]> git.sesse.net Git - x264/blobdiff - encoder/macroblock.c
Eliminate the possibility of CAVLC level code overflow
[x264] / encoder / macroblock.c
index 60e14b8fd3c2732acf7d8cdffb85919142455fc9..441df88ab1cc7539ef3de2b7703c740b5cc10bf6 100644 (file)
@@ -1,7 +1,7 @@
 /*****************************************************************************
- * macroblock.c: h264 encoder library
+ * macroblock.c: macroblock encoding
  *****************************************************************************
- * Copyright (C) 2003-2008 x264 project
+ * Copyright (C) 2003-2011 x264 project
  *
  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
  *          Loren Merritt <lorenm@u.washington.edu>
@@ -20,6 +20,9 @@
  * You should have received a copy of the GNU General Public License
  * along with this program; if not, write to the Free Software
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
+ *
+ * This program is also available under a commercial proprietary license.
+ * For more information, contact us at licensing@x264.com.
  *****************************************************************************/
 
 #include "common/common.h"
@@ -27,8 +30,8 @@
 
 /* These chroma DC functions don't have assembly versions and are only used here. */
 
-#define ZIG(i,y,x) level[i] = dct[x][y];
-static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
+#define ZIG(i,y,x) level[i] = dct[x*2+y];
+static inline void zigzag_scan_2x2_dc( dctcoef level[4], dctcoef dct[4] )
 {
     ZIG(0,0,0)
     ZIG(1,0,1)
@@ -38,76 +41,62 @@ static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
 #undef ZIG
 
 #define IDCT_DEQUANT_START \
-    int d0 = dct[0][0] + dct[0][1]; \
-    int d1 = dct[1][0] + dct[1][1]; \
-    int d2 = dct[0][0] - dct[0][1]; \
-    int d3 = dct[1][0] - dct[1][1]; \
-    int dmf = dequant_mf[i_qp%6][0][0]; \
-    int qbits = i_qp/6 - 5; \
-    if( qbits > 0 ) \
-    { \
-        dmf <<= qbits; \
-        qbits = 0; \
-    }
+    int d0 = dct[0] + dct[1]; \
+    int d1 = dct[2] + dct[3]; \
+    int d2 = dct[0] - dct[1]; \
+    int d3 = dct[2] - dct[3]; \
+    int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
 
-static inline void idct_dequant_2x2_dc( int16_t dct[2][2], int16_t dct4x4[4][4][4], int dequant_mf[6][4][4], int i_qp )
+static inline void idct_dequant_2x2_dc( dctcoef dct[4], dctcoef dct4x4[4][16], int dequant_mf[6][16], int i_qp )
 {
     IDCT_DEQUANT_START
-    dct4x4[0][0][0] = (d0 + d1) * dmf >> -qbits;
-    dct4x4[1][0][0] = (d0 - d1) * dmf >> -qbits;
-    dct4x4[2][0][0] = (d2 + d3) * dmf >> -qbits;
-    dct4x4[3][0][0] = (d2 - d3) * dmf >> -qbits;
+    dct4x4[0][0] = (d0 + d1) * dmf >> 5;
+    dct4x4[1][0] = (d0 - d1) * dmf >> 5;
+    dct4x4[2][0] = (d2 + d3) * dmf >> 5;
+    dct4x4[3][0] = (d2 - d3) * dmf >> 5;
 }
 
-static inline void idct_dequant_2x2_dconly( int16_t dct[2][2], int dequant_mf[6][4][4], int i_qp )
+static inline void idct_dequant_2x2_dconly( dctcoef out[4], dctcoef dct[4], int dequant_mf[6][16], int i_qp )
 {
     IDCT_DEQUANT_START
-    dct[0][0] = (d0 + d1) * dmf >> -qbits;
-    dct[0][1] = (d0 - d1) * dmf >> -qbits;
-    dct[1][0] = (d2 + d3) * dmf >> -qbits;
-    dct[1][1] = (d2 - d3) * dmf >> -qbits;
-}
-
-static inline void dct2x2dc( int16_t d[2][2], int16_t dct4x4[4][4][4] )
-{
-    int d0 = dct4x4[0][0][0] + dct4x4[1][0][0];
-    int d1 = dct4x4[2][0][0] + dct4x4[3][0][0];
-    int d2 = dct4x4[0][0][0] - dct4x4[1][0][0];
-    int d3 = dct4x4[2][0][0] - dct4x4[3][0][0];
-    d[0][0] = d0 + d1;
-    d[1][0] = d2 + d3;
-    d[0][1] = d0 - d1;
-    d[1][1] = d2 - d3;
-    dct4x4[0][0][0] = 0;
-    dct4x4[1][0][0] = 0;
-    dct4x4[2][0][0] = 0;
-    dct4x4[3][0][0] = 0;
+    out[0] = (d0 + d1) * dmf >> 5;
+    out[1] = (d0 - d1) * dmf >> 5;
+    out[2] = (d2 + d3) * dmf >> 5;
+    out[3] = (d2 - d3) * dmf >> 5;
 }
 
