Merge "Compressed/uncompressed frame header changes." into experimental
diff --git a/vp9/common/vp9_entropy.c b/vp9/common/vp9_entropy.c
index 1b7da6c..1ae3586 100644
--- a/vp9/common/vp9_entropy.c
+++ b/vp9/common/vp9_entropy.c
@@ -468,7 +468,7 @@
} else {
ctx = token_cache[scan[neighbors[MAX_NEIGHBORS * c + 0]]];
}
- return vp9_pt_energy_class[ctx];
+ return ctx;
}
};
@@ -642,6 +642,17 @@
#define COEF_COUNT_SAT_AFTER_KEY 24
#define COEF_MAX_UPDATE_FACTOR_AFTER_KEY 128
+void vp9_full_to_model_count(unsigned int *model_count,
+ unsigned int *full_count) {
+ int n;
+ model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
+ model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
+ model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
+ for (n = THREE_TOKEN; n < DCT_EOB_TOKEN; ++n)
+ model_count[TWO_TOKEN] += full_count[n];
+ model_count[DCT_EOB_MODEL_TOKEN] = full_count[DCT_EOB_TOKEN];
+}
+
void vp9_full_to_model_counts(
vp9_coeff_count_model *model_count, vp9_coeff_count *full_count) {
int i, j, k, l;
@@ -649,19 +660,10 @@
for (j = 0; j < REF_TYPES; ++j)
for (k = 0; k < COEF_BANDS; ++k)
for (l = 0; l < PREV_COEF_CONTEXTS; ++l) {
- int n;
if (l >= 3 && k == 0)
continue;
- model_count[i][j][k][l][ZERO_TOKEN] =
- full_count[i][j][k][l][ZERO_TOKEN];
- model_count[i][j][k][l][ONE_TOKEN] =
- full_count[i][j][k][l][ONE_TOKEN];
- model_count[i][j][k][l][TWO_TOKEN] =
- full_count[i][j][k][l][TWO_TOKEN];
- for (n = THREE_TOKEN; n < DCT_EOB_TOKEN; ++n)
- model_count[i][j][k][l][TWO_TOKEN] += full_count[i][j][k][l][n];
- model_count[i][j][k][l][DCT_EOB_MODEL_TOKEN] =
- full_count[i][j][k][l][DCT_EOB_TOKEN];
+ vp9_full_to_model_count(model_count[i][j][k][l],
+ full_count[i][j][k][l]);
}
}
diff --git a/vp9/common/vp9_entropy.h b/vp9/common/vp9_entropy.h
index 5d57f14..e76211a 100644
--- a/vp9/common/vp9_entropy.h
+++ b/vp9/common/vp9_entropy.h
@@ -174,6 +174,8 @@
typedef unsigned int vp9_coeff_stats_model[REF_TYPES][COEF_BANDS]
[PREV_COEF_CONTEXTS]
[UNCONSTRAINED_NODES][2];
+extern void vp9_full_to_model_count(unsigned int *model_count,
+ unsigned int *full_count);
extern void vp9_full_to_model_counts(
vp9_coeff_count_model *model_count, vp9_coeff_count *full_count);
diff --git a/vp9/common/vp9_reconintra.c b/vp9/common/vp9_reconintra.c
index dd60a76..b1f327b 100644
--- a/vp9/common/vp9_reconintra.c
+++ b/vp9/common/vp9_reconintra.c
@@ -402,7 +402,7 @@
(block_idx >> bwl) || xd->up_available;
const int have_left =
(block_idx & wmask) || xd->left_available;
- const int have_right = ((block_idx & wmask) != wmask);
+ int have_right = ((block_idx & wmask) != wmask);
const int txfm_block_size = 4 << tx_size;
assert(bwl >= 0);
diff --git a/vp9/decoder/vp9_detokenize.c b/vp9/decoder/vp9_detokenize.c
index 02900c0..890d5d0 100644
--- a/vp9/decoder/vp9_detokenize.c
+++ b/vp9/decoder/vp9_detokenize.c
@@ -58,13 +58,15 @@
254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129, 0
};
+DECLARE_ALIGNED(16, extern const uint8_t,
+ vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
#define INCREMENT_COUNT(token) \
do { \
coef_counts[type][ref][band][pt] \
[token >= TWO_TOKEN ? \
(token == DCT_EOB_TOKEN ? DCT_EOB_MODEL_TOKEN : TWO_TOKEN) : \
token]++; \
