Merge "Unified approach for backward probability update."
diff --git a/configure b/configure
index 621161c..729c986 100755
--- a/configure
+++ b/configure
@@ -249,7 +249,6 @@
unistd_h
"
EXPERIMENT_LIST="
- oneshotq
multiple_arf
non420
alpha
diff --git a/vp9/decoder/vp9_decodframe.c b/vp9/decoder/vp9_decodframe.c
index 63b889d..1fd9e97 100644
--- a/vp9/decoder/vp9_decodframe.c
+++ b/vp9/decoder/vp9_decodframe.c
@@ -41,6 +41,7 @@
VP9_COMMON *cm;
vp9_reader bit_reader;
DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+ DECLARE_ALIGNED(16, unsigned char, token_cache[1024]);
} TileWorkerData;
static int read_be32(const uint8_t *p) {
@@ -297,6 +298,7 @@
VP9_COMMON *cm;
MACROBLOCKD *xd;
vp9_reader *r;
+ unsigned char* token_cache;
};
static void predict_and_reconstruct_intra_block(int plane, int block,
@@ -326,7 +328,7 @@
if (!mi->mbmi.skip_coeff) {
vp9_decode_block_tokens(cm, xd, plane, block, plane_bsize, tx_size,
- args->r);
+ args->r, args->token_cache);
inverse_transform_block(xd, plane, block, plane_bsize, tx_size);
}
}
@@ -336,6 +338,7 @@
MACROBLOCKD *xd;
vp9_reader *r;
int *eobtotal;
+ unsigned char* token_cache;
};
static void reconstruct_inter_block(int plane, int block,
@@ -346,7 +349,8 @@
MACROBLOCKD *const xd = args->xd;
*args->eobtotal += vp9_decode_block_tokens(cm, xd, plane, block,
- plane_bsize, tx_size, args->r);
+ plane_bsize, tx_size,
+ args->r, args->token_cache);
inverse_transform_block(xd, plane, block, plane_bsize, tx_size);
}
@@ -398,7 +402,8 @@
static void decode_modes_b(VP9_COMMON *const cm, MACROBLOCKD *const xd,
const TileInfo *const tile,
int mi_row, int mi_col,
- vp9_reader *r, BLOCK_SIZE bsize) {
+ vp9_reader *r, BLOCK_SIZE bsize,
+ unsigned char *token_cache) {
const int less8x8 = bsize < BLOCK_8X8;
MB_MODE_INFO *mbmi;
@@ -420,7 +425,7 @@
}
if (!is_inter_block(mbmi)) {
- struct intra_args arg = { cm, xd, r };
+ struct intra_args arg = { cm, xd, r, token_cache };
foreach_transformed_block(xd, bsize, predict_and_reconstruct_intra_block,
&arg);
} else {
@@ -438,7 +443,7 @@
// Reconstruction
if (!mbmi->skip_coeff) {
int eobtotal = 0;
- struct inter_args arg = { cm, xd, r, &eobtotal };
+ struct inter_args arg = { cm, xd, r, &eobtotal, token_cache };
foreach_transformed_block(xd, bsize, reconstruct_inter_block, &arg);
if (!less8x8 && eobtotal == 0)
mbmi->skip_coeff = 1; // skip loopfilter
@@ -477,7 +482,8 @@
static void decode_modes_sb(VP9_COMMON *const cm, MACROBLOCKD *const xd,
const TileInfo *const tile,
int mi_row, int mi_col,
- vp9_reader* r, BLOCK_SIZE bsize) {
+ vp9_reader* r, BLOCK_SIZE bsize,
+ unsigned char *token_cache) {
const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
PARTITION_TYPE partition;
BLOCK_SIZE subsize;
@@ -488,27 +494,33 @@
partition = read_partition(cm, xd, hbs, mi_row, mi_col, bsize, r);
subsize = get_subsize(bsize, partition);
if (subsize < BLOCK_8X8) {
- decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize);
+ decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
} else {
switch (partition) {
case PARTITION_NONE:
- decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize);
+ decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
break;
case PARTITION_HORZ:
- decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize);
+ decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
if (mi_row + hbs < cm->mi_rows)
