hevc: optimize residual coding(cherry picked from commit 70692a44708157b4dfa50e402e446bfa2b27f55e)

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/hevc.h

@@ -653,10 +653,6 @@ typedef struct TransformUnit {
     int cur_intra_pred_mode;
 } TransformUnit;
 
-typedef struct ResidualCoding {
-    uint8_t significant_coeff_group_flag[8][8];
-} ResidualCoding;
-
 typedef struct SAOParams {
     uint8_t type_idx[3]; ///< sao_type_idx
 
@@ -726,16 +722,10 @@ typedef struct HEVCNAL {
 
 typedef struct HEVCLocalContext {
     uint8_t cabac_state[HEVC_CONTEXTS];
-    int ctx_set;
-    int greater1_ctx;
-    int last_coeff_abs_level_greater1_flag;
-    int c_rice_param;
-    int last_coeff_abs_level_remaining;
     GetBitContext gb;
     CABACContext cc;
     TransformTree tt;
     TransformUnit tu;
-    ResidualCoding rc;
     uint8_t first_qp_group;
     int8_t qp_y;
     int8_t curr_qp_y;
@@ -935,17 +925,12 @@ int ff_hevc_last_significant_coeff_y_prefix_decode(HEVCContext *s, int c_idx,
                                                    int log2_size);
 int ff_hevc_last_significant_coeff_suffix_decode(HEVCContext *s,
                                                  int last_significant_coeff_prefix);
-int ff_hevc_significant_coeff_group_flag_decode(HEVCContext *s, int c_idx, int x_cg,
-                                                int y_cg, int log2_trafo_size);
+int ff_hevc_significant_coeff_group_flag_decode(HEVCContext *s, int c_idx, int ctx_cg);
 int ff_hevc_significant_coeff_flag_decode(HEVCContext *s, int c_idx, int x_c, int y_c,
-                                          int log2_trafo_size, int scan_idx);
-int ff_hevc_coeff_abs_level_greater1_flag_decode(HEVCContext *s, int c_idx,
-                                                 int i, int n,
-                                                 int first_greater1_coeff_idx,
-                                                 int first_subset);
-int ff_hevc_coeff_abs_level_greater2_flag_decode(HEVCContext *s, int c_idx,
-                                                 int i, int n);
-int ff_hevc_coeff_abs_level_remaining(HEVCContext *s, int n, int base_level);
+                                          int log2_trafo_size, int scan_idx, int prev_sig);
+int ff_hevc_coeff_abs_level_greater1_flag_decode(HEVCContext *s, int c_idx, int ctx_set);
+int ff_hevc_coeff_abs_level_greater2_flag_decode(HEVCContext *s, int c_idx, int inc);
+int ff_hevc_coeff_abs_level_remaining(HEVCContext *s, int base_level, int rc_rice_param);
 int ff_hevc_coeff_sign_flag(HEVCContext *s, uint8_t nb);
 
 /**

libavcodec/hevc_cabac.c

@@ -764,24 +764,17 @@ int ff_hevc_last_significant_coeff_suffix_decode(HEVCContext *s,
     return value;
 }
 
-int ff_hevc_significant_coeff_group_flag_decode(HEVCContext *s, int c_idx, int x_cg,
-                                                int y_cg, int log2_trafo_size)
+int ff_hevc_significant_coeff_group_flag_decode(HEVCContext *s, int c_idx, int ctx_cg)
 {
-    int ctx_cg = 0;
     int inc;
 
-    if (x_cg < (1 << (log2_trafo_size - 2)) - 1)
-        ctx_cg += s->HEVClc.rc.significant_coeff_group_flag[x_cg + 1][y_cg];
-    if (y_cg < (1 << (log2_trafo_size - 2)) - 1)
-        ctx_cg += s->HEVClc.rc.significant_coeff_group_flag[x_cg][y_cg + 1];
-
     inc = FFMIN(ctx_cg, 1) + (c_idx>0 ? 2 : 0);
 
