On subgroup-size-tuned compute kernels: You mention that the kernel author will have to tune the kernel for each subgroup size but why not use specialization constants so you can emit 1 SPIRV and configure subgroup size at pipeline creation time? https://github.com/linebender/piet-gpu/blob/f56714a1ce3853b2b7d1c8d794c07235468ffc65/piet-gpu-hal/examples/shader/prefix.comp#L39 …
-
-
-
Yes, I should probably mention those, they're a good fit for this use case.
- 1 more reply
New conversation -
-
-
Fantastic post! Minor note though: vuda is definitely NOT meant to run CUDA workloads on Vulkan. Vuda just provides a wrapper around the Vulkan host API to make it resemble the CUDA host API, but your compute kernels still have to be Vulkan’s SPIRV, not CUDA’s PTX.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Nice post! Really useful for me as I am trying to write Morton codes and sort them in WebGPU, so I will most likely need a prefix sum at some point for the radix sort
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.