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 <random> 15 #include <vector> 16 17 #include <fp16.h> 18 19 #include <xnnpack.h> 20 #include <xnnpack/microfnptr.h> 21 #include <xnnpack/microparams-init.h> 22 23 24 class RAddStoreExpMinusMaxMicrokernelTester { 25 public: elements(size_t elements)26 inline RAddStoreExpMinusMaxMicrokernelTester& elements(size_t elements) { 27 assert(elements != 0); 28 this->elements_ = elements; 29 return *this; 30 } 31 elements()32 inline size_t elements() const { 33 return this->elements_; 34 } 35 iterations(size_t iterations)36 inline RAddStoreExpMinusMaxMicrokernelTester& iterations(size_t iterations) { 37 this->iterations_ = iterations; 38 return *this; 39 } 40 iterations()41 inline size_t iterations() const { 42 return this->iterations_; 43 } 44 Test(xnn_f16_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax,xnn_init_f16_expminus_params_fn init_params)45 void Test(xnn_f16_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax, xnn_init_f16_expminus_params_fn init_params) const { 46 std::random_device random_device; 47 auto rng = std::mt19937(random_device()); 48 // Choose such range that exph(x[i]) overflows, but exph(x[i] - x_max) doesn't. 49 // However, the range is still narrow enough that double-precision exp doesn't overflow. 50 std::uniform_real_distribution<float> f32dist(15.0f, 20.0f); 51 52 std::vector<uint16_t> x(elements() + XNN_EXTRA_BYTES / sizeof(uint16_t)); 53 std::vector<uint16_t> y(elements()); 54 std::vector<float> y_ref(elements()); 55 for (size_t iteration = 0; iteration < iterations(); iteration++) { 56 std::generate(x.begin(), x.end(), [&]() { return fp16_ieee_from_fp32_value(f32dist(rng)); }); 57 std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */); 58 59 // Compute reference results. 60 float sum_ref = 0.0f; 61 float x_max_as_float = -std::numeric_limits<float>::infinity(); 62 for (size_t i = 0; i < elements(); i++) { 63 x_max_as_float = std::max(x_max_as_float, fp16_ieee_to_fp32_value(x[i])); 64 } 65 const uint16_t x_max_as_half = fp16_ieee_from_fp32_value(x_max_as_float); 66 for (size_t i = 0; i < elements(); i++) { 67 const float y_ref_value = exp(fp16_ieee_to_fp32_value(x[i]) - x_max_as_float); 68 y_ref[i] = y_ref_value; 69 sum_ref += y_ref_value; 70 } 71 72 // Call optimized micro-kernel. 73 uint16_t sum = UINT16_C(0x7E00) /* NaN */; 74 xnn_f16_expminus_params params; 75 init_params(¶ms); 76 raddstoreexpminusmax(elements() * sizeof(uint16_t), x.data(), &x_max_as_half, y.data(), &sum, ¶ms); 77 78 // Verify results. 79 for (size_t i = 0; i < elements(); i++) { 80 ASSERT_NEAR(y_ref[i], fp16_ieee_to_fp32_value(y[i]), std::abs(y_ref[i]) * 5.0e-3f) 81 << "element " << i << " / " << elements() << ", x_max " << x_max_as_float; 82 } 83 ASSERT_NEAR(sum_ref, fp16_ieee_to_fp32_value(sum), std::abs(sum_ref) * 5.0e-3f) 84 << "batch " << elements() << ", x_max " << x_max_as_float; 85 } 86 } 87 Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax,xnn_init_f32_expminus_params_fn init_params)88 void Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax, xnn_init_f32_expminus_params_fn init_params) const { 89 std::random_device random_device; 90 auto rng = std::mt19937(random_device()); 91 // Choose such range that expf(x[i]) overflows, but expf(x[i] - x_max) doesn't. 92 // However, the range is still narrow enough that double-precision exp doesn't overflow. 93 std::uniform_real_distribution<float> f32dist(90.0f, 100.0f); 94 95 std::vector<float> x(elements() + XNN_EXTRA_BYTES / sizeof(float)); 96 std::vector<float> y(elements()); 97 std::vector<double> y_ref(elements()); 98 for (size_t iteration = 0; iteration < iterations(); iteration++) { 99 std::generate(x.begin(), x.end(), [&]() { return f32dist(rng); }); 100 std::fill(y.begin(), y.end(), std::nanf("")); 101 102 // Compute reference results. 103 double sum_ref = 0.0f; 104 const float x_max = *std::max_element(x.begin(), x.begin() + elements()); 105 for (size_t i = 0; i < elements(); i++) { 106 const double y_ref_value = exp(double(x[i]) - double(x_max)); 107 y_ref[i] = y_ref_value; 108 sum_ref += y_ref_value; 109 } 110 111 // Call optimized micro-kernel. 112 float sum = std::nanf(""); 113 xnn_f32_expminus_params params; 114 init_params(¶ms); 115 raddstoreexpminusmax(elements() * sizeof(float), x.data(), &x_max, y.data(), &sum, ¶ms); 116 117 // Verify results. 118 for (size_t i = 0; i < elements(); i++) { 119 ASSERT_NEAR(y_ref[i], double(y[i]), std::abs(y_ref[i]) * 1.0e-6) 120 << "element " << i << " / " << elements() << ", x_max " << x_max; 121 } 122 ASSERT_NEAR(sum_ref, double(sum), std::abs(sum_ref) * 1.0e-6) 123 << "batch " << elements() << ", x_max " << x_max; 124 } 125 } 126 127 private: 128 size_t elements_{1}; 129 size_t iterations_{15}; 130 }; 131