- decode_modes_b(cm, xd, tile, mi_row + hbs, mi_col, r, subsize);
+ decode_modes_b(cm, xd, tile, mi_row + hbs, mi_col, r, subsize,
+ token_cache);
break;
case PARTITION_VERT:
- decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize);
+ decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
if (mi_col + hbs < cm->mi_cols)
- decode_modes_b(cm, xd, tile, mi_row, mi_col + hbs, r, subsize);
+ decode_modes_b(cm, xd, tile, mi_row, mi_col + hbs, r, subsize,
+ token_cache);
break;
case PARTITION_SPLIT:
- decode_modes_sb(cm, xd, tile, mi_row, mi_col, r, subsize);
- decode_modes_sb(cm, xd, tile, mi_row, mi_col + hbs, r, subsize);
- decode_modes_sb(cm, xd, tile, mi_row + hbs, mi_col, r, subsize);
- decode_modes_sb(cm, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize);
+ decode_modes_sb(cm, xd, tile, mi_row, mi_col, r, subsize,
+ token_cache);
+ decode_modes_sb(cm, xd, tile, mi_row, mi_col + hbs, r, subsize,
+ token_cache);
+ decode_modes_sb(cm, xd, tile, mi_row + hbs, mi_col, r, subsize,
+ token_cache);
+ decode_modes_sb(cm, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize,
+ token_cache);
break;
default:
assert(!"Invalid partition type");
@@ -791,7 +803,8 @@
vp9_zero(xd->left_seg_context);
for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
mi_col += MI_BLOCK_SIZE)
- decode_modes_sb(cm, xd, tile, mi_row, mi_col, r, BLOCK_64X64);
+ decode_modes_sb(cm, xd, tile, mi_row, mi_col, r, BLOCK_64X64,
+ pbi->token_cache);
if (pbi->do_loopfilter_inline) {
const int lf_start = mi_row - MI_BLOCK_SIZE;
@@ -935,7 +948,7 @@
}
static int tile_worker_hook(void *arg1, void *arg2) {
- TileWorkerData *const tile_data = (TileWorkerData*)arg1;
+ TileWorkerData *tile_data = (TileWorkerData*)arg1;
const TileInfo *const tile = (TileInfo*)arg2;
int mi_row, mi_col;
@@ -944,9 +957,11 @@
vp9_zero(tile_data->xd.left_context);
vp9_zero(tile_data->xd.left_seg_context);
for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
- mi_col += MI_BLOCK_SIZE)
+ mi_col += MI_BLOCK_SIZE) {
decode_modes_sb(tile_data->cm, &tile_data->xd, tile,
- mi_row, mi_col, &tile_data->bit_reader, BLOCK_64X64);
+ mi_row, mi_col, &tile_data->bit_reader, BLOCK_64X64,
+ tile_data->token_cache);
+ }
}
return !tile_data->xd.corrupted;
}
diff --git a/vp9/decoder/vp9_detokenize.c b/vp9/decoder/vp9_detokenize.c
index 6ecce28..70d0d74 100644
--- a/vp9/decoder/vp9_detokenize.c
+++ b/vp9/decoder/vp9_detokenize.c
@@ -70,7 +70,6 @@
DCT_EOB_MODEL_TOKEN : TWO_TOKEN) : \
token]; \
} \
- token_cache[scan[c]] = vp9_pt_energy_class[token]; \
} while (0)
#define WRITE_COEF_CONTINUE(val, token) \
@@ -78,6 +77,7 @@
qcoeff_ptr[scan[c]] = vp9_read_and_apply_sign(r, val) * \
dq[c > 0] / (1 + (tx_size == TX_32X32)); \
INCREMENT_COUNT(token); \
+ token_cache[scan[c]] = vp9_pt_energy_class[token]; \
c++; \
continue; \
}
@@ -91,7 +91,8 @@
static int decode_coefs(VP9_COMMON *cm, const MACROBLOCKD *xd,
vp9_reader *r, int block_idx,
PLANE_TYPE type, int seg_eob, int16_t *qcoeff_ptr,
- TX_SIZE tx_size, const int16_t *dq, int pt) {
+ TX_SIZE tx_size, const int16_t *dq, int pt,
+ uint8_t *token_cache) {
const FRAME_CONTEXT *const fc = &cm->fc;
FRAME_COUNTS *const counts = &cm->counts;
const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
@@ -104,7 +105,6 @@
vp9_coeff_count_model *coef_counts = counts->coef[tx_size];
const int16_t *scan, *nb;
const uint8_t *const band_translate = get_band_translate(tx_size);
