// Copyright 2019 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #pragma once #include #include #include #include #include #include #include #include #include #include #include class RAddStoreExpMinusMaxMicrokernelTester { public: inline RAddStoreExpMinusMaxMicrokernelTester& elements(size_t elements) { assert(elements != 0); this->elements_ = elements; return *this; } inline size_t elements() const { return this->elements_; } inline RAddStoreExpMinusMaxMicrokernelTester& iterations(size_t iterations) { this->iterations_ = iterations; return *this; } inline size_t iterations() const { return this->iterations_; } void Test(xnn_f16_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax, xnn_init_f16_expminus_params_fn init_params) const { std::random_device random_device; auto rng = std::mt19937(random_device()); // Choose such range that exph(x[i]) overflows, but exph(x[i] - x_max) doesn't. // However, the range is still narrow enough that double-precision exp doesn't overflow. std::uniform_real_distribution f32dist(15.0f, 20.0f); std::vector x(elements() + XNN_EXTRA_BYTES / sizeof(uint16_t)); std::vector y(elements()); std::vector y_ref(elements()); for (size_t iteration = 0; iteration < iterations(); iteration++) { std::generate(x.begin(), x.end(), [&]() { return fp16_ieee_from_fp32_value(f32dist(rng)); }); std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */); // Compute reference results. float sum_ref = 0.0f; float x_max_as_float = -std::numeric_limits::infinity(); for (size_t i = 0; i < elements(); i++) { x_max_as_float = std::max(x_max_as_float, fp16_ieee_to_fp32_value(x[i])); } const uint16_t x_max_as_half = fp16_ieee_from_fp32_value(x_max_as_float); for (size_t i = 0; i < elements(); i++) { const float y_ref_value = exp(fp16_ieee_to_fp32_value(x[i]) - x_max_as_float); y_ref[i] = y_ref_value; sum_ref += y_ref_value; } // Call optimized micro-kernel. uint16_t sum = UINT16_C(0x7E00) /* NaN */; xnn_f16_expminus_params params; init_params(¶ms); raddstoreexpminusmax(elements() * sizeof(uint16_t), x.data(), &x_max_as_half, y.data(), &sum, ¶ms); // Verify results. for (size_t i = 0; i < elements(); i++) { ASSERT_NEAR(y_ref[i], fp16_ieee_to_fp32_value(y[i]), std::abs(y_ref[i]) * 5.0e-3f) << "element " << i << " / " << elements() << ", x_max " << x_max_as_float; } ASSERT_NEAR(sum_ref, fp16_ieee_to_fp32_value(sum), std::abs(sum_ref) * 5.0e-3f) << "batch " << elements() << ", x_max " << x_max_as_float; } } void Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax, xnn_init_f32_expminus_params_fn init_params) const { std::random_device random_device; auto rng = std::mt19937(random_device()); // Choose such range that expf(x[i]) overflows, but expf(x[i] - x_max) doesn't. // However, the range is still narrow enough that double-precision exp doesn't overflow. std::uniform_real_distribution f32dist(90.0f, 100.0f); std::vector x(elements() + XNN_EXTRA_BYTES / sizeof(float)); std::vector y(elements()); std::vector y_ref(elements()); for (size_t iteration = 0; iteration < iterations(); iteration++) { std::generate(x.begin(), x.end(), [&]() { return f32dist(rng); }); std::fill(y.begin(), y.end(), std::nanf("")); // Compute reference results. double sum_ref = 0.0f; const float x_max = *std::max_element(x.begin(), x.begin() + elements()); for (size_t i = 0; i < elements(); i++) { const double y_ref_value = exp(double(x[i]) - double(x_max)); y_ref[i] = y_ref_value; sum_ref += y_ref_value; } // Call optimized micro-kernel. float sum = std::nanf(""); xnn_f32_expminus_params params; init_params(¶ms); raddstoreexpminusmax(elements() * sizeof(float), x.data(), &x_max, y.data(), &sum, ¶ms); // Verify results. for (size_t i = 0; i < elements(); i++) { ASSERT_NEAR(y_ref[i], double(y[i]), std::abs(y_ref[i]) * 1.0e-6) << "element " << i << " / " << elements() << ", x_max " << x_max; } ASSERT_NEAR(sum_ref, double(sum), std::abs(sum_ref) * 1.0e-6) << "batch " << elements() << ", x_max " << x_max; } } private: size_t elements_{1}; size_t iterations_{15}; };