[x265] [PATCH] pixel-sse41.cpp: Modified PROCESS_SSE_SS4x1 macro with faster intrinsics

[chen]

At 2013-10-15 16:35:15,dnyaneshwar at multicorewareinc.com wrote:
># HG changeset patch
># User Dnyaneshwar Gorade <dnyaneshwar at multicorewareinc.com>
># Date 1381826069 -19800
>#      Tue Oct 15 14:04:29 2013 +0530
># Node ID 3cd533917aa110f7231abf6e0186e99b22dd4dcf
># Parent  1a85d8814346efdb984ea9eae24d1b06b973e9a8
>pixel-sse41.cpp: Modified PROCESS_SSE_SS4x1 macro with faster intrinsics.
>
>diff -r 1a85d8814346 -r 3cd533917aa1 source/common/vec/pixel-sse41.cpp
>--- a/source/common/vec/pixel-sse41.cpp Tue Oct 15 12:45:58 2013 +0530
>+++ b/source/common/vec/pixel-sse41.cpp Tue Oct 15 14:04:29 2013 +0530
>@@ -5331,10 +5331,8 @@
> #define PROCESS_SSE_SS4x1(BASE)\
>     m1 = _mm_loadu_si128((__m128i const*)(fenc + BASE)); \
>     n1 = _mm_loadu_si128((__m128i const*)(fref + BASE)); \
>-    sign1 = _mm_srai_epi16(m1, 15); \
>-    tmp1 = _mm_unpacklo_epi16(m1, sign1); \
>-    sign2 = _mm_srai_epi16(n1, 15); \
>-    tmp2 = _mm_unpacklo_epi16(n1, sign2); \
>+    tmp1= _mm_cvtepi16_epi32(m1); \
>+    tmp2= _mm_cvtepi16_epi32(n1); \
>     diff = _mm_sub_epi32(tmp1, tmp2); \
>     diff = _mm_mullo_epi32(diff, diff); \
>     sum = _mm_add_epi32(sum, diff)

two suggest:
1. be careful use SSE4 instruction with VS compiler, it have many bugs
2. are we have full of 16-bits dynamic range? if not, we may use instruction PMADDWD for more performance
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20131015/9b3dfc92/attachment.html>