rv34: joint coefficient decoding and dequantization

Perform dequantization while decoding coefficients instead of performing it
on the entire coefficients buffer.

Since quantized coefficients are very sparse, this usually causes a small
speedup. Speedup of around 1% on Panda board compared to the removed here
neon code. Global speedup is probably around 3%.

Signed-off-by: Kostya Shishkov <kostya.shishkov@gmail.com>

6 files changed, 44 insertions, 90 deletions

diff --git a/libavcodec/arm/rv34dsp_init_neon.c b/libavcodec/arm/rv34dsp_init_neon.c
index acf2a7dcd3..9a09fde7a9 100644
--- a/libavcodec/arm/rv34dsp_init_neon.c
+++ b/libavcodec/arm/rv34dsp_init_neon.c
@@ -25,12 +25,9 @@
 
 void ff_rv34_inv_transform_neon(DCTELEM *block);
 void ff_rv34_inv_transform_noround_neon(DCTELEM *block);
-void ff_rv34_dequant4x4_neon(DCTELEM *block, int Qdc, int Q);
 
 void ff_rv34dsp_init_neon(RV34DSPContext *c, DSPContext* dsp)
 {
     c->rv34_inv_transform_tab[0] = ff_rv34_inv_transform_neon;
     c->rv34_inv_transform_tab[1] = ff_rv34_inv_transform_noround_neon;
-
-    c->rv34_dequant4x4 = ff_rv34_dequant4x4_neon;
 }
diff --git a/libavcodec/arm/rv34dsp_neon.S b/libavcodec/arm/rv34dsp_neon.S
index 423b537fb9..f700f5c321 100644
--- a/libavcodec/arm/rv34dsp_neon.S
+++ b/libavcodec/arm/rv34dsp_neon.S
@@ -107,27 +107,3 @@ function ff_rv34_inv_transform_noround_neon, export=1
         vst4.16         {d0[3], d1[3], d2[3], d3[3]}, [r2,:64], r1
         bx              lr
 endfunc
-
-function ff_rv34_dequant4x4_neon, export=1
-        mov             r3,  r0
-        mov             r12, #16
-        vdup.16         q0,  r2
-        vmov.16         d0[0], r1
-        vld1.16         {d2},     [r0,:64], r12
-        vld1.16         {d4},     [r0,:64], r12
-        vld1.16         {d6},     [r0,:64], r12
-        vld1.16         {d16},    [r0,:64], r12
-        vmull.s16       q1,  d2,  d0
-        vmull.s16       q2,  d4,  d1
-        vmull.s16       q3,  d6,  d1
-        vmull.s16       q8,  d16, d1
-        vqrshrn.s32     d2,  q1,  #4
-        vqrshrn.s32     d4,  q2,  #4
-        vqrshrn.s32     d6,  q3,  #4
-        vqrshrn.s32     d16, q8,  #4
-        vst1.16         {d2},     [r3,:64], r12
-        vst1.16         {d4},     [r3,:64], r12
-        vst1.16         {d6},     [r3,:64], r12
-        vst1.16         {d16},    [r3,:64], r12
-        bx              lr
-endfunc
diff --git a/libavcodec/rv34.c b/libavcodec/rv34.c
index 1a126be2f9..db2e53aaa5 100644
--- a/libavcodec/rv34.c
+++ b/libavcodec/rv34.c
@@ -212,7 +212,7 @@ static int rv34_decode_cbp(GetBitContext *gb, RV34VLC *vlc, int table)
 /**
  * Get one coefficient value from the bistream and store it.
  */
-static inline void decode_coeff(DCTELEM *dst, int coef, int esc, GetBitContext *gb, VLC* vlc)
+static inline void decode_coeff(DCTELEM *dst, int coef, int esc, GetBitContext *gb, VLC* vlc, int q)
 {
     if(coef){
         if(coef == esc){
@@ -225,14 +225,14 @@ static inline void decode_coeff(DCTELEM *dst, int coef, int esc, GetBitContext *
         }
         if(get_bits1(gb))
             coef = -coef;
-        *dst = coef;
+        *dst = (coef*q + 8) >> 4;
     }
 }
 
 /**
  * Decode 2x2 subblock of coefficients.
  */
-static inline void decode_subblock(DCTELEM *dst, int code, const int is_block2, GetBitContext *gb, VLC *vlc)
+static inline void decode_subblock(DCTELEM *dst, int code, const int is_block2, GetBitContext *gb, VLC *vlc, int q)
 {
     int coeffs[4];
 
@@ -240,15 +240,35 @@ static inline void decode_subblock(DCTELEM *dst, int code, const int is_block2,
     coeffs[1] = modulo_three_table[code][1];
     coeffs[2] = modulo_three_table[code][2];
     coeffs[3] = modulo_three_table[code][3];
-    decode_coeff(dst  , coeffs[0], 3, gb, vlc);
+    decode_coeff(dst  , coeffs[0], 3, gb, vlc, q);
     if(is_block2){
-        decode_coeff(dst+8, coeffs[1], 2, gb, vlc);
-        decode_coeff(dst+1, coeffs[2], 2, gb, vlc);
+        decode_coeff(dst+8, coeffs[1], 2, gb, vlc, q);
+        decode_coeff(dst+1, coeffs[2], 2, gb, vlc, q);
     }else{
-        decode_coeff(dst+1, coeffs[1], 2, gb, vlc);
-        decode_coeff(dst+8, coeffs[2], 2, gb, vlc);
+        decode_coeff(dst+1, coeffs[1], 2, gb, vlc, q);
+        decode_coeff(dst+8, coeffs[2], 2, gb, vlc, q);
     }
-    decode_coeff(dst+9, coeffs[3], 2, gb, vlc);
+    decode_coeff(dst+9, coeffs[3], 2, gb, vlc, q);
+}
+
+static inline void decode_subblock3(DCTELEM *dst, int code, const int is_block2, GetBitContext *gb, VLC *vlc,
+                                    int q_dc, int q_ac1, int q_ac2)
+{
+    int coeffs[4];
+
+    coeffs[0] = modulo_three_table[code][0];
+    coeffs[1] = modulo_three_table[code][1];
+    coeffs[2] = modulo_three_table[code][2];
+    coeffs[3] = modulo_three_table[code][3];
+    decode_coeff(dst  , coeffs[0], 3, gb, vlc, q_dc);
+    if(is_block2){
+        decode_coeff(dst+8, coeffs[1], 2, gb, vlc, q_ac1);
+        decode_coeff(dst+1, coeffs[2], 2, gb, vlc, q_ac1);
+    }else{
+        decode_coeff(dst+1, coeffs[1], 2, gb, vlc, q_ac1);
+        decode_coeff(dst+8, coeffs[2], 2, gb, vlc, q_ac1);
+    }
+    decode_coeff(dst+9, coeffs[3], 2, gb, vlc, q_ac2);
 }
 
 /**
@@ -262,7 +282,7 @@ static inline void decode_subblock(DCTELEM *dst, int code, const int is_block2,
  *  o--o
  */
 
-static inline void rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *rvlc, int fc, int sc)
+static inline void rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *rvlc, int fc, int sc, int q_dc, int q_ac1, int q_ac2)
 {
     int code, pattern;
 
@@ -271,40 +291,24 @@ static inline void rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *r
     pattern = code & 0x7;
 
     code >>= 3;
-    decode_subblock(dst, code, 0, gb, &rvlc->coefficient);
+    decode_subblock3(dst, code, 0, gb, &rvlc->coefficient, q_dc, q_ac1, q_ac2);
 
     if(pattern & 4){
         code = get_vlc2(gb, rvlc->second_pattern[sc].table, 9, 2);
-        decode_subblock(dst + 2, code, 0, gb, &rvlc->coefficient);
+        decode_subblock(dst + 2, code, 0, gb, &rvlc->coefficient, q_ac2);
     }
     if(pattern & 2){ // Looks like coefficients 1 and 2 are swapped for this block
         code = get_vlc2(gb, rvlc->second_pattern[sc].table, 9, 2);
-        decode_subblock(dst + 8*2, code, 1, gb, &rvlc->coefficient);
+        decode_subblock(dst + 8*2, code, 1, gb, &rvlc->coefficient, q_ac2);
     }
     if(pattern & 1){
         code = get_vlc2(gb, rvlc->third_pattern[sc].table, 9, 2);
-        decode_subblock(dst + 8*2+2, code, 0, gb, &rvlc->coefficient);
+        decode_subblock(dst + 8*2+2, code, 0, gb, &rvlc->coefficient, q_ac2);
     }
 
 }
 
 /**
- * Dequantize 4x4 block of DC values for 16x16 macroblock.
- * @todo optimize
- */
-static inline void rv34_dequant4x4_16x16(DCTELEM *block, int Qdc, int Q)
-{
-    int i;
-
-    for(i = 0; i < 3; i++)
-         block[rv34_dezigzag[i]] = (block[rv34_dezigzag[i]] * Qdc + 8) >> 4;
-    for(; i < 16; i++)
-         block[rv34_dezigzag[i]] = (block[rv34_dezigzag[i]] * Q + 8) >> 4;
-}
-/** @} */ //block functions
-
-
-/**
  * @name RV30/40 bitstream parsing
  * @{
  */
@@ -1097,6 +1101,7 @@ static int rv34_decode_macroblock(RV34DecContext *r, int8_t *intra_types)
     MpegEncContext *s = &r->s;
     GetBitContext *gb = &s->gb;
     int cbp, cbp2;
+    int q_dc, q_ac;
     int i, blknum, blkoff;
     LOCAL_ALIGNED_16(DCTELEM, block16, [64]);
     int luma_dc_quant;
@@ -1133,31 +1138,34 @@ static int rv34_decode_macroblock(RV34DecContext *r, int8_t *intra_types)
 
     luma_dc_quant = r->block_type == RV34_MB_P_MIX16x16 ? r->luma_dc_quant_p[s->qscale] : r->luma_dc_quant_i[s->qscale];
     if(r->is16){
+        q_dc = rv34_qscale_tab[luma_dc_quant];
+        q_ac = rv34_qscale_tab[s->qscale];
         memset(block16, 0, 64 * sizeof(*block16));
-        rv34_decode_block(block16, gb, r->cur_vlcs, 3, 0);
-        rv34_dequant4x4_16x16(block16, rv34_qscale_tab[luma_dc_quant],rv34_qscale_tab[s->qscale]);
+        rv34_decode_block(block16, gb, r->cur_vlcs, 3, 0, q_dc, q_dc, q_ac);
         r->rdsp.rv34_inv_transform_tab[1](block16);
     }
 
