openvino icon indicating copy to clipboard operation
openvino copied to clipboard

Yi3/fix perf wenet

Open zhangYiIntel opened this issue 2 years ago • 1 comments

Details:

  • Improve performance for dynamic shape model such as Wenet
  • Implement Concat kernel to avoid frequent oneDNN kernel construction
  • Implement shape_infer for Cocant & Transpose
  • Use shape_infer for Eltwise rather than self-implementation
  • Simplify MatMul key formulation

Tickets:

  • 86353

zhangYiIntel avatar Sep 13 '22 05:09 zhangYiIntel

@a-sidorova , could you please review?

maxnick avatar Sep 19 '22 07:09 maxnick

@a-sidorova, could you please take a look once again?

maxnick avatar Oct 25 '22 12:10 maxnick

Hi @a-sidorova I think the trial for matmul key improvement is not mature enough and decide to drop that part so that we could focus on the implementation of concat kernel.

zhangYiIntel avatar Oct 28 '22 08:10 zhangYiIntel

@zhangYiIntel also I'd like to suggest you to rebase the branch to the master

a-sidorova avatar Nov 01 '22 14:11 a-sidorova

@dmitry-gorokhov Could you help review this PR finally ?

zhangYiIntel avatar Nov 02 '22 06:11 zhangYiIntel

@maxnick Could you help review again ?

zhangYiIntel avatar Nov 09 '22 08:11 zhangYiIntel