From dc5eec80851418f7165257224576518f19a2aabb Mon Sep 17 00:00:00 2001
From: "Ronald S. Bultje" <rsbultje@gmail.com>
Date: Thu, 22 Jul 2010 19:59:34 +0000
Subject: [PATCH] Use pextrw for SSE4 mbedge filter result writing, speedup
 5-10cycles on CPUs supporting it.

Originally committed as revision 24437 to svn://svn.ffmpeg.org/ffmpeg/trunk
---
 libavcodec/x86/vp8dsp-init.c |  4 ++++
 libavcodec/x86/vp8dsp.asm    | 35 ++++++++++++++++++++++++++++++-----
 2 files changed, 34 insertions(+), 5 deletions(-)

diff --git a/libavcodec/x86/vp8dsp-init.c b/libavcodec/x86/vp8dsp-init.c
index 66ae884705..fad399fba8 100644
--- a/libavcodec/x86/vp8dsp-init.c
+++ b/libavcodec/x86/vp8dsp-init.c
@@ -247,6 +247,7 @@ DECLARE_LOOP_FILTER(mmx)
 DECLARE_LOOP_FILTER(mmxext)
 DECLARE_LOOP_FILTER(sse2)
 DECLARE_LOOP_FILTER(ssse3)
+DECLARE_LOOP_FILTER(sse4)
 
 #endif
 
@@ -379,6 +380,9 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
 
     if (mm_flags & FF_MM_SSE4) {
         c->vp8_idct_dc_add                  = ff_vp8_idct_dc_add_sse4;
+
+        c->vp8_h_loop_filter16y       = ff_vp8_h_loop_filter16y_mbedge_sse4;
+        c->vp8_h_loop_filter8uv       = ff_vp8_h_loop_filter8uv_mbedge_sse4;
     }
 #endif
 }
diff --git a/libavcodec/x86/vp8dsp.asm b/libavcodec/x86/vp8dsp.asm
index c9d7d383fb..843873167d 100644
--- a/libavcodec/x86/vp8dsp.asm
+++ b/libavcodec/x86/vp8dsp.asm
@@ -1932,10 +1932,24 @@ INNER_LOOPFILTER ssse3,  h, 6,  8, 13
 
 ; write 4 or 8 words in the mmx/xmm registers as 8 lines
 ; 1 and 2 are the registers to write, this can be the same (for SSE2)
+; for pre-SSE4:
 ; 3 is a general-purpose register that we will clobber
+; for SSE4:
+; 3 is a pointer to the destination's 5th line
 ; 4 is a pointer to the destination's 4th line
-; 5 is -stride and +stride
-%macro WRITE_8W 6
+; 5/6 is -stride and +stride
+; 7 is optimization string
+%macro WRITE_8W 7
+%ifidn %7, sse4
+    pextrw    [%4+%5*4], %1, 0
+    pextrw    [%3+%5*4], %1, 1
+    pextrw    [%4+%5*2], %1, 2
+    pextrw    [%4+%5  ], %1, 3
+    pextrw    [%4     ], %1, 4
+    pextrw    [%3     ], %1, 5
+    pextrw    [%3+%6  ], %1, 6
+    pextrw    [%3+%6*2], %1, 7
+%else
     movd             %3, %1
 %if mmsize == 8
     punpckhdq        %1, %1
@@ -1974,6 +1988,7 @@ INNER_LOOPFILTER ssse3,  h, 6,  8, 13
 %if mmsize == 8
     add              %4, %5
 %endif
+%endif
 %endmacro
 
 %macro MBEDGE_LOOPFILTER 5
@@ -2509,14 +2524,17 @@ cglobal vp8_%2_loop_filter16y_mbedge_%1, 5, %3, %5
 %if mmsize == 8 ; mmx/mmxext (h)
     WRITE_4x2D        1, 2, 3, 4, dst_reg, dst2_reg, mstride_reg, stride_reg
     add         dst_reg, 4
-    WRITE_8W         m5, m6, dst2_reg, dst_reg, mstride_reg, stride_reg
+    WRITE_8W         m5, m6, dst2_reg, dst_reg, mstride_reg, stride_reg, %4
 %else ; sse2 (h)
     lea        dst8_reg, [dst8_reg+mstride_reg+1]
     WRITE_4x4D        1, 2, 3, 4, dst_reg, dst2_reg, dst8_reg, mstride_reg, stride_reg, %4
     lea         dst_reg, [dst2_reg+mstride_reg+4]
     lea        dst8_reg, [dst8_reg+mstride_reg+4]
-    WRITE_8W         m5, m5, dst2_reg, dst_reg,  mstride_reg, stride_reg
-    WRITE_8W         m6, m6, dst2_reg, dst8_reg, mstride_reg, stride_reg
+    WRITE_8W         m5, m5, dst2_reg, dst_reg,  mstride_reg, stride_reg, %2
+%ifidn %2, sse4
+    lea         dst_reg, [dst8_reg+ stride_reg]
+%endif
+    WRITE_8W         m6, m6, dst2_reg, dst8_reg, mstride_reg, stride_reg, %2
 %endif
 %endif
 
@@ -2574,3 +2592,10 @@ MBEDGE_LOOPFILTER ssse3,  h, 6, 16, 16
 %endif
 MBEDGE_LOOPFILTER ssse3,  v, 6,  8, 16
 MBEDGE_LOOPFILTER ssse3,  h, 6,  8, 16
+
+%ifdef m8
+MBEDGE_LOOPFILTER sse4,   h, 5, 16, 16
+%else
+MBEDGE_LOOPFILTER sse4,   h, 6, 16, 16
+%endif
+MBEDGE_LOOPFILTER sse4,   h, 6,  8, 16