I use NVRHI over Vulkan to do some headless processing, mixing many images into single texture atlas.
According to Nsight systems, I have performance issue caused by Async Compute. But I don’t use any compute shaders in my algorithm at all, so I can’t understand where this Async Compute workload comes from.
Some more facts:
- I recorded API trace using Vulkan configurator in text format (161 MB). It does not contain string “dispatch” at all. But I see vkCmdDrawIndexed and vkCmdCopyBufferToImage there, which I actually use.
- The Async Compute workload is shown in lane “GPU metrics” \ “Compute in Flight” \ “Async Compute in Flight” of Nsight Systems.
- The Async Compute workload happens exactly once per input image to be blended, so I’m rather convinced it is caused by my algorithm and is not just some other program on the computer.
- The Async Compute workload does not want to overlap with PCI-E traffic of image uploads on async transfer queue. That’s why I want to eliminate it, basically.
- The Async Compute workload supposedly happens simultaneously with the main draw call on graphics queue. This is the only draw call which uses imageLoad/imageStore GLSL commands, and the only one with framebuffer without attachments. In case it helps.
Unfortunately, I have little experience in any of these technologies, so maybe I miss something obvious.
The main question is: what async compute workloads can be implicitly started by NVIDIA Vulkan implementation , and how can I trace down what it is?
