110 lines
2.4 KiB
C++
110 lines
2.4 KiB
C++
#include <cmath>
|
|
#include <cfloat>
|
|
#include <vector>
|
|
#include <random>
|
|
#include <chrono>
|
|
#include <functional>
|
|
#include <algorithm>
|
|
|
|
#include <cpuinfo.h>
|
|
#include <nnpack/macros.h>
|
|
#include <nnpack/blas.h>
|
|
#include <nnpack/AlignedAllocator.h>
|
|
|
|
#include <benchmark/benchmark.h>
|
|
|
|
|
|
template<uint32_t mr_, uint32_t nr_>
|
|
class CONV1x1 : public benchmark::Fixture {
|
|
public:
|
|
inline CONV1x1() {
|
|
cpuinfo_initialize();
|
|
const size_t l1d_size = cpuinfo_get_l1d_cache(0)->size;
|
|
const size_t l1d_reserve = 512;
|
|
kc_ = ((l1d_size - l1d_reserve) / sizeof(float) - mr() * nr()) / (mr() + nr());
|
|
}
|
|
|
|
virtual void SetUp(const benchmark::State&) override {
|
|
const uint_fast32_t seed = std::chrono::system_clock::now().time_since_epoch().count();
|
|
auto rng = std::bind(std::uniform_real_distribution<float>(), std::mt19937(seed));
|
|
|
|
i_.resize(mr() * kc());
|
|
std::generate(i_.begin(), i_.end(), std::ref(rng));
|
|
k_.resize(mr() * kc() + nr());
|
|
std::fill(k_.begin(), k_.end(), std::nanf(""));
|
|
o_.resize(nr() * kc());
|
|
std::generate(o_.begin(), o_.end(), std::ref(rng));
|
|
}
|
|
|
|
virtual void TearDown(benchmark::State& state) override {
|
|
state.SetItemsProcessed(uint64_t(state.iterations()) * 2 * mr() * nr() * kc());
|
|
i_.clear();
|
|
k_.clear();
|
|
o_.clear();
|
|
}
|
|
|
|
inline const float* i() const {
|
|
return i_.data();
|
|
}
|
|
|
|
inline const float* k() const {
|
|
return k_.data();
|
|
}
|
|
|
|
inline float* o() {
|
|
return o_.data();
|
|
}
|
|
|
|
inline uint32_t mr() const {
|
|
return mr_;
|
|
}
|
|
|
|
inline uint32_t nr() const {
|
|
return nr_;
|
|
}
|
|
|
|
inline uint32_t kc() const {
|
|
return kc_;
|
|
}
|
|
|
|
private:
|
|
std::vector<float> i_;
|
|
std::vector<float> k_;
|
|
std::vector<float> o_;
|
|
uint32_t kc_;
|
|
};
|
|
|
|
#if NNP_BACKEND_X86_64
|
|
BENCHMARK_TEMPLATE_F(CONV1x1, fast__neon, 2, 4)(benchmark::State& state) {
|
|
for (auto _ : state) {
|
|
nnp_conv1x1_only_2x4__fma3(mr(), kc(), i(), k(), o());
|
|
}
|
|
}
|
|
#endif
|
|
|
|
#if NNP_BACKEND_ARM
|
|
BENCHMARK_TEMPLATE_F(CONV1x1, fast__neon, 4, 4)(benchmark::State& state) {
|
|
for (auto _ : state) {
|
|
nnp_conv1x1_only_4x4__neon(mr(), kc(), i(), k(), o());
|
|
}
|
|
}
|
|
#endif
|
|
|
|
#if NNP_BACKEND_PSIMD
|
|
BENCHMARK_TEMPLATE_F(CONV1x1, psimd, 2, 8)(benchmark::State& state) {
|
|
for (auto _ : state) {
|
|
nnp_conv1x1_only_2x4__psimd(mr(), kc(), i(), k(), o());
|
|
}
|
|
}
|
|
#endif
|
|
|
|
#if NNP_BACKEND_SCALAR
|
|
BENCHMARK_TEMPLATE_F(CONV1x1, scalar, 2, 4)(benchmark::State& state) {
|
|
for (auto _ : state) {
|
|
nnp_conv1x1_only_2x4__scalar(mr(), kc(), i(), k(), o());
|
|
}
|
|
}
|
|
#endif
|
|
|
|
BENCHMARK_MAIN();
|