This does not adjust the AVX2 asm. The asm clips in many places to the
required range (16-bit signed) for performance reason. No mismatch
observed with coefs generated by the forward transform in checkasm in
10 thousand runs.