Merge "Add sse2 version for vp9_quantize_fp"
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 9c50041..7175389 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1267,9 +1267,6 @@
$vp9_full_search_sad_sse3=vp9_full_search_sadx3;
$vp9_full_search_sad_sse4_1=vp9_full_search_sadx8;
-add_proto qw/int vp9_refining_search_sad/, "const struct macroblock *x, struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
-specialize qw/vp9_refining_search_sad/;
-
add_proto qw/int vp9_diamond_search_sad/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
specialize qw/vp9_diamond_search_sad/;
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index 4a3e79e..73ce789 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -500,7 +500,7 @@
static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
vp9_coeff_stats *coef_branch_ct,
vp9_coeff_probs_model *coef_probs) {
- vp9_coeff_count *coef_counts = cpi->coef_counts[tx_size];
+ vp9_coeff_count *coef_counts = cpi->frame_counts->coef_counts[tx_size];
unsigned int (*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
cpi->common.counts.eob_branch[tx_size];
int i, j, k, l, m;
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 3beceb4..f01f25f 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -602,7 +602,7 @@
int output_enabled) {
int i, x_idx, y;
VP9_COMMON *const cm = &cpi->common;
- RD_OPT *const rd_opt = &cpi->rd;
+ COUNTS *const counts = cpi->frame_counts;
MACROBLOCK *const x = cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = x->plane;
@@ -695,7 +695,7 @@
if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
for (i = 0; i < TX_MODES; i++)
- rd_opt->tx_select_diff[i] += ctx->tx_rd_diff[i];
+ counts->tx_select_diff[i] += ctx->tx_rd_diff[i];
}
#if CONFIG_INTERNAL_STATS
@@ -728,12 +728,12 @@
}
}
- rd_opt->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
- rd_opt->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
- rd_opt->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
+ counts->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
+ counts->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
+ counts->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
- rd_opt->filter_diff[i] += ctx->best_filter_diff[i];
+ counts->filter_diff[i] += ctx->best_filter_diff[i];
}
for (h = 0; h < y_mis; ++h) {
@@ -3495,15 +3495,16 @@
MACROBLOCK *const x = cpi->mb;
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
+ COUNTS *const counts = cpi->frame_counts;
xd->mi = cm->mi;
xd->mi[0].src_mi = &xd->mi[0];
vp9_zero(cm->counts);
- vp9_zero(cpi->coef_counts);
- vp9_zero(rd_opt->comp_pred_diff);
- vp9_zero(rd_opt->filter_diff);
- vp9_zero(rd_opt->tx_select_diff);
+ vp9_zero(counts->coef_counts);
+ vp9_zero(counts->comp_pred_diff);
+ vp9_zero(counts->filter_diff);
+ vp9_zero(counts->tx_select_diff);
vp9_zero(rd_opt->tx_select_threshes);
xd->lossless = cm->base_qindex == 0 &&
@@ -3610,6 +3611,7 @@
void vp9_encode_frame(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
RD_OPT *const rd_opt = &cpi->rd;
+ COUNTS *const counts = cpi->frame_counts;
// In the longer term the encoder should be generalized to match the
// decoder such that we allow compound where one of the 3 buffers has a
@@ -3667,13 +3669,13 @@
encode_frame_internal(cpi);
for (i = 0; i < REFERENCE_MODES; ++i)
- mode_thrs[i] = (mode_thrs[i] + rd_opt->comp_pred_diff[i] / cm->MBs) / 2;
+ mode_thrs[i] = (mode_thrs[i] + counts->comp_pred_diff[i] / cm->MBs) / 2;
for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
