vpx/vp8/decoder/error_concealment.c
Scott LaVarnway ed9c66f584 Remove usage of predict buffer for decode
Instead of using the predict buffer, the decoder now writes
the predictor into the recon buffer.  For blocks with eob=0,
unnecessary idcts can be eliminated.  This gave a performance
boost of ~1.8% for the HD clips used.

Tero: Added needed changes to ARM side and scheduled some
      assembly code to prevent interlocks.

Patch Set 6:  Merged (I1bcdca7a95aacc3a181b9faa6b10e3a71ee24df3)
into this commit because of similarities in the idct
functions.
Patch Set 7: EC bug fix.

Change-Id: Ie31d90b5d3522e1108163f2ac491e455e3f955e6
2011-10-18 12:06:50 -04:00

629 lines
22 KiB
C

/*
* Copyright (c) 2011 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 "error_concealment.h"
#include "onyxd_int.h"
#include "decodemv.h"
#include "vpx_mem/vpx_mem.h"
#include "vp8/common/recon.h"
#include "vp8/common/findnearmv.h"
#include <assert.h>
#define MIN(x,y) (((x)<(y))?(x):(y))
#define MAX(x,y) (((x)>(y))?(x):(y))
#define FLOOR(x,q) ((x) & -(1 << (q)))
#define NUM_NEIGHBORS 20
typedef struct ec_position
{
int row;
int col;
} EC_POS;
/*
* Regenerate the table in Matlab with:
* x = meshgrid((1:4), (1:4));
* y = meshgrid((1:4), (1:4))';
* W = round((1./(sqrt(x.^2 + y.^2))*2^7));
* W(1,1) = 0;
*/
static const int weights_q7[5][5] = {
{ 0, 128, 64, 43, 32 },
{128, 91, 57, 40, 31 },
{ 64, 57, 45, 36, 29 },
{ 43, 40, 36, 30, 26 },
{ 32, 31, 29, 26, 23 }
};
int vp8_alloc_overlap_lists(VP8D_COMP *pbi)
{
if (pbi->overlaps != NULL)
{
vpx_free(pbi->overlaps);
pbi->overlaps = NULL;
}
pbi->overlaps = vpx_calloc(pbi->common.mb_rows * pbi->common.mb_cols,
sizeof(MB_OVERLAP));
if (pbi->overlaps == NULL)
return -1;
vpx_memset(pbi->overlaps, 0,
sizeof(MB_OVERLAP) * pbi->common.mb_rows * pbi->common.mb_cols);
return 0;
}
void vp8_de_alloc_overlap_lists(VP8D_COMP *pbi)
{
vpx_free(pbi->overlaps);
pbi->overlaps = NULL;
}
/* Inserts a new overlap area value to the list of overlaps of a block */
static void assign_overlap(OVERLAP_NODE* overlaps,
union b_mode_info *bmi,
int overlap)
{
int i;
if (overlap <= 0)
return;
/* Find and assign to the next empty overlap node in the list of overlaps.
* Empty is defined as bmi == NULL */
for (i = 0; i < MAX_OVERLAPS; i++)
{
if (overlaps[i].bmi == NULL)
{
overlaps[i].bmi = bmi;
overlaps[i].overlap = overlap;
break;
}
}
}
/* Calculates the overlap area between two 4x4 squares, where the first
* square has its upper-left corner at (b1_row, b1_col) and the second
* square has its upper-left corner at (b2_row, b2_col). Doesn't
* properly handle squares which do not overlap.
*/
static int block_overlap(int b1_row, int b1_col, int b2_row, int b2_col)
{
const int int_top = MAX(b1_row, b2_row); // top
const int int_left = MAX(b1_col, b2_col); // left
/* Since each block is 4x4 pixels, adding 4 (Q3) to the left/top edge
* gives us the right/bottom edge.
