Initial open-source release
PiperOrigin-RevId: 271685289
diff --git a/bench/f32-igemm.cc b/bench/f32-igemm.cc
new file mode 100644
index 0000000..cd87bb7
--- /dev/null
+++ b/bench/f32-igemm.cc
@@ -0,0 +1,365 @@
+// 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 <cpuinfo.h>
+
+#include <benchmark/benchmark.h>
+#include "bench/conv.h"
+#include "bench/utils.h"
+#include <xnnpack/AlignedAllocator.h>
+#include <xnnpack/igemm.h>
+#include <xnnpack/indirection.h>
+#include <xnnpack/operator.h>
+#include <xnnpack/pack.h>
+#include <xnnpack/params.h>
+#include <xnnpack/requantization.h>
+
+
+static void IGEMMBenchmark(benchmark::State& state,
+ xnn_f32_igemm_ukernel_function f32_igemm,
+ uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr)
+{
+ if (!cpuinfo_initialize()) {
+ state.SkipWithError("cpuinfo initialization failed");
+ }
+
+ 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), rng);
+
+ 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);
+
+ std::vector<float> a(input_height * input_width * input_pixel_stride);
+ 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));
+
+ std::vector<float> z(group_input_channels);
+
+ 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>(cpuinfo_get_max_cache_size(),
+ sizeof(float) * (w_elements + c_elements) + sizeof(void*) * i_elements);
+
+ std::vector<float, AlignedAllocator<float, 32>> w(w_elements * num_buffers);
+ std::fill(w.begin(), w.end(), 0.0f);
+ xnn_pack_f32_conv_goki_w(
+ 1 /* groups */, group_output_channels, kernel_size, group_input_channels,
+ nr, kr, sr, k.data(), b.data(), w.data());
+ 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 float*> 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, 2 /* log2(sizeof(float)) */);
+ for (size_t n = 1; n < num_buffers; n++) {
+ std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements);
+ }
+
+ std::vector<float> c(c_elements * num_buffers);
+ std::fill(c.begin(), c.end(), std::nanf(""));
+
+ xnn_f32_output_params output_params =
+ xnn_compute_f32_output_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();
+
+ 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_igemm(
+ mb, nb, group_input_channels * sizeof(float), kernel_size * mr * sizeof(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(float), nr * sizeof(float),
+ 0, z.data(), &output_params);
+ }
+ }
+ }
+
+ state.counters["Freq"] = benchmark::utils::GetCurrentCpuFrequency();
+ 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 CPUINFO_ARCH_ARM || CPUINFO_ARCH_ARM64
+ static void f32_igemm_4x2__neon_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x2__neon_ld64, 4, 2, 1, 1);
+ }
+
+ static void f32_igemm_4x4__neon_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x4__neon_ld64, 4, 4, 1, 1);
+ }
+
+ static void f32_igemm_4x8__neon_ld128(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neon_ld128, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__neon_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neon_ld64, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x12__neon_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x12__neon_ld64, 4, 12, 1, 1);
+ }
+
+ static void f32_igemm_6x8__neon_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__neon_ld64, 6, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x2__neonfma_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x2__neonfma_ld64, 4, 2, 1, 1);
+ }
+
+ static void f32_igemm_4x4__neonfma_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x4__neonfma_ld64, 4, 4, 1, 1);
+ }
+
+ static void f32_igemm_4x8__neonfma_ld128(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neonfma_ld128, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__neonfma_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__neonfma_ld64, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x12__neonfma_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x12__neonfma_ld64, 4, 12, 1, 1);
+ }
+
+ static void f32_igemm_6x8__neonfma_ld64(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__neonfma_ld64, 6, 8, 1, 1);
+ }
+
+ BENCHMARK_CONV(f32_igemm_4x12__neon_ld64)
+ BENCHMARK_CONV(f32_igemm_4x12__neonfma_ld64)
+ BENCHMARK_CONV(f32_igemm_4x2__neon_ld64)
+ BENCHMARK_CONV(f32_igemm_4x2__neonfma_ld64)
+ BENCHMARK_CONV(f32_igemm_4x4__neon_ld64)
+ BENCHMARK_CONV(f32_igemm_4x4__neonfma_ld64)
+ BENCHMARK_CONV(f32_igemm_4x8__neon_ld128)
+ BENCHMARK_CONV(f32_igemm_4x8__neon_ld64)
+ BENCHMARK_CONV(f32_igemm_4x8__neonfma_ld128)
+ BENCHMARK_CONV(f32_igemm_4x8__neonfma_ld64)
+ BENCHMARK_CONV(f32_igemm_6x8__neon_ld64)
+ BENCHMARK_CONV(f32_igemm_6x8__neonfma_ld64)
+#endif
+
+#if CPUINFO_ARCH_ARM64
+ static void f32_igemm_1x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x12__aarch64_neonfma_cortex_a53, 1, 12, 1, 1);
+ }
+
+ static void f32_igemm_1x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a57, 1, 8, 1, 1);
+ }
+
+ static void f32_igemm_1x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, 1, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__aarch64_neonfma_cortex_a75, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_5x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_5x8__aarch64_neonfma_cortex_a75, 5, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x12__aarch64_neonfma_cortex_a53(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x12__aarch64_neonfma_cortex_a53, 4, 12, 1, 1);
+ }
+
+ static void f32_igemm_6x8__aarch64_neonfma_cortex_a57(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a57, 6, 8, 1, 1);
+ }
+
+ static void f32_igemm_6x8__aarch64_neonfma_cortex_a73(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a73, 6, 8, 1, 1);
+ }
+
+ static void f32_igemm_6x8__aarch64_neonfma_cortex_a75(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a75, 6, 8, 1, 1);
+ }
+
+ BENCHMARK_CONV(f32_igemm_1x12__aarch64_neonfma_cortex_a53)
+ BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a57)
+ BENCHMARK_CONV(f32_igemm_1x8__aarch64_neonfma_cortex_a75)
+ BENCHMARK_CONV(f32_igemm_4x12__aarch64_neonfma_cortex_a53)
+ BENCHMARK_CONV(f32_igemm_4x8__aarch64_neonfma_cortex_a75)
+ BENCHMARK_CONV(f32_igemm_5x8__aarch64_neonfma_cortex_a75)
+ BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a57)
+ BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a73)
+ BENCHMARK_CONV(f32_igemm_6x8__aarch64_neonfma_cortex_a75)
+#endif /* CPUINFO_ARCH_ARM64 */
+
+#if CPUINFO_ARCH_X86 || CPUINFO_ARCH_X86_64
+ static void f32_igemm_1x8__sse_load1(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__sse_load1, 1, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__sse_load1(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__sse_load1, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_1x8__sse_dup(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__sse_dup, 1, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__sse_dup(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__sse_dup, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_1x8s4__sse(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8s4__sse, 1, 8, 1, 4);
+ }
+
+ static void f32_igemm_4x8s4__sse(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8s4__sse, 4, 8, 1, 4);
+ }
+
+ BENCHMARK_CONV(f32_igemm_1x8__sse_load1)
+ BENCHMARK_CONV(f32_igemm_4x8__sse_load1)
+ BENCHMARK_CONV(f32_igemm_1x8__sse_dup)
+ BENCHMARK_CONV(f32_igemm_4x8__sse_dup)
+ BENCHMARK_CONV(f32_igemm_1x8s4__sse)
+ BENCHMARK_CONV(f32_igemm_4x8s4__sse)
+#endif /* CPUINFO_ARCH_X86 || CPUINFO_ARCH_X86_64 */
+
+#if !CPUINFO_ARCH_WASM && !CPUINFO_ARCH_ASMJS
+ static void f32_igemm_1x8__psimd_loadsplat(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__psimd_loadsplat, 1, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__psimd_loadsplat(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__psimd_loadsplat, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_6x8__psimd_loadsplat(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__psimd_loadsplat, 6, 8, 1, 1);
+ }
+
+ static void f32_igemm_1x8__psimd_splat(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8__psimd_splat, 1, 8, 1, 1);
+ }
+
+ static void f32_igemm_4x8__psimd_splat(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8__psimd_splat, 4, 8, 1, 1);
+ }
+
+ static void f32_igemm_6x8__psimd_splat(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8__psimd_splat, 6, 8, 1, 1);
+ }
+
+ static void f32_igemm_1x8s4__psimd(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x8s4__psimd, 1, 8, 1, 4);
+ }
+
+ static void f32_igemm_4x8s4__psimd(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x8s4__psimd, 4, 8, 1, 4);
+ }
+
+ static void f32_igemm_6x8s4__psimd(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_6x8s4__psimd, 6, 8, 1, 4);
+ }
+
+ BENCHMARK_CONV(f32_igemm_1x8__psimd_loadsplat)
+ BENCHMARK_CONV(f32_igemm_4x8__psimd_loadsplat)
+ BENCHMARK_CONV(f32_igemm_6x8__psimd_loadsplat)
+
+ BENCHMARK_CONV(f32_igemm_1x8__psimd_splat)
+ BENCHMARK_CONV(f32_igemm_4x8__psimd_splat)
+ BENCHMARK_CONV(f32_igemm_6x8__psimd_splat)
+
+ BENCHMARK_CONV(f32_igemm_1x8s4__psimd)
+ BENCHMARK_CONV(f32_igemm_4x8s4__psimd)
+ BENCHMARK_CONV(f32_igemm_6x8s4__psimd)
+#endif /* !CPUINFO_ARCH_WASM && !CPUINFO_ARCH_ASMJS */
+
+static void f32_igemm_1x4__scalar(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_1x4__scalar, 1, 4, 1, 1);
+}
+
+static void f32_igemm_2x4__scalar(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_2x4__scalar, 2, 4, 1, 1);
+}
+
+static void f32_igemm_4x4__scalar(benchmark::State& state, const char* net) {
+ IGEMMBenchmark(state, xnn_f32_igemm_ukernel_4x4__scalar, 4, 4, 1, 1);
+}
+
+BENCHMARK_CONV(f32_igemm_1x4__scalar)
+BENCHMARK_CONV(f32_igemm_2x4__scalar)
+BENCHMARK_CONV(f32_igemm_4x4__scalar)
+
+
+#ifndef XNNPACK_BENCHMARK_NO_MAIN
+BENCHMARK_MAIN();
+#endif