105 lines
3.7 KiB
C++
105 lines
3.7 KiB
C++
// Copyright 2022 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 <fp16/fp16.h>
|
|
#include "bench/utils.h"
|
|
|
|
#include <xnnpack.h>
|
|
#include <xnnpack/aligned-allocator.h>
|
|
#include <xnnpack/common.h>
|
|
#include <xnnpack/microfnptr.h>
|
|
#include <xnnpack/microparams-init.h>
|
|
#include <xnnpack/vunary.h>
|
|
|
|
|
|
static void f16_velu(
|
|
benchmark::State& state,
|
|
xnn_f16_velu_ukernel_function elu,
|
|
xnn_init_f16_elu_params_fn init_params,
|
|
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>(-9.0f, 9.0f), std::ref(rng));
|
|
auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
|
|
|
|
std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> x(num_elements);
|
|
std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> y(num_elements);
|
|
std::generate(x.begin(), x.end(), std::ref(f16rng));
|
|
std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */);
|
|
|
|
union xnn_f16_elu_params params;
|
|
init_params(¶ms,
|
|
UINT16_C(0x3C00) /* prescale = 1.0h */,
|
|
UINT16_C(0x3C00) /* alpha = 1.0h */,
|
|
UINT16_C(0x3C00) /* beta = 1.0h */);
|
|
for (auto _ : state) {
|
|
elu(num_elements * sizeof(uint16_t), x.data(), y.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 = 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(uint16_t);
|
|
state.counters["bytes"] =
|
|
benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
|
|
}
|
|
|
|
|
|
#if XNN_ENABLE_ARM_FP16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
|
|
BENCHMARK_CAPTURE(f16_velu, neonfp16arith_rr1_p3_x8,
|
|
xnn_f16_velu_ukernel__neonfp16arith_rr1_p3_x8,
|
|
xnn_init_f16_elu_neonfp16arith_rr1_p3_params,
|
|
benchmark::utils::CheckNEONFP16ARITH)
|
|
->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
|
|
->UseRealTime();
|
|
BENCHMARK_CAPTURE(f16_velu, neonfp16arith_rr1_p3_x16,
|
|
xnn_f16_velu_ukernel__neonfp16arith_rr1_p3_x16,
|
|
xnn_init_f16_elu_neonfp16arith_rr1_p3_params,
|
|
benchmark::utils::CheckNEONFP16ARITH)
|
|
->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
|
|
->UseRealTime();
|
|
#endif // XNN_ENABLE_ARM_FP16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
|
|
|
|
|
|
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
|
|
BENCHMARK_CAPTURE(f16_velu, avx2_rr1_p3_x8,
|
|
xnn_f16_velu_ukernel__avx2_rr1_p3_x8,
|
|
xnn_init_f16_elu_avx2_rr1_p3_params,
|
|
benchmark::utils::CheckAVX2)
|
|
->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
|
|
->UseRealTime();
|
|
BENCHMARK_CAPTURE(f16_velu, avx2_rr1_p3_x16,
|
|
xnn_f16_velu_ukernel__avx2_rr1_p3_x16,
|
|
xnn_init_f16_elu_avx2_rr1_p3_params,
|
|
benchmark::utils::CheckAVX2)
|
|
->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
|
|
->UseRealTime();
|
|
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
|
|
|
|
|
|
#ifndef XNNPACK_BENCHMARK_NO_MAIN
|
|
BENCHMARK_MAIN();
|
|
#endif
|