Help Needed with Understanding SIMD Filter Processing and GEMM in Convolution Implementation

[XinyiLiu577086410]

Hi,

I am a sophomore student participating in a competition on convolution and have been greatly inspired by your impressive work. I've learned a lot from it, but I'm having trouble understanding a few parts. Could you please help me?

The filter is packed and processed with SIMD over output channels. What about over input channels? Would this approach make it easier to perform the Hadamard product and then reduce?
I am struggling to understand the GEMM part. Can this be simply replaced with a standard SGEMM from BLAS, or are there specific tricks involved?

Thank you so much for your time and assistance!