ref: 60c58b52846ee9fa2672f7843ce7b22bc46a699f
parent: c7dc1d78bf26a2e883db485c5bba990e19db5f25
author: Ronald S. Bultje <rsbultje@gmail.com>
date: Mon Oct 12 13:54:25 EDT 2015
vp10: per-segment lossless coding. Some more testing of this patch would probably be useful, but I think the basics of it should work fine now. See issue 1035. Change-Id: I4a36d58f671c5391cb09d564581784a00ed26245
--- a/vp10/common/blockd.h
+++ b/vp10/common/blockd.h
@@ -211,7 +211,7 @@
int bd;
#endif
- int lossless;
+ int lossless[MAX_SEGMENTS];
int corrupted;
struct vpx_internal_error_info *error_info;
@@ -240,8 +240,8 @@
const MODE_INFO *const mi = xd->mi[0];
const MB_MODE_INFO *const mbmi = &mi->mbmi;
- if (plane_type != PLANE_TYPE_Y || xd->lossless || is_inter_block(mbmi) ||
- mbmi->tx_size >= TX_32X32)
+ if (plane_type != PLANE_TYPE_Y || xd->lossless[mbmi->segment_id] ||
+ is_inter_block(mbmi) || mbmi->tx_size >= TX_32X32)
return DCT_DCT;
return intra_mode_to_tx_type_lookup[get_y_mode(mi, block_idx)];
--- a/vp10/decoder/decodeframe.c
+++ b/vp10/decoder/decodeframe.c
@@ -213,6 +213,7 @@
int eob, int block) {struct macroblockd_plane *const pd = &xd->plane[plane];
TX_TYPE tx_type = get_tx_type(pd->plane_type, xd, block);
+ const int seg_id = xd->mi[0]->mbmi.segment_id;
if (eob > 0) {tran_low_t *const dqcoeff = pd->dqcoeff;
#if CONFIG_VP9_HIGHBITDEPTH
@@ -220,7 +221,7 @@
switch (tx_size) {case TX_4X4:
vp10_highbd_inv_txfm_add_4x4(dqcoeff, dst, stride, eob, xd->bd,
- tx_type, xd->lossless);
+ tx_type, xd->lossless[seg_id]);
break;
case TX_8X8:
vp10_highbd_inv_txfm_add_8x8(dqcoeff, dst, stride, eob, xd->bd,
@@ -243,7 +244,7 @@
switch (tx_size) {case TX_4X4:
vp10_inv_txfm_add_4x4(dqcoeff, dst, stride, eob, tx_type,
- xd->lossless);
+ xd->lossless[seg_id]);
break;
case TX_8X8:
vp10_inv_txfm_add_8x8(dqcoeff, dst, stride, eob, tx_type);
@@ -281,6 +282,7 @@
uint8_t *dst, int stride,
int eob) {struct macroblockd_plane *const pd = &xd->plane[plane];
+ const int seg_id = xd->mi[0]->mbmi.segment_id;
if (eob > 0) {tran_low_t *const dqcoeff = pd->dqcoeff;
#if CONFIG_VP9_HIGHBITDEPTH
@@ -288,7 +290,7 @@
switch (tx_size) {case TX_4X4:
vp10_highbd_inv_txfm_add_4x4(dqcoeff, dst, stride, eob, xd->bd,
- tx_type, xd->lossless);
+ tx_type, xd->lossless[seg_id]);
break;
case TX_8X8:
vp10_highbd_inv_txfm_add_8x8(dqcoeff, dst, stride, eob, xd->bd,
@@ -311,7 +313,7 @@
switch (tx_size) {case TX_4X4:
vp10_inv_txfm_add_4x4(dqcoeff, dst, stride, eob, tx_type,
- xd->lossless);
+ xd->lossless[seg_id]);
break;
case TX_8X8:
vp10_inv_txfm_add_8x8(dqcoeff, dst, stride, eob, tx_type);
@@ -1144,15 +1146,26 @@
static void setup_quantization(VP10_COMMON *const cm, MACROBLOCKD *const xd,
struct vpx_read_bit_buffer *rb) {+ int i;
+
cm->base_qindex = vpx_rb_read_literal(rb, QINDEX_BITS);
cm->y_dc_delta_q = read_delta_q(rb);
cm->uv_dc_delta_q = read_delta_q(rb);
cm->uv_ac_delta_q = read_delta_q(rb);
cm->dequant_bit_depth = cm->bit_depth;
- xd->lossless = cm->base_qindex == 0 &&
- cm->y_dc_delta_q == 0 &&
- cm->uv_dc_delta_q == 0 &&
- cm->uv_ac_delta_q == 0;
+ for (i = 0; i < (cm->seg.enabled ? MAX_SEGMENTS : 1); ++i) {+#if CONFIG_MISC_FIXES
+ const int qindex = vp10_get_qindex(&cm->seg, i, cm->base_qindex);
+#endif
+ xd->lossless[i] = cm->y_dc_delta_q == 0 &&
+#if CONFIG_MISC_FIXES
+ qindex == 0 &&
+#else
+ cm->base_qindex == 0 &&
+#endif
+ cm->uv_dc_delta_q == 0 &&
+ cm->uv_ac_delta_q == 0;
+ }
#if CONFIG_VP9_HIGHBITDEPTH
xd->bd = (int)cm->bit_depth;
@@ -2098,7 +2111,8 @@
setup_segmentation(cm, rb);
setup_segmentation_dequant(cm);
#if CONFIG_MISC_FIXES
- cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(rb);
+ cm->tx_mode = (!cm->seg.enabled && xd->lossless[0]) ? ONLY_4X4
+ : read_tx_mode(rb);
cm->reference_mode = read_frame_reference_mode(cm, rb);
#endif
@@ -2128,7 +2142,7 @@
"Failed to allocate bool decoder 0");
#if !CONFIG_MISC_FIXES
- cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(&r);
+ cm->tx_mode = xd->lossless[0] ? ONLY_4X4 : read_tx_mode(&r);
#endif
if (cm->tx_mode == TX_MODE_SELECT)
read_tx_mode_probs(&fc->tx_probs, &r);
--- a/vp10/encoder/bitstream.c
+++ b/vp10/encoder/bitstream.c
@@ -1326,7 +1326,7 @@
encode_quantization(cm, wb);
encode_segmentation(cm, xd, wb);
#if CONFIG_MISC_FIXES
- if (xd->lossless)
+ if (!cm->seg.enabled && xd->lossless[0])
cm->tx_mode = TX_4X4;
else
write_txfm_mode(cm->tx_mode, wb);
@@ -1356,7 +1356,7 @@
vpx_start_encode(&header_bc, data);
#if !CONFIG_MISC_FIXES
- if (cpi->td.mb.e_mbd.lossless)
+ if (cpi->td.mb.e_mbd.lossless[0])
cm->tx_mode = TX_4X4;
else
update_txfm_probs(cm, &header_bc, counts);
--- a/vp10/encoder/encodeframe.c
+++ b/vp10/encoder/encodeframe.c
@@ -2203,7 +2203,7 @@
// terminated for current branch of the partition search tree.
// The dist & rate thresholds are set to 0 at speed 0 to disable the
// early termination at that speed.
- if (!x->e_mbd.lossless &&
+ if (!x->e_mbd.lossless[xd->mi[0]->mbmi.segment_id] &&
(ctx->skippable && best_rdc.dist < dist_breakout_thr &&
best_rdc.rate < rate_breakout_thr)) {do_split = 0;
@@ -2588,7 +2588,7 @@
}
static TX_MODE select_tx_mode(const VP10_COMP *cpi, MACROBLOCKD *const xd) {- if (xd->lossless)
+ if (!cpi->common.seg.enabled && xd->lossless[0])
return ONLY_4X4;
if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
return ALLOW_32X32;
@@ -2695,6 +2695,7 @@
VP10_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
RD_COUNTS *const rdc = &cpi->td.rd_counts;
+ int i;
xd->mi = cm->mi_grid_visible;
xd->mi[0] = cm->mi;
@@ -2704,12 +2705,21 @@
vp10_zero(rdc->comp_pred_diff);
vp10_zero(rdc->filter_diff);
- xd->lossless = cm->base_qindex == 0 &&
- cm->y_dc_delta_q == 0 &&
- cm->uv_dc_delta_q == 0 &&
- cm->uv_ac_delta_q == 0;
+ for (i = 0; i < (cm->seg.enabled ? MAX_SEGMENTS : 1); ++i) {+#if CONFIG_MISC_FIXES
+ const int qindex = vp10_get_qindex(&cm->seg, i, cm->base_qindex);
+#endif
+ xd->lossless[i] = cm->y_dc_delta_q == 0 &&
+#if CONFIG_MISC_FIXES
+ qindex == 0 &&
+#else
+ cm->base_qindex == 0 &&
+#endif
+ cm->uv_dc_delta_q == 0 &&
+ cm->uv_ac_delta_q == 0;
+ }
- if (xd->lossless)
+ if (!cm->seg.enabled && xd->lossless[0])
x->optimize = 0;
cm->tx_mode = select_tx_mode(cpi, xd);
--- a/vp10/encoder/encodemb.c
+++ b/vp10/encoder/encodemb.c
@@ -367,7 +367,7 @@
scan_order->scan, scan_order->iscan);
break;
case TX_4X4:
- if (xd->lossless) {+ if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {vp10_highbd_fwht4x4(src_diff, coeff, diff_stride);
} else {vpx_highbd_fdct4x4(src_diff, coeff, diff_stride);
@@ -407,7 +407,7 @@
scan_order->scan, scan_order->iscan);
break;
case TX_4X4:
- if (xd->lossless) {+ if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {vp10_fwht4x4(src_diff, coeff, diff_stride);
} else {vpx_fdct4x4(src_diff, coeff, diff_stride);
@@ -461,7 +461,7 @@
pd->dequant[0], eob);
break;
case TX_4X4:
- if (xd->lossless) {+ if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {vp10_highbd_fwht4x4(src_diff, coeff, diff_stride);
} else {vpx_highbd_fdct4x4(src_diff, coeff, diff_stride);
@@ -497,7 +497,7 @@
pd->dequant[0], eob);
break;
case TX_4X4:
- if (xd->lossless) {+ if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {vp10_fwht4x4(src_diff, coeff, diff_stride);
} else {vpx_fdct4x4(src_diff, coeff, diff_stride);
@@ -702,7 +702,7 @@
break;
case TX_4X4:
vp10_highbd_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type,
- xd->lossless);
+ xd->lossless[xd->mi[0]->mbmi.segment_id]);
vpx_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, eob,
@@ -738,7 +738,8 @@
scan_order->scan, scan_order->iscan);
break;
case TX_4X4:
- vp10_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type, xd->lossless);
+ vp10_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type,
+ xd->lossless[xd->mi[0]->mbmi.segment_id]);
vpx_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, eob,
@@ -841,7 +842,7 @@
// case.
vp10_highbd_inv_txfm_add_4x4(dqcoeff, dst, pd->dst.stride,
p->eobs[block], xd->bd, tx_type,
- xd->lossless);
+ xd->lossless[xd->mi[0]->mbmi.segment_id]);
break;
default:
assert(0 && "Invalid transform size");
@@ -870,7 +871,7 @@
// which is significant (not just an optimization) for the lossless
// case.
vp10_inv_txfm_add_4x4(dqcoeff, dst, pd->dst.stride, p->eobs[block],
- tx_type, xd->lossless);
+ tx_type, xd->lossless[xd->mi[0]->mbmi.segment_id]);
break;
default:
assert(0 && "Invalid transform size");
@@ -895,7 +896,7 @@
if (p->eobs[block] > 0) {#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {- if (xd->lossless) {+ if (xd->lossless[0]) {vp10_highbd_iwht4x4_add(dqcoeff, dst, pd->dst.stride,
p->eobs[block], xd->bd);
} else {@@ -905,7 +906,7 @@
return;
}
#endif // CONFIG_VP9_HIGHBITDEPTH
- if (xd->lossless) {+ if (xd->lossless[0]) {vp10_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
} else {vp10_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
@@ -1031,7 +1032,7 @@
vpx_highbd_subtract_block(4, 4, src_diff, diff_stride,
src, src_stride, dst, dst_stride, xd->bd);
vp10_highbd_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type,
- xd->lossless);
+ xd->lossless[mbmi->segment_id]);
vpx_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, eob,
@@ -1043,7 +1044,7 @@
// eob<=1 which is significant (not just an optimization) for the
// lossless case.
vp10_highbd_inv_txfm_add_4x4(dqcoeff, dst, dst_stride, *eob, xd->bd,
- tx_type, xd->lossless);
+ tx_type, xd->lossless[mbmi->segment_id]);
break;
default:
assert(0);
@@ -1100,7 +1101,8 @@
if (!x->skip_recode) {vpx_subtract_block(4, 4, src_diff, diff_stride,
src, src_stride, dst, dst_stride);
- vp10_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type, xd->lossless);
+ vp10_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type,
+ xd->lossless[mbmi->segment_id]);
vpx_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
p->quant_shift, qcoeff, dqcoeff,
pd->dequant, eob, scan_order->scan,
@@ -1112,7 +1114,7 @@
// which is significant (not just an optimization) for the lossless
// case.
vp10_inv_txfm_add_4x4(dqcoeff, dst, dst_stride, *eob, tx_type,
- xd->lossless);
+ xd->lossless[mbmi->segment_id]);
}
break;
default:
--- a/vp10/encoder/encoder.c
+++ b/vp10/encoder/encoder.c
@@ -2663,7 +2663,7 @@
static void loopfilter_frame(VP10_COMP *cpi, VP10_COMMON *cm) {MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
struct loopfilter *lf = &cm->lf;
- if (xd->lossless) {+ if (is_lossless_requested(&cpi->oxcf)) {lf->filter_level = 0;
} else {struct vpx_usec_timer timer;
@@ -4119,7 +4119,7 @@
}
if (oxcf->pass == 1) {- cpi->td.mb.e_mbd.lossless = is_lossless_requested(oxcf);
+ cpi->td.mb.e_mbd.lossless[0] = is_lossless_requested(oxcf);
vp10_first_pass(cpi, source);
} else if (oxcf->pass == 2) {Pass2Encode(cpi, size, dest, frame_flags);
--- a/vp10/encoder/firstpass.c
+++ b/vp10/encoder/firstpass.c
@@ -614,6 +614,7 @@
cm->mi_rows, cm->mi_cols);
// Do intra 16x16 prediction.
+ xd->mi[0]->mbmi.segment_id = 0;
xd->mi[0]->mbmi.mode = DC_PRED;
xd->mi[0]->mbmi.tx_size = use_dc_pred ?
(bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
--- a/vp10/encoder/rdopt.c
+++ b/vp10/encoder/rdopt.c
@@ -533,7 +533,7 @@
rd = VPXMIN(rd1, rd2);
if (plane == 0)
x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
- (rd1 > rd2 && !xd->lossless);
+ (rd1 > rd2 && !xd->lossless[mbmi->segment_id]);
args->this_rate += rate;
args->this_dist += dist;
@@ -605,6 +605,21 @@
mbmi->tx_size, cpi->sf.use_fast_coef_costing);
}
+static void choose_smallest_tx_size(VP10_COMP *cpi, MACROBLOCK *x,
+ int *rate, int64_t *distortion,
+ int *skip, int64_t *sse,
+ int64_t ref_best_rd,
+ BLOCK_SIZE bs) {+ MACROBLOCKD *const xd = &x->e_mbd;
+ MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+ mbmi->tx_size = TX_4X4;
+
+ txfm_rd_in_plane(x, rate, distortion, skip,
+ sse, ref_best_rd, 0, bs,
+ mbmi->tx_size, cpi->sf.use_fast_coef_costing);
+}
+
static void choose_tx_size_from_rd(VP10_COMP *cpi, MACROBLOCK *x,
int *rate,
int64_t *distortion,
@@ -674,7 +689,8 @@
rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
}
- if (is_inter_block(mbmi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {+ if (is_inter_block(mbmi) && !xd->lossless[mbmi->segment_id] &&
+ !s[n] && sse[n] != INT64_MAX) {rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
}
@@ -709,7 +725,11 @@
assert(bs == xd->mi[0]->mbmi.sb_type);
- if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {+ if (CONFIG_MISC_FIXES && xd->lossless[xd->mi[0]->mbmi.segment_id]) {+ choose_smallest_tx_size(cpi, x, rate, distortion, skip, ret_sse,
+ ref_best_rd, bs);
+ } else if (cpi->sf.tx_size_search_method == USE_LARGESTALL ||
+ xd->lossless[xd->mi[0]->mbmi.segment_id]) {choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
bs);
} else {@@ -963,7 +983,7 @@
col + idx, row + idy, 0);
vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
dst, dst_stride, xd->bd);
- if (xd->lossless) {+ if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
const scan_order *so = get_scan(TX_4X4, tx_type);
vp10_highbd_fwd_txfm_4x4(src_diff, coeff, 8, DCT_DCT, 1);
@@ -1062,7 +1082,7 @@
dst, dst_stride, col + idx, row + idy, 0);
vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
- if (xd->lossless) {+ if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
const scan_order *so = get_scan(TX_4X4, tx_type);
vp10_fwd_txfm_4x4(src_diff, coeff, 8, DCT_DCT, 1);
@@ -1497,12 +1517,13 @@
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {- fwd_txm4x4 = xd->lossless ? vp10_highbd_fwht4x4 : vpx_highbd_fdct4x4;
+ fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? vp10_highbd_fwht4x4
+ : vpx_highbd_fdct4x4;
} else {- fwd_txm4x4 = xd->lossless ? vp10_fwht4x4 : vpx_fdct4x4;
+ fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? vp10_fwht4x4 : vpx_fdct4x4;
}
#else
- fwd_txm4x4 = xd->lossless ? vp10_fwht4x4 : vpx_fdct4x4;
+ fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? vp10_fwht4x4 : vpx_fdct4x4;
#endif // CONFIG_VP9_HIGHBITDEPTH
#if CONFIG_VP9_HIGHBITDEPTH
@@ -3460,7 +3481,7 @@
// Cost the skip mb case
rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
- } else if (ref_frame != INTRA_FRAME && !xd->lossless) {+ } else if (ref_frame != INTRA_FRAME && !xd->lossless[mbmi->segment_id]) {if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {// Add in the cost of the no skip flag.
@@ -4215,7 +4236,7 @@
// Skip is never coded at the segment level for sub8x8 blocks and instead
// always coded in the bitstream at the mode info level.
- if (ref_frame != INTRA_FRAME && !xd->lossless) {+ if (ref_frame != INTRA_FRAME && !xd->lossless[mbmi->segment_id]) {if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {// Add in the cost of the no skip flag.