1/*
2 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11
12#include "vp8/common/onyxc_int.h"
13#include "onyx_int.h"
14#include "vp8/common/systemdependent.h"
15#include "quantize.h"
16#include "vp8/common/alloccommon.h"
17#include "mcomp.h"
18#include "firstpass.h"
19#include "vpx_scale/vpx_scale.h"
20#include "vp8/common/extend.h"
21#include "ratectrl.h"
22#include "vp8/common/quant_common.h"
23#include "segmentation.h"
24#include "vpx_mem/vpx_mem.h"
25#include "vp8/common/swapyv12buffer.h"
26#include "vp8/common/threading.h"
27#include "vpx_ports/vpx_timer.h"
28
29#include <math.h>
30#include <limits.h>
31
32#define ALT_REF_MC_ENABLED 1    /* dis/enable MC in AltRef filtering */
33#define ALT_REF_SUBPEL_ENABLED 1 /* dis/enable subpel in MC AltRef filtering */
34
35#if VP8_TEMPORAL_ALT_REF
36
37static void vp8_temporal_filter_predictors_mb_c
38(
39    MACROBLOCKD *x,
40    unsigned char *y_mb_ptr,
41    unsigned char *u_mb_ptr,
42    unsigned char *v_mb_ptr,
43    int stride,
44    int mv_row,
45    int mv_col,
46    unsigned char *pred
47)
48{
49    int offset;
50    unsigned char *yptr, *uptr, *vptr;
51
52    /* Y */
53    yptr = y_mb_ptr + (mv_row >> 3) * stride + (mv_col >> 3);
54
55    if ((mv_row | mv_col) & 7)
56    {
57        x->subpixel_predict16x16(yptr, stride,
58                                    mv_col & 7, mv_row & 7, &pred[0], 16);
59    }
60    else
61    {
62        vp8_copy_mem16x16(yptr, stride, &pred[0], 16);
63    }
64
65    /* U & V */
66    mv_row >>= 1;
67    mv_col >>= 1;
68    stride = (stride + 1) >> 1;
69    offset = (mv_row >> 3) * stride + (mv_col >> 3);
70    uptr = u_mb_ptr + offset;
71    vptr = v_mb_ptr + offset;
72
73    if ((mv_row | mv_col) & 7)
74    {
75        x->subpixel_predict8x8(uptr, stride,
76                            mv_col & 7, mv_row & 7, &pred[256], 8);
77        x->subpixel_predict8x8(vptr, stride,
78                            mv_col & 7, mv_row & 7, &pred[320], 8);
79    }
80    else
81    {
82        vp8_copy_mem8x8(uptr, stride, &pred[256], 8);
83        vp8_copy_mem8x8(vptr, stride, &pred[320], 8);
84    }
85}
86void vp8_temporal_filter_apply_c
87(
88    unsigned char *frame1,
89    unsigned int stride,
90    unsigned char *frame2,
91    unsigned int block_size,
92    int strength,
93    int filter_weight,
94    unsigned int *accumulator,
95    unsigned short *count
96)
97{
98    unsigned int i, j, k;
99    int modifier;
100    int byte = 0;
101
102    for (i = 0,k = 0; i < block_size; i++)
103    {
104        for (j = 0; j < block_size; j++, k++)
105        {
106
107            int src_byte = frame1[byte];
108            int pixel_value = *frame2++;
109
110            modifier   = src_byte - pixel_value;
111            /* This is an integer approximation of:
112             * float coeff = (3.0 * modifer * modifier) / pow(2, strength);
113             * modifier =  (int)roundf(coeff > 16 ? 0 : 16-coeff);
114             */
115            modifier  *= modifier;
116            modifier  *= 3;
117            modifier  += 1 << (strength - 1);
118            modifier >>= strength;
119
120            if (modifier > 16)
121                modifier = 16;
122
123            modifier = 16 - modifier;
124            modifier *= filter_weight;
125
126            count[k] += modifier;
127            accumulator[k] += modifier * pixel_value;
128
129            byte++;
130        }
131
132        byte += stride - block_size;
133    }
134}
135
136#if ALT_REF_MC_ENABLED
137
138static int vp8_temporal_filter_find_matching_mb_c
139(
140    VP8_COMP *cpi,
141    YV12_BUFFER_CONFIG *arf_frame,
142    YV12_BUFFER_CONFIG *frame_ptr,
143    int mb_offset,
144    int error_thresh
145)
146{
147    MACROBLOCK *x = &cpi->mb;
148    int step_param;
149    int sadpb = x->sadperbit16;
150    int bestsme = INT_MAX;
151
152    BLOCK *b = &x->block[0];
153    BLOCKD *d = &x->e_mbd.block[0];
154    int_mv best_ref_mv1;
155    int_mv best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
156
157    /* Save input state */
158    unsigned char **base_src = b->base_src;
159    int src = b->src;
160    int src_stride = b->src_stride;
161    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
162    int pre = d->offset;
163    int pre_stride = x->e_mbd.pre.y_stride;
164
165    best_ref_mv1.as_int = 0;
166    best_ref_mv1_full.as_mv.col = best_ref_mv1.as_mv.col >>3;
167    best_ref_mv1_full.as_mv.row = best_ref_mv1.as_mv.row >>3;
168
169    /* Setup frame pointers */
170    b->base_src = &arf_frame->y_buffer;
171    b->src_stride = arf_frame->y_stride;
172    b->src = mb_offset;
173
174    x->e_mbd.pre.y_buffer = frame_ptr->y_buffer;
175    x->e_mbd.pre.y_stride = frame_ptr->y_stride;
176    d->offset = mb_offset;
177    (void)error_thresh;
178    /* Further step/diamond searches as necessary */
179    if (cpi->Speed < 8)
180    {
181        step_param = cpi->sf.first_step + (cpi->Speed > 5);
182    }
183    else
184    {
185        step_param = cpi->sf.first_step + 2;
186    }
187
188    /* TODO Check that the 16x16 vf & sdf are selected here */
189    /* Ignore mv costing by sending NULL cost arrays */
190    bestsme = vp8_hex_search(x, b, d, &best_ref_mv1_full, &d->bmi.mv,
191                             step_param, sadpb,
192                             &cpi->fn_ptr[BLOCK_16X16],
193                             NULL, NULL, &best_ref_mv1);
194
195#if ALT_REF_SUBPEL_ENABLED
196    /* Try sub-pixel MC? */
197    {
198        int distortion;
199        unsigned int sse;
200        /* Ignore mv costing by sending NULL cost array */
201        bestsme = cpi->find_fractional_mv_step(x, b, d,
202                                               &d->bmi.mv,
203                                               &best_ref_mv1,
204                                               x->errorperbit,
205                                               &cpi->fn_ptr[BLOCK_16X16],
206                                               NULL, &distortion, &sse);
207    }
208#endif
209
210    /* Save input state */
211    b->base_src = base_src;
212    b->src = src;
213    b->src_stride = src_stride;
214    x->e_mbd.pre.y_buffer = base_pre;
215    d->offset = pre;
216    x->e_mbd.pre.y_stride = pre_stride;
217
218    return bestsme;
219}
220#endif
221
222static void vp8_temporal_filter_iterate_c
223(
224    VP8_COMP *cpi,
225    int frame_count,
226    int alt_ref_index,
227    int strength
228)
229{
230    int byte;
231    int frame;
232    int mb_col, mb_row;
233    unsigned int filter_weight;
234    int mb_cols = cpi->common.mb_cols;
235    int mb_rows = cpi->common.mb_rows;
236    int mb_y_offset = 0;
237    int mb_uv_offset = 0;
238    DECLARE_ALIGNED_ARRAY(16, unsigned int, accumulator, 16*16 + 8*8 + 8*8);
239    DECLARE_ALIGNED_ARRAY(16, unsigned short, count, 16*16 + 8*8 + 8*8);
240    MACROBLOCKD *mbd = &cpi->mb.e_mbd;
241    YV12_BUFFER_CONFIG *f = cpi->frames[alt_ref_index];
242    unsigned char *dst1, *dst2;
243    DECLARE_ALIGNED_ARRAY(16, unsigned char,  predictor, 16*16 + 8*8 + 8*8);
244
245    /* Save input state */
246    unsigned char *y_buffer = mbd->pre.y_buffer;
247    unsigned char *u_buffer = mbd->pre.u_buffer;
248    unsigned char *v_buffer = mbd->pre.v_buffer;
249
250    for (mb_row = 0; mb_row < mb_rows; mb_row++)
251    {
252#if ALT_REF_MC_ENABLED
253        /* Source frames are extended to 16 pixels.  This is different than
254         *  L/A/G reference frames that have a border of 32 (VP8BORDERINPIXELS)
255         * A 6 tap filter is used for motion search.  This requires 2 pixels
256         *  before and 3 pixels after.  So the largest Y mv on a border would
257         *  then be 16 - 3.  The UV blocks are half the size of the Y and
258         *  therefore only extended by 8.  The largest mv that a UV block
259         *  can support is 8 - 3.  A UV mv is half of a Y mv.
260         *  (16 - 3) >> 1 == 6 which is greater than 8 - 3.
261         * To keep the mv in play for both Y and UV planes the max that it
262         *  can be on a border is therefore 16 - 5.
263         */
264        cpi->mb.mv_row_min = -((mb_row * 16) + (16 - 5));
265        cpi->mb.mv_row_max = ((cpi->common.mb_rows - 1 - mb_row) * 16)
266                                + (16 - 5);
267#endif
268
269        for (mb_col = 0; mb_col < mb_cols; mb_col++)
270        {
271            int i, j, k;
272            int stride;
273
274            vpx_memset(accumulator, 0, 384*sizeof(unsigned int));
275            vpx_memset(count, 0, 384*sizeof(unsigned short));
276
277#if ALT_REF_MC_ENABLED
278            cpi->mb.mv_col_min = -((mb_col * 16) + (16 - 5));
279            cpi->mb.mv_col_max = ((cpi->common.mb_cols - 1 - mb_col) * 16)
280                                    + (16 - 5);
281#endif
282
283            for (frame = 0; frame < frame_count; frame++)
284            {
285                if (cpi->frames[frame] == NULL)
286                    continue;
287
288                mbd->block[0].bmi.mv.as_mv.row = 0;
289                mbd->block[0].bmi.mv.as_mv.col = 0;
290
291                if (frame == alt_ref_index)
292                {
293                    filter_weight = 2;
294                }
295                else
296                {
297                    int err = 0;
298#if ALT_REF_MC_ENABLED
299#define THRESH_LOW   10000
300#define THRESH_HIGH  20000
301                    /* Find best match in this frame by MC */
302                    err = vp8_temporal_filter_find_matching_mb_c
303                              (cpi,
304                               cpi->frames[alt_ref_index],
305                               cpi->frames[frame],
306                               mb_y_offset,
307                               THRESH_LOW);
308#endif
309                    /* Assign higher weight to matching MB if it's error
310                     * score is lower. If not applying MC default behavior
311                     * is to weight all MBs equal.
312                     */
313                    filter_weight = err<THRESH_LOW
314                                       ? 2 : err<THRESH_HIGH ? 1 : 0;
315                }
316
317                if (filter_weight != 0)
318                {
319                    /* Construct the predictors */
320                    vp8_temporal_filter_predictors_mb_c
321                        (mbd,
322                         cpi->frames[frame]->y_buffer + mb_y_offset,
323                         cpi->frames[frame]->u_buffer + mb_uv_offset,
324                         cpi->frames[frame]->v_buffer + mb_uv_offset,
325                         cpi->frames[frame]->y_stride,
326                         mbd->block[0].bmi.mv.as_mv.row,
327                         mbd->block[0].bmi.mv.as_mv.col,
328                         predictor);
329
330                    /* Apply the filter (YUV) */
331                    vp8_temporal_filter_apply
332                        (f->y_buffer + mb_y_offset,
333                         f->y_stride,
334                         predictor,
335                         16,
336                         strength,
337                         filter_weight,
338                         accumulator,
339                         count);
340
341                    vp8_temporal_filter_apply
342                        (f->u_buffer + mb_uv_offset,
343                         f->uv_stride,
344                         predictor + 256,
345                         8,
346                         strength,
347                         filter_weight,
348                         accumulator + 256,
349                         count + 256);
350
351                    vp8_temporal_filter_apply
352                        (f->v_buffer + mb_uv_offset,
353                         f->uv_stride,
354                         predictor + 320,
355                         8,
356                         strength,
357                         filter_weight,
358                         accumulator + 320,
359                         count + 320);
360                }
361            }
362
363            /* Normalize filter output to produce AltRef frame */
364            dst1 = cpi->alt_ref_buffer.y_buffer;
365            stride = cpi->alt_ref_buffer.y_stride;
366            byte = mb_y_offset;
367            for (i = 0,k = 0; i < 16; i++)
368            {
369                for (j = 0; j < 16; j++, k++)
370                {
371                    unsigned int pval = accumulator[k] + (count[k] >> 1);
372                    pval *= cpi->fixed_divide[count[k]];
373                    pval >>= 19;
374
375                    dst1[byte] = (unsigned char)pval;
376
377                    /* move to next pixel */
378                    byte++;
379                }
380
381                byte += stride - 16;
382            }
383
384            dst1 = cpi->alt_ref_buffer.u_buffer;
385            dst2 = cpi->alt_ref_buffer.v_buffer;
386            stride = cpi->alt_ref_buffer.uv_stride;
387            byte = mb_uv_offset;
388            for (i = 0,k = 256; i < 8; i++)
389            {
390                for (j = 0; j < 8; j++, k++)
391                {
392                    int m=k+64;
393
394                    /* U */
395                    unsigned int pval = accumulator[k] + (count[k] >> 1);
396                    pval *= cpi->fixed_divide[count[k]];
397                    pval >>= 19;
398                    dst1[byte] = (unsigned char)pval;
399
400                    /* V */
401                    pval = accumulator[m] + (count[m] >> 1);
402                    pval *= cpi->fixed_divide[count[m]];
403                    pval >>= 19;
404                    dst2[byte] = (unsigned char)pval;
405
406                    /* move to next pixel */
407                    byte++;
408                }
409
410                byte += stride - 8;
411            }
412
413            mb_y_offset += 16;
414            mb_uv_offset += 8;
415        }
416
417        mb_y_offset += 16*(f->y_stride-mb_cols);
418        mb_uv_offset += 8*(f->uv_stride-mb_cols);
419    }
420
421    /* Restore input state */
422    mbd->pre.y_buffer = y_buffer;
423    mbd->pre.u_buffer = u_buffer;
424    mbd->pre.v_buffer = v_buffer;
425}
426
427void vp8_temporal_filter_prepare_c
428(
429    VP8_COMP *cpi,
430    int distance
431)
432{
433    int frame = 0;
434
435    int num_frames_backward = 0;
436    int num_frames_forward = 0;
437    int frames_to_blur_backward = 0;
438    int frames_to_blur_forward = 0;
439    int frames_to_blur = 0;
440    int start_frame = 0;
441
442    int strength = cpi->oxcf.arnr_strength;
443
444    int blur_type = cpi->oxcf.arnr_type;
445
446    int max_frames = cpi->active_arnr_frames;
447
448    num_frames_backward = distance;
449    num_frames_forward = vp8_lookahead_depth(cpi->lookahead)
450                         - (num_frames_backward + 1);
451
452    switch (blur_type)
453    {
454    case 1:
455        /* Backward Blur */
456
457        frames_to_blur_backward = num_frames_backward;
458
459        if (frames_to_blur_backward >= max_frames)
460            frames_to_blur_backward = max_frames - 1;
461
462        frames_to_blur = frames_to_blur_backward + 1;
463        break;
464
465    case 2:
466        /* Forward Blur */
467
468        frames_to_blur_forward = num_frames_forward;
469
470        if (frames_to_blur_forward >= max_frames)
471            frames_to_blur_forward = max_frames - 1;
472
473        frames_to_blur = frames_to_blur_forward + 1;
474        break;
475
476    case 3:
477    default:
478        /* Center Blur */
479        frames_to_blur_forward = num_frames_forward;
480        frames_to_blur_backward = num_frames_backward;
481
482        if (frames_to_blur_forward > frames_to_blur_backward)
483            frames_to_blur_forward = frames_to_blur_backward;
484
485        if (frames_to_blur_backward > frames_to_blur_forward)
486            frames_to_blur_backward = frames_to_blur_forward;
487
488        /* When max_frames is even we have 1 more frame backward than forward */
489        if (frames_to_blur_forward > (max_frames - 1) / 2)
490            frames_to_blur_forward = ((max_frames - 1) / 2);
491
492        if (frames_to_blur_backward > (max_frames / 2))
493            frames_to_blur_backward = (max_frames / 2);
494
495        frames_to_blur = frames_to_blur_backward + frames_to_blur_forward + 1;
496        break;
497    }
498
499    start_frame = distance + frames_to_blur_forward;
500
501    /* Setup frame pointers, NULL indicates frame not included in filter */
502    vpx_memset(cpi->frames, 0, max_frames*sizeof(YV12_BUFFER_CONFIG *));
503    for (frame = 0; frame < frames_to_blur; frame++)
504    {
505        int which_buffer =  start_frame - frame;
506        struct lookahead_entry* buf = vp8_lookahead_peek(cpi->lookahead,
507                                                         which_buffer,
508                                                         PEEK_FORWARD);
509        cpi->frames[frames_to_blur-1-frame] = &buf->img;
510    }
511
512    vp8_temporal_filter_iterate_c (
513        cpi,
514        frames_to_blur,
515        frames_to_blur_backward,
516        strength );
517}
518#endif
519