1 /**
2 * Copyright (c) 2016-present, Facebook, Inc.
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include <string>
18 #include <vector>
19
20 #include <ATen/ATen.h>
21 #include "caffe2/core/timer.h"
22 #include "caffe2/utils/string_utils.h"
23 #include <torch/csrc/autograd/grad_mode.h>
24 #include <torch/csrc/jit/mobile/module.h>
25 #include <torch/csrc/jit/mobile/import.h>
26 #include <torch/csrc/jit/serialization/import.h>
27 #include <torch/script.h>
28
29 #include <c10/mobile/CPUCachingAllocator.h>
30
31 #include <chrono>
32 using namespace std::chrono;
33
34 C10_DEFINE_string(model, "", "The given torch script model to benchmark.");
35 C10_DEFINE_int(iter, 10, "The number of iterations to run.");
36 C10_DEFINE_bool(
37 report_pep,
38 true,
39 "Whether to print performance stats for AI-PEP.");
40
main(int argc,char ** argv)41 int main(int argc, char** argv) {
42 c10::SetUsageMessage(
43 "Run model load time benchmark for pytorch model.\n"
44 "Example usage:\n"
45 "./load_benchmark_torch"
46 " --model=<model_file>"
47 " --iter=20");
48 if (!c10::ParseCommandLineFlags(&argc, &argv)) {
49 std::cerr << "Failed to parse command line flags!" << std::endl;
50 return 1;
51 }
52
53 std::cout << "Starting benchmark." << std::endl;
54 CAFFE_ENFORCE(
55 FLAGS_iter >= 0,
56 "Number of main runs should be non negative, provided ",
57 FLAGS_iter,
58 ".");
59
60 caffe2::Timer timer;
61 std::vector<long> times;
62
63 for (int i = 0; i < FLAGS_iter; ++i) {
64 auto start = high_resolution_clock::now();
65
66 #if BUILD_LITE_INTERPRETER
67 auto module = torch::jit::_load_for_mobile(FLAGS_model);
68 #else
69 auto module = torch::jit::load(FLAGS_model);
70 #endif
71
72 auto stop = high_resolution_clock::now();
73 auto duration = duration_cast<microseconds>(stop - start);
74 times.push_back(duration.count());
75 }
76
77 const double micros = static_cast<double>(timer.MicroSeconds());
78 if (FLAGS_report_pep) {
79 for (auto t : times) {
80 std::cout << R"(PyTorchObserver {"type": "NET", "unit": "us", )"
81 << R"("metric": "latency", "value": ")"
82 << t << R"("})" << std::endl;
83 }
84 }
85
86 const double iters = static_cast<double>(FLAGS_iter);
87 std::cout << "Main run finished. Microseconds per iter: "
88 << micros / iters
89 << ". Iters per second: " << 1000.0 * 1000 * iters / micros
90 << std::endl;
91
92 return 0;
93 }
94