Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
提交 77907a35 编写于 作者: T tensor-tang
浏览文件
操作

refine benchmark template

上级 8e785fec
显示空白变更内容
内联 并排
Showing with 85 addition and 318 deletion
paddle/fluid/operators/jit/benchmark.cc
浏览文件 @ 77907a35
...@@ -51,49 +51,41 @@ std::vector<int> TestSizes() { ...@@ -51,49 +51,41 @@ std::vector<int> TestSizes() {
return s; return s;
} }
// return this function avg time template <typename KernelTuples, typename... Args>
template <typename T, typename KernelTuples> struct BenchFunc {
double BenchXYZNFunc(const typename KernelTuples::func_type tgt, // return this function avg time
const std::vector<T>& x, const std::vector<T>& y, double operator()(const typename KernelTuples::func_type tgt, Args... args) {
std::vector<T>& z) { // NOLINT
const T* x_data = x.data();
const T* y_data = y.data();
const int d = z.size();
T* z_data = z.data();
for (int i = 0; i < FLAGS_burning; ++i) { for (int i = 0; i < FLAGS_burning; ++i) {
tgt(x_data, y_data, z_data, d); tgt(args...);
} }
auto start = GetCurrentUS(); auto start = GetCurrentUS();
for (int i = 0; i < FLAGS_repeat; ++i) { for (int i = 0; i < FLAGS_repeat; ++i) {
tgt(x_data, y_data, z_data, d); tgt(args...);
} }
auto end = GetCurrentUS(); auto end = GetCurrentUS();
return (end - start) / FLAGS_repeat; return (end - start) / FLAGS_repeat;
} }
};
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType> namespace jit = paddle::operators::jit;
void BenchXYZNKernel() {
namespace jit = paddle::operators::jit; template <jit::KernelType KT, typename KernelTuples, typename PlaceType,
for (int d : TestSizes()) { typename... Args>
void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
BenchFunc<KernelTuples, Args...> benchmark;
std::vector<std::pair<std::string, double>> infos; std::vector<std::pair<std::string, double>> infos;
std::vector<T> x(d), y(d), z(d); // test refer
RandomVec<T>(d, x.data()); auto refer = jit::GetRefer<KT, KernelTuples>();
RandomVec<T>(d, y.data()); if (!refer) {
// refer LOG(FATAL) << "Refer can not be empty!";
auto refer = jit::GetRefer<KT, jit::XYZNTuples<T>>();
if (refer) {
auto res = BenchXYZNFunc<T, jit::XYZNTuples<T>>(refer, x, y, z);
infos.push_back(std::make_pair("Refer", res));
} }
infos.push_back(std::make_pair("Refer", benchmark(refer, args...)));
// test jitcode // test jitcode
auto jitcode = jit::GetJitCode<KT, jit::XYZNTuples<T>, PlaceType>(d); auto jitcode = jit::GetJitCode<KT, KernelTuples, PlaceType>(attr);
if (jitcode) { if (jitcode) {
auto res = BenchXYZNFunc<T, jit::XYZNTuples<T>>(jitcode, x, y, z); infos.push_back(std::make_pair("JitCode", benchmark(jitcode, args...)));
infos.push_back(std::make_pair("JitCode", res));
} }
// test all impls in more // test all impls in more
jit::KernelKey kkey(KT, PlaceType()); jit::KernelKey kkey(KT, PlaceType());
auto& pool = jit::KernelPool().Instance().AllKernels(); auto& pool = jit::KernelPool().Instance().AllKernels();
...@@ -101,201 +93,66 @@ void BenchXYZNKernel() { ...@@ -101,201 +93,66 @@ void BenchXYZNKernel() {
if (iter != pool.end()) { if (iter != pool.end()) {
auto& impls = iter->second; auto& impls = iter->second;
for (auto& impl : impls) { for (auto& impl : impls) {
auto i = dynamic_cast<const jit::KernelImpl<jit::XYZNTuples<T>>*>( auto i = dynamic_cast<const jit::KernelImpl<KernelTuples>*>(impl.get());
impl.get()); if (i && i->UseMe(attr)) {
if (i && i->UseMe(d)) {
auto more = i->GetFunc(); auto more = i->GetFunc();
auto res = BenchXYZNFunc<T, jit::XYZNTuples<T>>(more, x, y, z); infos.push_back(std::make_pair("More", benchmark(more, args...)));
infos.push_back(std::make_pair("More", res));
