VP8: clear DCT blocks in iDCT instead of using clear_blocks.

~0.3% faster overall.

Originally committed as revision 24448 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/vp8.c

@@ -835,8 +835,6 @@ static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb
     int nnz_pred, nnz, nnz_total = 0;
     int segment = s->segment;
 
-    s->dsp.clear_blocks((DCTELEM *)s->block);
-
     if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
         AV_ZERO128(dc);
         AV_ZERO128(dc+8);

libavcodec/vp8dsp.c

@@ -69,6 +69,10 @@ static void vp8_idct_add_c(uint8_t *dst, DCTELEM block[16], int stride)
         t1 = block[0*4+i] - block[2*4+i];
         t2 = MUL_35468(block[1*4+i]) - MUL_20091(block[3*4+i]);
         t3 = MUL_20091(block[1*4+i]) + MUL_35468(block[3*4+i]);
+        block[0*4+i] = 0;
+        block[1*4+i] = 0;
+        block[2*4+i] = 0;
+        block[3*4+i] = 0;
 
         tmp[i*4+0] = t0 + t3;
         tmp[i*4+1] = t1 + t2;
@@ -94,6 +98,7 @@ static void vp8_idct_dc_add_c(uint8_t *dst, DCTELEM block[16], int stride)
 {
     int i, dc = (block[0] + 4) >> 3;
     uint8_t *cm = ff_cropTbl + MAX_NEG_CROP + dc;
+    block[0] = 0;
 
     for (i = 0; i < 4; i++) {
         dst[0] = cm[dst[0]];

libavcodec/x86/vp8dsp-init.c

@@ -222,6 +222,7 @@ 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_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);
 
 #define DECLARE_LOOP_FILTER(NAME)\
 extern void ff_vp8_v_loop_filter_simple_ ## NAME(uint8_t *dst, int stride, int flim);\
@@ -328,6 +329,7 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
     }
 
     if (mm_flags & FF_MM_SSE) {
+        c->vp8_idct_add                         = ff_vp8_idct_add_sse;
         c->put_vp8_epel_pixels_tab[0][0][0]     =
         c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_sse;
     }

libavcodec/x86/vp8dsp.asm

@@ -913,6 +913,7 @@ cglobal vp8_idct_dc_add_mmx, 3, 3
     paddw      mm0, [pw_4]
     pxor       mm1, mm1
     psraw      mm0, 3
+    movd      [r1], mm1
     psubw      mm1, mm0
     packuswb   mm0, mm0
     packuswb   mm1, mm1
@@ -944,11 +945,12 @@ cglobal vp8_idct_dc_add_mmx, 3, 3
 cglobal vp8_idct_dc_add_sse4, 3, 3, 6
     ; load data
     movd       xmm0, [r1]
-    lea          r1, [r0+r2*2]
     pxor       xmm1, xmm1
 
     ; calculate DC
     paddw      xmm0, [pw_4]
+    movd       [r1], xmm1
+    lea          r1, [r0+r2*2]
     movd       xmm2, [r0]
     movd       xmm3, [r0+r2]
     movd       xmm4, [r1]
@@ -1005,14 +1007,26 @@ cglobal vp8_idct_dc_add_sse4, 3, 3, 6
 %endmacro
 
 INIT_MMX
-cglobal vp8_idct_add_mmx, 3, 3
+%macro VP8_IDCT_ADD 1
+cglobal vp8_idct_add_%1, 3, 3
     ; load block data
-    movq         m0, [r1]
-    movq         m1, [r1+8]
+    movq         m0, [r1+ 0]
+    movq         m1, [r1+ 8]
     movq         m2, [r1+16]
     movq         m3, [r1+24]
     movq         m6, [pw_20091]
     movq         m7, [pw_17734]
+%ifidn %1, sse
+    xorps      xmm0, xmm0
+    movaps  [r1+ 0], xmm0
+    movaps  [r1+16], xmm0
+%else
+    pxor         m4, m4
+    movq    [r1+ 0], m4
+    movq    [r1+ 8], m4
+    movq    [r1+16], m4
+    movq    [r1+24], m4
+%endif
 
     ; actual IDCT
     VP8_IDCT_TRANSFORM4x4_1D 0, 1, 2, 3, 4, 5
@@ -1028,6 +1042,10 @@ cglobal vp8_idct_add_mmx, 3, 3
     STORE_DIFFx2 m2, m3, m6, m7, m4, 3, r1, r2
 
     RET
+%endmacro
+
+VP8_IDCT_ADD mmx
+VP8_IDCT_ADD sse
 
 ;-----------------------------------------------------------------------------
 ; void vp8_luma_dc_wht_mmxext(DCTELEM block[4][4][16], DCTELEM dc[16])