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 <cstddef> |
| 16 | #include <cstdlib> |
| 17 | #include <functional> |
| 18 | #include <random> |
| 19 | #include <vector> |
| 20 | |
| 21 | #include <xnnpack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 22 | #include <xnnpack/params-init.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 23 | #include <xnnpack/params.h> |
| 24 | #include <xnnpack/requantization.h> |
| 25 | |
| 26 | |
| 27 | class VAddMicrokernelTester { |
| 28 | public: |
| 29 | enum class Variant { |
| 30 | Native, |
| 31 | Scalar, |
| 32 | }; |
| 33 | |
| 34 | inline VAddMicrokernelTester& n(size_t n) { |
| 35 | assert(n != 0); |
| 36 | this->n_ = n; |
| 37 | return *this; |
| 38 | } |
| 39 | |
| 40 | inline size_t n() const { |
| 41 | return this->n_; |
| 42 | } |
| 43 | |
| 44 | inline VAddMicrokernelTester& inplace_a(bool inplace_a) { |
| 45 | this->inplace_a_ = inplace_a; |
| 46 | return *this; |
| 47 | } |
| 48 | |
| 49 | inline bool inplace_a() const { |
| 50 | return this->inplace_a_; |
| 51 | } |
| 52 | |
| 53 | inline VAddMicrokernelTester& inplace_b(bool inplace_b) { |
| 54 | this->inplace_b_ = inplace_b; |
| 55 | return *this; |
| 56 | } |
| 57 | |
| 58 | inline bool inplace_b() const { |
| 59 | return this->inplace_b_; |
| 60 | } |
| 61 | |
| 62 | inline VAddMicrokernelTester& a_scale(float a_scale) { |
| 63 | assert(a_scale > 0.0f); |
| 64 | assert(std::isnormal(a_scale)); |
| 65 | this->a_scale_ = a_scale; |
| 66 | return *this; |
| 67 | } |
| 68 | |
| 69 | inline float a_scale() const { |
| 70 | return this->a_scale_; |
| 71 | } |
| 72 | |
| 73 | inline VAddMicrokernelTester& a_zero_point(uint8_t a_zero_point) { |
| 74 | this->a_zero_point_ = a_zero_point; |
| 75 | return *this; |
| 76 | } |
| 77 | |
| 78 | inline uint8_t a_zero_point() const { |
| 79 | return this->a_zero_point_; |
| 80 | } |
| 81 | |
| 82 | inline VAddMicrokernelTester& b_scale(float b_scale) { |
| 83 | assert(b_scale > 0.0f); |
| 84 | assert(std::isnormal(b_scale)); |
| 85 | this->b_scale_ = b_scale; |
| 86 | return *this; |
| 87 | } |
| 88 | |
| 89 | inline float b_scale() const { |
| 90 | return this->b_scale_; |
| 91 | } |
| 92 | |
| 93 | inline VAddMicrokernelTester& b_zero_point(uint8_t b_zero_point) { |
| 94 | this->b_zero_point_ = b_zero_point; |
| 95 | return *this; |
| 96 | } |
| 97 | |
| 98 | inline uint8_t b_zero_point() const { |
| 99 | return this->b_zero_point_; |
| 100 | } |
| 101 | |
| 102 | inline VAddMicrokernelTester& y_scale(float y_scale) { |
| 103 | assert(y_scale > 0.0f); |
| 104 | assert(std::isnormal(y_scale)); |
| 105 | this->y_scale_ = y_scale; |
| 106 | return *this; |
| 107 | } |
| 108 | |
| 109 | inline float y_scale() const { |
| 110 | return this->y_scale_; |
| 111 | } |
| 112 | |
| 113 | inline VAddMicrokernelTester& y_zero_point(uint8_t y_zero_point) { |
| 114 | this->y_zero_point_ = y_zero_point; |
| 115 | return *this; |
| 116 | } |
| 117 | |
| 118 | inline uint8_t y_zero_point() const { |
| 119 | return this->y_zero_point_; |
| 120 | } |
| 121 | |
| 122 | inline VAddMicrokernelTester& qmin(uint8_t qmin) { |
| 123 | this->qmin_ = qmin; |
| 124 | return *this; |
| 125 | } |
| 126 | |
| 127 | inline uint8_t qmin() const { |
| 128 | return this->qmin_; |
| 129 | } |
| 130 | |
| 131 | inline VAddMicrokernelTester& qmax(uint8_t qmax) { |
| 132 | this->qmax_ = qmax; |
| 133 | return *this; |
| 134 | } |
| 135 | |
| 136 | inline uint8_t qmax() const { |
| 137 | return this->qmax_; |
| 138 | } |
| 139 | |
| 140 | inline VAddMicrokernelTester& iterations(size_t iterations) { |
| 141 | this->iterations_ = iterations; |
| 142 | return *this; |
| 143 | } |
| 144 | |
| 145 | inline size_t iterations() const { |
| 146 | return this->iterations_; |
| 147 | } |
| 148 | |
| 149 | void Test(xnn_q8_vadd_ukernel_function vadd, Variant variant = Variant::Native) const { |
