[–] janwas 20d ago ↗ Great writeup, especially liked the line by line explanation :)We actually did manage to get dynamic dispatch working in Compiler explorer with Matt's help: https://gcc.godbolt.org/z/KM3ben7ETMaybe interesting to also refer to Gemma.cpp's version: https://github.com/google/gemma.cpp/blob/dev/ops/ops-inl.h#L...We recently added a FastExpMinusOrZero, only valid for inputs <= 0, which is true for softmax. [–] euos 20d ago ↗ ThanksI have not yet looked in dynamic dispatch - I train on Neon and deploy to WASM so I only target a single arch on each target...
[–] euos 20d ago ↗ ThanksI have not yet looked in dynamic dispatch - I train on Neon and deploy to WASM so I only target a single arch on each target...
2 comments
[ 1.8 ms ] story [ 5.9 ms ] threadWe actually did manage to get dynamic dispatch working in Compiler explorer with Matt's help: https://gcc.godbolt.org/z/KM3ben7ET
Maybe interesting to also refer to Gemma.cpp's version: https://github.com/google/gemma.cpp/blob/dev/ops/ops-inl.h#L...
We recently added a FastExpMinusOrZero, only valid for inputs <= 0, which is true for softmax.
I have not yet looked in dynamic dispatch - I train on Neon and deploy to WASM so I only target a single arch on each target...