-static inline void dct2x2dc_dconly( int16_t d[2][2] )
+static inline void dct2x2dc( dctcoef d[4], dctcoef dct4x4[4][16] )
 {
-    int d0 = d[0][0] + d[0][1];
-    int d1 = d[1][0] + d[1][1];
-    int d2 = d[0][0] - d[0][1];
-    int d3 = d[1][0] - d[1][1];
-    d[0][0] = d0 + d1;
-    d[1][0] = d2 + d3;
-    d[0][1] = d0 - d1;
-    d[1][1] = d2 - d3;
+    int d0 = dct4x4[0][0] + dct4x4[1][0];
+    int d1 = dct4x4[2][0] + dct4x4[3][0];
+    int d2 = dct4x4[0][0] - dct4x4[1][0];
+    int d3 = dct4x4[2][0] - dct4x4[3][0];
+    d[0] = d0 + d1;
+    d[2] = d2 + d3;
+    d[1] = d0 - d1;
+    d[3] = d2 - d3;
+    dct4x4[0][0] = 0;
+    dct4x4[1][0] = 0;
+    dct4x4[2][0] = 0;
+    dct4x4[3][0] = 0;
 }
 
-static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, int16_t dct[4][4], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
+static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, dctcoef dct[16], int i_qp, int ctx_block_cat, int b_intra, int idx )
 {
     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
+    if( h->mb.b_noise_reduction && ctx_block_cat != DCT_LUMA_AC )
+        h->quantf.denoise_dct( dct, h->nr_residual_sum[0], h->nr_offset[0], 16 );
     if( h->mb.b_trellis )
-        return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
+        return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, ctx_block_cat, b_intra, 0, idx );
     else
         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
 }
 
-static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[8][8], int i_qp, int b_intra, int idx )
+static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, dctcoef dct[64], int i_qp, int b_intra, int idx )
 {
     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
+    if( h->mb.b_noise_reduction )
+        h->quantf.denoise_dct( dct, h->nr_residual_sum[1], h->nr_offset[1], 64 );
     if( h->mb.b_trellis )
         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
     else
@@ -128,9 +117,9 @@ static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[8][8], int i_qp,
 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 {
     int nz;
-    uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
-    uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
-    ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[4] );
+    pixel *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
+    pixel *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
+    ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
 
     if( h->mb.b_lossless )
     {
@@ -153,25 +142,35 @@ void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
     }
 }
 
-#define STORE_8x8_NNZ(idx,nz)\
+#define STORE_8x8_NNZ( s8, nz )\
+do\
+{\
+    M16( &h->mb.cache.non_zero_count[(s8) + 0*8] ) = (nz) * 0x0101;\
+    M16( &h->mb.cache.non_zero_count[(s8) + 1*8] ) = (nz) * 0x0101;\
+} while(0)
+
+#define CLEAR_16x16_NNZ \
 {\
-    *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] = nz * 0x0101;\
-    *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] = nz * 0x0101;\
+    M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;\
+    M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;\
+    M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;\
+    M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;\
 }
 
 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 {
-    int x = 8 * (idx&1);
-    int y = 8 * (idx>>1);
+    int x = idx&1;
+    int y = idx>>1;
+    int s8 = X264_SCAN8_0 + 2*x + 16*y;
     int nz;
-    uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
-    uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
-    ALIGNED_ARRAY_16( int16_t, dct8x8,[8],[8] );
+    pixel *p_src = &h->mb.pic.p_fenc[0][8*x + 8*y*FENC_STRIDE];
+    pixel *p_dst = &h->mb.pic.p_fdec[0][8*x + 8*y*FDEC_STRIDE];
+    ALIGNED_ARRAY_16( dctcoef, dct8x8,[64] );
 
     if( h->mb.b_lossless )
     {
         nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
-        STORE_8x8_NNZ(idx,nz);
+        STORE_8x8_NNZ( s8, nz );
         h->mb.i_cbp_luma |= nz<<idx;
         return;
     }
@@ -185,31 +184,30 @@ void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
         h->dctf.add8x8_idct8( p_dst, dct8x8 );
-        STORE_8x8_NNZ(idx,1);
+        STORE_8x8_NNZ( s8, 1 );
     }
     else
-        STORE_8x8_NNZ(idx,0);
+        STORE_8x8_NNZ( s8, 0 );
 }
 
 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 {
-    uint8_t  *p_src = h->mb.pic.p_fenc[0];
-    uint8_t  *p_dst = h->mb.pic.p_fdec[0];
+    pixel *p_src = h->mb.pic.p_fenc[0];
+    pixel *p_dst = h->mb.pic.p_fdec[0];
 
-    ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[4][4] );
-    ALIGNED_ARRAY_16( int16_t, dct_dc4x4,[4],[4] );
+    ALIGNED_ARRAY_16( dctcoef, dct4x4,[16],[16] );
+    ALIGNED_ARRAY_16( dctcoef, dct_dc4x4,[16] );
 
