File size: 9,899 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 |
// Copyright 2019 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 <random>
#include <vector>
#include <benchmark/benchmark.h>
#include <fp16/fp16.h>
#include "bench/conv.h"
#include "bench/utils.h"
#include <xnnpack.h>
#include <xnnpack/aligned-allocator.h>
#include <xnnpack/common.h>
#include <xnnpack/igemm.h>
#include <xnnpack/indirection.h>
#include <xnnpack/microfnptr.h>
#include <xnnpack/microparams-init.h>
#include <xnnpack/operator.h>
#include <xnnpack/pack.h>
static void f16_igemm(benchmark::State& state,
xnn_f16_igemm_minmax_ukernel_fn igemm,
uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr,
xnn_init_f16_minmax_params_fn init_params,
benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
if (isa_check != nullptr && !isa_check(state)) {
return;
}
const size_t input_height = state.range(0);
const size_t input_width = state.range(1);
const size_t kernel_height = state.range(2);
const size_t kernel_width = state.range(3);
const size_t kernel_size = kernel_height * kernel_width;
const size_t padding_height = state.range(4);
const size_t padding_width = state.range(5);
const size_t subsampling = state.range(6);
const size_t dilation = state.range(7);
const size_t group_input_channels = state.range(8);
const size_t group_output_channels = state.range(9);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
const size_t output_pixel_stride = group_output_channels;
const size_t input_pixel_stride = group_input_channels;
const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
const size_t padding_left = padding_width / 2;
const size_t padding_top = padding_height / 2;
const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
const size_t output_size = output_height * output_width;
const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr);
const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr);
const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr);
std::vector<uint16_t> a(input_height * input_width * input_pixel_stride + XNN_EXTRA_BYTES / sizeof(uint16_t));
std::generate(a.begin(), a.end(), std::ref(f16rng));
std::vector<uint16_t> k(group_output_channels * kernel_height * kernel_width * group_input_channels);
std::generate(k.begin(), k.end(), std::ref(f16rng));
std::vector<uint16_t> b(group_output_channels);
std::generate(b.begin(), b.end(), std::ref(f16rng));
std::vector<uint16_t> z(group_input_channels + XNN_EXTRA_BYTES / sizeof(uint16_t));
const size_t w_elements = (kernel_size * kc_stride + 1) * nc_stride;
const size_t i_elements = mc_stride * kernel_size;
const size_t c_elements = output_height * output_width * output_pixel_stride;
const size_t num_buffers = 1 +
benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
sizeof(uint16_t) * (w_elements + c_elements) + sizeof(void*) * i_elements);
std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> w(w_elements * num_buffers);
std::fill(w.begin(), w.end(), 0);
xnn_pack_f16_conv_goki_w(
1 /* groups */, group_output_channels, kernel_size, group_input_channels,
nr, kr, sr, k.data(), b.data(), w.data(), 0 /* extra bytes */, nullptr);
for (size_t n = 1; n < num_buffers; n++) {
std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements);
}
std::vector<const uint16_t*> i(i_elements * num_buffers);
xnn_operator convolution_op = { };
convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data());
convolution_op.input = a.data();
convolution_op.input_pixel_stride = input_pixel_stride;
convolution_op.zero_buffer = z.data();
convolution_op.groups = 1;
convolution_op.group_input_channels = group_input_channels;
convolution_op.batch_size = 1;
convolution_op.input_height = input_height;
convolution_op.input_width = input_width;
convolution_op.output_height = output_height;
convolution_op.output_width = output_width;
convolution_op.kernel_height = kernel_height;
convolution_op.kernel_width = kernel_width;
convolution_op.stride_height = subsampling;
convolution_op.stride_width = subsampling;
convolution_op.dilation_height = dilation;
convolution_op.dilation_width = dilation;
convolution_op.padding_top = padding_top;
convolution_op.padding_left = padding_left;
xnn_indirection_init_conv2d(&convolution_op, mr, XNN_LOG2_SIZEOF_HALF);
for (size_t n = 1; n < num_buffers; n++) {
std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements);
}
std::vector<uint16_t> c(c_elements * num_buffers);
std::fill(c.begin(), c.end(), UINT16_C(0x7E00) /* NaN */);
// Prepare minmax parameters.
xnn_f16_minmax_params params;
init_params(¶ms,
UINT16_C(0x7C00) /* inf */, UINT16_C(0xFC00) /* -inf */);
size_t buffer_index = 0;
for (auto _ : state) {
state.PauseTiming();
benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t));
buffer_index = (buffer_index + 1) % num_buffers;
state.ResumeTiming();
for (uint32_t m = 0; m < output_size; m += mr) {
const uint32_t mb = min(output_size - m, mr);
for (uint32_t n = 0; n < group_output_channels; n += nr) {
const uint32_t nb = min(group_output_channels - n, nr);
igemm(
mb, nb, group_input_channels * sizeof(uint16_t), kernel_size * mr * sizeof(void*),
reinterpret_cast<const void**>(i.data()) + buffer_index * i_elements + m,
w.data() + buffer_index * w_elements + n * (kc_stride * kernel_size + 1),
c.data() + buffer_index * c_elements + m * group_output_channels + n, group_output_channels * sizeof(uint16_t), nr * sizeof(uint16_t),
0, z.data(), ¶ms);
}
}
}
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["FLOPS"] = benchmark::Counter(
uint64_t(state.iterations()) * 2 *
output_height * output_width *
group_input_channels * group_output_channels *
kernel_height * kernel_width,
benchmark::Counter::kIsRate);
}
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
static void f16_f32acc_igemm_1x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_1x8__avx2_broadcast, 1, 8, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_4x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_4x8__avx2_broadcast, 4, 8, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_5x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_5x8__avx2_broadcast, 5, 8, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_6x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_6x8__avx2_broadcast, 6, 8, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_7x8__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_7x8__avx2_broadcast, 7, 8, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_1x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_1x16__avx2_broadcast, 1, 16, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_3x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_3x16__avx2_broadcast, 3, 16, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_4x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_4x16__avx2_broadcast, 4, 16, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
static void f16_f32acc_igemm_5x16__avx2_broadcast(benchmark::State& state, const char* net) {
f16_igemm(state, xnn_f16_f32acc_igemm_minmax_ukernel_5x16__avx2_broadcast, 5, 16, 1, 1,
xnn_init_f16_minmax_avx_params, benchmark::utils::CheckAVX2);
}
BENCHMARK_CONV(f16_f32acc_igemm_1x8__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_4x8__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_5x8__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_6x8__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_7x8__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_1x16__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_3x16__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_4x16__avx2_broadcast)
BENCHMARK_CONV(f16_f32acc_igemm_5x16__avx2_broadcast)
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif
|