ref: 0dccb6277c9350c157e866c243e9c93956635e6e
parent: cf3202132f74081292173e8cd6aff4bf9ca2c0af
author: Marco <marpan@google.com>
date: Mon Jan 5 12:13:13 EST 2015
Modify variance partition selection for low resolutions. For low spatial resolutions: bias partittion selection to smaller block sizes, and base the variance computation on 4x4 down-sampling. Also move the threshold computations into the choose_partitioning, so they are computed once for each sb block. On low-res clips (RTC_derf) PSNR/SSIMetrics increase by about 4-5%. No change for resolutions above CIF. Change-Id: I93f8ff742c8044786977bb6e31dcf8efda6dd1b0
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -401,55 +401,42 @@
void *data,
BLOCK_SIZE bsize,
int mi_row,
- int mi_col) {+ int mi_col,
+ int threshold,
+ BLOCK_SIZE bsize_min) {VP9_COMMON * const cm = &cpi->common;
variance_node vt;
const int block_width = num_8x8_blocks_wide_lookup[bsize];
const int block_height = num_8x8_blocks_high_lookup[bsize];
- // TODO(marpan): Adjust/tune these thresholds.
- const int threshold_multiplier = cm->frame_type == KEY_FRAME ? 80 : 4;
- int64_t threshold =
- (int64_t)(threshold_multiplier *
- vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth));
- int64_t threshold_bsize_ref = threshold << 6;
- int64_t threshold_low = threshold;
- BLOCK_SIZE bsize_ref = BLOCK_16X16;
assert(block_height == block_width);
tree_to_node(data, bsize, &vt);
- if (cm->frame_type == KEY_FRAME) {- bsize_ref = BLOCK_8X8;
- // Choose lower thresholds for key frame variance to favor split, but keep
- // threshold for splitting to 4x4 block still fairly high for now.
- threshold_bsize_ref = threshold << 2;
- threshold_low = threshold >> 2;
- }
-
- // For bsize=bsize_ref (16x16/8x8 for 8x8/4x4 downsampling), select if
+ // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
// variance is below threshold, otherwise split will be selected.
// No check for vert/horiz split as too few samples for variance.
- if (bsize == bsize_ref) {+ if (bsize == bsize_min) {get_variance(&vt.part_variances->none);
if (mi_col + block_width / 2 < cm->mi_cols &&
mi_row + block_height / 2 < cm->mi_rows &&
- vt.part_variances->none.variance < threshold_bsize_ref) {+ vt.part_variances->none.variance < threshold) {set_block_size(cpi, xd, mi_row, mi_col, bsize);
return 1;
}
return 0;
- } else if (bsize > bsize_ref) {+ } else if (bsize > bsize_min) {get_variance(&vt.part_variances->none);
- // For key frame, for bsize above 32X32, or very high variance, take split.
+ // For key frame or low_res: for bsize above 32X32 or very high variance,
+ // take split.
if (cm->frame_type == KEY_FRAME &&
(bsize > BLOCK_32X32 ||
- vt.part_variances->none.variance > (threshold << 2))) {+ vt.part_variances->none.variance > (threshold << 4))) {return 0;
}
// If variance is low, take the bsize (no split).
