[CUDA][HIP] Add Event Caching #1538
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
github-actions
bot
added
cuda
hdelan
reviewed
Apr 26, 2024
GeorgeWeb
reviewed
Apr 30, 2024
MartinWehking
force-pushed
the
event_cuda_amd_reuse
branch
from
2e18c74 to
c19b5ac
Compare
hdelan
reviewed
May 9, 2024
hdelan
reviewed
May 9, 2024
hdelan
suggested changes
May 9, 2024
hdelan
reviewed
May 9, 2024
hdelan
reviewed
May 9, 2024
hdelan
reviewed
May 10, 2024
MartinWehking
force-pushed
the
event_cuda_amd_reuse
branch
from
c3879cf to
590897e
Compare
hdelan
reviewed
May 10, 2024
hdelan
reviewed
May 10, 2024
MartinWehking
force-pushed
the
event_cuda_amd_reuse
branch
from
97e39dc to
ed75dc7
Compare
added 5 commits
May 13, 2024 12:18
Don't destroy events that are not needed anymore, but put them on a stack of its associated queue (If there is one). Use events from the stack before creating new ones. This caching mechanism ensures that the number of CUDA API calls gets significantly reduced for event creations and destructions.
Follow a similar logic as in #13f097c. Reduce the number of HIP API calls for event creations and destructions.
Make the retrieval and pushing of events onto the stack thread-safe by locking it with mutexes. Use these mutexes inside the queue, which is responsible for caching the events. Apply further minor changes, such as returning the result status for event releases directly, usage of the proper UR assertions and proper styling for variable names.
Copy thread-safety mechanism for event caching from CUDA to HIP. (#590897e9103c06ef1f1d86a6edf7c7eb525bd7d8) Use atomic deletion of cached events in queue destructor. Refactor minor styling/redundancy issues for the CUDA and HIP adapter.
Use a mutex to make the emptiness check and retrieval of cached events atomic.
MartinWehking
force-pushed
the
event_cuda_amd_reuse
branch
from
0938386 to
ffa30c7
Compare
Return UR_RESULT_ERROR_INVALID_EVENT only at the end of a try catch block and UR_RESULT_SUCCESS as default.
hdelan
reviewed
May 13, 2024
hdelan
reviewed
May 13, 2024
hdelan
reviewed
May 13, 2024
added 2 commits
May 13, 2024 15:36
Modify the return value inside urEventRelease.
Don't use a lock_guard inside another mutex locking function. This led to deadlocks for the CUDA and HIP adapters.
MartinWehking
changed the title
MartinWehking
changed the title
hdelan
reviewed
May 15, 2024
| @@ -57,6 +61,8 @@ struct ur_queue_handle_t_ { | |||
| std::mutex ComputeStreamMutex; | |||
| std::mutex TransferStreamMutex; | |||
| std::mutex BarrierMutex; | |||
| // The event cache might be accessed in multiple threads. | |||
| std::mutex CacheMutex; | |||
There was a problem hiding this comment.
I might be inclined to have this close in code to the CachedEvents
|
Small nit. Otherwise LGTM |
hdelan
approved these changes
May 15, 2024
MartinWehking
commented
May 15, 2024
| // Returns and removes an event from the CachedEvents stack. | ||
| ur_event_handle_t get_cached_event() { | ||
| std::lock_guard<std::mutex> CacheGuard(CacheMutex); | ||
| assert(!CachedEvents.empty()); |
There was a problem hiding this comment.
@konradkusiak97 spotted this. Need to change this to detail::ur::assertion
MartinWehking
commented
May 15, 2024
| // Returns and removes an event from the CachedEvents stack. | ||
| ur_event_handle_t get_cached_event() { | ||
| std::lock_guard<std::mutex> CacheGuard(CacheMutex); | ||
| assert(!CachedEvents.empty()); |
There was a problem hiding this comment.
@konradkusiak97 spotted this. Need to change this to detail::ur::assertion
|
Closed since it did not yield significant improvement in benchmarking |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Don't destroy events that are not needed anymore, but put them on a stack
of its associated queue (If there is one).
Use events from the stack before creating new ones.
Make the operations for putting events onto the stack and retrieving them
atomic to ensure thread safety.
This caching mechanism ensures that the number of CUDA API calls gets
significantly reduced for event creations and destructions.