} }
} }
} }
// Test result from Get function // Test result from Get function
auto tgt = jit::Get<KT, jit::XYZNTuples<T>, PlaceType>(d); auto tgt = jit::Get<KT, KernelTuples, PlaceType>(attr);
if (!tgt) { if (!tgt) {
LOG(ERROR) << "Target can not be empty!"; LOG(FATAL) << "Target can not be empty!";
} }
auto res = BenchXYZNFunc<T, jit::XYZNTuples<T>>(tgt, x, y, z); infos.push_back(std::make_pair("Target", benchmark(tgt, args...)));
infos.push_back(std::make_pair("Target", res));
// print // print
std::ostringstream loginfos; std::ostringstream loginfos;
loginfos << "Kernel Type: " << jit::to_string(KT) << ", size " << d << ": "; loginfos << "Kernel Type " << jit::to_string(KT) << ": " << attr << ": ";
for (auto pair : infos) { for (auto pair : infos) {
loginfos << pair.first << " takes " << pair.second << " us; "; loginfos << pair.first << " takes " << pair.second << " us; ";
} }
LOG(INFO) << loginfos.str(); LOG(INFO) << loginfos.str();
}
} }
// return this function avg time template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <typename T, typename KernelTuples> void BenchXYZNKernel() {
double BenchAXYNFunc(const typename KernelTuples::func_type tgt, const T a, for (int d : TestSizes()) {
const std::vector<T>& x, std::vector<T> x(d), y(d), z(d);
std::vector<T>& y) { // NOLINT RandomVec<T>(d, x.data());
const T* x_data = x.data(); RandomVec<T>(d, y.data());
T* y_data = y.data(); BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data(), y.data(),
const int d = y.size(); z.data(), d);
for (int i = 0; i < FLAGS_burning; ++i) {
tgt(&a, x_data, y_data, d);
}
auto start = GetCurrentUS();
for (int i = 0; i < FLAGS_repeat; ++i) {
tgt(&a, x_data, y_data, d);
} }
auto end = GetCurrentUS();
return (end - start) / FLAGS_repeat;
} }
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType> template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void BenchAXYNKernel() { void BenchAXYNKernel() {
namespace jit = paddle::operators::jit;
for (int d : TestSizes()) { for (int d : TestSizes()) {
std::vector<std::pair<std::string, double>> infos;
const T a = static_cast<T>(3); const T a = static_cast<T>(3);
std::vector<T> x(d), y(d); std::vector<T> x(d), y(d);
RandomVec<T>(d, x.data()); RandomVec<T>(d, x.data());
// test refer BenchAllImpls<KT, jit::AXYNTuples<T>, PlaceType>(d, &a, x.data(), y.data(),
auto refer = jit::GetRefer<KT, jit::AXYNTuples<T>>(); d);
if (refer) {
auto res = BenchAXYNFunc<T, jit::AXYNTuples<T>>(refer, a, x, y);
infos.push_back(std::make_pair("Refer", res));
}
// test jitcode
auto jitcode = jit::GetJitCode<KT, jit::AXYNTuples<T>, PlaceType>(d);
if (jitcode) {
auto res = BenchAXYNFunc<T, jit::AXYNTuples<T>>(jitcode, a, x, y);
infos.push_back(std::make_pair("JitCode", res));
}
// test all impls in more
jit::KernelKey kkey(KT, PlaceType());
auto& pool = jit::KernelPool().Instance().AllKernels();
auto iter = pool.find(kkey);
if (iter != pool.end()) {
auto& impls = iter->second;
for (auto& impl : impls) {
auto i = dynamic_cast<const jit::KernelImpl<jit::AXYNTuples<T>>*>(
impl.get());
if (i && i->UseMe(d)) {
auto more = i->GetFunc();
auto res = BenchAXYNFunc<T, jit::AXYNTuples<T>>(more, a, x, y);
infos.push_back(std::make_pair("More", res));
} }
}
}
// Test result from Get function
auto tgt = jit::Get<KT, jit::AXYNTuples<T>, PlaceType>(d);
if (!tgt) {
LOG(ERROR) << "Target can not be empty!";
}
auto res = BenchAXYNFunc<T, jit::AXYNTuples<T>>(tgt, a, x, y);
infos.push_back(std::make_pair("Target", res));
// print
std::ostringstream loginfos;
loginfos << "Kernel Type: " << jit::to_string(KT) << ", size " << d << ": ";
