Slow STFT/FFT/IFFT/RFFT layers

[georg95]

Current implementation is naive and for maxtix of 512x2048 takes 512x2048x2048 - 2 billion loops

tfjs/tfjs-backend-webgpu/src/fft_webgpu.ts

Line 58 in 0fc04d9

for (var i = 0; i < uniforms.realShape[1]; i = i + 1) {

STFT/RFFT current Framing window implementation also very slow because of hundreds gpu calls in a loop

tfjs/tfjs-core/src/ops/signal/frame.ts

Line 53 in 0fc04d9

while (start < signal.size) {

Made faster implementation for webgl/webgpu (up to 100+ times faster)
https://github.com/georg95/tfjs-fast-fft
Built it to speed up BirdNET, bird calls audio recognition.