XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #pragma once |
| 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
| 13 | #include <algorithm> |
| 14 | #include <cassert> |
| 15 | #include <cmath> |
| 16 | #include <cstddef> |
| 17 | #include <cstdlib> |
| 18 | #include <functional> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 19 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include <random> |
| 21 | #include <vector> |
| 22 | |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 23 | #include <fp16.h> |
| 24 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | #include <xnnpack.h> |
| 26 | #include <xnnpack/AlignedAllocator.h> |
| 27 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 28 | #include <xnnpack/params-init.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | #include <xnnpack/params.h> |
| 30 | #include <xnnpack/requantization.h> |
| 31 | |
| 32 | |
| 33 | class DWConvMicrokernelTester { |
| 34 | public: |
| 35 | enum class Variant { |
| 36 | Native, |
| 37 | Scalar, |
| 38 | }; |
| 39 | |
| 40 | inline DWConvMicrokernelTester& width(uint32_t width) { |
| 41 | assert(width >= 1); |
| 42 | this->width_ = width; |
| 43 | return *this; |
| 44 | } |
| 45 | |
| 46 | inline uint32_t width() const { |
| 47 | return this->width_; |
| 48 | } |
| 49 | |
| 50 | inline DWConvMicrokernelTester& step(uint32_t step) { |
| 51 | assert(step >= 1); |
| 52 | this->step_ = step; |
| 53 | return *this; |
| 54 | } |
| 55 | |
| 56 | inline uint32_t step() const { |
| 57 | return this->step_; |
| 58 | } |
| 59 | |
| 60 | inline DWConvMicrokernelTester& channels(uint32_t channels) { |
| 61 | assert(channels >= 1); |
| 62 | this->channels_ = channels; |
| 63 | return *this; |
| 64 | } |
| 65 | |
| 66 | inline uint32_t channels() const { |
| 67 | return this->channels_; |
| 68 | } |
| 69 | |
| 70 | inline DWConvMicrokernelTester& cr(uint32_t cr) { |
| 71 | assert(cr != 0); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 72 | this->cr_ = cr; |
| 73 | return *this; |
| 74 | } |
| 75 | |
| 76 | inline uint32_t cr() const { |
| 77 | return this->cr_; |
| 78 | } |
| 79 | |
| 80 | inline DWConvMicrokernelTester& kr(uint32_t kr) { |
| 81 | assert(kr != 0); |
| 82 | this->kr_ = kr; |
| 83 | return *this; |
| 84 | } |
| 85 | |
| 86 | inline uint32_t kr() const { |
| 87 | return this->kr_; |
| 88 | } |
| 89 | |
| 90 | inline uint32_t packed_channels() const { |
| 91 | return (channels() / cr() + !!(channels() % cr())) * cr(); |
| 92 | } |
| 93 | |
| 94 | inline DWConvMicrokernelTester& output_stride(uint32_t output_stride) { |
| 95 | assert(output_stride != 0); |
| 96 | this->output_stride_ = output_stride; |
| 97 | return *this; |
| 98 | } |
| 99 | |
| 100 | inline uint32_t output_stride() const { |
| 101 | if (this->output_stride_ == 0) { |
| 102 | return channels(); |
| 103 | } else { |
| 104 | assert(this->output_stride_ >= channels()); |
| 105 | return this->output_stride_; |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | inline DWConvMicrokernelTester& input_zero_point(uint8_t input_zero_point) { |
| 110 | this->input_zero_point_ = input_zero_point; |
| 111 | return *this; |
| 112 | } |
| 113 | |
| 114 | inline uint8_t input_zero_point() const { |
| 115 | return this->input_zero_point_; |
| 116 | } |
| 117 | |
| 118 | inline DWConvMicrokernelTester& kernel_zero_point(uint8_t kernel_zero_point) { |
| 119 | this->kernel_zero_point_ = kernel_zero_point; |
| 120 | return *this; |
| 121 | } |
| 122 | |
| 123 | inline uint8_t kernel_zero_point() const { |
| 124 | return this->kernel_zero_point_; |
| 125 | } |
| 126 | |
| 127 | inline DWConvMicrokernelTester& qmin(uint8_t qmin) { |
| 128 | this->qmin_ = qmin; |
| 129 | return *this; |
| 130 | } |
| 131 | |
| 132 | inline uint8_t qmin() const { |
| 133 | return this->qmin_; |
| 134 | } |
| 135 | |
| 136 | inline DWConvMicrokernelTester& qmax(uint8_t qmax) { |
| 137 | this->qmax_ = qmax; |
| 138 | return *this; |
| 139 | } |
| 140 | |
| 141 | inline uint8_t qmax() const { |
| 142 | return this->qmax_; |
| 143 | } |
| 144 | |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 145 | inline DWConvMicrokernelTester& input_offset(size_t input_offset) { |
| 146 | this->input_offset_ = input_offset; |