- token_cache[scan[c]] = token; \
+ token_cache[scan[c]] = vp9_pt_energy_class[token]; \
} while (0)
#define WRITE_COEF_CONTINUE(val, token) \
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index b19bf3a..13bfe21 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -1047,7 +1047,7 @@
fclose(f);
}
-static void build_tree_distribution(vp9_coeff_probs *coef_probs,
+static void build_tree_distribution(vp9_coeff_probs_model *coef_probs,
vp9_coeff_count *coef_counts,
unsigned int (*eob_branch_ct)[REF_TYPES]
[COEF_BANDS]
@@ -1056,12 +1056,13 @@
VP9_COMP *cpi,
vp9_coeff_accum *context_counters,
#endif
- vp9_coeff_stats *coef_branch_ct,
+ vp9_coeff_stats_model *coef_branch_ct,
int block_types) {
int i, j, k, l;
#ifdef ENTROPY_STATS
int t = 0;
#endif
+ unsigned int model_counts[UNCONSTRAINED_NODES + 1];
for (i = 0; i < block_types; ++i) {
for (j = 0; j < REF_TYPES; ++j) {
@@ -1069,10 +1070,11 @@
for (l = 0; l < PREV_COEF_CONTEXTS; ++l) {
if (l >= 3 && k == 0)
continue;
- vp9_tree_probs_from_distribution(vp9_coef_tree,
+ vp9_full_to_model_count(model_counts, coef_counts[i][j][k][l]);
+ vp9_tree_probs_from_distribution(vp9_coefmodel_tree,
coef_probs[i][j][k][l],
coef_branch_ct[i][j][k][l],
- coef_counts[i][j][k][l], 0);
+ model_counts, 0);
coef_branch_ct[i][j][k][l][0][1] = eob_branch_ct[i][j][k][l] -
coef_branch_ct[i][j][k][l][0][0];
coef_probs[i][j][k][l][0] =
@@ -1129,9 +1131,9 @@
#ifdef ENTROPY_STATS
vp9_coeff_stats *tree_update_hist,
#endif
- vp9_coeff_probs *new_frame_coef_probs,
+ vp9_coeff_probs_model *new_frame_coef_probs,
vp9_coeff_probs_model *old_frame_coef_probs,
- vp9_coeff_stats *frame_branch_ct,
+ vp9_coeff_stats_model *frame_branch_ct,
TX_SIZE tx_size) {
int i, j, k, l, t;
int update[2] = {0, 0};
diff --git a/vp9/encoder/vp9_block.h b/vp9/encoder/vp9_block.h
index 211eca4..0e9b680 100644
--- a/vp9/encoder/vp9_block.h
+++ b/vp9/encoder/vp9_block.h
@@ -117,7 +117,6 @@
int mbmode_cost[2][MB_MODE_COUNT];
int intra_uv_mode_cost[2][MB_MODE_COUNT];
int bmode_costs[VP9_BINTRAMODES][VP9_BINTRAMODES][VP9_BINTRAMODES];
- int inter_bmode_costs[INTRA_MODE_COUNT];
int switchable_interp_costs[VP9_SWITCHABLE_FILTERS + 1]
[VP9_SWITCHABLE_FILTERS];
diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c
index b7f60b1..755ff21 100644
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -20,6 +20,9 @@
#include "vp9/common/vp9_systemdependent.h"
#include "vp9_rtcd.h"
+DECLARE_ALIGNED(16, extern const uint8_t,
+ vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
+
void vp9_subtract_block(int rows, int cols,
int16_t *diff_ptr, int diff_stride,
const uint8_t *src_ptr, int src_stride,
@@ -105,7 +108,7 @@
uint8_t *token_cache,
int pad, int l) {
int bak = token_cache[scan[idx]], pt;
- token_cache[scan[idx]] = token;
+ token_cache[scan[idx]] = vp9_pt_energy_class[token];
pt = vp9_get_coef_context(scan, nb, pad, token_cache, idx + 1, l);
token_cache[scan[idx]] = bak;
return pt;
@@ -189,7 +192,8 @@
*(tokens[eob] + 1) = *(tokens[eob] + 0);
next = eob;
for (i = 0; i < eob; i++)
- token_cache[scan[i]] = vp9_dct_value_tokens_ptr[qcoeff_ptr[scan[i]]].token;
+ token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
+ qcoeff_ptr[scan[i]]].token];
nb = vp9_get_coef_neighbors_handle(scan, &pad);
for (i = eob; i-- > i0;) {
@@ -610,6 +614,7 @@
struct encode_b_args* const args = arg;
MACROBLOCK* const x = args->x;
MACROBLOCKD* const xd = &x->e_mbd;
+ MB_MODE_INFO* const mbmi = &xd->mode_info_context->mbmi;
const TX_SIZE tx_size = (TX_SIZE)(ss_txfrm_size / 2);
const int bw = 4 << (b_width_log2(bsize) - xd->plane[plane].subsampling_x);
const int raster_block = txfrm_block_to_raster_block(xd, bsize, plane,
@@ -634,9 +639,9 @@
TX_TYPE tx_type;
int mode, b_mode;
- mode = plane == 0? xd->mode_info_context->mbmi.mode:
- xd->mode_info_context->mbmi.uv_mode;
- if (bsize <= BLOCK_SIZE_SB8X8 && mode == I4X4_PRED && plane == 0)
+ mode = plane == 0? mbmi->mode: mbmi->uv_mode;
+ if (mbmi->sb_type < BLOCK_SIZE_SB8X8 && plane == 0 &&
+ mbmi->ref_frame == INTRA_FRAME)
b_mode = xd->mode_info_context->bmi[ib].as_mode.first;
else
b_mode = mode;
diff --git a/vp9/encoder/vp9_modecosts.c b/vp9/encoder/vp9_modecosts.c
index 171b44b..67658f5 100644
--- a/vp9/encoder/vp9_modecosts.c
+++ b/vp9/encoder/vp9_modecosts.c
@@ -17,7 +17,6 @@
void vp9_init_mode_costs(VP9_COMP *c) {
VP9_COMMON *x = &c->common;
- const vp9_tree_p T = vp9_bmode_tree;
const vp9_tree_p KT = vp9_bmode_tree;
int i, j;
@@ -28,8 +27,6 @@
}
}
- vp9_cost_tokens((int *)c->mb.inter_bmode_costs, x->fc.bmode_prob, T);
-
// TODO(rbultje) separate tables for superblock costing?
vp9_cost_tokens(c->mb.mbmode_cost[1], x->fc.sb_ymode_prob,
vp9_sb_ymode_tree);
diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h
index 15f9571..24a2acb 100644
--- a/vp9/encoder/vp9_onyx_int.h
+++ b/vp9/encoder/vp9_onyx_int.h
@@ -422,20 +422,20 @@
nmv_context_counts NMVcount;
vp9_coeff_count coef_counts_4x4[BLOCK_TYPES];
- vp9_coeff_probs frame_coef_probs_4x4[BLOCK_TYPES];
- vp9_coeff_stats frame_branch_ct_4x4[BLOCK_TYPES];
+ vp9_coeff_probs_model frame_coef_probs_4x4[BLOCK_TYPES];
+ vp9_coeff_stats_model frame_branch_ct_4x4[BLOCK_TYPES];
vp9_coeff_count coef_counts_8x8[BLOCK_TYPES];
- vp9_coeff_probs frame_coef_probs_8x8[BLOCK_TYPES];
- vp9_coeff_stats frame_branch_ct_8x8[BLOCK_TYPES];
+ vp9_coeff_probs_model frame_coef_probs_8x8[BLOCK_TYPES];
+ vp9_coeff_stats_model frame_branch_ct_8x8[BLOCK_TYPES];
vp9_coeff_count coef_counts_16x16[BLOCK_TYPES];
- vp9_coeff_probs frame_coef_probs_16x16[BLOCK_TYPES];
- vp9_coeff_stats frame_branch_ct_16x16[BLOCK_TYPES];
+ vp9_coeff_probs_model frame_coef_probs_16x16[BLOCK_TYPES];
+ vp9_coeff_stats_model frame_branch_ct_16x16[BLOCK_TYPES];
vp9_coeff_count coef_counts_32x32[BLOCK_TYPES];
- vp9_coeff_probs frame_coef_probs_32x32[BLOCK_TYPES];
- vp9_coeff_stats frame_branch_ct_32x32[BLOCK_TYPES];
+ vp9_coeff_probs_model frame_coef_probs_32x32[BLOCK_TYPES];
+ vp9_coeff_stats_model frame_branch_ct_32x32[BLOCK_TYPES];
int gfu_boost;
int last_boost;
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 2eb3f9b..e649192 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -46,6 +46,9 @@
/* Factor to weigh the rate for switchable interp filters */
#define SWITCHABLE_INTERP_RATE_FACTOR 1
+DECLARE_ALIGNED(16, extern const uint8_t,
+ vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
+
const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
{ZEROMV, LAST_FRAME, NONE},
{DC_PRED, INTRA_FRAME, NONE},
@@ -366,7 +369,7 @@
if (!c || token_cache[scan[c - 1]])
cost += vp9_cost_bit(coef_probs[band][pt][0], 1);
- token_cache[scan[c]] = t;
+ token_cache[scan[c]] = vp9_pt_energy_class[t];
}
if (c < seg_eob) {
if (c)
@@ -611,7 +614,6 @@
int64_t this_rd;
int ratey = 0;
- xd->mode_info_context->bmi[ib].as_mode.first = mode;
if (cm->frame_type == KEY_FRAME)
rate = bmode_costs[mode];
else
@@ -653,9 +655,6 @@
distortion += vp9_block_error(coeff, BLOCK_OFFSET(xd->plane[0].dqcoeff,
block, 16), 16) >> 2;
- vp9_intra4x4_predict(xd, block, BLOCK_SIZE_SB8X8, mode,
- dst, xd->plane[0].dst.stride);
-
if (best_tx_type != DCT_DCT)
vp9_short_iht4x4_add(BLOCK_OFFSET(xd->plane[0].dqcoeff, block, 16),
dst, xd->plane[0].dst.stride, best_tx_type);
@@ -731,7 +730,7 @@
vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
xd->mode_info_context->mbmi.mode = I4X4_PRED;
- bmode_costs = mb->inter_bmode_costs;
+ bmode_costs = mb->mbmode_cost[cpi->common.frame_type];
for (idy = 0; idy < 2; idy += bh) {
for (idx = 0; idx < 2; idx += bw) {
@@ -939,7 +938,7 @@
}
static int labels2mode(MACROBLOCK *x,
- int const *labelings, int which_label,
+ int const *labelings, int i,
MB_PREDICTION_MODE this_mode,
int_mv *this_mv, int_mv *this_second_mv,
int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
@@ -950,7 +949,7 @@
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *const mic = xd->mode_info_context;
MB_MODE_INFO * mbmi = &mic->mbmi;
- int i, cost = 0, thismvcost = 0;
+ int cost = 0, thismvcost = 0;
int idx, idy;
int bw = 1 << b_width_log2(mbmi->sb_type);
int bh = 1 << b_height_log2(mbmi->sb_type);
@@ -958,72 +957,65 @@
/* We have to be careful retrieving previously-encoded motion vectors.
Ones from this macroblock have to be pulled from the BLOCKD array
as they have not yet made it to the bmi array in our MB_MODE_INFO. */
- for (i = 0; i < 4; ++i) {
- MB_PREDICTION_MODE m;
+ MB_PREDICTION_MODE m;
- if (labelings[i] != which_label)
- continue;
-
- {
- // the only time we should do costing for new motion vector or mode
- // is when we are on a new label (jbb May 08, 2007)
- switch (m = this_mode) {
- case NEWMV:
- if (mbmi->second_ref_frame > 0) {
- this_mv->as_int = seg_mvs[mbmi->ref_frame - 1].as_int;
- this_second_mv->as_int =
- seg_mvs[mbmi->second_ref_frame - 1].as_int;
- }
-
- thismvcost = vp9_mv_bit_cost(this_mv, best_ref_mv, mvjcost, mvcost,
- 102, xd->allow_high_precision_mv);
- if (mbmi->second_ref_frame > 0) {
- thismvcost += vp9_mv_bit_cost(this_second_mv, second_best_ref_mv,
- mvjcost, mvcost, 102,
- xd->allow_high_precision_mv);
- }
- break;
- case NEARESTMV:
- this_mv->as_int = frame_mv[NEARESTMV][mbmi->ref_frame].as_int;
- if (mbmi->second_ref_frame > 0)
- this_second_mv->as_int =
- frame_mv[NEARESTMV][mbmi->second_ref_frame].as_int;
- break;
- case NEARMV:
- this_mv->as_int = frame_mv[NEARMV][mbmi->ref_frame].as_int;
- if (mbmi->second_ref_frame > 0)
- this_second_mv->as_int =
- frame_mv[NEARMV][mbmi->second_ref_frame].as_int;
- break;
- case ZEROMV:
- this_mv->as_int = 0;
- if (mbmi->second_ref_frame > 0)
- this_second_mv->as_int = 0;
- break;
- default:
- break;
+ // the only time we should do costing for new motion vector or mode
