Hello Qualcomm,
In the SNPE conversion documentation, matrix multiplication is still limited to fully connected layers where the second matrix in static. However , in the filed of NN's there is a lot of papers that utilize matrix multiplication as a novel approach to not only increase receptive fields, but to also also apply mathematical forumations to models at runtime. I think it would be extremely beneficial to the community if we can have a optmized MatMul operation for snpe, so that these state of the art networks can be run on device.
Thanks,
Edward Li