Skip to content

Latest commit

 

History

History

rgms

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
rgcn
rgcn
 
 
sparse_conv
sparse_conv
 
 
README.md
README.md
 
 

README.md

Relational Gather-Matmul-Scatter

This folder contains examples (RGCN and Sparse Convolution) of optimizing Relational Gather-Matmul-Scatter(RGMS) with SparseTIR.

rgms
├── rgcn
|     bench_rgcn_baseline.py            # rgcn implementations in DGL/PyG
|     bench_rgcn_non_composable.py      # fuse gather-matmul-scatter in SparseTIR, without using composable format/tensor cores.
|     bench_rgcn_composable.py          # fuse gather-matmul-scatter in SparseTIR, using composable format, no tensor cores.
|     bench_rgcn_tensorcore.py          # fuse gather-matmul-scatter in SparseTIR, using composable format and tensor cores.
├── sparse_conv
|     sparse_conv.py                    # apply SparseTIR's rgms operator to Sparse Convolution.

We haven't enabled software pipelining and async copy, nor any optimizations to avoid bank conflicts yet, and the performance could be further improved.