Merge branch 'improve_rdo_2022'

README.md

@@ -200,18 +200,22 @@ Compression tools:
       --(no-)signhide        : Sign hiding [disabled]
       --(no-)smp             : Symmetric motion partition [disabled]
       --(no-)amp             : Asymmetric motion partition [disabled]
-      --rd <integer>         : Intra mode search complexity [0]
+      --rd <integer>         : Mode search complexity [0]
                                    - 0: Skip intra if inter is good enough.
                                    - 1: Rough intra mode search with SATD.
-                                   - 2: Refine intra mode search with SSE.
-                                   - 3: Try all intra modes and enable intra
-                                        chroma mode search.
+                                   - 2: Refine mode search with SSE.
+                                   - 3: More SSE candidates for inter and
+                                        chroma mode search for 4x4 intra.
+                                   - 4: Even more SSE candidates for both.
+                                   - 5: Try all intra modes.
       --(no-)mv-rdo          : Rate-distortion optimized motion vector costs
                                [disabled]
       --(no-)zero-coeff-rdo  : If a CU is set inter, check if forcing zero
                                residual improves the RD cost. [enabled]
       --(no-)full-intra-search : Try all intra modes during rough search.
                                [disabled]
+      --(no-)intra-chroma-search : Test non-derived intra chroma modes.
+                                   [disabled]
       --(no-)transform-skip  : Try transform skip [disabled]
       --me <string>          : Integer motion estimation algorithm [hexbs]
                                    - hexbs: Hexagon Based Search
@@ -227,6 +231,7 @@ Compression tools:
                                    - 2: + 1/2-pixel diagonal
                                    - 3: + 1/4-pixel horizontal and vertical
                                    - 4: + 1/4-pixel diagonal
+      --(no-)fast-bipred     : Only perform fast bipred search. [enabled]
       --pu-depth-inter <int>-<int> : Inter prediction units sizes [0-3]
                                    - 0, 1, 2, 3: from 64x64 to 8x8
                                    - Accepts a list of values separated by ','

configure.ac

@@ -23,7 +23,7 @@ AC_CONFIG_SRCDIR([src/encmain.c])
 #
 # Here is a somewhat sane guide to lib versioning: http://apr.apache.org/versioning.html
 ver_major=7
-ver_minor=1
+ver_minor=2
 ver_release=0
 
 # Prevents configure from adding a lot of defines to the CFLAGS

doc/kvazaar.1

@@ -1,4 +1,4 @@
-.TH KVAZAAR "1" "June 2022" "kvazaar v2.1.0" "User Commands"
+.TH KVAZAAR "1" "September 2022" "kvazaar v2.1.0" "User Commands"
 .SH NAME
 kvazaar \- open source HEVC encoder
 .SH SYNOPSIS
@@ -252,12 +252,14 @@ Symmetric motion partition [disabled]
 Asymmetric motion partition [disabled]
 .TP
 \fB\-\-rd <integer>        
-Intra mode search complexity [0]
+Mode search complexity [0]
     \- 0: Skip intra if inter is good enough.
     \- 1: Rough intra mode search with SATD.
-    \- 2: Refine intra mode search with SSE.
-    \- 3: Try all intra modes and enable intra
-         chroma mode search.
+    \- 2: Refine mode search with SSE.
+    \- 3: More SSE candidates for inter and
+         chroma mode search for 4x4 intra.
+    \- 4: Even more SSE candidates for both.
+    \- 5: Try all intra modes.
 .TP
 \fB\-\-(no\-)mv\-rdo         
 Rate\-distortion optimized motion vector costs
@@ -271,6 +273,10 @@ residual improves the RD cost. [enabled]
 Try all intra modes during rough search.
 [disabled]
 .TP
+\fB\-\-(no\-)intra\-chroma\-search
+Test non\-derived intra chroma modes.
+    [disabled]
+.TP
 \fB\-\-(no\-)transform\-skip 
 Try transform skip [disabled]
 .TP
@@ -294,6 +300,9 @@ Fractional pixel motion estimation level [4]
     \- 3: + 1/4\-pixel horizontal and vertical
     \- 4: + 1/4\-pixel diagonal
 .TP
+\fB\-\-(no\-)fast\-bipred    
+Only perform fast bipred search. [enabled]
+.TP
 \fB\-\-pu\-depth\-inter <int>\-<int>
 Inter prediction units sizes [0\-3]
     \- 0, 1, 2, 3: from 64x64 to 8x8

src/cabac.c

@@ -272,15 +272,19 @@ void kvz_cabac_encode_bins_ep(cabac_data_t * const data, uint32_t bin_values, in
  * \param symbol Value of coeff_abs_level_minus3.
  * \param r_param Reference to Rice parameter.
  */
-void kvz_cabac_write_coeff_remain(cabac_data_t * const cabac, const uint32_t symbol, const uint32_t r_param)
+int kvz_cabac_write_coeff_remain(cabac_data_t* const cabac, const uint32_t symbol, const uint32_t r_param)
 {
   int32_t code_number = symbol;
   uint32_t length;
 
+  int bits = 0;
+
   if (code_number < (3 << r_param)) {
     length = code_number >> r_param;
     CABAC_BINS_EP(cabac, (1 << (length + 1)) - 2 , length + 1, "coeff_abs_level_remaining");
+    bits += length + 1;
     CABAC_BINS_EP(cabac, (code_number % (1 << r_param)), r_param, "coeff_abs_level_remaining");
+    bits += r_param;
   } else {
     length = r_param;
     code_number = code_number - (3 << r_param);
@@ -289,8 +293,11 @@ void kvz_cabac_write_coeff_remain(cabac_data_t * const cabac, const uint32_t sym
       ++length;
     }
     CABAC_BINS_EP(cabac, (1 << (3 + length + 1 - r_param)) - 2, 3 + length + 1 - r_param, "coeff_abs_level_remaining");
+    bits += 3 + length + 1 - r_param;
     CABAC_BINS_EP(cabac, code_number, length, "coeff_abs_level_remaining");
+    bits += length;
   }
+  return bits;
 }
 
 void kvz_cabac_write_coeff_remain_encry(struct encoder_state_t * const state, cabac_data_t * const cabac,const uint32_t symbol, const uint32_t r_param, int32_t base_level)

src/cabac.h

@@ -116,8 +116,8 @@ void kvz_cabac_encode_bins_ep(cabac_data_t *data, uint32_t bin_values, int num_b
 void kvz_cabac_encode_bin_trm(cabac_data_t *data, uint8_t bin_value);
 void kvz_cabac_write(cabac_data_t *data);
 void kvz_cabac_finish(cabac_data_t *data);
-void kvz_cabac_write_coeff_remain(cabac_data_t *cabac, uint32_t symbol,
-                              uint32_t r_param);
+int kvz_cabac_write_coeff_remain(cabac_data_t* cabac, uint32_t symbol,
+                                 uint32_t r_param);
 void kvz_cabac_write_coeff_remain_encry(struct encoder_state_t * const state, cabac_data_t * const cabac, const uint32_t symbol,
                                         const uint32_t r_param, int32_t base_level);
 uint32_t kvz_cabac_write_ep_ex_golomb(struct encoder_state_t * const state, cabac_data_t *data,

src/cfg.c

@@ -186,6 +186,8 @@ int kvz_config_init(kvz_config *cfg)
 
   cfg->combine_intra_cus = 1;
   cfg->force_inter = 0;
+  cfg->intra_chroma_search = 0;
+  cfg->fast_bipred = 1;
   return 1;
 }
 
@@ -703,7 +705,7 @@ int kvz_config_parse(kvz_config *cfg, const char *name, const char *value)
       },
       {
         "veryslow",
-        "rd", "2",
+        "rd", "3",
         "pu-depth-intra", "1-4",
         "pu-depth-inter", "0-3",
         "me", "tz",
@@ -731,7 +733,7 @@ int kvz_config_parse(kvz_config *cfg, const char *name, const char *value)
       },
       {
         "placebo",
-        "rd", "2",
+        "rd", "3",
         "pu-depth-intra", "1-4",
         "pu-depth-inter", "0-3",
         "me", "tz",
@@ -1399,6 +1401,12 @@ int kvz_config_parse(kvz_config *cfg, const char *name, const char *value)
   else if OPT("force-inter") {
     cfg->force_inter = atobool(value);
   }
+  else if OPT("intra-chroma-search") {
+    cfg->intra_chroma_search = atobool(value);
+  }
+  else if OPT("fast-bipred") {
+    cfg->fast_bipred = atobool(value);
+  }
   else {
     return 0;
   }
@@ -1590,8 +1598,8 @@ int kvz_config_validate(const kvz_config *const cfg)
     error = 1;
   }
 
