@khuasw さんによるSynthesizer Vが使用したニューラルネットワーク最適化手法を一般人向けに噛み砕く説明するツイートのまとめ
Kanru Hua
@khuasw
Our program builder itself is written in C++. It first scans what features the CPU supports and what kind of matrix is fed in; then it dumps those 0s and 1s into the memory and tells the CPU, “We got some fresh generated code here! Run it please!”.
2021-02-25 12:08:53
Kanru Hua
@khuasw
This method is called just-in-time (JIT) compilation, which is actually not something new. It’s the same technique used in modern web browsers to speed up Javascript execution. It is only until very recently did we see its application in machine learning.
2021-02-25 12:08:53
Kanru Hua
@khuasw
(For expert readers) You might have thought of FBGEMM. Yes, that’s the inspiration. Ours is like a GeMV/SpMV version of FBGEMM with support for older processors as well, also based on AsmJit.
2021-02-25 12:08:53