// Iterate over blocks within a superblock
static void av1_clpf_sb(const YV12_BUFFER_CONFIG *frame_buffer,
const AV1_COMMON *cm, MACROBLOCKD *xd,
MODE_INFO *const *mi_8x8, int xpos, int ypos) {
// Temporary buffer (to allow SIMD parallelism)
uint8_t buf_unaligned[BS * BS + 15];
uint8_t *buf = (uint8_t *)(((intptr_t)buf_unaligned + 15) & ~15);
int x, y, p;
for (p = 0; p < (CLPF_FILTER_ALL_PLANES ? MAX_MB_PLANE : 1); p++) {
for (y = 0; y < MI_BLOCK_SIZE && ypos + y < cm->mi_rows; y++) {
for (x = 0; x < MI_BLOCK_SIZE && xpos + x < cm->mi_cols; x++) {
const MB_MODE_INFO *mbmi =
&mi_8x8[(ypos + y) * cm->mi_stride + xpos + x]->mbmi;
// Do not filter if there is no residual
if (!mbmi->skip) {
// Do not filter frame edges
int has_top = ypos + y > 0;
int has_left = xpos + x > 0;
int has_bottom = ypos + y < cm->mi_rows - 1;
int has_right = xpos + x < cm->mi_cols - 1;
#if CLPF_ALLOW_BLOCK_PARALLELISM
// Do not filter superblock edges
has_top &= !!y;
has_left &= !!x;
has_bottom &= y != MI_BLOCK_SIZE - 1;
has_right &= x != MI_BLOCK_SIZE - 1;
#endif
av1_setup_dst_planes(xd->plane, frame_buffer, ypos + y, xpos + x);
clpf_block(
xd->plane[p].dst.buf, CLPF_ALLOW_PIXEL_PARALLELISM
? buf + y * MI_SIZE * BS + x * MI_SIZE
: xd->plane[p].dst.buf,
xd->plane[p].dst.stride,
CLPF_ALLOW_PIXEL_PARALLELISM ? BS : xd->plane[p].dst.stride,
has_top, has_left, has_bottom, has_right,
MI_SIZE >> xd->plane[p].subsampling_x,
MI_SIZE >> xd->plane[p].subsampling_y);
}
}
}
#if CLPF_ALLOW_PIXEL_PARALLELISM
for (y = 0; y < MI_BLOCK_SIZE && ypos + y < cm->mi_rows; y++) {
for (x = 0; x < MI_BLOCK_SIZE && xpos + x < cm->mi_cols; x++) {
const MB_MODE_INFO *mbmi =
&mi_8x8[(ypos + y) * cm->mi_stride + xpos + x]->mbmi;
av1_setup_dst_planes(xd->plane, frame_buffer, ypos + y, xpos + x);
if (!mbmi->skip) {
int i = 0;
for (i = 0; i<MI_SIZE>> xd->plane[p].subsampling_y; i++)
memcpy(xd->plane[p].dst.buf + i * xd->plane[p].dst.stride,
buf + (y * MI_SIZE + i) * BS + x * MI_SIZE,
MI_SIZE >> xd->plane[p].subsampling_x);
}
}
}
static int loop_filter_row_worker(AV1LfSync *const lf_sync,
LFWorkerData *const lf_data) {
const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
const int sb_cols =
mi_cols_aligned_to_sb(lf_data->cm) >> lf_data->cm->mib_size_log2;
int mi_row, mi_col;
#if !CONFIG_EXT_PARTITION_TYPES
enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
#endif // !CONFIG_EXT_PARTITION_TYPES
#if CONFIG_EXT_PARTITION
printf(
"STOPPING: This code has not been modified to work with the "
"extended coding unit size experiment");
exit(EXIT_FAILURE);
#endif // CONFIG_EXT_PARTITION
for (mi_row = lf_data->start; mi_row < lf_data->stop;
mi_row += lf_sync->num_workers * lf_data->cm->mib_size) {
MODE_INFO **const mi =
lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
for (mi_col = 0; mi_col < lf_data->cm->mi_cols;
mi_col += lf_data->cm->mib_size) {
const int r = mi_row >> lf_data->cm->mib_size_log2;
const int c = mi_col >> lf_data->cm->mib_size_log2;
#if !CONFIG_EXT_PARTITION_TYPES
LOOP_FILTER_MASK lfm;
#endif
int plane;
sync_read(lf_sync, r, c);
av1_setup_dst_planes(lf_data->planes, lf_data->cm->sb_size,
lf_data->frame_buffer, mi_row, mi_col);
#if CONFIG_EXT_PARTITION_TYPES
for (plane = 0; plane < num_planes; ++plane) {
av1_filter_block_plane_non420_ver(lf_data->cm, &lf_data->planes[plane],
mi + mi_col, mi_row, mi_col, plane);
av1_filter_block_plane_non420_hor(lf_data->cm, &lf_data->planes[plane],
mi + mi_col, mi_row, mi_col, plane);
}
#else
av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
lf_data->cm->mi_stride, &lfm);
for (plane = 0; plane < num_planes; ++plane) {
loop_filter_block_plane_ver(lf_data->cm, lf_data->planes, plane,
mi + mi_col, mi_row, mi_col, path, &lfm);
loop_filter_block_plane_hor(lf_data->cm, lf_data->planes, plane,
mi + mi_col, mi_row, mi_col, path, &lfm);
}
#endif // CONFIG_EXT_PARTITION_TYPES
sync_write(lf_sync, r, c, sb_cols);
}
}
return 1;
}
static int loop_filter_hor_row_worker(AV1LfSync *const lf_sync,
LFWorkerData *const lf_data) {
const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
const int sb_cols =
mi_cols_aligned_to_sb(lf_data->cm) >> lf_data->cm->mib_size_log2;
int mi_row, mi_col;
#if !CONFIG_EXT_PARTITION_TYPES
enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
#endif
for (mi_row = lf_data->start; mi_row < lf_data->stop;
mi_row += lf_sync->num_workers * lf_data->cm->mib_size) {
MODE_INFO **const mi =
lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
for (mi_col = 0; mi_col < lf_data->cm->mi_cols;
mi_col += lf_data->cm->mib_size) {
const int r = mi_row >> lf_data->cm->mib_size_log2;
const int c = mi_col >> lf_data->cm->mib_size_log2;
LOOP_FILTER_MASK lfm;
int plane;
// TODO([email protected]): For better parallelization, reorder
// the outer loop to column-based and remove the synchronizations here.
sync_read(lf_sync, r, c);
av1_setup_dst_planes(lf_data->planes, lf_data->cm->sb_size,
lf_data->frame_buffer, mi_row, mi_col);
av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
lf_data->cm->mi_stride, &lfm);
#if CONFIG_EXT_PARTITION_TYPES
for (plane = 0; plane < num_planes; ++plane)
av1_filter_block_plane_non420_hor(lf_data->cm, &lf_data->planes[plane],
mi + mi_col, mi_row, mi_col, plane);
#else
for (plane = 0; plane < num_planes; ++plane)
loop_filter_block_plane_hor(lf_data->cm, lf_data->planes, plane,
mi + mi_col, mi_row, mi_col, path, &lfm);
#endif
sync_write(lf_sync, r, c, sb_cols);
}
}
return 1;
}
static int loop_filter_ver_row_worker(AV1LfSync *const lf_sync,
LFWorkerData *const lf_data) {
const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
int mi_row, mi_col;
#if !CONFIG_EXT_PARTITION_TYPES
enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
#endif
for (mi_row = lf_data->start; mi_row < lf_data->stop;
mi_row += lf_sync->num_workers * lf_data->cm->mib_size) {
MODE_INFO **const mi =
lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
for (mi_col = 0; mi_col < lf_data->cm->mi_cols;
mi_col += lf_data->cm->mib_size) {
LOOP_FILTER_MASK lfm;
int plane;
av1_setup_dst_planes(lf_data->planes, lf_data->cm->sb_size,
lf_data->frame_buffer, mi_row, mi_col);
av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
lf_data->cm->mi_stride, &lfm);
#if CONFIG_EXT_PARTITION_TYPES
for (plane = 0; plane < num_planes; ++plane)
av1_filter_block_plane_non420_ver(lf_data->cm, &lf_data->planes[plane],
mi + mi_col, mi_row, mi_col, plane);
#else
for (plane = 0; plane < num_planes; ++plane)
loop_filter_block_plane_ver(lf_data->cm, lf_data->planes, plane,
mi + mi_col, mi_row, mi_col, path, &lfm);
#endif
}
}
return 1;
}
void av1_dering_frame(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
MACROBLOCKD *xd, int global_level) {
int r, c;
int sbr, sbc;
int nhsb, nvsb;
int16_t src[OD_DERING_INBUF_SIZE];
int16_t *linebuf[3];
int16_t colbuf[3][OD_BSIZE_MAX + 2 * OD_FILT_VBORDER][OD_FILT_HBORDER];
dering_list dlist[MAX_MIB_SIZE * MAX_MIB_SIZE];
unsigned char *row_dering, *prev_row_dering, *curr_row_dering;
int dering_count;
