search

TiledTensor / TiledCUDA

TiledCUDA is a highly efficient kernel template library designed to elevate CUDA C’s level of abstraction for processing tiles.
MIT License
157 stars 10 forks source link

refactor(cell): refactor the register to shared storer. #137

Closed haruhi55 closed 2 months ago

haruhi55 commented 2 months ago

This PR fixes two issues:

  1. To support using swizzled shared memory layout in storing tensor core's register tile, improved implementation is required to use layout for computing pointer offsets inside a BaseTile instead of manually computing offsets.
  2. Add a straightforward implementation to use vectorized instructions for accessing shared memory.