| 147 | return *this; |
| 148 | } |
| 149 | |
| 150 | inline size_t input_offset() const { |
| 151 | return this->input_offset_; |
| 152 | } |
| 153 | |
| 154 | inline DWConvMicrokernelTester& zero_index(size_t zero_index) { |
| 155 | this->zero_index_ = zero_index; |
| 156 | return *this; |
| 157 | } |
| 158 | |
| 159 | inline size_t zero_index() const { |
| 160 | return this->zero_index_; |
| 161 | } |
| 162 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 163 | inline DWConvMicrokernelTester& iterations(size_t iterations) { |
| 164 | this->iterations_ = iterations; |
| 165 | return *this; |
| 166 | } |
| 167 | |
| 168 | inline size_t iterations() const { |
| 169 | return this->iterations_; |
| 170 | } |
| 171 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 172 | void Test(xnn_qu8_dwconv_minmax_unipass_ukernel_function dwconv_minmax, Variant variant = Variant::Native) const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 173 | std::random_device random_device; |
| 174 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 175 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 176 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), rng); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 177 | |
| 178 | std::vector<const uint8_t*> indirection((width() - 1) * step() + kr()); |
| 179 | std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + indirection.size() * channels()); |
| 180 | std::vector<uint8_t> kernel(channels() * kr()); |
| 181 | std::vector<int32_t> bias(channels()); |
Marat Dukhan | 9594db0 | 2019-12-05 14:32:37 -0800 | [diff] [blame] | 182 | std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> packed_weights((kr() + sizeof(int32_t) / sizeof(uint8_t)) * packed_channels()); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 183 | std::vector<uint8_t> zero(channels() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 184 | std::vector<uint8_t> output((width() - 1) * output_stride() + channels()); |
| 185 | std::vector<int32_t> accumulators(width() * channels()); |
| 186 | std::vector<uint8_t> output_ref(width() * channels()); |
| 187 | |
| 188 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 189 | do { |
| 190 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 191 | } while (input.size() > 1 && *std::max_element(input.cbegin(), input.cend()) == *std::min_element(input.cbegin(), input.cend())); |
| 192 | do { |
| 193 | std::generate(kernel.begin(), kernel.end(), std::ref(u8rng)); |
| 194 | } while (kernel.size() > 1 && *std::max_element(kernel.cbegin(), kernel.cend()) == *std::min_element(kernel.cbegin(), kernel.cend())); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 195 | std::generate(bias.begin(), bias.end(), std::ref(i32rng)); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 196 | std::fill(zero.begin(), zero.end(), input_zero_point()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 197 | std::fill(output.begin(), output.end(), 0xA5); |
| 198 | |
| 199 | std::fill(packed_weights.begin(), packed_weights.end(), 0); |
Marat Dukhan | 1827af1 | 2020-07-16 10:58:55 -0700 | [diff] [blame] | 200 | const xnn_qu8_packing_params packing_params = { input_zero_point(), kernel_zero_point() }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 201 | xnn_pack_qu8_dwconv_ghw_w( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 202 | kr(), 1, channels(), cr(), |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 203 | kernel.data(), bias.data(), packed_weights.data(), &packing_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 204 | for (size_t i = 0; i < indirection.size(); i++) { |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 205 | indirection[i] = input.data() + i * channels() - input_offset(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 206 | } |
| 207 | std::shuffle(indirection.begin(), indirection.end(), rng); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 208 | if (zero_index() != SIZE_MAX) { |
| 209 | for (size_t i = 0; i < indirection.size(); i += kr()) { |
| 210 | indirection[i + zero_index()] = zero.data(); |
| 211 | } |
| 212 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 213 | |
| 214 | // Compute reference results, without renormalization. |