*/
const int int_right = MIN(b1_col + (4<<3), b2_col + (4<<3)); // right
const int int_bottom = MIN(b1_row + (4<<3), b2_row + (4<<3)); // bottom
return (int_bottom - int_top) * (int_right - int_left);
}
/* Calculates the overlap area for all blocks in a macroblock at position
* (mb_row, mb_col) in macroblocks, which are being overlapped by a given
* overlapping block at position (new_row, new_col) (in pixels, Q3). The
* first block being overlapped in the macroblock has position (first_blk_row,
* first_blk_col) in blocks relative the upper-left corner of the image.
*/
static void calculate_overlaps_mb(B_OVERLAP *b_overlaps, union b_mode_info *bmi,
int new_row, int new_col,
int mb_row, int mb_col,
int first_blk_row, int first_blk_col)
{
/* Find the blocks within this MB (defined by mb_row, mb_col) which are
* overlapped by bmi and calculate and assign overlap for each of those
* blocks. */
/* Block coordinates relative the upper-left block */
const int rel_ol_blk_row = first_blk_row - mb_row * 4;
const int rel_ol_blk_col = first_blk_col - mb_col * 4;
/* If the block partly overlaps any previous MB, these coordinates
* can be < 0. We don't want to access blocks in previous MBs.
*/
const int blk_idx = MAX(rel_ol_blk_row,0) * 4 + MAX(rel_ol_blk_col,0);
/* Upper left overlapping block */
B_OVERLAP *b_ol_ul = &(b_overlaps[blk_idx]);
/* Calculate and assign overlaps for all blocks in this MB
* which the motion compensated block overlaps
*/
/* Avoid calculating overlaps for blocks in later MBs */
int end_row = MIN(4 + mb_row * 4 - first_blk_row, 2);
int end_col = MIN(4 + mb_col * 4 - first_blk_col, 2);
int row, col;
/* Check if new_row and new_col are evenly divisible by 4 (Q3),
* and if so we shouldn't check neighboring blocks
*/
if (new_row >= 0 && (new_row & 0x1F) == 0)
end_row = 1;
if (new_col >= 0 && (new_col & 0x1F) == 0)
end_col = 1;
/* Check if the overlapping block partly overlaps a previous MB
* and if so, we're overlapping fewer blocks in this MB.
*/
if (new_row < (mb_row*16)<<3)
end_row = 1;
if (new_col < (mb_col*16)<<3)
end_col = 1;
for (row = 0; row < end_row; ++row)
{
for (col = 0; col < end_col; ++col)
{
/* input in Q3, result in Q6 */
const int overlap = block_overlap(new_row, new_col,
(((first_blk_row + row) *
4) << 3),
(((first_blk_col + col) *
4) << 3));
assign_overlap(b_ol_ul[row * 4 + col].overlaps, bmi, overlap);
}
}
}
void vp8_calculate_overlaps(MB_OVERLAP *overlap_ul,
int mb_rows, int mb_cols,
union b_mode_info *bmi,
int b_row, int b_col)
{
MB_OVERLAP *mb_overlap;
int row, col, rel_row, rel_col;
int new_row, new_col;
int end_row, end_col;