+ // is when we are on a new label (jbb May 08, 2007)
+ switch (m = this_mode) {
+ case NEWMV:
+ if (mbmi->second_ref_frame > 0) {
+ this_mv->as_int = seg_mvs[mbmi->ref_frame - 1].as_int;
+ this_second_mv->as_int =
+ seg_mvs[mbmi->second_ref_frame - 1].as_int;
}
- cost = vp9_cost_mv_ref(cpi, this_mode,
- mbmi->mb_mode_context[mbmi->ref_frame]);
- }
-
- mic->bmi[i].as_mv[0].as_int = this_mv->as_int;
- if (mbmi->second_ref_frame > 0)
- mic->bmi[i].as_mv[1].as_int = this_second_mv->as_int;
-
- x->partition_info->bmi[i].mode = m;
- x->partition_info->bmi[i].mv.as_int = this_mv->as_int;
- if (mbmi->second_ref_frame > 0)
- x->partition_info->bmi[i].second_mv.as_int = this_second_mv->as_int;
- for (idy = 0; idy < bh; ++idy) {
- for (idx = 0; idx < bw; ++idx) {
- vpx_memcpy(&mic->bmi[i + idy * 2 + idx],
- &mic->bmi[i], sizeof(mic->bmi[i]));
- vpx_memcpy(&x->partition_info->bmi[i + idy * 2 + idx],
- &x->partition_info->bmi[i],
- sizeof(x->partition_info->bmi[i]));
+ thismvcost = vp9_mv_bit_cost(this_mv, best_ref_mv, mvjcost, mvcost,
+ 102, xd->allow_high_precision_mv);
+ if (mbmi->second_ref_frame > 0) {
+ thismvcost += vp9_mv_bit_cost(this_second_mv, second_best_ref_mv,
+ mvjcost, mvcost, 102,
+ xd->allow_high_precision_mv);
}
+ break;
+ case NEARESTMV:
+ this_mv->as_int = frame_mv[NEARESTMV][mbmi->ref_frame].as_int;
+ if (mbmi->second_ref_frame > 0)
+ this_second_mv->as_int =
+ frame_mv[NEARESTMV][mbmi->second_ref_frame].as_int;
+ break;
+ case NEARMV:
+ this_mv->as_int = frame_mv[NEARMV][mbmi->ref_frame].as_int;
+ if (mbmi->second_ref_frame > 0)
+ this_second_mv->as_int =
+ frame_mv[NEARMV][mbmi->second_ref_frame].as_int;
+ break;
+ case ZEROMV:
+ this_mv->as_int = 0;
+ if (mbmi->second_ref_frame > 0)
+ this_second_mv->as_int = 0;
+ break;
+ default:
+ break;
+ }
+
+ cost = vp9_cost_mv_ref(cpi, this_mode,
+ mbmi->mb_mode_context[mbmi->ref_frame]);
+
+ mic->bmi[i].as_mv[0].as_int = this_mv->as_int;
+ if (mbmi->second_ref_frame > 0)
+ mic->bmi[i].as_mv[1].as_int = this_second_mv->as_int;
+
+ x->partition_info->bmi[i].mode = m;
+ x->partition_info->bmi[i].mv.as_int = this_mv->as_int;
+ if (mbmi->second_ref_frame > 0)
+ x->partition_info->bmi[i].second_mv.as_int = this_second_mv->as_int;
+ for (idy = 0; idy < bh; ++idy) {
+ for (idx = 0; idx < bw; ++idx) {
+ vpx_memcpy(&mic->bmi[i + idy * 2 + idx],
+ &mic->bmi[i], sizeof(mic->bmi[i]));
+ vpx_memcpy(&x->partition_info->bmi[i + idy * 2 + idx],
+ &x->partition_info->bmi[i],
+ sizeof(x->partition_info->bmi[i]));
}
}
@@ -1882,6 +1874,160 @@
return SWITCHABLE_INTERP_RATE_FACTOR * x->switchable_interp_costs[c][m];
}
+static void iterative_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+ BLOCK_SIZE_TYPE bsize,
+ int_mv *frame_mv,
+ YV12_BUFFER_CONFIG **scaled_ref_frame,
+ int mi_row, int mi_col,
+ int_mv single_newmv[MAX_REF_FRAMES]) {
+ int pw = 4 << b_width_log2(bsize), ph = 4 << b_height_log2(bsize);
+ MACROBLOCKD *xd = &x->e_mbd;
+ MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
+ int refs[2] = { mbmi->ref_frame,
+ (mbmi->second_ref_frame < 0 ? 0 : mbmi->second_ref_frame) };
+ int_mv ref_mv[2];
+ const enum BlockSize block_size = get_plane_block_size(bsize, &xd->plane[0]);
+ int ite;
+ // Prediction buffer from second frame.
+ uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
+
+ // Do joint motion search in compound mode to get more accurate mv.
+ struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
+ struct buf_2d backup_second_yv12[MAX_MB_PLANE] = {{0}};
+ struct buf_2d scaled_first_yv12;
+ int last_besterr[2] = {INT_MAX, INT_MAX};
+
+ ref_mv[0] = mbmi->ref_mvs[refs[0]][0];
+ ref_mv[1] = mbmi->ref_mvs[refs[1]][0];
+
+ if (scaled_ref_frame[0]) {
+ int i;
+
+ // Swap out the reference frame for a version that's been scaled to
+ // match the resolution of the current frame, allowing the existing
+ // motion search code to be used without additional modifications.
+ for (i = 0; i < MAX_MB_PLANE; i++)
+ backup_yv12[i] = xd->plane[i].pre[0];
+
+ setup_pre_planes(xd, scaled_ref_frame[0], NULL, mi_row, mi_col,
+ NULL, NULL);
+ }
+
+ if (scaled_ref_frame[1]) {
+ int i;
+
+ for (i = 0; i < MAX_MB_PLANE; i++)
+ backup_second_yv12[i] = xd->plane[i].pre[1];
+
+ setup_pre_planes(xd, scaled_ref_frame[1], NULL, mi_row, mi_col,
+ NULL, NULL);
+ }
+ xd->scale_factor[0].set_scaled_offsets(&xd->scale_factor[0],
+ mi_row, mi_col);
+ xd->scale_factor[1].set_scaled_offsets(&xd->scale_factor[1],
+ mi_row, mi_col);
+
+ scaled_first_yv12 = xd->plane[0].pre[0];
+
+ // Initialize mv using single prediction mode result.
+ frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
+ frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
+
+ // Allow joint search multiple times iteratively for each ref frame
+ // and break out the search loop if it couldn't find better mv.
+ for (ite = 0; ite < 4; ite++) {
+ struct buf_2d ref_yv12[2];
+ int bestsme = INT_MAX;
+ int sadpb = x->sadperbit16;
+ int_mv tmp_mv;
+ int search_range = 3;
+
+ int tmp_col_min = x->mv_col_min;
+ int tmp_col_max = x->mv_col_max;
+ int tmp_row_min = x->mv_row_min;
+ int tmp_row_max = x->mv_row_max;
+ int id = ite % 2;
+
+ // Initialized here because of compiler problem in Visual Studio.
+ ref_yv12[0] = xd->plane[0].pre[0];
+ ref_yv12[1] = xd->plane[0].pre[1];
+
+ // Get pred block from second frame.
+ vp9_build_inter_predictor(ref_yv12[!id].buf,
+ ref_yv12[!id].stride,
+ second_pred, pw,
+ &frame_mv[refs[!id]],
+ &xd->scale_factor[!id],
+ pw, ph, 0,
+ &xd->subpix);
+
+ // Compound motion search on first ref frame.
+ if (id)
+ xd->plane[0].pre[0] = ref_yv12[id];
+ vp9_clamp_mv_min_max(x, &ref_mv[id]);
+
+ // Use mv result from single mode as mvp.
