Skip to content

[SYCL][CUDA][PI] Improve performance of piQueueFinish #6201

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
steffenlarsen merged 5 commits into from
Jun 1, 2022
Merged

[SYCL][CUDA][PI] Improve performance of piQueueFinish #6201

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e971e91
improve performance of queue.wait()
t4c1 May 26, 2022
ac20914
bugfix
t4c1 May 26, 2022
08d2af1
format
t4c1 May 26, 2022
732fedc
addressed review comment
t4c1 May 27, 2022
5d4cf4a
format
t4c1 May 27, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion sycl/plugins/cuda/pi_cuda.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2393,7 +2393,7 @@ pi_result cuda_piQueueFinish(pi_queue command_queue) {
nullptr); // need PI_ERROR_INVALID_EXTERNAL_HANDLE error code
ScopedContext active(command_queue->get_context());

command_queue->for_each_stream([&result](CUstream s) {
command_queue->sync_streams([&result](CUstream s) {
result = PI_CHECK_ERROR(cuStreamSynchronize(s));
});

Expand Down
43 changes: 42 additions & 1 deletion sycl/plugins/cuda/pi_cuda.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -391,6 +391,8 @@ struct _pi_queue {
std::atomic_uint32_t transfer_stream_idx_;
unsigned int num_compute_streams_;
unsigned int num_transfer_streams_;
unsigned int last_sync_compute_streams_;
unsigned int last_sync_transfer_streams_;
unsigned int flags_;
std::mutex compute_stream_mutex_;
std::mutex transfer_stream_mutex_;
Expand All @@ -403,7 +405,9 @@ struct _pi_queue {
transfer_streams_{std::move(transfer_streams)}, context_{context},
device_{device}, properties_{properties}, refCount_{1}, eventCount_{0},
compute_stream_idx_{0}, transfer_stream_idx_{0},
num_compute_streams_{0}, num_transfer_streams_{0}, flags_(flags) {
num_compute_streams_{0}, num_transfer_streams_{0},
last_sync_compute_streams_{0}, last_sync_transfer_streams_{0},
flags_(flags) {
cuda_piContextRetain(context_);
cuda_piDeviceRetain(device_);
}
Expand Down Expand Up @@ -440,6 +444,43 @@ struct _pi_queue {
}
}

template <typename T> void sync_streams(T &&f) {
{
std::lock_guard<std::mutex> compute_guard(compute_stream_mutex_);
unsigned int size = static_cast<unsigned int>(compute_streams_.size());
unsigned int start = last_sync_compute_streams_;
unsigned int end = num_compute_streams_ < size
? num_compute_streams_
: compute_stream_idx_.load() % size;
Comment thread
steffenlarsen marked this conversation as resolved.
Outdated
if (size == 1 && end == 0) {
f(compute_streams_[0]);
} else {
for (unsigned int i = start; i != end; (++i < size) ? i : (i = 0)) {
f(compute_streams_[i]);
}
last_sync_compute_streams_ = end;
}
}
{
std::lock_guard<std::mutex> transfer_guard(transfer_stream_mutex_);
unsigned int size = static_cast<unsigned int>(transfer_streams_.size());
if (size > 0) {
unsigned int start = last_sync_transfer_streams_;
unsigned int end = num_transfer_streams_ < size
? num_transfer_streams_
: transfer_stream_idx_.load() % size;
if (size == 1 && end == 0) {
f(transfer_streams_[0]);
} else {
for (unsigned int i = start; i != end; (++i < size) ? i : (i = 0)) {
f(transfer_streams_[i]);
}
last_sync_transfer_streams_ = end;
}
}
}
}

_pi_context *get_context() const { return context_; };

pi_uint32 increment_reference_count() noexcept { return ++refCount_; }
Expand Down