ref: 8fa9281b7820baa255f712330fa2763e3112c4d9
parent: b8d86733e9d9c58e17028720751f96dad2df7a09
parent: 93c95bf3e17acf0452bc726d8c84e25c38fefc78
author: Angie Chiang <angiebird@google.com>
date: Fri Sep 20 18:52:11 EDT 2019
Merge "Remove USE_PQSORT and CHANGE_MV_SEARCH_ORDER"
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -2431,7 +2431,6 @@
cpi->kmeans_data_arr_alloc = 0;
#if CONFIG_NON_GREEDY_MV
- cpi->feature_score_loc_alloc = 0;
cpi->tpl_ready = 0;
#endif // CONFIG_NON_GREEDY_MV
for (i = 0; i < MAX_ARF_GOP_SIZE; ++i) cpi->tpl_stats[i].tpl_stats_ptr = NULL;
@@ -6699,42 +6698,6 @@
}
}
-static double get_feature_score(uint8_t *buf, ptrdiff_t stride, int rows,
- int cols) {- double IxIx = 0;
- double IxIy = 0;
- double IyIy = 0;
- double score;
- int r, c;
- vpx_clear_system_state();
- for (r = 0; r + 1 < rows; ++r) {- for (c = 0; c + 1 < cols; ++c) {- int diff_x = buf[r * stride + c] - buf[r * stride + c + 1];
- int diff_y = buf[r * stride + c] - buf[(r + 1) * stride + c];
- IxIx += diff_x * diff_x;
- IxIy += diff_x * diff_y;
- IyIy += diff_y * diff_y;
- }
- }
- IxIx /= (rows - 1) * (cols - 1);
- IxIy /= (rows - 1) * (cols - 1);
- IyIy /= (rows - 1) * (cols - 1);
- score = (IxIx * IyIy - IxIy * IxIy + 0.0001) / (IxIx + IyIy + 0.0001);
- return score;
-}
-
-static int compare_feature_score(const void *a, const void *b) {- const FEATURE_SCORE_LOC *aa = *(FEATURE_SCORE_LOC *const *)a;
- const FEATURE_SCORE_LOC *bb = *(FEATURE_SCORE_LOC *const *)b;
- if (aa->feature_score < bb->feature_score) {- return 1;
- } else if (aa->feature_score > bb->feature_score) {- return -1;
- } else {- return 0;
- }
-}
-
static void do_motion_search(VP9_COMP *cpi, ThreadData *td,
MotionField *motion_field, int frame_idx,
YV12_BUFFER_CONFIG *ref_frame, BLOCK_SIZE bsize,
@@ -6760,87 +6723,8 @@
}
}
-#define CHANGE_MV_SEARCH_ORDER 1
-#define USE_PQSORT 1
-
-#if CHANGE_MV_SEARCH_ORDER
-#if USE_PQSORT
-static void max_heap_pop(FEATURE_SCORE_LOC **heap, int *size,
- FEATURE_SCORE_LOC **output) {- if (*size > 0) {- *output = heap[0];
- --*size;
- if (*size > 0) {- int p, l, r;
- heap[0] = heap[*size];
- p = 0;
- l = 2 * p + 1;
- r = 2 * p + 2;
- while (l < *size) {- FEATURE_SCORE_LOC *tmp;
- int c = l;
- if (r < *size && heap[r]->feature_score > heap[l]->feature_score) {- c = r;
- }
- if (heap[p]->feature_score >= heap[c]->feature_score) {- break;
- }
- tmp = heap[p];
- heap[p] = heap[c];
- heap[c] = tmp;
- p = c;
- l = 2 * p + 1;
- r = 2 * p + 2;
- }
- }
- } else {- assert(0);
- }
-}
-
-static void max_heap_push(FEATURE_SCORE_LOC **heap, int *size,
- FEATURE_SCORE_LOC *input) {- int c, p;
- FEATURE_SCORE_LOC *tmp;
- input->visited = 1;
- heap[*size] = input;
- ++*size;
- c = *size - 1;
- p = c >> 1;
- while (c > 0 && heap[c]->feature_score > heap[p]->feature_score) {- tmp = heap[p];
- heap[p] = heap[c];
- heap[c] = tmp;
- c = p;
- p >>= 1;
- }
-}
-
-static void add_nb_blocks_to_heap(VP9_COMP *cpi, const TplDepFrame *tpl_frame,
- BLOCK_SIZE bsize, int mi_row, int mi_col,
- int *heap_size) {- const int mi_unit = num_8x8_blocks_wide_lookup[bsize];
- const int dirs[NB_MVS_NUM][2] = { { -1, 0 }, { 0, -1 }, { 1, 0 }, { 0, 1 } };- int i;
- for (i = 0; i < NB_MVS_NUM; ++i) {- int r = dirs[i][0] * mi_unit;
- int c = dirs[i][1] * mi_unit;
- if (mi_row + r >= 0 && mi_row + r < tpl_frame->mi_rows && mi_col + c >= 0 &&
- mi_col + c < tpl_frame->mi_cols) {- FEATURE_SCORE_LOC *fs_loc =
- &cpi->feature_score_loc_arr[(mi_row + r) * tpl_frame->stride +
- (mi_col + c)];
- if (fs_loc->visited == 0) {- max_heap_push(cpi->feature_score_loc_heap, heap_size, fs_loc);
- }
- }
- }
-}
-#endif // USE_PQSORT
-#endif // CHANGE_MV_SEARCH_ORDER
-
static void build_motion_field(
- VP9_COMP *cpi, MACROBLOCKD *xd, int frame_idx,
+ VP9_COMP *cpi, int frame_idx,
YV12_BUFFER_CONFIG *ref_frame[MAX_INTER_REF_FRAMES], BLOCK_SIZE bsize) {VP9_COMMON *cm = &cpi->common;
ThreadData *td = &cpi->td;
@@ -6849,8 +6733,6 @@
const int mi_width = num_8x8_blocks_wide_lookup[bsize];
const int pw = num_4x4_blocks_wide_lookup[bsize] << 2;
const int ph = num_4x4_blocks_high_lookup[bsize] << 2;
- int fs_loc_sort_size;
- int fs_loc_heap_size;
int mi_row, mi_col;
int rf_idx;
@@ -6857,31 +6739,7 @@
tpl_frame->lambda = (pw * ph) >> 2;
assert(pw * ph == tpl_frame->lambda << 2);
- fs_loc_sort_size = 0;
- for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) {- for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {- const int mb_y_offset =
- mi_row * MI_SIZE * xd->cur_buf->y_stride + mi_col * MI_SIZE;
- const int bw = 4 << b_width_log2_lookup[bsize];
- const int bh = 4 << b_height_log2_lookup[bsize];
- FEATURE_SCORE_LOC *fs_loc =
- &cpi->feature_score_loc_arr[mi_row * tpl_frame->stride + mi_col];
- fs_loc->feature_score = get_feature_score(
- xd->cur_buf->y_buffer + mb_y_offset, xd->cur_buf->y_stride, bw, bh);
- fs_loc->visited = 0;
- fs_loc->mi_row = mi_row;
- fs_loc->mi_col = mi_col;
- cpi->feature_score_loc_sort[fs_loc_sort_size] = fs_loc;
- ++fs_loc_sort_size;
- }
- }
-
- qsort(cpi->feature_score_loc_sort, fs_loc_sort_size,
- sizeof(*cpi->feature_score_loc_sort), compare_feature_score);
-
- // TODO(angiebird): Clean up this part.