     return GET_CABAC(elem_offset[SIGNIFICANT_COEFF_GROUP_FLAG] + inc);
 }
 
 int ff_hevc_significant_coeff_flag_decode(HEVCContext *s, int c_idx, int x_c, int y_c,
-                                          int log2_trafo_size, int scan_idx)
+                                          int log2_trafo_size, int scan_idx, int prev_sig)
 {
     static const uint8_t ctx_idx_map[] = {
         0, 1, 4, 5, 2, 3, 4, 5, 6, 6, 8, 8, 7, 7, 8, 8
@@ -796,13 +789,6 @@ int ff_hevc_significant_coeff_flag_decode(HEVCContext *s, int c_idx, int x_c, in
     } else if (log2_trafo_size == 2) {
         sig_ctx = ctx_idx_map[(y_c << 2) + x_c];
     } else {
-        int prev_sig = 0;
-
-        if (x_cg < ((1 << log2_trafo_size) - 1) >> 2)
-            prev_sig += s->HEVClc.rc.significant_coeff_group_flag[x_cg + 1][y_cg];
-        if (y_cg < ((1 << log2_trafo_size) - 1) >> 2)
-            prev_sig += (s->HEVClc.rc.significant_coeff_group_flag[x_cg][y_cg + 1] << 1);
-
         switch (prev_sig) {
         case 0: {
                 int x_off = x_c & 3;
@@ -839,82 +825,46 @@ int ff_hevc_significant_coeff_flag_decode(HEVCContext *s, int c_idx, int x_c, in
     return GET_CABAC(elem_offset[SIGNIFICANT_COEFF_FLAG] + inc);
 }
 
-int ff_hevc_coeff_abs_level_greater1_flag_decode(HEVCContext *s, int c_idx,
-                                                 int i, int n,
-                                                 int first_elem,
-                                                 int first_subset)
+int ff_hevc_coeff_abs_level_greater1_flag_decode(HEVCContext *s, int c_idx, int inc)
 {
 
-    int inc;
-
-    if (first_elem) {
-        s->HEVClc.ctx_set = (i > 0 && c_idx == 0) ? 2 : 0;
-
-        if (!first_subset && s->HEVClc.greater1_ctx == 0)
-            s->HEVClc.ctx_set++;
-        s->HEVClc.greater1_ctx = 1;
-    }
-
-    inc = (s->HEVClc.ctx_set << 2) + s->HEVClc.greater1_ctx;
     if (c_idx > 0)
         inc += 16;
 
-    s->HEVClc.last_coeff_abs_level_greater1_flag =
-        GET_CABAC(elem_offset[COEFF_ABS_LEVEL_GREATER1_FLAG] + inc);
-
-    if (s->HEVClc.last_coeff_abs_level_greater1_flag) {
-        s->HEVClc.greater1_ctx = 0;
-    } else if (s->HEVClc.greater1_ctx > 0 && s->HEVClc.greater1_ctx < 3) {
-        s->HEVClc.greater1_ctx++;
-    }
-
-    return s->HEVClc.last_coeff_abs_level_greater1_flag;
+    return GET_CABAC(elem_offset[COEFF_ABS_LEVEL_GREATER1_FLAG] + inc);
 }
 
-int ff_hevc_coeff_abs_level_greater2_flag_decode(HEVCContext *s, int c_idx,
-                                                 int i, int n)
+int ff_hevc_coeff_abs_level_greater2_flag_decode(HEVCContext *s, int c_idx, int inc)
 {
-    int inc;
-
-    inc = s->HEVClc.ctx_set;
     if (c_idx > 0)
         inc += 4;
 
     return GET_CABAC(elem_offset[COEFF_ABS_LEVEL_GREATER2_FLAG] + inc);
 }
 
-int ff_hevc_coeff_abs_level_remaining(HEVCContext *s, int first_elem, int base_level)
+int ff_hevc_coeff_abs_level_remaining(HEVCContext *s, int base_level, int rc_rice_param)
 {
-    int i;
-    HEVCLocalContext *lc = &s->HEVClc;
     int prefix = 0;
     int suffix = 0;
-
-    if (first_elem) {
-        lc->c_rice_param = 0;
-        lc->last_coeff_abs_level_remaining = 0;
-    }
+    int last_coeff_abs_level_remaining;
+    int i;
 
     while (prefix < CABAC_MAX_BIN && get_cabac_bypass(&s->HEVClc.cc))
         prefix++;
     if (prefix == CABAC_MAX_BIN)
         av_log(s->avctx, AV_LOG_ERROR, "CABAC_MAX_BIN : %d\n", prefix);
     if (prefix < 3) {
-        for (i = 0; i < lc->c_rice_param; i++)
+        for (i = 0; i < rc_rice_param; i++)
             suffix = (suffix << 1) | get_cabac_bypass(&s->HEVClc.cc);
-        lc->last_coeff_abs_level_remaining = (prefix << lc->c_rice_param) + suffix;
+        last_coeff_abs_level_remaining = (prefix << rc_rice_param) + suffix;
     } else {
-        for (i = 0; i < prefix - 3 + lc->c_rice_param; i++)
+        int prefix_minus3 = prefix - 3;
+        for (i = 0; i < prefix_minus3 + rc_rice_param; i++)
             suffix = (suffix << 1) | get_cabac_bypass(&s->HEVClc.cc);
-        lc->last_coeff_abs_level_remaining = (((1 << (prefix - 3)) + 3 - 1)
-                                              << lc->c_rice_param) + suffix;
+        last_coeff_abs_level_remaining = (((1 << prefix_minus3) + 3 - 1)
+                                              << rc_rice_param) + suffix;
     }
-
-    lc->c_rice_param = FFMIN(lc->c_rice_param +
-                             ((base_level + lc->last_coeff_abs_level_remaining) >
-                              (3 * (1 << lc->c_rice_param))), 4);
-
-    return lc->last_coeff_abs_level_remaining;
+    return last_coeff_abs_level_remaining;
 }
 
