| // Copyright 2020 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. |
| |
| #include <algorithm> |
| #include <cmath> |
| #include <functional> |
| #include <random> |
| #include <vector> |
| |
| #include <benchmark/benchmark.h> |
| #include "bench/utils.h" |
| |
| #include <xnnpack/AlignedAllocator.h> |
| #include <xnnpack/common.h> |
| #include <xnnpack/params.h> |
| #include <xnnpack/vunary.h> |
| |
| |
| static void f32_vrelu( |
| benchmark::State& state, |
| xnn_f32_vrelu_ukernel_function f32_vrelu, |
| benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| { |
| if (isa_check && !isa_check(state)) { |
| return; |
| } |
| |
| const size_t num_elements = state.range(0); |
| |
| std::random_device random_device; |
| auto rng = std::mt19937(random_device()); |
| auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
| |
| std::vector<float, AlignedAllocator<float, 64>> x(num_elements); |
| std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| std::vector<float, AlignedAllocator<float, 64>> y(num_elements); |
| std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| |
| for (auto _ : state) { |
| f32_vrelu(num_elements * sizeof(float), x.data(), y.data(), NULL); |
| } |
| |
| const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| if (cpu_frequency != 0) { |
| state.counters["cpufreq"] = cpu_frequency; |
| } |
| |
| const size_t elements_per_iteration = num_elements; |
| state.counters["elements"] = |
| benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| |
| const size_t bytes_per_iteration = 2 * num_elements * sizeof(float); |
| state.counters["bytes"] = |
| benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| } |
| |
| #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| BENCHMARK_CAPTURE(f32_vrelu, sse_x4, |
| xnn_f32_vrelu_ukernel__sse_x4) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, sse_x8, |
| xnn_f32_vrelu_ukernel__sse_x8) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, avx_x8, |
| xnn_f32_vrelu_ukernel__avx_x8, |
| benchmark::utils::CheckAVX) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, avx_x16, |
| xnn_f32_vrelu_ukernel__avx_x16, |
| benchmark::utils::CheckAVX) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, avx512f_x16, |
| xnn_f32_vrelu_ukernel__avx512f_x16, |
| benchmark::utils::CheckAVX512F) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, avx512f_x32, |
| xnn_f32_vrelu_ukernel__avx512f_x32, |
| benchmark::utils::CheckAVX512F) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| |
| #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| BENCHMARK_CAPTURE(f32_vrelu, neon_x4, |
| xnn_f32_vrelu_ukernel__neon_x4, |
| benchmark::utils::CheckNEON) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vrelu, neon_x8, |
| xnn_f32_vrelu_ukernel__neon_x8, |
| benchmark::utils::CheckNEON) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| |
| |
| #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| BENCHMARK_CAPTURE(f32_vrelu, wasm_x1, |
| xnn_f32_vrelu_ukernel__wasm_x1) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasm_x2, |
| xnn_f32_vrelu_ukernel__wasm_x2) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasm_x4, |
| xnn_f32_vrelu_ukernel__wasm_x4) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasm_x8, |
| xnn_f32_vrelu_ukernel__wasm_x8) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasm32_shr_x1, |
| xnn_f32_vrelu_ukernel__wasm32_shr_x1) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasm32_shr_x2, |
| xnn_f32_vrelu_ukernel__wasm32_shr_x2) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasm32_shr_x4, |
| xnn_f32_vrelu_ukernel__wasm32_shr_x4) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| |
| #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| BENCHMARK_CAPTURE(f32_vrelu, wasmsimd_x4, |
| xnn_f32_vrelu_ukernel__wasmsimd_x4) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasmsimd_x8, |
| xnn_f32_vrelu_ukernel__wasmsimd_x8) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, wasmsimd_x16, |
| xnn_f32_vrelu_ukernel__wasmsimd_x16) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| |
| BENCHMARK_CAPTURE(f32_vrelu, scalar_x1, |
| xnn_f32_vrelu_ukernel__scalar_x1) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, scalar_x2, |
| xnn_f32_vrelu_ukernel__scalar_x2) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, scalar_x4, |
| xnn_f32_vrelu_ukernel__scalar_x4) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vrelu, scalar_x8, |
| xnn_f32_vrelu_ukernel__scalar_x8) |
| ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>) |
| ->UseRealTime(); |
| |
| #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| BENCHMARK_MAIN(); |
| #endif |