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 <cassert> 12 #include <cstddef> 13 #include <cstdlib> 14 #include <functional> 15 #include <random> 16 #include <vector> 17 18 #include <xnnpack.h> 19 #include <xnnpack/params-init.h> 20 #include <xnnpack/params.h> 21 22 23 class VBinOpMicrokernelTester { 24 public: 25 enum class OpType { 26 Add, 27 Div, 28 Max, 29 Min, 30 Mul, 31 Sub, 32 }; 33 34 enum class Variant { 35 Native, 36 Scalar, 37 }; 38 batch_size(size_t batch_size)39 inline VBinOpMicrokernelTester& batch_size(size_t batch_size) { 40 assert(batch_size != 0); 41 this->batch_size_ = batch_size; 42 return *this; 43 } 44 batch_size()45 inline size_t batch_size() const { 46 return this->batch_size_; 47 } 48 inplace_a(bool inplace_a)49 inline VBinOpMicrokernelTester& inplace_a(bool inplace_a) { 50 this->inplace_a_ = inplace_a; 51 return *this; 52 } 53 inplace_a()54 inline bool inplace_a() const { 55 return this->inplace_a_; 56 } 57 inplace_b(bool inplace_b)58 inline VBinOpMicrokernelTester& inplace_b(bool inplace_b) { 59 this->inplace_b_ = inplace_b; 60 return *this; 61 } 62 inplace_b()63 inline bool inplace_b() const { 64 return this->inplace_b_; 65 } 66 qmin(uint8_t qmin)67 inline VBinOpMicrokernelTester& qmin(uint8_t qmin) { 68 this->qmin_ = qmin; 69 return *this; 70 } 71 qmin()72 inline uint8_t qmin() const { 73 return this->qmin_; 74 } 75 qmax(uint8_t qmax)76 inline VBinOpMicrokernelTester& qmax(uint8_t qmax) { 77 this->qmax_ = qmax; 78 return *this; 79 } 80 qmax()81 inline uint8_t qmax() const { 82 return this->qmax_; 83 } 84 iterations(size_t iterations)85 inline VBinOpMicrokernelTester& iterations(size_t iterations) { 86 this->iterations_ = iterations; 87 return *this; 88 } 89 iterations()90 inline size_t iterations() const { 91 return this->iterations_; 92 } 93 94 void Test(xnn_f32_vbinary_ukernel_function vbinary, OpType op_type, Variant variant = Variant::Native) const { 95 std::random_device random_device; 96 auto rng = std::mt19937(random_device()); 97 auto f32rng = std::bind(std::uniform_real_distribution<float>(0.01f, 1.0f), rng); 98 99 std::vector<float> a(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); 100 std::vector<float> b(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); 101 std::vector<float> y(batch_size() + (inplace_a() || inplace_b() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); 102 std::vector<float> y_ref(batch_size()); 103 for (size_t iteration = 0; iteration < iterations(); iteration++) { 104 std::generate(a.begin(), a.end(), std::ref(f32rng)); 105 std::generate(b.begin(), b.end(), std::ref(f32rng)); 106 if (inplace_a() || inplace_b()) { 107 std::generate(y.begin(), y.end(), std::ref(f32rng)); 108 } else { 109 std::fill(y.begin(), y.end(), nanf("")); 110 } 111 const float* a_data = inplace_a() ? y.data() : a.data(); 112 const float* b_data = inplace_b() ? y.data() : b.data(); 113 114 // Compute reference results. 115 for (size_t i = 0; i < batch_size(); i++) { 116 switch (op_type) { 117 case OpType::Add: 118 y_ref[i] = a_data[i] + b_data[i]; 119 break; 120 case OpType::Div: 121 y_ref[i] = a_data[i] / b_data[i]; 122 break; 123 case OpType::Max: 124 y_ref[i] = std::max<float>(a_data[i], b_data[i]); 125 break; 126 case OpType::Min: 127 y_ref[i] = std::min<float>(a_data[i], b_data[i]); 128 break; 129 case OpType::Mul: 130 y_ref[i] = a_data[i] * b_data[i]; 131 break; 132 case OpType::Sub: 133 y_ref[i] = a_data[i] - b_data[i]; 134 break; 135 } 136 } 137 const float accumulated_min = *std::min_element(y_ref.cbegin(), y_ref.cend()); 138 const float accumulated_max = *std::max_element(y_ref.cbegin(), y_ref.cend()); 139 const float accumulated_range = accumulated_max - accumulated_min; 140 const float y_max = accumulated_range > 0.0f ? 141 (accumulated_max - accumulated_range / 255.0f * float(255 - qmax())) : 142 +std::numeric_limits<float>::infinity(); 143 const float y_min = accumulated_range > 0.0f ? 144 (accumulated_min + accumulated_range / 255.0f * float(qmin())) : 145 -std::numeric_limits<float>::infinity(); 146 for (size_t i = 0; i < batch_size(); i++) { 147 y_ref[i] = std::max<float>(std::min<float>(y_ref[i], y_max), y_min); 148 } 149 150 // Prepare output parameters. 151 xnn_f32_output_params output_params = { }; 152 switch (variant) { 153 case Variant::Native: 154 output_params = xnn_init_f32_output_params(y_min, y_max); 155 break; 156 case Variant::Scalar: 157 output_params = xnn_init_scalar_f32_output_params(y_min, y_max); 158 break; 159 } 160 161 // Call optimized micro-kernel. 162 vbinary(batch_size() * sizeof(float), a_data, b_data, y.data(), &output_params); 163 164 // Verify results. 165 for (size_t i = 0; i < batch_size(); i++) { 166 ASSERT_NEAR(y[i], y_ref[i], std::abs(y_ref[i]) * 1.0e-6f) 167 << "at " << i << " / " << batch_size(); 168 } 169 } 170 } 171 172 private: 173 size_t batch_size_{1}; 174 bool inplace_a_{false}; 175 bool inplace_b_{false}; 176 uint8_t qmin_{0}; 177 uint8_t qmax_{255}; 178 size_t iterations_{15}; 179 }; 180