-    int i, nz;
-    int b_decimate = h->sh.i_type == SLICE_TYPE_B || (h->param.analyse.b_dct_decimate && h->sh.i_type == SLICE_TYPE_P);
-    int decimate_score = b_decimate ? 0 : 9;
+    int nz;
+    int decimate_score = h->mb.b_dct_decimate ? 0 : 9;
 
     if( h->mb.b_lossless )
     {
-        for( i = 0; i < 16; i++ )
+        for( int i = 0; i < 16; i++ )
         {
             int oe = block_idx_xy_fenc[i];
             int od = block_idx_xy_fdec[i];
-            nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[0][block_idx_yx_1d[i]] );
+            nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
             h->mb.i_cbp_luma |= nz;
         }
@@ -221,11 +219,13 @@ static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 
     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 
-    for( i = 0; i < 16; i++ )
+    for( int i = 0; i < 16; i++ )
     {
         /* copy dc coeff */
-        dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
-        dct4x4[i][0][0] = 0;
+        if( h->mb.b_noise_reduction )
+            h->quantf.denoise_dct( dct4x4[i], h->nr_residual_sum[0], h->nr_offset[0], 16 );
+        dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
+        dct4x4[i][0] = 0;
 
         /* quant/scan/dequant */
         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
@@ -244,15 +244,12 @@ static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
     if( decimate_score < 6 )
     {
         h->mb.i_cbp_luma = 0;
-        *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
-        *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
-        *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
-        *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
+        CLEAR_16x16_NNZ
     }
 
     h->dctf.dct4x4dc( dct_dc4x4 );
     if( h->mb.b_trellis )
-        nz = x264_quant_dc_trellis( h, (int16_t*)dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
+        nz = x264_quant_dc_trellis( h, dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
     else
         nz = h->quantf.quant_4x4_dc( dct_dc4x4, h->quant4_mf[CQM_4IY][i_qp][0]>>1, h->quant4_bias[CQM_4IY][i_qp][0]<<1 );
 
@@ -265,8 +262,8 @@ static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
         h->dctf.idct4x4dc( dct_dc4x4 );
         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
         if( h->mb.i_cbp_luma )
-            for( i = 0; i < 16; i++ )
-                dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
+            for( int i = 0; i < 16; i++ )
+                dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
     }
 
     /* put pixels to fdec */
@@ -276,21 +273,38 @@ static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 }
 
+/* Round down coefficients losslessly in DC-only chroma blocks.
+ * Unlike luma blocks, this can't be done with a lookup table or
+ * other shortcut technique because of the interdependencies
+ * between the coefficients due to the chroma DC transform. */
+static ALWAYS_INLINE int x264_mb_optimize_chroma_dc( x264_t *h, dctcoef dct2x2[4], int dequant_mf[6][16], int i_qp )
+{
+    int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
+
+    /* If the QP is too high, there's no benefit to rounding optimization. */
+    if( dmf > 32*64 )
+        return 1;
+
+    return h->quantf.optimize_chroma_dc( dct2x2, dmf );
+}
+
 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 {
-    int i, ch, nz, nz_dc;
-    int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
-    ALIGNED_ARRAY_16( int16_t, dct2x2,[2],[2] );
+    int nz, nz_dc;
+    int b_decimate = b_inter && h->mb.b_dct_decimate;
+    ALIGNED_ARRAY_16( dctcoef, dct2x2,[4] );
     h->mb.i_cbp_chroma = 0;
+    h->nr_count[2] += h->mb.b_noise_reduction * 4;
 
     /* Early termination: check variance of chroma residual before encoding.
      * Don't bother trying early termination at low QPs.
      * Values are experimentally derived. */
-    if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
+    if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) && !h->mb.b_noise_reduction )
     {
         int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
         int ssd[2];
-        int score  = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
+        int score = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
+        if( score < thresh*4 )
             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
         if( score < thresh*4 )
         {
@@ -302,24 +316,25 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
-            h->mb.cache.non_zero_count[x264_scan8[25]] = 0;
-            h->mb.cache.non_zero_count[x264_scan8[26]] = 0;
-            for( ch = 0; ch < 2; ch++ )
+            M16( &h->mb.cache.non_zero_count[x264_scan8[25]] ) = 0;
+
+            for( int ch = 0; ch < 2; ch++ )
             {
                 if( ssd[ch] > thresh )
                 {
                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
-                    dct2x2dc_dconly( dct2x2 );
                     if( h->mb.b_trellis )
-                        nz_dc = x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
+                        nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
                     else
-                        nz_dc = h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<
-    1 );
+                        nz_dc = h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<1 );
+
                     if( nz_dc )
                     {
+                        if( !x264_mb_optimize_chroma_dc( h, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp ) )
+                            continue;
                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
-                        idct_dequant_2x2_dconly( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
+                        idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
                         h->mb.i_cbp_chroma = 1;
                     }
@@ -329,18 +344,18 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
         }
     }
 