| 215 | for (size_t x = 0; x < width(); x++) { |
| 216 | for (size_t c = 0; c < channels(); c++) { |
| 217 | float acc = bias[c]; |
| 218 | for (size_t k = 0; k < kr(); k++) { |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 219 | if (indirection[x * step() + k] != zero.data()) { |
| 220 | acc += |
| 221 | (int32_t(indirection[x * step() + k][c + input_offset()]) - int32_t(input_zero_point())) * |
| 222 | (int32_t(kernel[c * kr() + k]) - int32_t(kernel_zero_point())); |
| 223 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 224 | } |
| 225 | accumulators[x * channels() + c] = acc; |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | // Compute renormalization parameters. |
| 230 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 231 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 232 | const uint32_t accumulated_range = uint32_t(accumulated_max) - uint32_t(accumulated_min); |
| 233 | const double output_scale = accumulated_range >= 256 ? double(accumulated_range) / 255.0 : 1.00001; |
| 234 | const uint8_t output_zero_point = uint8_t(std::max(std::min( |
| 235 | lrint(127.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 236 | long(std::numeric_limits<uint8_t>::max())), long(std::numeric_limits<uint8_t>::min()))); |
| 237 | |
Frank Barchard | 9f3a843 | 2020-06-02 13:59:35 -0700 | [diff] [blame] | 238 | // Prepare parameters. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 239 | const float requantization_scale = 1.0f / float(output_scale); |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 240 | union xnn_qu8_gemm_params quantization_params = { }; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 241 | switch (variant) { |
| 242 | case Variant::Native: |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 243 | quantization_params = xnn_init_qu8_gemm_params( |
Marat Dukhan | 9199246 | 2020-07-30 00:06:34 -0700 | [diff] [blame] | 244 | kernel_zero_point(), requantization_scale, output_zero_point, qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 245 | break; |
| 246 | case Variant::Scalar: |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 247 | quantization_params = xnn_init_scalar_qu8_gemm_params( |
Marat Dukhan | 9199246 | 2020-07-30 00:06:34 -0700 | [diff] [blame] | 248 | kernel_zero_point(), requantization_scale, output_zero_point, qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 249 | break; |
| 250 | } |
Marat Dukhan | ec88e27 | 2020-07-30 15:02:09 -0700 | [diff] [blame] | 251 | const union xnn_qu8_requantization_params scalar_requantization_params = |
| 252 | xnn_init_scalar_qu8_requantization_params(requantization_scale, output_zero_point, qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 253 | |
| 254 | // Renormalize reference results. |
| 255 | for (size_t x = 0; x < width(); x++) { |
| 256 | for (size_t c = 0; c < channels(); c++) { |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 257 | output_ref[x * channels() + c] = xnn_qu8_requantize_q31(accumulators[x * channels() + c], scalar_requantization_params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 258 | } |
| 259 | } |
| 260 | |
| 261 | // Call optimized micro-kernel. |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 262 | dwconv_minmax( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 263 | channels(), width(), |
| 264 | indirection.data(), packed_weights.data(), output.data(), |
| 265 | step() * sizeof(void*), |
| 266 | (output_stride() - channels()) * sizeof(uint8_t), |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 267 | input_offset() * sizeof(uint8_t), zero.data(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 268 | &quantization_params); |
| 269 | |
| 270 | // Verify results. |
| 271 | for (size_t x = 0; x < width(); x++) { |
| 272 | for (size_t c = 0; c < channels(); c++) { |
| 273 | ASSERT_GE(uint32_t(output[x * output_stride() + c]), uint32_t(qmin())) |
| 274 | << "x = " << x << ", channel = " << c; |
| 275 | ASSERT_LE(uint32_t(output[x * output_stride() + c]), uint32_t(qmax())) |
| 276 | << "x = " << x << ", channel = " << c; |
| 277 | ASSERT_EQ(uint32_t(output[x * output_stride() + c]), uint32_t(output_ref[x * channels() + c])) |