for (auto pair : infos) {
loginfos << pair.first << " takes " << pair.second << " us; ";
}
LOG(INFO) << loginfos.str();
}
}
// return this function avg time
template <typename T, typename KernelTuples>
double BenchXYNFunc(const typename KernelTuples::func_type tgt,
const std::vector<T>& x,
std::vector<T>& y) { // NOLINT
const T* x_data = x.data();
T* y_data = y.data();
const int d = y.size();
for (int i = 0; i < FLAGS_burning; ++i) {
tgt(x_data, y_data, d);
}
auto start = GetCurrentUS();
for (int i = 0; i < FLAGS_repeat; ++i) {
tgt(x_data, y_data, d);
}
auto end = GetCurrentUS();
return (end - start) / FLAGS_repeat;
} }
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType> template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void BenchXYNKernel() { void BenchXYNKernel() {
namespace jit = paddle::operators::jit;
for (int d : TestSizes()) { for (int d : TestSizes()) {
std::vector<std::pair<std::string, double>> infos;
std::vector<T> x(d), y(d); std::vector<T> x(d), y(d);
RandomVec<T>(d, x.data()); RandomVec<T>(d, x.data());
// test refer BenchAllImpls<KT, jit::XYNTuples<T>, PlaceType>(d, x.data(), y.data(), d);
auto refer = jit::GetRefer<KT, jit::XYNTuples<T>>();
if (refer) {
auto res = BenchXYNFunc<T, jit::XYNTuples<T>>(refer, x, y);
infos.push_back(std::make_pair("Refer", res));
}
// test jitcode
auto jitcode = jit::GetJitCode<KT, jit::XYNTuples<T>, PlaceType>(d);
if (jitcode) {
auto res = BenchXYNFunc<T, jit::XYNTuples<T>>(jitcode, x, y);
infos.push_back(std::make_pair("JitCode", res));
}
// test all impls in more
jit::KernelKey kkey(KT, PlaceType());
auto& pool = jit::KernelPool().Instance().AllKernels();
auto iter = pool.find(kkey);
if (iter != pool.end()) {
auto& impls = iter->second;
for (auto& impl : impls) {
auto i =
dynamic_cast<const jit::KernelImpl<jit::XYNTuples<T>>*>(impl.get());
if (i && i->UseMe(d)) {
auto more = i->GetFunc();
auto res = BenchXYNFunc<T, jit::XYNTuples<T>>(more, x, y);
infos.push_back(std::make_pair("More", res));
}
}
}
// Test result from Get function
auto tgt = jit::Get<KT, jit::XYNTuples<T>, PlaceType>(d);
if (!tgt) {
LOG(ERROR) << "Target can not be empty!";
}
auto res = BenchXYNFunc<T, jit::XYNTuples<T>>(tgt, x, y);
infos.push_back(std::make_pair("Target", res));
// print
std::ostringstream loginfos;
loginfos << "Kernel Type: " << jit::to_string(KT) << ", size " << d << ": ";
for (auto pair : infos) {
loginfos << pair.first << " takes " << pair.second << " us; ";
}
LOG(INFO) << loginfos.str();
} }
} }
// return this function avg time
template <typename T, typename KernelTuples>
double BenchLSTMFunc(const typename KernelTuples::func_type tgt,
const paddle::operators::jit::lstm_attr_t* attr,
paddle::operators::jit::lstm_t* step) {
for (int i = 0; i < FLAGS_burning; ++i) {
tgt(step, attr);
}
auto start = GetCurrentUS();
for (int i = 0; i < FLAGS_repeat; ++i) {
tgt(step, attr);
}
auto end = GetCurrentUS();
return (end - start) / FLAGS_repeat;
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType> template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void BenchLSTMKernel() { void BenchLSTMKernel() {
namespace jit = paddle::operators::jit;
for (bool use_peephole : {true, false}) { for (bool use_peephole : {true, false}) {
for (int d : TestSizes()) { for (int d : TestSizes()) {
const jit::lstm_attr_t attr(d, jit::vsigmoid, jit::vtanh, jit::vtanh, const jit::lstm_attr_t attr(d, jit::vsigmoid, jit::vtanh, jit::vtanh,
use_peephole); use_peephole);
std::vector<std::pair<std::string, double>> infos;
std::vector<T> x(4 * d), ct_1(d), ct(d), ht(d), wp(3 * d), checked(2 * d); std::vector<T> x(4 * d), ct_1(d), ct(d), ht(d), wp(3 * d), checked(2 * d);
RandomVec<T>(4 * d, x.data(), -2.f, 2.f); RandomVec<T>(4 * d, x.data(), -2.f, 2.f);
RandomVec<T>(3 * d, wp.data(), -2.f, 2.f); RandomVec<T>(3 * d, wp.data(), -2.f, 2.f);
...@@ -315,77 +172,15 @@ void BenchLSTMKernel() { ...@@ -315,77 +172,15 @@ void BenchLSTMKernel() {
step.wp = wp_data; step.wp = wp_data;
step.checked = checked_data; step.checked = checked_data;
} }
BenchAllImpls<KT, jit::LSTMTuples<T>, PlaceType>(attr, &step, &attr);
// test refer
auto refer = jit::GetRefer<KT, jit::LSTMTuples<T>>();
if (refer) {
auto res = BenchLSTMFunc<T, jit::LSTMTuples<T>>(refer, &attr, &step);
infos.push_back(std::make_pair("Refer", res));
} }
// test jitcode
auto jitcode = jit::GetJitCode<KT, jit::LSTMTuples<T>, PlaceType>(attr);
if (jitcode) {
auto res = BenchLSTMFunc<T, jit::LSTMTuples<T>>(jitcode, &attr, &step);
infos.push_back(std::make_pair("JitCode", res));
}
// test all impls in more
jit::KernelKey kkey(KT, PlaceType());
auto& pool = jit::KernelPool().Instance().AllKernels();
auto iter = pool.find(kkey);
if (iter != pool.end()) {
auto& impls = iter->second;
for (auto& impl : impls) {
auto i = dynamic_cast<const jit::KernelImpl<jit::LSTMTuples<T>>*>(
impl.get());
if (i && i->UseMe(attr)) {
auto more = i->GetFunc();
auto res = BenchLSTMFunc<T, jit::LSTMTuples<T>>(more, &attr, &step);
infos.push_back(std::make_pair("More", res));
}
}
}
// Test result from Get function
auto tgt = jit::Get<KT, jit::LSTMTuples<T>, PlaceType>(attr);
if (!tgt) {
LOG(ERROR) << "Target can not be empty!";
}
auto res = BenchLSTMFunc<T, jit::LSTMTuples<T>>(tgt, &attr, &step);
infos.push_back(std::make_pair("Target", res));
// print
std::ostringstream loginfos;
loginfos << "Kernel Type: " << jit::to_string(KT)
<< ", Sigmoid,Tanh,Tanh, " << (use_peephole ? "Peephole_" : "")
<< " size " << d << ": ";
for (auto pair : infos) {
loginfos << pair.first << " takes " << pair.second << " us; ";
} }
LOG(INFO) << loginfos.str();
}
}
}
// return this function avg time
template <typename T, typename KernelTuples>
double BenchGRUFunc(const typename KernelTuples::func_type tgt,
const paddle::operators::jit::gru_attr_t* attr,
paddle::operators::jit::gru_t* step) {
for (int i = 0; i < FLAGS_burning; ++i) {
tgt(step, attr);
}
auto start = GetCurrentUS();
for (int i = 0; i < FLAGS_repeat; ++i) {
tgt(step, attr);
}
auto end = GetCurrentUS();
return (end - start) / FLAGS_repeat;
} }
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType> template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void BenchGRUKernel() { void BenchGRUKernel() {
namespace jit = paddle::operators::jit;
for (int d : TestSizes()) { for (int d : TestSizes()) {
const jit::gru_attr_t attr(d, jit::vsigmoid, jit::vtanh); const jit::gru_attr_t attr(d, jit::vsigmoid, jit::vtanh);
std::vector<std::pair<std::string, double>> infos;
std::vector<T> x(3 * d), ht_1(d), ht(d); std::vector<T> x(3 * d), ht_1(d), ht(d);
RandomVec<T>(3 * d, x.data(), -2.f, 2.f); RandomVec<T>(3 * d, x.data(), -2.f, 2.f);
RandomVec<T>(d, ht_1.data(), -2.f, 2.f); RandomVec<T>(d, ht_1.data(), -2.f, 2.f);
...@@ -396,50 +191,7 @@ void BenchGRUKernel() { ...@@ -396,50 +191,7 @@ void BenchGRUKernel() {
step.gates = x_data; step.gates = x_data;
step.ht_1 = ht_1_data; step.ht_1 = ht_1_data;
step.ht = ht_data; step.ht = ht_data;
BenchAllImpls<KT, jit::GRUTuples<T>, PlaceType>(attr, &step, &attr);
// test refer
auto refer = jit::GetRefer<KT, jit::GRUTuples<T>>();
if (refer) {
auto res = BenchGRUFunc<T, jit::GRUTuples<T>>(refer, &attr, &step);
infos.push_back(std::make_pair("Refer", res));
}
// test jitcode