int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
int stride;
int bsize[3];
int dec[3];
int pli;
int dering_left;
int coeff_shift = AOMMAX(cm->bit_depth - 8, 0);
int nplanes;
if (xd->plane[1].subsampling_x == xd->plane[1].subsampling_y &&
xd->plane[2].subsampling_x == xd->plane[2].subsampling_y)
nplanes = 3;
else
nplanes = 1;
nvsb = (cm->mi_rows + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
nhsb = (cm->mi_cols + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
av1_setup_dst_planes(xd->plane, frame, 0, 0);
row_dering = aom_malloc(sizeof(*row_dering) * nhsb * 2);
memset(row_dering, 1, sizeof(*row_dering) * (nhsb + 2) * 2);
prev_row_dering = row_dering + 1;
curr_row_dering = prev_row_dering + nhsb + 2;
for (pli = 0; pli < nplanes; pli++) {
dec[pli] = xd->plane[pli].subsampling_x;
bsize[pli] = OD_DERING_SIZE_LOG2 - dec[pli];
}
stride = (cm->mi_cols << bsize[0]) + 2 * OD_FILT_HBORDER;
for (pli = 0; pli < nplanes; pli++) {
linebuf[pli] = aom_malloc(sizeof(*linebuf) * OD_FILT_VBORDER * stride);
}
for (sbr = 0; sbr < nvsb; sbr++) {
for (pli = 0; pli < nplanes; pli++) {
for (r = 0; r < (MAX_MIB_SIZE << bsize[pli]) + 2 * OD_FILT_VBORDER; r++) {
for (c = 0; c < OD_FILT_HBORDER; c++) {
colbuf[pli][r][c] = OD_DERING_VERY_LARGE;
}
}
}
dering_left = 1;
for (sbc = 0; sbc < nhsb; sbc++) {
int level;
int nhb, nvb;
int cstart = 0;
if (!dering_left) cstart = -OD_FILT_HBORDER;
nhb = AOMMIN(MAX_MIB_SIZE, cm->mi_cols - MAX_MIB_SIZE * sbc);
nvb = AOMMIN(MAX_MIB_SIZE, cm->mi_rows - MAX_MIB_SIZE * sbr);
level = compute_level_from_index(
global_level, cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
MAX_MIB_SIZE * sbc]
->mbmi.dering_gain);
curr_row_dering[sbc] = 0;
if (level == 0 ||
(dering_count = sb_compute_dering_list(
cm, sbr * MAX_MIB_SIZE, sbc * MAX_MIB_SIZE, dlist)) == 0) {
dering_left = 0;
continue;
}
curr_row_dering[sbc] = 1;
for (pli = 0; pli < nplanes; pli++) {
int16_t dst[OD_BSIZE_MAX * OD_BSIZE_MAX];
int threshold;
int coffset;
int rend, cend;
if (sbc == nhsb - 1)
cend = (nhb << bsize[pli]);
else
cend = (nhb << bsize[pli]) + OD_FILT_HBORDER;
if (sbr == nvsb - 1)
rend = (nvb << bsize[pli]);
else
rend = (nvb << bsize[pli]) + OD_FILT_VBORDER;
coffset = sbc * MAX_MIB_SIZE << bsize[pli];
if (sbc == nhsb - 1) {
/* On the last superblock column, fill in the right border with
OD_DERING_VERY_LARGE to avoid filtering with the outside. */
for (r = 0; r < rend + OD_FILT_VBORDER; r++) {
for (c = cend; c < (nhb << bsize[pli]) + OD_FILT_HBORDER; ++c) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
OD_DERING_VERY_LARGE;
}
}
}
if (sbr == nvsb - 1) {
/* On the last superblock row, fill in the bottom border with
OD_DERING_VERY_LARGE to avoid filtering with the outside. */
for (r = rend; r < rend + OD_FILT_VBORDER; r++) {
for (c = 0; c < (nhb << bsize[pli]) + 2 * OD_FILT_HBORDER; c++) {
src[(r + OD_FILT_VBORDER) * OD_FILT_BSTRIDE + c] =
OD_DERING_VERY_LARGE;
}
}
}
/* Copy in the pixels we need from the current superblock for
deringing.*/
copy_sb8_16(
cm,
&src[OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER + cstart],
OD_FILT_BSTRIDE, xd->plane[pli].dst.buf,
(MAX_MIB_SIZE << bsize[pli]) * sbr, coffset + cstart,
xd->plane[pli].dst.stride, rend, cend - cstart);
if (!prev_row_dering[sbc]) {
copy_sb8_16(cm, &src[OD_FILT_HBORDER], OD_FILT_BSTRIDE,
xd->plane[pli].dst.buf,
(MAX_MIB_SIZE << bsize[pli]) * sbr - OD_FILT_VBORDER,
coffset, xd->plane[pli].dst.stride, OD_FILT_VBORDER,
nhb << bsize[pli]);
} else if (sbr > 0) {
for (r = 0; r < OD_FILT_VBORDER; r++) {
for (c = 0; c < nhb << bsize[pli]; c++) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
linebuf[pli][r * stride + coffset + c];
}
}
} else {
for (r = 0; r < OD_FILT_VBORDER; r++) {
for (c = 0; c < nhb << bsize[pli]; c++) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
OD_DERING_VERY_LARGE;
}
}
}
if (!prev_row_dering[sbc - 1]) {
copy_sb8_16(cm, src, OD_FILT_BSTRIDE, xd->plane[pli].dst.buf,
(MAX_MIB_SIZE << bsize[pli]) * sbr - OD_FILT_VBORDER,
coffset - OD_FILT_HBORDER, xd->plane[pli].dst.stride,
OD_FILT_VBORDER, OD_FILT_HBORDER);
} else if (sbr > 0 && sbc > 0) {
for (r = 0; r < OD_FILT_VBORDER; r++) {
for (c = -OD_FILT_HBORDER; c < 0; c++) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
linebuf[pli][r * stride + coffset + c];
}
}
} else {
for (r = 0; r < OD_FILT_VBORDER; r++) {
for (c = -OD_FILT_HBORDER; c < 0; c++) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
OD_DERING_VERY_LARGE;
}
}
}
if (!prev_row_dering[sbc + 1]) {
copy_sb8_16(cm, &src[OD_FILT_HBORDER + (nhb << bsize[pli])],
OD_FILT_BSTRIDE, xd->plane[pli].dst.buf,
(MAX_MIB_SIZE << bsize[pli]) * sbr - OD_FILT_VBORDER,
coffset + (nhb << bsize[pli]), xd->plane[pli].dst.stride,
OD_FILT_VBORDER, OD_FILT_HBORDER);
} else if (sbr > 0 && sbc < nhsb - 1) {
for (r = 0; r < OD_FILT_VBORDER; r++) {
for (c = nhb << bsize[pli];
c < (nhb << bsize[pli]) + OD_FILT_HBORDER; c++) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
linebuf[pli][r * stride + coffset + c];
}
}
} else {
for (r = 0; r < OD_FILT_VBORDER; r++) {
for (c = nhb << bsize[pli];
c < (nhb << bsize[pli]) + OD_FILT_HBORDER; c++) {
src[r * OD_FILT_BSTRIDE + c + OD_FILT_HBORDER] =
OD_DERING_VERY_LARGE;
}
}
}
if (dering_left) {
/* If we deringed the superblock on the left then we need to copy in
saved pixels. */
for (r = 0; r < rend + OD_FILT_VBORDER; r++) {
for (c = 0; c < OD_FILT_HBORDER; c++) {
src[r * OD_FILT_BSTRIDE + c] = colbuf[pli][r][c];
}
}
}
for (r = 0; r < rend + OD_FILT_VBORDER; r++) {
for (c = 0; c < OD_FILT_HBORDER; c++) {
/* Saving pixels in case we need to dering the superblock on the
right. */
colbuf[pli][r][c] =
src[r * OD_FILT_BSTRIDE + c + (nhb << bsize[pli])];
}
}
copy_sb8_16(cm, &linebuf[pli][coffset], stride, xd->plane[pli].dst.buf,
(MAX_MIB_SIZE << bsize[pli]) * (sbr + 1) - OD_FILT_VBORDER,
coffset, xd->plane[pli].dst.stride, OD_FILT_VBORDER,
(nhb << bsize[pli]));
/* FIXME: This is a temporary hack that uses more conservative
deringing for chroma. */
if (pli)
threshold = (level * 5 + 4) >> 3 << coeff_shift;
else
threshold = level << coeff_shift;
if (threshold == 0) continue;
od_dering(
dst, &src[OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER],
dec[pli], dir, pli, dlist, dering_count, threshold, coeff_shift);
#if CONFIG_AOM_HIGHBITDEPTH
if (cm->use_highbitdepth) {
copy_dering_16bit_to_16bit(
(int16_t *)&CONVERT_TO_SHORTPTR(
xd->plane[pli]
.dst.buf)[xd->plane[pli].dst.stride *
(MAX_MIB_SIZE * sbr << bsize[pli]) +
(sbc * MAX_MIB_SIZE << bsize[pli])],
xd->plane[pli].dst.stride, dst, dlist, dering_count,
3 - dec[pli]);
} else {
#endif
copy_dering_16bit_to_8bit(
&xd->plane[pli].dst.buf[xd->plane[pli].dst.stride *
(MAX_MIB_SIZE * sbr << bsize[pli]) +
(sbc * MAX_MIB_SIZE << bsize[pli])],
xd->plane[pli].dst.stride, dst, dlist, dering_count, bsize[pli]);
#if CONFIG_AOM_HIGHBITDEPTH
}
#endif
}
dering_left = 1;
}
{
unsigned char *tmp;
tmp = prev_row_dering;
prev_row_dering = curr_row_dering;
curr_row_dering = tmp;
}
}
