hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-27 11:24:05 +00:00
Code Issues Actions Projects Releases Wiki Activity

[isp] Count isp cbfs.

Browse source
This commit is contained in:
siivonek 2022-08-18 14:03:53 +03:00 committed by Marko Viitanen
parent 75175ee2e2
commit 031a758d6c
7 changed files with 69 additions and 34 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/intra.c
View file

@ -1665,12 +1665,15 @@ void uvg_intra_recon_cu(
return; return;
} }
if (search_data->pred_cu.intra.isp_mode != ISP_MODE_NO_ISP && recon_luma ) { if (search_data->pred_cu.intra.isp_mode != ISP_MODE_NO_ISP && recon_luma ) {
search_data->best_isp_cbfs = 0;
// ISP split is done horizontally or vertically depending on ISP mode, 2 or 4 times depending on block dimensions. // ISP split is done horizontally or vertically depending on ISP mode, 2 or 4 times depending on block dimensions.
// Small blocks are split only twice. // Small blocks are split only twice.
int split_type = search_data->pred_cu.intra.isp_mode; int split_type = search_data->pred_cu.intra.isp_mode;
int part_dim = uvg_get_isp_split_dim(width, height, split_type); int part_dim = uvg_get_isp_split_dim(width, height, split_type);
int limit = split_type == ISP_MODE_HOR ? height : width; int limit = split_type == ISP_MODE_HOR ? height : width;
int split_num = 0;
for (int part = 0; part < limit; part += part_dim) { for (int part = 0; part < limit; part += part_dim) {
cbf_clear(&cur_cu->cbf, depth, COLOR_Y);
const int part_x = split_type == ISP_MODE_HOR ? x : x + part; const int part_x = split_type == ISP_MODE_HOR ? x : x + part;
const int part_y = split_type == ISP_MODE_HOR ? y + part: y; const int part_y = split_type == ISP_MODE_HOR ? y + part: y;
const int part_w = split_type == ISP_MODE_HOR ? part_dim : width; const int part_w = split_type == ISP_MODE_HOR ? part_dim : width;
@ -1683,6 +1686,8 @@ void uvg_intra_recon_cu(
uvg_quantize_lcu_residual(state, true, false, false, uvg_quantize_lcu_residual(state, true, false, false,
&loc, depth, cur_cu, lcu, &loc, depth, cur_cu, lcu,
false, tree_type); false, tree_type);
search_data->best_isp_cbfs |= cbf_is_set(cur_cu->cbf, depth, COLOR_Y) << (split_num++);
} }
} }
const bool has_luma = recon_luma && search_data->pred_cu.intra.isp_mode == ISP_MODE_NO_ISP; const bool has_luma = recon_luma && search_data->pred_cu.intra.isp_mode == ISP_MODE_NO_ISP;
@ -1700,6 +1705,7 @@ void uvg_intra_recon_cu(
intra_recon_tb_leaf(state, &loc, lcu, COLOR_V, search_data, tree_type); intra_recon_tb_leaf(state, &loc, lcu, COLOR_V, search_data, tree_type);
} }
// TODO: not necessary to call if only luma and ISP is on
uvg_quantize_lcu_residual(state, has_luma, has_chroma && !(search_data->pred_cu.joint_cb_cr & 3), uvg_quantize_lcu_residual(state, has_luma, has_chroma && !(search_data->pred_cu.joint_cb_cr & 3),
search_data->pred_cu.joint_cb_cr & 3 && state->encoder_control->cfg.jccr && has_chroma, search_data->pred_cu.joint_cb_cr & 3 && state->encoder_control->cfg.jccr && has_chroma,
&loc, depth, cur_cu, lcu, &loc, depth, cur_cu, lcu,

2
src/intra.h
View file

@ -71,6 +71,7 @@ typedef struct {
double coeff_bits; double coeff_bits;
double distortion; double distortion;
double lfnst_costs[3]; double lfnst_costs[3];
uint8_t best_isp_cbfs;
} intra_search_data_t ; } intra_search_data_t ;
@ -168,5 +169,6 @@ int uvg_get_mip_flag_context(int x, int y, int width, int height, const lcu_t* l
#define SPLIT_TYPE_HOR 1 #define SPLIT_TYPE_HOR 1
#define SPLIT_TYPE_VER 2 #define SPLIT_TYPE_VER 2
int uvg_get_isp_split_dim(const int width, const int height, const int split_type);
bool uvg_can_use_isp(const int width, const int height, const int max_tr_size); bool uvg_can_use_isp(const int width, const int height, const int max_tr_size);
bool uvg_can_use_isp_with_lfnst(const int width, const int height, const int isp_mode, const enum uvg_tree_type tree_type); bool uvg_can_use_isp_with_lfnst(const int width, const int height, const int isp_mode, const enum uvg_tree_type tree_type);

85
src/search.c
View file

@ -307,7 +307,8 @@ static void downsample_cclm_rec(encoder_state_t *state, int x, int y, int width,
double uvg_cu_rd_cost_luma(const encoder_state_t *const state, double uvg_cu_rd_cost_luma(const encoder_state_t *const state,
const int x_px, const int y_px, const int depth, const int x_px, const int y_px, const int depth,
const cu_info_t *const pred_cu, const cu_info_t *const pred_cu,
lcu_t *const lcu) lcu_t *const lcu,
uint8_t isp_cbf)
{ {
const int width = LCU_WIDTH >> depth; const int width = LCU_WIDTH >> depth;
const int skip_residual_coding = pred_cu->skipped || (pred_cu->type != CU_INTRA && pred_cu->cbf == 0); const int skip_residual_coding = pred_cu->skipped || (pred_cu->type != CU_INTRA && pred_cu->cbf == 0);
@ -329,29 +330,40 @@ double uvg_cu_rd_cost_luma(const encoder_state_t *const state,
int offset = width / 2; int offset = width / 2;
double sum = 0; double sum = 0;
sum += uvg_cu_rd_cost_luma(state, x_px, y_px, depth + 1, pred_cu, lcu); sum += uvg_cu_rd_cost_luma(state, x_px, y_px, depth + 1, pred_cu, lcu, isp_cbf);
sum += uvg_cu_rd_cost_luma(state, x_px + offset, y_px, depth + 1, pred_cu, lcu); sum += uvg_cu_rd_cost_luma(state, x_px + offset, y_px, depth + 1, pred_cu, lcu, isp_cbf);
sum += uvg_cu_rd_cost_luma(state, x_px, y_px + offset, depth + 1, pred_cu, lcu); sum += uvg_cu_rd_cost_luma(state, x_px, y_px + offset, depth + 1, pred_cu, lcu, isp_cbf);
sum += uvg_cu_rd_cost_luma(state, x_px + offset, y_px + offset, depth + 1, pred_cu, lcu); sum += uvg_cu_rd_cost_luma(state, x_px + offset, y_px + offset, depth + 1, pred_cu, lcu, isp_cbf);
return sum + tr_tree_bits * state->lambda; return sum + tr_tree_bits * state->lambda;
} }
// Add transform_tree cbf_luma bit cost. // Add transform_tree cbf_luma bit cost.
const int is_tr_split = tr_cu->tr_depth - tr_cu->depth; if (pred_cu->intra.isp_mode == ISP_MODE_NO_ISP) {
int is_set = cbf_is_set(pred_cu->cbf, depth, COLOR_Y); const int is_tr_split = tr_cu->tr_depth - tr_cu->depth;
if (pred_cu->type == CU_INTRA || int is_set = cbf_is_set(pred_cu->cbf, depth, COLOR_Y);
if (pred_cu->type == CU_INTRA ||
is_tr_split || is_tr_split ||
cbf_is_set(tr_cu->cbf, depth, COLOR_U) || cbf_is_set(tr_cu->cbf, depth, COLOR_U) ||
cbf_is_set(tr_cu->cbf, depth, COLOR_V)) cbf_is_set(tr_cu->cbf, depth, COLOR_V))
{ {
cabac_ctx_t *ctx = &(cabac->ctx.qt_cbf_model_luma[0]); cabac_ctx_t* ctx = &(cabac->ctx.qt_cbf_model_luma[0]);
CABAC_FBITS_UPDATE(cabac, ctx, is_set, tr_tree_bits, "cbf_y_search"); CABAC_FBITS_UPDATE(cabac, ctx, is_set, tr_tree_bits, "cbf_y_search");
}
if (is_set && state->encoder_control->cfg.trskip_enable && width <= (1 << state->encoder_control->cfg.trskip_max_size)) {
CABAC_FBITS_UPDATE(cabac, &cabac->ctx.transform_skip_model_luma, pred_cu->tr_idx == MTS_SKIP, tr_tree_bits, "transform_skip_flag");
}
} }
else {
if (is_set && state->encoder_control->cfg.trskip_enable && width <= (1 << state->encoder_control->cfg.trskip_max_size)) { cabac_ctx_t* ctx = &(cabac->ctx.qt_cbf_model_luma[0]);
CABAC_FBITS_UPDATE(cabac, &cabac->ctx.transform_skip_model_luma, pred_cu->tr_idx == MTS_SKIP, tr_tree_bits, "transform_skip_flag"); // TODO: 8x4 CUs
for (int i = 0; i < 4; i++) {
if (i != 3 && isp_cbf != 0x8) {
CABAC_FBITS_UPDATE(cabac, ctx, (isp_cbf >> i) & 1, tr_tree_bits, "cbf_y_search");
}
}
} }
// SSD between reconstruction and original // SSD between reconstruction and original
@ -478,7 +490,8 @@ static double cu_rd_cost_tr_split_accurate(
const int depth, const int depth,
const cu_info_t* const pred_cu, const cu_info_t* const pred_cu,
lcu_t* const lcu, lcu_t* const lcu,
enum uvg_tree_type tree_type) { enum uvg_tree_type tree_type,
uint8_t isp_cbf) {
const int width = LCU_WIDTH >> depth; const int width = LCU_WIDTH >> depth;
const int skip_residual_coding = pred_cu->skipped || (pred_cu->type != CU_INTRA && pred_cu->cbf == 0); const int skip_residual_coding = pred_cu->skipped || (pred_cu->type != CU_INTRA && pred_cu->cbf == 0);
@ -523,25 +536,37 @@ static double cu_rd_cost_tr_split_accurate(
int offset = LCU_WIDTH >> (depth + 1); int offset = LCU_WIDTH >> (depth + 1);
double sum = 0; double sum = 0;
sum += cu_rd_cost_tr_split_accurate(state, x_px, y_px, depth + 1, pred_cu, lcu, tree_type); sum += cu_rd_cost_tr_split_accurate(state, x_px, y_px, depth + 1, pred_cu, lcu, tree_type, isp_cbf);
sum += cu_rd_cost_tr_split_accurate(state, x_px + offset, y_px, depth + 1, pred_cu, lcu, tree_type); sum += cu_rd_cost_tr_split_accurate(state, x_px + offset, y_px, depth + 1, pred_cu, lcu, tree_type, isp_cbf);
sum += cu_rd_cost_tr_split_accurate(state, x_px, y_px + offset, depth + 1, pred_cu, lcu, tree_type); sum += cu_rd_cost_tr_split_accurate(state, x_px, y_px + offset, depth + 1, pred_cu, lcu, tree_type, isp_cbf);
sum += cu_rd_cost_tr_split_accurate(state, x_px + offset, y_px + offset, depth + 1, pred_cu, lcu, tree_type); sum += cu_rd_cost_tr_split_accurate(state, x_px + offset, y_px + offset, depth + 1, pred_cu, lcu, tree_type, isp_cbf);
return sum + tr_tree_bits * state->lambda; return sum + tr_tree_bits * state->lambda;
} }
const int cb_flag_y = cbf_is_set(tr_cu->cbf, depth, COLOR_Y) && tree_type != UVG_CHROMA_T; const int cb_flag_y = cbf_is_set(tr_cu->cbf, depth, COLOR_Y) && tree_type != UVG_CHROMA_T;