-    for( ch = 0; ch < 2; ch++ )
+    for( int ch = 0; ch < 2; ch++ )
     {
-        uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
-        uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
+        pixel *p_src = h->mb.pic.p_fenc[1+ch];
+        pixel *p_dst = h->mb.pic.p_fdec[1+ch];
         int i_decimate_score = 0;
         int nz_ac = 0;
 
-        ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[4][4] );
+        ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
 
         if( h->mb.b_lossless )
         {
-            for( i = 0; i < 4; i++ )
+            for( int i = 0; i < 4; i++ )
             {
                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
@@ -353,9 +368,12 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
         }
 
         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
+        if( h->mb.b_noise_reduction )
+            for( int i = 0; i < 4; i++ )
+                h->quantf.denoise_dct( dct4x4[i], h->nr_residual_sum[2], h->nr_offset[2], 16 );
         dct2x2dc( dct2x2, dct4x4 );
         /* calculate dct coeffs */
-        for( i = 0; i < 4; i++ )
+        for( int i = 0; i < 4; i++ )
         {
             if( h->mb.b_trellis )
                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
@@ -373,7 +391,7 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
         }
 
         if( h->mb.b_trellis )
-            nz_dc = x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
+            nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
         else
             nz_dc = h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<1 );
 
@@ -388,9 +406,14 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
             if( !nz_dc ) /* Whole block is empty */
                 continue;
+            if( !x264_mb_optimize_chroma_dc( h, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp  ) )
+            {
+                h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
+                continue;
+            }
             /* DC-only */
             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
-            idct_dequant_2x2_dconly( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
+            idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
         }
         else
@@ -405,19 +428,20 @@ void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
         }
     }
 
-    if( h->mb.i_cbp_chroma )
-        h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
-    else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
-             h->mb.cache.non_zero_count[x264_scan8[26]] )
-        h->mb.i_cbp_chroma = 1;    /* dc only */
+    /* 0 = none, 1 = DC only, 2 = DC+AC */
+    h->mb.i_cbp_chroma = ((!!M16( &h->mb.cache.non_zero_count[x264_scan8[25]] )) | h->mb.i_cbp_chroma) + h->mb.i_cbp_chroma;
 }
 
 static void x264_macroblock_encode_skip( x264_t *h )
 {
-    h->mb.i_cbp_luma = 0x00;
-    h->mb.i_cbp_chroma = 0x00;
-    memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
-    /* store cbp */
+    M32( &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] ) = 0;
+    M32( &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] ) = 0;
+    M32( &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] ) = 0;
+    M32( &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] ) = 0;
+    for( int i = 16; i < 24; i++ )
+        h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
+    h->mb.i_cbp_luma = 0;
+    h->mb.i_cbp_chroma = 0;
     h->mb.cbp[h->mb.i_mb_xy] = 0;
 }
 
@@ -427,25 +451,34 @@ static void x264_macroblock_encode_skip( x264_t *h )
  *****************************************************************************/
 static void x264_macroblock_encode_pskip( x264_t *h )
 {
-    const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
-                                h->mb.mv_min[0], h->mb.mv_max[0] );
-    const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
-                                h->mb.mv_min[1], h->mb.mv_max[1] );
-
     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
     if( !h->mb.b_skip_mc )
     {
+        int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
+                              h->mb.mv_min[0], h->mb.mv_max[0] );
+        int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
+                              h->mb.mv_min[1], h->mb.mv_max[1] );
+
         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 );
+                       mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 
-        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 );
+        /* Special case for mv0, which is (of course) very common in P-skip mode. */
+        if( mvx | mvy )
+            h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
+                             h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
+                             mvx, mvy, 8, 8 );
+        else
+            h->mc.load_deinterleave_8x8x2_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1] );
+
+        if( h->sh.weight[0][1].weightfn )
+            h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
+                                               h->mb.pic.p_fdec[1], FDEC_STRIDE,
+                                               &h->sh.weight[0][1], 8 );
+        if( h->sh.weight[0][2].weightfn )
+            h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
+                                               h->mb.pic.p_fdec[2], FDEC_STRIDE,
+                                               &h->sh.weight[0][2], 8 );
     }
 
     x264_macroblock_encode_skip( h );
@@ -455,25 +488,21 @@ static void x264_macroblock_encode_pskip( x264_t *h )
  * Intra prediction for predictive lossless mode.
  *****************************************************************************/
 
-/* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
- * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
- * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
- * lossless mode cannot be mixed with lossy mode within a frame. */
-/* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
- * need to be done unless we decide to allow mixing lossless and lossy compression. */
-
 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 {
-    int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
     if( i_mode == I_PRED_CHROMA_V )
     {
-        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
-        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
+        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1]-FENC_STRIDE, FENC_STRIDE, 8 );
+        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2]-FENC_STRIDE, FENC_STRIDE, 8 );
+        memcpy( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[1]-FDEC_STRIDE, 8*sizeof(pixel) );
+        memcpy( h->mb.pic.p_fdec[2], h->mb.pic.p_fdec[2]-FDEC_STRIDE, 8*sizeof(pixel) );
     }
     else if( i_mode == I_PRED_CHROMA_H )
     {
-        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
-        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
+        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1]-1, FENC_STRIDE, 8 );
+        h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2]-1, FENC_STRIDE, 8 );
+        x264_copy_column8( h->mb.pic.p_fdec[1]+4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+4*FDEC_STRIDE-1 );
+        x264_copy_column8( h->mb.pic.p_fdec[2]+4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+4*FDEC_STRIDE-1 );
     }
     else
     {
@@ -482,10 +511,10 @@ void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
     }
 }
 
-void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
+void x264_predict_lossless_4x4( x264_t *h, pixel *p_dst, int idx, int i_mode )
 {
     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
-    uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
+    pixel *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 
     if( i_mode == I_PRED_4x4_V )
         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
@@ -495,10 +524,10 @@ void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
         h->predict_4x4[i_mode]( p_dst );
 }
 
-void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
+void x264_predict_lossless_8x8( x264_t *h, pixel *p_dst, int idx, int i_mode, pixel edge[33] )
 {
     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
-    uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
+    pixel *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 
     if( i_mode == I_PRED_8x8_V )
         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
@@ -524,14 +553,22 @@ void x264_predict_lossless_16x16( x264_t *h, int i_mode )
  *****************************************************************************/
 void x264_macroblock_encode( x264_t *h )
 {
-    int i_cbp_dc = 0;
     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_decimate = h->mb.b_dct_decimate;
     int b_force_no_skip = 0;
-    int i,idx,nz;
+    int nz;
     h->mb.i_cbp_luma = 0;
     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 
+    if( h->mb.i_type == I_PCM )
+    {
+        /* if PCM is chosen, we need to store reconstructed frame data */
+        h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE, 16 );
+        h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, 8 );
+        h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, 8 );
+        return;
+    }
+
     if( h->sh.b_mbaff
         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
@@ -579,25 +616,25 @@ void x264_macroblock_encode( x264_t *h )
     }
     else if( h->mb.i_type == I_8x8 )
     {
-        ALIGNED_ARRAY_16( uint8_t, edge,[33] );
+        ALIGNED_ARRAY_16( pixel, edge,[33] );
         h->mb.b_transform_8x8 = 1;
         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
         if( h->mb.i_skip_intra )
         {
             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = h->mb.pic.i8x8_nnz_buf[0];
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = h->mb.pic.i8x8_nnz_buf[1];
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = h->mb.pic.i8x8_nnz_buf[2];
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = h->mb.pic.i8x8_nnz_buf[3];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
             /* In RD mode, restore the now-overwritten DCT data. */
             if( h->mb.i_skip_intra == 2 )
                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
         }
-        for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
+        for( int i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
         {
-            uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
-            int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
+            pixel *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
+            int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 
             if( h->mb.b_lossless )
@@ -615,23 +652,23 @@ void x264_macroblock_encode( x264_t *h )
         if( h->mb.i_skip_intra )
         {
             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = h->mb.pic.i4x4_nnz_buf[0];
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = h->mb.pic.i4x4_nnz_buf[1];
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = h->mb.pic.i4x4_nnz_buf[2];
-            *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = h->mb.pic.i4x4_nnz_buf[3];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
+            M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
             /* In RD mode, restore the now-overwritten DCT data. */
             if( h->mb.i_skip_intra == 2 )
                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
         }
-        for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
+        for( int i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
         {
-            uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
-            int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
+            pixel *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
+            int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 
             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
                 /* emulate missing topright samples */
-                *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
+                MPIXEL_X4( &p_dst[4-FDEC_STRIDE] ) = PIXEL_SPLAT_X4( p_dst[3-FDEC_STRIDE] );
 
             if( h->mb.b_lossless )
                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
@@ -642,7 +679,6 @@ void x264_macroblock_encode( x264_t *h )
     }
     else    /* Inter MB */
     {
-        int i8x8, i4x4;
         int i_decimate_mb = 0;
 
         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
@@ -652,18 +688,19 @@ void x264_macroblock_encode( x264_t *h )
         if( h->mb.b_lossless )
         {
             if( h->mb.b_transform_8x8 )
-                for( i8x8 = 0; i8x8 < 4; i8x8++ )
+                for( int i8x8 = 0; i8x8 < 4; i8x8++ )
                 {
-                    int x = 8*(i8x8&1);
-                    int y = 8*(i8x8>>1);
-                    nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
-                                        h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
-                                        h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
-                    STORE_8x8_NNZ(i8x8,nz);
+                    int x = i8x8&1;
+                    int y = i8x8>>1;
+                    int s8 = X264_SCAN8_0 + 2*x + 16*y;
+
+                    nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8], h->mb.pic.p_fenc[0] + 8*x + 8*y*FENC_STRIDE,
+                                                                   h->mb.pic.p_fdec[0] + 8*x + 8*y*FDEC_STRIDE );
+                    STORE_8x8_NNZ( s8, nz );
                     h->mb.i_cbp_luma |= nz << i8x8;
                 }
             else
-                for( i4x4 = 0; i4x4 < 16; i4x4++ )
+                for( int i4x4 = 0; i4x4 < 16; i4x4++ )
                 {
                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
@@ -674,15 +711,13 @@ void x264_macroblock_encode( x264_t *h )
         }
         else if( h->mb.b_transform_8x8 )
         {
-            ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[8][8] );
-            b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
+            ALIGNED_ARRAY_16( dctcoef, dct8x8,[4],[64] );
+            b_decimate &= !h->mb.b_trellis || !h->param.b_cabac; // 8x8 trellis is inherently optimal decimation for CABAC
             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 
-            for( idx = 0; idx < 4; idx++ )
+            for( int idx = 0; idx < 4; idx++ )
             {
-                if( h->mb.b_noise_reduction )
-                    h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 
                 if( nz )
@@ -703,44 +738,43 @@ void x264_macroblock_encode( x264_t *h )
             if( i_decimate_mb < 6 && b_decimate )
             {
                 h->mb.i_cbp_luma = 0;
-                *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
-                *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
-                *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
-                *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
+                CLEAR_16x16_NNZ
             }
             else
             {
-                for( idx = 0; idx < 4; idx++ )
+                for( int idx = 0; idx < 4; idx++ )
                 {
+                    int x = idx&1;
+                    int y = idx>>1;
+                    int s8 = X264_SCAN8_0 + 2*x + 16*y;
+
                     if( h->mb.i_cbp_luma&(1<<idx) )
                     {
                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
-                        h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
-                        STORE_8x8_NNZ(idx,1);
+                        h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][8*x + 8*y*FDEC_STRIDE], dct8x8[idx] );
+                        STORE_8x8_NNZ( s8, 1 );
                     }
                     else
-                        STORE_8x8_NNZ(idx,0);
+                        STORE_8x8_NNZ( s8, 0 );
                 }
             }
         }
         else
         {
-            ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[4][4] );
+            ALIGNED_ARRAY_16( dctcoef, dct4x4,[16],[16] );
             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 
-            for( i8x8 = 0; i8x8 < 4; i8x8++ )
+            for( int i8x8 = 0; i8x8 < 4; i8x8++ )
             {
                 int i_decimate_8x8 = 0;
                 int cbp = 0;
 
                 /* encode one 4x4 block */
-                for( i4x4 = 0; i4x4 < 4; i4x4++ )
+                for( int i4x4 = 0; i4x4 < 4; i4x4++ )
                 {
-                    idx = i8x8 * 4 + i4x4;
+                    int idx = i8x8 * 4 + i4x4;
 
-                    if( h->mb.b_noise_reduction )
-                        h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 
@@ -754,18 +788,24 @@ void x264_macroblock_encode( x264_t *h )
                     }
                 }
 
+                int x = i8x8&1;
+                int y = i8x8>>1;
+
                 /* decimate this 8x8 block */
                 i_decimate_mb += i_decimate_8x8;
                 if( b_decimate )
                 {
                     if( i_decimate_8x8 < 4 )
-                        STORE_8x8_NNZ(i8x8,0)
+                    {
+                        int s8 = X264_SCAN8_0 + 2*x + 16*y;
+                        STORE_8x8_NNZ( s8, 0 );
+                    }
                     else
                         h->mb.i_cbp_luma |= 1<<i8x8;
                 }
                 else if( cbp )
                 {
-                    h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
+                    h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][8*x + 8*y*FDEC_STRIDE], &dct4x4[i8x8*4] );
                     h->mb.i_cbp_luma |= 1<<i8x8;
                 }
             }
@@ -775,14 +815,11 @@ void x264_macroblock_encode( x264_t *h )
                 if( i_decimate_mb < 6 )
                 {
                     h->mb.i_cbp_luma = 0;
-                    *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
-                    *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
-                    *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
-                    *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
+                    CLEAR_16x16_NNZ
                 }
                 else
                 {
-                    for( i8x8 = 0; i8x8 < 4; i8x8++ )
+                    for( int i8x8 = 0; i8x8 < 4; i8x8++ )
                         if( h->mb.i_cbp_luma&(1<<i8x8) )
                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
                 }
@@ -806,15 +843,13 @@ void x264_macroblock_encode( x264_t *h )
     /* encode the 8x8 blocks */
     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 
-    if( h->param.b_cabac )
-    {
-        i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
-                 | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
-                 | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
-    }
-
     /* store cbp */
-    h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
+    int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
+    if( h->param.b_cabac )
+        cbp |= h->mb.cache.non_zero_count[x264_scan8[24]] << 8
+            |  h->mb.cache.non_zero_count[x264_scan8[25]] << 9
+            |  h->mb.cache.non_zero_count[x264_scan8[26]] << 10;
+    h->mb.cbp[h->mb.i_mb_xy] = cbp;
 