-  if (cfg->rdo < 0 || cfg->rdo > 3) {
-    fprintf(stderr, "Input error: --rd parameter out of range [0..3]\n");
+  if (cfg->rdo < 0 || cfg->rdo > 6) {
+    fprintf(stderr, "Input error: --rd parameter out of range [0..5]\n");
     error = 1;
   }
 

src/cli.c

@@ -172,6 +172,10 @@ static const struct option long_options[] = {
   { "no-combine-intra-cus",     no_argument, NULL, 0 },
   { "force-inter",              no_argument, NULL, 0 },
   { "no-force-inter",           no_argument, NULL, 0 },
+  { "intra-chroma-search",      no_argument, NULL, 0 },
+  { "no-intra-chroma-search",   no_argument, NULL, 0 },
+  { "fast-bipred",              no_argument, NULL, 0 },
+  { "no-fast-bipred",           no_argument, NULL, 0 },
   {0, 0, 0, 0}
 };
 
@@ -550,18 +554,22 @@ void print_help(void)
     "      --(no-)signhide        : Sign hiding [disabled]\n"
     "      --(no-)smp             : Symmetric motion partition [disabled]\n"
     "      --(no-)amp             : Asymmetric motion partition [disabled]\n"
-    "      --rd <integer>         : Intra mode search complexity [0]\n"
+    "      --rd <integer>         : Mode search complexity [0]\n"
     "                                   - 0: Skip intra if inter is good enough.\n"
     "                                   - 1: Rough intra mode search with SATD.\n"
-    "                                   - 2: Refine intra mode search with SSE.\n"
-    "                                   - 3: Try all intra modes and enable intra\n"
-    "                                        chroma mode search.\n"
+    "                                   - 2: Refine mode search with SSE.\n"
+    "                                   - 3: More SSE candidates for inter and\n"
+    "                                        chroma mode search for 4x4 intra.\n"
+    "                                   - 4: Even more SSE candidates for both.\n"
+    "                                   - 5: Try all intra modes.\n"
     "      --(no-)mv-rdo          : Rate-distortion optimized motion vector costs\n"
     "                               [disabled]\n"
     "      --(no-)zero-coeff-rdo  : If a CU is set inter, check if forcing zero\n"
     "                               residual improves the RD cost. [enabled]\n"
     "      --(no-)full-intra-search : Try all intra modes during rough search.\n"
     "                               [disabled]\n"
+    "      --(no-)intra-chroma-search : Test non-derived intra chroma modes.\n"
+    "                                   [disabled]\n"
     "      --(no-)transform-skip  : Try transform skip [disabled]\n"
     "      --me <string>          : Integer motion estimation algorithm [hexbs]\n"
     "                                   - hexbs: Hexagon Based Search\n"
@@ -577,6 +585,7 @@ void print_help(void)
     "                                   - 2: + 1/2-pixel diagonal\n"
     "                                   - 3: + 1/4-pixel horizontal and vertical\n"
     "                                   - 4: + 1/4-pixel diagonal\n"
+    "      --(no-)fast-bipred     : Only perform fast bipred search. [enabled]\n"
     "      --pu-depth-inter <int>-<int> : Inter prediction units sizes [0-3]\n"
     "                                   - 0, 1, 2, 3: from 64x64 to 8x8\n"
     "                                   - Accepts a list of values separated by ','\n"

src/encode_coding_tree.c

@@ -63,11 +63,12 @@
 void kvz_encode_last_significant_xy(cabac_data_t * const cabac,
                                     uint8_t lastpos_x, uint8_t lastpos_y,
                                     uint8_t width, uint8_t height,
-                                    uint8_t type, uint8_t scan)
+                                    uint8_t type, uint8_t scan, double* bits_out)
 {
   const int index = kvz_math_floor_log2(width) - 2;
   uint8_t ctx_offset = type ? 0 : (index * 3 + (index + 1) / 4);
   uint8_t shift = type ? index : (index + 3) / 4;
+  double bits = 0;
 
   cabac_ctx_t *base_ctx_x = (type ? cabac->ctx.cu_ctx_last_x_chroma : cabac->ctx.cu_ctx_last_x_luma);
   cabac_ctx_t *base_ctx_y = (type ? cabac->ctx.cu_ctx_last_y_chroma : cabac->ctx.cu_ctx_last_y_luma);
@@ -81,37 +82,36 @@ void kvz_encode_last_significant_xy(cabac_data_t * const cabac,
 