| 278 | << "x = " << x << ", channel = " << c << ", accumulator = " << accumulators[x * channels() + c]; |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | } |
| 283 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 284 | void Test(xnn_qs8_dwconv_minmax_unipass_ukernel_function dwconv_minmax, Variant variant = Variant::Native) const { |
| 285 | std::random_device random_device; |
| 286 | auto rng = std::mt19937(random_device()); |
| 287 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
| 288 | auto i8rng = std::bind( |
| 289 | std::uniform_int_distribution<uint32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), rng); |
| 290 | |
| 291 | std::vector<const int8_t*> indirection((width() - 1) * step() + kr()); |
| 292 | std::vector<int8_t> input(XNN_EXTRA_BYTES / sizeof(int8_t) + indirection.size() * channels()); |
| 293 | std::vector<int8_t> kernel(channels() * kr()); |
| 294 | std::vector<int32_t> bias(channels()); |
| 295 | std::vector<int8_t, AlignedAllocator<int8_t, 64>> packed_weights((kr() + sizeof(int32_t) / sizeof(int8_t)) * packed_channels()); |
| 296 | std::vector<int8_t> zero(channels() + XNN_EXTRA_BYTES / sizeof(int8_t)); |
| 297 | std::vector<int8_t> output((width() - 1) * output_stride() + channels()); |
| 298 | std::vector<int32_t> accumulators(width() * channels()); |
| 299 | std::vector<int8_t> output_ref(width() * channels()); |
| 300 | |
| 301 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 302 | do { |
| 303 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 304 | } while (input.size() > 1 && *std::max_element(input.cbegin(), input.cend()) == *std::min_element(input.cbegin(), input.cend())); |
| 305 | do { |
| 306 | std::generate(kernel.begin(), kernel.end(), std::ref(i8rng)); |
| 307 | } while (kernel.size() > 1 && *std::max_element(kernel.cbegin(), kernel.cend()) == *std::min_element(kernel.cbegin(), kernel.cend())); |
| 308 | std::generate(bias.begin(), bias.end(), std::ref(i32rng)); |
| 309 | std::fill(zero.begin(), zero.end(), int8_t(input_zero_point() - 0x80)); |
| 310 | std::fill(output.begin(), output.end(), 0xA5); |
| 311 | |
| 312 | std::fill(packed_weights.begin(), packed_weights.end(), 0); |
| 313 | const xnn_qs8_packing_params packing_params = { int8_t(input_zero_point() - 0x80) }; |
| 314 | xnn_pack_qs8_dwconv_ghw_w( |
| 315 | kr(), 1, channels(), cr(), |
| 316 | kernel.data(), bias.data(), packed_weights.data(), &packing_params); |
| 317 | for (size_t i = 0; i < indirection.size(); i++) { |
| 318 | indirection[i] = input.data() + i * channels() - input_offset(); |
| 319 | } |
| 320 | std::shuffle(indirection.begin(), indirection.end(), rng); |
| 321 | if (zero_index() != SIZE_MAX) { |
| 322 | for (size_t i = 0; i < indirection.size(); i += kr()) { |
| 323 | indirection[i + zero_index()] = zero.data(); |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | // Compute reference results, without renormalization. |
| 328 | for (size_t x = 0; x < width(); x++) { |
| 329 | for (size_t c = 0; c < channels(); c++) { |
| 330 | float acc = bias[c]; |
| 331 | for (size_t k = 0; k < kr(); k++) { |
| 332 | if (indirection[x * step() + k] != zero.data()) { |
| 333 | acc += |
| 334 | (int32_t(indirection[x * step() + k][c + input_offset()]) - int32_t(input_zero_point() - 0x80)) * |
| 335 | int32_t(kernel[c * kr() + k]); |
| 336 | } |
| 337 | } |
| 338 | accumulators[x * channels() + c] = acc; |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | // Compute renormalization parameters. |
| 343 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 344 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 345 | const uint32_t accumulated_range = uint32_t(accumulated_max) - uint32_t(accumulated_min); |
| 346 | const double output_scale = accumulated_range >= 256 ? double(accumulated_range) / 255.0 : 1.00001; |
| 347 | const int8_t output_zero_point = int8_t(std::max(std::min( |
| 348 | lrint(-0.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 349 | long(std::numeric_limits<int8_t>::max())), long(std::numeric_limits<int8_t>::min()))); |
| 350 | |
| 351 | // Prepare parameters. |
| 352 | const float requantization_scale = 1.0f / float(output_scale); |
| 353 | union xnn_qs8_gemm_params quantization_params = { }; |
| 354 | switch (variant) { |
| 355 | case Variant::Native: |
