VkFFT VS VkResample

Not quite what I asked for, but close enough for now...

https://github.com/DTolm/VkFFT/discussions/126

Hello, I am the creator of the VkFFT - GPU Fast Fourier Transform library for Vulkan/CUDA/HIP/OpenCL and Level Zero. In the latest update, I have added support for Apple Metal API, which will allow VkFFT to run natively on modern Apple SoC. I have tested it on MacBook Pro with an M1 Pro 8c CPU/14c GPU SoC single precision on 1D batched FFT test of all systems from 2 to 4096. Achieved bandwidth is calculated as 2*system size divided by the time taken per FFT - minimum memory that has to be transferred between DRAM and GPU:

Another possible project to look at is https://github.com/DTolm/VkFFT

Source: I have some of these things implemented in VkFFT that confirm the mentioned scaling of execution times.

https://github.com/DTolm/VkFFT is a project to look at.

Incorrect. Vulkan has compute shaders[1], which are generally usable. Libraries like VkFFT[2] demonstrate basic signal processing in Vulkan. This can certainly be expanded upon, & there are numerous other non-graphical uses.

There is a Vulkan ML TSG (Technical Subgroup), which is supposed to be working on ML. Even Nvidia is participating, with extensions like VK_NV_cooperative_matrix, which specifically target the ml tensor cores.

There's people could probably say this better/more specifically, but I'll give it a try: Vulkan is, above all, an general standard for dispatching & orchestrating work usually on a GPU. Right now that work is most of often graphics, but that is far from a limit.

SYCL is, imo, the opposite of where we need to go. It's the old historical legacy that CUDA has, of writing really dumb ignorant code & hoping the tools can make it run well on a GPU. Vulkan, on the other hand, asks us to consider deeply what near-to-the-metal resources we are going to need, and demands that we define, dispatch, & complete the actual processing engines on the GPU that will do the work. It's a much much much harder task, but it invites in fantastic levels of close optimization & tuning, allows for far more advanced pipelining & possibilities. If the future is good, it should abandon SYCL and CUDA, and bother to get good at Vulkan.

[1] https://vkguide.dev/docs/gpudriven/compute_shaders/

[2] https://github.com/DTolm/VkFFT

[3] https://www.khronos.org/assets/uploads/developers/presentati...

VkFFT supports convolution calculations - see samples 7, 8 and 9 in the VkFFT repository.

There is already an upscaling reshader using DFTs and zero-padding, but those tend to generate fairly blurry results: vkResample

You can also check https://github.com/DTolm/VkResample for an FFT-based image upscaling example, though it is not exactly a convolution

Real-time FFT upscaling on GPUs VkResample