if (mi_col + block_width / 2 < cm->mi_cols &&
mi_row + block_height / 2 < cm->mi_rows &&
- vt.part_variances->none.variance < threshold_low) {+ vt.part_variances->none.variance < threshold) {set_block_size(cpi, xd, mi_row, mi_col, bsize);
return 1;
}
@@ -458,8 +445,8 @@
if (mi_row + block_height / 2 < cm->mi_rows) {get_variance(&vt.part_variances->vert[0]);
get_variance(&vt.part_variances->vert[1]);
- if (vt.part_variances->vert[0].variance < threshold_low &&
- vt.part_variances->vert[1].variance < threshold_low) {+ if (vt.part_variances->vert[0].variance < threshold &&
+ vt.part_variances->vert[1].variance < threshold) {BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
set_block_size(cpi, xd, mi_row, mi_col, subsize);
set_block_size(cpi, xd, mi_row, mi_col + block_width / 2, subsize);
@@ -470,8 +457,8 @@
if (mi_col + block_width / 2 < cm->mi_cols) {get_variance(&vt.part_variances->horz[0]);
get_variance(&vt.part_variances->horz[1]);
- if (vt.part_variances->horz[0].variance < threshold_low &&
- vt.part_variances->horz[1].variance < threshold_low) {+ if (vt.part_variances->horz[0].variance < threshold &&
+ vt.part_variances->horz[1].variance < threshold) {BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
set_block_size(cpi, xd, mi_row, mi_col, subsize);
set_block_size(cpi, xd, mi_row + block_height / 2, mi_col, subsize);
@@ -485,8 +472,7 @@
}
// This function chooses partitioning based on the variance between source and
-// reconstructed last, where variance is computed for downsampled inputs.
-// Currently 8x8 downsampling is used for delta frames, 4x4 for key frames.
+// reconstructed last, where variance is computed for downs-sampled inputs.
static void choose_partitioning(VP9_COMP *cpi,
const TileInfo *const tile,
MACROBLOCK *x,
@@ -496,6 +482,7 @@
int i, j, k, m;
v64x64 vt;
+ v16x16 vt2[16];
uint8_t *s;
const uint8_t *d;
int sp;
@@ -504,6 +491,27 @@
const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
const struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
+ // Always use 4x4 partition for key frame.
+ int use_4x4_partition = (cm->frame_type == KEY_FRAME);
+
+ int variance4x4downsample[16];
+ int low_res = (cm->width <= 352 && cm->height <= 288) ? 1 : 0;
+ const int threshold_multiplier = cm->frame_type == KEY_FRAME ? 80 : 4;
+ int64_t threshold_base = (int64_t)(threshold_multiplier *
+ vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth));
+ int64_t threshold = threshold_base;
+ int64_t threshold_bsize_min = threshold_base << 6;
+ int64_t threshold_bsize_max = threshold_base;
+ // Modify thresholds for key frame and for low-resolutions (set lower
+ // thresholds to favor split).
+ if (cm->frame_type == KEY_FRAME) {+ threshold = threshold_base >> 2;
+ threshold_bsize_min = threshold_base << 2;
+ } else if (low_res) {+ threshold_bsize_min = threshold_base << 3;
+ threshold_bsize_max = threshold_base >> 2;
+ }
+
vp9_clear_system_state();
set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
@@ -546,44 +554,62 @@
#endif // CONFIG_VP9_HIGHBITDEPTH
}
- // Fill in the entire tree of 8x8 variances for splits.
+ // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
+ // for splits.
for (i = 0; i < 4; i++) {const int x32_idx = ((i & 1) << 5);
const int y32_idx = ((i >> 1) << 5);
+ const int i2 = i << 2;
for (j = 0; j < 4; j++) {const int x16_idx = x32_idx + ((j & 1) << 4);
const int y16_idx = y32_idx + ((j >> 1) << 4);
v16x16 *vst = &vt.split[i].split[j];
- for (k = 0; k < 4; k++) {- int x8_idx = x16_idx + ((k & 1) << 3);
- int y8_idx = y16_idx + ((k >> 1) << 3);
- if (cm->frame_type != KEY_FRAME) {- unsigned int sse = 0;
- int sum = 0;
- if (x8_idx < pixels_wide && y8_idx < pixels_high) {- int s_avg, d_avg;
+ variance4x4downsample[i2 + j] = 0;
+ if (cm->frame_type != KEY_FRAME) {+ for (k = 0; k < 4; k++) {+ int x8_idx = x16_idx + ((k & 1) << 3);
+ int y8_idx = y16_idx + ((k >> 1) << 3);
+ unsigned int sse = 0;
+ int sum = 0;
+ if (x8_idx < pixels_wide && y8_idx < pixels_high) {+ int s_avg, d_avg;
#if CONFIG_VP9_HIGHBITDEPTH
- if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {- s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
- d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
- } else {+ if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {+ s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
+ d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
+ } else {+ s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
+ d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
+ }
+#else
s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
- }
-#else
- s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
- d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
#endif
- sum = s_avg - d_avg;
- sse = sum * sum;
- }
- // If variance is based on 8x8 downsampling, we stop here and have
- // one sample for 8x8 block (so use 1 for count in fill_variance),
- // which of course means variance = 0 for 8x8 block.
- fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
- } else {- // For key frame, go down to 4x4.
- v8x8 *vst2 = &vst->split[k];
+ sum = s_avg - d_avg;
+ sse = sum * sum;
+ }
+ // If variance is based on 8x8 downsampling, we stop here and have
+ // one sample for 8x8 block (so use 1 for count in fill_variance),
+ // which of course means variance = 0 for 8x8 block.
+ fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
+ }
+ fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
+ // For low-resolution, compute the variance based on 8x8 down-sampling,
+ // and if it is large (above the threshold) we go down for 4x4.
+ // For key frame we always go down to 4x4.
+ if (low_res)
+ get_variance(&vt.split[i].split[j].part_variances.none);
+ }
+ if (cm->frame_type == KEY_FRAME || (low_res &&
+ vt.split[i].split[j].part_variances.none.variance >
+ (threshold << 1))) {+ // Go down to 4x4 down-sampling for variance.
+ variance4x4downsample[i2 + j] = 1;
+ for (k = 0; k < 4; k++) {+ int x8_idx = x16_idx + ((k & 1) << 3);
+ int y8_idx = y16_idx + ((k >> 1) << 3);
+ v8x8 *vst2 = (cm->frame_type == KEY_FRAME) ? &vst->split[k] :
+ &vt2[i2 + j].split[k];
for (m = 0; m < 4; m++) {int x4_idx = x8_idx + ((m & 1) << 2);
int y4_idx = y8_idx + ((m >> 1) << 2);
@@ -590,79 +616,100 @@
unsigned int sse = 0;
int sum = 0;
if (x4_idx < pixels_wide && y4_idx < pixels_high) {+ int d_avg = 128;
#if CONFIG_VP9_HIGHBITDEPTH
int s_avg;
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {s_avg = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
+ if (cm->frame_type != KEY_FRAME)
+ d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
} else {s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
+ if (cm->frame_type != KEY_FRAME)
+ d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
}
#else
int s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
+ if (cm->frame_type != KEY_FRAME)
+ d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
#endif
- // For key frame, reference is set to 128.
- sum = s_avg - 128;
+ sum = s_avg - d_avg;
sse = sum * sum;
}
- // If variance is based on 4x4 downsampling, we stop here and have
+ // If variance is based on 4x4 down-sampling, we stop here and have
// one sample for 4x4 block (so use 1 for count in fill_variance),
// which of course means variance = 0 for 4x4 block.
- fill_variance(sse, sum, 0, &vst2->split[m].part_variances.none);
+ fill_variance(sse, sum, 0, &vst2->split[m].part_variances.none);
}
}
}
}
}
+
// Fill the rest of the variance tree by summing split partition values.
for (i = 0; i < 4; i++) {+ const int i2 = i << 2;
for (j = 0; j < 4; j++) {- if (cm->frame_type == KEY_FRAME) {+ if (variance4x4downsample[i2 + j] == 1) {+ v16x16 *vtemp = (cm->frame_type != KEY_FRAME) ? &vt2[i2 + j] :
+ &vt.split[i].split[j];
for (m = 0; m < 4; m++) {- fill_variance_tree(&vt.split[i].split[j].split[m], BLOCK_8X8);
+ fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
}
+ fill_variance_tree(vtemp, BLOCK_16X16);
}
- fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
}
fill_variance_tree(&vt.split[i], BLOCK_32X32);
}
fill_variance_tree(&vt, BLOCK_64X64);
+
// Now go through the entire structure, splitting every block size until
// we get to one that's got a variance lower than our threshold.