     /* Check for P_SKIP
      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
@@ -823,7 +858,7 @@ void x264_macroblock_encode( x264_t *h )
     {
         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
             !(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
+            M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
         {
             h->mb.i_type = P_SKIP;
@@ -839,21 +874,17 @@ void x264_macroblock_encode( x264_t *h )
 
 /*****************************************************************************
  * x264_macroblock_probe_skip:
- *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
- *  the previous QP
+ *  Check if the current MB could be encoded as a [PB]_SKIP
  *****************************************************************************/
 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 {
-    ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[4][4] );
-    ALIGNED_ARRAY_16( int16_t, dct2x2,[2],[2] );
-    ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
+    ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
+    ALIGNED_ARRAY_16( dctcoef, dct2x2,[4] );
+    ALIGNED_ARRAY_16( dctcoef, dctscan,[16] );
+    ALIGNED_4( int16_t mvp[2] );
 
     int i_qp = h->mb.i_qp;
-    int mvp[2];
-    int ch, thresh, ssd;
-
-    int i8x8, i4x4;
-    int i_decimate_mb;
+    int thresh, ssd;
 
     if( !b_bidir )
     {
@@ -864,10 +895,10 @@ int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
         /* Motion compensation */
         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],
-                       mvp[0], mvp[1], 16, 16 );
+                       mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
     }
 
-    for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
+    for( int 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;
@@ -875,8 +906,10 @@ int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
         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++ )
+        for( int i4x4 = 0; i4x4 < 4; i4x4++ )
         {
+            if( h->mb.b_noise_reduction )
+                h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[0], h->nr_offset[0], 16 );
             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
                 continue;
             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
@@ -890,17 +923,26 @@ int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
     i_qp = h->mb.i_chroma_qp;
     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 
-    for( ch = 0; ch < 2; ch++ )
+    if( !b_bidir )
     {
-        uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
-        uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
-
-        if( !b_bidir )
-        {
-            h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
-                             h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
+        /* Special case for mv0, which is (of course) very common in P-skip mode. */
+        if( M32( mvp ) )
+            h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
+                             h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
                              mvp[0], mvp[1], 8, 8 );
-        }
+        else
+            h->mc.load_deinterleave_8x8x2_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1] );
+    }
+
+    for( int ch = 0; ch < 2; ch++ )
+    {
+        pixel *p_src = h->mb.pic.p_fenc[1+ch];
+        pixel *p_dst = h->mb.pic.p_fdec[1+ch];
+
+        if( !b_bidir && h->sh.weight[0][1+ch].weightfn )
+            h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
+                                                  h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
+                                                  &h->sh.weight[0][1+ch], 8 );
 
         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
         /* so instead we check SSD and skip the actual check if the score is low enough. */
@@ -908,10 +950,20 @@ int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
         if( ssd < thresh )
             continue;
 
-        h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
+        /* The vast majority of chroma checks will terminate during the DC check or the higher
+         * threshold check, so we can save time by doing a DC-only DCT. */
+        if( h->mb.b_noise_reduction )
+        {
+            h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
+            for( int i4x4 = 0; i4x4 < 4; i4x4++ )
+            {
+                h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[2], h->nr_offset[2], 16 );
+                dct2x2[i4x4] = dct4x4[i4x4][0];
+            }
+        }
+        else
+            h->dctf.sub8x8_dct_dc( dct2x2, p_src, p_dst );
 
-        /* calculate dct DC */
-        dct2x2dc( dct2x2, dct4x4 );
         if( h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 ) )
             return 0;
 
@@ -919,9 +971,15 @@ int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
         if( ssd < thresh*4 )
             continue;
 
+        if( !h->mb.b_noise_reduction )
+            h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
+
         /* calculate dct coeffs */
-        for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
+        for( int i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
         {
+            dct4x4[i4x4][0] = 0;
+            if( h->mb.b_noise_reduction )
+                h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[2], h->nr_offset[2], 16 );
             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
                 continue;
             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
@@ -942,24 +1000,30 @@ int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 
 void x264_noise_reduction_update( x264_t *h )
 {
-    int cat, i;
-    for( cat = 0; cat < 2; cat++ )
+    h->nr_offset = h->nr_offset_denoise;
+    h->nr_residual_sum = h->nr_residual_sum_buf[0];
+    h->nr_count = h->nr_count_buf[0];
+    for( int cat = 0; cat < 3; cat++ )
     {
-        int size = cat ? 64 : 16;
-        const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
+        int dct8x8 = cat == 1;
+        int size = dct8x8 ? 64 : 16;
+        const uint16_t *weight = dct8x8 ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 
-        if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
+        if( h->nr_count[cat] > (dct8x8 ? (1<<16) : (1<<18)) )
         {
-            for( i = 0; i < size; i++ )
+            for( int i = 0; i < size; i++ )
                 h->nr_residual_sum[cat][i] >>= 1;
             h->nr_count[cat] >>= 1;
         }
 
