Use ref_type as skip_ref_frame_mask's index
Use ref_type instead of mix single ref and
compound ref can skip more unwanted combinations.
For encoder, about 1.5% faster shows by encoding
20 frame of BasketballDrill_832x480_50.y4m at
800kbps on speed 1. (234709 ms -> 231096 ms)
The coding performance is about 0.02% loss on
average.
STATS_CHANGED expected
Change-Id: I0aa0c624994098cc89d46fc698d96f4980bd95f4
diff --git a/av1/encoder/encodeframe.c b/av1/encoder/encodeframe.c
index e9a93ce..05b45f3 100644
--- a/av1/encoder/encodeframe.c
+++ b/av1/encoder/encodeframe.c
@@ -1731,31 +1731,20 @@
pc_tree->partitioning = partition;
}
}
-
- pc_tree->horizontal[0].skip_ref_frame_mask = 0;
- pc_tree->horizontal[1].skip_ref_frame_mask = 0;
- pc_tree->vertical[0].skip_ref_frame_mask = 0;
- pc_tree->vertical[1].skip_ref_frame_mask = 0;
- pc_tree->horizontala[0].skip_ref_frame_mask = 0;
- pc_tree->horizontala[1].skip_ref_frame_mask = 0;
- pc_tree->horizontala[2].skip_ref_frame_mask = 0;
- pc_tree->horizontalb[0].skip_ref_frame_mask = 0;
- pc_tree->horizontalb[1].skip_ref_frame_mask = 0;
- pc_tree->horizontalb[2].skip_ref_frame_mask = 0;
- pc_tree->verticala[0].skip_ref_frame_mask = 0;
- pc_tree->verticala[1].skip_ref_frame_mask = 0;
- pc_tree->verticala[2].skip_ref_frame_mask = 0;
- pc_tree->verticalb[0].skip_ref_frame_mask = 0;
- pc_tree->verticalb[1].skip_ref_frame_mask = 0;
- pc_tree->verticalb[2].skip_ref_frame_mask = 0;
- pc_tree->horizontal4[0].skip_ref_frame_mask = 0;
- pc_tree->horizontal4[1].skip_ref_frame_mask = 0;
- pc_tree->horizontal4[2].skip_ref_frame_mask = 0;
- pc_tree->horizontal4[3].skip_ref_frame_mask = 0;
- pc_tree->vertical4[0].skip_ref_frame_mask = 0;
- pc_tree->vertical4[1].skip_ref_frame_mask = 0;
- pc_tree->vertical4[2].skip_ref_frame_mask = 0;
- pc_tree->vertical4[3].skip_ref_frame_mask = 0;
+ for (int b = 0; b < 2; ++b) {
+ pc_tree->horizontal[b].skip_ref_frame_mask = 0;
+ pc_tree->vertical[b].skip_ref_frame_mask = 0;
+ }
+ for (int b = 0; b < 3; ++b) {
+ pc_tree->horizontala[b].skip_ref_frame_mask = 0;
+ pc_tree->horizontalb[b].skip_ref_frame_mask = 0;
+ pc_tree->verticala[b].skip_ref_frame_mask = 0;
+ pc_tree->verticalb[b].skip_ref_frame_mask = 0;
+ }
+ for (int b = 0; b < 4; ++b) {
+ pc_tree->horizontal4[b].skip_ref_frame_mask = 0;
+ pc_tree->vertical4[b].skip_ref_frame_mask = 0;
+ }
switch (partition) {
case PARTITION_NONE:
rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
@@ -3321,7 +3310,7 @@
// Ref frames picked in the [i_th] quarter subblock during square partition
// RD search. It may be used to prune ref frame selection of rect partitions.
