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> |
Frank Barchard | e0601b5 | 2019-10-25 17:43:34 -0700 | [diff] [blame] | 23 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | |
| 25 | |
| 26 | class ClampMicrokernelTester { |
| 27 | public: |
| 28 | enum class Variant { |
| 29 | Native, |
| 30 | Scalar, |
| 31 | }; |
| 32 | |
| 33 | inline ClampMicrokernelTester& n(size_t n) { |
| 34 | assert(n != 0); |
| 35 | this->n_ = n; |
| 36 | return *this; |
| 37 | } |
| 38 | |
| 39 | inline size_t n() const { |
| 40 | return this->n_; |
| 41 | } |
| 42 | |
| 43 | inline ClampMicrokernelTester& inplace(bool inplace) { |
| 44 | this->inplace_ = inplace; |
| 45 | return *this; |
| 46 | } |
| 47 | |
| 48 | inline bool inplace() const { |
| 49 | return this->inplace_; |
| 50 | } |
| 51 | |
| 52 | inline ClampMicrokernelTester& qmin(uint8_t qmin) { |
| 53 | this->qmin_ = qmin; |
| 54 | return *this; |
| 55 | } |
| 56 | |
| 57 | inline uint8_t qmin() const { |
| 58 | return this->qmin_; |
| 59 | } |
| 60 | |
| 61 | inline ClampMicrokernelTester& qmax(uint8_t qmax) { |
| 62 | this->qmax_ = qmax; |
| 63 | return *this; |
| 64 | } |
| 65 | |
| 66 | inline uint8_t qmax() const { |
| 67 | return this->qmax_; |
| 68 | } |
| 69 | |
| 70 | inline ClampMicrokernelTester& iterations(size_t iterations) { |
| 71 | this->iterations_ = iterations; |
| 72 | return *this; |
| 73 | } |
| 74 | |
| 75 | inline size_t iterations() const { |
| 76 | return this->iterations_; |
| 77 | } |
| 78 | |
| 79 | void Test(xnn_u8_clamp_ukernel_function clamp, Variant variant = Variant::Native) const { |
| 80 | std::random_device random_device; |
| 81 | auto rng = std::mt19937(random_device()); |
| 82 | auto u8rng = std::bind(std::uniform_int_distribution<uint8_t>(), rng); |
| 83 | |
| 84 | std::vector<uint8_t> x(n() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 85 | std::vector<uint8_t> y(n() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint8_t) : 0)); |
| 86 | std::vector<uint8_t> y_ref(n()); |
| 87 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 88 | std::generate(x.begin(), x.end(), std::ref(u8rng)); |
| 89 | if (inplace()) { |
| 90 | std::generate(y.begin(), y.end(), std::ref(u8rng)); |
| 91 | } else { |
| 92 | std::fill(y.begin(), y.end(), 0xA5); |
| 93 | } |
| 94 | const uint8_t* x_data = inplace() ? y.data() : x.data(); |
| 95 | |
| 96 | // Prepare clamping parameters. |
| 97 | union xnn_u8_output_params output_params = { }; |
| 98 | switch (variant) { |
| 99 | case Variant::Native: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 100 | output_params = xnn_init_u8_output_params(qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 101 | break; |
| 102 | case Variant::Scalar: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 103 | output_params = xnn_init_scalar_u8_output_params(qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 104 | break; |
| 105 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 106 | |
| 107 | // Compute reference results. |
| 108 | for (size_t i = 0; i < n(); i++) { |
| 109 | y_ref[i] = std::max(std::min(x_data[i], qmax()), qmin()); |
| 110 | } |
| 111 | |
| 112 | // Call optimized micro-kernel. |
| 113 | clamp(n() * sizeof(uint8_t), x_data, y.data(), &output_params); |
| 114 | |
| 115 | // Verify results. |
| 116 | for (size_t i = 0; i < n(); i++) { |
| 117 | ASSERT_LE(uint32_t(y[i]), uint32_t(qmax())) |
| 118 | << "at position " << i << ", n = " << n(); |
| 119 | ASSERT_GE(uint32_t(y[i]), uint32_t(qmin())) |
| 120 | << "at position " << i << ", n = " << n(); |
| 121 | ASSERT_EQ(uint32_t(y_ref[i]), uint32_t(y[i])) |
| 122 | << "at position " << i << ", n = " << n() |
| 123 | << ", qmin = " << uint32_t(qmin()) << ", qmax = " << uint32_t(qmax()); |
| 124 | } |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | void Test(xnn_f32_clamp_ukernel_function clamp, Variant variant = Variant::Native) const { |
| 129 | std::random_device random_device; |
| 130 | auto rng = std::mt19937(random_device()); |
| 131 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 255.0f), rng); |
| 132 | |
| 133 | std::vector<float> x(n() + XNN_EXTRA_BYTES / sizeof(float)); |
| 134 | std::vector<float> y(n() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 135 | std::vector<float> y_ref(n()); |
| 136 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 137 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 138 | if (inplace()) { |
| 139 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 140 | } else { |
| 141 | std::fill(y.begin(), y.end(), std::nanf("")); |
| 142 | } |
| 143 | const float* x_data = inplace() ? y.data() : x.data(); |
| 144 | |
| 145 | // Prepare output parameters. |
| 146 | xnn_f32_output_params output_params = { }; |
| 147 | switch (variant) { |
| 148 | case Variant::Native: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 149 | output_params = xnn_init_f32_output_params(float(qmin()), float(qmax())); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 150 | break; |
| 151 | case Variant::Scalar: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 152 | output_params = xnn_init_scalar_f32_output_params(float(qmin()), float(qmax())); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 153 | break; |
| 154 | } |
| 155 | |
| 156 | // Compute reference results. |
| 157 | for (size_t i = 0; i < n(); i++) { |
| 158 | y_ref[i] = std::max(std::min(x_data[i], float(qmax())), float(qmin())); |
| 159 | } |
| 160 | |
| 161 | // Call optimized micro-kernel. |
| 162 | clamp(n() * sizeof(float), x_data, y.data(), &output_params); |
| 163 | |
| 164 | // Verify results. |
| 165 | for (size_t i = 0; i < n(); i++) { |
| 166 | ASSERT_LE(y[i], float(qmax())) |
| 167 | << "at position " << i << ", n = " << n(); |
| 168 | ASSERT_GE(y[i], float(qmin())) |
| 169 | << "at position " << i << ", n = " << n(); |
| 170 | ASSERT_EQ(y_ref[i], y[i]) |
| 171 | << "at position " << i << ", n = " << n() |
| 172 | << ", qmin = " << uint32_t(qmin()) << ", qmax = " << uint32_t(qmax()); |
| 173 | } |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | private: |
| 178 | size_t n_{1}; |
| 179 | bool inplace_{false}; |
| 180 | uint8_t qmin_{5}; |
| 181 | uint8_t qmax_{250}; |
| 182 | size_t iterations_{15}; |
| 183 | }; |