int overlap_b_row, overlap_b_col;
int overlap_mb_row, overlap_mb_col;
/* mb subpixel position */
row = (4 * b_row) << 3; /* Q3 */
col = (4 * b_col) << 3; /* Q3 */
/* reverse compensate for motion */
new_row = row - bmi->mv.as_mv.row;
new_col = col - bmi->mv.as_mv.col;
if (new_row >= ((16*mb_rows) << 3) || new_col >= ((16*mb_cols) << 3))
{
/* the new block ended up outside the frame */
return;
}
if (new_row <= (-4 << 3) || new_col <= (-4 << 3))
{
/* outside the frame */
return;
}
/* overlapping block's position in blocks */
overlap_b_row = FLOOR(new_row / 4, 3) >> 3;
overlap_b_col = FLOOR(new_col / 4, 3) >> 3;
/* overlapping block's MB position in MBs
* operations are done in Q3
*/
overlap_mb_row = FLOOR((overlap_b_row << 3) / 4, 3) >> 3;
overlap_mb_col = FLOOR((overlap_b_col << 3) / 4, 3) >> 3;
end_row = MIN(mb_rows - overlap_mb_row, 2);
end_col = MIN(mb_cols - overlap_mb_col, 2);
/* Don't calculate overlap for MBs we don't overlap */
/* Check if the new block row starts at the last block row of the MB */
if (abs(new_row - ((16*overlap_mb_row) << 3)) < ((3*4) << 3))
end_row = 1;
/* Check if the new block col starts at the last block col of the MB */
if (abs(new_col - ((16*overlap_mb_col) << 3)) < ((3*4) << 3))
end_col = 1;
/* find the MB(s) this block is overlapping */
for (rel_row = 0; rel_row < end_row; ++rel_row)
{
for (rel_col = 0; rel_col < end_col; ++rel_col)
{
if (overlap_mb_row + rel_row < 0 ||
overlap_mb_col + rel_col < 0)
continue;
mb_overlap = overlap_ul + (overlap_mb_row + rel_row) * mb_cols +
overlap_mb_col + rel_col;
calculate_overlaps_mb(mb_overlap->overlaps, bmi,
new_row, new_col,
overlap_mb_row + rel_row,
overlap_mb_col + rel_col,
overlap_b_row + rel_row,
overlap_b_col + rel_col);
}
}
}
/* Estimates a motion vector given the overlapping blocks' motion vectors.
* Filters out all overlapping blocks which do not refer to the correct
* reference frame type.
*/
static void estimate_mv(const OVERLAP_NODE *overlaps, union b_mode_info *bmi)
{
int i;
int overlap_sum = 0;
int row_acc = 0;
int col_acc = 0;
bmi->mv.as_int = 0;
for (i=0; i < MAX_OVERLAPS; ++i)
{
if (overlaps[i].bmi == NULL)
break;
col_acc += overlaps[i].overlap * overlaps[i].bmi->mv.as_mv.col;
row_acc += overlaps[i].overlap * overlaps[i].bmi->mv.as_mv.row;
overlap_sum += overlaps[i].overlap;
}
if (overlap_sum > 0)
{
/* Q9 / Q6 = Q3 */
bmi->mv.as_mv.col = col_acc / overlap_sum;
bmi->mv.as_mv.row = row_acc / overlap_sum;
}
else
{
bmi->mv.as_mv.col = 0;
bmi->mv.as_mv.row = 0;
}
}
/* Estimates all motion vectors for a macroblock given the lists of
* overlaps for each block. Decides whether or not the MVs must be clamped.
