Skip to content

Commit e767a0b

Browse files
lmonetalmoneta
committed
Add RDF benchmark
Add also all needed input files
1 parent 09ca844 commit e767a0b

14 files changed

+46110
-1688
lines changed

‎root/tmva/sofie/CMakeLists.txt‎

Lines changed: 33 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -128,6 +128,38 @@ if(ROOT_tmva_FOUND AND ROOT_tmva-sofie_FOUND)
128128
#LIBRARIES TMVA ROOTTMVASofie /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.14.sdk/System/Library/Frameworks/Accelerate.framework
129129
LIBRARIES TMVA ROOTTMVASofie /home/moneta/intel/mkl/lib/intel64/libmkl_intel_lp64.so /home/moneta/intel/mkl/lib/intel64/libmkl_sequential.so /home/moneta/intel/mkl/lib/intel64/libmkl_core.so -lpthread
130130
)
131-
target_include_directories(SOFIEInference PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
131+
# target_include_directories(SOFIEInference PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
132132

133+
set(SOFIE_BLAS_LIBS /home/moneta/intel/mkl/lib/intel64/libmkl_intel_lp64.so /home/moneta/intel/mkl/lib/intel64/libmkl_sequential.so /home/moneta/intel/mkl/lib/intel64/libmkl_core.so -lpthread)
134+
135+
configure_file(input_models/compiled/higgs_model_dense.hxx higgs_model_dense.hxx COPYONLY)
136+
configure_file(input_models/compiled/higgs_model_dense.dat higgs_model_dense.dat COPYONLY)
137+
RB_ADD_GBENCHMARK(RDF_SOFIE_Inference
138+
RDF_SOFIE_Inference.cxx
139+
LABEL short
140+
#LIBRARIES TMVA ROOTTMVASofie openblas
141+
#LIBRARIES TMVA ROOTTMVASofie /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.14.sdk/System/Library/Frameworks/Accelerate.framework
142+
LIBRARIES Core Hist Imt RIO Tree TreePlayer ROOTDataFrame ROOTVecOps TMVA ROOTTMVASofie ${SOFIE_BLAS_LIBS}
143+
)
144+
133145
endif()
146+
147+
configure_file(input_models/higgs_model_dense.json higgs_model_dense.json COPYONLY)
148+
set(LWTNN_INCLUDE_DIR /home/moneta/cernbox/root/tests/tmva/sofie/lwtnn-build/include)
149+
set(LWTNN_LIBS /home/moneta/cernbox/root/tests/tmva/sofie/lwtnn-build/lib/liblwtnn.so)
150+
RB_ADD_GBENCHMARK(RDF_lwtnn_Inference
151+
RDF_lwtnn_Inference.cxx
152+
LABEL short
153+
LIBRARIES Core Hist Imt RIO Tree TreePlayer ROOTDataFrame ROOTVecOps TMVA ROOTTMVASofie ${LWTNN_LIBS}
154+
)
155+
target_include_directories(RDF_lwtnn_Inference PRIVATE ${LWTNN_INCLUDE_DIR})
156+
157+
158+
configure_file(input_models/higgs_model_dense.onnx higgs_model_dense.onnx COPYONLY)
159+
RB_ADD_GBENCHMARK(RDF_ONNXRuntime_Inference
160+
RDF_ONNXRuntime_Inference.cxx
161+
LABEL short
162+
LIBRARIES Core Hist Imt RIO Tree TreePlayer ROOTDataFrame ROOTVecOps TMVA ROOTTMVASofie onnxruntime
163+
)
164+
target_link_directories(RDF_ONNXRuntime_Inference PRIVATE ${ONNXRuntime_LIBRARIES})
165+
target_include_directories(RDF_ONNXRuntime_Inference PRIVATE ${ONNXRuntime_INCLUDE_DIR})

