Implemented proper intra mode cost calculation

This does not currently affect the search primary search defining the used block sizes, only the refining second intra search. Gain 1.9% BD-rate on All Intra 600f of BQMall QP 22,27,32,37.
This commit is contained in:
Marko Viitanen 2014-02-24 15:28:37 +02:00
parent b067b2f802
commit fb9e3ad1d2
4 changed files with 41 additions and 9 deletions

View file

@ -2623,6 +2623,12 @@ void encode_block_residual(encoder_control *encoder,
static vector2d offsets[4] = {{0,0},{1,0},{0,1},{1,1}}; static vector2d offsets[4] = {{0,0},{1,0},{0,1},{1,1}};
int num_pu = (cur_cu->part_size == SIZE_2Nx2N ? 1 : 4); int num_pu = (cur_cu->part_size == SIZE_2Nx2N ? 1 : 4);
int i; int i;
int8_t merge[3];
intra_get_dir_luma_predictor(encoder->in.cur_pic,
x_ctb * 2,
y_ctb * 2,
merge);
if (cur_cu->part_size == SIZE_NxN) { if (cur_cu->part_size == SIZE_NxN) {
width = width_c; width = width_c;
@ -2645,7 +2651,7 @@ void encode_block_residual(encoder_control *encoder,
x_ctb * (LCU_WIDTH >> (MAX_DEPTH)), x_ctb * (LCU_WIDTH >> (MAX_DEPTH)),
y_ctb * (LCU_WIDTH >> (MAX_DEPTH)), y_ctb * (LCU_WIDTH >> (MAX_DEPTH)),
width, pred_y, width, width, pred_y, width,
&cur_cu->intra[0].cost); &cur_cu->intra[0].cost,merge);
intra_set_block_mode(encoder->in.cur_pic, x_ctb, y_ctb, depth, intra_set_block_mode(encoder->in.cur_pic, x_ctb, y_ctb, depth,
cur_cu->intra[0].mode, cur_cu->part_size); cur_cu->intra[0].mode, cur_cu->part_size);

View file

@ -233,6 +233,25 @@ void intra_filter(pixel *ref, int32_t stride,int32_t width, int8_t mode)
#undef FWIDTH #undef FWIDTH
} }
/**
* \brief Helper function to find intra merge costs
* \returns intra mode coding cost in bits
*/
static uint32_t intra_merge_cost(int8_t mode, int8_t *merge_modes)
{
// merge mode -1 means they are not used -> cost 0
if(merge_modes[0] == -1) return 0;
// First candidate needs only one bit and two other need two
if(merge_modes[0] == mode) {
return 1;
} else if(merge_modes[1] == mode || merge_modes[2] == mode) {
return 2;
}
// Without merging the cost is 5 bits
return 5;
}
/** /**
* \brief Function to test best intra prediction mode * \brief Function to test best intra prediction mode
* \param orig original picture data * \param orig original picture data
@ -250,13 +269,14 @@ void intra_filter(pixel *ref, int32_t stride,int32_t width, int8_t mode)
This function derives the prediction samples for planar mode (intra coding). This function derives the prediction samples for planar mode (intra coding).
*/ */
int16_t intra_prediction(pixel *orig, int32_t origstride, pixel *rec, int16_t recstride, uint16_t xpos, int16_t intra_prediction(pixel *orig, int32_t origstride, pixel *rec, int16_t recstride, uint16_t xpos,
uint16_t ypos, uint8_t width, pixel *dst, int32_t dststride, uint32_t *sad_out) uint16_t ypos, uint8_t width, pixel *dst, int32_t dststride, uint32_t *sad_out, int8_t *merge)
{ {
uint32_t best_sad = 0xffffffff; uint32_t best_sad = 0xffffffff;
uint32_t sad = 0; uint32_t sad = 0;
int16_t best_mode = 1; int16_t best_mode = 1;
int32_t x,y; int32_t x,y;
int16_t i; int16_t i;
uint32_t ratecost = 0;
cost_16bit_nxn_func cost_func = get_sad_16bit_nxn_func(width); cost_16bit_nxn_func cost_func = get_sad_16bit_nxn_func(width);
@ -306,7 +326,8 @@ int16_t intra_prediction(pixel *orig, int32_t origstride, pixel *rec, int16_t re
for (i = 0; i < (int32_t)(width*width); i++) { for (i = 0; i < (int32_t)(width*width); i++) {
pred[i] = val; pred[i] = val;
} }
CHECK_FOR_BEST(1,0); ratecost = intra_merge_cost(1,merge)*(int)(g_cur_lambda_cost+0.5);
CHECK_FOR_BEST(1,ratecost);
} }
// Check angular not requiring filtering // Check angular not requiring filtering
@ -314,7 +335,8 @@ int16_t intra_prediction(pixel *orig, int32_t origstride, pixel *rec, int16_t re
int distance = MIN(abs(i - 26),abs(i - 10)); //!< Distance from top and left predictions int distance = MIN(abs(i - 26),abs(i - 10)); //!< Distance from top and left predictions
if(distance <= threshold) { if(distance <= threshold) {
intra_get_angular_pred(rec, recstride, pred, width, width, i, filter); intra_get_angular_pred(rec, recstride, pred, width, width, i, filter);
CHECK_FOR_BEST(i,0); ratecost = intra_merge_cost(i,merge)*(int)(g_cur_lambda_cost+0.5);
CHECK_FOR_BEST(i,ratecost);
} }
} }
@ -322,7 +344,8 @@ int16_t intra_prediction(pixel *orig, int32_t origstride, pixel *rec, int16_t re
// Test planar mode (always filtered) // Test planar mode (always filtered)
intra_get_planar_pred(rec_filtered, recstride, width, pred, width); intra_get_planar_pred(rec_filtered, recstride, width, pred, width);
CHECK_FOR_BEST(0,0); ratecost = intra_merge_cost(0,merge)*(int)(g_cur_lambda_cost+0.5);
CHECK_FOR_BEST(0,ratecost);
// Check angular predictions which require filtered samples // Check angular predictions which require filtered samples
// TODO: add conditions to skip some modes on borders // TODO: add conditions to skip some modes on borders
@ -331,7 +354,8 @@ int16_t intra_prediction(pixel *orig, int32_t origstride, pixel *rec, int16_t re
int distance = MIN(abs(i-26),abs(i-10)); //!< Distance from top and left predictions int distance = MIN(abs(i-26),abs(i-10)); //!< Distance from top and left predictions
if(distance > threshold) { if(distance > threshold) {
intra_get_angular_pred(rec_filtered, recstride, pred, width, width, i, filter); intra_get_angular_pred(rec_filtered, recstride, pred, width, width, i, filter);
CHECK_FOR_BEST(i,0); ratecost = intra_merge_cost(i,merge)*(int)(g_cur_lambda_cost+0.5);
CHECK_FOR_BEST(i,ratecost);
} }
} }

View file

@ -38,7 +38,7 @@ void intra_build_reference_border(picture* pic, const pixel *src, int32_t x_ctb,
void intra_filter(pixel* ref, int32_t stride, int32_t width, int8_t mode); void intra_filter(pixel* ref, int32_t stride, int32_t width, int8_t mode);
/* Predictions */ /* Predictions */
int16_t intra_prediction(pixel* orig, int32_t orig_stride, pixel* rec, int16_t rec_stride, uint16_t x_pos, uint16_t ypos, uint8_t width, pixel* dst, int32_t dst_stride, uint32_t *sad); int16_t intra_prediction(pixel* orig, int32_t orig_stride, pixel* rec, int16_t rec_stride, uint16_t x_pos, uint16_t ypos, uint8_t width, pixel* dst, int32_t dst_stride, uint32_t *sad, int8_t *merge);
pixel intra_get_dc_pred(pixel* pic, uint16_t pic_width, uint8_t width); pixel intra_get_dc_pred(pixel* pic, uint16_t pic_width, uint8_t width);
void intra_get_planar_pred(pixel* src,int32_t srcstride, uint32_t width, pixel* dst, int32_t dststride); void intra_get_planar_pred(pixel* src,int32_t srcstride, uint32_t width, pixel* dst, int32_t dststride);

View file

@ -664,13 +664,15 @@ static void search_intra(encoder_control *encoder, uint16_t x_ctb,
pixel rec[(LCU_WIDTH * 2 + 1) * (LCU_WIDTH * 2 + 1)]; pixel rec[(LCU_WIDTH * 2 + 1) * (LCU_WIDTH * 2 + 1)];
pixel *recShift = &rec[(LCU_WIDTH >> (depth)) * 2 + 8 + 1]; pixel *recShift = &rec[(LCU_WIDTH >> (depth)) * 2 + 8 + 1];
int8_t merge[3] = {-1,-1,-1};
// Build reconstructed block to use in prediction with extrapolated borders // Build reconstructed block to use in prediction with extrapolated borders
intra_build_reference_border(cur_pic, cur_pic->y_data, intra_build_reference_border(cur_pic, cur_pic->y_data,
x, y, x, y,
(int16_t)width * 2 + 8, rec, (int16_t)width * 2 + 8, 0); (int16_t)width * 2 + 8, rec, (int16_t)width * 2 + 8, 0);
cur_cu->intra[0].mode = (int8_t)intra_prediction(encoder->in.cur_pic->y_data, cur_cu->intra[0].mode = (int8_t)intra_prediction(encoder->in.cur_pic->y_data,
encoder->in.width, recShift, width * 2 + 8, x, y, encoder->in.width, recShift, width * 2 + 8, x, y,
width, pred, width, &cur_cu->intra[0].cost); width, pred, width, &cur_cu->intra[0].cost,merge);
cur_cu->part_size = SIZE_2Nx2N; cur_cu->part_size = SIZE_2Nx2N;
// Do search for NxN split. // Do search for NxN split.
@ -692,7 +694,7 @@ static void search_intra(encoder_control *encoder, uint16_t x_ctb,
(int16_t)width * 2 + 8, rec, (int16_t)width * 2 + 8, 0); (int16_t)width * 2 + 8, rec, (int16_t)width * 2 + 8, 0);
cur_cu->intra[i].mode = (int8_t)intra_prediction(encoder->in.cur_pic->y_data, cur_cu->intra[i].mode = (int8_t)intra_prediction(encoder->in.cur_pic->y_data,
encoder->in.width, recShift, width * 2 + 8, (int16_t)x_pos, (int16_t)y_pos, encoder->in.width, recShift, width * 2 + 8, (int16_t)x_pos, (int16_t)y_pos,
width, pred, width, &cur_cu->intra[i].cost); width, pred, width, &cur_cu->intra[i].cost,merge);
cost += cur_cu->intra[i].cost; cost += cur_cu->intra[i].cost;
} }