ref: 1a4d8f20336195f560e29a2a96def8b3fdedfb54
parent: 66a96fd3de6426f8a7ec5293a858d97009ae00c4
parent: 3d6b0cb825d18fba31649c4ada500b9ed15b0d4a
author: Jerome Jiang <jianj@google.com>
date: Tue Jul 11 15:44:22 EDT 2017
Merge "vp9: Move skinmap computation into multithreading loop."
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -4129,6 +4129,10 @@
(*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
sb_col_in_tile);
+ if (cpi->use_skin_detection) {
+ vp9_compute_skin_sb(cpi, BLOCK_16X16, mi_row, mi_col);
+ }
+
x->source_variance = UINT_MAX;
vp9_zero(x->pred_mv);
vp9_rd_cost_init(&dummy_rdc);
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -3475,7 +3475,6 @@
cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
cpi->use_skin_detection = 1;
- vp9_compute_skin_map(cpi, BLOCK_16X16);
}
vp9_set_quantizer(cm, q);
--- a/vp9/encoder/vp9_noise_estimate.c
+++ b/vp9/encoder/vp9_noise_estimate.c
@@ -172,6 +172,7 @@
int mi_row, mi_col;
int num_low_motion = 0;
int frame_low_motion = 1;
+ if (cpi->use_skin_detection) vp9_compute_skin_map(cpi, BLOCK_16X16);
for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
int bl_index = mi_row * cm->mi_cols + mi_col;
--- a/vp9/encoder/vp9_skin_detection.c
+++ b/vp9/encoder/vp9_skin_detection.c
@@ -37,8 +37,9 @@
}
}
-void vp9_compute_skin_map(VP9_COMP *const cpi, BLOCK_SIZE bsize) {
- int mi_row, mi_col, num_bl;
+void vp9_compute_skin_sb(VP9_COMP *const cpi, BLOCK_SIZE bsize, int mi_row,
+ int mi_col) {
+ int i, j, num_bl;
VP9_COMMON *const cm = &cpi->common;
const uint8_t *src_y = cpi->Source->y_buffer;
const uint8_t *src_u = cpi->Source->u_buffer;
@@ -50,13 +51,17 @@
const int shy = (y_bsize == 8) ? 3 : 4;
const int shuv = shy - 1;
const int fac = y_bsize / 8;
- // Loop through blocks and set skin map based on center pixel of block.
- // Ignore rightmost/bottom boundary blocks.
- for (mi_row = 0; mi_row < cm->mi_rows - 1; mi_row += fac) {
+ const int y_shift = src_ystride * (mi_row << 3) + (mi_col << 3);
+ const int uv_shift = src_uvstride * (mi_row << 2) + (mi_col << 2);
+ src_y += y_shift;
+ src_u += uv_shift;
+ src_v += uv_shift;
+
+ for (i = mi_row; i < VPXMIN(mi_row + 7, cm->mi_rows - 1); i += fac) {
num_bl = 0;
- for (mi_col = 0; mi_col < cm->mi_cols - 1; mi_col += fac) {
+ for (j = mi_col; j < VPXMIN(mi_col + 7, cm->mi_cols - 1); j += fac) {
int consec_zeromv = 0;
- int bl_index = mi_row * cm->mi_cols + mi_col;
+ int bl_index = i * cm->mi_cols + j;
int bl_index1 = bl_index + 1;
int bl_index2 = bl_index + cm->mi_cols;
int bl_index3 = bl_index2 + 1;
@@ -67,7 +72,7 @@
VPXMIN(cpi->consec_zero_mv[bl_index1],
VPXMIN(cpi->consec_zero_mv[bl_index2],
cpi->consec_zero_mv[bl_index3])));
- cpi->skin_map[mi_row * cm->mi_cols + mi_col] =
+ cpi->skin_map[bl_index] =
vp9_compute_skin_block(src_y, src_u, src_v, src_ystride, src_uvstride,
bsize, consec_zeromv, 0);
num_bl++;
@@ -78,6 +83,18 @@
src_y += (src_ystride << shy) - (num_bl << shy);
src_u += (src_uvstride << shuv) - (num_bl << shuv);
src_v += (src_uvstride << shuv) - (num_bl << shuv);
+ }
+}
+
+void vp9_compute_skin_map(VP9_COMP *const cpi, BLOCK_SIZE bsize) {
+ int mi_row, mi_col;
+ VP9_COMMON *const cm = &cpi->common;
+ // Loop through blocks and set skin map based on center pixel of block.
+ // Ignore rightmost/bottom boundary blocks.
+ for (mi_row = 0; mi_row < cm->mi_rows - 1; mi_row += MI_BLOCK_SIZE) {
+ for (mi_col = 0; mi_col < cm->mi_cols - 1; mi_col += MI_BLOCK_SIZE) {
+ vp9_compute_skin_sb(cpi, bsize, mi_row, mi_col);
+ }
}
}
--- a/vp9/encoder/vp9_skin_detection.h
+++ b/vp9/encoder/vp9_skin_detection.h
@@ -25,6 +25,9 @@
int stride, int strideuv, int bsize,
int consec_zeromv, int curr_motion_magn);
+void vp9_compute_skin_sb(struct VP9_COMP *const cpi, BLOCK_SIZE bsize,
+ int mi_row, int mi_col);
+
void vp9_compute_skin_map(struct VP9_COMP *const cpi, BLOCK_SIZE bsize);
#ifdef OUTPUT_YUV_SKINMAP