X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=encoder%2Fmacroblock.c;h=7d03e41b21c6008950ea831b3ab6ae8e1de5eda5;hb=ec3d09554addbcecb8cf82f3ff33ac737a6f996b;hp=0eafd586c5accb9506ec0dcdbd178285653d4f7f;hpb=27b73b3b86524ec9b0bdf8310a55081898b408c0;p=x264

diff --git a/encoder/macroblock.c b/encoder/macroblock.c
index 0eafd586..7d03e41b 100644
--- a/encoder/macroblock.c
+++ b/encoder/macroblock.c
@@ -64,7 +64,7 @@ static int x264_mb_decimate_score( int16_t *dct, int i_max )
     {
         int i_run;
 
-        if( abs( dct[idx--] ) > 1 )
+        if( (unsigned)(dct[idx--] + 1) > 2 )
             return 9;
 
         i_run = 0;
@@ -144,6 +144,7 @@ static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
             int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
             dct4x4[0][block_idx_x[i]][block_idx_y[i]] = h->dct.luma4x4[i][0];
+            h->dct.luma4x4[i][0] = 0;
         }
         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
         return;
@@ -154,6 +155,7 @@ static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
     {
         /* copy dc coeff */
         dct4x4[0][block_idx_y[i]][block_idx_x[i]] = dct4x4[1+i][0][0];
+        dct4x4[1+i][0][0] = 0;
 
         /* quant/scan/dequant */
         if( h->mb.b_trellis )
@@ -205,6 +207,7 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
+                h->dct.luma4x4[16+i+ch*4][0] = 0;
             }
             continue;
         }
@@ -215,6 +218,7 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
         {
             /* copy dc coeff */
             dct2x2[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];
+            dct4x4[i][0][0] = 0;
 
             /* no trellis; it doesn't seem to help chroma noticeably */
             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qscale], h->quant4_bias[CQM_4IC+b_inter][i_qscale] );
@@ -257,7 +261,7 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
     h->mb.i_cbp_chroma = 0;
     for( i = 0; i < 8; i++ )
     {
-        int nz = array_non_zero_count( h->dct.luma4x4[16+i]+1, 15 );
+        int nz = array_non_zero( h->dct.luma4x4[16+i] );
         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
         h->mb.i_cbp_chroma |= nz;
     }
@@ -269,15 +273,9 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
 
 static void x264_macroblock_encode_skip( x264_t *h )
 {
-    int i;
     h->mb.i_cbp_luma = 0x00;
     h->mb.i_cbp_chroma = 0x00;
-
-    for( i = 0; i < 16+8; i++ )
-    {
-        h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
-    }
-
+    memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
     /* store cbp */
     h->mb.cbp[h->mb.i_mb_xy] = 0;
 }
@@ -293,19 +291,21 @@ void x264_macroblock_encode_pskip( x264_t *h )
     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 
-    /* Motion compensation XXX probably unneeded */
-    h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
-                   h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
-                   mvx, mvy, 16, 16 );
+    /* don't do pskip motion compensation if it was already done in macroblock_analyse */
+    if( !h->mb.b_skip_mc )
+    {
+        h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
+                       h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
+                       mvx, mvy, 16, 16 );
 
-    /* Chroma MC */
-    h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
-                     h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
-                     mvx, mvy, 8, 8 );
+        h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
+                         h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
+                         mvx, mvy, 8, 8 );
 
-    h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
-                     h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
-                     mvx, mvy, 8, 8 );
+        h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
+                         h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
+                         mvx, mvy, 8, 8 );
+    }
 
     x264_macroblock_encode_skip( h );
 }
@@ -319,7 +319,8 @@ void x264_macroblock_encode( x264_t *h )
     int i_qp = h->mb.i_qp;
     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
     int b_force_no_skip = 0;
-    int i;
+    int i,j,idx;
+    uint8_t nnz8x8[4] = {1,1,1,1};
 
     if( h->sh.b_mbaff
         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
@@ -346,8 +347,9 @@ void x264_macroblock_encode( x264_t *h )
     }
     if( h->mb.i_type == B_SKIP )
     {
-        /* XXX motion compensation is probably unneeded */
-        x264_mb_mc( h );
+        /* don't do bskip motion compensation if it was already done in macroblock_analyse */
+        if( !h->mb.b_skip_mc )
+            x264_mb_mc( h );
         x264_macroblock_encode_skip( h );
         return;
     }
@@ -383,6 +385,8 @@ void x264_macroblock_encode( x264_t *h )
             h->predict_8x8[i_mode]( p_dst, edge );
             x264_mb_encode_i8x8( h, i, i_qp );
         }
+        for( i = 0; i < 4; i++ )
+            nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
     }
     else if( h->mb.i_type == I_4x4 )
     {
@@ -410,11 +414,12 @@ void x264_macroblock_encode( x264_t *h )
     }
     else    /* Inter MB */
     {
-        int i8x8, i4x4, idx;
+        int i8x8, i4x4;
         int i_decimate_mb = 0;
 
