Commit 5668dcc7 authored by Frank Galligan's avatar Frank Galligan

Refactor dec_build_inter_predictors

Refactor the loops in dec_build_inter_predictors to try and decrease
the number of instructions. Limited testing saw about 1% perf
increase on x86 and about 0.67 % perf increase on Arm.

Change-Id: I69cfe6335bb562fbaaebf43fb3f5c5a2a28882a2
parent 59681be0
......@@ -1807,162 +1807,134 @@ static void high_build_mc_border(const uint8_t *src8, int src_stride,
#endif // CONFIG_VP9_HIGHBITDEPTH
void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
int plane, int block, int bw, int bh, int x,
int y, int w, int h, int mi_x, int mi_y) {
int plane, int bw, int bh, int x,
int y, int w, int h, int mi_x, int mi_y,
const InterpKernel *kernel,
const struct scale_factors *sf,
struct buf_2d *pre_buf, struct buf_2d *dst_buf,
const MV* mv, RefCntBuffer *ref_frame_buf,
int is_scaled, int ref) {
struct macroblockd_plane *const pd = &xd->plane[plane];
const MODE_INFO *mi = xd->mi[0].src_mi;
const int is_compound = has_second_ref(&mi->mbmi);
const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
int ref;
for (ref = 0; ref < 1 + is_compound; ++ref) {
const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
struct buf_2d *const pre_buf = &pd->pre[ref];
struct buf_2d *const dst_buf = &pd->dst;
uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
const MV mv = mi->mbmi.sb_type < BLOCK_8X8
? average_split_mvs(pd, mi, ref, block)
: mi->mbmi.mv[ref].as_mv;
const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
pd->subsampling_x,
pd->subsampling_y);
MV32 scaled_mv;
int xs, ys, x0, y0, x0_16, y0_16, y1, frame_width, frame_height,
buf_stride, subpel_x, subpel_y;
uint8_t *ref_frame, *buf_ptr;
const int idx = xd->block_refs[ref]->idx;
BufferPool *const pool = pbi->common.buffer_pool;
RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
const int is_scaled = vp9_is_scaled(sf);
// Get reference frame pointer, width and height.
if (plane == 0) {
frame_width = ref_frame_buf->buf.y_crop_width;
frame_height = ref_frame_buf->buf.y_crop_height;
ref_frame = ref_frame_buf->buf.y_buffer;
} else {
frame_width = ref_frame_buf->buf.uv_crop_width;
frame_height = ref_frame_buf->buf.uv_crop_height;
ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
: ref_frame_buf->buf.v_buffer;
}
uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, mv, bw, bh,
pd->subsampling_x,
pd->subsampling_y);
MV32 scaled_mv;
int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height,
buf_stride, subpel_x, subpel_y;
uint8_t *ref_frame, *buf_ptr;
// Get reference frame pointer, width and height.
if (plane == 0) {
frame_width = ref_frame_buf->buf.y_crop_width;
frame_height = ref_frame_buf->buf.y_crop_height;
ref_frame = ref_frame_buf->buf.y_buffer;
} else {
frame_width = ref_frame_buf->buf.uv_crop_width;
frame_height = ref_frame_buf->buf.uv_crop_height;
ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
: ref_frame_buf->buf.v_buffer;
}
if (is_scaled) {
// Co-ordinate of containing block to pixel precision.
int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
// Co-ordinate of the block to 1/16th pixel precision.
x0_16 = (x_start + x) << SUBPEL_BITS;
y0_16 = (y_start + y) << SUBPEL_BITS;
// Co-ordinate of current block in reference frame
// to 1/16th pixel precision.
x0_16 = sf->scale_value_x(x0_16, sf);
y0_16 = sf->scale_value_y(y0_16, sf);
// Map the top left corner of the block into the reference frame.
x0 = sf->scale_value_x(x_start + x, sf);
y0 = sf->scale_value_y(y_start + y, sf);
// Scale the MV and incorporate the sub-pixel offset of the block
// in the reference frame.
scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
xs = sf->x_step_q4;
ys = sf->y_step_q4;
} else {
// Co-ordinate of containing block to pixel precision.
x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
if (is_scaled) {
// Co-ordinate of containing block to pixel precision.
int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
// Co-ordinate of the block to 1/16th pixel precision.
x0_16 = (x_start + x) << SUBPEL_BITS;
y0_16 = (y_start + y) << SUBPEL_BITS;
// Co-ordinate of current block in reference frame
// to 1/16th pixel precision.
