diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c
index 448c1c4102cc6be8aff4d0797350f9211d65339a..a0e7b37da033c4a694bb2cb1ec2910f48077d9e4 100644
--- a/vp9/encoder/vp9_onyx_if.c
+++ b/vp9/encoder/vp9_onyx_if.c
@@ -736,18 +736,18 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
sf->quarter_pixel_search = 1;
sf->half_pixel_search = 1;
sf->iterative_sub_pixel = 1;
- sf->no_skip_block4x4_search = 1;
sf->optimize_coefficients = !cpi->oxcf.lossless;
sf->first_step = 0;
sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
sf->comp_inter_joint_search_thresh = BLOCK_SIZE_AB4X4;
+ sf->adpative_rd_thresh = 0;
+
#if CONFIG_MULTIPLE_ARF
// Switch segmentation off.
sf->static_segmentation = 0;
#else
sf->static_segmentation = 0;
#endif
- sf->mb16_breakout = 0;
switch (mode) {
case 0: // best quality mode
@@ -761,13 +761,11 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
#else
sf->static_segmentation = 0;
#endif
- sf->mb16_breakout = 0;
sf->comp_inter_joint_search_thresh = BLOCK_SIZE_SB8X8;
-
+ sf->adpative_rd_thresh = 1;
if (speed > 0) {
sf->comp_inter_joint_search_thresh = BLOCK_SIZE_TYPES;
sf->optimize_coefficients = 0;
- sf->no_skip_block4x4_search = 0;
sf->first_step = 1;
}
break;
diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h
index 670f41f3b648ff68c066be83129c85fa7d5c027d..de9d18c167ddbfb2943383ee3d0e89f12fb4ef83 100644
--- a/vp9/encoder/vp9_onyx_int.h
+++ b/vp9/encoder/vp9_onyx_int.h
@@ -213,11 +213,10 @@ typedef struct {
int max_step_search_steps;
int first_step;
int optimize_coefficients;
- int no_skip_block4x4_search;
int search_best_filter;
- int mb16_breakout;
int static_segmentation;
int comp_inter_joint_search_thresh;
+ int adpative_rd_thresh;
} SPEED_FEATURES;
enum BlockSize {
@@ -320,8 +319,10 @@ typedef struct VP9_COMP {
unsigned int mode_chosen_counts[MAX_MODES];
int rd_thresh_mult[MAX_MODES];
- int rd_baseline_thresh[MAX_MODES];
- int rd_threshes[MAX_MODES];
+ int rd_baseline_thresh[BLOCK_SIZE_TYPES][MAX_MODES];
+ int rd_threshes[BLOCK_SIZE_TYPES][MAX_MODES];
+ int rd_thresh_freq_fact[BLOCK_SIZE_TYPES][MAX_MODES];
+
int64_t rd_comp_pred_diff[NB_PREDICTION_TYPES];
int rd_prediction_type_threshes[4][NB_PREDICTION_TYPES];
int comp_pred_count[COMP_PRED_CONTEXTS];
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index a8a29d98abc68464f80cd1dddf06f9c8d7d10e2c..930eae2f71de6bc58d81cfb7f49c89dbde807a16 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -111,6 +111,17 @@ const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
{SPLITMV, GOLDEN_FRAME, ALTREF_FRAME},
};
+// The baseline rd thresholds for breaking out of the rd loop for
+// certain modes are assumed to be based on 8x8 blocks.
+// This table is used to correct for blocks size.
+// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
+static int rd_thresh_block_size_factor[BLOCK_SIZE_TYPES] =
+ {2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32};
+
+#define BASE_RD_THRESH_FREQ_FACT 16
+#define MAX_RD_THRESH_FREQ_FACT 32
+#define MAX_RD_THRESH_FREQ_INC 1
+
static void fill_token_costs(vp9_coeff_count (*c)[BLOCK_TYPES],
vp9_coeff_count (*cnoskip)[BLOCK_TYPES],
vp9_coeff_probs_model (*p)[BLOCK_TYPES]) {
@@ -175,7 +186,7 @@ void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex) {
void vp9_initialize_rd_consts(VP9_COMP *cpi, int qindex) {
- int q, i;
+ int q, i, bsize;
vp9_clear_system_state(); // __asm emms;
@@ -207,24 +218,43 @@ void vp9_initialize_rd_consts(VP9_COMP *cpi, int qindex) {
cpi->RDDIV = 1;
cpi->RDMULT /= 100;
- for (i = 0; i < MAX_MODES; i++) {
- if (cpi->sf.thresh_mult[i] < INT_MAX) {
- cpi->rd_threshes[i] = cpi->sf.thresh_mult[i] * q / 100;
- } else {
- cpi->rd_threshes[i] = INT_MAX;
+ for (bsize = 0; bsize < BLOCK_SIZE_TYPES; ++bsize) {
+ for (i = 0; i < MAX_MODES; ++i) {
