rv40dsp: implement prescaled versions for biweight.

Quite often, the original weights are multiple of 512. By prescaling them
by 1/512 when they are computed (once per frame), no intermediate shifting
is needed, and no prescaling on each call either.

The x86 code already used that trick.

Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>

7 files changed, 112 insertions, 79 deletions

diff --git a/libavcodec/arm/rv40dsp_init_neon.c b/libavcodec/arm/rv40dsp_init_neon.c
index 650ef61878..2ce50a2073 100644
--- a/libavcodec/arm/rv40dsp_init_neon.c
+++ b/libavcodec/arm/rv40dsp_init_neon.c
@@ -128,8 +128,8 @@ void ff_rv40dsp_init_neon(RV34DSPContext *c, DSPContext* dsp)
     c->avg_chroma_pixels_tab[0] = ff_avg_rv40_chroma_mc8_neon;
     c->avg_chroma_pixels_tab[1] = ff_avg_rv40_chroma_mc4_neon;
 
-    c->rv40_weight_pixels_tab[0] = ff_rv40_weight_func_16_neon;
-    c->rv40_weight_pixels_tab[1] = ff_rv40_weight_func_8_neon;
+    c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_16_neon;
+    c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_8_neon;
 
     c->rv40_loop_filter_strength[0] = ff_rv40_h_loop_filter_strength_neon;
     c->rv40_loop_filter_strength[1] = ff_rv40_v_loop_filter_strength_neon;
diff --git a/libavcodec/rv34.c b/libavcodec/rv34.c
index 3ad1717d13..12475692c6 100644
--- a/libavcodec/rv34.c
+++ b/libavcodec/rv34.c
@@ -521,7 +521,7 @@ static void rv34_pred_mv(RV34DecContext *r, int block_type, int subblock_no, int
  */
 static int calc_add_mv(RV34DecContext *r, int dir, int val)
 {
-    int mul = dir ? -r->weight2 : r->weight1;
+    int mul = dir ? -r->mv_weight2 : r->mv_weight1;
 
     return (val * mul + 0x2000) >> 14;
 }
@@ -776,24 +776,24 @@ static void rv34_mc_1mv(RV34DecContext *r, const int block_type,
 
 static void rv4_weight(RV34DecContext *r)
 {
-    r->rdsp.rv40_weight_pixels_tab[0](r->s.dest[0],
-                                      r->tmp_b_block_y[0],
-                                      r->tmp_b_block_y[1],
-                                      r->weight1,
-                                      r->weight2,
-                                      r->s.linesize);
-    r->rdsp.rv40_weight_pixels_tab[1](r->s.dest[1],
-                                      r->tmp_b_block_uv[0],
-                                      r->tmp_b_block_uv[2],
-                                      r->weight1,
-                                      r->weight2,
-                                      r->s.uvlinesize);
-    r->rdsp.rv40_weight_pixels_tab[1](r->s.dest[2],
-                                      r->tmp_b_block_uv[1],
-                                      r->tmp_b_block_uv[3],
-                                      r->weight1,
-                                      r->weight2,
-                                      r->s.uvlinesize);
+    r->rdsp.rv40_weight_pixels_tab[r->scaled_weight][0](r->s.dest[0],
+                                                        r->tmp_b_block_y[0],
+                                                        r->tmp_b_block_y[1],
+                                                        r->weight1,
+                                                        r->weight2,
+                                                        r->s.linesize);
+    r->rdsp.rv40_weight_pixels_tab[r->scaled_weight][1](r->s.dest[1],
+                                                        r->tmp_b_block_uv[0],
+                                                        r->tmp_b_block_uv[2],
+                                                        r->weight1,
+                                                        r->weight2,
+                                                        r->s.uvlinesize);
+    r->rdsp.rv40_weight_pixels_tab[r->scaled_weight][1](r->s.dest[2],
+                                                        r->tmp_b_block_uv[1],
+                                                        r->tmp_b_block_uv[3],
+                                                        r->weight1,
+                                                        r->weight2,
+                                                        r->s.uvlinesize);
 }
 
 static void rv34_mc_2mv(RV34DecContext *r, const int block_type)
@@ -1703,11 +1703,21 @@ int ff_rv34_decode_frame(AVCodecContext *avctx,
             int dist0   = GET_PTS_DIFF(r->cur_pts,  r->last_pts);
             int dist1   = GET_PTS_DIFF(r->next_pts, r->cur_pts);
 
-            if (!refdist) {
-                r->weight1 = r->weight2 = 8192;
-            } else {
-                r->weight1 = (dist0 << 14) / refdist;
-                r->weight2 = (dist1 << 14) / refdist;
+            if(!refdist){
+                r->mv_weight1 = r->mv_weight2 = r->weight1 = r->weight2 = 8192;
+                r->scaled_weight = 0;
+            }else{
+                r->mv_weight1 = (dist0 << 14) / refdist;
+                r->mv_weight2 = (dist1 << 14) / refdist;
+                if((r->mv_weight1|r->mv_weight2) & 511){
+                    r->weight1 = r->mv_weight1;
+                    r->weight2 = r->mv_weight2;
+                    r->scaled_weight = 0;
+                }else{
+                    r->weight1 = r->mv_weight1 >> 9;
+                    r->weight2 = r->mv_weight2 >> 9;
+                    r->scaled_weight = 1;
+                }
             }
         }
         s->mb_x = s->mb_y = 0;
diff --git a/libavcodec/rv34.h b/libavcodec/rv34.h
index 76232145c5..e7a59c4bed 100644
--- a/libavcodec/rv34.h
+++ b/libavcodec/rv34.h
@@ -106,7 +106,9 @@ typedef struct RV34DecContext{
     int rpr;                 ///< one field size in RV30 slice header
 
     int cur_pts, last_pts, next_pts;
+    int scaled_weight;
     int weight1, weight2;    ///< B frame distance fractions (0.14) used in motion compensation
+    int mv_weight1, mv_weight2;
 
     uint16_t *cbp_luma;      ///< CBP values for luma subblocks
     uint8_t  *cbp_chroma;    ///< CBP values for chroma subblocks
diff --git a/libavcodec/rv34dsp.h b/libavcodec/rv34dsp.h
index c70194cc20..58da59f038 100644
--- a/libavcodec/rv34dsp.h
+++ b/libavcodec/rv34dsp.h
@@ -58,7 +58,12 @@ typedef struct RV34DSPContext {
     qpel_mc_func avg_pixels_tab[4][16];
     h264_chroma_mc_func put_chroma_pixels_tab[3];
     h264_chroma_mc_func avg_chroma_pixels_tab[3];
-    rv40_weight_func rv40_weight_pixels_tab[2];
+    /**
+     * Biweight functions, first dimension is transform size (16/8),
+     * second is whether the weight is prescaled by 1/512 to skip
+     * the intermediate shifting.
+     */
+    rv40_weight_func rv40_weight_pixels_tab[2][2];
     rv34_inv_transform_func rv34_inv_transform;
     rv34_inv_transform_func rv34_inv_transform_dc;
     rv34_idct_add_func rv34_idct_add;
diff --git a/libavcodec/rv40dsp.c b/libavcodec/rv40dsp.c
index c12958a89c..19a18d37a5 100644
--- a/libavcodec/rv40dsp.c
+++ b/libavcodec/rv40dsp.c
@@ -278,7 +278,7 @@ RV40_CHROMA_MC(put_, op_put)
 RV40_CHROMA_MC(avg_, op_avg)
 
 #define RV40_WEIGHT_FUNC(size) \
-static void rv40_weight_func_ ## size (uint8_t *dst, uint8_t *src1, uint8_t *src2, int w1, int w2, ptrdiff_t stride)\
+static void rv40_weight_func_rnd_ ## size (uint8_t *dst, uint8_t *src1, uint8_t *src2, int w1, int w2, ptrdiff_t stride)\
 {\
     int i, j;\
 \
@@ -289,6 +289,18 @@ static void rv40_weight_func_ ## size (uint8_t *dst, uint8_t *src1, uint8_t *src
         src2 += stride;\
         dst  += stride;\
     }\
+}\
+static void rv40_weight_func_nornd_ ## size (uint8_t *dst, uint8_t *src1, uint8_t *src2, int w1, int w2, ptrdiff_t stride)\
+{\
+    int i, j;\
+\
+    for (j = 0; j < size; j++) {\
+        for (i = 0; i < size; i++)\
+            dst[i] = (w2 * src1[i] + w1 * src2[i] + 0x10) >> 5;\
+        src1 += stride;\
+        src2 += stride;\
+        dst  += stride;\
+    }\
 }
 
 RV40_WEIGHT_FUNC(16)
@@ -578,8 +590,10 @@ av_cold void ff_rv40dsp_init(RV34DSPContext *c, DSPContext* dsp) {
     c->avg_chroma_pixels_tab[0] = avg_rv40_chroma_mc8_c;
     c->avg_chroma_pixels_tab[1] = avg_rv40_chroma_mc4_c;
 
-    c->rv40_weight_pixels_tab[0] = rv40_weight_func_16;
-    c->rv40_weight_pixels_tab[1] = rv40_weight_func_8;
+    c->rv40_weight_pixels_tab[0][0] = rv40_weight_func_rnd_16;
+    c->rv40_weight_pixels_tab[0][1] = rv40_weight_func_rnd_8;
+    c->rv40_weight_pixels_tab[1][0] = rv40_weight_func_nornd_16;
+    c->rv40_weight_pixels_tab[1][1] = rv40_weight_func_nornd_8;
 
     c->rv40_weak_loop_filter[0]     = rv40_h_weak_loop_filter;
     c->rv40_weak_loop_filter[1]     = rv40_v_weak_loop_filter;
diff --git a/libavcodec/x86/rv40dsp.asm b/libavcodec/x86/rv40dsp.asm
index bff3e7b96a..9028e74024 100644
--- a/libavcodec/x86/rv40dsp.asm
+++ b/libavcodec/x86/rv40dsp.asm
@@ -139,69 +139,61 @@ SECTION .text
 
 ; rv40_weight_func_%1(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w1, int w2, int stride)
 ; %1=size  %2=num of xmm regs
-%macro RV40_WEIGHT  2
-cglobal rv40_weight_func_%1, 6, 7, %2
+; The weights are FP0.14 notation of fractions depending on pts.
+; For timebases without rounding error (i.e. PAL), the fractions
+; can be simplified, and several operations can be avoided.
+; Therefore, we check here whether they are multiples of 2^9 for
+; those simplifications to occur.
+%macro RV40_WEIGHT  3
+cglobal rv40_weight_func_%1_%2, 6, 7, %3
 %if cpuflag(ssse3)
     mova       m1, [shift_round]
 %else
     mova       m1, [pw_16]
 %endif
     pxor       m0, m0
-    mov        r6, r3
-    or         r6, r4
-    ; The weights are FP0.14 notation of fractions depending on pts.
-    ; For timebases without rounding error (i.e. PAL), the fractions
-    ; can be simplified, and several operations can be avoided.
-    ; Therefore, we check here whether they are multiples of 2^9 for
-    ; those simplifications to occur.
-    and        r6, 0x1FF
     ; Set loop counter and increments
 %if mmsize == 8
-    mov        r6, %1
+    mov        r6, %2
 %else
-    mov        r6, (%1 * %1) / mmsize
+    mov        r6, (%2 * %2) / mmsize
 %endif
 
-    ; Use result of test now
-    jz .loop_512
     movd       m2, r3
     movd       m3, r4
+%ifidn %1,rnd
+%define  RND   0
     SPLATW     m2, m2
-    SPLATW     m3, m3
-
-.loop:
-    MAIN_LOOP  %1, 0
-    jnz        .loop
-    REP_RET
-
-    ; Weights are multiple of 512, which allows some shortcuts
-.loop_512:
-    sar        r3, 9
-    sar        r4, 9
-    movd       m2, r3
-    movd       m3, r4
+%else
+%define  RND   1
 %if cpuflag(ssse3)
     punpcklbw  m3, m2
-    SPLATW     m3, m3
 %else
     SPLATW     m2, m2
-    SPLATW     m3, m3
 %endif
-.loop2:
-    MAIN_LOOP  %1, 1
-    jnz        .loop2
-    REP_RET
+%endif
+    SPLATW     m3, m3
 
+.loop:
+    MAIN_LOOP  %2, RND
+    jnz        .loop
+    REP_RET
 %endmacro
 
 INIT_MMX mmx
-RV40_WEIGHT    8, 0
-RV40_WEIGHT   16, 0
+RV40_WEIGHT   rnd,    8, 3
+RV40_WEIGHT   rnd,   16, 4
+RV40_WEIGHT   nornd,  8, 3
+RV40_WEIGHT   nornd, 16, 4
 
 INIT_XMM sse2
-RV40_WEIGHT    8, 8
-RV40_WEIGHT   16, 8
+RV40_WEIGHT   rnd,    8, 3
+RV40_WEIGHT   rnd,   16, 4
+RV40_WEIGHT   nornd,  8, 3
+RV40_WEIGHT   nornd, 16, 4
 
 INIT_XMM ssse3
-RV40_WEIGHT    8, 8
-RV40_WEIGHT   16, 8
+RV40_WEIGHT   rnd,    8, 3
+RV40_WEIGHT   rnd,   16, 4
+RV40_WEIGHT   nornd,  8, 3
+RV40_WEIGHT   nornd, 16, 4
diff --git a/libavcodec/x86/rv40dsp_init.c b/libavcodec/x86/rv40dsp_init.c
index 79c70f78c3..df468aa9e5 100644
--- a/libavcodec/x86/rv40dsp_init.c
+++ b/libavcodec/x86/rv40dsp_init.c
@@ -41,10 +41,14 @@ void ff_avg_rv40_chroma_mc4_3dnow(uint8_t *dst, uint8_t *src,
                                   int stride, int h, int x, int y);
 
 #define DECLARE_WEIGHT(opt) \
-void ff_rv40_weight_func_16_##opt(uint8_t *dst, uint8_t *src1, uint8_t *src2, \
-                                  int w1, int w2, ptrdiff_t stride); \
-void ff_rv40_weight_func_8_##opt (uint8_t *dst, uint8_t *src1, uint8_t *src2, \
-                                  int w1, int w2, ptrdiff_t stride);
+void ff_rv40_weight_func_rnd_16_##opt(uint8_t *dst, uint8_t *src1, uint8_t *src2, \
+                                      int w1, int w2, ptrdiff_t stride); \
+void ff_rv40_weight_func_rnd_8_##opt (uint8_t *dst, uint8_t *src1, uint8_t *src2, \
+                                      int w1, int w2, ptrdiff_t stride); \
+void ff_rv40_weight_func_nornd_16_##opt(uint8_t *dst, uint8_t *src1, uint8_t *src2, \
+                                        int w1, int w2, ptrdiff_t stride); \
+void ff_rv40_weight_func_nornd_8_##opt (uint8_t *dst, uint8_t *src1, uint8_t *src2, \
+                                        int w1, int w2, ptrdiff_t stride);
 DECLARE_WEIGHT(mmx)
 DECLARE_WEIGHT(sse2)
 DECLARE_WEIGHT(ssse3)
@@ -57,8 +61,10 @@ void ff_rv40dsp_init_x86(RV34DSPContext *c, DSPContext *dsp)
     if (mm_flags & AV_CPU_FLAG_MMX) {
         c->put_chroma_pixels_tab[0] = ff_put_rv40_chroma_mc8_mmx;
         c->put_chroma_pixels_tab[1] = ff_put_rv40_chroma_mc4_mmx;
-        c->rv40_weight_pixels_tab[0] = ff_rv40_weight_func_16_mmx;
-        c->rv40_weight_pixels_tab[1] = ff_rv40_weight_func_8_mmx;
+        c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_mmx;
+        c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_mmx;
+        c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_mmx;
+        c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_mmx;
     }
     if (mm_flags & AV_CPU_FLAG_MMX2) {
         c->avg_chroma_pixels_tab[0] = ff_avg_rv40_chroma_mc8_mmx2;
@@ -68,12 +74,16 @@ void ff_rv40dsp_init_x86(RV34DSPContext *c, DSPContext *dsp)
         c->avg_chroma_pixels_tab[1] = ff_avg_rv40_chroma_mc4_3dnow;
     }
     if (mm_flags & AV_CPU_FLAG_SSE2) {
-        c->rv40_weight_pixels_tab[0] = ff_rv40_weight_func_16_sse2;
-        c->rv40_weight_pixels_tab[1] = ff_rv40_weight_func_8_sse2;
+        c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_sse2;
+        c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_sse2;
+        c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_sse2;
+        c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_sse2;
     }
     if (mm_flags & AV_CPU_FLAG_SSSE3) {
-        c->rv40_weight_pixels_tab[0] = ff_rv40_weight_func_16_ssse3;
-        c->rv40_weight_pixels_tab[1] = ff_rv40_weight_func_8_ssse3;
+        c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_ssse3;
+        c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_ssse3;
+        c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_ssse3;
+        c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_ssse3;
     }
 #endif
 }