Closed FrancescoConti closed 3 years ago
Added sources and headers of standard convolution and matmul kernels which are named pulp_nn_conv_Co_parallel and pulp_nn_matmul_Co_parallel respectively at the 8bit library.
Super-fast! I'll keep the issue open till we test it a bit in DORY...
In some corner cases, it could be more advantageous to parallelize convolutions over the Kout dimension instead of the Hout/Wout remainder (like in depthwise conv's). Potential advantages:
Potential disadvantages:
@NBruschi @anga93 what do you think?