-        /* Motion compensation */
-        x264_mb_mc( h );
+        /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
+        if( !h->mb.b_skip_mc )
+            x264_mb_mc( h );
 
         if( h->mb.b_lossless )
         {
@@ -430,7 +435,6 @@ void x264_macroblock_encode( x264_t *h )
         else if( h->mb.b_transform_8x8 )
         {
             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
-            int nnz8x8[4] = {1,1,1,1};
             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 
@@ -450,18 +454,14 @@ void x264_macroblock_encode( x264_t *h )
                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
                     i_decimate_mb += i_decimate_8x8;
                     if( i_decimate_8x8 < 4 )
-                    {
-                        memset( h->dct.luma8x8[idx], 0, sizeof( h->dct.luma8x8[idx] ) );
-                        memset( dct8x8[idx], 0, sizeof( dct8x8[idx] ) );
                         nnz8x8[idx] = 0;
-                    }
                 }
                 else
                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
             }
 
             if( i_decimate_mb < 6 && b_decimate )
-                memset( h->dct.luma8x8, 0, sizeof( h->dct.luma8x8 ) );
+                *(uint32_t*)nnz8x8 = 0;
             else
             {
                 for( idx = 0; idx < 4; idx++ )
@@ -475,7 +475,6 @@ void x264_macroblock_encode( x264_t *h )
         else
         {
             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
-            int nnz8x8[4] = {1,1,1,1};
             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 
             for( i8x8 = 0; i8x8 < 4; i8x8++ )
@@ -496,23 +495,19 @@ void x264_macroblock_encode( x264_t *h )
                         h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 
                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
-                    
-                    if( b_decimate )
+
+                    if( b_decimate && i_decimate_8x8 <= 6 )
                         i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[idx], 16 );
                 }
 
                 /* decimate this 8x8 block */
                 i_decimate_mb += i_decimate_8x8;
                 if( i_decimate_8x8 < 4 && b_decimate )
-                {
-                    memset( &dct4x4[i8x8*4], 0, 4 * sizeof( *dct4x4 ) );
-                    memset( &h->dct.luma4x4[i8x8*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
                     nnz8x8[i8x8] = 0;
-                }
             }
 
             if( i_decimate_mb < 6 && b_decimate )
-                memset( h->dct.luma4x4, 0, 16 * sizeof( *h->dct.luma4x4 ) );
+                *(uint32_t*)nnz8x8 = 0;
             else
             {
                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
@@ -543,34 +538,38 @@ void x264_macroblock_encode( x264_t *h )
     {
         for( i = 0; i < 16; i++ )
         {
-            const int nz = array_non_zero_count( h->dct.luma4x4[i]+1, 15 );
+            int nz = array_non_zero( h->dct.luma4x4[i] );
             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
-            if( nz > 0 )
-                h->mb.i_cbp_luma = 0x0f;
-        }
-    }
-    else if( h->mb.b_transform_8x8 )
-    {
-        /* coded_block_flag is enough for CABAC.
-         * the full non_zero_count is done only in CAVLC. */
-        for( i = 0; i < 4; i++ )
-        {
-            const int nz = array_non_zero( h->dct.luma8x8[i] );
-            int j;
-            for( j = 0; j < 4; j++ )
-                h->mb.cache.non_zero_count[x264_scan8[4*i+j]] = nz;
-            if( nz > 0 )
-                h->mb.i_cbp_luma |= 1 << i;
+            h->mb.i_cbp_luma |= nz;
         }
+        h->mb.i_cbp_luma *= 0xf;
     }
     else
     {
-        for( i = 0; i < 16; i++ )
+        for( i = 0; i < 4; i++)
         {
-            const int nz = array_non_zero_count( h->dct.luma4x4[i], 16 );
-            h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
-            if( nz > 0 )
-                h->mb.i_cbp_luma |= 1 << (i/4);
+            if(!nnz8x8[i])
+            {
+                *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+i*4]] = 0;
+                *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+i*4]] = 0;
+            }
+            else if( h->mb.b_transform_8x8 )
+            {
+                *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+4*i]] = nnz8x8[i] * 0x0101;
+                *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+4*i]] = nnz8x8[i] * 0x0101;
+                h->mb.i_cbp_luma |= nnz8x8[i] << i;
+            }
+            else
+            {
+                int nz, cbp = 0;
+                for( j = 0; j < 4; j++ )
+                {
+                    nz = array_non_zero( h->dct.luma4x4[j+4*i] );
+                    h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
+                    cbp |= nz;
+                }
+                h->mb.i_cbp_luma |= cbp << i;
+            }
         }
     }
 