const bool is_isp = !(pred_cu->type == CU_INTER || pred_cu->intra.isp_mode == ISP_MODE_NO_ISP);
// Add transform_tree cbf_luma bit cost. // Add transform_tree cbf_luma bit cost.
const int is_tr_split = depth - tr_cu->depth; if (is_isp) {
if ((pred_cu->type == CU_INTRA || const int is_tr_split = depth - tr_cu->depth;
is_tr_split || if ((pred_cu->type == CU_INTRA ||
cb_flag_u || is_tr_split ||
cb_flag_v) cb_flag_u ||
cb_flag_v)
&& !skip_residual_coding && tree_type != UVG_CHROMA_T) && !skip_residual_coding && tree_type != UVG_CHROMA_T)
{ {
cabac_ctx_t* ctx = &(cabac->ctx.qt_cbf_model_luma[0]); cabac_ctx_t* ctx = &(cabac->ctx.qt_cbf_model_luma[0]);
CABAC_FBITS_UPDATE(cabac, ctx, cb_flag_y, tr_tree_bits, "cbf_y_search"); CABAC_FBITS_UPDATE(cabac, ctx, cb_flag_y, tr_tree_bits, "cbf_y_search");
}
}
else {
cabac_ctx_t* ctx = &(cabac->ctx.qt_cbf_model_luma[0]);
// TODO: 8x4 CUs
for (int i = 0; i < 4; i++) {
if (i != 3 && isp_cbf != 0x8) {
CABAC_FBITS_UPDATE(cabac, ctx, (isp_cbf >> i) & 1, tr_tree_bits, "cbf_y_search");
}
}
} }
if (cb_flag_y || cb_flag_u || cb_flag_v) { if (cb_flag_y || cb_flag_u || cb_flag_v) {
@ -563,7 +588,7 @@ static double cu_rd_cost_tr_split_accurate(
} }
// Chroma transform skip enable/disable is non-normative, so we need to count the chroma // Chroma transform skip enable/disable is non-normative, so we need to count the chroma
// tr-skip bits even when we are never using it. // tr-skip bits even when we are never using it.
const bool can_use_tr_skip = state->encoder_control->cfg.trskip_enable && width <= (1 << state->encoder_control->cfg.trskip_max_size); const bool can_use_tr_skip = state->encoder_control->cfg.trskip_enable && width <= (1 << state->encoder_control->cfg.trskip_max_size) && !is_isp;
if(cb_flag_y){ if(cb_flag_y){
if (can_use_tr_skip) { if (can_use_tr_skip) {
@ -1144,7 +1169,7 @@ static double search_cu(
cost = bits * state->lambda; cost = bits * state->lambda;
cost += cu_rd_cost_tr_split_accurate(state, x_local, y_local, depth, cur_cu, lcu, tree_type); cost += cu_rd_cost_tr_split_accurate(state, x_local, y_local, depth, cur_cu, lcu, tree_type, 0);
if (ctrl->cfg.zero_coeff_rdo && inter_zero_coeff_cost <= cost) { if (ctrl->cfg.zero_coeff_rdo && inter_zero_coeff_cost <= cost) {
cost = inter_zero_coeff_cost; cost = inter_zero_coeff_cost;
@ -1301,7 +1326,7 @@ static double search_cu(
double mode_bits = calc_mode_bits(state, lcu, cur_cu, x, y, depth) + bits; double mode_bits = calc_mode_bits(state, lcu, cur_cu, x, y, depth) + bits;
cost += mode_bits * state->lambda; cost += mode_bits * state->lambda;
cost += cu_rd_cost_tr_split_accurate(state, x_local, y_local, depth, cur_cu, lcu, tree_type); cost += cu_rd_cost_tr_split_accurate(state, x_local, y_local, depth, cur_cu, lcu, tree_type, 0);
memcpy(&post_seach_cabac, &state->search_cabac, sizeof(post_seach_cabac)); memcpy(&post_seach_cabac, &state->search_cabac, sizeof(post_seach_cabac));
memcpy(&state->search_cabac, &temp_cabac, sizeof(temp_cabac)); memcpy(&state->search_cabac, &temp_cabac, sizeof(temp_cabac));