   // x prefix
   for (int last_x = 0; last_x < group_idx_x; last_x++) {
-    cabac->cur_ctx = &base_ctx_x[ctx_offset + (last_x >> shift)];
-    CABAC_BIN(cabac, 1, "last_sig_coeff_x_prefix");
+    CABAC_FBITS_UPDATE(cabac, &base_ctx_x[ctx_offset + (last_x >> shift)], 1, bits, "last_sig_coeff_x_prefix");
   }
   if (group_idx_x < g_group_idx[width - 1]) {
-    cabac->cur_ctx = &base_ctx_x[ctx_offset + (group_idx_x >> shift)];
-    CABAC_BIN(cabac, 0, "last_sig_coeff_x_prefix");
+    CABAC_FBITS_UPDATE(cabac, &base_ctx_x[ctx_offset + (group_idx_x >> shift)], 0, bits, "last_sig_coeff_x_prefix");
   }
 
   // y prefix
   for (int last_y = 0; last_y < group_idx_y; last_y++) {
-    cabac->cur_ctx = &base_ctx_y[ctx_offset + (last_y >> shift)];
-    CABAC_BIN(cabac, 1, "last_sig_coeff_y_prefix");
+    CABAC_FBITS_UPDATE(cabac, &base_ctx_y[ctx_offset + (last_y >> shift)], 1, bits, "last_sig_coeff_y_prefix");
   }
   if (group_idx_y < g_group_idx[height - 1]) {
-    cabac->cur_ctx = &base_ctx_y[ctx_offset + (group_idx_y >> shift)];
-    CABAC_BIN(cabac, 0, "last_sig_coeff_y_prefix");
+    CABAC_FBITS_UPDATE(cabac, &base_ctx_y[ctx_offset + (group_idx_y >> shift)], 0, bits, "last_sig_coeff_y_prefix");
   }
 
   // last_sig_coeff_x_suffix
   if (group_idx_x > 3) {
     const int suffix = lastpos_x - g_min_in_group[group_idx_x];
-    const int bits = (group_idx_x - 2) / 2;
-    CABAC_BINS_EP(cabac, suffix, bits, "last_sig_coeff_x_suffix");
+    const int write_bits = (group_idx_x - 2) / 2;
+    CABAC_BINS_EP(cabac, suffix, write_bits, "last_sig_coeff_x_suffix");
+    if (cabac->only_count) bits += write_bits;
   }
 
   // last_sig_coeff_y_suffix
   if (group_idx_y > 3) {
     const int suffix = lastpos_y - g_min_in_group[group_idx_y];
-    const int bits = (group_idx_y - 2) / 2;
-    CABAC_BINS_EP(cabac, suffix, bits, "last_sig_coeff_y_suffix");
+    const int write_bits = (group_idx_y - 2) / 2;
+    CABAC_BINS_EP(cabac, suffix, write_bits, "last_sig_coeff_y_suffix");
+    if (cabac->only_count) bits += write_bits;
   }
+  if (cabac->only_count && bits_out) *bits_out += bits;
 }
 
 static void encode_transform_unit(encoder_state_t * const state,
@@ -142,7 +142,7 @@ static void encode_transform_unit(encoder_state_t * const state,
                          width,
                          0,
                          scan_idx,
-                         cur_pu->tr_skip);
+                         cur_pu->tr_skip, NULL);
   }
 
   if (depth == MAX_DEPTH + 1) {
@@ -172,11 +172,11 @@ static void encode_transform_unit(encoder_state_t * const state,
     const coeff_t *coeff_v = &state->coeff->v[xy_to_zorder(LCU_WIDTH_C, x_local, y_local)];
 
     if (cbf_is_set(cur_pu->cbf, depth, COLOR_U)) {
-      kvz_encode_coeff_nxn(state, &state->cabac, coeff_u, width_c, 2, scan_idx, 0);
+      kvz_encode_coeff_nxn(state, &state->cabac, coeff_u, width_c, 2, scan_idx, 0, NULL);
     }
 
     if (cbf_is_set(cur_pu->cbf, depth, COLOR_V)) {
-      kvz_encode_coeff_nxn(state, &state->cabac, coeff_v, width_c, 2, scan_idx, 0);
+      kvz_encode_coeff_nxn(state, &state->cabac, coeff_v, width_c, 2, scan_idx, 0, NULL);
     }
   }
 }

src/encode_coding_tree.h

@@ -74,6 +74,6 @@ void kvz_encode_inter_prediction_unit(encoder_state_t* const state,
 void kvz_encode_last_significant_xy(cabac_data_t * const cabac,
                                     uint8_t lastpos_x, uint8_t lastpos_y,
                                     uint8_t width, uint8_t height,
-                                    uint8_t type, uint8_t scan);
+                                    uint8_t type, uint8_t scan, double* bits_out);
 
 #endif // ENCODE_CODING_TREE_H_