Correctly compute loop bounds in pack_rhs_mem for convolutions.
(1) Correctly compute maxCols/maxRows/maxPlanes for the fast path. (2) Use fast divisors in a hot path. ~8-20% improvement for cuboid convolutions (Conv3D). ~1-5% improvement for spatial convolutions (Conv2D). PiperOrigin-RevId: 214061787
Loading
Please sign in to comment