x0_16 = sf->scale_value_x(x0_16, sf);
y0_16 = sf->scale_value_y(y0_16, sf);
// Map the top left corner of the block into the reference frame.
x0 = sf->scale_value_x(x_start + x, sf);
y0 = sf->scale_value_y(y_start + y, sf);
// Scale the MV and incorporate the sub-pixel offset of the block
// in the reference frame.
scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
xs = sf->x_step_q4;
ys = sf->y_step_q4;
} else {
// Co-ordinate of containing block to pixel precision.
x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
// Co-ordinate of the block to 1/16th pixel precision.
x0_16 = x0 << SUBPEL_BITS;
y0_16 = y0 << SUBPEL_BITS;
// Co-ordinate of the block to 1/16th pixel precision.
x0_16 = x0 << SUBPEL_BITS;
y0_16 = y0 << SUBPEL_BITS;
scaled_mv.row = mv_q4.row;
scaled_mv.col = mv_q4.col;
xs = ys = 16;
scaled_mv.row = mv_q4.row;
scaled_mv.col = mv_q4.col;
xs = ys = 16;
}
subpel_x = scaled_mv.col & SUBPEL_MASK;
subpel_y = scaled_mv.row & SUBPEL_MASK;
// Calculate the top left corner of the best matching block in the
// reference frame.
x0 += scaled_mv.col >> SUBPEL_BITS;
y0 += scaled_mv.row >> SUBPEL_BITS;
x0_16 += scaled_mv.col;
y0_16 += scaled_mv.row;
// Get reference block pointer.
buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
buf_stride = pre_buf->stride;
// Do border extension if there is motion or the
// width/height is not a multiple of 8 pixels.
if (is_scaled || scaled_mv.col || scaled_mv.row ||
(frame_width & 0x7) || (frame_height & 0x7)) {
int y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
// Get reference block bottom right horizontal coordinate.
int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
int x_pad = 0, y_pad = 0;
if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
x0 -= VP9_INTERP_EXTEND - 1;
x1 += VP9_INTERP_EXTEND;
x_pad = 1;
}
subpel_x = scaled_mv.col & SUBPEL_MASK;
subpel_y = scaled_mv.row & SUBPEL_MASK;
// Calculate the top left corner of the best matching block in the
// reference frame.
x0 += scaled_mv.col >> SUBPEL_BITS;
y0 += scaled_mv.row >> SUBPEL_BITS;
x0_16 += scaled_mv.col;
y0_16 += scaled_mv.row;
// Get reference block pointer.
buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
buf_stride = pre_buf->stride;
// Get reference block bottom right vertical coordinate.
y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
// Do border extension if there is motion or the
// width/height is not a multiple of 8 pixels.
if (is_scaled || scaled_mv.col || scaled_mv.row ||
(frame_width & 0x7) || (frame_height & 0x7)) {
// Get reference block bottom right horizontal coordinate.
int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
int x_pad = 0, y_pad = 0;
if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
x0 -= VP9_INTERP_EXTEND - 1;
x1 += VP9_INTERP_EXTEND;
x_pad = 1;
}
if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
y0 -= VP9_INTERP_EXTEND - 1;
y1 += VP9_INTERP_EXTEND;
y_pad = 1;
}
if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
y0 -= VP9_INTERP_EXTEND - 1;
y1 += VP9_INTERP_EXTEND;
y_pad = 1;
}
// Wait until reference block is ready. Pad 7 more pixels as last 7
// pixels of each superblock row can be changed by next superblock row.
if (pbi->frame_parallel_decode)
vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
// Skip border extension if block is inside the frame.
if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
// Extend the border.
// Wait until reference block is ready. Pad 7 more pixels as last 7
// pixels of each superblock row can be changed by next superblock row.
if (pbi->frame_parallel_decode)
vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
// Skip border extension if block is inside the frame.
if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
// Extend the border.
