Viewing profile — DTolm

Hello, I am the author of VkFFT, Tolmachev Dmitrii. I remember VkFFT got a lot of initial traction thanks to Hacker News three years ago. Back then VkFFT was a simple collection of…

The white paper of VkFFT is out. It can be interesting to people who want to know more about performant GPU algorithms in HPC and how VkFFT is designed. VkFFT is an efficient GPU-a…

In VkFFT FFT is done as a part of the so called compute queue. There is no need to transfer data through PCI to show it on screen - there is a graphics queue that has this as its m…

Big 1D FFTs also take a lot of memory by themselves (i.e. 2^28 takes 2GB just to store complex data). Multiple smaller batches can be used in ML applications for example for big ke…

There are surely many different ways to get the job done. VkFFT.h file by itself desn't do any computations btw - it is more like a configurator that launches shaders (sth similar …

1k FFT size in single precision is 1024 x 2 x sizeof(float) = 8KB. If we don't think that it won't utilize full GPU (not even one compute unit) and assume that it scales similarly …

FFT is an extremely bandwidth limited problem, so if most time is taken by one upload by both algorithms, the overall time will be similar. More in-depth analysis of how VkFFT and …

The library only includes vkFFT.h file (in C) and a set of shaders (C-like language compiled to SPIR-V). Vulkan_FFT.cpp is only an example that shows how VkFFT can be used. It also…

Actually, it is still best to aim at zero transfers between GPU and CPU during the execution. The GPU is limited by VRAM-chip bandwidth which is much bigger than the PCI-E bandwidt…

Yes, this is indeed something I would like to add in the future. While adding different radix kernels support for small prime factors is not that hard, writing efficient scheduler …

It is a great open-source license for library projects. For example, Eigen uses it: https://eigen.tuxfamily.org/index.php?title=News:Relicensing... !

Hello! Since the last post VkFFT has experienced a number of huge improvements and optimizations. Namely: -It now supports sequences up to 2^32 in all dimensions (algorithmically, …