GitLab

9ms --> 1ms on waymo open dataset

unroll loop for common params

2. working on batch indice conv

1. use dispatch instead of template
2. set use_hash default to False

[for torch 1.4]
torch::jit::RegisterOperators --> torch::RegisterOperators

ref) https://github.com/pytorch/pytorch/commit/243298668caca7ee7de454ebc4d99c487ec9b272

1. workaround for gcc 5.5 avx512 problem
2. workaround for cuda 9 invalid device function problem

1. add cuda hash support for cuda indice generation.
2. use hash table instead of dense table in CPU code.
3. add CPU-only build support.

2. update install command