For example in AVX2 implementation of mul_vec_by_scalar_then_add_into function, we can optimize the throughput by issuing more instructions per iteration of the loop. We just have to manually unroll the loop to get it working.

This is possible because AVX2 or AVX512 instructions may have higher latency, but they allow issuing multiple of them. When pipelined nicely, it can give us quite good performance boost.

So explore that and see what we can achieve.