/**
* Copy all non-reference CU data from depth+1 to depth.
*/
static void work_tree_copy_up(int x_px, int y_px, int depth, lcu_t work_tree[MAX_PU_DEPTH + 1])
{
assert(depth >= 0 && depth < MAX_PU_DEPTH);
// Copy non-reference CUs.
{
const int x_orig = SUB_SCU(x_px);
const int y_orig = SUB_SCU(y_px);
const int width_cu = LCU_WIDTH >> depth;
for (int y = y_orig; y < y_orig + width_cu; y += SCU_WIDTH) {
for (int x = x_orig; x < x_orig + width_cu; x += SCU_WIDTH) {
const cu_info_t *from_cu = LCU_GET_CU_AT_PX(&work_tree[depth + 1], x, y);
cu_info_t *to_cu = LCU_GET_CU_AT_PX(&work_tree[depth], x, y);
memcpy(to_cu, from_cu, sizeof(*to_cu));
}
}
}
// Copy reconstructed pixels.
{
const int x = SUB_SCU(x_px);
const int y = SUB_SCU(y_px);
const int width_px = LCU_WIDTH >> depth;
const int luma_index = x + y * LCU_WIDTH;
const int chroma_index = (x / 2) + (y / 2) * (LCU_WIDTH / 2);
const lcu_yuv_t *from = &work_tree[depth + 1].rec;
lcu_yuv_t *to = &work_tree[depth].rec;
const lcu_coeff_t *from_coeff = &work_tree[depth + 1].coeff;
lcu_coeff_t *to_coeff = &work_tree[depth].coeff;
kvz_pixels_blit(&from->y[luma_index], &to->y[luma_index],
width_px, width_px, LCU_WIDTH, LCU_WIDTH);
if (from->chroma_format != KVZ_CSP_400) {
kvz_pixels_blit(&from->u[chroma_index], &to->u[chroma_index],
width_px / 2, width_px / 2, LCU_WIDTH / 2, LCU_WIDTH / 2);
kvz_pixels_blit(&from->v[chroma_index], &to->v[chroma_index],
width_px / 2, width_px / 2, LCU_WIDTH / 2, LCU_WIDTH / 2);
}
// Copy coefficients up. They do not have to be copied down because they
// are not used for the search.
kvz_coefficients_blit(&from_coeff->y[luma_index], &to_coeff->y[luma_index],
width_px, width_px, LCU_WIDTH, LCU_WIDTH);
if (from->chroma_format != KVZ_CSP_400) {
kvz_coefficients_blit(&from_coeff->u[chroma_index], &to_coeff->u[chroma_index],
width_px / 2, width_px / 2, LCU_WIDTH / 2, LCU_WIDTH / 2);
kvz_coefficients_blit(&from_coeff->v[chroma_index], &to_coeff->v[chroma_index],
width_px / 2, width_px / 2, LCU_WIDTH / 2, LCU_WIDTH / 2);
}
}
}
/**
* \brief Like kvz_quantize_residual except that this uses trskip if that is better.
*
* Using this function saves one step of quantization and inverse quantization
* compared to doing the decision separately from the actual operation.
*
* \param width Transform width.
* \param color Color.
* \param scan_order Coefficient scan order.
* \param trskip_out Whether transform skip is used.
* \param stride Stride for ref_in, pred_in rec_out and coeff_out.
* \param ref_in Reference pixels.
* \param pred_in Predicted pixels.
* \param rec_out Reconstructed pixels.
* \param coeff_out Coefficients used for reconstruction of rec_out.
*
* \returns Whether coeff_out contains any non-zero coefficients.
*/
int kvz_quantize_residual_trskip(
encoder_state_t *const state,
const cu_info_t *const cur_cu, const int width, const color_t color,
const coeff_scan_order_t scan_order, int8_t *trskip_out,
const int in_stride, const int out_stride,
const kvz_pixel *const ref_in, const kvz_pixel *const pred_in,
kvz_pixel *rec_out, coeff_t *coeff_out)
{
struct {
kvz_pixel rec[4*4];
coeff_t coeff[4*4];
uint32_t cost;
int has_coeffs;
} skip, noskip, *best;
const int bit_cost = (int)(state->global->cur_lambda_cost+0.5);
noskip.has_coeffs = kvz_quantize_residual(
state, cur_cu, width, color, scan_order,
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;
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;
if (noskip.cost <= skip.cost) {
*trskip_out = 0;
best = &noskip;
} else {
*trskip_out = 1;
best = &skip;
}
if (best->has_coeffs || rec_out != pred_in) {
// If there is no residual and reconstruction is already in rec_out,
// we can skip this.
kvz_pixels_blit(best->rec, rec_out, width, width, 4, out_stride);
}
kvz_coefficients_blit(best->coeff, coeff_out, width, width, 4, out_stride);
return best->has_coeffs;
}
/**
* Calculate RD cost for a Coding Unit.
