Statistics
| Branch: | Revision:

ffmpeg / libavcodec / imgresample.c @ 1c2a8c7f

History | View | Annotate | Download (16.8 KB)

1 de6d9b64 Fabrice Bellard
/*
2
 * High quality image resampling with polyphase filters 
3
 * Copyright (c) 2001 Gerard Lantau.
4
 *
5
 * This program is free software; you can redistribute it and/or modify
6
 * it under the terms of the GNU General Public License as published by
7
 * the Free Software Foundation; either version 2 of the License, or
8
 * (at your option) any later version.
9
 *
10
 * This program is distributed in the hope that it will be useful,
11
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13
 * GNU General Public License for more details.
14
 *
15
 * You should have received a copy of the GNU General Public License
16
 * along with this program; if not, write to the Free Software
17
 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
18
 */
19
#include <stdlib.h>
20
#include <stdio.h>
21
#include <string.h>
22
#include <math.h>
23
#include "dsputil.h"
24
#include "avcodec.h"
25
26 54329dd5 Nick Kurshev
#ifdef USE_FASTMEMCPY
27
#include "fastmemcpy.h"
28
#endif
29
30
31 de6d9b64 Fabrice Bellard
#define NB_COMPONENTS 3
32
33
#define PHASE_BITS 4
34
#define NB_PHASES  (1 << PHASE_BITS)
35
#define NB_TAPS    4
36
#define FCENTER    1  /* index of the center of the filter */
37
38
#define POS_FRAC_BITS 16
39
#define POS_FRAC      (1 << POS_FRAC_BITS)
40
/* 6 bits precision is needed for MMX */
41
#define FILTER_BITS   8
42
43
#define LINE_BUF_HEIGHT (NB_TAPS * 4)
44
45
struct ImgReSampleContext {
46
    int iwidth, iheight, owidth, oheight;
47
    int h_incr, v_incr;
48
    INT16 h_filters[NB_PHASES][NB_TAPS] __align8; /* horizontal filters */
49
    INT16 v_filters[NB_PHASES][NB_TAPS] __align8; /* vertical filters */
50
    UINT8 *line_buf;
51
};
52
53
static inline int get_phase(int pos)
54
{
55
    return ((pos) >> (POS_FRAC_BITS - PHASE_BITS)) & ((1 << PHASE_BITS) - 1);
56
}
57
58
/* This function must be optimized */
59
static void h_resample_fast(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
60
                            int src_start, int src_incr, INT16 *filters)
61
{
62
    int src_pos, phase, sum, i;
63
    UINT8 *s;
64
    INT16 *filter;
65
66
    src_pos = src_start;
67
    for(i=0;i<dst_width;i++) {
68
#ifdef TEST
69
        /* test */
70
        if ((src_pos >> POS_FRAC_BITS) < 0 ||
71
            (src_pos >> POS_FRAC_BITS) > (src_width - NB_TAPS))
72
            abort();
73
#endif
74
        s = src + (src_pos >> POS_FRAC_BITS);
75
        phase = get_phase(src_pos);
76
        filter = filters + phase * NB_TAPS;
77
#if NB_TAPS == 4
78
        sum = s[0] * filter[0] +
79
            s[1] * filter[1] +
80
            s[2] * filter[2] +
81
            s[3] * filter[3];
82
#else
83
        {
84
            int j;
85
            sum = 0;
86
            for(j=0;j<NB_TAPS;j++)
87
                sum += s[j] * filter[j];
88
        }
89
#endif
90
        sum = sum >> FILTER_BITS;
91
        if (sum < 0)
92
            sum = 0;
93
        else if (sum > 255)
94
            sum = 255;
95
        dst[0] = sum;
96
        src_pos += src_incr;
97
        dst++;
98
    }
99
}
100
101
/* This function must be optimized */
102
static void v_resample(UINT8 *dst, int dst_width, UINT8 *src, int wrap, 
103
                       INT16 *filter)
104
{
105
    int sum, i;
106
    UINT8 *s;
107
108
    s = src;
109
    for(i=0;i<dst_width;i++) {
110
#if NB_TAPS == 4
111
        sum = s[0 * wrap] * filter[0] +
112
            s[1 * wrap] * filter[1] +
113
            s[2 * wrap] * filter[2] +
114
            s[3 * wrap] * filter[3];
115
#else
116
        {
117
            int j;
118
            UINT8 *s1 = s;
119
120
            sum = 0;
121
            for(j=0;j<NB_TAPS;j++) {
122
                sum += s1[0] * filter[j];
123
                s1 += wrap;
124
            }
125
        }
126
#endif
127
        sum = sum >> FILTER_BITS;
128
        if (sum < 0)
129
            sum = 0;
130
        else if (sum > 255)
131
            sum = 255;
132
        dst[0] = sum;
133
        dst++;
134
        s++;
135
    }
136
}
137
138 980fc7b8 Fabrice Bellard
#ifdef HAVE_MMX
139 de6d9b64 Fabrice Bellard
140
#include "i386/mmx.h"
141
142
#define FILTER4(reg) \
143
{\
144
        s = src + (src_pos >> POS_FRAC_BITS);\
145
        phase = get_phase(src_pos);\
146
        filter = filters + phase * NB_TAPS;\
147
        movq_m2r(*s, reg);\
148
        punpcklbw_r2r(mm7, reg);\
149
        movq_m2r(*filter, mm6);\
150
        pmaddwd_r2r(reg, mm6);\
151
        movq_r2r(mm6, reg);\
152
        psrlq_i2r(32, reg);\
153
        paddd_r2r(mm6, reg);\
154
        psrad_i2r(FILTER_BITS, reg);\
155
        src_pos += src_incr;\
156
}
157
158
#define DUMP(reg) movq_r2m(reg, tmp); printf(#reg "=%016Lx\n", tmp.uq);
159
160
/* XXX: do four pixels at a time */
161
static void h_resample_fast4_mmx(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
162
                                 int src_start, int src_incr, INT16 *filters)
163
{
164
    int src_pos, phase;
165
    UINT8 *s;
166
    INT16 *filter;
167
    mmx_t tmp;
168
    
