Merge "Add vp10_inv_txfm2d" into nextgenv2
diff --git a/configure b/configure
index 315c427..27d618b 100755
--- a/configure
+++ b/configure
@@ -267,6 +267,7 @@
fp_mb_stats
emulate_hardware
var_tx
+ ref_mv
ext_tx
misc_fixes
ext_intra
diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 5d59e83..d44a64a 100644
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -191,9 +191,14 @@
INTRA_PRED_TEST(SSE, TestIntraPred4, vpx_dc_predictor_4x4_sse,
vpx_dc_left_predictor_4x4_sse, vpx_dc_top_predictor_4x4_sse,
vpx_dc_128_predictor_4x4_sse, vpx_v_predictor_4x4_sse, NULL,
- NULL, NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_4x4_sse)
+ NULL, NULL, NULL, NULL, NULL, NULL, NULL)
#endif // HAVE_SSE && CONFIG_USE_X86INC
+#if HAVE_SSE2 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSE2, TestIntraPred4, NULL, NULL, NULL, NULL, NULL, NULL,
+ NULL, NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_4x4_sse2)
+#endif // HAVE_SSE2 && CONFIG_USE_X86INC
+
#if HAVE_SSSE3 && CONFIG_USE_X86INC
INTRA_PRED_TEST(SSSE3, TestIntraPred4, NULL, NULL, NULL, NULL, NULL,
vpx_h_predictor_4x4_ssse3, vpx_d45_predictor_4x4_ssse3, NULL,
diff --git a/vp10/common/blockd.h b/vp10/common/blockd.h
index 3b94cc5..bb94397 100644
--- a/vp10/common/blockd.h
+++ b/vp10/common/blockd.h
@@ -274,6 +274,8 @@
};
#if CONFIG_EXT_TX
+#define ALLOW_INTRA_EXT_TX 1
+
static const int num_ext_tx_set_inter[EXT_TX_SETS_INTER] = {
1, 17, 10, 2
};
@@ -368,7 +370,8 @@
return DCT_DCT;
#if CONFIG_EXT_TX
- if (mbmi->sb_type >= BLOCK_8X8 && plane_type == PLANE_TYPE_Y)
+ if (mbmi->sb_type >= BLOCK_8X8 && plane_type == PLANE_TYPE_Y &&
+ ALLOW_INTRA_EXT_TX)
return mbmi->tx_type;
#endif // CONFIG_EXT_TX
@@ -405,8 +408,10 @@
#endif
return DCT_DCT;
if (mbmi->sb_type >= BLOCK_8X8) {
- if (plane_type == PLANE_TYPE_Y)
+ if (plane_type == PLANE_TYPE_Y) {
+ if (is_inter_block(mbmi) || ALLOW_INTRA_EXT_TX)
return mbmi->tx_type;
+ }
if (is_inter_block(mbmi))
// UV Inter only
return (mbmi->tx_type == IDTX && tx_size == TX_32X32 ?
diff --git a/vp10/decoder/decodemv.c b/vp10/decoder/decodemv.c
index 34d2c73..d66ddd9 100644
--- a/vp10/decoder/decodemv.c
+++ b/vp10/decoder/decodemv.c
@@ -422,7 +422,8 @@
#if CONFIG_EXT_TX
if (get_ext_tx_types(mbmi->tx_size, mbmi->sb_type, 0) > 1 &&
cm->base_qindex > 0 && !mbmi->skip &&
- !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+ !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP) &&
+ ALLOW_INTRA_EXT_TX) {
FRAME_COUNTS *counts = xd->counts;
int eset = get_ext_tx_set(mbmi->tx_size, mbmi->sb_type, 0);
if (eset > 0) {
@@ -887,7 +888,7 @@
if (counts)
++counts->inter_ext_tx[eset][mbmi->tx_size][mbmi->tx_type];
}
- } else {
+ } else if (ALLOW_INTRA_EXT_TX) {
if (eset > 0) {
mbmi->tx_type = vpx_read_tree(r, vp10_ext_tx_intra_tree[eset],
cm->fc->intra_ext_tx_prob[eset]
diff --git a/vp10/encoder/bitstream.c b/vp10/encoder/bitstream.c
index 59edec2..a80e5bd 100644
--- a/vp10/encoder/bitstream.c
+++ b/vp10/encoder/bitstream.c
@@ -693,7 +693,7 @@
vp10_write_token(w, vp10_ext_tx_inter_tree[eset],
cm->fc->inter_ext_tx_prob[eset][mbmi->tx_size],
&ext_tx_inter_encodings[eset][mbmi->tx_type]);
- } else {
+ } else if (ALLOW_INTRA_EXT_TX) {
if (eset > 0)
vp10_write_token(
w, vp10_ext_tx_intra_tree[eset],
@@ -790,7 +790,8 @@
#if CONFIG_EXT_TX
if (get_ext_tx_types(mbmi->tx_size, bsize, 0) > 1 &&
cm->base_qindex > 0 && !mbmi->skip &&
- !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+ !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP) &&
+ ALLOW_INTRA_EXT_TX) {
int eset = get_ext_tx_set(mbmi->tx_size, bsize, 0);
if (eset > 0)
vp10_write_token(
diff --git a/vp10/encoder/rdopt.c b/vp10/encoder/rdopt.c
index e0912b5..7562b60 100644
--- a/vp10/encoder/rdopt.c
+++ b/vp10/encoder/rdopt.c
@@ -691,6 +691,10 @@
if (!ext_tx_used_inter[ext_tx_set][tx_type])
continue;
} else {
+ if (!ALLOW_INTRA_EXT_TX && bs >= BLOCK_8X8) {
+ if (tx_type != intra_mode_to_tx_type_lookup[mbmi->mode])
+ continue;
+ }
if (!ext_tx_used_intra[ext_tx_set][tx_type])
continue;
}
@@ -719,7 +723,7 @@
r += cpi->inter_tx_type_costs[ext_tx_set]
[mbmi->tx_size][mbmi->tx_type];
} else {
- if (ext_tx_set > 0)
+ if (ext_tx_set > 0 && ALLOW_INTRA_EXT_TX)
r += cpi->intra_tx_type_costs[ext_tx_set][mbmi->tx_size]
[mbmi->mode][mbmi->tx_type];
}
@@ -754,12 +758,16 @@
if (get_ext_tx_types(mbmi->tx_size, bs, is_inter) > 1 &&
!xd->lossless[mbmi->segment_id] && *rate != INT_MAX) {
int ext_tx_set = get_ext_tx_set(mbmi->tx_size, bs, is_inter);
- if (is_inter)
- *rate += cpi->inter_tx_type_costs[ext_tx_set][mbmi->tx_size]
- [mbmi->tx_type];
- else
- *rate += cpi->intra_tx_type_costs[ext_tx_set][mbmi->tx_size]
- [mbmi->mode][mbmi->tx_type];
+ if (is_inter) {
+ if (ext_tx_set > 0)
+ *rate += cpi->inter_tx_type_costs[ext_tx_set][mbmi->tx_size]
+ [mbmi->tx_type];
+ } else {
+ if (ext_tx_set > 0 && ALLOW_INTRA_EXT_TX)
+ *rate +=
+ cpi->intra_tx_type_costs[ext_tx_set][mbmi->tx_size]
+ [mbmi->mode][mbmi->tx_type];
+ }
}
#endif // CONFIG_EXT_TX
}
@@ -849,6 +857,10 @@
if (!ext_tx_used_inter[ext_tx_set][tx_type])
continue;
} else {
+ if (!ALLOW_INTRA_EXT_TX && bs >= BLOCK_8X8) {
+ if (tx_type != intra_mode_to_tx_type_lookup[mbmi->mode])
+ continue;
+ }
if (!ext_tx_used_intra[ext_tx_set][tx_type])
continue;
}
@@ -874,7 +886,7 @@
r += cpi->inter_tx_type_costs[ext_tx_set]
[mbmi->tx_size][mbmi->tx_type];
} else {
- if (ext_tx_set > 0)
+ if (ext_tx_set > 0 && ALLOW_INTRA_EXT_TX)
r += cpi->intra_tx_type_costs[ext_tx_set][mbmi->tx_size]
[mbmi->mode][mbmi->tx_type];
}
@@ -2227,6 +2239,10 @@
if (!ext_tx_used_inter[ext_tx_set][tx_type])
continue;
} else {
+ if (!ALLOW_INTRA_EXT_TX && bsize >= BLOCK_8X8) {
+ if (tx_type != intra_mode_to_tx_type_lookup[mbmi->mode])
+ continue;
+ }
if (!ext_tx_used_intra[ext_tx_set][tx_type])
continue;
}
@@ -2251,7 +2267,7 @@
this_rate += cpi->inter_tx_type_costs[ext_tx_set]
[max_tx_size][mbmi->tx_type];
} else {
- if (ext_tx_set > 0)
+ if (ext_tx_set > 0 && ALLOW_INTRA_EXT_TX)
this_rate += cpi->intra_tx_type_costs[ext_tx_set][max_tx_size]
[mbmi->mode][mbmi->tx_type];
}
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index f41ee09..890b638 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -314,9 +314,6 @@
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 avx/;
-add_proto qw/int vp9_full_range_search/, "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_full_range_search/;
-
add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
specialize qw/vp9_temporal_filter_apply sse2 msa/;
diff --git a/vp9/encoder/vp9_block.h b/vp9/encoder/vp9_block.h
index fc34786..ca5d746 100644
--- a/vp9/encoder/vp9_block.h
+++ b/vp9/encoder/vp9_block.h
@@ -71,6 +71,8 @@
int rddiv;
int rdmult;
int mb_energy;
+ int * m_search_count_ptr;
+ int * ex_search_count_ptr;
// These are set to their default values at the beginning, and then adjusted
// further in the encoding process.
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 0475883..f9c28f6 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -3839,6 +3839,10 @@
TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
int mi_row;
+ // Set up pointers to per thread motion search counters.
+ td->mb.m_search_count_ptr = &td->rd_counts.m_search_count;
+ td->mb.ex_search_count_ptr = &td->rd_counts.ex_search_count;
+
for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
mi_row += MI_BLOCK_SIZE) {
if (cpi->sf.use_nonrd_pick_mode)
@@ -3895,6 +3899,9 @@
vp9_zero(rdc->coef_counts);
vp9_zero(rdc->comp_pred_diff);
vp9_zero(rdc->filter_diff);
+ rdc->m_search_count = 0; // Count of motion search hits.
+ rdc->ex_search_count = 0; // Exhaustive mesh search hits.
+
xd->lossless = cm->base_qindex == 0 &&
cm->y_dc_delta_q == 0 &&
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index eebd7c5..a57cf87 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -2995,7 +2995,7 @@
recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
if (cpi->twopass.total_left_stats.coded_error != 0.0)
- fprintf(f, "%10u %dx%d %d %d %10d %10d %10d %10d"
+ fprintf(f, "%10u %dx%d %10d %10d %d %d %10d %10d %10d %10d"
"%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
"%10"PRId64" %10"PRId64" %10d "
"%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
@@ -3004,6 +3004,8 @@
"%10lf %8u %10"PRId64" %10d %10d %10d\n",
cpi->common.current_video_frame,
cm->width, cm->height,
+ cpi->td.rd_counts.m_search_count,
+ cpi->td.rd_counts.ex_search_count,
cpi->rc.source_alt_ref_pending,
cpi->rc.source_alt_ref_active,
cpi->rc.this_frame_target,
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index 975d9f4..f6d8931 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -260,6 +260,8 @@
vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
int64_t comp_pred_diff[REFERENCE_MODES];
int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+ int m_search_count;
+ int ex_search_count;
} RD_COUNTS;
typedef struct ThreadData {
diff --git a/vp9/encoder/vp9_ethread.c b/vp9/encoder/vp9_ethread.c
index ad25712..1d1926c 100644
--- a/vp9/encoder/vp9_ethread.c
+++ b/vp9/encoder/vp9_ethread.c
@@ -30,6 +30,10 @@
for (n = 0; n < ENTROPY_TOKENS; n++)
td->rd_counts.coef_counts[i][j][k][l][m][n] +=
td_t->rd_counts.coef_counts[i][j][k][l][m][n];
+
+ // Counts of all motion searches and exhuastive mesh searches.
+ td->rd_counts.m_search_count += td_t->rd_counts.m_search_count;
+ td->rd_counts.ex_search_count += td_t->rd_counts.ex_search_count;
}
static int enc_worker_hook(EncWorkerData *const thread_data, void *unused) {
diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c
index b9a104a..327ac19 100644
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -1517,69 +1517,83 @@
#undef CHECK_BETTER
-int vp9_full_range_search_c(const MACROBLOCK *x,
- const search_site_config *cfg,
- MV *ref_mv, MV *best_mv,
- int search_param, int sad_per_bit, int *num00,
- const vp9_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv) {
+// Exhuastive motion search around a given centre position with a given
+// step size.
+static int exhuastive_mesh_search(const MACROBLOCK *x,
+ MV *ref_mv, MV *best_mv,
+ int range, int step, int sad_per_bit,
+ const vp9_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv) {
const MACROBLOCKD *const xd = &x->e_mbd;
const struct buf_2d *const what = &x->plane[0].src;
const struct buf_2d *const in_what = &xd->plane[0].pre[0];
- const int range = 64;
- const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+ MV fcenter_mv = {center_mv->row, center_mv->col};
unsigned int best_sad = INT_MAX;
int r, c, i;
int start_col, end_col, start_row, end_row;
+ int col_step = (step > 1) ? step : 4;
- // The cfg and search_param parameters are not used in this search variant
- (void)cfg;
- (void)search_param;
+ assert(step >= 1);
- clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
- *best_mv = *ref_mv;
- *num00 = 11;
+ clamp_mv(&fcenter_mv, x->mv_col_min, x->mv_col_max,
+ x->mv_row_min, x->mv_row_max);
+ *best_mv = fcenter_mv;
best_sad = fn_ptr->sdf(what->buf, what->stride,
- get_buf_from_mv(in_what, ref_mv), in_what->stride) +
- mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
- start_row = VPXMAX(-range, x->mv_row_min - ref_mv->row);
- start_col = VPXMAX(-range, x->mv_col_min - ref_mv->col);
- end_row = VPXMIN(range, x->mv_row_max - ref_mv->row);
- end_col = VPXMIN(range, x->mv_col_max - ref_mv->col);
+ get_buf_from_mv(in_what, &fcenter_mv), in_what->stride) +
+ mvsad_err_cost(x, &fcenter_mv, ref_mv, sad_per_bit);
+ start_row = VPXMAX(-range, x->mv_row_min - fcenter_mv.row);
+ start_col = VPXMAX(-range, x->mv_col_min - fcenter_mv.col);
+ end_row = VPXMIN(range, x->mv_row_max - fcenter_mv.row);
+ end_col = VPXMIN(range, x->mv_col_max - fcenter_mv.col);
- for (r = start_row; r <= end_row; ++r) {
- for (c = start_col; c <= end_col; c += 4) {
- if (c + 3 <= end_col) {
- unsigned int sads[4];
- const uint8_t *addrs[4];
- for (i = 0; i < 4; ++i) {
- const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
- addrs[i] = get_buf_from_mv(in_what, &mv);
- }
-
- fn_ptr->sdx4df(what->buf, what->stride, addrs, in_what->stride, sads);
-
- for (i = 0; i < 4; ++i) {
- if (sads[i] < best_sad) {
- const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
- const unsigned int sad = sads[i] +
- mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
- if (sad < best_sad) {
- best_sad = sad;
- *best_mv = mv;
- }
+ for (r = start_row; r <= end_row; r += step) {
+ for (c = start_col; c <= end_col; c += col_step) {
+ // Step > 1 means we are not checking every location in this pass.
+ if (step > 1) {
+ const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c};
+ unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &mv), in_what->stride);
+ if (sad < best_sad) {
+ sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+ if (sad < best_sad) {
+ best_sad = sad;
+ *best_mv = mv;
}
}
} else {
- for (i = 0; i < end_col - c; ++i) {
- const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
- unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
- get_buf_from_mv(in_what, &mv), in_what->stride);
- if (sad < best_sad) {
- sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+ // 4 sads in a single call if we are checking every location
+ if (c + 3 <= end_col) {
+ unsigned int sads[4];
+ const uint8_t *addrs[4];
+ for (i = 0; i < 4; ++i) {
+ const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
+ addrs[i] = get_buf_from_mv(in_what, &mv);
+ }
+ fn_ptr->sdx4df(what->buf, what->stride, addrs,
+ in_what->stride, sads);
+
+ for (i = 0; i < 4; ++i) {
+ if (sads[i] < best_sad) {
+ const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
+ const unsigned int sad = sads[i] +
+ mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+ if (sad < best_sad) {
+ best_sad = sad;
+ *best_mv = mv;
+ }
+ }
+ }
+ } else {
+ for (i = 0; i < end_col - c; ++i) {
+ const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
+ unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &mv), in_what->stride);
if (sad < best_sad) {
- best_sad = sad;
- *best_mv = mv;
+ sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+ if (sad < best_sad) {
+ best_sad = sad;
+ *best_mv = mv;
+ }
}
}
}
@@ -2011,6 +2025,70 @@
return bestsme;
}
+#define MIN_RANGE 7
+#define MAX_RANGE 256
+#define MIN_INTERVAL 1
+// Runs an limited range exhaustive mesh search using a pattern set
+// according to the encode speed profile.
+static int full_pixel_exhaustive(VP9_COMP *cpi, MACROBLOCK *x,
+ MV *centre_mv_full, int sadpb, int *cost_list,
+ const vp9_variance_fn_ptr_t *fn_ptr,
+ const MV *ref_mv, MV *dst_mv) {
+ const SPEED_FEATURES *const sf = &cpi->sf;
+ MV temp_mv = {centre_mv_full->row, centre_mv_full->col};
+ MV f_ref_mv = {ref_mv->row >> 3, ref_mv->col >> 3};
+ int bestsme;
+ int i;
+ int interval = sf->mesh_patterns[0].interval;
+ int range = sf->mesh_patterns[0].range;
+ int baseline_interval_divisor;
+
+ // Keep track of number of exhaustive calls (this frame in this thread).
+ ++(*x->ex_search_count_ptr);
+
+ // Trap illegal values for interval and range for this function.
+ if ((range < MIN_RANGE) || (range > MAX_RANGE) ||
+ (interval < MIN_INTERVAL) || (interval > range))
+ return INT_MAX;
+
+ baseline_interval_divisor = range / interval;
+
+ // Check size of proposed first range against magnitude of the centre
+ // value used as a starting point.
+ range = VPXMAX(range, (5 * VPXMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
+ range = VPXMIN(range, MAX_RANGE);
+ interval = VPXMAX(interval, range / baseline_interval_divisor);
+
+ // initial search
+ bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv, range,
+ interval, sadpb, fn_ptr, &temp_mv);
+
+ if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
+ // Progressive searches with range and step size decreasing each time
+ // till we reach a step size of 1. Then break out.
+ for (i = 1; i < MAX_MESH_STEP; ++i) {
+ // First pass with coarser step and longer range
+ bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv,
+ sf->mesh_patterns[i].range,
+ sf->mesh_patterns[i].interval,
+ sadpb, fn_ptr, &temp_mv);
+
+ if (sf->mesh_patterns[i].interval == 1)
+ break;
+ }
+ }
+
+ if (bestsme < INT_MAX)
+ bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+ *dst_mv = temp_mv;
+
+ // Return cost list.
+ if (cost_list) {
+ calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
+ }
+ return bestsme;
+}
+
int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
int sad_per_bit, int distance,
const vp9_variance_fn_ptr_t *fn_ptr,
@@ -2324,6 +2402,18 @@
return best_sad;
}
+#define MIN_EX_SEARCH_LIMIT 128
+static int is_exhaustive_allowed(VP9_COMP *cpi, MACROBLOCK *x) {
+ const SPEED_FEATURES *const sf = &cpi->sf;
+ const int max_ex = VPXMAX(MIN_EX_SEARCH_LIMIT,
+ (*x->m_search_count_ptr * sf->max_exaustive_pct) / 100);
+
+ return sf->allow_exhaustive_searches &&
+ (sf->exhaustive_searches_thresh < INT_MAX) &&
+ (*x->ex_search_count_ptr <= max_ex) &&
+ !cpi->rc.is_src_frame_alt_ref;
+}
+
int vp9_full_pixel_search(VP9_COMP *cpi, MACROBLOCK *x,
BLOCK_SIZE bsize, MV *mvp_full,
int step_param, int error_per_bit,
@@ -2342,6 +2432,9 @@
cost_list[4] = INT_MAX;
}
+ // Keep track of number of searches (this frame in this thread).
+ ++(*x->m_search_count_ptr);
+
switch (method) {
case FAST_DIAMOND:
var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
@@ -2367,6 +2460,27 @@
var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
MAX_MVSEARCH_STEPS - 1 - step_param,
1, cost_list, fn_ptr, ref_mv, tmp_mv);
+
+ // Should we allow a follow on exhaustive search?
