Deterministic cuDNN picked FFT for my chess net's 128-channel convolutions: 2.4x slower than needed on an 8x8 board.
4 fixes (a PyTorch trick, TorchScript, a CUDA kernel, a C++ cuDNN hook): 1.5x faster, bit-identical runs, one 2017 GTX 1080 Ti.
https://t.co/9Z7MTDpqvn