auto jitcode = jit::GetJitCode<KT, jit::GRUTuples<T>, PlaceType>(attr);
if (jitcode) {
auto res = BenchGRUFunc<T, jit::GRUTuples<T>>(jitcode, &attr, &step);
infos.push_back(std::make_pair("JitCode", res));
}
// test all impls in more
jit::KernelKey kkey(KT, PlaceType());
auto& pool = jit::KernelPool().Instance().AllKernels();
auto iter = pool.find(kkey);
if (iter != pool.end()) {
auto& impls = iter->second;
for (auto& impl : impls) {
auto i =
dynamic_cast<const jit::KernelImpl<jit::GRUTuples<T>>*>(impl.get());
if (i && i->UseMe(attr)) {
auto more = i->GetFunc();
auto res = BenchGRUFunc<T, jit::GRUTuples<T>>(more, &attr, &step);
infos.push_back(std::make_pair("More", res));
}
}
}
// Test result from Get function
auto tgt = jit::Get<KT, jit::GRUTuples<T>, PlaceType>(attr);
if (!tgt) {
LOG(ERROR) << "Target can not be empty!";
}
auto res = BenchGRUFunc<T, jit::GRUTuples<T>>(tgt, &attr, &step);
infos.push_back(std::make_pair("Target", res));
// print
std::ostringstream loginfos;
loginfos << "Kernel Type: " << jit::to_string(KT) << ", Sigmoid,Tanh, size "
<< d << ": ";
for (auto pair : infos) {
loginfos << pair.first << " takes " << pair.second << " us; ";
}
LOG(INFO) << loginfos.str();
} }
} }
...@@ -456,16 +208,17 @@ int main(int argc, char* argv[]) { ...@@ -456,16 +208,17 @@ int main(int argc, char* argv[]) {
<< " times."; << " times.";
using T = float; using T = float;
using PlaceType = paddle::platform::CPUPlace; using PlaceType = paddle::platform::CPUPlace;
namespace jit = paddle::operators::jit; // xyzn
BenchXYZNKernel<jit::vmul, T, PlaceType>(); BenchXYZNKernel<jit::vmul, T, PlaceType>();
BenchXYZNKernel<jit::vadd, T, PlaceType>(); BenchXYZNKernel<jit::vadd, T, PlaceType>();
BenchXYZNKernel<jit::vaddrelu, T, PlaceType>(); BenchXYZNKernel<jit::vaddrelu, T, PlaceType>();
BenchXYZNKernel<jit::vsub, T, PlaceType>(); BenchXYZNKernel<jit::vsub, T, PlaceType>();
// axyn
BenchAXYNKernel<jit::vscal, T, PlaceType>(); BenchAXYNKernel<jit::vscal, T, PlaceType>();
BenchAXYNKernel<jit::vaddbias, T, PlaceType>(); BenchAXYNKernel<jit::vaddbias, T, PlaceType>();
// act // xyn
BenchXYNKernel<jit::vrelu, T, PlaceType>(); BenchXYNKernel<jit::vrelu, T, PlaceType>();
BenchXYNKernel<jit::videntity, T, PlaceType>(); BenchXYNKernel<jit::videntity, T, PlaceType>();
BenchXYNKernel<jit::vexp, T, PlaceType>(); BenchXYNKernel<jit::vexp, T, PlaceType>();
......
paddle/fluid/operators/jit/helper.h
浏览文件 @ 77907a35
...@@ -14,6 +14,7 @@ ...@@ -14,6 +14,7 @@
#pragma once #pragma once
#include <iostream>
#include <string> #include <string>
#include <vector> #include <vector>
#include "paddle/fluid/operators/jit/gen_base.h" #include "paddle/fluid/operators/jit/gen_base.h"
...@@ -124,6 +125,19 @@ const char* to_string(KernelType kt); ...@@ -124,6 +125,19 @@ const char* to_string(KernelType kt);
KernelType to_kerneltype(const std::string& act); KernelType to_kerneltype(const std::string& act);
inline std::ostream& operator<<(std::ostream& os, const lstm_attr_t& attr) {
os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate)
<< "],act_cand[" << to_string(attr.act_cand) << "],act_cell["
<< to_string(attr.act_cell) << "],use_peephole["
<< (attr.use_peephole ? "True" : "False") << "]";
return os;
}
inline std::ostream& operator<<(std::ostream& os, const gru_attr_t& attr) {
os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate)
<< "],act_cand[" << to_string(attr.act_cand) << "]";
return os;
}
} // namespace jit } // namespace jit
} // namespace operators } // namespace operators
} // namespace paddle } // namespace paddle
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或