 int ff_hevc_coeff_sign_flag(HEVCContext *s, uint8_t nb)

libavcodec/hevc_filter.c

@@ -182,7 +182,7 @@ static void copy_CTB(uint8_t *dst, uint8_t *src, int width, int height, int stri
 
 #define CTB(tab, x, y) ((tab)[(y) * s->sps->ctb_width + (x)])
 
-static void sao_filter_CTB(HEVCContext *s, int x, int y, int c_idx_min, int c_idx_max)
+static void sao_filter_CTB(HEVCContext *s, int x, int y)
 {
     //  TODO: This should be easily parallelizable
     //  TODO: skip CBs when (cu_transquant_bypass_flag || (pcm_loop_filter_disable_flag && pcm_flag))
@@ -283,8 +283,8 @@ static void sao_filter_CTB(HEVCContext *s, int x, int y, int c_idx_min, int c_id
                  (edges[2] ? width  + (x_shift >> chroma) : width)  << s->sps->pixel_shift,
                  (edges[3] ? height + (y_shift >> chroma) : height), stride);
 
-        for (class_index = 0; class_index < class && c_idx >= c_idx_min &&
-                              c_idx < c_idx_max; class_index++) {
+        for (class_index = 0; class_index < class; class_index++) {
+
             switch (sao[class_index]->type_idx[c_idx]) {
             case SAO_BAND:
                 s->hevcdsp.sao_band_filter[classes[class_index]](dst, src, stride, sao[class_index], edges, width, height, c_idx);
@@ -334,11 +334,8 @@ static void deblocking_filter_CTB(HEVCContext *s, int x0, int y0)
     int pcmf        = (s->sps->pcm_enabled_flag && s->sps->pcm.loop_filter_disable_flag) ||
                       s->pps->transquant_bypass_enable_flag;
 
-    if (s->deblock[ctb].disable)
-        return;
-
-    if (x0) {
-        left_tc_offset   = s->deblock[ctb-1].tc_offset;
+    if(x0) {
+        left_tc_offset = s->deblock[ctb-1].tc_offset;
         left_beta_offset = s->deblock[ctb-1].beta_offset;
     }
 
@@ -649,7 +646,7 @@ void ff_hevc_deblocking_boundary_strengths(HEVCContext *s, int x0, int y0, int l
     }
 
     // bs for TU internal vertical PU boundaries
-    if (log2_trafo_size > s->sps->log2_min_pu_size && !is_intra)
+    if (log2_trafo_size > log2_min_pu_size && !is_intra)
         for (j = 0; j < (1 << log2_trafo_size); j += 4) {
             int y_pu = (y0 + j) >> log2_min_pu_size;
             int y_tu = (y0 + j) >> log2_min_tu_size;
@@ -665,7 +662,6 @@ void ff_hevc_deblocking_boundary_strengths(HEVCContext *s, int x0, int y0, int l
                 uint8_t curr_cbf_luma = s->cbf_luma[y_tu * pic_width_in_min_tu + xq_tu];
                 RefPicList* left_refPicList = ff_hevc_get_ref_list(s, s->ref, x0 + i - 1, y0 + j);
 
-
                 bs = boundary_strength(s, curr, curr_cbf_luma, left, left_cbf_luma, left_refPicList, 0);
                 if (s->sh.disable_deblocking_filter_flag == 1)
                     bs = 0;
@@ -680,11 +676,9 @@ void ff_hevc_deblocking_boundary_strengths(HEVCContext *s, int x0, int y0, int l
 
 void ff_hevc_hls_filter(HEVCContext *s, int x, int y)
 {
-    int c_idx_min = s->sh.slice_sample_adaptive_offset_flag[0] != 0 ? 0 : 1;
-    int c_idx_max = s->sh.slice_sample_adaptive_offset_flag[1] != 0 ? 3 : 1;
     deblocking_filter_CTB(s, x, y);
     if (s->sps->sao_enabled)
-        sao_filter_CTB(s, x, y, c_idx_min, c_idx_max);
+        sao_filter_CTB(s, x, y);
 }
 
 void ff_hevc_hls_filters(HEVCContext *s, int x_ctb, int y_ctb, int ctb_size)