if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
- !set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col)) {+ !set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col,
+ threshold_bsize_max, BLOCK_16X16)) { for (i = 0; i < 4; ++i) {const int x32_idx = ((i & 1) << 2);
const int y32_idx = ((i >> 1) << 2);
+ const int i2 = i << 2;
if (!set_vt_partitioning(cpi, xd, &vt.split[i], BLOCK_32X32,
- (mi_row + y32_idx), (mi_col + x32_idx))) {+ (mi_row + y32_idx), (mi_col + x32_idx),
+ threshold, BLOCK_16X16)) { for (j = 0; j < 4; ++j) {const int x16_idx = ((j & 1) << 1);
const int y16_idx = ((j >> 1) << 1);
- // Note: If 8x8 downsampling is used for variance calculation we
- // cannot really select block size 8x8 (or even 8x16/16x8), since we
- // don't have sufficient samples for variance. So on delta frames,
- // 8x8 partition is only set if variance of the 16x16 block is very
- // high. For key frames, 4x4 downsampling is used, so we can better
- // select 8x16/16x8 and 8x8. 4x4 partition can potentially be set
- // used here too, but for now 4x4 is not allowed.
- if (!set_vt_partitioning(cpi, xd, &vt.split[i].split[j],
- BLOCK_16X16,
+ // TODO(marpan): Allow 4x4 partitions for inter-frames.
+ // use_4x4_partition = (variance4x4downsample[i2 + j] == 1);
+ // If 4x4 partition is not used, then 8x8 partition will be selected
+ // if variance of 16x16 block is very high, so use larger threshold
+ // for 16x16 (threshold_bsize_min) in that case.
+ uint64_t threshold_16x16 = (use_4x4_partition) ? threshold :
+ threshold_bsize_min;
+ BLOCK_SIZE bsize_min = (use_4x4_partition) ? BLOCK_8X8 : BLOCK_16X16;
+ // For inter frames: if variance4x4downsample[] == 1 for this 16x16
+ // block, then the variance is based on 4x4 down-sampling, so use vt2
+ // in set_vt_partioning(), otherwise use vt.
+ v16x16 *vtemp = (cm->frame_type != KEY_FRAME &&
+ variance4x4downsample[i2 + j] == 1) ?
+ &vt2[i2 + j] : &vt.split[i].split[j];
+ if (!set_vt_partitioning(cpi, xd, vtemp, BLOCK_16X16,
mi_row + y32_idx + y16_idx,
- mi_col + x32_idx + x16_idx)) {+ mi_col + x32_idx + x16_idx,
+ threshold_16x16, bsize_min)) { for (k = 0; k < 4; ++k) {const int x8_idx = (k & 1);
const int y8_idx = (k >> 1);
- if (cm->frame_type == KEY_FRAME) {- if (!set_vt_partitioning(cpi, xd,
- &vt.split[i].split[j].split[k],
+ if (use_4x4_partition) {+ if (!set_vt_partitioning(cpi, xd, &vtemp->split[k],
BLOCK_8X8,
mi_row + y32_idx + y16_idx + y8_idx,
- mi_col + x32_idx + x16_idx + x8_idx)) {- set_block_size(cpi, xd,
- (mi_row + y32_idx + y16_idx + y8_idx),
- (mi_col + x32_idx + x16_idx + x8_idx),
- BLOCK_4X4);
+ mi_col + x32_idx + x16_idx + x8_idx,
+ threshold_bsize_min, BLOCK_8X8)) {+ set_block_size(cpi, xd,
+ (mi_row + y32_idx + y16_idx + y8_idx),
+ (mi_col + x32_idx + x16_idx + x8_idx),
+ BLOCK_4X4);
}
} else {set_block_size(cpi, xd,
@@ -669,7 +716,7 @@
(mi_row + y32_idx + y16_idx + y8_idx),
(mi_col + x32_idx + x16_idx + x8_idx),
BLOCK_8X8);
- }
+ }
}
}
}