Issue with Swapchain when rendering is done with compute shaders

[fortmeier]

Hi!

I am experiencing troubles with the swap chain when my output is from a compute shader.
What I see that I get unfinished images like these

It should look like this:

My solution to create the second screenshot was to add another a render pass after the compute shader.

Could this be an issue with MoltenVK? On Windows, I did not notice this problem.

Vulkan version is 1.4.321

(for reference, the issue in my repo: fortmeier/klartraum#15).

[cdavis5e]

Your screenshot suggests the actual presentation to the window is occurring before the compute pass completes. Is your compute shader outputting directly to the swapchain image? If so, did you remember to set VK_IMAGE_USAGE_STORAGE_BIT in the imageUsage member of the VkSwapchainCreateInfoKHR?

[fortmeier]

Thanks for the reply, yes, the compute shader directly writes to the swap chain image. The storage bit is set, see

https://github.com/fortmeier/klartraum/blob/3d1bcfae5c21e7025ceb458202adad542ee6e023/src/vulkan_context.cpp#L301

[cdavis5e]

OK, so that's not it...

From your code, your engine may start multiple compute passes--however, only one semaphore is passed to endRender(). Does it work reliably if you have presentation wait for all the compute passes, i.e. with one semaphore for each one, instead of just the last one?

[fortmeier]

yes, you are right, it could be multiple compute graphs (which are directed a-cyclical graphs of compute and render passes). however, currently I only really use a single compute graph that consists of a render pass for rendering a axis visualization and several compute shader calls for performing the rendering. (all chained by semaphores, but that is not really needed since I use only a single queue at the moment).

In the case of multiple compute graphs I would have to handle that differently and pass all the semaphores to endRender().

[cdavis5e]

So that's not it either.

Looks like we have a real bug here, either in MoltenVK or in Metal, but I'll need to investigate this further. Can you capture a GPU trace and upload it either here or on some file hosting site if it's too big for here? If not, that's fine, since you pointed out the source is available here on GitHub.

[fortmeier]

I am rather new to MacOS graphics development, can you point me to so resource how to caputure a GPU trace?

[cdavis5e]

It's all there in the MoltenVK User Guide.

[cdavis5e]

I should point out one other thing: the resulting GPU trace will be contained in a bundle (i.e. a directory that the Finder treats as a single file), which you will likely have to compress into a zipfile or a tarball for uploading.

[fortmeier]

I captured a GPU trace and it looks suspicious:

#1337 [MTLDrawable present] (green dot) has as dependency only vkCmdBeginRenderPass (orange dot) and no dependency from the following compute shaders.

In the list, #1337 [MTLDrawable present] is listed after #1328 vkQueuePresentKHR MTLCommandBuffer (which waits for the right semaphores) so I guess this is actually a bug in MoltenVK.

[fortmeier]

Btw. I also created the trace of what happens when I add a second render pass after the compute graphs. The #1376 [MTLDrawable present] in that case then depends on the second render pass and then is after the compute passes and all is fine.

This hints to me that there is an issue in MoltenVK when presenting results from compute shaders.