*/
static void estimate_mb_mvs(const B_OVERLAP *block_overlaps,
MODE_INFO *mi,
int mb_to_left_edge,
int mb_to_right_edge,
int mb_to_top_edge,
int mb_to_bottom_edge)
{
int row, col;
int non_zero_count = 0;
MV * const filtered_mv = &(mi->mbmi.mv.as_mv);
union b_mode_info * const bmi = mi->bmi;
filtered_mv->col = 0;
filtered_mv->row = 0;
mi->mbmi.need_to_clamp_mvs = 0;
for (row = 0; row < 4; ++row)
{
int this_b_to_top_edge = mb_to_top_edge + ((row*4)<<3);
int this_b_to_bottom_edge = mb_to_bottom_edge - ((row*4)<<3);
for (col = 0; col < 4; ++col)
{
int i = row * 4 + col;
int this_b_to_left_edge = mb_to_left_edge + ((col*4)<<3);
int this_b_to_right_edge = mb_to_right_edge - ((col*4)<<3);
/* Estimate vectors for all blocks which are overlapped by this */
/* type. Interpolate/extrapolate the rest of the block's MVs */
estimate_mv(block_overlaps[i].overlaps, &(bmi[i]));
mi->mbmi.need_to_clamp_mvs |= vp8_check_mv_bounds(
&bmi[i].mv,
this_b_to_left_edge,
this_b_to_right_edge,
this_b_to_top_edge,
this_b_to_bottom_edge);
if (bmi[i].mv.as_int != 0)
{
++non_zero_count;
filtered_mv->col += bmi[i].mv.as_mv.col;
filtered_mv->row += bmi[i].mv.as_mv.row;
}
}
}
if (non_zero_count > 0)
{
filtered_mv->col /= non_zero_count;
filtered_mv->row /= non_zero_count;
}
}
static void calc_prev_mb_overlaps(MB_OVERLAP *overlaps, MODE_INFO *prev_mi,
int mb_row, int mb_col,
int mb_rows, int mb_cols)
{
int sub_row;
int sub_col;
for (sub_row = 0; sub_row < 4; ++sub_row)
{
for (sub_col = 0; sub_col < 4; ++sub_col)
{
vp8_calculate_overlaps(
overlaps, mb_rows, mb_cols,
&(prev_mi->bmi[sub_row * 4 + sub_col]),
4 * mb_row + sub_row,
4 * mb_col + sub_col);
}
}
}
/* Estimate all missing motion vectors. This function does the same as the one
* above, but has different input arguments. */
static void estimate_missing_mvs(MB_OVERLAP *overlaps,
MODE_INFO *mi, MODE_INFO *prev_mi,
int mb_rows, int mb_cols,
unsigned int first_corrupt)
{
int mb_row, mb_col;
vpx_memset(overlaps, 0, sizeof(MB_OVERLAP) * mb_rows * mb_cols);
/* First calculate the overlaps for all blocks */
for (mb_row = 0; mb_row < mb_rows; ++mb_row)
{
for (mb_col = 0; mb_col < mb_cols; ++mb_col)
{
/* We're only able to use blocks referring to the last frame
* when extrapolating new vectors.
*/
if (prev_mi->mbmi.ref_frame == LAST_FRAME)
{
calc_prev_mb_overlaps(overlaps, prev_mi,
mb_row, mb_col,
mb_rows, mb_cols);
}
++prev_mi;
}
++prev_mi;
}
mb_row = first_corrupt / mb_cols;
mb_col = first_corrupt - mb_row * mb_cols;
mi += mb_row*(mb_cols + 1) + mb_col;
/* Go through all macroblocks in the current image with missing MVs
* and calculate new MVs using the overlaps.
*/
for (; mb_row < mb_rows; ++mb_row)
{
int mb_to_top_edge = -((mb_row * 16)) << 3;
int mb_to_bottom_edge = ((mb_rows - 1 - mb_row) * 16) << 3;
for (; mb_col < mb_cols; ++mb_col)
{
int mb_to_left_edge = -((mb_col * 16) << 3);
int mb_to_right_edge = ((mb_cols - 1 - mb_col) * 16) << 3;
const B_OVERLAP *block_overlaps =
overlaps[mb_row*mb_cols + mb_col].overlaps;
mi->mbmi.ref_frame = LAST_FRAME;
mi->mbmi.mode = SPLITMV;
mi->mbmi.uv_mode = DC_PRED;
mi->mbmi.partitioning = 3;
mi->mbmi.segment_id = 0;
estimate_mb_mvs(block_overlaps,
mi,
mb_to_left_edge,
mb_to_right_edge,
mb_to_top_edge,
mb_to_bottom_edge);
++mi;
}
mb_col = 0;
++mi;
}
}
void vp8_estimate_missing_mvs(VP8D_COMP *pbi)
{
VP8_COMMON * const pc = &pbi->common;
estimate_missing_mvs(pbi->overlaps,
pc->mi, pc->prev_mi,
pc->mb_rows, pc->mb_cols,
pbi->mvs_corrupt_from_mb);
}
static void assign_neighbor(EC_BLOCK *neighbor, MODE_INFO *mi, int block_idx)
{
assert(mi->mbmi.ref_frame < MAX_REF_FRAMES);
neighbor->ref_frame = mi->mbmi.ref_frame;
neighbor->mv = mi->bmi[block_idx].mv.as_mv;
}
/* Finds the neighboring blocks of a macroblocks. In the general case
* 20 blocks are found. If a fewer number of blocks are found due to
* image boundaries, those positions in the EC_BLOCK array are left "empty".
