Luwei Zhou
Luwei Zhou
> > Deconv only has one ref entry(CPU_INSTANCE(ref_deconvolution_fwd_t) for JIT,gemm and c++ reference implement. So one layout can't use RTinfo to filter implement. Update test cases with 1x1 related or...
@maxnick , Applied review comments on test cases. pls help to review. Thx!
> > Change stride to 1 unit because of brgemm limitation. > > What exactly happens if the stride is other than 1? stride 1 would cause some cases fall...
@maxnick , Applied review comments.
https://jira.devtools.intel.com/browse/CVS-140606 to track the acc issue on 1x1 kernel with details.
> > The gemm conv kernel diverges most with the upstream version, because fork enabled jit_pp kernel for jit gemm kernel but upstream still use ref_pp kernel. > > https://github.com/openvinotoolkit/oneDNN/blob/26633ae49edd4353a29b7170d9fcef6b2d79f4b3/src/cpu/gemm_convolution.hpp#L77...
@maxnick , Applied review comment. Pls help to review. Thx!
@sitabulaixizawaluduo , Thanks! I think you are doing OP level perf test on Split. Spitting the [256,279,81] input along "279" dimension to 96 outputs with planar . 1. About the...
@yuxu42 @mg-intel , I don't think this is a bug. It is just some internal supported descriptor behavior changed inside CPU plugin split node. Thanks!