‎root/tmva/sofie/RDF_ONNXRuntime_Inference.cxx‎

Lines changed: 173 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,173 @@
1+
#include "higgs_model_dense.hxx"
2+
#include <iostream>
3+
#include "TROOT.h"
4+
#include "TSystem.h"
5+
#include "ROOT/RDataFrame.hxx"
6+
7+
#include <onnxruntime_cxx_api.h>
8+
9+
10+
#include <string>
11+
#include <fstream>
12+
#include <stdlib.h>
13+
14+
#include <benchmark/benchmark.h>
15+
16+
// template <typename Func>
17+
struct ONNXFunctor {
18+
19+
// std::vector<float> input;
20+
// std::vector<std::shared_ptr<Func>> sessions;
21+
22+
std::map<std::string, double> inputs;
23+
std::vector<std::string> names;
24+
25+
std::shared_ptr<Ort::Session> session;
26+
27+
//td::vector<Ort::Value> input_tensors;
28+
29+
//Ort::Value * ort_input = nullptr;
30+
31+
//float *input_arr = nullptr;
32+
33+
std::vector<const char *> input_node_names;
34+
std::vector<const char *> output_node_names;
35+
36+
std::vector<float> input_tensor_values;
37+
38+
std::vector<int64_t> input_node_dims;
39+
std::vector<int64_t> output_node_dims;
40+
41+
ONNXFunctor(unsigned nslots)
42+
{
43+
44+
Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "benchmark");
45+
46+
std::string model_path = "higgs_model_dense.onnx";
47+
48+
Ort::SessionOptions session_options;
49+
session_options.SetIntraOpNumThreads(1);
50+
session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);
51+
52+
// std::cout << "benchmarking model " << model_path << std::endl;
53+
session = std::make_shared<Ort::Session>(env, model_path.c_str(), session_options);
54+
55+
56+
57+
Ort::AllocatorWithDefaultOptions allocator;
58+
input_node_names.push_back(session->GetInputName(0, allocator));
59+
output_node_names.push_back( session->GetOutputName(0, allocator));
60+
61+
// Getting the shapes
62+
63+
input_node_dims = session->GetInputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
64+
output_node_dims = session->GetOutputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
65+
66+
// Calculating the dimension of the input tensor
67+
68+
//int bsize = input_node_dims[0];
69+
// std::cout << "Using bsize = " << bsize << std::endl;
70+
//int nbatches = nevt / bsize;
71+
72+
size_t input_tensor_size = std::accumulate(input_node_dims.begin(), input_node_dims.end(), 1, std::multiplies<int>());
73+
//std::vector<float> input_tensor_values(input_tensor_size );
74+
75+
input_tensor_values.resize(input_tensor_size);
76+
77+
auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
78+
79+
//input_tensors.push_back(Ort::Value::CreateTensor<float>(
80+
// memory_info, input_tensor_values.data(), input_tensor_size, input_node_dims.data(), input_node_dims.size()) );
81+
82+
83+
// Ort::Value
84+
// Ort::Value *ort_input = new Ort::Value(nullptr);
85+
// // input_tensor =
86+
// *ort_input = Ort::Value::CreateTensor<float>(memory_info, input_tensor_values.data(), input_tensor_values.size(),
87+
// input_node_dims.data(), input_node_dims.size());
88+
89+
//input_arr = input_tensor.GetTensorMutableData<float>();
90+
91+
// Running the model
92+
//input_arr = input_tensors[0].GetTensorMutableData<float>();
93+
94+
///////
95+
96+
97+
// // Load inputs from argv
98+
// std::cout << "input size is " << config.inputs.size() << std::endl;
99+
// for (size_t n = 0; n < config.inputs.size(); n++) {
100+
// inputs[config.inputs.at(n).name] = 0.0;
101+
// names.push_back(config.inputs.at(n).name);
102+
// }
103+
}
104+
105+
double operator()(unsigned nslots, float x0, float x1, float x2, float x3, float x4, float x5, float x6)
106+
{
107+
108+
// not sure how to cache input ort tensor
109+
auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
110+
Ort::Value
111+
input_tensor = Ort::Value::CreateTensor<float>(
112+
memory_info, input_tensor_values.data(), input_tensor_values.size(), input_node_dims.data(), input_node_dims.size());
113+
float * input_arr = input_tensor.GetTensorMutableData<float>();
114+
115+
int off = 0;
116+
input_arr[off] = x0;
117+
input_arr[off + 1] = x1;
118+
input_arr[off + 2] = x2;
119+
input_arr[off + 3] = x3;
120+
input_arr[off + 4] = x4;
121+
input_arr[off + 5] = x5;
122+
input_arr[off + 6] = x6;
123+
124+
125+
126+
auto output_tensors = session->Run(Ort::RunOptions{nullptr}, input_node_names.data(), &input_tensor, 1, output_node_names.data(), 1);
127+
float * floatarr = output_tensors.front().GetTensorMutableData<float>();
128+
return floatarr[0];
129+
}
130+
};
131+
132+
void BM_RDF_ONNX_Inference(benchmark::State &state)
133+
{
134+
135+
int nslot = 1;
136+
if (nslot > 1)
137+
ROOT::EnableImplicitMT(nslot);
138+
139+
auto fileName = "Higgs_data_full.root";
140+
// file is available at "https://cernbox.cern.ch/index.php/s/YuSHwTXBa0UBEhD/download";
141+
// do curl https://cernbox.cern.ch/index.php/s/XaPBtaGrnN38wU0 -o Higgs_data_full.root
142+
if (gSystem->AccessPathName(fileName)) {
143+
std::string cmd = "curl https://cernbox.cern.ch/index.php/s/YuSHwTXBa0UBEhD/download -o ";
144+
cmd += fileName;
145+
gSystem->Exec(cmd.c_str());
146+
}
147+
auto treeName = "test_tree";
148+
ROOT::RDataFrame df(treeName, fileName);
149+
150+
ONNXFunctor functor(nslot);
151+
152+
for (auto _ : state) {
153+
154+
auto h1 = df.DefineSlot("DNN_Value", functor, {"m_jj", "m_jjj", "m_lv", "m_jlv", "m_bb", "m_wbb", "m_wwbb"})
155+
.Histo1D("DNN_Value");
156+
157+
auto t1 = std::chrono::high_resolution_clock::now();
158+
159+
auto n = h1->GetEntries();
160+
auto t2 = std::chrono::high_resolution_clock::now();
161+
auto duration = std::chrono::duration_cast<std::chrono::microseconds>(t2 - t1).count();
162+
163+
std::cout << " Processed " << n << " entries "
164+
<< " time = " << duration / 1.E6 << " (sec) time/event = " << duration / double(n) << " musec"
165+
<< std::endl;
166+
167+
// h1->DrawClone();
168+
}
169+
}
170+
171+
172+
BENCHMARK(BM_RDF_ONNX_Inference)->Unit(benchmark::kMillisecond);
173+
BENCHMARK_MAIN();

