ref: f00d157c125a5362b48037a2c65d2f0eee4f5f84
parent: 3c34619125ea8a5ea1657d06d7c4d970d0a30599
author: Dmitry Kovalev <dkovalev@google.com>
date: Tue Dec 3 12:59:32 EST 2013
Moving eob array to the encoder. In the decoder we don't need to save eobs, we can pass eob as an argument. That's why removing eob arrays from VP9Decompressor and TileWorkerData, and moving eob pointer from macroblockd_plane to macroblock_plane. Change-Id: I8eb919acc837acfb3abdd8319af63d1bbca8217a
--- a/vp9/common/vp9_blockd.h
+++ b/vp9/common/vp9_blockd.h
@@ -199,7 +199,6 @@
struct macroblockd_plane {int16_t *dqcoeff;
- uint16_t *eobs;
PLANE_TYPE plane_type;
int subsampling_x;
int subsampling_y;
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -42,7 +42,6 @@
vp9_reader bit_reader;
DECLARE_ALIGNED(16, MACROBLOCKD, xd);
DECLARE_ALIGNED(16, int16_t, dqcoeff[MAX_MB_PLANE][64 * 64]);
- DECLARE_ALIGNED(16, uint16_t, eobs[MAX_MB_PLANE][256]);
} TileWorkerData;
static int read_be32(const uint8_t *p) {@@ -238,9 +237,9 @@
}
static void inverse_transform_block(MACROBLOCKD* xd, int plane, int block,
- TX_SIZE tx_size, uint8_t *dst, int stride) {+ TX_SIZE tx_size, uint8_t *dst, int stride,
+ int eob) {struct macroblockd_plane *const pd = &xd->plane[plane];
- const int eob = pd->eobs[block];
if (eob > 0) {TX_TYPE tx_type;
const int plane_type = pd->plane_type;
@@ -313,9 +312,11 @@
dst, pd->dst.stride, dst, pd->dst.stride);
if (!mi->mbmi.skip_coeff) {- vp9_decode_block_tokens(cm, xd, plane, block, plane_bsize, x, y, tx_size,
- args->r);
- inverse_transform_block(xd, plane, block, tx_size, dst, pd->dst.stride);
+ const int eob = vp9_decode_block_tokens(cm, xd, plane, block,
+ plane_bsize, x, y, tx_size,
+ args->r);
+ inverse_transform_block(xd, plane, block, tx_size, dst, pd->dst.stride,
+ eob);
}
}
@@ -333,14 +334,14 @@
VP9_COMMON *const cm = args->cm;
MACROBLOCKD *const xd = args->xd;
struct macroblockd_plane *const pd = &xd->plane[plane];
- int x, y;
+ int x, y, eob;
txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
- *args->eobtotal += vp9_decode_block_tokens(cm, xd, plane, block,
- plane_bsize, x, y, tx_size,
- args->r);
+ eob = vp9_decode_block_tokens(cm, xd, plane, block, plane_bsize, x, y,
+ tx_size, args->r);
inverse_transform_block(xd, plane, block, tx_size,
&pd->dst.buf[4 * y * pd->dst.stride + 4 * x],
- pd->dst.stride);
+ pd->dst.stride, eob);
+ *args->eobtotal += eob;
}
static void set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
@@ -925,7 +926,6 @@
for (i = 0; i < MAX_MB_PLANE; ++i) {pd[i].dqcoeff = tile_data->dqcoeff[i];
- pd[i].eobs = tile_data->eobs[i];
vpx_memset(xd->plane[i].dqcoeff, 0, 64 * 64 * sizeof(int16_t));
}
}
--- a/vp9/decoder/vp9_detokenize.c
+++ b/vp9/decoder/vp9_detokenize.c
@@ -213,7 +213,6 @@
BLOCK_OFFSET(pd->dqcoeff, block), tx_size,
pd->dequant, pt);
set_contexts(xd, pd, plane_bsize, tx_size, eob > 0, x, y);
- pd->eobs[block] = eob;
return eob;
}
--- a/vp9/decoder/vp9_onyxd_if.c
+++ b/vp9/decoder/vp9_onyxd_if.c
@@ -112,10 +112,8 @@
struct macroblockd_plane *const pd = xd->plane;
int i;
- for (i = 0; i < MAX_MB_PLANE; ++i) {+ for (i = 0; i < MAX_MB_PLANE; ++i)
pd[i].dqcoeff = pbi->dqcoeff[i];
- pd[i].eobs = pbi->eobs[i];
- }
}
VP9D_PTR vp9_create_decompressor(VP9D_CONFIG *oxcf) {--- a/vp9/decoder/vp9_onyxd_int.h
+++ b/vp9/decoder/vp9_onyxd_int.h
@@ -23,7 +23,6 @@
DECLARE_ALIGNED(16, VP9_COMMON, common);
