tkonolige commented on code in PR #11448:
URL: https://github.com/apache/tvm/pull/11448#discussion_r893899782
##########
src/runtime/thread_pool.cc:
##########
@@ -116,9 +115,16 @@ class ParallelLauncher {
num_pending_.fetch_sub(1);
par_errors_[task_id] = TVMGetLastError();
has_error_.store(true);
+ std::unique_lock<std::mutex> lock(finish_mutex_);
+ finish_cv_.notify_all();
}
// Signal that one job has finished.
- void SignalJobFinish() { num_pending_.fetch_sub(1); }
+ void SignalJobFinish() {
+ if (num_pending_.fetch_sub(1) <= 1) {
Review Comment:
If I understand this correctly, you are adding a mutex `finish_cv_` that
guards `num_pending_`. The mutex serves as a wait to have threads wait until
`num_pending_` hits zero. Given that `num_pending_` is now guarded, can we turn
it into just a raw `int32_t` to avoid additional overhead?
It seems like c++20 would allow a cleaner solution by allowing us to wait on
an atomic https://en.cppreference.com/w/cpp/atomic/atomic/wait.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
