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 HSwishMicrokernelTester { 24 public: 25 enum class Variant { 26 Native, 27 Scalar, 28 }; 29 batch_size(size_t batch_size)30 inline HSwishMicrokernelTester& batch_size(size_t batch_size) { 31 assert(batch_size != 0); 32 this->batch_size_ = batch_size; 33 return *this; 34 } 35 batch_size()36 inline size_t batch_size() const { 37 return this->batch_size_; 38 } 39 inplace(bool inplace)40 inline HSwishMicrokernelTester& inplace(bool inplace) { 41 this->inplace_ = inplace; 42 return *this; 43 } 44 inplace()45 inline bool inplace() const { 46 return this->inplace_; 47 } 48 iterations(size_t iterations)49 inline HSwishMicrokernelTester& iterations(size_t iterations) { 50 this->iterations_ = iterations; 51 return *this; 52 } 53 iterations()54 inline size_t iterations() const { 55 return this->iterations_; 56 } 57 58 void Test(xnn_f32_hswish_ukernel_function hswish, Variant variant = Variant::Native) const { 59 std::random_device random_device; 60 auto rng = std::mt19937(random_device()); 61 auto f32rng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), rng); 62 63 std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); 64 std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); 65 std::vector<float> y_ref(batch_size()); 66 for (size_t iteration = 0; iteration < iterations(); iteration++) { 67 std::generate(x.begin(), x.end(), std::ref(f32rng)); 68 if (inplace()) { 69 std::generate(y.begin(), y.end(), std::ref(f32rng)); 70 } else { 71 std::fill(y.begin(), y.end(), std::nanf("")); 72 } 73 const float* x_data = inplace() ? y.data() : x.data(); 74 75 // Prepare micro-kernel parameters. 76 union xnn_f32_hswish_params params = { }; 77 switch (variant) { 78 case Variant::Native: 79 params = xnn_init_f32_hswish_params(); 80 break; 81 case Variant::Scalar: 82 params = xnn_init_scalar_f32_hswish_params(); 83 break; 84 } 85 86 // Compute reference results. 87 for (size_t i = 0; i < batch_size(); i++) { 88 y_ref[i] = x_data[i] * std::max(std::min(x_data[i] + 3.0f, 6.0f), 0.0f) / 6.0f; 89 } 90 91 // Call optimized micro-kernel. 92 hswish(batch_size() * sizeof(float), x_data, y.data(), ¶ms); 93 94 // Verify results. 95 for (size_t i = 0; i < batch_size(); i++) { 96 ASSERT_NEAR(y_ref[i], y[i], std::abs(y_ref[i]) * 1.0e-6f) 97 << "at position " << i << ", batch_size = " << batch_size(); 98 } 99 } 100 } 101 102 private: 103 size_t batch_size_{1}; 104 bool inplace_{false}; 105 size_t iterations_{5}; 106 }; 107