pytorch/torch/csrc/jit/codegen/cuda/instrumentation.cpp
Chester Liu 8177f63c91 Reorganize and refine the Windows.h import in C++ files (#48009)
Summary:
This PR aims to reduce the import overhead and symbol noises from the `windows.h` headers.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/48009

Reviewed By: gchanan

Differential Revision: D25045840

Pulled By: ezyang

fbshipit-source-id: 01fda70f433ba2dd0cd2d7cd676ab6ffe9d98b90
2020-11-20 14:21:09 -08:00

73 lines
1.8 KiB
C++

#include <torch/csrc/jit/codegen/cuda/instrumentation.h>
#include <torch/csrc/WindowsTorchApiMacro.h>
#ifdef _WIN32
#include <c10/util/win32-headers.h>
#else
#include <pthread.h>
#include <unistd.h>
#endif
namespace torch {
namespace jit {
namespace fuser {
namespace cuda {
namespace inst {
Trace::Trace() {
const char* trace_filename = getenv("PYTORCH_CUDA_FUSER_TRACE");
if (trace_filename != nullptr) {
log_file_ = fopen(trace_filename, "w");
TORCH_CHECK(log_file_ != nullptr, "Can't open trace file");
// Disable the file stream buffering, since it may result
// in torn writes in multi-threaded tracing
setbuf(log_file_, nullptr);
// Print the trace prologue
// (including a dummy TRACE_START event)
fprintf(log_file_, "{\n\"traceEvents\": [\n");
start_timestamp_ = Clock::now();
logEvent('I', "TRACE_START");
}
}
Trace::~Trace() {
if (log_file_ != nullptr) {
// Print trace epilogue
logEvent('I', "TRACE_END", ' ');
fprintf(log_file_, "],\n\"displayTimeUnit\": \"ms\"\n}\n");
fclose(log_file_);
}
}
void Trace::logEvent(char ph, const char* name, char sep) {
const std::chrono::duration<double> d = Clock::now() - start_timestamp_;
const double elapsed = d.count() * 1e6;
#ifdef _WIN32
const unsigned int pid = GetCurrentProcessId();
const unsigned int tid = GetCurrentThreadId();
#else
const unsigned int pid = getpid();
const unsigned int tid = std::hash<pthread_t>{}(pthread_self());
#endif // _WIN32
fprintf(
log_file_,
"{ \"name\": \"%s\", \"ph\": \"%c\", \"pid\": %u, \"tid\": %u, \"ts\": %.0f }%c\n",
name,
ph,
pid,
tid,
elapsed,
sep);
}
} // namespace inst
} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch