generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
57e995ff9c
vpx/vp9/common/vp9_reconintra4x4.c
Ronald S. Bultje 1407bdc243 [WIP] Add column-based tiling. 
This patch adds column-based tiling. The idea is to make each tile
independently decodable (after reading the common frame header) and
also independendly encodable (minus within-frame cost adjustments in
the RD loop) to speed-up hardware & software en/decoders if they used
multi-threading. Column-based tiling has the added advantage (over
other tiling methods) that it minimizes realtime use-case latency,
since all threads can start encoding data as soon as the first SB-row
worth of data is available to the encoder.

There is some test code that does random tile ordering in the decoder,
to confirm that each tile is indeed independently decodable from other
tiles in the same frame. At tile edges, all contexts assume default
values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode),
and motion vector search and ordering do not cross tiles in the same
frame.
t log

Tile independence is not maintained between frames ATM, i.e. tile 0 of
frame 1 is free to use motion vectors that point into any tile of frame
0. We support 1 (i.e. no tiling), 2 or 4 column-tiles.

The loopfilter crosses tile boundaries. I discussed this briefly with Aki
and he says that's OK. An in-loop loopfilter would need to do some sync
between tile threads, but that shouldn't be a big issue.

Resuls: with tiling disabled, we go up slightly because of improved edge
use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf,
~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5%
on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is
concentrated in the low-bitrate end of clips, and most of it is because
of the loss of edges at tile boundaries and the resulting loss of intra
predictors.

TODO:
- more tiles (perhaps allow row-based tiling also, and max. 8 tiles)?
- maybe optionally (for EC purposes), motion vectors themselves
  should not cross tile edges, or we should emulate such borders as
  if they were off-frame, to limit error propagation to within one
  tile only. This doesn't have to be the default behaviour but could
  be an optional bitstream flag.

Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-05 15:43:03 -08:00

494 lines
15 KiB
C
Raw Blame History

/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "./vpx_config.h"
#include "vpx_mem/vpx_mem.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9_rtcd.h"
#if CONFIG_NEWBINTRAMODES
static int find_grad_measure(uint8_t *x, int stride, int n, int t,
int dx, int dy) {
int i, j;
int count = 0, gsum = 0, gdiv;
/* TODO: Make this code more efficient by breaking up into two loops */
for (i = -t; i < n; ++i)
for (j = -t; j < n; ++j) {
int g;
if (i >= 0 && j >= 0) continue;
if (i + dy >= 0 && j + dx >= 0) continue;
if (i + dy < -t || i + dy >= n || j + dx < -t || j + dx >= n) continue;
g = abs(x[(i + dy) * stride + j + dx] - x[i * stride + j]);
gsum += g * g;
count++;
}
gdiv = (dx * dx + dy * dy) * count;
return ((gsum << 8) + (gdiv >> 1)) / gdiv;
}
#if CONTEXT_PRED_REPLACEMENTS == 6
B_PREDICTION_MODE vp9_find_dominant_direction(uint8_t *ptr,
int stride, int n) {
int g[8], i, imin, imax;
g[1] = find_grad_measure(ptr, stride, n, 4, 2, 1);
g[2] = find_grad_measure(ptr, stride, n, 4, 1, 1);
g[3] = find_grad_measure(ptr, stride, n, 4, 1, 2);
g[5] = find_grad_measure(ptr, stride, n, 4, -1, 2);
g[6] = find_grad_measure(ptr, stride, n, 4, -1, 1);
g[7] = find_grad_measure(ptr, stride, n, 4, -2, 1);
imin = 1;
for (i = 2; i < 8; i += 1 + (i == 3))
imin = (g[i] < g[imin] ? i : imin);
imax = 1;
for (i = 2; i < 8; i += 1 + (i == 3))
imax = (g[i] > g[imax] ? i : imax);
/*
printf("%d %d %d %d %d %d = %d %d\n",
g[1], g[2], g[3], g[5], g[6], g[7], imin, imax);
*/
switch (imin) {
case 1:
return B_HD_PRED;
case 2:
return B_RD_PRED;
case 3:
return B_VR_PRED;
case 5:
return B_VL_PRED;
case 6:
return B_LD_PRED;
case 7:
return B_HU_PRED;
default:
assert(0);
}
}
#elif CONTEXT_PRED_REPLACEMENTS == 4
B_PREDICTION_MODE vp9_find_dominant_direction(uint8_t *ptr,
int stride, int n) {
int g[8], i, imin, imax;
g[1] = find_grad_measure(ptr, stride, n, 4, 2, 1);
g[3] = find_grad_measure(ptr, stride, n, 4, 1, 2);
g[5] = find_grad_measure(ptr, stride, n, 4, -1, 2);
g[7] = find_grad_measure(ptr, stride, n, 4, -2, 1);
imin = 1;
for (i = 3; i < 8; i+=2)
imin = (g[i] < g[imin] ? i : imin);
imax = 1;
for (i = 3; i < 8; i+=2)
imax = (g[i] > g[imax] ? i : imax);
/*
printf("%d %d %d %d = %d %d\n",
g[1], g[3], g[5], g[7], imin, imax);
*/
switch (imin) {
case 1:
return B_HD_PRED;
case 3:
return B_VR_PRED;
case 5:
return B_VL_PRED;
case 7:
return B_HU_PRED;
default:
assert(0);
}
}
#elif CONTEXT_PRED_REPLACEMENTS == 0
B_PREDICTION_MODE vp9_find_dominant_direction(uint8_t *ptr,
int stride, int n) {
int g[8], i, imin, imax;
g[0] = find_grad_measure(ptr, stride, n, 4, 1, 0);
g[1] = find_grad_measure(ptr, stride, n, 4, 2, 1);
g[2] = find_grad_measure(ptr, stride, n, 4, 1, 1);
g[3] = find_grad_measure(ptr, stride, n, 4, 1, 2);
g[4] = find_grad_measure(ptr, stride, n, 4, 0, 1);
g[5] = find_grad_measure(ptr, stride, n, 4, -1, 2);
g[6] = find_grad_measure(ptr, stride, n, 4, -1, 1);
g[7] = find_grad_measure(ptr, stride, n, 4, -2, 1);
imax = 0;
for (i = 1; i < 8; i++)
imax = (g[i] > g[imax] ? i : imax);
imin = 0;
for (i = 1; i < 8; i++)
imin = (g[i] < g[imin] ? i : imin);
switch (imin) {
case 0:
return B_HE_PRED;
case 1:
return B_HD_PRED;
case 2:
return B_RD_PRED;
case 3:
return B_VR_PRED;
case 4:
return B_VE_PRED;
case 5:
return B_VL_PRED;
case 6:
return B_LD_PRED;
case 7:
return B_HU_PRED;
default:
assert(0);
}
}
#endif
B_PREDICTION_MODE vp9_find_bpred_context(BLOCKD *x) {
uint8_t *ptr = *(x->base_dst) + x->dst;
int stride = x->dst_stride;
return vp9_find_dominant_direction(ptr, stride, 4);
}
#endif
void vp9_intra4x4_predict(MACROBLOCKD *xd,
BLOCKD *x,
int b_mode,
uint8_t *predictor) {
int i, r, c;
const int block_idx = x - xd->block;
const int have_top = (block_idx >> 2) || xd->up_available;
const int have_left = (block_idx & 3) || xd->left_available;
const int have_right = (block_idx & 3) != 3 || xd->right_available;
uint8_t left[4], above[8], top_left;
/*
* 127 127 127 .. 127 127 127 127 127 127
* 129 A B .. Y Z
* 129 C D .. W X
* 129 E F .. U V
* 129 G H .. S T T T T T
* ..
*/
if (have_left) {
uint8_t *left_ptr = *(x->base_dst) + x->dst - 1;
const int stride = x->dst_stride;
left[0] = left_ptr[0 * stride];
left[1] = left_ptr[1 * stride];
left[2] = left_ptr[2 * stride];
left[3] = left_ptr[3 * stride];
} else {
left[0] = left[1] = left[2] = left[3] = 129;
}
if (have_top) {
uint8_t *above_ptr = *(x->base_dst) + x->dst - x->dst_stride;
if (have_left) {
top_left = above_ptr[-1];
} else {
top_left = 127;
}
above[0] = above_ptr[0];
above[1] = above_ptr[1];
above[2] = above_ptr[2];
above[3] = above_ptr[3];
if (((block_idx & 3) != 3) ||
(have_right && block_idx == 3 &&
((xd->mb_index != 3 && xd->sb_index != 3) ||
((xd->mb_index & 1) == 0 && xd->sb_index == 3)))) {
above[4] = above_ptr[4];
above[5] = above_ptr[5];
above[6] = above_ptr[6];
above[7] = above_ptr[7];
} else if (have_right) {
uint8_t *above_right = above_ptr + 4;
if (xd->sb_index == 3 && (xd->mb_index & 1))
above_right -= 32 * x->dst_stride;
if (xd->mb_index == 3)
above_right -= 16 * x->dst_stride;
above_right -= (block_idx & ~3) * x->dst_stride;
/* use a more distant above-right (from closest available top-right
* corner), but with a "localized DC" (similar'ish to TM-pred):
*
* A B C D E F G H
* I J K L
* M N O P
* Q R S T
* U V W X x1 x2 x3 x4
*
* Where:
* x1 = clip_pixel(E + X - D)
* x2 = clip_pixel(F + X - D)
* x3 = clip_pixel(G + X - D)
* x4 = clip_pixel(H + X - D)
*
* This is applied anytime when we use a "distant" above-right edge
* that is not immediately top-right to the block that we're going
* to do intra prediction for.
*/
above[4] = clip_pixel(above_right[0] + above_ptr[3] - above_right[-1]);
above[5] = clip_pixel(above_right[1] + above_ptr[3] - above_right[-1]);
above[6] = clip_pixel(above_right[2] + above_ptr[3] - above_right[-1]);
above[7] = clip_pixel(above_right[3] + above_ptr[3] - above_right[-1]);
} else {
// extend edge
above[4] = above[5] = above[6] = above[7] = above[3];
}
} else {
above[0] = above[1] = above[2] = above[3] = 127;
above[4] = above[5] = above[6] = above[7] = 127;
top_left = 127;
}
#if CONFIG_NEWBINTRAMODES
if (b_mode == B_CONTEXT_PRED)
b_mode = x->bmi.as_mode.context;
#endif
switch (b_mode) {
case B_DC_PRED: {
int expected_dc = 0;
for (i = 0; i < 4; i++) {
expected_dc += above[i];
expected_dc += left[i];
}
expected_dc = (expected_dc + 4) >> 3;
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
predictor[c] = expected_dc;
}
predictor += 16;
}
}
break;
case B_TM_PRED: {
/* prediction similar to true_motion prediction */
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
predictor[c] = clip_pixel(above[c] - top_left + left[r]);
}
predictor += 16;
}
}
break;
case B_VE_PRED: {
unsigned int ap[4];
ap[0] = above[0];
ap[1] = above[1];
ap[2] = above[2];
ap[3] = above[3];
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
predictor[c] = ap[c];
}
predictor += 16;
}
}
break;
case B_HE_PRED: {
unsigned int lp[4];
lp[0] = left[0];
lp[1] = left[1];
lp[2] = left[2];
lp[3] = left[3];
for (r = 0; r < 4; r++) {
for (c = 0; c < 4; c++) {
predictor[c] = lp[r];
}
predictor += 16;
}
}
break;
case B_LD_PRED: {
uint8_t *ptr = above;
predictor[0 * 16 + 0] = (ptr[0] + ptr[1] * 2 + ptr[2] + 2) >> 2;
predictor[0 * 16 + 1] =
predictor[1 * 16 + 0] = (ptr[1] + ptr[2] * 2 + ptr[3] + 2) >> 2;
predictor[0 * 16 + 2] =
predictor[1 * 16 + 1] =
predictor[2 * 16 + 0] = (ptr[2] + ptr[3] * 2 + ptr[4] + 2) >> 2;
predictor[0 * 16 + 3] =
predictor[1 * 16 + 2] =
predictor[2 * 16 + 1] =
predictor[3 * 16 + 0] = (ptr[3] + ptr[4] * 2 + ptr[5] + 2) >> 2;
predictor[1 * 16 + 3] =
predictor[2 * 16 + 2] =
predictor[3 * 16 + 1] = (ptr[4] + ptr[5] * 2 + ptr[6] + 2) >> 2;
predictor[2 * 16 + 3] =
predictor[3 * 16 + 2] = (ptr[5] + ptr[6] * 2 + ptr[7] + 2) >> 2;
predictor[3 * 16 + 3] = (ptr[6] + ptr[7] * 2 + ptr[7] + 2) >> 2;
}
break;
case B_RD_PRED: {
uint8_t pp[9];
pp[0] = left[3];
pp[1] = left[2];
pp[2] = left[1];
pp[3] = left[0];
pp[4] = top_left;
pp[5] = above[0];
pp[6] = above[1];
pp[7] = above[2];
pp[8] = above[3];
predictor[3 * 16 + 0] = (pp[0] + pp[1] * 2 + pp[2] + 2) >> 2;
predictor[3 * 16 + 1] =
predictor[2 * 16 + 0] = (pp[1] + pp[2] * 2 + pp[3] + 2) >> 2;
predictor[3 * 16 + 2] =
predictor[2 * 16 + 1] =
predictor[1 * 16 + 0] = (pp[2] + pp[3] * 2 + pp[4] + 2) >> 2;
predictor[3 * 16 + 3] =
predictor[2 * 16 + 2] =
predictor[1 * 16 + 1] =
predictor[0 * 16 + 0] = (pp[3] + pp[4] * 2 + pp[5] + 2) >> 2;
predictor[2 * 16 + 3] =
predictor[1 * 16 + 2] =
predictor[0 * 16 + 1] = (pp[4] + pp[5] * 2 + pp[6] + 2) >> 2;
predictor[1 * 16 + 3] =
predictor[0 * 16 + 2] = (pp[5] + pp[6] * 2 + pp[7] + 2) >> 2;
predictor[0 * 16 + 3] = (pp[6] + pp[7] * 2 + pp[8] + 2) >> 2;
}
break;
case B_VR_PRED: {
uint8_t pp[9];
pp[0] = left[3];
pp[1] = left[2];
pp[2] = left[1];
pp[3] = left[0];
pp[4] = top_left;
pp[5] = above[0];
pp[6] = above[1];
pp[7] = above[2];
pp[8] = above[3];
predictor[3 * 16 + 0] = (pp[1] + pp[2] * 2 + pp[3] + 2) >> 2;
predictor[2 * 16 + 0] = (pp[2] + pp[3] * 2 + pp[4] + 2) >> 2;
predictor[3 * 16 + 1] =
predictor[1 * 16 + 0] = (pp[3] + pp[4] * 2 + pp[5] + 2) >> 2;
predictor[2 * 16 + 1] =
predictor[0 * 16 + 0] = (pp[4] + pp[5] + 1) >> 1;
predictor[3 * 16 + 2] =
predictor[1 * 16 + 1] = (pp[4] + pp[5] * 2 + pp[6] + 2) >> 2;
predictor[2 * 16 + 2] =
predictor[0 * 16 + 1] = (pp[5] + pp[6] + 1) >> 1;
predictor[3 * 16 + 3] =
predictor[1 * 16 + 2] = (pp[5] + pp[6] * 2 + pp[7] + 2) >> 2;
predictor[2 * 16 + 3] =
predictor[0 * 16 + 2] = (pp[6] + pp[7] + 1) >> 1;
predictor[1 * 16 + 3] = (pp[6] + pp[7] * 2 + pp[8] + 2) >> 2;
predictor[0 * 16 + 3] = (pp[7] + pp[8] + 1) >> 1;
}
break;
case B_VL_PRED: {
uint8_t *pp = above;
predictor[0 * 16 + 0] = (pp[0] + pp[1] + 1) >> 1;
predictor[1 * 16 + 0] = (pp[0] + pp[1] * 2 + pp[2] + 2) >> 2;
predictor[2 * 16 + 0] =
predictor[0 * 16 + 1] = (pp[1] + pp[2] + 1) >> 1;
predictor[1 * 16 + 1] =
predictor[3 * 16 + 0] = (pp[1] + pp[2] * 2 + pp[3] + 2) >> 2;
predictor[2 * 16 + 1] =
predictor[0 * 16 + 2] = (pp[2] + pp[3] + 1) >> 1;
predictor[3 * 16 + 1] =
predictor[1 * 16 + 2] = (pp[2] + pp[3] * 2 + pp[4] + 2) >> 2;
predictor[0 * 16 + 3] =
predictor[2 * 16 + 2] = (pp[3] + pp[4] + 1) >> 1;
predictor[1 * 16 + 3] =
predictor[3 * 16 + 2] = (pp[3] + pp[4] * 2 + pp[5] + 2) >> 2;
predictor[2 * 16 + 3] = (pp[4] + pp[5] * 2 + pp[6] + 2) >> 2;
predictor[3 * 16 + 3] = (pp[5] + pp[6] * 2 + pp[7] + 2) >> 2;
}
break;
case B_HD_PRED: {
uint8_t pp[9];
pp[0] = left[3];
pp[1] = left[2];
pp[2] = left[1];
pp[3] = left[0];
pp[4] = top_left;
pp[5] = above[0];
pp[6] = above[1];
pp[7] = above[2];
pp[8] = above[3];
predictor[3 * 16 + 0] = (pp[0] + pp[1] + 1) >> 1;
predictor[3 * 16 + 1] = (pp[0] + pp[1] * 2 + pp[2] + 2) >> 2;
predictor[2 * 16 + 0] =
predictor[3 * 16 + 2] = (pp[1] + pp[2] + 1) >> 1;
predictor[2 * 16 + 1] =
predictor[3 * 16 + 3] = (pp[1] + pp[2] * 2 + pp[3] + 2) >> 2;
predictor[2 * 16 + 2] =
predictor[1 * 16 + 0] = (pp[2] + pp[3] + 1) >> 1;
predictor[2 * 16 + 3] =
predictor[1 * 16 + 1] = (pp[2] + pp[3] * 2 + pp[4] + 2) >> 2;
predictor[1 * 16 + 2] =
predictor[0 * 16 + 0] = (pp[3] + pp[4] + 1) >> 1;
predictor[1 * 16 + 3] =
predictor[0 * 16 + 1] = (pp[3] + pp[4] * 2 + pp[5] + 2) >> 2;
predictor[0 * 16 + 2] = (pp[4] + pp[5] * 2 + pp[6] + 2) >> 2;
predictor[0 * 16 + 3] = (pp[5] + pp[6] * 2 + pp[7] + 2) >> 2;
}
break;
case B_HU_PRED: {
uint8_t *pp = left;
predictor[0 * 16 + 0] = (pp[0] + pp[1] + 1) >> 1;
predictor[0 * 16 + 1] = (pp[0] + pp[1] * 2 + pp[2] + 2) >> 2;
predictor[0 * 16 + 2] =
predictor[1 * 16 + 0] = (pp[1] + pp[2] + 1) >> 1;
predictor[0 * 16 + 3] =
predictor[1 * 16 + 1] = (pp[1] + pp[2] * 2 + pp[3] + 2) >> 2;
predictor[1 * 16 + 2] =
predictor[2 * 16 + 0] = (pp[2] + pp[3] + 1) >> 1;
predictor[1 * 16 + 3] =
predictor[2 * 16 + 1] = (pp[2] + pp[3] * 2 + pp[3] + 2) >> 2;
predictor[2 * 16 + 2] =
predictor[2 * 16 + 3] =
predictor[3 * 16 + 0] =
predictor[3 * 16 + 1] =
predictor[3 * 16 + 2] =
predictor[3 * 16 + 3] = pp[3];
}
break;
#if CONFIG_NEWBINTRAMODES
case B_CONTEXT_PRED:
break;
/*
case B_CORNER_PRED:
corner_predictor(predictor, 16, 4, above, left);
break;
*/
#endif
}
}
Reference in New Issue View Git Blame Copy Permalink