Merge branch 'cross_comp_pred_444' into 444-format-revive

Author: fador

.gitlab-ci.yml

@@ -63,3 +63,8 @@ test-10bit:
   <<: *test-template
   variables:
     CFLAGS: '-DKVZ_BIT_DEPTH=10'
+    
+test-12bit:
+  <<: *test-template
+  variables:
+    CFLAGS: '-DKVZ_BIT_DEPTH=12'

README.md

@@ -77,7 +77,7 @@ Input:
                                    - progressive: Progressive scan
                                    - tff: Top field first
                                    - bff: Bottom field first
-      --input-format <string> : P420 or P400 [P420]
+      --input-format <string> : P420, P444 or P400 [P420]
       --input-bitdepth <int> : 8-16 [8]
       --loop-input           : Re-read input file forever.
       --input-file-format <string> : Input file format [auto]
@@ -293,6 +293,7 @@ Compression tools:
       --(no-)implicit-rdpcm  : Implicit residual DPCM. Currently only supported
                                with lossless coding. [disabled]
       --(no-)tmvp            : Temporal motion vector prediction [enabled]
+      --(no-)cross-comp-pred : Cross-component prediction for 4:4:4 [disabled]
 
 Parallel processing:
       --threads <integer>    : Number of threads to use [auto]

doc/kvazaar.1

@@ -44,7 +44,7 @@ Source scan type [progressive]
     \- bff: Bottom field first
 .TP
 \fB\-\-input\-format <string>
-P420 or P400 [P420]
+P420, P444 or P400 [P420]
 .TP
 \fB\-\-input\-bitdepth <int>
 8\-16 [8]
@@ -403,6 +403,9 @@ with lossless coding. [disabled]
 .TP
 \fB\-\-(no\-)tmvp           
 Temporal motion vector prediction [enabled]
+.TP
+\fB\-\-(no\-)cross\-comp\-pred
+Cross\-component prediction for 4:4:4 [disabled]
 
 .SS "Parallel processing:"
 .TP

src/cabac.h

@@ -97,6 +97,7 @@ typedef struct
     cabac_ctx_t cu_qt_root_cbf_model;
     cabac_ctx_t transform_skip_model_luma;
     cabac_ctx_t transform_skip_model_chroma;
+    cabac_ctx_t cross_component_prediction[10];
   } ctx;
 } cabac_data_t;
 

src/cfg.c

@@ -197,6 +197,8 @@ int kvz_config_init(kvz_config *cfg)
   cfg->chroma_shift = cfg->chroma_shift_w = 1;
   cfg->chroma_shift_h = 1;
 
+  cfg->enable_cross_component_prediction = 0;
+
   return 1;
 }
 
@@ -1442,6 +1444,9 @@ int kvz_config_parse(kvz_config *cfg, const char *name, const char *value)
   else if OPT("enable-logging") {
     cfg->enable_logging_output = atobool(value);
   }
+  else if OPT("cross-comp-pred") {
+    cfg->enable_cross_component_prediction = atobool(value);
+  }
   else {
     return 0;
   }
@@ -1785,6 +1790,16 @@ int kvz_config_validate(const kvz_config *const cfg)
     error = 1;
   }
 
+  if (cfg->enable_cross_component_prediction && cfg->chroma_format != KVZ_CSP_444) {
+    fprintf(stderr, "Cross-component prediction is only supported in 4:4:4 chroma format.\n");
+    error = 1;
+  }
+
+  if (cfg->chroma_format == KVZ_CSP_422) {
+    fprintf(stderr, "4:2:2 chroma format is not supported yet.\n");
+    error = 1;
+  }
+
   return !error;
 }
 

src/cli.c

@@ -178,6 +178,8 @@ static const struct option long_options[] = {
   { "no-fast-bipred",           no_argument, NULL, 0 },
   { "enable-logging",           no_argument, NULL, 0 },
   { "no-enable-logging",        no_argument, NULL, 0 },
+  { "cross-comp-pred",          no_argument, NULL, 0 },
+  { "no-cross-comp-pred",       no_argument, NULL, 0 },
   {0, 0, 0, 0}
 };
 
@@ -653,6 +655,7 @@ void print_help(void)
     "      --(no-)implicit-rdpcm  : Implicit residual DPCM. Currently only supported\n"
     "                               with lossless coding. [disabled]\n"
     "      --(no-)tmvp            : Temporal motion vector prediction [enabled]\n"
+    "      --(no-)cross-comp-pred : Cross-component prediction for 4:4:4 [disabled]\n"
     "\n"
     /* Word wrap to this width to stay under 80 characters (including ") *************/
     "Parallel processing:\n"

src/context.c

@@ -193,7 +193,12 @@ static const uint8_t INIT_TRANSFORMSKIP_FLAG[3][2] =
   { 139,  139},
 };
 
-
+static const uint8_t INIT_CROSS_COMPONENT_PREDICTION[3][10] =
+{
+  { 154, 154, 154, 154, 154, 154, 154, 154, 154, 154 },
+  { 154, 154, 154, 154, 154, 154, 154, 154, 154, 154 },
+  { 154, 154, 154, 154, 154, 154, 154, 154, 154, 154 },
+};
 
 
 /**
@@ -285,6 +290,12 @@ void kvz_init_contexts(encoder_state_t *state, int8_t QP, int8_t slice)
     kvz_ctx_init(&cabac->ctx.cu_one_model_chroma[i], QP, INIT_ONE_FLAG[slice][i+16]);
   }
 
+  if (state->encoder_control->cfg.chroma_format == KVZ_CSP_444) {
+    for (i = 0; i < 10; i++) {
+      kvz_ctx_init(&cabac->ctx.cross_component_prediction[i], QP, INIT_CROSS_COMPONENT_PREDICTION[slice][i]);
+    }
+  }
+
   for (i = 0; i < 15; i++) {
     kvz_ctx_init(&cabac->ctx.cu_ctx_last_y_luma[i], QP, INIT_LAST[slice][i] );
     kvz_ctx_init(&cabac->ctx.cu_ctx_last_x_luma[i], QP, INIT_LAST[slice][i] );

src/cu.h

@@ -128,14 +128,18 @@ typedef struct {
  */
 typedef struct
 {
-  uint8_t type      : 2; //!< \brief block type, one of cu_type_t values
-  uint8_t depth     : 3; //!< \brief depth / size of this block
-  uint8_t part_size : 3; //!< \brief partition mode, one of part_mode_t values
-  uint8_t tr_depth  : 3; //!< \brief transform depth
-  uint8_t skipped   : 1; //!< \brief flag to indicate this block is skipped
-  uint8_t merged    : 1; //!< \brief flag to indicate this block is merged
-  uint8_t merge_idx : 3; //!< \brief merge index
-  uint8_t tr_skip   : 1; //!< \brief transform skip flag
+  uint8_t type       : 2; //!< \brief block type, one of cu_type_t values
+  uint8_t depth      : 3; //!< \brief depth / size of this block
+  uint8_t part_size  : 3; //!< \brief partition mode, one of part_mode_t values
+  uint8_t tr_depth   : 3; //!< \brief transform depth
+  uint8_t skipped    : 1; //!< \brief flag to indicate this block is skipped
+  uint8_t merged     : 1; //!< \brief flag to indicate this block is merged
+  uint8_t merge_idx  : 3; //!< \brief merge index
+  uint8_t tr_skip    : 1; //!< \brief transform skip flag
+  uint8_t alpha_u_s  : 1; // !< \brief cross-component prediction sign flag for U channel
+  uint8_t alpha_v_s  : 1; // !< \brief cross-component prediction sign flag for V channel
+  uint8_t alpha_u    : 3; // !< \brief cross-component prediction value for U channel
+  uint8_t alpha_v    : 3; // !< \brief cross-component prediction value for V channel
 
   uint16_t cbf;
 

src/encode_coding_tree.c

@@ -47,6 +47,30 @@
 #include "tables.h"
 #include "videoframe.h"
 