* The neighbors are enumerated with the upper-left neighbor as the first
* element, the second element refers to the neighbor to right of the previous
* neighbor, and so on. The last element refers to the neighbor below the first
* neighbor.
*/
static void find_neighboring_blocks(MODE_INFO *mi,
EC_BLOCK *neighbors,
int mb_row, int mb_col,
int mb_rows, int mb_cols,
int mi_stride)
{
int i = 0;
int j;
if (mb_row > 0)
{
/* upper left */
if (mb_col > 0)
assign_neighbor(&neighbors[i], mi - mi_stride - 1, 15);
++i;
/* above */
for (j = 12; j < 16; ++j, ++i)
assign_neighbor(&neighbors[i], mi - mi_stride, j);
}
else
i += 5;
if (mb_col < mb_cols - 1)
{
/* upper right */
if (mb_row > 0)
assign_neighbor(&neighbors[i], mi - mi_stride + 1, 12);
++i;
/* right */
for (j = 0; j <= 12; j += 4, ++i)
assign_neighbor(&neighbors[i], mi + 1, j);
}
else
i += 5;
if (mb_row < mb_rows - 1)
{
/* lower right */
if (mb_col < mb_cols - 1)
assign_neighbor(&neighbors[i], mi + mi_stride + 1, 0);
++i;
/* below */
for (j = 0; j < 4; ++j, ++i)
assign_neighbor(&neighbors[i], mi + mi_stride, j);
}
else
i += 5;
if (mb_col > 0)
{
/* lower left */
if (mb_row < mb_rows - 1)
assign_neighbor(&neighbors[i], mi + mi_stride - 1, 4);
++i;
/* left */
for (j = 3; j < 16; j += 4, ++i)
{
assign_neighbor(&neighbors[i], mi - 1, j);
}
}
else
i += 5;
assert(i == 20);
}
/* Calculates which reference frame type is dominating among the neighbors */
static MV_REFERENCE_FRAME dominant_ref_frame(EC_BLOCK *neighbors)
{
/* Default to referring to "skip" */
MV_REFERENCE_FRAME dom_ref_frame = LAST_FRAME;
int max_ref_frame_cnt = 0;
int ref_frame_cnt[MAX_REF_FRAMES] = {0};
int i;
/* Count neighboring reference frames */
for (i = 0; i < NUM_NEIGHBORS; ++i)
{
if (neighbors[i].ref_frame < MAX_REF_FRAMES &&
neighbors[i].ref_frame != INTRA_FRAME)
++ref_frame_cnt[neighbors[i].ref_frame];
}
/* Find maximum */
for (i = 0; i < MAX_REF_FRAMES; ++i)
{
if (ref_frame_cnt[i] > max_ref_frame_cnt)
{
dom_ref_frame = i;
max_ref_frame_cnt = ref_frame_cnt[i];
}
}
return dom_ref_frame;
}
/* Interpolates all motion vectors for a macroblock from the neighboring blocks'
* motion vectors.
*/
static void interpolate_mvs(MACROBLOCKD *mb,
EC_BLOCK *neighbors,
MV_REFERENCE_FRAME dom_ref_frame)
{
int row, col, i;
MODE_INFO * const mi = mb->mode_info_context;
/* Table with the position of the neighboring blocks relative the position
* of the upper left block of the current MB. Starting with the upper left
* neighbor and going to the right.
