neural-compressor icon indicating copy to clipboard operation
neural-compressor copied to clipboard

Detailed information about sparsity algorithm?

Open wenjingk-xilinx opened this issue 3 years ago • 1 comments

@airMeng Hi, can you provide some more detailed information about the sparsity method on this page ? what does the sparsity pattern mean? (ex, 2x1, 16x4 )

wenjingk-xilinx avatar Aug 25 '22 10:08 wenjingk-xilinx

Hi wenjingk, we prefer tile-wise sparsity in INC, which is balance between accuracy and performance. Tile-wise sparsity means to divide the whole matrix into tiles, if there are non-zero elements in the tile, the tile is called non-zero tiles. the 2x1, 16x4 corresponds to shapes of tile, like 2x1 means the tile is in 2x1shape, and the below image demonstrates so-called 4x1 image

For more details like why we choose 2x1, you can refer to some of my gist, not totally the same but similar. We will give more details soon, and feel free to raise any issues!

cc @ftian1 @XinyuYe-Intel

airMeng avatar Aug 29 '22 08:08 airMeng

Closed due to no further question.

chensuyue avatar Oct 25 '22 05:10 chensuyue