+/*
+* \brief Encode cross component prediction for U or V channel
+*/
+static void encode_cross_component_prediction(const cu_info_t* cur_pu, cabac_data_t* const cabac, color_t channel)
+{
+  cabac_ctx_t* ctx = &(cabac->ctx.cross_component_prediction[channel == COLOR_V ? 5 : 0]);
+  cabac->cur_ctx = ctx;
+  int8_t alpha = (channel == COLOR_V ? cur_pu->alpha_v : cur_pu->alpha_u);
+  int8_t alpha_sign = (channel == COLOR_V ? cur_pu->alpha_v_s : cur_pu->alpha_u_s);
+  
+  CABAC_BIN(cabac, (alpha != 0)? 1: 0, "cross_component_prediction_flag");
+
+  if(alpha != 0) {
+    alpha--;
+    cabac->cur_ctx = &ctx[1];
+    CABAC_BIN(cabac, (alpha > 0)? 1: 0, "cross_component_prediction_alpha");
+    if(alpha > 0) {
+      kvz_cabac_write_unary_max_symbol(cabac, &ctx[2], alpha - 1, 1, 2, NULL);      
+    }
+    cabac->cur_ctx = &ctx[4];
+    CABAC_BIN(cabac, alpha_sign, "cross_component_prediction_sign");
+  }
+}
+
 /**
  * \brief Encode (X,Y) position of the last significant coefficient
  *
@@ -163,24 +187,33 @@ static void encode_transform_unit(encoder_state_t * const state,
     }
   }
 
-  bool chroma_cbf_set = cbf_is_set(cur_pu->cbf, depth, COLOR_U) ||
-                        cbf_is_set(cur_pu->cbf, depth, COLOR_V);
-  if (chroma_cbf_set) {
+  if (state->encoder_control->cfg.chroma_format != KVZ_CSP_400) {
     int x_local = (x >> SHIFT_W) % (LCU_WIDTH >> SHIFT_W);
     int y_local = (y >> SHIFT_H) % (LCU_WIDTH >> SHIFT_H);
     scan_idx = kvz_get_scan_order(cur_pu->type, cur_pu->intra.mode_chroma, depth);
 
-    const coeff_t *coeff_u = &state->coeff->u[xy_to_zorder(LCU_WIDTH >> SHIFT_W, x_local, y_local)];
-    const coeff_t *coeff_v = &state->coeff->v[xy_to_zorder(LCU_WIDTH >> SHIFT_W, x_local, y_local)];
+    bool cross_component_prediction = cbf_y && state->encoder_control->cfg.enable_cross_component_prediction &&
+      (cur_pu->type == CU_INTER || cur_pu->intra.mode_chroma == cur_pu->intra.mode);
+
+    if (cross_component_prediction) {
+      encode_cross_component_prediction(cur_pu, &state->cabac, COLOR_U);
+    }
 
     if (cbf_is_set(cur_pu->cbf, depth, COLOR_U)) {
+      const coeff_t* coeff_u = &state->coeff->u[xy_to_zorder(LCU_WIDTH >> SHIFT_W, x_local, y_local)];
       kvz_encode_coeff_nxn(state, &state->cabac, coeff_u, width_c, 2, scan_idx, 0, NULL);
     }
 
+    if (cross_component_prediction) {
+      encode_cross_component_prediction(cur_pu, &state->cabac, COLOR_V);
+    }
+
     if (cbf_is_set(cur_pu->cbf, depth, COLOR_V)) {
+      const coeff_t* coeff_v = &state->coeff->v[xy_to_zorder(LCU_WIDTH >> SHIFT_W, x_local, y_local)];
       kvz_encode_coeff_nxn(state, &state->cabac, coeff_v, width_c, 2, scan_idx, 0, NULL);
     }
   }
+  
 }
 
