Commit a220b931 authored by Jerome Jiang's avatar Jerome Jiang

vp9: compute skinmap only once before encoding.

This could save some cycles since skin detection is used in multiple
places in vp9.

1~2% speed up on ARM.

Change-Id: I86b731945f85215bbb0976021cd0f2040ff2687c
parent e187b274
......@@ -743,16 +743,7 @@ static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
for (i = 0; i < ymis; i += 2) {
for (j = 0; j < xmis; j += 2) {
int bl_index = block_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;
int consec_zeromv =
VPXMIN(cpi->consec_zero_mv[bl_index],
VPXMIN(cpi->consec_zero_mv[bl_index1],
VPXMIN(cpi->consec_zero_mv[bl_index2],
cpi->consec_zero_mv[bl_index3])));
int is_skin = vp9_compute_skin_block(
ysignal, usignal, vsignal, sp, spuv, BLOCK_16X16, consec_zeromv, 0);
int is_skin = cpi->skin_map[bl_index];
num_16x16_skin += is_skin;
num_16x16_nonskin += (1 - is_skin);
if (num_16x16_nonskin > 3) {
......
......@@ -779,6 +779,9 @@ static void dealloc_compressor_data(VP9_COMP *cpi) {
cpi->nmvsadcosts_hp[0] = NULL;
cpi->nmvsadcosts_hp[1] = NULL;
vpx_free(cpi->skin_map);
cpi->skin_map = NULL;
vpx_free(cpi->prev_partition);
cpi->prev_partition = NULL;
......@@ -2046,6 +2049,9 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
realloc_segmentation_maps(cpi);
CHECK_MEM_ERROR(cm, cpi->skin_map, vpx_calloc(cm->mi_rows * cm->mi_cols,
sizeof(cpi->skin_map[0])));
CHECK_MEM_ERROR(cm, cpi->alt_ref_aq, vp9_alt_ref_aq_create());
CHECK_MEM_ERROR(
......@@ -3550,6 +3556,7 @@ static void encode_without_recode_loop(VP9_COMP *cpi, size_t *size,
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);
......@@ -4424,7 +4431,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, size_t *size,
#endif
#ifdef OUTPUT_YUV_SKINMAP
if (cpi->common.current_video_frame > 1) {
vp9_compute_skin_map(cpi, yuv_skinmap_file);
vp9_output_skin_map(cpi, yuv_skinmap_file);
}
#endif
......
......@@ -542,6 +542,8 @@ typedef struct VP9_COMP {
uint8_t *segmentation_map;
uint8_t *skin_map;
// segment threashold for encode breakout
int segment_encode_breakout[MAX_SEGMENTS];
......
......@@ -190,21 +190,12 @@ void vp9_update_noise_estimate(VP9_COMP *const cpi) {
int bl_index1 = bl_index + 1;
int bl_index2 = bl_index + cm->mi_cols;
int bl_index3 = bl_index2 + 1;
int is_skin = 0;
if (cpi->use_skin_detection) is_skin = cpi->skin_map[bl_index];
// Only consider blocks that are likely steady background. i.e, have
// been encoded as zero/low motion x (= thresh_consec_zeromv) frames
// in a row. consec_zero_mv[] defined for 8x8 blocks, so consider all
// 4 sub-blocks for 16x16 block. Also, avoid skin blocks.
int consec_zeromv =
VPXMIN(cpi->consec_zero_mv[bl_index],
VPXMIN(cpi->consec_zero_mv[bl_index1],
VPXMIN(cpi->consec_zero_mv[bl_index2],
cpi->consec_zero_mv[bl_index3])));
int is_skin = 0;
if (cpi->use_skin_detection) {
is_skin =
vp9_compute_skin_block(src_y, src_u, src_v, src_ystride,
src_uvstride, bsize, consec_zeromv, 0);
}
if (frame_low_motion &&
cpi->consec_zero_mv[bl_index] > thresh_consec_zeromv &&
cpi->consec_zero_mv[bl_index1] > thresh_consec_zeromv &&
......