+ if (is_exhaustive_allowed(cpi, x)) {
+ int64_t exhuastive_thr = sf->exhaustive_searches_thresh;
+ exhuastive_thr >>= 8 - (b_width_log2_lookup[bsize] +
+ b_height_log2_lookup[bsize]);
+
+ // Threshold variance for an exhaustive full search.
+ if (var > exhuastive_thr) {
+ int var_ex;
+ MV tmp_mv_ex;
+ var_ex = full_pixel_exhaustive(cpi, x, tmp_mv,
+ error_per_bit, cost_list, fn_ptr,
+ ref_mv, &tmp_mv_ex);
+
+ if (var_ex < var) {
+ var = var_ex;
+ *tmp_mv = tmp_mv_ex;
+ }
+ }
+ }
break;
default:
assert(0 && "Invalid search method.");
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index d78137a..e32130e 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -1750,8 +1750,9 @@
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
ENTROPY_CONTEXT t_above[2], t_left[2];
int subpelmv = 1, have_ref = 0;
+ SPEED_FEATURES *const sf = &cpi->sf;
const int has_second_rf = has_second_ref(mbmi);
- const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
+ const int inter_mode_mask = sf->inter_mode_mask[bsize];
MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
vp9_zero(*bsi);
@@ -1820,7 +1821,7 @@
seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
int step_param = 0;
- int thissme, bestsme = INT_MAX;
+ int bestsme = INT_MAX;
int sadpb = x->sadperbit4;
MV mvp_full;
int max_mv;
@@ -1845,7 +1846,7 @@
max_mv =
VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
- if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
+ if (sf->mv.auto_mv_step_size && cm->show_frame) {
// Take wtd average of the step_params based on the last frame's
// max mv magnitude and the best ref mvs of the current block for
// the given reference.
@@ -1858,7 +1859,7 @@
mvp_full.row = bsi->mvp.as_mv.row >> 3;
mvp_full.col = bsi->mvp.as_mv.col >> 3;
- if (cpi->sf.adaptive_motion_search) {
+ if (sf->adaptive_motion_search) {
mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
step_param = VPXMAX(step_param, 8);
@@ -1871,31 +1872,10 @@
bestsme = vp9_full_pixel_search(
cpi, x, bsize, &mvp_full, step_param, sadpb,
- cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
+ sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
&bsi->ref_mv[0]->as_mv, new_mv,
INT_MAX, 1);
- // Should we do a full search (best quality only)
- if (cpi->oxcf.mode == BEST) {
- int_mv *const best_mv = &mi->bmi[i].as_mv[0];
- /* Check if mvp_full is within the range. */
- clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
- x->mv_row_min, x->mv_row_max);
- thissme = cpi->full_search_sad(x, &mvp_full,
- sadpb, 16, &cpi->fn_ptr[bsize],
- &bsi->ref_mv[0]->as_mv,
- &best_mv->as_mv);
- cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
- if (thissme < bestsme) {
- bestsme = thissme;
- *new_mv = best_mv->as_mv;
- } else {
- // The full search result is actually worse so re-instate the
- // previous best vector
- best_mv->as_mv = *new_mv;
- }
- }
-
if (bestsme < INT_MAX) {
int distortion;
cpi->find_fractional_mv_step(
@@ -1904,8 +1884,8 @@
&bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
- cpi->sf.mv.subpel_force_stop,
- cpi->sf.mv.subpel_iters_per_step,
+ sf->mv.subpel_force_stop,
+ sf->mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&distortion,
@@ -1916,7 +1896,7 @@
seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
}
- if (cpi->sf.adaptive_motion_search)
+ if (sf->adaptive_motion_search)
x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
// restore src pointers
@@ -1933,7 +1913,7 @@
mbmi->interp_filter == EIGHTTAP) {
// adjust src pointers
mi_buf_shift(x, i);
- if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
+ if (sf->comp_inter_joint_search_thresh <= bsize) {
int rate_mv;
joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
mi_row, mi_col, seg_mvs[i],
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index a539629..318d810 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -15,6 +15,22 @@
#include "vp9/encoder/vp9_rdopt.h"
#include "vpx_dsp/vpx_dsp_common.h"
+// Mesh search patters for various speed settings
+static MESH_PATTERN best_quality_mesh_pattern[MAX_MESH_STEP] =
+ {{64, 4}, {28, 2}, {15, 1}, {7, 1}};
+
+#define MAX_MESH_SPEED 5 // Max speed setting for mesh motion method
+static MESH_PATTERN good_quality_mesh_patterns[MAX_MESH_SPEED + 1]
+ [MAX_MESH_STEP] =
+ {{{64, 8}, {28, 4}, {15, 1}, {7, 1}},
+ {{64, 8}, {28, 4}, {15, 1}, {7, 1}},
+ {{64, 8}, {14, 2}, {7, 1}, {7, 1}},
+ {{64, 16}, {24, 8}, {12, 4}, {7, 1}},
+ {{64, 16}, {24, 8}, {12, 4}, {7, 1}},
+ {{64, 16}, {24, 8}, {12, 4}, {7, 1}},
+ };
+static unsigned char good_quality_max_mesh_pct[MAX_MESH_SPEED + 1] =
+ {50, 25, 15, 5, 1, 1};
// Intra only frames, golden frames (except alt ref overlays) and
// alt ref frames tend to be coded at a higher than ambient quality
@@ -259,6 +275,8 @@
sf->static_segmentation = 0;
sf->adaptive_rd_thresh = 1;
sf->use_fast_coef_costing = 1;
+ sf->allow_exhaustive_searches = 0;
+ sf->exhaustive_searches_thresh = INT_MAX;
if (speed >= 1) {
sf->use_square_partition_only = !frame_is_intra_only(cm);
@@ -460,7 +478,6 @@
sf->mv.auto_mv_step_size = 0;
sf->mv.fullpel_search_step_param = 6;
sf->comp_inter_joint_search_thresh = BLOCK_4X4;
- sf->adaptive_rd_thresh = 0;
sf->tx_size_search_method = USE_FULL_RD;
sf->use_lp32x32fdct = 0;
sf->adaptive_motion_search = 0;
@@ -516,19 +533,50 @@
// Recode loop tolerance %.
sf->recode_tolerance = 25;
sf->default_interp_filter = SWITCHABLE;
- sf->tx_size_search_breakout = 0;
- sf->partition_search_breakout_dist_thr = 0;
- sf->partition_search_breakout_rate_thr = 0;
sf->simple_model_rd_from_var = 0;
+ // Some speed-up features even for best quality as minimal impact on quality.
+ sf->adaptive_rd_thresh = 1;
+ sf->tx_size_search_breakout = 1;
+ sf->partition_search_breakout_dist_thr = (1 << 19);
+ sf->partition_search_breakout_rate_thr = 80;
+
if (oxcf->mode == REALTIME)
set_rt_speed_feature(cpi, sf, oxcf->speed, oxcf->content);
else if (oxcf->mode == GOOD)
set_good_speed_feature(cpi, cm, sf, oxcf->speed);
cpi->full_search_sad = vp9_full_search_sad;
- cpi->diamond_search_sad = oxcf->mode == BEST ? vp9_full_range_search
- : vp9_diamond_search_sad;
+ cpi->diamond_search_sad = vp9_diamond_search_sad;
+
+ sf->allow_exhaustive_searches = 1;
+ if (oxcf->mode == BEST) {
+ if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
+ sf->exhaustive_searches_thresh = (1 << 20);
+ else
+ sf->exhaustive_searches_thresh = (1 << 21);
+ sf->max_exaustive_pct = 100;
+ for (i = 0; i < MAX_MESH_STEP; ++i) {
+ sf->mesh_patterns[i].range = best_quality_mesh_pattern[i].range;
+ sf->mesh_patterns[i].interval = best_quality_mesh_pattern[i].interval;
+ }
+ } else {
+ int speed = (oxcf->speed > MAX_MESH_SPEED) ? MAX_MESH_SPEED : oxcf->speed;
+ if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
+ sf->exhaustive_searches_thresh = (1 << 22);
+ else
+ sf->exhaustive_searches_thresh = (1 << 23);
+ sf->max_exaustive_pct = good_quality_max_mesh_pct[speed];
+ if (speed > 0)
+ sf->exhaustive_searches_thresh = sf->exhaustive_searches_thresh << 1;
+
+ for (i = 0; i < MAX_MESH_STEP; ++i) {
+ sf->mesh_patterns[i].range =
+ good_quality_mesh_patterns[speed][i].range;
+ sf->mesh_patterns[i].interval =
+ good_quality_mesh_patterns[speed][i].interval;
+ }
+ }
// Slow quant, dct and trellis not worthwhile for first pass
// so make sure they are always turned off.
diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h
index 575e98c..ad7b64a 100644
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -195,6 +195,13 @@
int fullpel_search_step_param;
} MV_SPEED_FEATURES;
+#define MAX_MESH_STEP 4
+
+typedef struct MESH_PATTERN {
+ int range;
+ int interval;
+} MESH_PATTERN;
+
typedef struct SPEED_FEATURES {
MV_SPEED_FEATURES mv;
@@ -299,6 +306,18 @@
// point for this motion search and limits the search range around it.
int adaptive_motion_search;
+ // Flag for allowing some use of exhaustive searches;
+ int allow_exhaustive_searches;
+
+ // Threshold for allowing exhaistive motion search.
+ int exhaustive_searches_thresh;
+
+ // Maximum number of exhaustive searches for a frame.
+ int max_exaustive_pct;
+
+ // Pattern to be used for any exhaustive mesh searches.
+ MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
+
int schedule_mode_search;
// Allows sub 8x8 modes to use the prediction filter that was determined
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index 1dfc45c..13da155 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -278,7 +278,8 @@
cpi->alt_ref_source = lc->alt_ref_source;
// Reset the frames_since_key and frames_to_key counters to their values
// before the layer restore. Keep these defined for the stream (not layer).
- if (cpi->svc.number_temporal_layers > 1) {
+ if (cpi->svc.number_temporal_layers > 1 ||
+ cpi->svc.number_spatial_layers > 1) {
cpi->rc.frames_since_key = old_frame_since_key;
cpi->rc.frames_to_key = old_frame_to_key;
}
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index b369b05..03dcd4e 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -97,7 +97,7 @@
specialize qw/vpx_ve_predictor_4x4/;
add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse_x86inc";
+specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse_x86inc";
diff --git a/vpx_dsp/x86/intrapred_sse2.asm b/vpx_dsp/x86/intrapred_sse2.asm
index 22b5731..04b39a5 100644
--- a/vpx_dsp/x86/intrapred_sse2.asm
+++ b/vpx_dsp/x86/intrapred_sse2.asm
@@ -515,35 +515,34 @@
jnz .loop
REP_RET
-INIT_MMX sse
-cglobal tm_predictor_4x4, 4, 4, 4, dst, stride, above, left
+INIT_XMM sse2
+cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left
pxor m1, m1
- movd m2, [aboveq-1]
- movd m0, [aboveq]
- punpcklbw m2, m1
+ movq m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x
punpcklbw m0, m1
- pshufw m2, m2, 0x0
- DEFINE_ARGS dst, stride, line, left
- mov lineq, -2
- add leftq, 4
- psubw m0, m2
-.loop:
- movd m2, [leftq+lineq*2]
- movd m3, [leftq+lineq*2+1]
+ pshuflw m2, m0, 0x0 ; [63:0] tl tl tl tl [word]
+ psrldq m0, 2
+ psubw m0, m2 ; [63:0] t1-tl t2-tl t3-tl t4-tl [word]
+ movd m2, [leftq]
punpcklbw m2, m1
- punpcklbw m3, m1
- pshufw m2, m2, 0x0
- pshufw m3, m3, 0x0
- paddw m2, m0
+ pshuflw m4, m2, 0x0 ; [63:0] l1 l1 l1 l1 [word]
+ pshuflw m3, m2, 0x55 ; [63:0] l2 l2 l2 l2 [word]
+ paddw m4, m0
paddw m3, m0
- packuswb m2, m2
+ packuswb m4, m4
packuswb m3, m3
- movd [dstq ], m2
+ movd [dstq ], m4
movd [dstq+strideq], m3
lea dstq, [dstq+strideq*2]
- inc lineq
- jnz .loop
- REP_RET
+ pshuflw m4, m2, 0xaa
+ pshuflw m3, m2, 0xff
+ paddw m4, m0
+ paddw m3, m0
+ packuswb m4, m4
+ packuswb m3, m3
+ movd [dstq ], m4
+ movd [dstq+strideq], m3
+ RET
INIT_XMM sse2
cglobal tm_predictor_8x8, 4, 4, 4, dst, stride, above, left