+    q_ac = rv34_qscale_tab[s->qscale];
     for(i = 0; i < 16; i++, cbp >>= 1){
         if(!r->is16 && !(cbp & 1)) continue;
         blknum = ((i & 2) >> 1) + ((i & 8) >> 2);
         blkoff = ((i & 1) << 2) + ((i & 4) << 3);
         if(cbp & 1)
-            rv34_decode_block(s->block[blknum] + blkoff, gb, r->cur_vlcs, r->luma_vlc, 0);
-        r->rdsp.rv34_dequant4x4(s->block[blknum] + blkoff, rv34_qscale_tab[s->qscale],rv34_qscale_tab[s->qscale]);
+            rv34_decode_block(s->block[blknum] + blkoff, gb,
+                              r->cur_vlcs, r->luma_vlc, 0, q_ac, q_ac, q_ac);
         if(r->is16) //FIXME: optimize
             s->block[blknum][blkoff] = block16[(i & 3) | ((i & 0xC) << 1)];
         r->rdsp.rv34_inv_transform_tab[0](s->block[blknum] + blkoff);
     }
     if(r->block_type == RV34_MB_P_MIX16x16)
         r->cur_vlcs = choose_vlc_set(r->si.quant, r->si.vlc_set, 1);
+    q_dc = rv34_qscale_tab[rv34_chroma_quant[1][s->qscale]];
+    q_ac = rv34_qscale_tab[rv34_chroma_quant[0][s->qscale]];
     for(; i < 24; i++, cbp >>= 1){
         if(!(cbp & 1)) continue;
         blknum = ((i & 4) >> 2) + 4;
         blkoff = ((i & 1) << 2) + ((i & 2) << 4);
-        rv34_decode_block(s->block[blknum] + blkoff, gb, r->cur_vlcs, r->chroma_vlc, 1);
-        r->rdsp.rv34_dequant4x4(s->block[blknum] + blkoff, rv34_qscale_tab[rv34_chroma_quant[1][s->qscale]],rv34_qscale_tab[rv34_chroma_quant[0][s->qscale]]);
+        rv34_decode_block(s->block[blknum] + blkoff, gb, r->cur_vlcs, r->chroma_vlc, 1, q_dc, q_ac, q_ac);
         r->rdsp.rv34_inv_transform_tab[0](s->block[blknum] + blkoff);
     }
     if (IS_INTRA(s->current_picture_ptr->f.mb_type[mb_pos]))
diff --git a/libavcodec/rv34data.h b/libavcodec/rv34data.h
index f8f941d061..fa41a882a6 100644
--- a/libavcodec/rv34data.h
+++ b/libavcodec/rv34data.h
@@ -101,16 +101,6 @@ static const uint16_t rv34_qscale_tab[32] = {
 };
 
 /**
- * 4x4 dezigzag pattern
- */
-static const uint8_t rv34_dezigzag[16] = {
-  0,  1,  8, 16,
-  9,  2,  3, 10,
- 17, 24, 25, 18,
- 11, 19, 26, 27
-};
-
-/**
  * tables used to translate a quantizer value into a VLC set for decoding
  * The first table is used for intraframes.
  */
diff --git a/libavcodec/rv34dsp.c b/libavcodec/rv34dsp.c
index 974bf9ec16..1f4cea8544 100644
--- a/libavcodec/rv34dsp.c
+++ b/libavcodec/rv34dsp.c
@@ -100,26 +100,10 @@ static void rv34_inv_transform_noround_c(DCTELEM *block){
 /** @} */ // transform
 
 
-/**
- * Dequantize ordinary 4x4 block.
- */
-void ff_rv34_dequant4x4_neon(DCTELEM *block, int Qdc, int Q);
-static void rv34_dequant4x4_c(DCTELEM *block, int Qdc, int Q)
-{
-    int i, j;
-
-    block[0] = (block[0] * Qdc + 8) >> 4;
-    for (i = 0; i < 4; i++)
-        for (j = !i; j < 4; j++)
-            block[j + i*8] = (block[j + i*8] * Q + 8) >> 4;
-}
-
 av_cold void ff_rv34dsp_init(RV34DSPContext *c, DSPContext* dsp) {
     c->rv34_inv_transform_tab[0] = rv34_inv_transform_c;
     c->rv34_inv_transform_tab[1] = rv34_inv_transform_noround_c;
 
-    c->rv34_dequant4x4 = rv34_dequant4x4_c;
-
     if (HAVE_NEON)
         ff_rv34dsp_init_neon(c, dsp);
 }
diff --git a/libavcodec/rv34dsp.h b/libavcodec/rv34dsp.h
index 01352ea793..f2bc20e911 100644
--- a/libavcodec/rv34dsp.h
+++ b/libavcodec/rv34dsp.h
@@ -56,7 +56,6 @@ typedef struct RV34DSPContext {
     h264_chroma_mc_func avg_chroma_pixels_tab[3];
     rv40_weight_func rv40_weight_pixels_tab[2];
     rv34_inv_transform_func rv34_inv_transform_tab[2];
-    void (*rv34_dequant4x4)(DCTELEM *block, int Qdc, int Q);
     rv40_weak_loop_filter_func rv40_weak_loop_filter[2];
     rv40_strong_loop_filter_func rv40_strong_loop_filter[2];
     rv40_loop_filter_strength_func rv40_loop_filter_strength[2];