
| 151 | flags, &prelu_op)); |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 152 | ASSERT_NE(nullptr, prelu_op); |
| 153 | |
| 154 | // Smart pointer to automatically delete prelu_op. |
| 155 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_prelu_op(prelu_op, xnn_delete_operator); |
| 156 | |
| 157 | ASSERT_EQ(xnn_status_success, |
| 158 | xnn_setup_prelu_nc_f16( |
| 159 | prelu_op, |
| 160 | batch_size(), |
| 161 | x.data(), y.data(), |
| 162 | nullptr /* thread pool */)); |
| 163 | |
| 164 | ASSERT_EQ(xnn_status_success, |
| 165 | xnn_run_operator(prelu_op, nullptr /* thread pool */)); |
| 166 | |
| 167 | // Verify results. |
| 168 | for (size_t i = 0; i < batch_size(); i++) { |
| 169 | for (size_t c = 0; c < channels(); c++) { |
| 170 | ASSERT_NEAR( |
| 171 | fp16_ieee_to_fp32_value(y[i * y_stride() + c]), |
| 172 | y_ref[i * channels() + c], |
| 173 | std::max(1.0e-4f, std::abs(y_ref[i * channels() + c]) * 1.0e-4f)) |
| 174 | << "at position " << i << " / " << batch_size() << ", channel " << c << " / " << channels(); |
| 175 | } |
| 176 | } |
| 177 | } |
| 178 | } |
| 179 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 180 | void TestF32() const { |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 181 | ASSERT_EQ(weights_type(), WeightsType::Default); |
| 182 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 183 | std::random_device random_device; |
| 184 | auto rng = std::mt19937(random_device()); |
| 185 | auto f32irng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), rng); |
| 186 | auto f32wrng = std::bind(std::uniform_real_distribution<float>(0.25f, 0.75f), rng); |
| 187 | |
| 188 | std::vector<float> x((batch_size() - 1) * x_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float)); |
| 189 | std::vector<float> w(channels()); |
| 190 | std::vector<float> y((batch_size() - 1) * y_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float)); |
| 191 | std::vector<float> y_ref(batch_size() * channels()); |
| 192 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 193 | std::generate(x.begin(), x.end(), std::ref(f32irng)); |
| 194 | std::generate(w.begin(), w.end(), std::ref(f32wrng)); |
| 195 | std::fill(y.begin(), y.end(), nanf("")); |
| 196 | |
| 197 | // Compute reference results, without clamping. |
| 198 | for (size_t i = 0; i < batch_size(); i++) { |
| 199 | for (size_t c = 0; c < channels(); c++) { |
Marat Dukhan | 629a33e | 2019-10-01 10:39:14 -0700 | [diff] [blame] | 200 | y_ref[i * channels() + c] = std::signbit(x[i * x_stride() + c]) ? x[i * x_stride() + c] * w[c] : x[i * x_stride() + c]; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 201 | } |
| 202 | } |
| 203 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 204 | // Create, setup, run, and destroy PReLU operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 205 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 206 | xnn_operator_t prelu_op = nullptr; |
| 207 | |
| 208 | ASSERT_EQ(xnn_status_success, |
| 209 | xnn_create_prelu_nc_f32( |
| 210 | channels(), x_stride(), y_stride(), |
| 211 | w.data(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 212 | 0, &prelu_op)); |
| 213 | ASSERT_NE(nullptr, prelu_op); |
| 214 | |
| 215 | // Smart pointer to automatically delete prelu_op. |
| 216 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_prelu_op(prelu_op, xnn_delete_operator); |
| 217 | |
| 218 | ASSERT_EQ(xnn_status_success, |
| 219 | xnn_setup_prelu_nc_f32( |
| 220 | prelu_op, |
| 221 | batch_size(), |
| 222 | x.data(), y.data(), |
| 223 | nullptr /* thread pool */)); |
| 224 | |
| 225 | ASSERT_EQ(xnn_status_success, |
| 226 | xnn_run_operator(prelu_op, nullptr /* thread pool */)); |
| 227 | |
| 228 | // Verify results. |
| 229 | for (size_t i = 0; i < batch_size(); i++) { |
| 230 | for (size_t c = 0; c < channels(); c++) { |
Marat Dukhan | 0a756b5 | 2022-02-03 23:08:50 -0800 | [diff] [blame] | 231 | ASSERT_NEAR( |
| 232 | y[i * y_stride() + c], |
| 233 | y_ref[i * channels() + c], |
| 234 | std::max(1.0e-6f, std::abs(y_ref[i * channels() + c]) * 1.0e-6f)) |
| 235 | << "at position " << i << " / " << batch_size() << ", channel " << c << " / " << channels(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 236 | } |
| 237 | } |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | private: |
| 242 | size_t batch_size_{1}; |
| 243 | size_t channels_{1}; |
| 244 | size_t x_stride_{0}; |
| 245 | size_t y_stride_{0}; |
Marat Dukhan | af1671a | 2022-02-04 00:32:09 -0800 | [diff] [blame] | 246 | WeightsType weights_type_{WeightsType::Default}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 247 | size_t iterations_{15}; |
| 248 | }; |