- filter_thrs[i] = (filter_thrs[i] + rd_opt->filter_diff[i] / cm->MBs) / 2;
+ filter_thrs[i] = (filter_thrs[i] + counts->filter_diff[i] / cm->MBs) / 2;
for (i = 0; i < TX_MODES; ++i) {
- int64_t pd = rd_opt->tx_select_diff[i];
+ int64_t pd = counts->tx_select_diff[i];
if (i == TX_MODE_SELECT)
pd -= RDCOST(cpi->mb->rdmult, cpi->mb->rddiv, 2048 * (TX_SIZES - 1), 0);
tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 4425d2f..9ece2d3 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -235,6 +235,9 @@
cpi->nmvsadcosts_hp[0] = NULL;
cpi->nmvsadcosts_hp[1] = NULL;
+ vpx_free(cpi->frame_counts);
+ cpi->frame_counts = NULL;
+
vp9_cyclic_refresh_free(cpi->cyclic_refresh);
cpi->cyclic_refresh = NULL;
@@ -1407,7 +1410,7 @@
VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf) {
- unsigned int i, j;
+ unsigned int i;
VP9_COMP *const cpi = vpx_memalign(32, sizeof(VP9_COMP));
VP9_COMMON *const cm = cpi != NULL ? &cpi->common : NULL;
@@ -1476,6 +1479,9 @@
CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
+ CHECK_MEM_ERROR(cm, cpi->frame_counts, vpx_calloc(1,
+ sizeof(*cpi->frame_counts)));
+
for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
sizeof(cpi->mbgraph_stats[0])); i++) {
CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
@@ -1640,14 +1646,6 @@
cpi->source_var_thresh = 0;
cpi->frames_till_next_var_check = 0;
- // Default rd threshold factors for mode selection
- for (i = 0; i < BLOCK_SIZES; ++i) {
- for (j = 0; j < MAX_MODES; ++j) {
- cpi->rd.thresh_freq_fact[i][j] = 32;
- cpi->rd.mode_map[i][j] = j;
- }
- }
-
#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
cpi->fn_ptr[BT].sdf = SDF; \
cpi->fn_ptr[BT].sdaf = SDAF; \
@@ -3289,7 +3287,7 @@
vp9_update_reference_frames(cpi);
for (t = TX_4X4; t <= TX_32X32; t++)
- full_to_model_counts(cm->counts.coef[t], cpi->coef_counts[t]);
+ full_to_model_counts(cm->counts.coef[t], cpi->frame_counts->coef_counts[t]);
if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
vp9_adapt_coef_probs(cm);
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index ccac59d..85b606d 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -241,6 +241,13 @@
int mode_map[BLOCK_SIZES][MAX_MODES];
} TileDataEnc;
+typedef struct {
+ vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
+ int64_t comp_pred_diff[REFERENCE_MODES];
+ int64_t tx_select_diff[TX_MODES];
+ int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+} COUNTS;
+
typedef struct VP9_COMP {
QUANTS quants;
MACROBLOCK *mb;
@@ -287,6 +294,7 @@
int ambient_err;
RD_OPT rd;
+ COUNTS *frame_counts;
CODING_CONTEXT coding_context;
@@ -305,7 +313,6 @@
RATE_CONTROL rc;
double framerate;
- vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE];
struct vpx_codec_pkt_list *output_pkt_list;
@@ -338,7 +345,6 @@
fractional_mv_step_fp *find_fractional_mv_step;
vp9_full_search_fn_t full_search_sad;
- vp9_refining_search_fn_t refining_search_sad;
vp9_diamond_search_fn_t diamond_search_sad;
vp9_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
uint64_t time_receive_data;
diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c
index 28598f1..252be95 100644
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -1752,7 +1752,7 @@
if (do_refine) {
const int search_range = 8;
MV best_mv = *dst_mv;
- thissme = cpi->refining_search_sad(x, &best_mv, sadpb, search_range,
+ thissme = vp9_refining_search_sad(x, &best_mv, sadpb, search_range,
fn_ptr, ref_mv);
if (thissme < INT_MAX)
thissme = vp9_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
@@ -1956,11 +1956,11 @@
return best_sad;
}
-int vp9_refining_search_sad_c(const MACROBLOCK *x,
- MV *ref_mv, int error_per_bit,
- int search_range,
- const vp9_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv) {
+int vp9_refining_search_sad(const MACROBLOCK *x,
+ MV *ref_mv, int error_per_bit,
+ int search_range,
+ const vp9_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv) {
const MACROBLOCKD *const xd = &x->e_mbd;
const MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}};
const struct buf_2d *const what = &x->plane[0].src;
diff --git a/vp9/encoder/vp9_mcomp.h b/vp9/encoder/vp9_mcomp.h
index 9ddca25..bba48fd 100644
--- a/vp9/encoder/vp9_mcomp.h
+++ b/vp9/encoder/vp9_mcomp.h
@@ -66,6 +66,12 @@
int vp9_init_search_range(int size);
+int vp9_refining_search_sad(const struct macroblock *x,
+ struct mv *ref_mv,
+ int sad_per_bit, int distance,
+ const struct vp9_variance_vtable *fn_ptr,
+ const struct mv *center_mv);
+
// Runs sequence of diamond searches in smaller steps for RD
int vp9_full_pixel_diamond(const struct VP9_COMP *cpi, MACROBLOCK *x,
MV *mvp_full, int step_param,
diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c
index c19602a..f1d4774 100644
--- a/vp9/encoder/vp9_ratectrl.c
+++ b/vp9/encoder/vp9_ratectrl.c
@@ -430,7 +430,8 @@
adjustment_limit = 0.75;
break;
case 1:
- adjustment_limit = 0.125 + 0.5 * MIN(1, fabs(log10(0.01 * correction_factor)));
+ adjustment_limit = 0.25 +
+ 0.5 * MIN(1, fabs(log10(0.01 * correction_factor)));
break;
case 2:
default:
@@ -438,12 +439,21 @@
break;
}
+ cpi->rc.q_2_frame = cpi->rc.q_1_frame;
+ cpi->rc.q_1_frame = cm->base_qindex;
+ cpi->rc.rc_2_frame = cpi->rc.rc_1_frame;
+ if (correction_factor > 110)
+ cpi->rc.rc_1_frame = -1;
+ else if (correction_factor < 90)
+ cpi->rc.rc_1_frame = 1;
+ else
+ cpi->rc.rc_1_frame = 0;
+
if (correction_factor > 102) {
// We are not already at the worst allowable quality
correction_factor = (int)(100 + ((correction_factor - 100) *
adjustment_limit));
rate_correction_factor = (rate_correction_factor * correction_factor) / 100;
-
// Keep rate_correction_factor within limits
if (rate_correction_factor > MAX_BPB_FACTOR)
rate_correction_factor = MAX_BPB_FACTOR;
@@ -494,6 +504,14 @@
}
} while (++i <= active_worst_quality);
+ // In CBR mode, this makes sure q is between oscillating Qs to prevent
+ // resonance.
+ if (cpi->oxcf.rc_mode == VPX_CBR &&
+ (cpi->rc.rc_1_frame * cpi->rc.rc_2_frame == -1) &&
+ cpi->rc.q_1_frame != cpi->rc.q_2_frame) {
+ q = clamp(q, MIN(cpi->rc.q_1_frame, cpi->rc.q_2_frame),
+ MAX(cpi->rc.q_1_frame, cpi->rc.q_2_frame));
+ }
return q;
}
@@ -563,18 +581,23 @@
const VP9_COMMON *const cm = &cpi->common;
const RATE_CONTROL *rc = &cpi->rc;
// Buffer level below which we push active_worst to worst_quality.
- int64_t critical_level = rc->optimal_buffer_level >> 2;
+ int64_t critical_level = rc->optimal_buffer_level >> 3;
int64_t buff_lvl_step = 0;
int adjustment = 0;
int active_worst_quality;
+ int ambient_qp;
if (cm->frame_type == KEY_FRAME)
- return rc->worst_quality * 4 / 5;
- if (cm->current_video_frame > 1)
- active_worst_quality = MIN(rc->worst_quality,
- rc->avg_frame_qindex[INTER_FRAME] * 5 / 4);
- else
- active_worst_quality = MIN(rc->worst_quality,
- rc->avg_frame_qindex[KEY_FRAME] * 3 / 2);
+ return rc->worst_quality;
+ // For ambient_qp we use minimum of avg_frame_qindex[KEY_FRAME/INTER_FRAME]
+ // for the first few frames following key frame. These are both initialized
+ // to worst_quality and updated with (3/4, 1/4) average in postencode_update.
