Rework semaphore waiting #2226
Open
Rework semaphore waiting #2226
+401
−241
Conversation
This file contains 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
First step in moving wait logic to the logical queue. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
We still need to support waits on the physical queue for memory clears and sparse initialization. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
This was only necessary due to what seems to be a Halo Infinite game bug. Still flush before waiting on a shared fence to preserve the order for debug purposes. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Use the submission timeline instead of a round-trip with the binary semaphore. Reduces number of queue submissions, and potentially reduces bubbles in single-queue scenarios. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
May in some cases avoid a submission and subsequent signal delays. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Any memory clear or sparse init that happens after command execution on a virtual queue should not delay fence signals on said queue. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
doitsujin
force-pushed
the
queue-waiter-rework
branch
from
79f79cb
to
7589c03
Compare
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
doitsujin
force-pushed
the
queue-waiter-rework
branch
2 times, most recently
from
56c1b59
to
ebe548a
Compare
Reduces number of vkQueueSubmit calls. Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
doitsujin
force-pushed
the
queue-waiter-rework
branch
from
ebe548a
to
67679ed
Compare
| @@ -3298,6 +3298,10 @@ struct d3d12_command_queue | |||
| struct vkd3d_fence_virtual_wait *wait_fences; | |||
| size_t wait_fences_size; | |||
| size_t wait_fence_count; | |||
|
|
|||
| VkSemaphoreSubmitInfo *wait_semaphores; | |||
| @@ -18142,6 +18154,9 @@ static void d3d12_command_queue_flush_waiters(struct d3d12_command_queue *comman | |||
|
|
|||
| if ((vr = VK_CALL(vkQueueSubmit2(vk_queue, 1, &submit_info, VK_NULL_HANDLE)))) | |||
| ERR("Failed to submit semaphore waits, vr %d.\n"); | |||
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.
Workaround for #2215.
This moves fence wait tracking from
vkd3d_queueto the virtual queues and optimizes away most of the empty submissions that only do a semaphore wait and a semaphore signal, which is what caused the massive performance degradation in Cyberpunk since AMDGPU does not handle this very well.Additionally, sparse binding serialization now uses the queue's timeline instead of the binary semaphore. At worst, there will now be one extra
vkQueueSubmiton the first sparse binding call in order to flush waiters, rather than doing a round-trip over the graphics queue each time.One potential concern is that this may delay fence signals in scenarios where we set
NO_STAGGERED_SUBMITin case the app does something likequeue->Wait(...); nothing for quite some time; queue->Signal(...)with no commands being submitted. The only solution I can think of would involve stalling the signaling queue worker until all pending waits have completed in that case, which is something that we're kind of trying to avoid, but I'm not sure if this is even an issue in practice - Black Myth Wukong doesn't seem to run into this problem with FSR framegen enabled at all, so at least UE5 should be fine.Most of this code was written on four hours of sleep, so this will need extensive testing.