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.h> 20 21 22 class RAddStoreExpMinusMaxMicrokernelTester { 23 public: elements(size_t elements)24 inline RAddStoreExpMinusMaxMicrokernelTester& elements(size_t elements) { 25 assert(elements != 0); 26 this->elements_ = elements; 27 return *this; 28 } 29 elements()30 inline size_t elements() const { 31 return this->elements_; 32 } 33 iterations(size_t iterations)34 inline RAddStoreExpMinusMaxMicrokernelTester& iterations(size_t iterations) { 35 this->iterations_ = iterations; 36 return *this; 37 } 38 iterations()39 inline size_t iterations() const { 40 return this->iterations_; 41 } 42 Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax)43 void Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax) const { 44 std::random_device random_device; 45 auto rng = std::mt19937(random_device()); 46 // Choose such range that expf(x[i]) overflows, but expf(x[i] - x_max) doesn't. 47 // However, the range is still narrow enough that double-precision exp doesn't overflow. 48 auto f32rng = std::bind(std::uniform_real_distribution<float>(90.0f, 100.0f), rng); 49 50 std::vector<float> x(elements() + XNN_EXTRA_BYTES / sizeof(float)); 51 std::vector<float> y(elements()); 52 std::vector<double> y_ref(elements()); 53 for (size_t iteration = 0; iteration < iterations(); iteration++) { 54 std::generate(x.begin(), x.end(), std::ref(f32rng)); 55 std::fill(y.begin(), y.end(), std::nanf("")); 56 57 // Compute reference results. 58 double sum_ref = 0.0f; 59 const float x_max = *std::max_element(x.begin(), x.begin() + elements()); 60 for (size_t i = 0; i < elements(); i++) { 61 const double y_ref_value = exp(double(x[i]) - double(x_max)); 62 y_ref[i] = y_ref_value; 63 sum_ref += y_ref_value; 64 } 65 66 // Call optimized micro-kernel. 67 float sum = std::nanf(""); 68 raddstoreexpminusmax(elements() * sizeof(float), x.data(), y.data(), &sum, x_max); 69 70 // Verify results. 71 for (size_t i = 0; i < elements(); i++) { 72 ASSERT_NEAR(y_ref[i], double(y[i]), std::abs(y_ref[i]) * 1.0e-6) 73 << "i = " << i << ", elements = " << elements() << ", x_max = " << x_max; 74 } 75 ASSERT_NEAR(sum_ref, double(sum), std::abs(sum_ref) * 1.0e-6) 76 << "elements = " << elements() << ", x_max = " << x_max; 77 } 78 } 79 80 private: 81 size_t elements_{1}; 82 size_t iterations_{15}; 83 }; 84