Add CUDA profiler tools in new framework. #5954
Conversation
There was a problem hiding this comment.
@qingqing01 Overall looks good to me.
In the issue description(#5953), could you please provide a couple of examples where this feature is useful? Especially where nvprof is not sufficient.
paddle/platform/cuda_profiler.h Outdated
| memcpy(buf.data(), tmpl.data(), tmpl.size()); | ||
| auto result = mktemp(buf.data()); | ||
| PADDLE_ENFORCE(strlen(result) != 0); | ||
| std::string config = result; |
There was a problem hiding this comment.
probably config -> config_file would be a better name.
paddle/platform/cuda_profiler.h Outdated
| | ||
| void CudaProfilerStart() { PADDLE_ENFORCE(cudaProfilerStart()); } | ||
| | ||
| void CudaProfilerStop() { PADDLE_ENFORCE((cudaProfilerStop())); } |
python/paddle/v2/fluid/profiler.py Outdated
| ] | ||
| | ||
| | ||
| def nvporf_init(output_file, output_mode=None, flags=None): |
There was a problem hiding this comment.
We currently use google style for Python comment.
http://sphinxcontrib-napoleon.readthedocs.io/en/latest/example_google.html
There was a problem hiding this comment.
Done. Move the comments to the refined CudaProfiler.
python/paddle/v2/fluid/profiler.py Outdated
| core.nvprof_stop() | ||
| | ||
| | ||
| class CudaProfiler(object): |
There was a problem hiding this comment.
This a good implementation.
I recently found a wonderful package named contextmanager. It can make the profiler to one function.
There was a problem hiding this comment.
Done. Cool! Thanks!
python/paddle/v2/fluid/profiler.py Outdated
| core.nvprof_init(output_file, output_mode, flags) | ||
| | ||
| | ||
| def nvporf_start(): |
| """ | ||
| Enables profiler collection by the active CUDA profiling tool. | ||
| """ | ||
| core.nvprof_start() |
There was a problem hiding this comment.
I think this function may not need wrap again in Python code.
There was a problem hiding this comment.
Done. Remove them.
python/paddle/v2/fluid/profiler.py Outdated
| self.out_file = output_file | ||
| nvporf_init(output_file, output_mode, flags) | ||
| | ||
| def __enter__(self): |
There was a problem hiding this comment.
these two functions can be replaced with contextmanager.
| @tonyyang-svail I update the issue description(#5953) and give a simple example. And the output file needs to be made more human-readable later. |
4 participants
Fix #5953
The usage is: