Fix and enable I_PCM macroblock support

In RD mode, always consider PCM as a macroblock mode possibility
Fix bitstream writing for PCM blocks in CAVLC and CABAC, and a few other minor changes to make PCM work.
PCM macroblocks improve compression at very low QPs (1-5) and in lossless mode.

common/common.h

@@ -50,6 +50,7 @@
 #define X264_THREAD_MAX 128
 #define X264_SLICE_MAX 4
 #define X264_NAL_MAX (4 + X264_SLICE_MAX)
+#define X264_PCM_COST (386*8)
 
 // number of pixels (per thread) in progress at any given time.
 // 16 for the macroblock in progress + 3 for deblocking + 3 for motion compensation filter + 2 for extra safety

common/macroblock.c

@@ -1399,7 +1399,7 @@ void x264_macroblock_cache_save( x264_t *h )
 
     if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
         h->mb.i_qp = h->mb.i_last_qp;
-    h->mb.qp[i_mb_xy] = h->mb.i_qp;
+    h->mb.qp[i_mb_xy] = i_mb_type != I_PCM ? h->mb.i_qp : 0;
 
     h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
     h->mb.i_last_qp = h->mb.i_qp;
@@ -1418,7 +1418,10 @@ void x264_macroblock_cache_save( x264_t *h )
 
     if( i_mb_type == I_PCM )
     {
+        h->mb.i_cbp_chroma = 2;
+        h->mb.i_cbp_luma = 0xf;
         h->mb.cbp[i_mb_xy] = 0x72f;   /* all set */
+        h->mb.b_transform_8x8 = 0;
         for( i = 0; i < 16 + 2*4; i++ )
             non_zero_count[i] = 16;
     }

common/macroblock.h

@@ -55,7 +55,7 @@ static const uint8_t x264_pred_i4x4_neighbors[12] =
 
 
 /* XXX mb_type isn't the one written in the bitstream -> only internal usage */
-#define IS_INTRA(type) ( (type) == I_4x4 || (type) == I_8x8 || (type) == I_16x16 )
+#define IS_INTRA(type) ( (type) == I_4x4 || (type) == I_8x8 || (type) == I_16x16 || (type) == I_PCM )
 #define IS_SKIP(type)  ( (type) == P_SKIP || (type) == B_SKIP )
 #define IS_DIRECT(type)  ( (type) == B_DIRECT )
 enum mb_class_e

encoder/analyse.c

@@ -96,6 +96,8 @@ typedef struct
     int i_satd_i4x4;
     int i_predict4x4[16];
 
+    int i_satd_pcm;
+
     /* Chroma part */
     int i_satd_i8x8chroma;
     int i_satd_i8x8chroma_dir[4];
@@ -223,6 +225,9 @@ static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp )
     a->i_satd_i4x4   =
     a->i_satd_i8x8chroma = COST_MAX;
 
+    /* non-RD PCM decision is inaccurate, so don't do it */
+    a->i_satd_pcm = a->b_mbrd ? ((uint64_t)X264_PCM_COST*a->i_lambda2 + 128) >> 8 : COST_MAX;
+
     a->b_fast_intra = 0;
     h->mb.i_skip_intra =
         h->mb.b_lossless ? 0 :
@@ -2066,15 +2071,12 @@ void x264_macroblock_analyse( x264_t *h )
 
         i_cost = analysis.i_satd_i16x16;
         h->mb.i_type = I_16x16;
-        if( analysis.i_satd_i4x4 < i_cost )
-        {
-            i_cost = analysis.i_satd_i4x4;
-            h->mb.i_type = I_4x4;
-        }
-        if( analysis.i_satd_i8x8 < i_cost )
-            h->mb.i_type = I_8x8;
+        COPY2_IF_LT( i_cost, analysis.i_satd_i4x4, h->mb.i_type, I_4x4 );
+        COPY2_IF_LT( i_cost, analysis.i_satd_i8x8, h->mb.i_type, I_8x8 );
+        if( analysis.i_satd_pcm < i_cost )
+            h->mb.i_type = I_PCM;
 
-        if( h->mb.i_subpel_refine >= 7 )
+        else if( h->mb.i_subpel_refine >= 7 )
             x264_intra_rd_refine( h, &analysis );
     }
     else if( h->sh.i_type == SLICE_TYPE_P )
@@ -2285,6 +2287,7 @@ void x264_macroblock_analyse( x264_t *h )
             i_intra_cost = analysis.i_satd_i16x16;
             COPY2_IF_LT( i_intra_cost, analysis.i_satd_i8x8, i_intra_type, I_8x8 );
             COPY2_IF_LT( i_intra_cost, analysis.i_satd_i4x4, i_intra_type, I_4x4 );
+            COPY2_IF_LT( i_intra_cost, analysis.i_satd_pcm, i_intra_type, I_PCM );
             COPY2_IF_LT( i_cost, i_intra_cost, i_type, i_intra_type );
 
             if( i_intra_cost == COST_MAX )
@@ -2295,7 +2298,7 @@ void x264_macroblock_analyse( x264_t *h )
             h->stat.frame.i_inter_cost += i_cost;
             h->stat.frame.i_mbs_analysed++;
 
-            if( h->mb.i_subpel_refine >= 7 )
+            if( h->mb.i_subpel_refine >= 7 && h->mb.i_type != I_PCM )
             {
                 if( IS_INTRA( h->mb.i_type ) )
                 {
@@ -2566,11 +2569,12 @@ void x264_macroblock_analyse( x264_t *h )
             COPY2_IF_LT( i_cost, analysis.i_satd_i16x16, i_type, I_16x16 );
             COPY2_IF_LT( i_cost, analysis.i_satd_i8x8, i_type, I_8x8 );
             COPY2_IF_LT( i_cost, analysis.i_satd_i4x4, i_type, I_4x4 );
+            COPY2_IF_LT( i_cost, analysis.i_satd_pcm, i_type, I_PCM );
 
             h->mb.i_type = i_type;
             h->mb.i_partition = i_partition;
 
-            if( analysis.b_mbrd && h->mb.i_subpel_refine >= 7 && IS_INTRA( i_type ) )
+            if( analysis.b_mbrd && h->mb.i_subpel_refine >= 7 && IS_INTRA( i_type ) && i_type != I_PCM )
                 x264_intra_rd_refine( h, &analysis );
             else if( h->param.analyse.b_bidir_me )
                 refine_bidir( h, &analysis );
@@ -2612,6 +2616,9 @@ static void x264_analyse_update_cache( x264_t *h, x264_mb_analysis_t *a  )
             x264_mb_analyse_intra_chroma( h, a );
             break;
 
+        case I_PCM:
+            break;
+
         case P_L0:
             switch( h->mb.i_partition )
             {

encoder/cabac.c

@@ -543,14 +543,12 @@ static int x264_cabac_mb_cbf_ctxidxinc( x264_t *h, int i_cat, int i_idx )
         if( h->mb.i_neighbour & MB_LEFT )
         {
             i_mba_xy = h->mb.i_mb_xy - 1;
-            if( h->mb.i_mb_type_left == I_16x16 )
-                i_nza = h->mb.cbp[i_mba_xy] & 0x100;
+            i_nza = h->mb.cbp[i_mba_xy] & 0x100;
         }
         if( h->mb.i_neighbour & MB_TOP )
         {
             i_mbb_xy = h->mb.i_mb_top_xy;
-            if( h->mb.i_mb_type_top == I_16x16 )
-                i_nzb = h->mb.cbp[i_mbb_xy] & 0x100;
+            i_nzb = h->mb.cbp[i_mbb_xy] & 0x100;
         }
     }
     else if( i_cat == DCT_LUMA_AC || i_cat == DCT_LUMA_4x4 )
@@ -785,36 +783,35 @@ void x264_macroblock_write_cabac( x264_t *h, x264_cabac_t *cb )
     /* Write the MB type */
     x264_cabac_mb_type( h, cb );
 