-        for( i = 0; i < size; i++ )
+        for( int i = 0; i < size; i++ )
             h->nr_offset[cat][i] =
                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
                  + h->nr_residual_sum[cat][i]/2)
               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
+
+        /* Don't denoise DC coefficients */
+        h->nr_offset[cat][0] = 0;
     }
 }
 
@@ -970,28 +1034,29 @@ void x264_noise_reduction_update( x264_t *h )
 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 {
     int i_qp = h->mb.i_qp;
-    uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
-    uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
-    int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
+    int x = i8&1;
+    int y = i8>>1;
+    int s8 = X264_SCAN8_0 + 2*x + 16*y;
+    pixel *p_fenc = h->mb.pic.p_fenc[0] + 8*x + 8*y*FENC_STRIDE;
+    pixel *p_fdec = h->mb.pic.p_fdec[0] + 8*x + 8*y*FDEC_STRIDE;
+    int b_decimate = h->mb.b_dct_decimate;
     int nnz8x8 = 0;
-    int ch, nz;
+    int nz;
 
     if( !h->mb.b_skip_mc )
         x264_mb_mc_8x8( h, i8 );
 
     if( h->mb.b_lossless )
     {
-        int i4;
         if( h->mb.b_transform_8x8 )
         {
             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
-            STORE_8x8_NNZ(i8,nnz8x8);
+            STORE_8x8_NNZ( s8, nnz8x8 );
         }
         else
         {
-            for( i4 = i8*4; i4 < i8*4+4; i4++ )
+            for( int i4 = i8*4; i4 < i8*4+4; i4++ )
             {
-                int nz;
                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
@@ -999,11 +1064,11 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
                 nnz8x8 |= nz;
             }
         }
-        for( ch = 0; ch < 2; ch++ )
+        for( int ch = 0; ch < 2; ch++ )
         {
-            int16_t dc;
-            p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
-            p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
+            dctcoef dc;
+            p_fenc = h->mb.pic.p_fenc[1+ch] + 4*x + 4*y*FENC_STRIDE;
+            p_fdec = h->mb.pic.p_fdec[1+ch] + 4*x + 4*y*FDEC_STRIDE;
             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
         }
@@ -1012,7 +1077,7 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
     {
         if( h->mb.b_transform_8x8 )
         {
-            ALIGNED_ARRAY_16( int16_t, dct8x8,[8],[8] );
+            ALIGNED_ARRAY_16( dctcoef, dct8x8,[64] );
             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
             if( nnz8x8 )
@@ -1026,21 +1091,20 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
                 {
                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
-                    STORE_8x8_NNZ(i8,1);
+                    STORE_8x8_NNZ( s8, 1 );
                 }
                 else
-                    STORE_8x8_NNZ(i8,0);
+                    STORE_8x8_NNZ( s8, 0 );
             }
             else
-                STORE_8x8_NNZ(i8,0);
+                STORE_8x8_NNZ( s8, 0 );
         }
         else
         {
-            int i4;
             int i_decimate_8x8 = 0;
-            ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[4][4] );
+            ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
-            for( i4 = 0; i4 < 4; i4++ )
+            for( int i4 = 0; i4 < 4; i4++ )
             {
                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
@@ -1060,19 +1124,20 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
             if( nnz8x8 )
                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
             else
-                STORE_8x8_NNZ(i8,0);
+                STORE_8x8_NNZ( s8, 0 );
         }
 
         i_qp = h->mb.i_chroma_qp;
 
-        for( ch = 0; ch < 2; ch++ )
+        for( int ch = 0; ch < 2; ch++ )
         {
-            ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[4] );
-            p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
-            p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
-
+            ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
+            p_fenc = h->mb.pic.p_fenc[1+ch] + 4*x + 4*y*FENC_STRIDE;
+            p_fdec = h->mb.pic.p_fdec[1+ch] + 4*x + 4*y*FDEC_STRIDE;
             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
-            dct4x4[0][0] = 0;
+            if( h->mb.b_noise_reduction )
+                h->quantf.denoise_dct( dct4x4, h->nr_residual_sum[2], h->nr_offset[2], 16 );
+            dct4x4[0] = 0;
 
             if( h->mb.b_trellis )
                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
@@ -1099,14 +1164,11 @@ void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
 {
     int i_qp = h->mb.i_qp;
-    uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
-    uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
-    const int i_ref = h->mb.cache.ref[0][x264_scan8[i4]];
-    const int mvx   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][0], h->mb.mv_min[0], h->mb.mv_max[0] );
-    const int mvy   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][1], h->mb.mv_min[1], h->mb.mv_max[1] );
+    pixel *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
+    pixel *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
     int nz;
 
-    h->mc.mc_luma( p_fdec, FDEC_STRIDE, h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0], mvx + 4*4*block_idx_x[i4], mvy + 4*4*block_idx_y[i4], 4, 4 );
+    /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
 
     if( h->mb.b_lossless )
     {
@@ -1115,7 +1177,7 @@ void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
     }
     else
     {
-        ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[4] );
+        ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;