| // 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/vunary.h> |
| #include <xnnpack/params.h> |
| #include <xnnpack/params-init.h> |
| |
| |
| static void f32_vsqrt( |
| benchmark::State& state, |
| xnn_f32_vsqrt_ukernel_function vsqrt, |
| benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| { |
| if (isa_check && !isa_check(state)) { |
| return; |
| } |
| |
| const size_t elements = state.range(0); |
| std::vector<float, AlignedAllocator<float, 64>> input(elements); |
| std::vector<float, AlignedAllocator<float, 64>> output(elements); |
| |
| std::random_device random_device; |
| auto rng = std::mt19937(random_device()); |
| auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 10.0f), std::ref(rng)); |
| std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| std::fill(output.begin(), output.end(), std::nanf("")); |
| |
| const union xnn_f32_sqrt_params params = xnn_init_f32_sqrt_params(); |
| for (auto _ : state) { |
| vsqrt(elements * sizeof(float), input.data(), output.data(), ¶ms); |
| } |
| |
| const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
| if (cpu_frequency != 0) { |
| state.counters["cpufreq"] = cpu_frequency; |
| } |
| |
| const size_t elements_per_iteration = elements; |
| state.counters["elements"] = |
| benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate); |
| |
| const size_t bytes_per_iteration = 2 * elements * sizeof(float); |
| state.counters["bytes"] = |
| benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate); |
| } |
| |
| #if XNN_ARCH_ARM64 |
| BENCHMARK_CAPTURE(f32_vsqrt, neon_sqrt_x4, xnn_f32_vsqrt_ukernel__neon_sqrt_x4) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neon_sqrt_x8, xnn_f32_vsqrt_ukernel__neon_sqrt_x8) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_ARM64 |
| |
| #if XNN_ARCH_ARM64 || XNN_ARCH_ARM64 |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x4, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x8, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x12, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x16, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x20, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x24, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x28, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x32, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x36, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr1rsqrts1fma1adj_x40, xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x4, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x4, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x8, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x12, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x16, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x20, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x24, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x28, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x32, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x36, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, neonfma_nr2fma1adj_x40, xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40, benchmark::utils::CheckNEONFMA) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_ARM64 || XNN_ARCH_ARM64 |
| |
| #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| BENCHMARK_CAPTURE(f32_vsqrt, sse_sqrt_x4, xnn_f32_vsqrt_ukernel__sse_sqrt_x4) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, sse_sqrt_x8, xnn_f32_vsqrt_ukernel__sse_sqrt_x8) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vsqrt, avx_sqrt_x8, xnn_f32_vsqrt_ukernel__avx_sqrt_x8, benchmark::utils::CheckAVX) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx_sqrt_x16, xnn_f32_vsqrt_ukernel__avx_sqrt_x16, benchmark::utils::CheckAVX) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x8, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x16, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x24, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x32, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x40, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x48, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x56, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, fma3_nr1fma1adj_x64, xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64, benchmark::utils::CheckFMA3) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x16, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x32, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x48, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x64, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x80, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x96, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x112, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, avx512f_nr1fma1adj_x128, xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128, benchmark::utils::CheckAVX512F) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| |
| #if XNN_ARCH_WASMSIMD |
| BENCHMARK_CAPTURE(f32_vsqrt, wasmsimd_sqrt_x4, xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, wasmsimd_sqrt_x8, xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| #endif // XNN_ARCH_WASMSIMD |
| |
| BENCHMARK_CAPTURE(f32_vsqrt, scalar_sqrt_x1, xnn_f32_vsqrt_ukernel__scalar_sqrt_x1) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, scalar_sqrt_x2, xnn_f32_vsqrt_ukernel__scalar_sqrt_x2) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| BENCHMARK_CAPTURE(f32_vsqrt, scalar_sqrt_x4, xnn_f32_vsqrt_ukernel__scalar_sqrt_x4) |
| ->RangeMultiplier(10) |
| ->Range(1000, 1000000) |
| ->UseRealTime(); |
| |
| #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| BENCHMARK_MAIN(); |
| #endif |