File size: 9,042 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 |
// 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 <cfloat>
#include <cmath>
#include <functional>
#include <memory>
#include <numeric>
#include <random>
#include <vector>
#include <cpuinfo.h>
#include <pthreadpool.h>
#include <benchmark/benchmark.h>
#include "bench/utils.h"
#include <xnnpack/aligned-allocator.h>
#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/math-stubs.h>
struct ComputeErrorContext {
const float* input;
const float* output;
float* error;
};
static void ComputeError(
struct ComputeErrorContext* context,
size_t start,
size_t range)
{
const float* input = context->input;
const float* output = context->output;
float* error = context->error;
for (size_t i = start; i < start + range; i++) {
const double output_ref = std::expm1(double(input[i]));
const double abs_error = std::abs(output_ref - double(output[i]));
const float output_abs = std::abs(output_ref);
const float output_ulp = uint32_as_float(float_as_uint32(output_abs) + 1) - output_abs;
error[i] = float(abs_error / output_ulp);
}
}
static void ExpM1Error(benchmark::State& state,
xnn_f32_unary_math_fn expm1,
benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
if (!cpuinfo_initialize()) {
state.SkipWithError("failed cpuinfo init");
return;
}
if (isa_check != nullptr && !isa_check(state)) {
return;
}
// The smallest x for which expm1f(x) is not saturated at -1 (-0x1.154244p+4f).
const uint32_t min_input = UINT32_C(0xC18AA122);
// Number of elements in one block of inputs/outputs.
// Combining multiple elements in a block reduce function call overhead.
const size_t block_size = 1048576;
// Number of elements in one parallelization tile. Worker threads process this many elements in each task.
const size_t tile_size = 64;
uint32_t num_threads = cpuinfo_get_processors_count();
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
// Use all cores except for the least performant cluster
if (cpuinfo_get_clusters_count() > 1) {
num_threads -= cpuinfo_get_cluster(cpuinfo_get_clusters_count() - 1)->core_count;
}
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
std::unique_ptr<pthreadpool, decltype(&pthreadpool_destroy)> threadpool(
pthreadpool_create(num_threads), pthreadpool_destroy);
std::vector<float, AlignedAllocator<float, 64>> x(block_size);
std::vector<float, AlignedAllocator<float, 64>> y(block_size);
std::vector<float> ulp_error(block_size);
float max_ulp_error = 0.0f;
ComputeErrorContext context;
context.input = x.data();
context.output = y.data();
context.error = ulp_error.data();
for (auto _ : state) {
for (uint32_t n = min_input; int32_t(n) < 0; n -= block_size) {
for (uint32_t i = 0; i < block_size; i++) {
x[i] = uint32_as_float(std::max<uint32_t>(n - i, 0x80000000));
}
std::fill(y.begin(), y.end(), std::nanf(""));
expm1(block_size * sizeof(float), x.data(), y.data());
pthreadpool_parallelize_1d_tile_1d(
threadpool.get(),
reinterpret_cast<pthreadpool_task_1d_tile_1d_t>(ComputeError),
static_cast<void*>(&context),
block_size, tile_size, 0 /* flags */);
max_ulp_error = std::accumulate(ulp_error.cbegin(), ulp_error.cend(), max_ulp_error,
static_cast<const float& (*)(const float&, const float&)>(std::max<float>));
}
}
state.counters["ULPERROR"] = benchmark::Counter(max_ulp_error);
}
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
BENCHMARK_CAPTURE(ExpM1Error, neon_rr2_lut16_p3,
xnn_math_f32_expm1minus__neon_rr2_lut16_p3,
benchmark::utils::CheckNEON)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, neon_rr2_p6,
xnn_math_f32_expm1minus__neon_rr2_p6,
benchmark::utils::CheckNEON)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, neonfma_rr1_lut16_p3,
xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3,
benchmark::utils::CheckNEONFMA)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, neonfma_rr1_p6,
xnn_math_f32_expm1minus__neonfma_rr1_p6,
benchmark::utils::CheckNEONFMA)
->Unit(benchmark::kMillisecond)
->Iterations(1);
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
BENCHMARK_CAPTURE(ExpM1Error, avx512f_rr1_lut16_p3_perm,
xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm,
benchmark::utils::CheckAVX512F)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx512f_rr1_p6,
xnn_math_f32_expm1minus__avx512f_rr1_p6,
benchmark::utils::CheckAVX512F)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_lut4_p4_perm,
xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm,
benchmark::utils::CheckAVX2)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_lut8_p4_perm,
xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm,
benchmark::utils::CheckAVX2)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_lut16_p3_gather,
xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather,
benchmark::utils::CheckAVX2)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_p6,
xnn_math_f32_expm1minus__avx2_rr1_p6,
benchmark::utils::CheckAVX2)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx_rr2_lut4_p4_perm,
xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm,
benchmark::utils::CheckAVX)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx_rr2_lut16_p3,
xnn_math_f32_expm1minus__avx_rr2_lut16_p3,
benchmark::utils::CheckAVX)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, avx_rr2_p6,
xnn_math_f32_expm1minus__avx_rr2_p6,
benchmark::utils::CheckAVX)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, sse2_rr2_lut16_p3,
xnn_math_f32_expm1minus__sse2_rr2_lut16_p3)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, sse2_rr2_p6,
xnn_math_f32_expm1minus__sse2_rr2_p6)
->Unit(benchmark::kMillisecond)
->Iterations(1);
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
#if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_lut16_p3_andnot,
xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_lut16_p3_max,
xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_p6_andnot,
xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_p6_max,
xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max)
->Unit(benchmark::kMillisecond)
->Iterations(1);
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut4_p4,
xnn_math_f32_expm1minus__scalar_rr2_lut4_p4)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut8_p3,
xnn_math_f32_expm1minus__scalar_rr2_lut8_p3)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut8_p4,
xnn_math_f32_expm1minus__scalar_rr2_lut8_p4)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut16_p3,
xnn_math_f32_expm1minus__scalar_rr2_lut16_p3)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut16_p4,
xnn_math_f32_expm1minus__scalar_rr2_lut16_p4)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_p5,
xnn_math_f32_expm1minus__scalar_rr2_p5)
->Unit(benchmark::kMillisecond)
->Iterations(1);
BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_p6,
xnn_math_f32_expm1minus__scalar_rr2_p6)
->Unit(benchmark::kMillisecond)
->Iterations(1);
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif
|