......@@ -36,44 +36,31 @@ int vp9_compute_skin_block(const uint8_t *y, const uint8_t *u, const uint8_t *v,
}
}
#ifdef OUTPUT_YUV_SKINMAP
// For viewing skin map on input source.
void vp9_compute_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file) {
int i, j, mi_row, mi_col, num_bl;
void vp9_compute_skin_map(VP9_COMP *const cpi, BLOCK_SIZE bsize) {
int mi_row, mi_col, num_bl;
VP9_COMMON *const cm = &cpi->common;
uint8_t *y;
const uint8_t *src_y = cpi->Source->y_buffer;
const uint8_t *src_u = cpi->Source->u_buffer;
const uint8_t *src_v = cpi->Source->v_buffer;
const int src_ystride = cpi->Source->y_stride;
const int src_uvstride = cpi->Source->uv_stride;
int y_bsize = 16; // Use 8x8 or 16x16.
int uv_bsize = y_bsize >> 1;
int ypos = y_bsize >> 1;
int uvpos = uv_bsize >> 1;
int shy = (y_bsize == 8) ? 3 : 4;
int shuv = shy - 1;
int fac = y_bsize / 8;
const int y_bsize = 4 << b_width_log2_lookup[bsize];
const int uv_bsize = y_bsize >> 1;
const int shy = (y_bsize == 8) ? 3 : 4;
const int shuv = shy - 1;
const int fac = y_bsize / 8;
// Use center pixel or average of center 2x2 pixels.
int mode_filter = 0;
YV12_BUFFER_CONFIG skinmap;
memset(&skinmap, 0, sizeof(YV12_BUFFER_CONFIG));
if (vpx_alloc_frame_buffer(&skinmap, cm->width, cm->height, cm->subsampling_x,
cm->subsampling_y, VP9_ENC_BORDER_IN_PIXELS,
cm->byte_alignment)) {
vpx_free_frame_buffer(&skinmap);
return;
}
memset(skinmap.buffer_alloc, 128, skinmap.frame_size);
y = skinmap.y_buffer;
const int mode_filter = 0;
// Loop through blocks and set skin map based on center pixel of block.
// Set y to white for skin block, otherwise set to source with gray scale.
// Ignore rightmost/bottom boundary blocks.
for (mi_row = 0; mi_row < cm->mi_rows - 1; mi_row += fac) {
num_bl = 0;
for (mi_col = 0; mi_col < cm->mi_cols - 1; mi_col += fac) {
int is_skin = 0;
if (mode_filter == 1) {
const int ypos = y_bsize >> 1;
const int uvpos = uv_bsize >> 1;
// Use 2x2 average at center.
uint8_t ysource = src_y[ypos * src_ystride + ypos];
uint8_t usource = src_u[uvpos * src_uvstride + uvpos];
......@@ -92,13 +79,12 @@ void vp9_compute_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file) {
vsource = (vsource + vsource2 + vsource3 + vsource4) >> 2;
is_skin = vpx_skin_pixel(ysource, usource, vsource, 1);
} else {
int block_size = BLOCK_8X8;
int consec_zeromv = 0;
int bl_index = mi_row * cm->mi_cols + mi_col;
int bl_index1 = bl_index + 1;
int bl_index2 = bl_index + cm->mi_cols;
int bl_index3 = bl_index2 + 1;
if (y_bsize == 8)
if (bsize == BLOCK_8X8)
consec_zeromv = cpi->consec_zero_mv[bl_index];
else
consec_zeromv =
......@@ -106,29 +92,63 @@ void vp9_compute_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file) {
VPXMIN(cpi->consec_zero_mv[bl_index1],
VPXMIN(cpi->consec_zero_mv[bl_index2],
cpi->consec_zero_mv[bl_index3])));
if (y_bsize == 16) block_size = BLOCK_16X16;
is_skin =
vp9_compute_skin_block(src_y, src_u, src_v, src_ystride,
src_uvstride, block_size, consec_zeromv, 0);
is_skin = vp9_compute_skin_block(src_y, src_u, src_v, src_ystride,
src_uvstride, bsize, consec_zeromv, 0);
}
cpi->skin_map[mi_row * cm->mi_cols + mi_col] = is_skin;
num_bl++;
src_y += y_bsize;
src_u += uv_bsize;
src_v += uv_bsize;
}
src_y += (src_ystride << shy) - (num_bl << shy);
src_u += (src_uvstride << shuv) - (num_bl << shuv);
src_v += (src_uvstride << shuv) - (num_bl << shuv);
}
}
#ifdef OUTPUT_YUV_SKINMAP
// For viewing skin map on input source.
