search

eloimoliner / CQT_pytorch

Pytorch implementation of the invertible CQT based on Non-stationary Gabor filters
29 stars 0 forks source link

Suboptimal GPU efficiency #1

Closed eloimoliner closed 1 year ago

eloimoliner commented 1 year ago

The inverse transform implementation is suboptimal. This for loop should be optimized somehow.

https://github.com/eloimoliner/CQT_pytorch/blob/0a976b80f4bb8523cd2dd79cbe335ca8439e3e13/src/CQT_nsgt.py#L224

eloimoliner commented 1 year ago

I mamage to vectorize the multiplications outside the loop. There is a significant performance increase. Still not perfect, though. The variable indexing is something hard to work with.