[MPlayer-dev-eng] [PATCH]: af_resample SIMD optimization (MMX/SSE2/3DNOW!/SSE/SSE3)

[Loren Merritt]

On Thu, 12 Mar 2009, Zhou Zongyi wrote:

> This patch implements SIMD optimized FIR filter in resample.

Filter weights should be aligned, and multiplied directly from memory.

pshufd is slow, pshuflw is preferred for the horizontal sum.

There's a division in the inner loop: wi=(wi+dn)%up

I would consider transposing the filter weights, eliminate most of the 
sample loads (use palignr to effectively keep them in a long 
shift-register instead), and eliminate the horizontal sums.
Failing that, do multiple horizontal sums at once to make full use of simd 
width.

--Loren Merritt