x86: sbrdsp: implement SSE2 hf_apply_noise
authorChristophe Gisquet <christophe.gisquet@gmail.com>
Tue, 9 Apr 2013 20:16:36 +0000 (22:16 +0200)
committerMichael Niedermayer <michaelni@gmx.at>
Fri, 19 Apr 2013 11:19:45 +0000 (13:19 +0200)
commit76c7277385120741914168d02b57a7a1acf87a9b
tree1a0821b6a1b2aae976515ed3423a05ac9cd245f5
parent380cfce2b2138a0513f7c054134458b3b4c92fd4
x86: sbrdsp: implement SSE2 hf_apply_noise

233 to 105 cycles on Arrandale and Win64.
Replacing the multiplication by s_m[m] by a pand and a pxor with
appropriate vectors is slower. Unrolling is a 15 cycles win.
A SSE version was 4 cycles slower.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
libavcodec/aacsbrdata.h
libavcodec/x86/sbrdsp.asm
libavcodec/x86/sbrdsp_init.c