You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

162 lines
6.2 KiB

// 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 "bench/conv.h"
#include "bench/utils.h"
#include <xnnpack/AlignedAllocator.h>
#include <xnnpack/common.h>
#include <xnnpack/gemm.h>
#include <xnnpack/im2col.h>
#include <xnnpack/pack.h>
#include <xnnpack/params-init.h>
#include <xnnpack/params.h>
static void Im2ColGEMMBenchmark(benchmark::State& state,
xnn_f32_gemm_minmax_ukernel_function f32_gemm,
uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr,
benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
if (isa_check && !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>(0.0f, 1.0f), std::ref(rng));
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 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);
std::vector<float> a(input_height * input_width * group_input_channels);
std::generate(a.begin(), a.end(), std::ref(f32rng));
std::vector<float> k(group_output_channels * kernel_height * kernel_width * group_input_channels);
std::generate(k.begin(), k.end(), std::ref(f32rng));
std::vector<float> b(group_output_channels);
std::generate(b.begin(), b.end(), std::ref(f32rng));
const size_t w_elements = (kernel_size * kc_stride + 1) * nc_stride;
const size_t c_elements = output_size * group_output_channels;
const size_t num_buffers = 1 +
benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
sizeof(float) * (w_elements + c_elements));
std::vector<float, AlignedAllocator<float, 32>> w(w_elements * num_buffers);
std::fill(w.begin(), w.end(), 0.0f);
xnn_pack_f32_gemm_goi_w(1 /* groups */, group_output_channels, group_input_channels * kernel_size,
nr, kr, sr, k.data(), b.data(), w.data(), 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<float> im2col_buffer(output_size * group_input_channels * kernel_size * group_output_channels);
std::vector<float> c(c_elements * num_buffers);
std::fill(c.begin(), c.end(), std::nanf(""));
xnn_f32_minmax_params params =
xnn_init_f32_minmax_params(-std::numeric_limits<float>::infinity(), +std::numeric_limits<float>::infinity());
size_t buffer_index = 0;
for (auto _ : state) {
state.PauseTiming();
benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(float));
buffer_index = (buffer_index + 1) % num_buffers;
state.ResumeTiming();
const float* inputData = a.data();
if (kernel_size != 1 || subsampling != 1) {
xnn_im2col_conv2d(
output_height, output_width,
kernel_height, kernel_width,
subsampling, subsampling,
dilation, dilation,
input_width, padding_top, padding_left,
group_input_channels * sizeof(float) /* input channels */,
group_input_channels * sizeof(float) /* input stride */,
a.data(), im2col_buffer.data());
inputData = im2col_buffer.data();
}
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);
f32_gemm(
mb, nb, kernel_size * group_input_channels * sizeof(float),
inputData + m * kernel_size * group_input_channels, kernel_size * group_input_channels * sizeof(float),
w.data() + (buffer_index * nc_stride + n) * (kernel_size * kc_stride + 1),
c.data() + (buffer_index * output_size + m) * group_output_channels + n, group_output_channels * sizeof(float), nr * sizeof(float),
&params);
}
}
}
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_ARM64 && XNN_ENABLE_ASSEMBLY
static void f32_gemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) {
Im2ColGEMMBenchmark(state, xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75, 4, 8, 1, 1);
}
BENCHMARK_CONV(f32_gemm_4x8__aarch64_neonfma_cortex_a75)
#endif // XNN_ARCH_ARM64
static void f32_gemm_2x4__scalar(benchmark::State& state, const char* net) {
Im2ColGEMMBenchmark(state, xnn_f32_gemm_minmax_ukernel_2x4__scalar, 2, 4, 1, 1);
}
static void f32_gemm_4x4__scalar(benchmark::State& state, const char* net) {
Im2ColGEMMBenchmark(state, xnn_f32_gemm_minmax_ukernel_4x4__scalar, 4, 4, 1, 1);
}
BENCHMARK_CONV(f32_gemm_2x4__scalar)
BENCHMARK_CONV(f32_gemm_4x4__scalar)
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif