ref: ffd221366067ceda66e68d74778bf1b2d077188a
parent: 6214038be94fe8a5648a9588ca2985650541fe2e
parent: 00fc0e3ff5e34472efc652686703493387fe36ee
author: Jingning Han <[email protected]>
date: Mon Jul 7 08:07:09 EDT 2014
Merge "Tune SSSE3 implementation of fast path quantization"
--- a/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
+++ b/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -264,7 +264,6 @@
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
pcmpeqw m7, m7
- pcmpeqw m12, m12
paddsw m6, m1 ; m6 += round
punpckhqdq m1, m1
@@ -296,7 +295,7 @@
mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
psubw m6, m7 ; m6 = scan[i] + 1
- psubw m11, m12 ; m11 = scan[i] + 1
+ psubw m11, m7 ; m11 = scan[i] + 1
pandn m8, m6 ; m8 = max(eob)
pandn m13, m11 ; m13 = max(eob)
pmaxsw m8, m13
@@ -309,10 +308,9 @@
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
pcmpeqw m7, m7
- pcmpeqw m12, m12
%ifidn %1, b_32x32
pmovmskb r6, m7
- pmovmskb r2, m12
+ pmovmskb r2, m7
or r6, r2
jz .skip_iter
%endif
@@ -343,7 +341,7 @@
mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
psubw m6, m7 ; m6 = scan[i] + 1
- psubw m11, m12 ; m11 = scan[i] + 1
+ psubw m11, m7 ; m11 = scan[i] + 1
pandn m14, m6 ; m14 = max(eob)
pandn m13, m11 ; m13 = max(eob)
pmaxsw m8, m14