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

Tuned cost calculations for "final cost"

Browse source 
By adding chroma cost and using 1.5x coeff cost instead of 0.5x, additional 3% (LP) and 1.8% (AI) drop in BD-rate noticed.
This commit is contained in:
Marko Viitanen 2014-03-12 10:35:56 +02:00
parent 75dd1f48ef
commit b5756821dd
1 changed files with 14 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/search.c
View file

@ -775,7 +775,7 @@ static int search_cu_intra(encoder_control *encoder,
* *
* Take SAD between reconstruction and original and add cost from * Take SAD between reconstruction and original and add cost from
* coding (bitcost * lambda) and cost for coding coefficients (estimated * coding (bitcost * lambda) and cost for coding coefficients (estimated
* here as (coefficient_sum / 2) * lambda * here as (coefficient_sum * 1.5) * lambda)
*/ */
static int lcu_get_final_cost(encoder_control *encoder, static int lcu_get_final_cost(encoder_control *encoder,
const int x_px, const int y_px, const int x_px, const int y_px,
@ -789,19 +789,29 @@ static int lcu_get_final_cost(encoder_control *encoder,
int x,y; int x,y;
cur_cu = &lcu->cu[LCU_CU_OFFSET+(x_local>>3) + (y_local>>3)*LCU_T_CU_WIDTH]; cur_cu = &lcu->cu[LCU_CU_OFFSET+(x_local>>3) + (y_local>>3)*LCU_T_CU_WIDTH];
// SAD between reconstruction and original // SAD between reconstruction and original + sum of coeffs
for (y = y_local; y < y_local+width; ++y) { for (y = y_local; y < y_local+width; ++y) {
for (x = x_local; x < x_local+width; ++x) { for (x = x_local; x < x_local+width; ++x) {
cost += abs((int)lcu->rec.y[y * LCU_WIDTH + x] - (int)lcu->ref.y[y * LCU_WIDTH + x]); cost += abs((int)lcu->rec.y[y * LCU_WIDTH + x] - (int)lcu->ref.y[y * LCU_WIDTH + x]);
coeff_cost += abs((int)lcu->coeff.y[y * LCU_WIDTH + x]); coeff_cost += abs((int)lcu->coeff.y[y * LCU_WIDTH + x]);
} }
} }
// Chroma SAD + sum of coeffs
for (y = y_local>>1; y < (y_local+width)>>1; ++y) {
for (x = x_local>>1; x < (x_local+width)>>1; ++x) {
cost += abs((int)lcu->rec.u[y * (LCU_WIDTH>>1) + x] - (int)lcu->ref.u[y * (LCU_WIDTH>>1) + x]);
cost += abs((int)lcu->rec.v[y * (LCU_WIDTH>>1) + x] - (int)lcu->ref.v[y * (LCU_WIDTH>>1) + x]);
coeff_cost += abs((int)lcu->coeff.u[y * (LCU_WIDTH>>1) + x]);
coeff_cost += abs((int)lcu->coeff.v[y * (LCU_WIDTH>>1) + x]);
}
}
// Bitcost // Bitcost
cost += (cur_cu->type == CU_INTER ? cur_cu->inter.bitcost : cur_cu->intra[0].bitcost)*(int32_t)(g_cur_lambda_cost+0.5); cost += (cur_cu->type == CU_INTER ? cur_cu->inter.bitcost : cur_cu->intra[0].bitcost)*(int32_t)(g_cur_lambda_cost+0.5);
// Coefficient costs (TODO: check cost) // Coefficient costs (TODO: more tuning of the cost)
cost += (coeff_cost>>1) * (int32_t)(g_cur_lambda_cost+0.5); cost += (coeff_cost + (coeff_cost>>1)) * (int32_t)(g_cur_lambda_cost+0.5);
return cost; return cost;
} }