ICL Technical Report ICL-UT-21-03 (Benchmarking FFT...)

[juha_p]

FYII, recent
Interim Report on Benchmarking FFT Libraries on High Performance Systems (PDF, Size: 2.8MB)

EDIT: FYII = For You If Interests

[2DaT]

How relevant is that? Nobody runs audio plugins on HPC hardware .

[mystran]

How relevant is that? Nobody runs audio plugins on HPC hardware .

They also seem to be mostly focused on 1024^3 transforms, which in (complex) double-precision is about 16GB of data. This is not a typical audio dataset at all.

FFT is one of those algorithms where you might choose a different radix based on the number of registers available in a given ISA. It's one of those algorithms where you'll typically switch to a new strategy every time you blow past another level of cache. It's very much an algorithm where at the end of the day it's all about optimizing specific transform sizes for specific set of hardware and the whole problem of optimizing large multi-dimensional transforms is pretty much completely different from the problem of optimizing (relatively) small audio transforms (or even large 1D transforms as you might find in some audio situations).

I could go on and on about this, but the point is that there is no such thing as a high-performance FFT routine. Every realistic FFT implementation is going to be a large collection of routines together with some sort of a planner to choose what to do in any given situation and this means that looking at some random benchmark of a random transform size on a specific set of hardware won't tell you absolutely anything about a drastically different transform size on a completely different set of hardware.

[BertKoor]

Perhaps these guys will benefit:
https://en.m.wikipedia.org/wiki/Low-Fre ... ay_(LOFAR)

I talked to one of its software engineers some years ago. Each small antenna is connected to a Linux box which performs FFT on a standard GPU. It still generates terabytes of data in no time.

[juha_p]

How relevant is that? Nobody runs audio plugins on HPC hardware .

Not relevant directly but, aren't most of those FFT packages available for personal computers as well. I know FFTW is kind of standard in circles... .

Here's one analysis done between FFTE and FFTW: https://www.semanticscholar.org/paper/A ... e1e5d3907d

Are there any open source C/C++ implementations available for FFTE?