generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Speed up motion estimation using small partitions' result(experiment)

Browse Source 
Current partition checking starts from small sizes, and then goes up
to large sizes. This experiment uses the small partitions' motion
estimation result, which is already available, to speed up the
large partition's motion estimation. We can decide to skip some
patition checkings if they are unlikely choices. We could use the
motion vector(MV) result as current partition's prediction MV, limit
the search range and reference frame.

Current result at speed 1:
psnr loss: 1.19% for stdhd, 0.287% for derf.
speed gain: 14% for sunflower(hd), 11% for akiyo.

Further improvement will be done later.

Change-Id: I5abfd070e9cace2e91e2a0247d1325df313887ab
This commit is contained in:
Yunqing Wang
2013-07-03 14:43:23 -07:00
parent a33086f925
commit df90d58f4f
5 changed files with 179 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

132
vp9/encoder/vp9_encodeframe.c
View File

@@ -1466,6 +1466,138 @@ static void rd_pick_partition(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row,
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
}
}
x->fast_ms = 0;
x->pred_mv.as_int = 0;
x->subblock_ref = 0;
// Use 4 subblocks' motion estimation results to speed up current
// partition's checking.
if (cpi->sf.using_small_partition_info) {
// Only use 8x8 result for non HD videos.
// int use_8x8 = (MIN(cpi->common.width, cpi->common.height) < 720) ? 1 : 0;
int use_8x8 = 1;
if (cm->frame_type && !cpi->is_src_frame_alt_ref &&
((use_8x8 && bsize == BLOCK_SIZE_MB16X16) ||
bsize == BLOCK_SIZE_SB32X32 || bsize == BLOCK_SIZE_SB64X64)) {
int ref0 = 0, ref1 = 0, ref2 = 0, ref3 = 0;
if (bsize == BLOCK_SIZE_MB16X16) {
ref0 = x->sb8x8_context[xd->sb_index][xd->mb_index][0].mic.mbmi.
ref_frame[0];
ref1 = x->sb8x8_context[xd->sb_index][xd->mb_index][1].mic.mbmi.
ref_frame[0];
ref2 = x->sb8x8_context[xd->sb_index][xd->mb_index][2].mic.mbmi.
ref_frame[0];
ref3 = x->sb8x8_context[xd->sb_index][xd->mb_index][3].mic.mbmi.
ref_frame[0];
} else if (bsize == BLOCK_SIZE_SB32X32) {
ref0 = x->mb_context[xd->sb_index][0].mic.mbmi.ref_frame[0];
ref1 = x->mb_context[xd->sb_index][1].mic.mbmi.ref_frame[0];
ref2 = x->mb_context[xd->sb_index][2].mic.mbmi.ref_frame[0];
ref3 = x->mb_context[xd->sb_index][3].mic.mbmi.ref_frame[0];
} else if (bsize == BLOCK_SIZE_SB64X64) {
ref0 = x->sb32_context[0].mic.mbmi.ref_frame[0];
ref1 = x->sb32_context[1].mic.mbmi.ref_frame[0];
ref2 = x->sb32_context[2].mic.mbmi.ref_frame[0];
ref3 = x->sb32_context[3].mic.mbmi.ref_frame[0];
}
// Currently, only consider 4 inter ref frames.
if (ref0 && ref1 && ref2 && ref3) {
int16_t mvr0 = 0, mvc0 = 0, mvr1 = 0, mvc1 = 0, mvr2 = 0, mvc2 = 0,
mvr3 = 0, mvc3 = 0;
int d01, d23, d02, d13; // motion vector distance between 2 blocks
// Get each subblock's motion vectors.
if (bsize == BLOCK_SIZE_MB16X16) {
mvr0 = x->sb8x8_context[xd->sb_index][xd->mb_index][0].mic.mbmi.mv[0].
as_mv.row;
mvc0 = x->sb8x8_context[xd->sb_index][xd->mb_index][0].mic.mbmi.mv[0].
as_mv.col;
mvr1 = x->sb8x8_context[xd->sb_index][xd->mb_index][1].mic.mbmi.mv[0].
as_mv.row;
mvc1 = x->sb8x8_context[xd->sb_index][xd->mb_index][1].mic.mbmi.mv[0].
as_mv.col;
mvr2 = x->sb8x8_context[xd->sb_index][xd->mb_index][2].mic.mbmi.mv[0].
as_mv.row;
mvc2 = x->sb8x8_context[xd->sb_index][xd->mb_index][2].mic.mbmi.mv[0].
as_mv.col;
mvr3 = x->sb8x8_context[xd->sb_index][xd->mb_index][3].mic.mbmi.mv[0].
as_mv.row;
mvc3 = x->sb8x8_context[xd->sb_index][xd->mb_index][3].mic.mbmi.mv[0].
as_mv.col;
} else if (bsize == BLOCK_SIZE_SB32X32) {
mvr0 = x->mb_context[xd->sb_index][0].mic.mbmi.mv[0].as_mv.row;
mvc0 = x->mb_context[xd->sb_index][0].mic.mbmi.mv[0].as_mv.col;
mvr1 = x->mb_context[xd->sb_index][1].mic.mbmi.mv[0].as_mv.row;
mvc1 = x->mb_context[xd->sb_index][1].mic.mbmi.mv[0].as_mv.col;
mvr2 = x->mb_context[xd->sb_index][2].mic.mbmi.mv[0].as_mv.row;
mvc2 = x->mb_context[xd->sb_index][2].mic.mbmi.mv[0].as_mv.col;
mvr3 = x->mb_context[xd->sb_index][3].mic.mbmi.mv[0].as_mv.row;
mvc3 = x->mb_context[xd->sb_index][3].mic.mbmi.mv[0].as_mv.col;
} else if (bsize == BLOCK_SIZE_SB64X64) {
mvr0 = x->sb32_context[0].mic.mbmi.mv[0].as_mv.row;
mvc0 = x->sb32_context[0].mic.mbmi.mv[0].as_mv.col;
mvr1 = x->sb32_context[1].mic.mbmi.mv[0].as_mv.row;
mvc1 = x->sb32_context[1].mic.mbmi.mv[0].as_mv.col;
mvr2 = x->sb32_context[2].mic.mbmi.mv[0].as_mv.row;
mvc2 = x->sb32_context[2].mic.mbmi.mv[0].as_mv.col;
mvr3 = x->sb32_context[3].mic.mbmi.mv[0].as_mv.row;
mvc3 = x->sb32_context[3].mic.mbmi.mv[0].as_mv.col;
}
// Adjust sign if ref is alt_ref
if (cm->ref_frame_sign_bias[ref0]) {
mvr0 *= -1;
mvc0 *= -1;
}
if (cm->ref_frame_sign_bias[ref1]) {
mvr1 *= -1;
mvc1 *= -1;
}
if (cm->ref_frame_sign_bias[ref2]) {
mvr2 *= -1;
mvc2 *= -1;
}
if (cm->ref_frame_sign_bias[ref3]) {
mvr3 *= -1;
mvc3 *= -1;
}
// Calculate mv distances.
d01 = MAX(abs(mvr0 - mvr1), abs(mvc0 - mvc1));
d23 = MAX(abs(mvr2 - mvr3), abs(mvc2 - mvc3));
d02 = MAX(abs(mvr0 - mvr2), abs(mvc0 - mvc2));
d13 = MAX(abs(mvr1 - mvr3), abs(mvc1 - mvc3));
if (d01 < 24 && d23 < 24 && d02 < 24 && d13 < 24) {
// Set fast motion search level.
x->fast_ms = 1;
// Calculate prediction MV
x->pred_mv.as_mv.row = (mvr0 + mvr1 + mvr2 + mvr3) >> 2;
x->pred_mv.as_mv.col = (mvc0 + mvc1 + mvc2 + mvc3) >> 2;
if (ref0 == ref1 && ref1 == ref2 && ref2 == ref3 &&
d01 < 2 && d23 < 2 && d02 < 2 && d13 < 2) {
// Set fast motion search level.
x->fast_ms = 2;
if (!d01 && !d23 && !d02 && !d13) {
x->fast_ms = 3;
x->subblock_ref = ref0;
}
}
}
}
}
}
if (!cpi->sf.use_partitions_less_than
|| (cpi->sf.use_partitions_less_than
&& bsize <= cpi->sf.less_than_block_size)) {