 /**

src/encoder_state-bitstream.c

@@ -76,7 +76,16 @@ static void encoder_state_write_bitstream_PTL(bitstream_t *stream,
   WRITE_U(stream, 0, 2, "general_profile_space");
   WRITE_U(stream, state->encoder_control->cfg.high_tier, 1, "general_tier_flag");
   // Main Profile == 1,  Main 10 profile == 2
-  WRITE_U(stream, (state->encoder_control->bitdepth == 8) ? 1 : 2, 5, "general_profile_idc");
+  int8_t profile = 1;
+  if (state->encoder_control->cfg.chroma_format == KVZ_CSP_444 || state->encoder_control->cfg.chroma_format == KVZ_CSP_422 ||
+      state->encoder_control->bitdepth > 10) {
+    profile = 4; // 4:2:2 and 4:4:4 range extension profiles
+  } else if (state->encoder_control->bitdepth == 10) {
+    profile = 2; // Main 10 profile
+  } else {
+    profile = 1; // Main profile
+  }
+  WRITE_U(stream, profile, 5, "general_profile_idc");
   /* Compatibility flags should be set at general_profile_idc
    *  (so with general_profile_idc = 1, compatibility_flag[1] should be 1)
    * According to specification, when compatibility_flag[1] is set,
@@ -358,6 +367,36 @@ static void encoder_state_write_bitstream_SPS_extension(bitstream_t *stream,
   }
 }
 
+static void encoder_state_write_bitstream_PPS_extension(bitstream_t* stream,
+                                                        encoder_state_t* const state)
+{
+  const kvz_config* cfg = &state->encoder_control->cfg;
+  bool need_pps_extension = cfg->enable_cross_component_prediction;
+  WRITE_U(stream, need_pps_extension, 1, "pps_extension_present_flag");
+  if (need_pps_extension) {
+    WRITE_U(stream, 1, 1, "pps_range_extension_flag");
+    WRITE_U(stream, 0, 1, "pps_multilayer_extension_flag");
+    WRITE_U(stream, 0, 1, "pps_3d_extension_flag");
+    WRITE_U(stream, 0, 1, "pps_scc_extension_flag");
+    WRITE_U(stream, 0, 4, "pps_extension_4bits");
+
+    // pps_range_extension_flag
+    if(cfg->trskip_enable) WRITE_UE(stream, 0, "log2_max_transform_skip_block_size_minus2");
+    WRITE_U(stream, cfg->enable_cross_component_prediction, 1, "cross_component_prediction_enabled_flag");
+    WRITE_U(stream, 0, 1, "chroma_qp_offset_list_enabled_flag");
+    //IF chroma_qp_offset_list_enabled_flag
+      //WRITE_UE(stream, 0, "diff_cu_chroma_qp_offset_depth");
+      //WRITE_UE(stream, 0, "chroma_qp_offset_list_len_minus1");
+      //for (i = 0; i <= chroma_qp_offset_list_len_minus1; i++)
+        //WRITE_SE(stream, 0, "cb_qp_offset_list[i]");
+        //WRITE_SE(stream, 0, "cr_qp_offset_list[i]");
+      //end for
+    //ENDIF
+    WRITE_UE(stream, 0, "log2_sao_offset_scale_luma");
+    WRITE_UE(stream, 0, "log2_sao_offset_scale_chroma");
+  }
+}
+
 static void encoder_state_write_bitstream_seq_parameter_set(bitstream_t* stream,
                                                             encoder_state_t * const state)
 {
@@ -569,7 +608,8 @@ static void encoder_state_write_bitstream_pic_parameter_set(bitstream_t* stream,
   WRITE_U(stream, 0, 1, "lists_modification_present_flag");
   WRITE_UE(stream, 0, "log2_parallel_merge_level_minus2");
   WRITE_U(stream, 0, 1, "slice_segment_header_extension_present_flag");
-  WRITE_U(stream, 0, 1, "pps_extension_flag");
+
+  encoder_state_write_bitstream_PPS_extension(stream, state);
 
   kvz_bitstream_add_rbsp_trailing_bits(stream);
 }