‎root/tmva/sofie/RDF_SOFIE_Inference.cxx‎

Lines changed: 76 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,76 @@
1+
#include "higgs_model_dense.hxx"
2+
#include <iostream>
3+
#include <memory>
4+
#include "TROOT.h"
5+
#include "TSystem.h"
6+
#include "ROOT/RDataFrame.hxx"
7+
8+
#include <benchmark/benchmark.h>
9+
10+
// Functor to wrap SOFIE session to RDF functor signature
11+
template <typename Func>
12+
struct SofieFunctor {
13+
std::vector<float> input;
14+
std::vector<std::shared_ptr<Func>> sessions;
15+
SofieFunctor(unsigned nslots) :
16+
input(7*nslots,0) {
17+
for (int i = 0; i < nslots; i++) {
18+
sessions.push_back(std::make_shared<Func>());
19+
}
20+
}
21+
22+
double operator()(unsigned nslots, float x0, float x1, float x2, float x3, float x4, float x5, float x6) {
23+
int off = nslots*7;
24+
input[off] = x0;
25+
input[off+1] = x1;
26+
input[off+2] = x2;
27+
input[off+3] = x3;
28+
input[off+4] = x4;
29+
input[off+5] = x5;
30+
input[off+6] = x6;
31+
auto y = sessions[nslots]->infer(input.data()+nslots*7);
32+
return y[0];
33+
}
34+
35+
};
36+
37+
void BM_RDF_SOFIE_Inference(benchmark::State &state)
38+
{
39+
int nslot = 1;
40+
41+
if (nslot > 1) ROOT::EnableImplicitMT(nslot);
42+
auto fileName = "Higgs_data_full.root";
43+
//file is available at "https://cernbox.cern.ch/index.php/s/YuSHwTXBa0UBEhD/download";
44+
// do curl https://cernbox.cern.ch/index.php/s/XaPBtaGrnN38wU0 -o Higgs_data_full.root
45+
if (gSystem->AccessPathName(fileName) ) {
46+
std::string cmd = "curl https://cernbox.cern.ch/index.php/s/YuSHwTXBa0UBEhD/download -o ";
47+
cmd += fileName;
48+
gSystem->Exec(cmd.c_str());
49+
}
50+
auto treeName = "test_tree";
51+
ROOT::RDataFrame df(treeName, fileName);
52+
53+
54+
SofieFunctor<TMVA_SOFIE_higgs_model_dense::Session> functor(nslot);
55+
56+
for (auto _ : state) {
57+
58+
auto h1 = df.DefineSlot("DNN_Value", functor, {"m_jj", "m_jjj", "m_lv", "m_jlv", "m_bb", "m_wbb", "m_wwbb"})
59+
.Histo1D("DNN_Value");
60+
61+
auto t1 = std::chrono::high_resolution_clock::now();
62+
63+
auto n = h1->GetEntries();
64+
auto t2 = std::chrono::high_resolution_clock::now();
65+
auto duration = std::chrono::duration_cast<std::chrono::microseconds>(t2 - t1).count();
66+
67+
std::cout << " Processed " << n << " entries "
68+
<< " time = " << duration / 1.E6 << " (sec) time/event = " << duration / double(n) << " musec"
69+
<< std::endl;
70+
}
71+
// h1->DrawClone();
72+
}
73+
74+
75+
BENCHMARK(BM_RDF_SOFIE_Inference)->Unit(benchmark::kMillisecond);
76+
BENCHMARK_MAIN();