- uint8_t token_cache[1024];
get_scan(xd, tx_size, type, block_idx, &scan, &nb);
while (1) {
@@ -131,6 +131,7 @@
if (!vp9_read(r, prob[ZERO_CONTEXT_NODE])) {
INCREMENT_COUNT(ZERO_TOKEN);
+ token_cache[scan[c]] = vp9_pt_energy_class[ZERO_TOKEN];
++c;
goto SKIP_START;
}
@@ -212,7 +213,8 @@
int vp9_decode_block_tokens(VP9_COMMON *cm, MACROBLOCKD *xd,
int plane, int block, BLOCK_SIZE plane_bsize,
- TX_SIZE tx_size, vp9_reader *r) {
+ TX_SIZE tx_size, vp9_reader *r,
+ uint8_t *token_cache) {
struct macroblockd_plane *const pd = &xd->plane[plane];
const int seg_eob = get_tx_eob(&cm->seg, xd->mi_8x8[0]->mbmi.segment_id,
tx_size);
@@ -223,7 +225,7 @@
eob = decode_coefs(cm, xd, r, block,
pd->plane_type, seg_eob, BLOCK_OFFSET(pd->qcoeff, block),
- tx_size, pd->dequant, pt);
+ tx_size, pd->dequant, pt, token_cache);
set_contexts(xd, pd, plane_bsize, tx_size, eob > 0, aoff, loff);
diff --git a/vp9/decoder/vp9_detokenize.h b/vp9/decoder/vp9_detokenize.h
index 94dd8e4..04939ea 100644
--- a/vp9/decoder/vp9_detokenize.h
+++ b/vp9/decoder/vp9_detokenize.h
@@ -17,6 +17,7 @@
int vp9_decode_block_tokens(VP9_COMMON *cm, MACROBLOCKD *xd,
int plane, int block, BLOCK_SIZE plane_bsize,
- TX_SIZE tx_size, vp9_reader *r);
+ TX_SIZE tx_size, vp9_reader *r,
+ uint8_t *token_cache);
#endif // VP9_DECODER_VP9_DETOKENIZE_H_
diff --git a/vp9/decoder/vp9_onyxd_int.h b/vp9/decoder/vp9_onyxd_int.h
index 83ea967..7c4c9db 100644
--- a/vp9/decoder/vp9_onyxd_int.h
+++ b/vp9/decoder/vp9_onyxd_int.h
@@ -49,6 +49,8 @@
ENTROPY_CONTEXT *above_context[MAX_MB_PLANE];
PARTITION_CONTEXT *above_seg_context;
+
+ DECLARE_ALIGNED(16, unsigned char, token_cache[1024]);
} VP9D_COMP;
#endif // VP9_DECODER_VP9_ONYXD_INT_H_
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index c83954e..c3dbc86 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -2169,17 +2169,14 @@
cpi->ni_av_qi = tmp_q;
cpi->avg_q = vp9_convert_qindex_to_q(tmp_q);
-#ifndef ONE_SHOT_Q_ESTIMATE
// Limit the maxq value returned subsequently.
// This increases the risk of overspend or underspend if the initial
// estimate for the clip is bad, but helps prevent excessive
// variation in Q, especially near the end of a clip
// where for example a small overspend may cause Q to crash
adjust_maxq_qrange(cpi);
-#endif
}
-#ifndef ONE_SHOT_Q_ESTIMATE
// The last few frames of a clip almost always have to few or too many
// bits and for the sake of over exact rate control we dont want to make
// radical adjustments to the allowed quantizer range just to use up a
@@ -2202,7 +2199,6 @@
cpi->active_worst_quality =
adjust_active_maxq(cpi->active_worst_quality, tmp_q);
}
-#endif
}
vp9_zero(this_frame);
if (EOF == input_stats(cpi, &this_frame))
diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c
index 0b05cf2..1d3170a 100644
--- a/vp9/encoder/vp9_onyx_if.c
+++ b/vp9/encoder/vp9_onyx_if.c
@@ -1180,7 +1180,6 @@
int i;
cpi->oxcf = *oxcf;
- cpi->goldfreq = 7;
cm->version = oxcf->version;
@@ -2851,19 +2850,11 @@
if (cpi->oxcf.end_usage == USAGE_CONSTANT_QUALITY) {
cpi->active_best_quality = cpi->cq_target_quality;
} else {
-#ifdef ONE_SHOT_Q_ESTIMATE
-#ifdef STRICT_ONE_SHOT_Q
- cpi->active_best_quality = q;
-#else
- cpi->active_best_quality = inter_minq[q];
-#endif
-#else
cpi->active_best_quality = inter_minq[q];
// 1-pass: for now, use the average Q for the active_best, if its lower
// than active_worst.