+ // Threshold here seem unecessarily harsh but fine given actual
+ // range of values used for cpi->sf.thresh_mult[]
+ int thresh_max = INT_MAX / (q * rd_thresh_block_size_factor[bsize]);
+
+ // *4 relates to the scaling of rd_thresh_block_size_factor[]
+ if ((int64_t)cpi->sf.thresh_mult[i] < thresh_max) {
+ cpi->rd_threshes[bsize][i] =
+ cpi->sf.thresh_mult[i] * q *
+ rd_thresh_block_size_factor[bsize] / (4 * 100);
+ } else {
+ cpi->rd_threshes[bsize][i] = INT_MAX;
+ }
+ cpi->rd_baseline_thresh[bsize][i] = cpi->rd_threshes[bsize][i];
+ cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
}
- cpi->rd_baseline_thresh[i] = cpi->rd_threshes[i];
}
} else {
cpi->RDDIV = 100;
- for (i = 0; i < MAX_MODES; i++) {
- if (cpi->sf.thresh_mult[i] < (INT_MAX / q)) {
- cpi->rd_threshes[i] = cpi->sf.thresh_mult[i] * q;
- } else {
- cpi->rd_threshes[i] = INT_MAX;
+ for (bsize = 0; bsize < BLOCK_SIZE_TYPES; ++bsize) {
+ for (i = 0; i < MAX_MODES; i++) {
+ // Threshold here seem unecessarily harsh but fine given actual
+ // range of values used for cpi->sf.thresh_mult[]
+ int thresh_max = INT_MAX / (q * rd_thresh_block_size_factor[bsize]);
+
+ if (cpi->sf.thresh_mult[i] < thresh_max) {
+ cpi->rd_threshes[bsize][i] =
+ cpi->sf.thresh_mult[i] * q *
+ rd_thresh_block_size_factor[bsize] / 4;
+ } else {
+ cpi->rd_threshes[bsize][i] = INT_MAX;
+ }
+ cpi->rd_baseline_thresh[bsize][i] = cpi->rd_threshes[bsize][i];
+ cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
}
- cpi->rd_baseline_thresh[i] = cpi->rd_threshes[i];
}
}
@@ -2619,9 +2649,9 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
txfm_cache[i] = INT64_MAX;
// Test best rd so far against threshold for trying this mode.
- if (bsize >= BLOCK_SIZE_SB8X8 &&
- (best_rd < cpi->rd_threshes[mode_index] ||
- cpi->rd_threshes[mode_index] == INT_MAX))
+ if ((best_rd < ((cpi->rd_threshes[bsize][mode_index] *
+ cpi->rd_thresh_freq_fact[bsize][mode_index]) >> 4)) ||
+ cpi->rd_threshes[bsize][mode_index] == INT_MAX)
continue;
x->skip = 0;
@@ -2812,9 +2842,10 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
int uv_skippable;
this_rd_thresh = (mbmi->ref_frame == LAST_FRAME) ?
- cpi->rd_threshes[THR_NEWMV] : cpi->rd_threshes[THR_NEWA];
+ cpi->rd_threshes[bsize][THR_NEWMV] :
+ cpi->rd_threshes[bsize][THR_NEWA];
this_rd_thresh = (mbmi->ref_frame == GOLDEN_FRAME) ?
- cpi->rd_threshes[THR_NEWG] : this_rd_thresh;
+ cpi->rd_threshes[bsize][THR_NEWG] : this_rd_thresh;
xd->mode_info_context->mbmi.txfm_size = TX_4X4;
for (switchable_filter_index = 0;
@@ -3155,7 +3186,27 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
if (is_inter_mode(best_mode))
++cpi->best_switchable_interp_count[vp9_switchable_interp_map[best_filter]];
- // TODO(rbultje) integrate with RD thresholding
+ // Updating rd_thresh_freq_fact[] here means that the differnt
+ // partition/block sizes are handled independently based on the best
+ // choice for the current partition. It may well be better to keep a scaled
+ // best rd so far value and update rd_thresh_freq_fact based on the mode/size
+ // combination that wins out.
+ if (cpi->sf.adpative_rd_thresh) {
+ for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
+ if (mode_index == best_mode_index) {
+ cpi->rd_thresh_freq_fact[bsize][mode_index] = BASE_RD_THRESH_FREQ_FACT;
+ } else {
+ cpi->rd_thresh_freq_fact[bsize][mode_index] += MAX_RD_THRESH_FREQ_INC;
+ if (cpi->rd_thresh_freq_fact[bsize][mode_index] >
+ (cpi->sf.adpative_rd_thresh * MAX_RD_THRESH_FREQ_FACT)) {
+ cpi->rd_thresh_freq_fact[bsize][mode_index] =
+ cpi->sf.adpative_rd_thresh * MAX_RD_THRESH_FREQ_FACT;
+ }
+ }
+ }
+ }
+
+ // TODO(rbultje) integrate with RD trd_thresh_freq_facthresholding
#if 0
// Reduce the activation RD thresholds for the best choice mode
if ((cpi->rd_baseline_thresh[best_mode_index] > 0) &&