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
high_build_mc_border(buf_ptr1,
pre_buf->stride,
xd->mc_buf_high,
x1 - x0 + 1,
x0,
y0,
x1 - x0 + 1,
y1 - y0 + 1,
frame_width,
frame_height);
buf_stride = x1 - x0 + 1;
buf_ptr = CONVERT_TO_BYTEPTR(xd->mc_buf_high) +
y_pad * 3 * buf_stride + x_pad * 3;
} else {
build_mc_border(buf_ptr1,
pre_buf->stride,
xd->mc_buf,
x1 - x0 + 1,
x0,
y0,
x1 - x0 + 1,
y1 - y0 + 1,
frame_width,
frame_height);
buf_stride = x1 - x0 + 1;
buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
}
#else
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
high_build_mc_border(buf_ptr1,
pre_buf->stride,
xd->mc_buf_high,
x1 - x0 + 1,
x0,
y0,
x1 - x0 + 1,
y1 - y0 + 1,
frame_width,
frame_height);
buf_stride = x1 - x0 + 1;
buf_ptr = CONVERT_TO_BYTEPTR(xd->mc_buf_high) +
y_pad * 3 * buf_stride + x_pad * 3;
} else {
build_mc_border(buf_ptr1,
pre_buf->stride,
xd->mc_buf,
......@@ -1975,28 +1947,43 @@ void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
frame_height);
buf_stride = x1 - x0 + 1;
buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
#endif // CONFIG_VP9_HIGHBITDEPTH
}
} else {
// Wait until reference block is ready. Pad 7 more pixels as last 7
// pixels of each superblock row can be changed by next superblock row.
if (pbi->frame_parallel_decode)
vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
#else
build_mc_border(buf_ptr1,
pre_buf->stride,
xd->mc_buf,
x1 - x0 + 1,
x0,
y0,
x1 - x0 + 1,
y1 - y0 + 1,
frame_width,
frame_height);
buf_stride = x1 - x0 + 1;
buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
#endif // CONFIG_VP9_HIGHBITDEPTH
}
} else {
// Wait until reference block is ready. Pad 7 more pixels as last 7
// pixels of each superblock row can be changed by next superblock row.
if (pbi->frame_parallel_decode) {
const int y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
}
}
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
} else {
inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
subpel_y, sf, w, h, ref, kernel, xs, ys);
}
#else
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
} else {
inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
subpel_y, sf, w, h, ref, kernel, xs, ys);
#endif // CONFIG_VP9_HIGHBITDEPTH
}
#else
inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
subpel_y, sf, w, h, ref, kernel, xs, ys);
#endif // CONFIG_VP9_HIGHBITDEPTH
}
void vp9_dec_build_inter_predictors_sb(VP9Decoder *const pbi, MACROBLOCKD *xd,
......@@ -2005,24 +1992,50 @@ void vp9_dec_build_inter_predictors_sb(VP9Decoder *const pbi, MACROBLOCKD *xd,
int plane;
const int mi_x = mi_col * MI_SIZE;
const int mi_y = mi_row * MI_SIZE;
const MODE_INFO *mi = xd->mi[0].src_mi;
const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
const BLOCK_SIZE sb_type = mi->mbmi.sb_type;
const int is_compound = has_second_ref(&mi->mbmi);
for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
&xd->plane[plane]);
struct macroblockd_plane *const pd = &xd->plane[plane];
struct buf_2d *const dst_buf = &pd->dst;
const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
const int bw = 4 * num_4x4_w;
const int bh = 4 * num_4x4_h;
if (xd->mi[0].src_mi->mbmi.sb_type < BLOCK_8X8) {
int i = 0, x, y;
assert(bsize == BLOCK_8X8);
for (y = 0; y < num_4x4_h; ++y)
for (x = 0; x < num_4x4_w; ++x)
dec_build_inter_predictors(pbi, xd, plane, i++, bw, bh,
4 * x, 4 * y, 4, 4, mi_x, mi_y);
} else {
dec_build_inter_predictors(pbi, xd, plane, 0, bw, bh,
0, 0, bw, bh, mi_x, mi_y);
int ref;
for (ref = 0; ref < 1 + is_compound; ++ref) {
const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
struct buf_2d *const pre_buf = &pd->pre[ref];
const int idx = xd->block_refs[ref]->idx;
BufferPool *const pool = pbi->common.buffer_pool;
RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
const int is_scaled = vp9_is_scaled(sf);
if (sb_type < BLOCK_8X8) {
int i = 0, x, y;
assert(bsize == BLOCK_8X8);
for (y = 0; y < num_4x4_h; ++y) {
for (x = 0; x < num_4x4_w; ++x) {
const MV mv = average_split_mvs(pd, mi, ref, i++);
dec_build_inter_predictors(pbi, xd, plane, bw, bh,
4 * x, 4 * y, 4, 4, mi_x, mi_y, kernel,
sf, pre_buf, dst_buf, &mv,
ref_frame_buf, is_scaled, ref);
}
}
} else {
const MV mv = mi->mbmi.mv[ref].as_mv;
dec_build_inter_predictors(pbi, xd, plane, bw, bh,
0, 0, bw, bh, mi_x, mi_y, kernel,
sf, pre_buf, dst_buf, &mv, ref_frame_buf,
is_scaled, ref);
}
}
}
}
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