*/
const EC_POS neigh_pos[NUM_NEIGHBORS] = {
{-1,-1}, {-1,0}, {-1,1}, {-1,2}, {-1,3},
{-1,4}, {0,4}, {1,4}, {2,4}, {3,4},
{4,4}, {4,3}, {4,2}, {4,1}, {4,0},
{4,-1}, {3,-1}, {2,-1}, {1,-1}, {0,-1}
};
mi->mbmi.need_to_clamp_mvs = 0;
for (row = 0; row < 4; ++row)
{
int mb_to_top_edge = mb->mb_to_top_edge + ((row*4)<<3);
int mb_to_bottom_edge = mb->mb_to_bottom_edge - ((row*4)<<3);
for (col = 0; col < 4; ++col)
{
int mb_to_left_edge = mb->mb_to_left_edge + ((col*4)<<3);
int mb_to_right_edge = mb->mb_to_right_edge - ((col*4)<<3);
int w_sum = 0;
int mv_row_sum = 0;
int mv_col_sum = 0;
int_mv * const mv = &(mi->bmi[row*4 + col].mv);
mv->as_int = 0;
for (i = 0; i < NUM_NEIGHBORS; ++i)
{
/* Calculate the weighted sum of neighboring MVs referring
* to the dominant frame type.
*/
const int w = weights_q7[abs(row - neigh_pos[i].row)]
[abs(col - neigh_pos[i].col)];
if (neighbors[i].ref_frame != dom_ref_frame)
continue;
w_sum += w;
/* Q7 * Q3 = Q10 */
mv_row_sum += w*neighbors[i].mv.row;
mv_col_sum += w*neighbors[i].mv.col;
}
if (w_sum > 0)
{
/* Avoid division by zero.
* Normalize with the sum of the coefficients
* Q3 = Q10 / Q7
*/
mv->as_mv.row = mv_row_sum / w_sum;
mv->as_mv.col = mv_col_sum / w_sum;
mi->mbmi.need_to_clamp_mvs |= vp8_check_mv_bounds(
mv,
mb_to_left_edge,
mb_to_right_edge,
mb_to_top_edge,
mb_to_bottom_edge);
}
}
}
}
void vp8_interpolate_motion(MACROBLOCKD *mb,
int mb_row, int mb_col,
int mb_rows, int mb_cols,
int mi_stride)
{
/* Find relevant neighboring blocks */
EC_BLOCK neighbors[NUM_NEIGHBORS];
MV_REFERENCE_FRAME dom_ref_frame;
int i;
/* Initialize the array. MAX_REF_FRAMES is interpreted as "doesn't exist" */
for (i = 0; i < NUM_NEIGHBORS; ++i)
{
neighbors[i].ref_frame = MAX_REF_FRAMES;
neighbors[i].mv.row = neighbors[i].mv.col = 0;
}
find_neighboring_blocks(mb->mode_info_context,
neighbors,
mb_row, mb_col,
mb_rows, mb_cols,
mb->mode_info_stride);
/* Determine the dominant block type */
dom_ref_frame = dominant_ref_frame(neighbors);
/* Interpolate MVs for the missing blocks
* from the dominating MVs */
interpolate_mvs(mb, neighbors, dom_ref_frame);
mb->mode_info_context->mbmi.ref_frame = dom_ref_frame;
mb->mode_info_context->mbmi.mode = SPLITMV;
mb->mode_info_context->mbmi.uv_mode = DC_PRED;
mb->mode_info_context->mbmi.partitioning = 3;
mb->mode_info_context->mbmi.segment_id = 0;
}
void vp8_conceal_corrupt_mb(MACROBLOCKD *xd)
{
/* This macroblock has corrupt residual, use the motion compensated
image (predictor) for concealment */
/* The build predictor functions now output directly into the dst buffer,
* so the copies are no longer necessary */
}