hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-24 02:24:07 +00:00
Code Issues Actions Projects Releases Wiki Activity

Replace simple coefficient cost estimation with CABAC. Substantial improvement.

Browse source 
Approximation proved to be too inaccurate while not giving actually that much speedup.
This commit is contained in:
Ari Lemmetti 2017-12-10 01:23:48 +02:00 committed by Ari Lemmetti
parent ea79069dc8
commit 14892fda00
4 changed files with 12 additions and 38 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
src/rdo.c
View file

@ -41,8 +41,6 @@
#define LOG2_SCAN_SET_SIZE 4
#define SBH_THRESHOLD 4
static const double COEFF_SUM_MULTIPLIER = 1.9;
const uint32_t kvz_g_go_rice_range[5] = { 7, 14, 26, 46, 78 };
const uint32_t kvz_g_go_rice_prefix_len[5] = { 8, 7, 6, 5, 4 };
@ -154,7 +152,7 @@ struct sh_rates_t {
*
* \returns bits needed to code input coefficients
*/
static INLINE uint32_t get_coeff_cabac_cost(
uint32_t kvz_get_coeff_cabac_cost(
const encoder_state_t * const state,
const coeff_t *coeff,
int32_t width,
@ -195,31 +193,6 @@ static INLINE uint32_t get_coeff_cabac_cost(
return (23 - cabac_copy.bits_left) + (cabac_copy.num_buffered_bytes << 3);
}
/**
* \brief Estimate bitcost for coding coefficients.
*
* \param coeff coefficient array
* \param width coeff block width
* \param type data type (0 == luma)
*
* \returns number of bits needed to code coefficients
*/
uint32_t kvz_get_coeff_cost(const encoder_state_t * const state,
const coeff_t *coeff,
int32_t width,
int32_t type,
int8_t scan_mode)
{
if (state->encoder_control->cfg.rdo > 0) {
return get_coeff_cabac_cost(state, coeff, width, type, scan_mode);
} else {
return COEFF_SUM_MULTIPLIER * kvz_coeff_abs_sum(coeff, width * width) + 0.5;
}
}
#define COEF_REMAIN_BIN_REDUCTION 3
/** Calculates the cost for specific absolute transform level
* \param abs_level scaled quantized level

11
src/rdo.h
View file

@ -39,11 +39,12 @@ extern const uint32_t kvz_g_go_rice_prefix_len[5];
void kvz_rdoq(encoder_state_t *state, coeff_t *coef, coeff_t *dest_coeff, int32_t width,
int32_t height, int8_t type, int8_t scan_mode, int8_t block_type, int8_t tr_depth);
uint32_t kvz_get_coeff_cost(const encoder_state_t *state,
const coeff_t *coeff,
int32_t width,
int32_t type,
int8_t scan_mode);
uint32_t kvz_get_coeff_cabac_cost(
const encoder_state_t * const state,
const coeff_t *coeff,
int32_t width,
int32_t type,
int8_t scan_mode);
int32_t kvz_get_ic_rate(encoder_state_t *state, uint32_t abs_level, uint16_t ctx_num_one, uint16_t ctx_num_abs,
uint16_t abs_go_rice, uint32_t c1_idx, uint32_t c2_idx, int8_t type);

6
src/search.c
View file

@ -262,7 +262,7 @@ double kvz_cu_rd_cost_luma(const encoder_state_t *const state,
int8_t luma_scan_mode = kvz_get_scan_order(pred_cu->type, pred_cu->intra.mode, depth);
const coeff_t *coeffs = &lcu->coeff.y[xy_to_zorder(LCU_WIDTH, x_px, y_px)];
coeff_bits += kvz_get_coeff_cost(state, coeffs, width, 0, luma_scan_mode);
coeff_bits += kvz_get_coeff_cabac_cost(state, coeffs, width, 0, luma_scan_mode);
}
double bits = tr_tree_bits + coeff_bits;
@ -331,8 +331,8 @@ double kvz_cu_rd_cost_chroma(const encoder_state_t *const state,
int8_t scan_order = kvz_get_scan_order(pred_cu->type, pred_cu->intra.mode_chroma, depth);
const int index = xy_to_zorder(LCU_WIDTH_C, lcu_px.x, lcu_px.y);
coeff_bits += kvz_get_coeff_cost(state, &lcu->coeff.u[index], width, 2, scan_order);
coeff_bits += kvz_get_coeff_cost(state, &lcu->coeff.v[index], width, 2, scan_order);
coeff_bits += kvz_get_coeff_cabac_cost(state, &lcu->coeff.u[index], width, 2, scan_order);
coeff_bits += kvz_get_coeff_cabac_cost(state, &lcu->coeff.v[index], width, 2, scan_order);
}
double bits = tr_tree_bits + coeff_bits;

4
src/transform.c
View file

@ -238,14 +238,14 @@ int kvz_quantize_residual_trskip(
0, in_stride, 4,
ref_in, pred_in, noskip.rec, noskip.coeff);
noskip.cost = kvz_pixels_calc_ssd(ref_in, noskip.rec, in_stride, 4, 4);
noskip.cost += kvz_get_coeff_cost(state, noskip.coeff, 4, 0, scan_order) * bit_cost;
noskip.cost += kvz_get_coeff_cabac_cost(state, noskip.coeff, 4, 0, scan_order) * bit_cost;
skip.has_coeffs = kvz_quantize_residual(
state, cur_cu, width, color, scan_order,
1, in_stride, 4,
ref_in, pred_in, skip.rec, skip.coeff);
skip.cost = kvz_pixels_calc_ssd(ref_in, skip.rec, in_stride, 4, 4);
skip.cost += kvz_get_coeff_cost(state, skip.coeff, 4, 0, scan_order) * bit_cost;
skip.cost += kvz_get_coeff_cabac_cost(state, skip.coeff, 4, 0, scan_order) * bit_cost;
if (noskip.cost <= skip.cost) {
*trskip_out = 0;