* \return Cost of block
* \param ref_cu CU used for prediction parameters.
*
* Calculates the RDO cost of a single CU that will not be split further.
* Takes into account SSD of reconstruction and the cost of encoding whatever
* prediction unit data needs to be coded.
*/
double kvz_cu_rd_cost_luma(const encoder_state_t *const state,
const int x_px, const int y_px, const int depth,
const cu_info_t *const pred_cu,
lcu_t *const lcu)
{
const int width = LCU_WIDTH >> depth;
// cur_cu is used for TU parameters.
cu_info_t *const tr_cu = LCU_GET_CU_AT_PX(lcu, x_px, y_px);
double coeff_bits = 0;
double tr_tree_bits = 0;
// Check that lcu is not in
assert(x_px >= 0 && x_px < LCU_WIDTH);
assert(y_px >= 0 && y_px < LCU_WIDTH);
const uint8_t tr_depth = tr_cu->tr_depth - depth;
// Add transform_tree split_transform_flag bit cost.
bool intra_split_flag = pred_cu->type == CU_INTRA && pred_cu->part_size == SIZE_NxN && depth == 3;
if (width <= TR_MAX_WIDTH
&& width > TR_MIN_WIDTH
&& !intra_split_flag)
{
const cabac_ctx_t *ctx = &(state->cabac.ctx.trans_subdiv_model[5 - (6 - depth)]);
tr_tree_bits += CTX_ENTROPY_FBITS(ctx, tr_depth > 0);
}
if (tr_depth > 0) {
int offset = width / 2;
double sum = 0;
sum += kvz_cu_rd_cost_luma(state, x_px, y_px, depth + 1, pred_cu, lcu);
sum += kvz_cu_rd_cost_luma(state, x_px + offset, y_px, depth + 1, pred_cu, lcu);
sum += kvz_cu_rd_cost_luma(state, x_px, y_px + offset, depth + 1, pred_cu, lcu);
sum += kvz_cu_rd_cost_luma(state, x_px + offset, y_px + offset, depth + 1, pred_cu, lcu);
return sum + tr_tree_bits * state->lambda;
}
// Add transform_tree cbf_luma bit cost.
if (pred_cu->type == CU_INTRA ||
tr_depth > 0 ||
cbf_is_set(tr_cu->cbf, depth, COLOR_U) ||
cbf_is_set(tr_cu->cbf, depth, COLOR_V))
{
const cabac_ctx_t *ctx = &(state->cabac.ctx.qt_cbf_model_luma[!tr_depth]);
tr_tree_bits += CTX_ENTROPY_FBITS(ctx, cbf_is_set(pred_cu->cbf, depth, COLOR_Y));
}
// SSD between reconstruction and original
int ssd = 0;
if (!state->encoder_control->cfg->lossless) {
int index = y_px * LCU_WIDTH + x_px;
ssd = kvz_pixels_calc_ssd(&lcu->ref.y[index], &lcu->rec.y[index],
LCU_WIDTH, LCU_WIDTH,
width);
}
{
coeff_t coeff_temp[32 * 32];
int8_t luma_scan_mode = kvz_get_scan_order(pred_cu->type, pred_cu->intra.mode, depth);
// Code coeffs using cabac to get a better estimate of real coding costs.
kvz_coefficients_blit(&lcu->coeff.y[(y_px*LCU_WIDTH) + x_px], coeff_temp, width, width, LCU_WIDTH, width);
coeff_bits += kvz_get_coeff_cost(state, coeff_temp, width, 0, luma_scan_mode);
}
double bits = tr_tree_bits + coeff_bits;
return (double)ssd * LUMA_MULT + bits * state->lambda;
}
