RFE: capture Vulkan compute shaders in a headless application

Is there any support for capturing Vulkan profiling info similar to what’s produced by Nsight/CUDA or nvprof or nvvp?

I write a lot of compute shaders that aren’t associated with a swapchain/frame until late in the development cycle and after a large amount of integration.

For this reason, it would be valuable to profile a standalone Vulkan compute shader being launched on a queue.

I suspect many CUDA devs are porting kernels to Vulkan 1.1 now that it has first class support for compute primitives.

Let me know if you need more info.

Hi Allanmac,

Currently, we don’t support a custom “frame” delimiter on Vulkan, but we do with OpenGL.

https://docs.nvidia.com/nsight-graphics/UserGuide/index.html#connect_launch_tab (Scroll down to OGL launch options)

I’ve added a story to our backlog to support a custom delimiter which would allow you to capture. (DG-4479)

I’ll ask the development team if there is an internal work around as well.


Ah, a frame delimiter in Vulkan sounds perfect.

It would be great if there was a workaround in the interim.

Thanks Seth!