pytorch/torch/csrc/jit/python/python_tracer.h
Ilia Cherniavskii f5c95d5cf1 Source code level attribution in profiler (#43898)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/43898

Adding with_source parameter to enable tracking source code
(filename and line) in profiler for eager, torchscript and autograd
modes

Test Plan:
python test/test_profiler.py
```
Name                                 Self CPU total %  Self CPU total   CPU total %      CPU total        CPU time avg     Number of Calls  Source Location
-----------------------------------  ---------------  ---------------  ---------------  ---------------  ---------------  ---------------  --------------------------------------------
ts_method_1                          10.43%           235.364us        36.46%           822.920us        822.920us        1                test/test_profiler.py(70): test_source
aten::add                            7.52%            169.833us        8.88%            200.439us        200.439us        1                test/test_profiler.py(69): test_source
aten::normal_                        6.26%            141.380us        6.26%            141.380us        141.380us        1                test/test_profiler.py(67): test_source
aten::add                            5.80%            130.830us        8.41%            189.800us        63.267us         3                test/test_profiler.py(72): test_source
aten::sum                            5.02%            113.340us        8.39%            189.475us        189.475us        1                test/test_profiler.py(64): ts_method_1
aten::add                            4.58%            103.346us        6.33%            142.847us        142.847us        1                test/test_profiler.py(62): ts_method_1
aten::mul                            4.05%            91.498us         9.62%            217.113us        217.113us        1                test/test_profiler.py(71): test_source
aten::add                            4.03%            90.880us         5.60%            126.405us        126.405us        1                test/test_profiler.py(58): ts_method_2
aten::empty                          3.49%            78.735us         3.49%            78.735us         19.684us         4                test/test_profiler.py(72): test_source
```

Reviewed By: ngimel

Differential Revision: D23432664

Pulled By: ilia-cher

fbshipit-source-id: 83ad7ebe0c2502494d3b48c4e687802db9c77615
2020-09-30 00:57:35 -07:00

37 lines
854 B
C++

#pragma once
#include <torch/csrc/jit/frontend/source_range.h>
#include <torch/csrc/jit/frontend/tracer.h>
#include <torch/csrc/python_headers.h>
#include <torch/csrc/utils/pybind.h>
#include <memory>
#include <string>
namespace torch {
namespace jit {
struct Module;
namespace tracer {
void initPythonTracerBindings(PyObject* module);
SourceRange getPythonInterpreterSourceRange();
Node* preRecordPythonTrace(
THPObjectPtr pyobj,
const std::string& arg_types,
at::ArrayRef<autograd::Variable> inputs,
std::vector<THPObjectPtr> scalar_args);
std::pair<std::shared_ptr<Graph>, Stack> createGraphByTracing(
const py::function& func,
Stack inputs,
const py::function& var_name_lookup_fn,
bool strict,
bool force_outplace,
Module* self = nullptr);
} // namespace tracer
} // namespace jit
} // namespace torch