Implicit command queue flushes triggered by clReleaseCommandQueue

[kpet]

The specification currently states

clReleaseCommandQueue performs an implicit flush to issue any previously queued OpenCL
commands in command_queue.

A few implementers observed in the 2024/09/03 teleconference that this leads to unwanted flushes when RAII objects that hold command queues frequently call clRetainCommandQueue and clReleaseCommandQueue. The intent may have been to solve the issue described in #1238.

Do people rely on this behaviour? Is this something we want to relax? Would we want explicit opt-in/opt-out control for this behaviour?

[bashbaug]

Do people rely on this behaviour? Is this something we want to relax? Would we want explicit opt-in/opt-out control for this behaviour?

I think the currently specified behavior makes sense when releasing last reference to the command-queue, because this causes the command-queue to become inaccessible thereby eliminating the ability to explicitly flush the command-queue in the future.

Would it make sense to relax the behavior so there is no implicit flush if the reference count is still nonzero after releasing? This avoid the unwanted (implicit) flushes with RAII objects but retains the implicit flush behavior when the last reference is released.

Depending how we resolve #1238 we may want to consider strengthening the behavior when the last reference is released. It would be interesting to understand how other APIs (e.g. Vulkan, SYCL) handle this case.

[kpet]

Discussed in 2024/09/24 teleconference. Requiring the flush on each call to clReleaseCommandQueue has unwanted side-effects. Consensus that the intent likely was to free applications from having to insert explicit flushes before they relinquish their last refcount on a queue in multi-queue scenarios. Consensus (+1 from Arm, Intel, Qualcomm) to change the behaviour in the next minor version of OpenCL to only require the flush when the app is giving up its last refcount on the queue. Implementations would still be free to insert additional implicit flushes if they wanted.

[bashbaug]

Discussed in the November 5th teleconference. Pending what we find in the tests added by KhronosGroup/OpenCL-CTS#2134, we may revisit whether fixing this will require a new minor OpenCL version.