DECLARE_ALIGNED(16, int16_t, dqcoeff[MAX_MB_PLANE][64 * 64]);
- DECLARE_ALIGNED(16, uint16_t, eobs[MAX_MB_PLANE][256]);
VP9D_CONFIG oxcf;
--- a/vp9/encoder/vp9_block.h
+++ b/vp9/encoder/vp9_block.h
@@ -71,6 +71,7 @@
DECLARE_ALIGNED(16, int16_t, src_diff[64 * 64]);
int16_t *qcoeff;
int16_t *coeff;
+ uint16_t *eobs;
struct buf_2d src;
// Quantizer setings
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -441,7 +441,7 @@
p[i].coeff = ctx->coeff_pbuf[i][1];
p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
- pd[i].eobs = ctx->eobs_pbuf[i][1];
+ p[i].eobs = ctx->eobs_pbuf[i][1];
}
for (i = max_plane; i < MAX_MB_PLANE; ++i) {@@ -448,7 +448,7 @@
p[i].coeff = ctx->coeff_pbuf[i][2];
p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
- pd[i].eobs = ctx->eobs_pbuf[i][2];
+ p[i].eobs = ctx->eobs_pbuf[i][2];
}
// Restore the coding context of the MB to that that was in place
@@ -677,7 +677,7 @@
p[i].coeff = ctx->coeff_pbuf[i][0];
p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
- pd[i].eobs = ctx->eobs_pbuf[i][0];
+ p[i].eobs = ctx->eobs_pbuf[i][0];
}
ctx->is_coded = 0;
x->skip_recode = 0;
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -144,7 +144,7 @@
const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
int16_t *qcoeff_ptr;
int16_t *dqcoeff_ptr;
- int eob = pd->eobs[block], final_eob, sz = 0;
+ int eob = p->eobs[block], final_eob, sz = 0;
const int i0 = 0;
int rc, x, next, i;
int64_t rdmult, rddiv, rd_cost0, rd_cost1;
@@ -334,7 +334,7 @@
}
final_eob++;
- xd->plane[plane].eobs[block] = final_eob;
+ mb->plane[plane].eobs[block] = final_eob;
*a = *l = (final_eob > 0);
}
@@ -372,7 +372,7 @@
int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
const scan_order *so;
- uint16_t *eob = &pd->eobs[block];
+ uint16_t *eob = &p->eobs[block];
const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
int i, j;
int16_t *src_diff;
@@ -423,6 +423,7 @@
MACROBLOCK *const x = args->x;
MACROBLOCKD *const xd = &x->e_mbd;
struct optimize_ctx *const ctx = args->ctx;
+ struct macroblock_plane *const p = &x->plane[plane];
struct macroblockd_plane *const pd = &xd->plane[plane];
int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
int i, j;
@@ -433,7 +434,7 @@
// TODO(jingning): per transformed block zero forcing only enabled for
// luma component. will integrate chroma components as well.
if (x->zcoeff_blk[tx_size][block] && plane == 0) {- pd->eobs[block] = 0;
+ p->eobs[block] = 0;
ctx->ta[plane][i] = 0;
ctx->tl[plane][j] = 0;
return;
@@ -445,28 +446,28 @@
if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
} else {- ctx->ta[plane][i] = pd->eobs[block] > 0;
- ctx->tl[plane][j] = pd->eobs[block] > 0;
+ ctx->ta[plane][i] = p->eobs[block] > 0;
+ ctx->tl[plane][j] = p->eobs[block] > 0;
}
- if (x->skip_encode || pd->eobs[block] == 0)
+ if (x->skip_encode || p->eobs[block] == 0)
return;
switch (tx_size) {case TX_32X32:
- vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
+ vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
break;
case TX_16X16:
- vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
+ vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
break;
case TX_8X8:
- vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
+ vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
break;
case TX_4X4:
// this is like vp9_short_idct4x4 but has a special case around eob<=1
// which is significant (not just an optimization) for the lossless
// case.
- xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
+ xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
break;
default:
assert(!"Invalid transform size");
@@ -478,6 +479,7 @@
struct encode_b_args *const args = arg;
MACROBLOCK *const x = args->x;
MACROBLOCKD *const xd = &x->e_mbd;
+ struct macroblock_plane *const p = &x->plane[plane];
struct macroblockd_plane *const pd = &xd->plane[plane];
int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
int i, j;
@@ -487,10 +489,10 @@
vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
- if (pd->eobs[block] == 0)
+ if (p->eobs[block] == 0)
return;
- xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
+ xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
}
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {@@ -540,7 +542,7 @@
const int diff_stride = 4 * (1 << bwl);
uint8_t *src, *dst;
int16_t *src_diff;
- uint16_t *eob = &pd->eobs[block];
+ uint16_t *eob = &p->eobs[block];
int i, j;
txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
dst = &pd->dst.buf[4 * (j * pd->dst.stride + i)];
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -545,7 +545,7 @@
p[i].coeff = ctx->coeff_pbuf[i][1];
p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
- pd[i].eobs = ctx->eobs_pbuf[i][1];
+ p[i].eobs = ctx->eobs_pbuf[i][1];
}
x->skip_recode = 0;
--- a/vp9/encoder/vp9_quantize.c
+++ b/vp9/encoder/vp9_quantize.c
@@ -148,7 +148,7 @@
p->zbin, p->round, p->quant, p->quant_shift,
BLOCK_OFFSET(p->qcoeff, block),
BLOCK_OFFSET(pd->dqcoeff, block),
- pd->dequant, p->zbin_extra, &pd->eobs[block], scan, iscan);
+ pd->dequant, p->zbin_extra, &p->eobs[block], scan, iscan);
}
static void invert_quant(int16_t *quant, int16_t *shift, int d) {--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -525,7 +525,7 @@
struct macroblockd_plane *pd = &xd->plane[plane];
const PLANE_TYPE type = pd->plane_type;
const int16_t *band_count = &band_counts[tx_size][1];
- const int eob = pd->eobs[block];
+ const int eob = p->eobs[block];
const int16_t *const qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
const int ref = mbmi->ref_frame[0] != INTRA_FRAME;
unsigned int (*token_costs)[2][PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS] =
@@ -643,7 +643,7 @@
// TODO(jingning): temporarily enabled only for luma component
rd = MIN(rd1, rd2);
if (plane == 0)
- x->zcoeff_blk[tx_size][block] = !xd->plane[plane].eobs[block] ||
+ x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
(rd1 > rd2 && !xd->lossless);
args->this_rate += args->rate;
@@ -739,7 +739,7 @@
*distortion = rd_stack->this_dist;
*rate = rd_stack->this_rate;
*sse = rd_stack->this_sse;
- *skippable = vp9_is_skippable_in_plane(xd, bsize, plane);
+ *skippable = vp9_is_skippable_in_plane(x, bsize, plane);
}
}
@@ -1329,7 +1329,7 @@
p[i].coeff = ctx->coeff_pbuf[i][2];
p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
- pd[i].eobs = ctx->eobs_pbuf[i][2];
+ p[i].eobs = ctx->eobs_pbuf[i][2];
ctx->coeff_pbuf[i][2] = ctx->coeff_pbuf[i][0];
ctx->qcoeff_pbuf[i][2] = ctx->qcoeff_pbuf[i][0];
@@ -1339,7 +1339,7 @@
ctx->coeff_pbuf[i][0] = p[i].coeff;
ctx->qcoeff_pbuf[i][0] = p[i].qcoeff;
ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
- ctx->eobs_pbuf[i][0] = pd[i].eobs;
+ ctx->eobs_pbuf[i][0] = p[i].eobs;
}
}
}
@@ -1630,6 +1630,7 @@
MB_PREDICTION_MODE this_mode;
MODE_INFO *mi = x->e_mbd.mi_8x8[0];
MB_MODE_INFO *const mbmi = &mi->mbmi;
+ struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
const int label_count = 4;
int64_t this_segment_rd = 0;
@@ -1958,11 +1959,11 @@
bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
bsi->rdstat[i][mode_idx].brate, 0);
bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
- bsi->rdstat[i][mode_idx].eobs = pd->eobs[i];
+ bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
if (num_4x4_blocks_wide > 1)
- bsi->rdstat[i + 1][mode_idx].eobs = pd->eobs[i + 1];
+ bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
if (num_4x4_blocks_high > 1)
- bsi->rdstat[i + 2][mode_idx].eobs = pd->eobs[i + 2];
+ bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
}
if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {@@ -2060,7 +2061,7 @@
mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
if (has_second_ref(mbmi))
mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
- xd->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
+ x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
mi->bmi[i].as_mode = bsi->modes[i];
}
@@ -2070,7 +2071,7 @@
*returntotrate = bsi->r;
*returndistortion = bsi->d;
*returnyrate = bsi->segment_yrate;
- *skippable = vp9_is_skippable_in_plane(&x->e_mbd, BLOCK_8X8, 0);
+ *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
*psse = bsi->sse;
mbmi->mode = bsi->modes[3];
@@ -3005,7 +3006,7 @@
p[i].coeff = ctx->coeff_pbuf[i][1];
p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
- pd[i].eobs = ctx->eobs_pbuf[i][1];
+ p[i].eobs = ctx->eobs_pbuf[i][1];
ctx->coeff_pbuf[i][1] = ctx->coeff_pbuf[i][0];
ctx->qcoeff_pbuf[i][1] = ctx->qcoeff_pbuf[i][0];
@@ -3015,7 +3016,7 @@
ctx->coeff_pbuf[i][0] = p[i].coeff;
ctx->qcoeff_pbuf[i][0] = p[i].qcoeff;
ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
- ctx->eobs_pbuf[i][0] = pd[i].eobs;
+ ctx->eobs_pbuf[i][0] = p[i].eobs;
}
}
@@ -4133,7 +4134,7 @@
tmp_best_mbmode = *mbmi;
for (i = 0; i < 4; i++) {tmp_best_bmodes[i] = xd->mi_8x8[0]->bmi[i];
- x->zcoeff_blk[TX_4X4][i] = !xd->plane[0].eobs[i];
+ x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
}
pred_exists = 1;
if (switchable_filter_index == 0 &&
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -168,10 +168,11 @@
TX_SIZE tx_size, void *arg) {struct tokenize_b_args* const args = arg;
MACROBLOCKD *const xd = args->xd;
+ struct macroblock_plane *p = &args->cpi->mb.plane[plane];
struct macroblockd_plane *pd = &xd->plane[plane];
int aoff, loff;
txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
- set_contexts(xd, pd, plane_bsize, tx_size, pd->eobs[block] > 0, aoff, loff);
+ set_contexts(xd, pd, plane_bsize, tx_size, p->eobs[block] > 0, aoff, loff);
}
static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
@@ -181,16 +182,15 @@
MACROBLOCKD *xd = args->xd;
TOKENEXTRA **tp = args->tp;
uint8_t *token_cache = args->token_cache;
+ struct macroblock_plane *p = &cpi->mb.plane[plane];
struct macroblockd_plane *pd = &xd->plane[plane];
MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
int pt; /* near block/prev token context index */
int c = 0, rc = 0;
TOKENEXTRA *t = *tp; /* store tokens starting here */
- const int eob = pd->eobs[block];
+ const int eob = p->eobs[block];
const PLANE_TYPE type = pd->plane_type;
- struct macroblock_plane *p = &cpi->mb.plane[plane];
const int16_t *qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
-
const int segment_id = mbmi->segment_id;
const int16_t *scan, *nb;
const scan_order *so;
@@ -249,7 +249,7 @@
}
struct is_skippable_args {- MACROBLOCKD *xd;
+ MACROBLOCK *x;
int *skippable;
};
@@ -257,21 +257,21 @@
BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
void *argv) {struct is_skippable_args *args = argv;
- args->skippable[0] &= (!args->xd->plane[plane].eobs[block]);
+ args->skippable[0] &= (!args->x->plane[plane].eobs[block]);
}
-int vp9_sb_is_skippable(MACROBLOCKD *xd, BLOCK_SIZE bsize) {+static int sb_is_skippable(MACROBLOCK *x, BLOCK_SIZE bsize) {int result = 1;
- struct is_skippable_args args = {xd, &result};- foreach_transformed_block(xd, bsize, is_skippable, &args);
+ struct is_skippable_args args = {x, &result};+ foreach_transformed_block(&x->e_mbd, bsize, is_skippable, &args);
return result;
}
-int vp9_is_skippable_in_plane(MACROBLOCKD *xd, BLOCK_SIZE bsize,
- int plane) {+int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {int result = 1;
- struct is_skippable_args args = {xd, &result};- foreach_transformed_block_in_plane(xd, bsize, plane, is_skippable, &args);
+ struct is_skippable_args args = {x, &result};+ foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane, is_skippable,
+ &args);
return result;
}
@@ -286,7 +286,7 @@
SEG_LVL_SKIP);
struct tokenize_b_args arg = {cpi, xd, t, mbmi->tx_size, cpi->mb.token_cache};- mbmi->skip_coeff = vp9_sb_is_skippable(xd, bsize);
+ mbmi->skip_coeff = sb_is_skippable(&cpi->mb, bsize);
if (mbmi->skip_coeff) {if (!dry_run)
cm->counts.mbskip[mb_skip_context][1] += skip_inc;
--- a/vp9/encoder/vp9_tokenize.h
+++ b/vp9/encoder/vp9_tokenize.h
@@ -32,9 +32,8 @@
extern const vp9_tree_index vp9_coef_con_tree[];
extern struct vp9_token vp9_coef_encodings[];
-int vp9_sb_is_skippable(MACROBLOCKD *xd, BLOCK_SIZE bsize);
-int vp9_is_skippable_in_plane(MACROBLOCKD *xd, BLOCK_SIZE bsize,
- int plane);
+int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
struct VP9_COMP;
void vp9_tokenize_sb(struct VP9_COMP *cpi, TOKENEXTRA **t, int dry_run,