+ tmp_mv.as_int = frame_mv[refs[id]].as_int;
+
+ tmp_mv.as_mv.col >>= 3;
+ tmp_mv.as_mv.row >>= 3;
+
+ // Small-range full-pixel motion search
+ bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
+ search_range,
+ &cpi->fn_ptr[block_size],
+ x->nmvjointcost, x->mvcost,
+ &ref_mv[id], second_pred,
+ pw, ph);
+
+ x->mv_col_min = tmp_col_min;
+ x->mv_col_max = tmp_col_max;
+ x->mv_row_min = tmp_row_min;
+ x->mv_row_max = tmp_row_max;
+
+ if (bestsme < INT_MAX) {
+ int dis; /* TODO: use dis in distortion calculation later. */
+ unsigned int sse;
+
+ bestsme = vp9_find_best_sub_pixel_comp(x, &tmp_mv,
+ &ref_mv[id],
+ x->errorperbit,
+ &cpi->fn_ptr[block_size],
+ x->nmvjointcost, x->mvcost,
+ &dis, &sse, second_pred,
+ pw, ph);
+ }
+
+ if (id)
+ xd->plane[0].pre[0] = scaled_first_yv12;
+
+ if (bestsme < last_besterr[id]) {
+ frame_mv[refs[id]].as_int =
+ xd->mode_info_context->bmi[0].as_mv[1].as_int = tmp_mv.as_int;
+ last_besterr[id] = bestsme;
+ } else {
+ break;
+ }
+ }
+
+ // restore the predictor
+ if (scaled_ref_frame[0]) {
+ int i;
+
+ for (i = 0; i < MAX_MB_PLANE; i++)
+ xd->plane[i].pre[0] = backup_yv12[i];
+ }
+
+ if (scaled_ref_frame[1]) {
+ int i;
+
+ for (i = 0; i < MAX_MB_PLANE; i++)
+ xd->plane[i].pre[1] = backup_second_yv12[i];
+ }
+
+ vpx_free(second_pred);
+}
+
static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize,
int64_t txfm_cache[],
@@ -1924,145 +2070,9 @@
ref_mv[1] = mbmi->ref_mvs[refs[1]][0];
if (is_comp_pred) {
- if (cpi->sf.comp_inter_joint_serach) {
- int pw = 4 << b_width_log2(bsize), ph = 4 << b_height_log2(bsize);
- int ite;
- // Prediction buffer from second frame.
- uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
-
- // Do joint motion search in compound mode to get more accurate mv.
- struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
- struct buf_2d backup_second_yv12[MAX_MB_PLANE] = {{0}};
- struct buf_2d scaled_first_yv12;
- int last_besterr[2] = {INT_MAX, INT_MAX};
-
- if (scaled_ref_frame[0]) {
- int i;
-
- // Swap out the reference frame for a version that's been scaled to
- // match the resolution of the current frame, allowing the existing
- // motion search code to be used without additional modifications.
- for (i = 0; i < MAX_MB_PLANE; i++)
- backup_yv12[i] = xd->plane[i].pre[0];
-
- setup_pre_planes(xd, scaled_ref_frame[0], NULL, mi_row, mi_col,
- NULL, NULL);
- }
-
- if (scaled_ref_frame[1]) {
- int i;
-
- for (i = 0; i < MAX_MB_PLANE; i++)
- backup_second_yv12[i] = xd->plane[i].pre[1];
-
- setup_pre_planes(xd, scaled_ref_frame[1], NULL, mi_row, mi_col,
- NULL, NULL);
- }
- xd->scale_factor[0].set_scaled_offsets(&xd->scale_factor[0],
- mi_row, mi_col);
- xd->scale_factor[1].set_scaled_offsets(&xd->scale_factor[1],
- mi_row, mi_col);
-
- scaled_first_yv12 = xd->plane[0].pre[0];
-
- // Initialize mv using single prediction mode result.
- frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
- frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
-
- // Allow joint search multiple times iteratively for each ref frame
- // and break out the search loop if it couldn't find better mv.
- for (ite = 0; ite < 4; ite++) {
- struct buf_2d ref_yv12[2];
- int bestsme = INT_MAX;
- int sadpb = x->sadperbit16;
- int_mv tmp_mv;
- int search_range = 3;
-
- int tmp_col_min = x->mv_col_min;
- int tmp_col_max = x->mv_col_max;
- int tmp_row_min = x->mv_row_min;
- int tmp_row_max = x->mv_row_max;
- int id = ite % 2;
-
- // Initialized here because of compiler problem in Visual Studio.