- uint8_t ref_frames_used[4] = {
+ int ref_frames_used[4] = {
0,
};
@@ -3352,11 +3341,9 @@
if (none_rd) *none_rd = this_rdc.rdcost;
if (this_rdc.rate != INT_MAX) {
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- const int ref1 = ctx_none->mic.ref_frame[0];
- const int ref2 = ctx_none->mic.ref_frame[1];
+ const int ref_type = av1_ref_frame_type(ctx_none->mic.ref_frame);
for (int i = 0; i < 4; ++i) {
- ref_frames_used[i] |= (1 << ref1);
- if (ref2 > 0) ref_frames_used[i] |= (1 << ref2);
+ ref_frames_used[i] |= (1 << ref_type);
}
}
if (bsize_at_least_8x8) {
@@ -3491,10 +3478,9 @@
sum_rdc.rdcost += this_rdc.rdcost;
if (cpi->sf.prune_ref_frame_for_rect_partitions &&
pc_tree->split[idx]->none.rate != INT_MAX) {
- const int ref1 = pc_tree->split[idx]->none.mic.ref_frame[0];
- const int ref2 = pc_tree->split[idx]->none.mic.ref_frame[1];
- ref_frames_used[idx] |= (1 << ref1);
- if (ref2 > 0) ref_frames_used[idx] |= (1 << ref2);
+ const int ref_type =
+ av1_ref_frame_type(pc_tree->split[idx]->none.mic.ref_frame);
+ ref_frames_used[idx] |= (1 << ref_type);
}
if (idx <= 1 && (bsize <= BLOCK_8X8 ||
pc_tree->split[idx]->partitioning == PARTITION_NONE)) {
@@ -3532,7 +3518,7 @@
pc_tree->vertical[0].skip_ref_frame_mask = 0;
pc_tree->vertical[1].skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- uint8_t used_frames;
+ int used_frames;
used_frames = ref_frames_used[0] | ref_frames_used[1];
if (used_frames) pc_tree->horizontal[0].skip_ref_frame_mask = ~used_frames;
used_frames = ref_frames_used[2] | ref_frames_used[3];
@@ -3789,7 +3775,7 @@
pc_tree->horizontala[1].skip_ref_frame_mask = 0;
pc_tree->horizontala[2].skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- uint8_t used_frames;
+ int used_frames;
used_frames = ref_frames_used[0];
if (used_frames)
pc_tree->horizontala[0].skip_ref_frame_mask = ~used_frames;
@@ -3822,7 +3808,7 @@
pc_tree->horizontalb[1].skip_ref_frame_mask = 0;
pc_tree->horizontalb[2].skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- uint8_t used_frames;
+ int used_frames;
used_frames = ref_frames_used[0] | ref_frames_used[1];
if (used_frames)
pc_tree->horizontalb[0].skip_ref_frame_mask = ~used_frames;
@@ -3856,7 +3842,7 @@
pc_tree->verticala[1].skip_ref_frame_mask = 0;
pc_tree->verticala[2].skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- uint8_t used_frames;
+ int used_frames;
used_frames = ref_frames_used[0];
if (used_frames) pc_tree->verticala[0].skip_ref_frame_mask = ~used_frames;
used_frames = ref_frames_used[2];
@@ -3886,7 +3872,7 @@
pc_tree->verticalb[1].skip_ref_frame_mask = 0;
pc_tree->verticalb[2].skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- uint8_t used_frames;
+ int used_frames;
used_frames = ref_frames_used[0] | ref_frames_used[2];
if (used_frames) pc_tree->verticalb[0].skip_ref_frame_mask = ~used_frames;
used_frames = ref_frames_used[1];
@@ -3956,9 +3942,9 @@
ctx_this->rd_mode_is_ready = 0;
ctx_this->skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- const uint8_t used_frames =
- i <= 1 ? (ref_frames_used[0] | ref_frames_used[1])
- : (ref_frames_used[2] | ref_frames_used[3]);
+ const int used_frames = i <= 1
+ ? (ref_frames_used[0] | ref_frames_used[1])