‎root/tmva/sofie/RDF_lwtnn_Inference.cxx‎

Lines changed: 101 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,101 @@
1+
#include "higgs_model_dense.hxx"
2+
#include <iostream>
3+
#include "TROOT.h"
4+
#include "TSystem.h"
5+
#include "ROOT/RDataFrame.hxx"
6+
7+
#include "lwtnn/LightweightNeuralNetwork.hh"
8+
#include "lwtnn/parse_json.hh"
9+
10+
#include <string>
11+
#include <fstream>
12+
#include <stdlib.h>
13+
14+
#include <benchmark/benchmark.h>
15+
16+
// template <typename Func>
17+
struct LWTNNFunctor {
18+
19+
// std::vector<float> input;
20+
// std::vector<std::shared_ptr<Func>> sessions;
21+
22+
std::map<std::string, double> inputs;
23+
std::vector<std::string> names;
24+
25+
std::shared_ptr<lwt::LightweightNeuralNetwork> model;
26+
27+
LWTNNFunctor(unsigned nslots)
28+
{
29+
std::string config_filename("higgs_model_dense.json");
30+
std::ifstream config_file(config_filename);
31+
auto config = lwt::parse_json(config_file);
32+
33+
// Set up neural network model from config
34+
model = std::make_shared<lwt::LightweightNeuralNetwork>(config.inputs, config.layers, config.outputs);
35+
36+
// Load inputs from argv
37+
std::cout << "input size is " << config.inputs.size() << std::endl;
38+
for (size_t n = 0; n < config.inputs.size(); n++) {
39+
inputs[config.inputs.at(n).name] = 0.0;
40+
names.push_back(config.inputs.at(n).name);
41+
}
42+
}
43+
44+
double operator()(unsigned nslots, float x0, float x1, float x2, float x3, float x4, float x5, float x6)
45+
{
46+
// int off = 0;//nslots*7;
47+
inputs[names[0]] = x0;
48+
inputs[names[1]] = x1;
49+
inputs[names[2]] = x2;
50+
inputs[names[3]] = x3;
51+
inputs[names[4]] = x4;
52+
inputs[names[5]] = x5;
53+
inputs[names[6]] = x6;
54+
55+
auto y = model->compute(inputs);
56+
return y.begin()->second;
57+
}
58+
};
59+
60+
void BM_RDF_LWTNN_Inference(benchmark::State &state)
61+
{
62+
63+
int nslot = 1;
64+
if (nslot > 1)
65+
ROOT::EnableImplicitMT(nslot);
66+
67+
auto fileName = "Higgs_data_full.root";
68+
// file is available at "https://cernbox.cern.ch/index.php/s/YuSHwTXBa0UBEhD/download";
69+
// do curl https://cernbox.cern.ch/index.php/s/XaPBtaGrnN38wU0 -o Higgs_data_full.root
70+
if (gSystem->AccessPathName(fileName)) {
71+
std::string cmd = "curl https://cernbox.cern.ch/index.php/s/YuSHwTXBa0UBEhD/download -o ";
72+
cmd += fileName;
73+
gSystem->Exec(cmd.c_str());
74+
}
75+
auto treeName = "test_tree";
76+
ROOT::RDataFrame df(treeName, fileName);
77+
78+
LWTNNFunctor functor(nslot);
79+
80+
for (auto _ : state) {
81+
82+
auto h1 = df.DefineSlot("DNN_Value", functor, {"m_jj", "m_jjj", "m_lv", "m_jlv", "m_bb", "m_wbb", "m_wwbb"})
83+
.Histo1D("DNN_Value");
84+
85+
auto t1 = std::chrono::high_resolution_clock::now();
86+
87+
auto n = h1->GetEntries();
88+
auto t2 = std::chrono::high_resolution_clock::now();
89+
auto duration = std::chrono::duration_cast<std::chrono::microseconds>(t2 - t1).count();
90+
91+
std::cout << " Processed " << n << " entries "
92+
<< " time = " << duration / 1.E6 << " (sec) time/event = " << duration / double(n) << " musec"
93+
<< std::endl;
94+
95+
// h1->DrawClone();
96+
}
97+
}
98+
99+
100+
BENCHMARK(BM_RDF_LWTNN_Inference)->Unit(benchmark::kMillisecond);
101+
BENCHMARK_MAIN();

0 commit comments

Comments
 (0)