File size: 3,710 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 |
// Copyright 2023 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/reduce.h>
static void f16_rsum(
benchmark::State& state,
xnn_f16_rsum_ukernel_fn rsum,
xnn_init_f16_scale_params_fn init_params,
benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
if (isa_check != nullptr && !isa_check(state)) {
return;
}
const size_t elements = state.range(0);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32rng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), std::ref(rng));
auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> input(elements);
std::generate(input.begin(), input.end(), std::ref(f16rng));
xnn_f16_scale_params params;
init_params(¶ms, /*scale=*/fp16_ieee_from_fp32_value(0.1f));
uint16_t output = UINT16_C(0x7E00); /* NaN */
for (auto _ : state) {
rsum(elements * sizeof(uint16_t), input.data(), &output, ¶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 = elements * sizeof(uint16_t);
state.counters["bytes"] =
benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
}
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
BENCHMARK_CAPTURE(f16_rsum, neonfp16arith_x8,
xnn_f16_rsum_ukernel__neonfp16arith_x8,
xnn_init_f16_scale_fp16arith_params,
benchmark::utils::CheckNEONFP16ARITH)
->Apply(benchmark::utils::ReductionParameters<uint16_t>)
->UseRealTime();
BENCHMARK_CAPTURE(f16_rsum, neonfp16arith_x16_acc2,
xnn_f16_rsum_ukernel__neonfp16arith_x16_acc2,
xnn_init_f16_scale_fp16arith_params,
benchmark::utils::CheckNEONFP16ARITH)
->Apply(benchmark::utils::ReductionParameters<uint16_t>)
->UseRealTime();
BENCHMARK_CAPTURE(f16_rsum, neonfp16arith_x24_acc3,
xnn_f16_rsum_ukernel__neonfp16arith_x24_acc3,
xnn_init_f16_scale_fp16arith_params,
benchmark::utils::CheckNEONFP16ARITH)
->Apply(benchmark::utils::ReductionParameters<uint16_t>)
->UseRealTime();
BENCHMARK_CAPTURE(f16_rsum, neonfp16arith_x32_acc2,
xnn_f16_rsum_ukernel__neonfp16arith_x32_acc2,
xnn_init_f16_scale_fp16arith_params,
benchmark::utils::CheckNEONFP16ARITH)
->Apply(benchmark::utils::ReductionParameters<uint16_t>)
->UseRealTime();
BENCHMARK_CAPTURE(f16_rsum, neonfp16arith_x32_acc4,
xnn_f16_rsum_ukernel__neonfp16arith_x32_acc4,
xnn_init_f16_scale_fp16arith_params,
benchmark::utils::CheckNEONFP16ARITH)
->Apply(benchmark::utils::ReductionParameters<uint16_t>)
->UseRealTime();
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif
|