+ : (ref_frames_used[2] | ref_frames_used[3]);
if (used_frames) ctx_this->skip_ref_frame_mask = ~used_frames;
}
if (!rd_try_subblock(cpi, td, tile_data, tp, (i == 0), (i == 3),
@@ -3999,9 +3985,9 @@
ctx_this->rd_mode_is_ready = 0;
ctx_this->skip_ref_frame_mask = 0;
if (cpi->sf.prune_ref_frame_for_rect_partitions) {
- const uint8_t used_frames =
- i <= 1 ? (ref_frames_used[0] | ref_frames_used[2])
- : (ref_frames_used[1] | ref_frames_used[3]);
+ const int used_frames = i <= 1
+ ? (ref_frames_used[0] | ref_frames_used[2])
+ : (ref_frames_used[1] | ref_frames_used[3]);
if (used_frames) ctx_this->skip_ref_frame_mask = ~used_frames;
}
if (!rd_try_subblock(cpi, td, tile_data, tp, (i == 0), (i == 3), mi_row,
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index d82833b..7353f4a 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -7719,6 +7719,7 @@
int ref_frame_cost;
int single_comp_cost;
int64_t (*simple_rd)[MAX_REF_MV_SERCH][REF_FRAMES];
+ int skip_motion_mode;
} HandleInterModeArgs;
static INLINE int clamp_and_check_mv(int_mv *out_mv, int_mv in_mv,
@@ -8409,6 +8410,7 @@
RD_STATS best_rd_stats, best_rd_stats_y, best_rd_stats_uv;
MB_MODE_INFO base_mbmi, best_mbmi;
uint8_t best_blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE];
+
int interintra_allowed = cm->seq_params.enable_interintra_compound &&
is_interintra_allowed(mbmi) && mbmi->compound_idx;
int pts0[SAMPLES_ARRAY_SIZE], pts_inref0[SAMPLES_ARRAY_SIZE];
@@ -8419,11 +8421,11 @@
av1_invalid_rd_stats(&best_rd_stats);
aom_clear_system_state();
mbmi->num_proj_ref = 1; // assume num_proj_ref >=1
- MOTION_MODE last_motion_mode_allowed =
- cm->switchable_motion_mode
- ? motion_mode_allowed(xd->global_motion, xd, mbmi,
- cm->allow_warped_motion)
- : SIMPLE_TRANSLATION;
+ MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
+ if (cm->switchable_motion_mode) {
+ last_motion_mode_allowed = motion_mode_allowed(xd->global_motion, xd, mbmi,
+ cm->allow_warped_motion);
+ }
if (last_motion_mode_allowed == WARPED_CAUSAL) {
mbmi->num_proj_ref = findSamples(cm, xd, mi_row, mi_col, pts0, pts_inref0);
}
@@ -8439,6 +8441,7 @@
for (int mode_index = (int)SIMPLE_TRANSLATION;
mode_index <= (int)last_motion_mode_allowed + interintra_allowed;
mode_index++) {
+ if (args->skip_motion_mode && mode_index) continue;
int64_t tmp_rd = INT64_MAX;
int tmp_rate2 = rate2_nocoeff;
int is_interintra_mode = mode_index > (int)last_motion_mode_allowed;
@@ -10113,9 +10116,20 @@
mbmi_ext->ref_mv_count[ref_frame] = UINT8_MAX;
if (cpi->ref_frame_flags & ref_frame_flag_list[ref_frame]) {
if (mbmi->partition != PARTITION_NONE &&
- mbmi->partition != PARTITION_SPLIT &&
- (skip_ref_frame_mask & (1 << ref_frame))) {
- continue;
+ mbmi->partition != PARTITION_SPLIT) {
+ if (skip_ref_frame_mask & (1 << ref_frame)) {
+ int skip = 1;
+ for (int r = ALTREF_FRAME + 1; r < MODE_CTX_REF_FRAMES; ++r) {
+ if (!(skip_ref_frame_mask & (1 << r))) {
+ const MV_REFERENCE_FRAME *rf = ref_frame_map[r - REF_FRAMES];
+ if (rf[0] == ref_frame || rf[1] == ref_frame) {
+ skip = 0;
+ break;
+ }
+ }
+ }
+ if (skip) continue;
+ }
}
assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
setup_buffer_ref_mvs_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
@@ -10131,10 +10145,10 @@
(cpi->ref_frame_flags & ref_frame_flag_list[rf[1]]))) {
continue;
}
+
if (mbmi->partition != PARTITION_NONE &&
mbmi->partition != PARTITION_SPLIT) {
- if ((skip_ref_frame_mask & (1 << rf[0])) ||
- (skip_ref_frame_mask & (1 << rf[1]))) {
+ if (skip_ref_frame_mask & (1 << ref_frame)) {
continue;
}
}
@@ -10448,6 +10462,9 @@
av1_zero(search_state->single_state_modelled_cnt);
}
+// Case 1: return 0, means don't skip this mode
+// Case 2: return 1, means skip this mode completely
+// Case 3: return 2, means skip compound only, but still try single motion modes
static int inter_mode_search_order_independent_skip(
const AV1_COMP *cpi, const PICK_MODE_CONTEXT *ctx, const MACROBLOCK *x,
BLOCK_SIZE bsize, int mode_index, int mi_row, int mi_col,
@@ -10460,10 +10477,31 @@
const unsigned char segment_id = mbmi->segment_id;
const MV_REFERENCE_FRAME *ref_frame = av1_mode_order[mode_index].ref_frame;
const PREDICTION_MODE this_mode = av1_mode_order[mode_index].mode;
+ int skip_motion_mode = 0;
if (mbmi->partition != PARTITION_NONE && mbmi->partition != PARTITION_SPLIT) {
- if (ctx->skip_ref_frame_mask & (1 << ref_frame[0])) return 1;
- if (ref_frame[1] > 0 && (ctx->skip_ref_frame_mask & (1 << ref_frame[1])))
- return 1;
+ const int ref_type = av1_ref_frame_type(ref_frame);
+ int skip_ref = ctx->skip_ref_frame_mask & (1 << ref_type);
+ if (ref_type <= ALTREF_FRAME && skip_ref) {
+ // Since the compound ref modes depends on the motion estimation result of
+ // two single ref modes( best mv of single ref modes as the start point )
+ // If current single ref mode is marked skip, we need to check if it will
+ // be used in compound ref modes.
+ for (int r = ALTREF_FRAME + 1; r < MODE_CTX_REF_FRAMES; ++r) {
+ if (!(ctx->skip_ref_frame_mask & (1 << r))) {
+ const MV_REFERENCE_FRAME *rf = ref_frame_map[r - REF_FRAMES];
+ if (rf[0] == ref_type || rf[1] == ref_type) {
+ // Found a not skipped compound ref mode which contains current
+ // single ref. So this single ref can't be skipped completly
+ // Just skip it's motion mode search, still try it's simple
+ // transition mode.
+ skip_motion_mode = 1;
+ skip_ref = 0;
+ break;
+ }
+ }
+ }
+ }
+ if (skip_ref) return 1;
}
if (cpi->sf.mode_pruning_based_on_two_pass_partition_search &&
@@ -10580,6 +10618,9 @@
if (skip_repeated_mv(cm, x, this_mode, ref_frame)) {
return 1;
}
+ if (skip_motion_mode) {
+ return 2;
+ }
return 0;
}
@@ -11116,7 +11157,8 @@
NULL, NULL,
NULL, NULL,
{ { 0 } }, INT_MAX,
- INT_MAX, NULL
+ INT_MAX, NULL,
+ 0
};
for (i = 0; i < REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
@@ -11170,12 +11212,11 @@
analyze_single_states(cpi, &search_state);
reach_first_comp_mode = 1;
}
-
- if (inter_mode_search_order_independent_skip(cpi, ctx, x, bsize, mode_index,
- mi_row, mi_col, mode_skip_mask,
- ref_frame_skip_mask))
- continue;
-
+ const int ret = inter_mode_search_order_independent_skip(
+ cpi, ctx, x, bsize, mode_index, mi_row, mi_col, mode_skip_mask,
+ ref_frame_skip_mask);
+ if (ret == 1) continue;
+ args.skip_motion_mode = (ret == 2);
if (sf->prune_comp_search_by_single_result > 0 &&
second_ref_frame > INTRA_FRAME) {
if (compound_skip_by_single_states(cpi, &search_state, this_mode,