+ // So for first few frames following key, the qp of that key frame is weighted
+ // into the active_worst_quality setting.
+ ambient_qp = (cm->current_video_frame < 5) ?
+ MIN(rc->avg_frame_qindex[INTER_FRAME], rc->avg_frame_qindex[KEY_FRAME]) :
+ rc->avg_frame_qindex[INTER_FRAME];
+ active_worst_quality = MIN(rc->worst_quality,
+ ambient_qp * 5 / 4);
if (rc->buffer_level > rc->optimal_buffer_level) {
// Adjust down.
// Maximum limit for down adjustment, ~30%.
@@ -592,12 +615,11 @@
if (critical_level) {
buff_lvl_step = (rc->optimal_buffer_level - critical_level);
if (buff_lvl_step) {
- adjustment =
- (int)((rc->worst_quality - rc->avg_frame_qindex[INTER_FRAME]) *
- (rc->optimal_buffer_level - rc->buffer_level) /
- buff_lvl_step);
+ adjustment = (int)((rc->worst_quality - ambient_qp) *
+ (rc->optimal_buffer_level - rc->buffer_level) /
+ buff_lvl_step);
}
- active_worst_quality = rc->avg_frame_qindex[INTER_FRAME] + adjustment;
+ active_worst_quality = ambient_qp + adjustment;
}
} else {
// Set to worst_quality if buffer is below critical level.
diff --git a/vp9/encoder/vp9_ratectrl.h b/vp9/encoder/vp9_ratectrl.h
index 2bc5b59..6200b39 100644
--- a/vp9/encoder/vp9_ratectrl.h
+++ b/vp9/encoder/vp9_ratectrl.h
@@ -99,7 +99,14 @@
int64_t starting_buffer_level;
int64_t optimal_buffer_level;
int64_t maximum_buffer_size;
- // int active_best_quality;
+ // rate control history for last frame(1) and the frame before(2).
+ // -1: undershot
+ // 1: overshoot
+ // 0: not initialized.
+ int rc_1_frame;
+ int rc_2_frame;
+ int q_1_frame;
+ int q_2_frame;
} RATE_CONTROL;
struct VP9_COMP;
diff --git a/vp9/encoder/vp9_rd.h b/vp9/encoder/vp9_rd.h
index ebbe821..fe000c0 100644
--- a/vp9/encoder/vp9_rd.h
+++ b/vp9/encoder/vp9_rd.h
@@ -101,17 +101,11 @@
int thresh_mult_sub8x8[MAX_REFS];
int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
- int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
- int mode_map[BLOCK_SIZES][MAX_MODES];
-
- int64_t comp_pred_diff[REFERENCE_MODES];
int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
- int64_t tx_select_diff[TX_MODES];
// TODO(agrange): can this overflow?
int tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
- int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
int64_t mask_filter;
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index 1994f48..126d142 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -458,8 +458,6 @@
cpi->full_search_sad = vp9_full_search_sad;
cpi->diamond_search_sad = oxcf->mode == BEST ? vp9_full_range_search
: vp9_diamond_search_sad;
- cpi->refining_search_sad = vp9_refining_search_sad;
-
// Slow quant, dct and trellis not worthwhile for first pass
// so make sure they are always turned off.
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index 8d3ca0d..184322f 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -77,6 +77,7 @@
lc->target_bandwidth = oxcf->ts_target_bitrate[layer];
lrc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
lrc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
+ lrc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
} else {
lc->target_bandwidth = oxcf->ss_target_bitrate[layer];
lrc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 74c9cfd..dfde606 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -311,7 +311,7 @@
const scan_order *so;
const int ref = is_inter_block(mbmi);
unsigned int (*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
- cpi->coef_counts[tx_size][type][ref];
+ cpi->frame_counts->coef_counts[tx_size][type][ref];
vp9_prob (*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
cpi->common.fc->coef_probs[tx_size][type][ref];
unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
