Skip to content
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

Jump to bottom

[SYCL] Fix reductions not working inside graph #50

Merged
merged 3 commits into from
Dec 13, 2022
Merged

[SYCL] Fix reductions not working inside graph #50

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f195d51
[SYCL] Fix reductions not working inside graph
Bensuo Dec 8, 2022
ca55c3d
[SYCL] Add todo comment to remove code duplication in piQueueFlush
Bensuo Dec 9, 2022
fd0d114
[SYCL] Remove code duplication in piQueueFlush
Bensuo Dec 9, 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
32 changes: 8 additions & 24 deletions sycl/plugins/level_zero/pi_level_zero.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1583,9 +1583,11 @@ void _pi_queue::CaptureIndirectAccesses() {
pi_result _pi_queue::executeCommandList(pi_command_list_ptr_t CommandList,
bool IsBlocking,
bool OKToBatchCommand) {
// When executing a Graph, defer execution
if( this->Properties & PI_EXT_ONEAPI_QUEUE_LAZY_EXECUTION ) return PI_SUCCESS;

// When executing a Graph, defer execution if this is a command
// which could be batched (i.e. likely a kernel submission)
if (this->Properties & PI_EXT_ONEAPI_QUEUE_LAZY_EXECUTION && OKToBatchCommand)
return PI_SUCCESS;

bool UseCopyEngine = CommandList->second.isCopy(this);

// If the current LastCommandEvent is the nullptr, then it means
Expand Down Expand Up @@ -3827,31 +3829,13 @@ pi_result piQueueFinish(pi_queue Queue) {
// so this can be left as a no-op.
pi_result piQueueFlush(pi_queue Queue) {
if( Queue->Properties & PI_EXT_ONEAPI_QUEUE_LAZY_EXECUTION ) {

pi_command_list_ptr_t CommandList{};
// TODO:
CommandList = Queue->LazyCommandListMap.begin();

auto &ZeCommandQueue = CommandList->second.ZeQueue;
// Scope of the lock must be till the end of the function, otherwise new mem
// allocs can be created between the moment when we made a snapshot and the
// moment when command list is closed and executed. But mutex is locked only
// if indirect access tracking enabled, because std::defer_lock is used.
// unique_lock destructor at the end of the function will unlock the mutex
// if it was locked (which happens only if IndirectAccessTrackingEnabled is
// true).
std::unique_lock<pi_shared_mutex> ContextsLock(
Queue->Device->Platform->ContextsMutex, std::defer_lock);

// Close the command list and have it ready for dispatch.
ZE_CALL(zeCommandListClose, (CommandList->first));

// Offload command list to the GPU for asynchronous execution
auto ZeCommandList = CommandList->first;
auto ZeResult = ZE_CALL_NOCHECK(
zeCommandQueueExecuteCommandLists,
(ZeCommandQueue, 1, &ZeCommandList, CommandList->second.ZeFence));

Queue->executeCommandList(CommandList, false, false);
}
(void)Queue;
return PI_SUCCESS;
}

Expand Down