Contraction mapper for cuboid convolutions.
Directly pack rhs memory for the gebp kernes with a gemm_pack_rhs specialization. It's similar to optimized memory packing in eigen_spatial_convolutions. Works for: 1. CuboidConvolution 2. CuboidConvolutionBackwardInput ~2x-4x speedup when compiled with AVX (depends on tensor&patch dimensions). PiperOrigin-RevId: 212488060
Loading
Please sign in to comment