Open zhuohoudeputao opened 11 months ago
Since increasingly deep learning or ai related computation relies on FFT and calculates half precision, is there any plan to support fp16 and bf16 precision for fftw?
It should be straightforward to add, in principle. The scalar case would be trivial to add, while SIMD support will take a bit more work.
libbench should also support testing for half precision, too.
Since increasingly deep learning or ai related computation relies on FFT and calculates half precision, is there any plan to support fp16 and bf16 precision for fftw?