@@ -590,17 +589,15 @@ void x264_macroblock_encode( x264_t *h )
     if( !b_force_no_skip )
     {
         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
-            h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma == 0x00 &&
-            h->mb.cache.mv[0][x264_scan8[0]][0] == h->mb.cache.pskip_mv[0] &&
-            h->mb.cache.mv[0][x264_scan8[0]][1] == h->mb.cache.pskip_mv[1] &&
-            h->mb.cache.ref[0][x264_scan8[0]] == 0 )
+            !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) && 
+            *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
+            && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
         {
             h->mb.i_type = P_SKIP;
         }
 
         /* Check for B_SKIP */
-        if( h->mb.i_type == B_DIRECT &&
-            h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma== 0x00 )
+        if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
         {
             h->mb.i_type = B_SKIP;
         }
@@ -614,7 +611,7 @@ void x264_macroblock_encode( x264_t *h )
  *****************************************************************************/
 int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
 {
-    DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
+    DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 
@@ -637,27 +634,23 @@ int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
                        mvp[0], mvp[1], 16, 16 );
     }
 
-    /* get luma diff */
-    h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0],
-                                  h->mb.pic.p_fdec[0] );
-
     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
     {
+        int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
+        int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
+        /* get luma diff */
+        h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
+                                    h->mb.pic.p_fdec[0] + fdec_offset );
         /* encode one 4x4 block */
         for( i4x4 = 0; i4x4 < 4; i4x4++ )
         {
-            const int idx = i8x8 * 4 + i4x4;
-
-            h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
-            h->zigzagf.scan_4x4( dctscan, dct4x4[idx] );
-
+            h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
+            if( !array_non_zero(dct4x4[i4x4]) )
+                continue;
+            h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
-
             if( i_decimate_mb >= 6 )
-            {
-                /* not as P_SKIP */
                 return 0;
-            }
         }
     }
 
@@ -685,26 +678,23 @@ int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
         dct2x2[1][1] = dct4x4[3][0][0];
         h->dctf.dct2x2dc( dct2x2 );
         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
-        if( dct2x2[0][0] || dct2x2[0][1] || dct2x2[1][0] || dct2x2[1][1]  )
-        {
-            /* can't be */
+        if( array_non_zero(dct2x2) )
             return 0;
-        }
 
         /* calculate dct coeffs */
         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
         {
             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
+            if( !array_non_zero(dct4x4[i4x4]) )
+                continue;
             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
-
             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
             if( i_decimate_mb >= 7 )
-            {
                 return 0;
-            }
         }
     }
 
+    h->mb.b_skip_mc = 1;
     return 1;
 }
 
@@ -805,10 +795,8 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
         int i4;
         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
         h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
-        h->quantf.quant_4x4( dct4x4[0], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
-        h->quantf.quant_4x4( dct4x4[1], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
-        h->quantf.quant_4x4( dct4x4[2], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
-        h->quantf.quant_4x4( dct4x4[3], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
+        for( i4 = 0; i4 < 4; i4++ )
+            h->quantf.quant_4x4( dct4x4[i4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
         for( i4 = 0; i4 < 4; i4++ )
             h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 
@@ -848,9 +836,7 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
         }
     }
 
-    if( nnz8x8 )
-        h->mb.i_cbp_luma |= (1 << i8);
-    else
-        h->mb.i_cbp_luma &= ~(1 << i8);
+    h->mb.i_cbp_luma &= ~(1 << i8);
+    h->mb.i_cbp_luma |= nnz8x8 << i8;
     h->mb.i_cbp_chroma = 0x02;
 }