void vp9_output_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file) {
int i, j, mi_row, mi_col, num_bl;
VP9_COMMON *const cm = &cpi->common;
uint8_t *y;
const uint8_t *src_y = cpi->Source->y_buffer;
const int src_ystride = cpi->Source->y_stride;
const int y_bsize = 16; // Use 8x8 or 16x16.
const int shy = (y_bsize == 8) ? 3 : 4;
const int fac = y_bsize / 8;
YV12_BUFFER_CONFIG skinmap;
memset(&skinmap, 0, sizeof(YV12_BUFFER_CONFIG));
if (vpx_alloc_frame_buffer(&skinmap, cm->width, cm->height, cm->subsampling_x,
cm->subsampling_y, VP9_ENC_BORDER_IN_PIXELS,
cm->byte_alignment)) {
vpx_free_frame_buffer(&skinmap);
return;
}
memset(skinmap.buffer_alloc, 128, skinmap.frame_size);
y = skinmap.y_buffer;
// Loop through blocks and set skin map based on center pixel of block.
// Set y to white for skin block, otherwise set to source with gray scale.
// Ignore rightmost/bottom boundary blocks.
for (mi_row = 0; mi_row < cm->mi_rows - 1; mi_row += fac) {
num_bl = 0;
for (mi_col = 0; mi_col < cm->mi_cols - 1; mi_col += fac) {
const int block_index = mi_row * cm->mi_cols + mi_col;
const int is_skin = cpi->skin_map[block_index];
for (i = 0; i < y_bsize; i++) {
for (j = 0; j < y_bsize; j++) {
if (is_skin)
y[i * src_ystride + j] = 255;
else
y[i * src_ystride + j] = src_y[i * src_ystride + j];
y[i * src_ystride + j] = is_skin ? 255 : src_y[i * src_ystride + j];
}
}
num_bl++;
y += y_bsize;
src_y += y_bsize;
src_u += uv_bsize;
src_v += uv_bsize;
}
y += (src_ystride << shy) - (num_bl << shy);
src_y += (src_ystride << shy) - (num_bl << shy);
src_u += (src_uvstride << shuv) - (num_bl << shuv);
src_v += (src_uvstride << shuv) - (num_bl << shuv);
}
vpx_write_yuv_frame(yuv_skinmap_file, &skinmap);
vpx_free_frame_buffer(&skinmap);
......
......@@ -25,9 +25,11 @@ int vp9_compute_skin_block(const uint8_t *y, const uint8_t *u, const uint8_t *v,
int stride, int strideuv, int bsize,
int consec_zeromv, int curr_motion_magn);
void vp9_compute_skin_map(struct VP9_COMP *const cpi, BLOCK_SIZE bsize);
#ifdef OUTPUT_YUV_SKINMAP
// For viewing skin map on input source.
void vp9_compute_skin_map(struct VP9_COMP *const cpi, FILE *yuv_skinmap_file);
void vp9_output_skin_map(struct VP9_COMP *const cpi, FILE *yuv_skinmap_file);
#endif
#ifdef __cplusplus
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment