Increase the unrolling depth for AVX+
Hey all,
As requested in !111 (comment 2220110)
I'm filing this issue to analyse how best to increase the unroll_shift
parameter for the
AVX backend.
This is because the tests added in !136 (merged) do not tolerate a higher unrolling value (they
fail register assignment), which is also the state of orc_quantdequant3_s16
in 32-bit
AVX. This latter example also fails to build with the C backend altogether under the
same circumstances.