Prune uv intra mode with hog
Performance on hdres over 70 frames, vbr mode
SPD_SET | AVG_PSNR | OVR_PSNR | SSIM | SPD
6 | +0.187% | +0.185% | +0.229% | +3.0%
5 | +0.123% | +0.125% | +0.195% | +2.0%
3 | +0.088% | +0.094% | +0.088% | +2.2%
STATS_CHANGED
Change-Id: I4fd8a3b643e67dbf826d7b248af5f6a409fd2081
diff --git a/av1/encoder/intra_mode_search.c b/av1/encoder/intra_mode_search.c
index 0f9ecfe..b9f311e 100644
--- a/av1/encoder/intra_mode_search.c
+++ b/av1/encoder/intra_mode_search.c
@@ -9,11 +9,13 @@
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
+#include "av1/common/av1_common_int.h"
#include "av1/common/reconintra.h"
#include "av1/encoder/intra_mode_search.h"
#include "av1/encoder/intra_mode_search_utils.h"
#include "av1/encoder/palette.h"
+#include "av1/encoder/speed_features.h"
#include "av1/encoder/tx_search.h"
/*!\cond */
@@ -472,6 +474,8 @@
cpi->optimize_seg_arr[mbmi->segment_id]);
xd->cfl.store_y = 0;
}
+ IntraModeSearchState intra_search_state;
+ init_intra_mode_search_state(&intra_search_state);
// Search through all non-palette modes.
for (int mode_idx = 0; mode_idx < UV_INTRA_MODES; ++mode_idx) {
@@ -503,6 +507,26 @@
if (is_directional_mode && av1_use_angle_delta(mbmi->bsize) &&
intra_mode_cfg->enable_angle_delta) {
+ const SPEED_FEATURES *sf = &cpi->sf;
+ if (sf->intra_sf.chroma_intra_pruning_with_hog &&
+ !intra_search_state.dir_mode_skip_mask_ready) {
+ static const float thresh[2][4] = {
+ { -1.2f, 0.0f, 0.0f, 1.2f }, // Interframe
+ { -1.2f, -1.2f, -0.6f, 0.4f }, // Intraframe
+ };
+ const int is_chroma = 1;
+ const int is_intra_frame = frame_is_intra_only(cm);
+ prune_intra_mode_with_hog(
+ x, bsize,
+ thresh[is_intra_frame]
+ [sf->intra_sf.chroma_intra_pruning_with_hog - 1],
+ intra_search_state.directional_mode_skip_mask, is_chroma);
+ intra_search_state.dir_mode_skip_mask_ready = 1;
+ }
+ if (intra_search_state.directional_mode_skip_mask[mode]) {
+ continue;
+ }
+
// Search through angle delta
const int rate_overhead =
mode_costs->intra_uv_mode_cost[is_cfl_allowed(xd)][mbmi->mode][mode];
@@ -884,9 +908,10 @@
if (sf->intra_sf.intra_pruning_with_hog &&
!intra_search_state->dir_mode_skip_mask_ready) {
const float thresh[4] = { -1.2f, 0.0f, 0.0f, 1.2f };
- prune_intra_mode_with_hog(x, bsize,
- thresh[sf->intra_sf.intra_pruning_with_hog - 1],
- intra_search_state->directional_mode_skip_mask);
+ const int is_chroma = 0;
+ prune_intra_mode_with_hog(
+ x, bsize, thresh[sf->intra_sf.intra_pruning_with_hog - 1],
+ intra_search_state->directional_mode_skip_mask, is_chroma);
intra_search_state->dir_mode_skip_mask_ready = 1;
}
if (intra_search_state->directional_mode_skip_mask[mode]) return 0;
@@ -1030,9 +1055,10 @@
// Less aggressive thresholds are used here than those used in inter frame
// encoding.
const float thresh[4] = { -1.2f, -1.2f, -0.6f, 0.4f };
+ const int is_chroma = 0;
prune_intra_mode_with_hog(
x, bsize, thresh[cpi->sf.intra_sf.intra_pruning_with_hog - 1],
- directional_mode_skip_mask);
+ directional_mode_skip_mask, is_chroma);
}
mbmi->filter_intra_mode_info.use_filter_intra = 0;
pmi->palette_size[0] = 0;
diff --git a/av1/encoder/intra_mode_search.h b/av1/encoder/intra_mode_search.h
index 86a546e..cc2a87b0 100644
--- a/av1/encoder/intra_mode_search.h
+++ b/av1/encoder/intra_mode_search.h
@@ -254,6 +254,14 @@
int *val_count_8bit, int *num_color_bins,
int *num_colors);
+/*! \brief Initializes the \ref IntraModeSearchState struct.
+ */
+static AOM_INLINE void init_intra_mode_search_state(
+ IntraModeSearchState *intra_search_state) {
+ memset(intra_search_state, 0, sizeof(*intra_search_state));
+ intra_search_state->rate_uv_intra = INT_MAX;
+}
+
#ifdef __cplusplus
} // extern "C"
#endif
diff --git a/av1/encoder/intra_mode_search_utils.h b/av1/encoder/intra_mode_search_utils.h
index f901130..8ec7cb3 100644
--- a/av1/encoder/intra_mode_search_utils.h
+++ b/av1/encoder/intra_mode_search_utils.h
@@ -191,18 +191,24 @@
static AOM_INLINE void prune_intra_mode_with_hog(
const MACROBLOCK *x, BLOCK_SIZE bsize, float th,
- uint8_t *directional_mode_skip_mask) {
+ uint8_t *directional_mode_skip_mask, int is_chroma) {
aom_clear_system_state();
+ const MACROBLOCKD *xd = &x->e_mbd;
+ const int plane = is_chroma ? AOM_PLANE_U : AOM_PLANE_Y;
+ const struct macroblockd_plane *const pd = &xd->plane[plane];
+ const int ss_x = pd->subsampling_x;
+ const int ss_y = pd->subsampling_y;
const int bh = block_size_high[bsize];
const int bw = block_size_wide[bsize];
- const MACROBLOCKD *xd = &x->e_mbd;
const int rows =
- (xd->mb_to_bottom_edge >= 0) ? bh : (xd->mb_to_bottom_edge >> 3) + bh;
+ ((xd->mb_to_bottom_edge >= 0) ? bh : (xd->mb_to_bottom_edge >> 3) + bh) >>
+ ss_y;
const int cols =
- (xd->mb_to_right_edge >= 0) ? bw : (xd->mb_to_right_edge >> 3) + bw;
- const int src_stride = x->plane[0].src.stride;
- const uint8_t *src = x->plane[0].src.buf;
+ ((xd->mb_to_right_edge >= 0) ? bw : (xd->mb_to_right_edge >> 3) + bw) >>
+ ss_x;
+ const int src_stride = x->plane[plane].src.stride;
+ const uint8_t *src = x->plane[plane].src.buf;
float hist[BINS] = { 0.0f };
if (is_cur_buf_hbd(xd)) {
generate_hog_hbd(src, src_stride, rows, cols, hist);
@@ -210,6 +216,10 @@
generate_hog(src, src_stride, rows, cols, hist);
}
+ for (int b = 0; b < BINS; ++b) {
+ hist[b] *= (1 + ss_x) * (1 + ss_y);
+ }
+
for (int i = 0; i < DIRECTIONAL_MODES; ++i) {
float this_score = intra_hog_model_bias[i];
const float *weights = &intra_hog_model_weights[i * BINS];
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index d3c0a77..c6b2d16 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -3876,16 +3876,6 @@
x->comp_rd_stats_idx = 0;
}
-static AOM_INLINE void init_intra_mode_search_state(
- IntraModeSearchState *intra_search_state) {
- intra_search_state->skip_intra_modes = 0;
- intra_search_state->best_intra_mode = DC_PRED;
- intra_search_state->dir_mode_skip_mask_ready = 0;
- av1_zero(intra_search_state->directional_mode_skip_mask);
- intra_search_state->rate_uv_intra = INT_MAX;
- av1_zero(intra_search_state->pmi_uv);
-}
-
static AOM_INLINE void init_inter_mode_search_state(
InterModeSearchState *search_state, const AV1_COMP *cpi,
const MACROBLOCK *x, BLOCK_SIZE bsize, int64_t best_rd_so_far) {
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index fb27aca..ee1f890 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -460,6 +460,7 @@
sf->inter_sf.skip_repeated_newmv = 1;
sf->interp_sf.use_interp_filter = 1;
+
sf->intra_sf.prune_palette_search_level = 1;
sf->tx_sf.adaptive_txb_search_level = 2;
@@ -557,6 +558,10 @@
sf->inter_sf.txfm_rd_gate_level =
boosted ? 0 : (is_boosted_arf2_bwd_type ? 1 : 2);
+ // TODO(chiyotsai@google.com): the thresholds chosen for intra hog are
+ // inherited directly from luma hog with some minor tweaking. Eventually we
+ // should run this with a bayesian optimizer to find the Pareto frontier.
+ sf->intra_sf.chroma_intra_pruning_with_hog = 2;
sf->intra_sf.intra_pruning_with_hog = 3;
sf->intra_sf.prune_palette_search_level = 2;
@@ -666,6 +671,8 @@
sf->inter_sf.prune_inter_modes_if_skippable = 1;
sf->inter_sf.txfm_rd_gate_level = boosted ? 0 : 5;
+ sf->intra_sf.chroma_intra_pruning_with_hog = 3;
+
// TODO(any): Extend multi-winner mode processing support for inter frames
sf->winner_mode_sf.multi_winner_mode_type =
frame_is_intra_only(&cpi->common) ? MULTI_WINNER_MODE_FAST
@@ -690,6 +697,7 @@
sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 3;
sf->inter_sf.prune_nearmv_using_neighbors = 1;
+ sf->intra_sf.chroma_intra_pruning_with_hog = 4;
sf->intra_sf.intra_pruning_with_hog = 4;
sf->part_sf.prune_rectangular_split_based_on_qidx =
@@ -1182,6 +1190,7 @@
}
static AOM_INLINE void init_intra_sf(INTRA_MODE_SPEED_FEATURES *intra_sf) {
+ intra_sf->chroma_intra_pruning_with_hog = 0;
intra_sf->skip_intra_in_interframe = 1;
intra_sf->intra_pruning_with_hog = 0;
intra_sf->src_var_thresh_intra_skip = 1;
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index 7aae39e..9f2a99b 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -806,8 +806,13 @@
unsigned int src_var_thresh_intra_skip;
// Prune intra mode candidates based on source block histogram of gradient.
+ // Applies to luma plane only.
int intra_pruning_with_hog;
+ // Prune intra mode candidates based on source block histogram of gradient.
+ // Applies to chroma plane only.
+ int chroma_intra_pruning_with_hog;
+
// Enable/disable smooth intra modes.
int disable_smooth_intra;