| 356 | quantization_params = xnn_init_qs8_gemm_params( |
| 357 | requantization_scale, output_zero_point, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80)); |
| 358 | break; |
| 359 | case Variant::Scalar: |
| 360 | quantization_params = xnn_init_scalar_qs8_gemm_params( |
| 361 | requantization_scale, output_zero_point, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80)); |
| 362 | break; |
| 363 | } |
| 364 | const union xnn_qs8_requantization_params scalar_requantization_params = |
| 365 | xnn_init_scalar_qs8_requantization_params(requantization_scale, output_zero_point, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80)); |
| 366 | |
| 367 | // Renormalize reference results. |
| 368 | for (size_t x = 0; x < width(); x++) { |
| 369 | for (size_t c = 0; c < channels(); c++) { |
| 370 | output_ref[x * channels() + c] = xnn_qs8_requantize_q31(accumulators[x * channels() + c], scalar_requantization_params); |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | // Call optimized micro-kernel. |
| 375 | dwconv_minmax( |
| 376 | channels(), width(), |
| 377 | indirection.data(), packed_weights.data(), output.data(), |
| 378 | step() * sizeof(void*), |
| 379 | (output_stride() - channels()) * sizeof(int8_t), |
| 380 | input_offset() * sizeof(int8_t), zero.data(), |
| 381 | &quantization_params); |
| 382 | |
| 383 | // Verify results. |
| 384 | for (size_t x = 0; x < width(); x++) { |
| 385 | for (size_t c = 0; c < channels(); c++) { |
| 386 | ASSERT_GE(int32_t(output[x * output_stride() + c]), int32_t(qmin()) - 0x80) |
| 387 | << "x = " << x << ", channel = " << c; |
| 388 | ASSERT_LE(int32_t(output[x * output_stride() + c]), int32_t(qmax()) - 0x80) |
| 389 | << "x = " << x << ", channel = " << c; |
| 390 | ASSERT_EQ(int32_t(output[x * output_stride() + c]), int32_t(output_ref[x * channels() + c])) |
| 391 | << "x = " << x << ", channel = " << c << ", accumulator = " << accumulators[x * channels() + c]; |
| 392 | } |
| 393 | } |
| 394 | } |
| 395 | } |
| 396 | |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 397 | void Test(xnn_f16_dwconv_minmax_unipass_ukernel_function dwconv_minmax, Variant variant = Variant::Native) const { |
| 398 | std::random_device random_device; |
| 399 | auto rng = std::mt19937(random_device()); |
| 400 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 401 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 402 | |
| 403 | std::vector<const uint16_t*> indirection((width() - 1) * step() + kr()); |
| 404 | std::vector<uint16_t> input(XNN_EXTRA_BYTES / sizeof(uint16_t) + indirection.size() * channels()); |
| 405 | std::vector<uint16_t> kernel(channels() * kr()); |
| 406 | std::vector<uint16_t> bias(channels()); |
| 407 | std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> packed_weights((kr() + 1) * packed_channels()); |
| 408 | std::vector<uint16_t> zero(channels() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 409 | std::vector<uint16_t> output((width() - 1) * output_stride() + channels()); |
| 410 | std::vector<float> output_ref(width() * channels()); |
| 411 | |
| 412 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 413 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 414 | std::generate(kernel.begin(), kernel.end(), std::ref(f16rng)); |
| 415 | std::generate(bias.begin(), bias.end(), std::ref(f16rng)); |
| 416 | std::fill(zero.begin(), zero.end(), 0); |
| 417 | std::fill(output_ref.begin(), output_ref.end(), 0.0f); |
| 418 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 419 | |
| 420 | std::fill(packed_weights.begin(), packed_weights.end(), 0); |
| 421 | xnn_pack_f16_dwconv_ghw_w( |
| 422 | kr(), 1, channels(), cr(), |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 423 | kernel.data(), bias.data(), packed_weights.data(), nullptr); |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 424 | for (size_t i = 0; i < indirection.size(); i++) { |
| 425 | indirection[i] = input.data() + i * channels() - input_offset(); |
| 426 | } |
| 427 | std::shuffle(indirection.begin(), indirection.end(), rng); |
| 428 | if (zero_index() != SIZE_MAX) { |
| 429 | for (size_t i = 0; i < indirection.size(); i += kr()) { |
| 430 | indirection[i + zero_index()] = zero.data(); |
| 431 | } |