for (rf_idx = 0; rf_idx < MAX_INTER_REF_FRAMES; ++rf_idx) {- int i;
MotionField *motion_field = vp9_motion_field_info_get_motion_field(
&cpi->motion_field_info, frame_idx, rf_idx, bsize);
if (ref_frame[rf_idx] == NULL) {@@ -6888,34 +6746,6 @@
continue;
}
vp9_motion_field_reset_mvs(motion_field);
-#if CHANGE_MV_SEARCH_ORDER
-#if !USE_PQSORT
- for (i = 0; i < fs_loc_sort_size; ++i) {- FEATURE_SCORE_LOC *fs_loc = cpi->feature_score_loc_sort[i];
- do_motion_search(cpi, td, motion_field, frame_idx, ref_frame[rf_idx],
- bsize, fs_loc->mi_row, fs_loc->mi_col);
- }
-#else // !USE_PQSORT
- fs_loc_heap_size = 0;
- max_heap_push(cpi->feature_score_loc_heap, &fs_loc_heap_size,
- cpi->feature_score_loc_sort[0]);
-
- for (i = 0; i < fs_loc_sort_size; ++i) {- cpi->feature_score_loc_sort[i]->visited = 0;
- }
-
- while (fs_loc_heap_size > 0) {- FEATURE_SCORE_LOC *fs_loc;
- max_heap_pop(cpi->feature_score_loc_heap, &fs_loc_heap_size, &fs_loc);
-
- do_motion_search(cpi, td, motion_field, frame_idx, ref_frame[rf_idx],
- bsize, fs_loc->mi_row, fs_loc->mi_col);
-
- add_nb_blocks_to_heap(cpi, tpl_frame, bsize, fs_loc->mi_row,
- fs_loc->mi_col, &fs_loc_heap_size);
- }
-#endif // !USE_PQSORT
-#else // CHANGE_MV_SEARCH_ORDER
for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) { for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {do_motion_search(cpi, td, motion_field, frame_idx, ref_frame[rf_idx],
@@ -6922,7 +6752,6 @@
bsize, mi_row, mi_col);
}
}
-#endif // CHANGE_MV_SEARCH_ORDER
}
}
#endif // CONFIG_NON_GREEDY_MV
@@ -7004,7 +6833,7 @@
for (square_block_idx = 0; square_block_idx < SQUARE_BLOCK_SIZES;
++square_block_idx) {BLOCK_SIZE square_bsize = square_block_idx_to_bsize(square_block_idx);
- build_motion_field(cpi, xd, frame_idx, ref_frame, square_bsize);
+ build_motion_field(cpi, frame_idx, ref_frame, square_bsize);
}
for (rf_idx = 0; rf_idx < MAX_INTER_REF_FRAMES; ++rf_idx) {int ref_frame_idx = gf_picture[frame_idx].ref_frame[rf_idx];
@@ -7131,21 +6960,6 @@
#if CONFIG_NON_GREEDY_MV
int rf_idx;
- if (cpi->feature_score_loc_alloc == 0) {- // The smallest block size of motion field is 4x4, but the mi_unit is 8x8,
- // therefore the number of units is "mi_rows * mi_cols * 4" here.
- CHECK_MEM_ERROR(
- cm, cpi->feature_score_loc_arr,
- vpx_calloc(mi_rows * mi_cols * 4, sizeof(*cpi->feature_score_loc_arr)));
- CHECK_MEM_ERROR(cm, cpi->feature_score_loc_sort,
- vpx_calloc(mi_rows * mi_cols * 4,
- sizeof(*cpi->feature_score_loc_sort)));
- CHECK_MEM_ERROR(cm, cpi->feature_score_loc_heap,
- vpx_calloc(mi_rows * mi_cols * 4,
- sizeof(*cpi->feature_score_loc_heap)));
-
- cpi->feature_score_loc_alloc = 1;
- }
vpx_free(cpi->select_mv_arr);
CHECK_MEM_ERROR(
cm, cpi->select_mv_arr,
@@ -7195,9 +7009,6 @@
int frame;
#if CONFIG_NON_GREEDY_MV
vp9_free_motion_field_info(&cpi->motion_field_info);
- vpx_free(cpi->feature_score_loc_arr);
- vpx_free(cpi->feature_score_loc_sort);
- vpx_free(cpi->feature_score_loc_heap);
vpx_free(cpi->select_mv_arr);
#endif
for (frame = 0; frame < MAX_ARF_GOP_SIZE; ++frame) {--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -505,15 +505,6 @@
// Maximum operating frame buffer size needed for a GOP using ARF reference.
#define MAX_ARF_GOP_SIZE (2 * MAX_LAG_BUFFERS)
-#if CONFIG_NON_GREEDY_MV
-typedef struct FEATURE_SCORE_LOC {- int visited;
- double feature_score;
- int mi_row;
- int mi_col;
-} FEATURE_SCORE_LOC;
-#endif
-
#define MAX_KMEANS_GROUPS 8
typedef struct KMEANS_DATA {@@ -563,10 +554,6 @@
#if CONFIG_NON_GREEDY_MV
MotionFieldInfo motion_field_info;
int tpl_ready;
- int feature_score_loc_alloc;
- FEATURE_SCORE_LOC *feature_score_loc_arr;
- FEATURE_SCORE_LOC **feature_score_loc_sort;
- FEATURE_SCORE_LOC **feature_score_loc_heap;
int_mv *select_mv_arr;
#endif