[L0v2] fix unbounded memory growth of queue's submitted kernels #20847
+29
−8
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
Contributor
Author
|
I've created a PR against 6.2, since this functionality changed a little in the latest version, and will require a different patch. |
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
force-pushed
the
cleanup-large-submitted-kernels
branch
from
753bc9e to
f99e1cd
Compare
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
force-pushed
the
cleanup-large-submitted-kernels
branch
from
f99e1cd to
57aef87
Compare
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
force-pushed
the
cleanup-large-submitted-kernels
branch
from
57aef87 to
156235d
Compare
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
pbalcer
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
L0v2 avoids internally tracking each kernel submission through an event for lifetime management. Instead, when a kernel is submitted to the queue, its handle is added to a vector, to be removed at the next queue synchronization point, urQueueFinish(). This is a much more efficient way of handling kernel tracking, since it avoids taking and storing an event. However, if the application never synchronizes the queue, this vector of submitted kernels will grow unbounded. This patch forcibly synchronizes the queue once the submitted kernels vector reaches a threshold.
nrspruit
approved these changes
Dec 6, 2025
MichalMrozek
reviewed
Dec 8, 2025
| ur_kernel_handle_t hKernel) { | ||
| locked<ur_command_list_manager> &commandList, ur_kernel_handle_t hKernel) { | ||
| if (submittedKernels.size() > MAX_QUEUE_SUBMITTED_KERNELS) { | ||
| synchronize(commandList); |
Contributor
There was a problem hiding this comment.
There is very low chance that we will have 1000 unique kernels, most likely we have duplicates here.
RIght now this will always synchronize every 1k submits which is undesired.
Can you compact the vector by finding duplicates and releasing those?
You just need one kernel instance in container to hold the object all additional ones are not neeeded.
Also what I would consider is not adding kernel to the vector if it is already there.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
L0v2 avoids internally tracking each kernel submission through an event for lifetime management. Instead, when a kernel is submitted to the queue, its handle is added to a vector, to be removed at the next queue synchronization point, urQueueFinish(). This is a much more efficient way of handling kernel tracking, since it avoids taking and storing an event. However, if the application never synchronizes the queue, this vector of submitted kernels will grow unbounded.
This patch forcibly synchronizes the queue once the submitted kernels vector reaches a threshold.