VP8: optimize DC-only chroma case in the same way as luma.

Add MMX idct_dc_add4uv function for this case.
~40% faster chroma idct.

Originally committed as revision 24455 to svn://svn.ffmpeg.org/ffmpeg/trunk

2 files changed, 53 insertions, 10 deletions

diff --git a/libavcodec/x86/vp8dsp-init.c b/libavcodec/x86/vp8dsp-init.c
index 5da70824fc..4f40bdd11f 100644
--- a/libavcodec/x86/vp8dsp-init.c
+++ b/libavcodec/x86/vp8dsp-init.c
@@ -220,8 +220,9 @@ HVBILIN(ssse3,  8, 16, 16)
 
 extern void ff_vp8_idct_dc_add_mmx(uint8_t *dst, DCTELEM block[16], int stride);
 extern void ff_vp8_idct_dc_add_sse4(uint8_t *dst, DCTELEM block[16], int stride);
-extern void ff_vp8_idct_dc_add4_mmx(uint8_t *dst, DCTELEM block[4][16], int stride);
-extern void ff_vp8_idct_dc_add4_sse2(uint8_t *dst, DCTELEM block[4][16], int stride);
+extern void ff_vp8_idct_dc_add4y_mmx(uint8_t *dst, DCTELEM block[4][16], int stride);
+extern void ff_vp8_idct_dc_add4y_sse2(uint8_t *dst, DCTELEM block[4][16], int stride);
+extern void ff_vp8_idct_dc_add4uv_mmx(uint8_t *dst, DCTELEM block[2][16], int stride);
 extern void ff_vp8_luma_dc_wht_mmx(DCTELEM block[4][4][16], DCTELEM dc[16]);
 extern void ff_vp8_idct_add_mmx(uint8_t *dst, DCTELEM block[16], int stride);
 extern void ff_vp8_idct_add_sse(uint8_t *dst, DCTELEM block[16], int stride);
@@ -284,10 +285,11 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
 
 #if HAVE_YASM
     if (mm_flags & FF_MM_MMX) {
-        c->vp8_idct_dc_add                  = ff_vp8_idct_dc_add_mmx;
-        c->vp8_idct_dc_add4                 = ff_vp8_idct_dc_add4_mmx;
-        c->vp8_idct_add                     = ff_vp8_idct_add_mmx;
-        c->vp8_luma_dc_wht                  = ff_vp8_luma_dc_wht_mmx;
+        c->vp8_idct_dc_add    = ff_vp8_idct_dc_add_mmx;
+        c->vp8_idct_dc_add4y  = ff_vp8_idct_dc_add4y_mmx;
+        c->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmx;
+        c->vp8_idct_add       = ff_vp8_idct_add_mmx;
+        c->vp8_luma_dc_wht    = ff_vp8_luma_dc_wht_mmx;
         c->put_vp8_epel_pixels_tab[0][0][0]     =
         c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_mmx;
         c->put_vp8_epel_pixels_tab[1][0][0]     =
@@ -354,7 +356,7 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
     }
 
     if (mm_flags & FF_MM_SSE2) {
-        c->vp8_idct_dc_add4           = ff_vp8_idct_dc_add4_sse2;
+        c->vp8_idct_dc_add4y          = ff_vp8_idct_dc_add4y_sse2;
 
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_sse2;
         c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_sse2;
diff --git a/libavcodec/x86/vp8dsp.asm b/libavcodec/x86/vp8dsp.asm
index 70fe9d3bfe..4aa901e27f 100644
--- a/libavcodec/x86/vp8dsp.asm
+++ b/libavcodec/x86/vp8dsp.asm
@@ -976,11 +976,11 @@ cglobal vp8_idct_dc_add_sse4, 3, 3, 6
     RET
 
 ;-----------------------------------------------------------------------------
-; void vp8_idct_dc_add4_<opt>(uint8_t *dst, DCTELEM block[4][16], int stride);
+; void vp8_idct_dc_add4y_<opt>(uint8_t *dst, DCTELEM block[4][16], int stride);
 ;-----------------------------------------------------------------------------
 
 INIT_MMX
-cglobal vp8_idct_dc_add4_mmx, 3, 3
+cglobal vp8_idct_dc_add4y_mmx, 3, 3
     ; load data
     movd      m0, [r1+32*0] ; A
     movd      m1, [r1+32*2] ; C
@@ -1015,7 +1015,7 @@ cglobal vp8_idct_dc_add4_mmx, 3, 3
     RET
 
 INIT_XMM
-cglobal vp8_idct_dc_add4_sse2, 3, 3
+cglobal vp8_idct_dc_add4y_sse2, 3, 3, 6
     ; load data
     movd      m0, [r1+32*0] ; A
     movd      m1, [r1+32*2] ; C
@@ -1045,6 +1045,47 @@ cglobal vp8_idct_dc_add4_sse2, 3, 3
     RET
 
 ;-----------------------------------------------------------------------------
+; void vp8_idct_dc_add4uv_<opt>(uint8_t *dst, DCTELEM block[4][16], int stride);
+;-----------------------------------------------------------------------------
+
+INIT_MMX
+cglobal vp8_idct_dc_add4uv_mmx, 3, 3
+    ; load data
+    movd      m0, [r1+32*0] ; A
+    movd      m1, [r1+32*2] ; C
+    punpcklwd m0, [r1+32*1] ; A B
+    punpcklwd m1, [r1+32*3] ; C D
+    punpckldq m0, m1        ; A B C D
+    pxor      m6, m6
+
+    ; calculate DC
+    paddw     m0, [pw_4]
+    movd [r1+32*0], m6
+    movd [r1+32*1], m6
+    movd [r1+32*2], m6
+    movd [r1+32*3], m6
+    psraw     m0, 3
+    psubw     m6, m0
+    packuswb  m0, m0
+    packuswb  m6, m6
+    punpcklbw m0, m0 ; AABBCCDD
+    punpcklbw m6, m6 ; AABBCCDD
+    movq      m1, m0
+    movq      m7, m6
+    punpcklbw m0, m0 ; AAAABBBB
+    punpckhbw m1, m1 ; CCCCDDDD
+    punpcklbw m6, m6 ; AAAABBBB
+    punpckhbw m7, m7 ; CCCCDDDD
+
+    ; add DC
+    lea       r1, [r0+r2*2]
+    ADD_DC    m0, m6, 0, mova
+    lea       r0, [r0+r2*4]
+    lea       r1, [r1+r2*4]
+    ADD_DC    m1, m7, 0, mova
+    RET
+
+;-----------------------------------------------------------------------------
 ; void vp8_idct_add_<opt>(uint8_t *dst, DCTELEM block[16], int stride);
 ;-----------------------------------------------------------------------------