-    /* PCM special block type UNTESTED */
+#ifndef RDO_SKIP_BS
     if( i_mb_type == I_PCM )
     {
-#ifdef RDO_SKIP_BS
-        cb->f8_bits_encoded += (384*8) << 8;
-#else
-        if( cb->p + 385 >= cb->p_end )
-            return; //FIXME throw an error
-        /* Luma */
-        for( i = 0; i < 16; i++ )
-        {
-            memcpy( cb->p, h->fenc->plane[0] + i*h->mb.pic.i_stride[0], 16 );
-            cb->p += 16;
-        }
-        /* Cb */
+        i_mb_pos_tex = x264_cabac_pos( cb );
+        h->stat.frame.i_hdr_bits += i_mb_pos_tex - i_mb_pos_start;
+
+        memcpy( cb->p, h->mb.pic.p_fenc[0], 256 );
+        cb->p += 256;
         for( i = 0; i < 8; i++ )
-        {
-            memcpy( cb->p, h->fenc->plane[1] + i*h->mb.pic.i_stride[1], 8 );
-            cb->p += 8;
-        }
-        /* Cr */
+            memcpy( cb->p + i*8, h->mb.pic.p_fenc[1] + i*FENC_STRIDE, 8 );
+        cb->p += 64;
         for( i = 0; i < 8; i++ )
-        {
-            memcpy( cb->p, h->fenc->plane[2] + i*h->mb.pic.i_stride[2], 8 );
-            cb->p += 8;
-        }
-        x264_cabac_encode_init( cb, cb->p, cb->p_end );
-#endif
+            memcpy( cb->p + i*8, h->mb.pic.p_fenc[2] + i*FENC_STRIDE, 8 );
+        cb->p += 64;
+
+        cb->i_low   = 0;
+        cb->i_range = 0x01FE;
+        cb->i_queue = -1;
+        cb->i_bytes_outstanding = 0;
+
+        /* 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 );
+
+        h->stat.frame.i_itex_bits += x264_cabac_pos( cb ) - i_mb_pos_tex;
         return;
     }
+#endif
 
     if( IS_INTRA( i_mb_type ) )
     {

encoder/cavlc.c

@@ -339,44 +339,39 @@ void x264_macroblock_write_cavlc( x264_t *h, bs_t *s )
         bs_write1( s, h->mb.b_interlaced );
     }
 
-    /* Write:
-      - type
-      - prediction
-      - mv */
-    if( i_mb_type == I_PCM )
+#ifndef RDO_SKIP_BS
+    if( i_mb_type == I_PCM)
     {
-        /* Untested */
         bs_write_ue( s, i_mb_i_offset + 25 );
+        i_mb_pos_tex = bs_pos( s );
+        h->stat.frame.i_hdr_bits += i_mb_pos_tex - i_mb_pos_start;
 
-#ifdef RDO_SKIP_BS
-        s->i_bits_encoded += 384*8;
-#else
         bs_align_0( s );
-        /* Luma */
-        for( i = 0; i < 16*16; i++ )
-        {
-            const int x = 16 * h->mb.i_mb_x + (i % 16);
-            const int y = 16 * h->mb.i_mb_y + (i / 16);
-            bs_write( s, 8, h->fenc->plane[0][y*h->mb.pic.i_stride[0]+x] );
-        }
-        /* Cb */
-        for( i = 0; i < 8*8; i++ )
-        {
-            const int x = 8 * h->mb.i_mb_x + (i % 8);
-            const int y = 8 * h->mb.i_mb_y + (i / 8);
-            bs_write( s, 8, h->fenc->plane[1][y*h->mb.pic.i_stride[1]+x] );
-        }
-        /* Cr */
-        for( i = 0; i < 8*8; i++ )
-        {
-            const int x = 8 * h->mb.i_mb_x + (i % 8);
-            const int y = 8 * h->mb.i_mb_y + (i / 8);
-            bs_write( s, 8, h->fenc->plane[2][y*h->mb.pic.i_stride[2]+x] );
-        }
-#endif
+
+        memcpy( s->p, h->mb.pic.p_fenc[0], 256 );
+        s->p += 256;
+        for( i = 0; i < 8; i++ )
+            memcpy( s->p + i*8, h->mb.pic.p_fenc[1] + i*FENC_STRIDE, 8 );
+        s->p += 64;
+        for( i = 0; i < 8; i++ )
+            memcpy( s->p + i*8, h->mb.pic.p_fenc[2] + i*FENC_STRIDE, 8 );
+        s->p += 64;
+
+        /* 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 );
+
+        h->stat.frame.i_itex_bits += bs_pos(s) - i_mb_pos_tex;
         return;
     }
-    else if( i_mb_type == I_4x4 || i_mb_type == I_8x8 )
+#endif
+
+    /* Write:
+      - type
+      - prediction
+      - mv */
+    if( i_mb_type == I_4x4 || i_mb_type == I_8x8 )
     {
         int di = i_mb_type == I_8x8 ? 4 : 1;
         bs_write_ue( s, i_mb_i_offset + 0 );

encoder/encoder.c

@@ -1841,10 +1841,11 @@ void    x264_encoder_close  ( x264_t *h )
         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
         x264_log( h, X264_LOG_INFO,
-                  "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
+                  "mb I  I16..4..PCM: %4.1f%% %4.1f%% %4.1f%% %4.1f%%\n",
                   i_mb_count[I_16x16]/ i_count,
                   i_mb_count[I_8x8]  / i_count,
-                  i_mb_count[I_4x4]  / i_count );
+                  i_mb_count[I_4x4]  / i_count,
+                  i_mb_count[I_PCM]  / i_count );
     }
     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
     {
@@ -1852,10 +1853,11 @@ void    x264_encoder_close  ( x264_t *h )
         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
         x264_log( h, X264_LOG_INFO,
-                  "mb P  I16..4: %4.1f%% %4.1f%% %4.1f%%  P16..4: %4.1f%% %4.1f%% %4.1f%% %4.1f%% %4.1f%%    skip:%4.1f%%\n",
+                  "mb P  I16..4..PCM: %4.1f%% %4.1f%% %4.1f%% %4.1f%%  P16..4: %4.1f%% %4.1f%% %4.1f%% %4.1f%% %4.1f%%    skip:%4.1f%%\n",
                   i_mb_count[I_16x16]/ i_count,
                   i_mb_count[I_8x8]  / i_count,
                   i_mb_count[I_4x4]  / i_count,
+                  i_mb_count[I_PCM]  / i_count,
                   i_mb_size[PIXEL_16x16] / (i_count*4),
                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
                   i_mb_size[PIXEL_8x8] / (i_count*4),
@@ -1869,10 +1871,11 @@ void    x264_encoder_close  ( x264_t *h )
         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
         x264_log( h, X264_LOG_INFO,
-                  "mb B  I16..4: %4.1f%% %4.1f%% %4.1f%%  B16..8: %4.1f%% %4.1f%% %4.1f%%  direct:%4.1f%%  skip:%4.1f%%\n",
+                  "mb B  I16..4..PCM: %4.1f%% %4.1f%% %4.1f%% %4.1f%%  B16..8: %4.1f%% %4.1f%% %4.1f%%  direct:%4.1f%%  skip:%4.1f%%\n",
                   i_mb_count[I_16x16]  / i_count,
                   i_mb_count[I_8x8]    / i_count,
                   i_mb_count[I_4x4]    / i_count,
+                  i_mb_count[I_PCM]    / i_count,
                   i_mb_size[PIXEL_16x16] / (i_count*4),
                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
                   i_mb_size[PIXEL_8x8] / (i_count*4),