| 150 | std::random_device random_device; |
| 151 | auto rng = std::mt19937(random_device()); |
| 152 | auto u8rng = std::bind(std::uniform_int_distribution<uint8_t>(), rng); |
| 153 | |
| 154 | std::vector<uint8_t> a(n() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 155 | std::vector<uint8_t> b(n() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 156 | std::vector<uint8_t> y(n() + (inplace_a() || inplace_b() ? XNN_EXTRA_BYTES / sizeof(uint8_t) : 0)); |
| 157 | std::vector<float> y_fp(n()); |
| 158 | std::vector<uint8_t> y_ref(n()); |
| 159 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 160 | std::generate(a.begin(), a.end(), std::ref(u8rng)); |
| 161 | std::generate(b.begin(), b.end(), std::ref(u8rng)); |
| 162 | if (inplace_a() || inplace_b()) { |
| 163 | std::generate(y.begin(), y.end(), std::ref(u8rng)); |
| 164 | } else { |
| 165 | std::fill(y.begin(), y.end(), 0xA5); |
| 166 | } |
| 167 | const uint8_t* a_data = inplace_a() ? y.data() : a.data(); |
| 168 | const uint8_t* b_data = inplace_b() ? y.data() : b.data(); |
| 169 | |
| 170 | // Prepare quantization parameters. |
| 171 | xnn_q8_add_params quantization_params = { }; |
| 172 | switch (variant) { |
| 173 | case Variant::Native: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 174 | quantization_params = xnn_init_q8_add_params( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 175 | a_zero_point(), b_zero_point(), y_zero_point(), |
| 176 | a_scale() / y_scale(), b_scale() / y_scale(), |
| 177 | qmin(), qmax()); |
| 178 | break; |
| 179 | case Variant::Scalar: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 180 | quantization_params = xnn_init_scalar_q8_add_params( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 181 | a_zero_point(), b_zero_point(), y_zero_point(), |
| 182 | a_scale() / y_scale(), b_scale() / y_scale(), |
| 183 | qmin(), qmax()); |
| 184 | break; |
| 185 | } |
| 186 | const xnn_q8_add_params scalar_quantization_params = |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 187 | xnn_init_scalar_q8_add_params( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 188 | a_zero_point(), b_zero_point(), y_zero_point(), |
| 189 | a_scale() / y_scale(), b_scale() / y_scale(), |
| 190 | qmin(), qmax()); |
| 191 | |
| 192 | // Compute reference results. |
| 193 | for (size_t i = 0; i < n(); i++) { |
| 194 | y_fp[i] = float(y_zero_point()) + |
| 195 | float(int32_t(a_data[i]) - int32_t(a_zero_point())) * (a_scale() / y_scale()) + |
| 196 | float(int32_t(b_data[i]) - int32_t(b_zero_point())) * (b_scale() / y_scale()); |
| 197 | y_fp[i] = std::min<float>(y_fp[i], float(qmax())); |
| 198 | y_fp[i] = std::max<float>(y_fp[i], float(qmin())); |
| 199 | y_ref[i] = xnn_add_quantize(a_data[i], b_data[i], scalar_quantization_params); |
| 200 | } |
| 201 | |
| 202 | // Call optimized micro-kernel. |
| 203 | vadd(n(), a_data, b_data, y.data(), &quantization_params); |
| 204 | |
| 205 | // Verify results. |
| 206 | for (size_t i = 0; i < n(); i++) { |
| 207 | ASSERT_LE(uint32_t(y[i]), uint32_t(qmax())) |
| 208 | << "at " << i << ", n = " << n(); |
| 209 | ASSERT_GE(uint32_t(y[i]), uint32_t(qmin())) |
| 210 | << "at " << i << ", n = " << n(); |
| 211 | ASSERT_NEAR(float(int32_t(y[i])), y_fp[i], 0.6f) |
| 212 | << "at " << i << ", n = " << n(); |
| 213 | ASSERT_EQ(uint32_t(y_ref[i]), uint32_t(y[i])) |
| 214 | << "at " << i << ", n = " << n(); |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 219 | private: |
| 220 | size_t n_{1}; |
| 221 | bool inplace_a_{false}; |
| 222 | bool inplace_b_{false}; |
| 223 | float a_scale_{0.75f}; |
| 224 | float b_scale_{1.25f}; |
| 225 | float y_scale_{0.96875f}; |
| 226 | uint8_t a_zero_point_{121}; |
| 227 | uint8_t b_zero_point_{127}; |
| 228 | uint8_t y_zero_point_{133}; |
| 229 | uint8_t qmin_{0}; |
| 230 | uint8_t qmax_{255}; |
| 231 | size_t iterations_{15}; |
| 232 | }; |