- ref_yv12[0] = xd->plane[0].pre[0];
- ref_yv12[1] = xd->plane[0].pre[1];
-
- // Get pred block from second frame.
- vp9_build_inter_predictor(ref_yv12[!id].buf,
- ref_yv12[!id].stride,
- second_pred, pw,
- &frame_mv[refs[!id]],
- &xd->scale_factor[!id],
- pw, ph, 0,
- &xd->subpix);
-
- // Compound motion search on first ref frame.
- if (id)
- xd->plane[0].pre[0] = ref_yv12[id];
- vp9_clamp_mv_min_max(x, &ref_mv[id]);
-
- // Use mv result from single mode as mvp.
- tmp_mv.as_int = frame_mv[refs[id]].as_int;
-
- tmp_mv.as_mv.col >>= 3;
- tmp_mv.as_mv.row >>= 3;
-
- // Small-range full-pixel motion search
- bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
- search_range,
- &cpi->fn_ptr[block_size],
- x->nmvjointcost, x->mvcost,
- &ref_mv[id], second_pred,
- pw, ph);
-
- x->mv_col_min = tmp_col_min;
- x->mv_col_max = tmp_col_max;
- x->mv_row_min = tmp_row_min;
- x->mv_row_max = tmp_row_max;
-
- if (bestsme < INT_MAX) {
- int dis; /* TODO: use dis in distortion calculation later. */
- unsigned int sse;
-
- bestsme = vp9_find_best_sub_pixel_comp(x, &tmp_mv,
- &ref_mv[id],
- x->errorperbit,
- &cpi->fn_ptr[block_size],
- x->nmvjointcost, x->mvcost,
- &dis, &sse, second_pred,
- pw, ph);
- }
-
- if (id)
- xd->plane[0].pre[0] = scaled_first_yv12;
-
- if (bestsme < last_besterr[id]) {
- frame_mv[refs[id]].as_int =
- xd->mode_info_context->bmi[0].as_mv[1].as_int = tmp_mv.as_int;
- last_besterr[id] = bestsme;
- } else {
- break;
- }
- }
-
- // restore the predictor
- if (scaled_ref_frame[0]) {
- int i;
-
- for (i = 0; i < MAX_MB_PLANE; i++)
- xd->plane[i].pre[0] = backup_yv12[i];
- }
-
- if (scaled_ref_frame[1]) {
- int i;
-
- for (i = 0; i < MAX_MB_PLANE; i++)
- xd->plane[i].pre[1] = backup_second_yv12[i];
- }
-
- vpx_free(second_pred);
- }
+ if (cpi->sf.comp_inter_joint_serach)
+ iterative_motion_search(cpi, x, bsize, frame_mv, scaled_ref_frame,
+ mi_row, mi_col, single_newmv);
if (frame_mv[refs[0]].as_int == INVALID_MV ||
frame_mv[refs[1]].as_int == INVALID_MV)
@@ -2684,8 +2694,6 @@
if (this_mode == I4X4_PRED) {
int rate;
- // Note the rate value returned here includes the cost of coding
- // the I4X4_PRED mode : x->mbmode_cost[xd->frame_type][I4X4_PRED];
mbmi->txfm_size = TX_4X4;
rd_pick_intra4x4mby_modes(cpi, x, &rate, &rate_y,
&distortion_y, INT64_MAX);
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 08efc84..eb79de1 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -36,6 +36,9 @@
extern vp9_coeff_stats tree_update_hist_32x32[BLOCK_TYPES];
#endif /* ENTROPY_STATS */
+DECLARE_ALIGNED(16, extern const uint8_t,
+ vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
+
static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2];
const TOKENVALUE *vp9_dct_value_tokens_ptr;
static int dct_value_cost[DCT_MAX_VALUE * 2];
@@ -228,7 +231,7 @@
if (!t->skip_eob_node)
++cpi->common.fc.eob_branch_counts[tx_size][type][ref][band][pt];
}
- token_cache[scan[c]] = token;
+ token_cache[scan[c]] = vp9_pt_energy_class[token];
++t;
} while (c < eob && ++c < seg_eob);