| 432 | } |
| 433 | |
| 434 | // Compute reference results, without clamping. |
| 435 | for (size_t x = 0; x < width(); x++) { |
| 436 | for (size_t c = 0; c < channels(); c++) { |
| 437 | float acc = fp16_ieee_to_fp32_value(bias[c]); |
| 438 | for (size_t k = 0; k < kr(); k++) { |
| 439 | if (indirection[x * step() + k] != zero.data()) { |
| 440 | acc += fp16_ieee_to_fp32_value(indirection[x * step() + k][c + input_offset()]) * fp16_ieee_to_fp32_value(kernel[c * kr() + k]); |
| 441 | } |
| 442 | } |
| 443 | output_ref[x * channels() + c] = acc; |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | // Compute clamping parameters. |
| 448 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 449 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 450 | const float accumulated_range = accumulated_max - accumulated_min; |
| 451 | const float output_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_min + accumulated_range / 255.0f * float(qmin()))); |
| 452 | const float output_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_max - accumulated_range / 255.0f * float(255 - qmax()))); |
| 453 | |
| 454 | // Prepare parameters. |
| 455 | xnn_f16_minmax_params params = xnn_init_f16_minmax_params( |
| 456 | fp16_ieee_from_fp32_value(output_min), |
| 457 | fp16_ieee_from_fp32_value(output_max)); |
| 458 | |
| 459 | // Clamp reference results. |
| 460 | for (float& output_val : output_ref) { |
| 461 | output_val = std::max(std::min(output_val, output_max), output_min); |
| 462 | } |
| 463 | |
| 464 | // Call optimized micro-kernel. |
| 465 | dwconv_minmax( |
| 466 | channels(), width(), |
| 467 | reinterpret_cast<const void**>(indirection.data()), packed_weights.data(), output.data(), |
| 468 | step() * sizeof(void*), |
| 469 | (output_stride() - channels()) * sizeof(uint16_t), |
| 470 | input_offset() * sizeof(uint16_t), zero.data(), |
| 471 | ¶ms); |
| 472 | |
| 473 | // Verify results. |
| 474 | for (size_t x = 0; x < width(); x++) { |
| 475 | for (size_t c = 0; c < channels(); c++) { |
| 476 | ASSERT_GE(fp16_ieee_to_fp32_value(output[x * output_stride() + c]), output_min) |
| 477 | << "x = " << x << ", channel = " << c; |
| 478 | ASSERT_LE(fp16_ieee_to_fp32_value(output[x * output_stride() + c]), output_max) |
| 479 | << "x = " << x << ", channel = " << c; |
| 480 | ASSERT_NEAR( |
| 481 | output_ref[x * channels() + c], |
| 482 | fp16_ieee_to_fp32_value(output[x * output_stride() + c]), |
| 483 | std::abs(output_ref[x * channels() + c]) * 1.0e-2) |
| 484 | << "x = " << x << ", channel = " << c; |
| 485 | } |
| 486 | } |
| 487 | } |
| 488 | } |
| 489 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 490 | void Test(xnn_f32_dwconv_unipass_ukernel_function dwconv) const { |
| 491 | std::random_device random_device; |
| 492 | auto rng = std::mt19937(random_device()); |
| 493 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 494 | |
| 495 | std::vector<const float*> indirection((width() - 1) * step() + kr()); |
| 496 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + indirection.size() * channels()); |
| 497 | std::vector<float> kernel(channels() * kr()); |
| 498 | std::vector<float> bias(channels()); |
| 499 | std::vector<float, AlignedAllocator<float, 64>> packed_weights((kr() + 1) * packed_channels()); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 500 | std::vector<float> zero(channels() + XNN_EXTRA_BYTES / sizeof(float)); |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 501 | std::vector<float> output((width() - 1) * output_stride() + channels()); |
| 502 | std::vector<float> output_ref(width() * channels()); |
| 503 | |
| 504 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 505 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 506 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 507 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 508 | std::fill(zero.begin(), zero.end(), 0.0f); |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 509 | std::fill(output_ref.begin(), output_ref.end(), nanf("")); |
| 510 | std::fill(output.begin(), output.end(), nanf("")); |
| 511 | |
| 512 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 513 | xnn_pack_f32_dwconv_ghw_w( |
| 514 | kr(), 1, channels(), cr(), |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 515 | kernel.data(), bias.data(), packed_weights.data(), nullptr); |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 516 | for (size_t i = 0; i < indirection.size(); i++) { |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 517 | indirection[i] = input.data() + i * channels() - input_offset(); |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 518 | } |
| 519 | std::shuffle(indirection.begin(), indirection.end(), rng); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 520 | if (zero_index() != SIZE_MAX) { |
| 521 | for (size_t i = 0; i < indirection.size(); i += kr()) { |
| 522 | indirection[i + zero_index()] = zero.data(); |
| 523 | } |
| 524 | } |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 525 | |
| 526 | // Compute reference results, without clamping. |
| 527 | for (size_t x = 0; x < width(); x++) { |
| 528 | for (size_t c = 0; c < channels(); c++) { |
| 529 | float acc = bias[c]; |
| 530 | for (size_t k = 0; k < kr(); k++) { |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 531 | if (indirection[x * step() + k] != zero.data()) { |
| 532 | acc += indirection[x * step() + k][c + input_offset()] * kernel[c * kr() + k]; |
| 533 | } |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 534 | } |
| 535 | output_ref[x * channels() + c] = acc; |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | // Call optimized micro-kernel. |
| 540 | dwconv( |
| 541 | channels(), width(), |
| 542 | indirection.data(), packed_weights.data(), output.data(), |
| 543 | step() * sizeof(void*), |
| 544 | (output_stride() - channels()) * sizeof(float), |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 545 | input_offset() * sizeof(float), zero.data(), |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 546 | nullptr); |
| 547 | |
| 548 | // Verify results. |
| 549 | for (size_t x = 0; x < width(); x++) { |
| 550 | for (size_t c = 0; c < channels(); c++) { |
| 551 | ASSERT_NEAR( |
| 552 | output_ref[x * channels() + c], |
| 553 | output[x * output_stride() + c], |
| 554 | std::abs(output_ref[x * channels() + c]) * 1.0e-5) |
| 555 | << "x = " << x << ", channel = " << c; |
| 556 | } |
| 557 | } |
| 558 | } |
| 559 | } |
| 560 | |
| 561 | void Test(xnn_f32_dwconv_minmax_unipass_ukernel_function dwconv_minmax, Variant variant = Variant::Native) const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 562 | std::random_device random_device; |
| 563 | auto rng = std::mt19937(random_device()); |
| 564 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 565 | |
| 566 | std::vector<const float*> indirection((width() - 1) * step() + kr()); |
| 567 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + indirection.size() * channels()); |
| 568 | std::vector<float> kernel(channels() * kr()); |
| 569 | std::vector<float> bias(channels()); |
Marat Dukhan | 9594db0 | 2019-12-05 14:32:37 -0800 | [diff] [blame] | 570 | std::vector<float, AlignedAllocator<float, 64>> packed_weights((kr() + 1) * packed_channels()); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 571 | std::vector<float> zero(channels() + XNN_EXTRA_BYTES / sizeof(float)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 572 | std::vector<float> output((width() - 1) * output_stride() + channels()); |
| 573 | std::vector<float> output_ref(width() * channels()); |
| 574 | |
| 575 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 576 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 577 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 578 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 579 | std::fill(zero.begin(), zero.end(), 0.0f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 580 | std::fill(output_ref.begin(), output_ref.end(), nanf("")); |
| 581 | std::fill(output.begin(), output.end(), nanf("")); |
| 582 | |
| 583 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 584 | xnn_pack_f32_dwconv_ghw_w( |
| 585 | kr(), 1, channels(), cr(), |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 586 | kernel.data(), bias.data(), packed_weights.data(), nullptr); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 587 | for (size_t i = 0; i < indirection.size(); i++) { |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 588 | indirection[i] = input.data() + i * channels() - input_offset(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 589 | } |