3
src/search.h
View file

@ -87,7 +87,8 @@ void uvg_search_lcu(encoder_state_t *state, int x, int y, const yuv_t *hor_buf,
double uvg_cu_rd_cost_luma(const encoder_state_t *const state, double uvg_cu_rd_cost_luma(const encoder_state_t *const state,
const int x_px, const int y_px, const int depth, const int x_px, const int y_px, const int depth,
const cu_info_t *const pred_cu, const cu_info_t *const pred_cu,
lcu_t *const lcu); lcu_t *const lcu,
uint8_t isp_cbf);
double uvg_cu_rd_cost_chroma(const encoder_state_t *const state, double uvg_cu_rd_cost_chroma(const encoder_state_t *const state,
const int x_px, const int y_px, const int depth, const int x_px, const int y_px, const int depth,
cu_info_t *const pred_cu, cu_info_t *const pred_cu,

2
src/search_inter.c
View file

@ -2286,7 +2286,7 @@ void uvg_cu_cost_inter_rd2(encoder_state_t * const state,
int cbf = cbf_is_set_any(cur_cu->cbf, depth); int cbf = cbf_is_set_any(cur_cu->cbf, depth);
if(cbf) { if(cbf) {
*inter_cost = uvg_cu_rd_cost_luma(state, x_px, y_px, depth, cur_cu, lcu); *inter_cost = uvg_cu_rd_cost_luma(state, x_px, y_px, depth, cur_cu, lcu, 0);
if (reconstruct_chroma) { if (reconstruct_chroma) {
if (cur_cu->depth != cur_cu->tr_depth || !state->encoder_control->cfg.jccr) { if (cur_cu->depth != cur_cu->tr_depth || !state->encoder_control->cfg.jccr) {
*inter_cost += uvg_cu_rd_cost_chroma(state, x_px, y_px, depth, cur_cu, lcu); *inter_cost += uvg_cu_rd_cost_chroma(state, x_px, y_px, depth, cur_cu, lcu);

4
src/search_intra.c
View file

@ -406,6 +406,7 @@ static double search_intra_trdepth(
UVG_LUMA_T, UVG_LUMA_T,
true, true,
false); false);
if (pred_cu->intra.isp_mode != ISP_MODE_NO_ISP && search_data->best_isp_cbfs == 0) continue;
if (trafo != 0 && !cbf_is_set(pred_cu->cbf, depth, COLOR_Y)) continue; if (trafo != 0 && !cbf_is_set(pred_cu->cbf, depth, COLOR_Y)) continue;
@ -445,7 +446,8 @@ static double search_intra_trdepth(
lcu_px.y, lcu_px.y,
depth, depth,
pred_cu, pred_cu,
lcu); lcu,
search_data->best_isp_cbfs);
double transform_bits = 0; double transform_bits = 0;
if (state->encoder_control->cfg.lfnst && depth == pred_cu->tr_depth && if (state->encoder_control->cfg.lfnst && depth == pred_cu->tr_depth &&
trafo != MTS_SKIP) { trafo != MTS_SKIP) {

1
src/transform.c
View file

@ -1269,7 +1269,6 @@ static void quantize_tr_residual(
} }
// ISP_TODO: when other ISP things work, ask Joose about this
cbf_clear(&cur_pu->cbf, depth, color); cbf_clear(&cur_pu->cbf, depth, color);
if (has_coeffs) { if (has_coeffs) {
cbf_set(&cur_pu->cbf, depth, color); cbf_set(&cur_pu->cbf, depth, color);