if (cpi->pass == 0 && (cpi->avg_frame_qindex < q))
cpi->active_best_quality = inter_minq[cpi->avg_frame_qindex];
-#endif
// For the constrained quality mode we don't want
// q to fall below the cq level.
diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h
index 44f1e26..9429c7f 100644
--- a/vp9/encoder/vp9_onyx_int.h
+++ b/vp9/encoder/vp9_onyx_int.h
@@ -29,11 +29,6 @@
#include "vp9/common/vp9_findnearmv.h"
#include "vp9/encoder/vp9_lookahead.h"
-// Experimental rate control switches
-#if CONFIG_ONESHOTQ
-#define ONE_SHOT_Q_ESTIMATE 0
-#define STRICT_ONE_SHOT_Q 0
-#endif
#define DISABLE_RC_LONG_TERM_MEM 0
// #define MODE_TEST_HIT_STATS
@@ -506,14 +501,9 @@
int decimation_count;
// for real time encoding
- int avg_encode_time; // microsecond
- int avg_pick_mode_time; // microsecond
int speed;
- unsigned int cpu_freq; // Mhz
int compressor_speed;
- int interquantizer;
- int goldfreq;
int auto_worst_q;
int cpu_used;
int pass;
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 05928e0..993919e 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -1092,7 +1092,7 @@
else
x->fwd_txm4x4(src_diff, coeff, 8);
- vp9_regular_quantize_b_4x4(x, 16, block, scan, get_iscan_4x4(tx_type));
+ vp9_regular_quantize_b_4x4(x, 4, block, scan, get_iscan_4x4(tx_type));
ratey += cost_coeffs(x, 0, block,
tempa + idx, templ + idy, TX_4X4, scan, nb);
@@ -1559,7 +1559,7 @@
coeff = BLOCK_OFFSET(p->coeff, k);
x->fwd_txm4x4(raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
coeff, 8);
- vp9_regular_quantize_b_4x4(x, 16, k, get_scan_4x4(DCT_DCT),
+ vp9_regular_quantize_b_4x4(x, 4, k, get_scan_4x4(DCT_DCT),
get_iscan_4x4(DCT_DCT));
thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
16, &ssz);
@@ -1871,12 +1871,14 @@
mi_buf_restore(x, orig_src, orig_pre);
}
- if (has_second_rf && this_mode == NEWMV &&
- mbmi->interp_filter == EIGHTTAP) {
+ if (has_second_rf) {
if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
continue;
+ }
+ if (has_second_rf && this_mode == NEWMV &&
+ mbmi->interp_filter == EIGHTTAP) {
// adjust src pointers
mi_buf_shift(x, i);
if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
@@ -2660,6 +2662,12 @@
int orig_dst_stride[MAX_MB_PLANE];
int rs = 0;
+ if (is_comp_pred) {
+ if (frame_mv[refs[0]].as_int == INVALID_MV ||
+ frame_mv[refs[1]].as_int == INVALID_MV)
+ return INT64_MAX;
+ }
+
if (this_mode == NEWMV) {
int rate_mv;
if (is_comp_pred) {
@@ -2678,9 +2686,6 @@
&mbmi->ref_mvs[refs[1]][0].as_mv,
x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
}
- if (frame_mv[refs[0]].as_int == INVALID_MV ||
- frame_mv[refs[1]].as_int == INVALID_MV)
- return INT64_MAX;
*rate2 += rate_mv;
} else {
int_mv tmp_mv;