| 590 | std::shuffle(indirection.begin(), indirection.end(), rng); |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 591 | if (zero_index() != SIZE_MAX) { |
| 592 | for (size_t i = 0; i < indirection.size(); i += kr()) { |
| 593 | indirection[i + zero_index()] = zero.data(); |
| 594 | } |
| 595 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 596 | |
| 597 | // Compute reference results, without clamping. |
| 598 | for (size_t x = 0; x < width(); x++) { |
| 599 | for (size_t c = 0; c < channels(); c++) { |
| 600 | float acc = bias[c]; |
| 601 | for (size_t k = 0; k < kr(); k++) { |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 602 | if (indirection[x * step() + k] != zero.data()) { |
| 603 | acc += indirection[x * step() + k][c + input_offset()] * kernel[c * kr() + k]; |
| 604 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 605 | } |
| 606 | output_ref[x * channels() + c] = acc; |
| 607 | } |
| 608 | } |
| 609 | |
| 610 | // Compute clamping parameters. |
| 611 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 612 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 613 | const float accumulated_range = accumulated_max - accumulated_min; |
| 614 | const float output_min = accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 615 | const float output_max = accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 616 | |
Frank Barchard | 9f3a843 | 2020-06-02 13:59:35 -0700 | [diff] [blame] | 617 | // Prepare parameters. |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 618 | xnn_f32_minmax_params params = { }; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 619 | switch (variant) { |
| 620 | case Variant::Native: |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 621 | params = xnn_init_f32_minmax_params(output_min, output_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 622 | break; |
| 623 | case Variant::Scalar: |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 624 | params = xnn_init_scalar_f32_minmax_params(output_min, output_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 625 | break; |
| 626 | } |
| 627 | |
| 628 | // Clamp reference results. |
| 629 | for (float& output_val : output_ref) { |
| 630 | output_val = std::max(std::min(output_val, output_max), output_min); |
| 631 | } |
| 632 | |
| 633 | // Call optimized micro-kernel. |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 634 | dwconv_minmax( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 635 | channels(), width(), |
| 636 | indirection.data(), packed_weights.data(), output.data(), |
| 637 | step() * sizeof(void*), |
| 638 | (output_stride() - channels()) * sizeof(float), |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 639 | input_offset() * sizeof(float), zero.data(), |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 640 | ¶ms); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 641 | |
| 642 | // Verify results. |
| 643 | for (size_t x = 0; x < width(); x++) { |
| 644 | for (size_t c = 0; c < channels(); c++) { |
| 645 | ASSERT_GE(output[x * output_stride() + c], output_min) |
| 646 | << "x = " << x << ", channel = " << c; |
| 647 | ASSERT_LE(output[x * output_stride() + c], output_max) |
| 648 | << "x = " << x << ", channel = " << c; |
| 649 | ASSERT_NEAR( |
| 650 | output_ref[x * channels() + c], |
| 651 | output[x * output_stride() + c], |
| 652 | std::abs(output_ref[x * channels() + c]) * 1.0e-5) |
| 653 | << "x = " << x << ", channel = " << c; |
| 654 | } |
| 655 | } |
| 656 | } |
| 657 | } |
| 658 | |
| 659 | private: |
| 660 | uint32_t channels_{1}; |
| 661 | uint32_t cr_{1}; |
| 662 | uint32_t kr_{1}; |
| 663 | uint32_t width_{1}; |
| 664 | uint32_t step_{1}; |
| 665 | uint32_t output_stride_{0}; |
| 666 | uint8_t input_zero_point_{127}; |
| 667 | uint8_t kernel_zero_point_{127}; |
| 668 | uint8_t qmin_{0}; |
| 669 | uint8_t qmax_{255}; |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 670 | size_t input_offset_{0}; |
| 671 | size_t zero_index_{SIZE_MAX}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 672 | size_t iterations_{3}; |
| 673 | }; |