169
    src_pos = src_start;
170
    pxor_r2r(mm7, mm7);
171
172
    while (dst_width >= 4) {
173
174
        FILTER4(mm0);
175
        FILTER4(mm1);
176
        FILTER4(mm2);
177
        FILTER4(mm3);
178
179
        packuswb_r2r(mm7, mm0);
180
        packuswb_r2r(mm7, mm1);
181
        packuswb_r2r(mm7, mm3);
182
        packuswb_r2r(mm7, mm2);
183
        movq_r2m(mm0, tmp);
184
        dst[0] = tmp.ub[0];
185
        movq_r2m(mm1, tmp);
186
        dst[1] = tmp.ub[0];
187
        movq_r2m(mm2, tmp);
188
        dst[2] = tmp.ub[0];
189
        movq_r2m(mm3, tmp);
190
        dst[3] = tmp.ub[0];
191
        dst += 4;
192
        dst_width -= 4;
193
    }
194
    while (dst_width > 0) {
195
        FILTER4(mm0);
196
        packuswb_r2r(mm7, mm0);
197
        movq_r2m(mm0, tmp);
198
        dst[0] = tmp.ub[0];
199
        dst++;
200
        dst_width--;
201
    }
202
    emms();
203
}
204
205
static void v_resample4_mmx(UINT8 *dst, int dst_width, UINT8 *src, int wrap, 
206
                            INT16 *filter)
207
{
208
    int sum, i, v;
209
    UINT8 *s;
210
    mmx_t tmp;
211
    mmx_t coefs[4];
212
    
213
    for(i=0;i<4;i++) {
214
        v = filter[i];
215
        coefs[i].uw[0] = v;
216
        coefs[i].uw[1] = v;
217
        coefs[i].uw[2] = v;
218
        coefs[i].uw[3] = v;
219
    }
220
    
221
    pxor_r2r(mm7, mm7);
222
    s = src;
223
    while (dst_width >= 4) {
224
        movq_m2r(s[0 * wrap], mm0);
225
        punpcklbw_r2r(mm7, mm0);
226
        movq_m2r(s[1 * wrap], mm1);
227
        punpcklbw_r2r(mm7, mm1);
228
        movq_m2r(s[2 * wrap], mm2);
229
        punpcklbw_r2r(mm7, mm2);
230
        movq_m2r(s[3 * wrap], mm3);
231
        punpcklbw_r2r(mm7, mm3);
232
233
        pmullw_m2r(coefs[0], mm0);
234
        pmullw_m2r(coefs[1], mm1);
235
        pmullw_m2r(coefs[2], mm2);
236
        pmullw_m2r(coefs[3], mm3);
237
238
        paddw_r2r(mm1, mm0);
239
        paddw_r2r(mm3, mm2);
240
        paddw_r2r(mm2, mm0);
241
        psraw_i2r(FILTER_BITS, mm0);
242
        
243
        packuswb_r2r(mm7, mm0);
244
        movq_r2m(mm0, tmp);
245
246
        *(UINT32 *)dst = tmp.ud[0];
247
        dst += 4;
248
        s += 4;
249
        dst_width -= 4;
250
    }
251
    while (dst_width > 0) {
252
        sum = s[0 * wrap] * filter[0] +
253
            s[1 * wrap] * filter[1] +
254
            s[2 * wrap] * filter[2] +
255
            s[3 * wrap] * filter[3];
256
        sum = sum >> FILTER_BITS;
257
        if (sum < 0)
258
            sum = 0;
259
        else if (sum > 255)
260
            sum = 255;
261
        dst[0] = sum;
262
        dst++;
263
        s++;
264
        dst_width--;
265
    }
266
    emms();
267
}
268
#endif
269
270
/* slow version to handle limit cases. Does not need optimisation */
271
static void h_resample_slow(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
272
                            int src_start, int src_incr, INT16 *filters)
273
{
274
    int src_pos, phase, sum, j, v, i;
275
    UINT8 *s, *src_end;
276
    INT16 *filter;
277
278
    src_end = src + src_width;
279
    src_pos = src_start;
280
    for(i=0;i<dst_width;i++) {
281
        s = src + (src_pos >> POS_FRAC_BITS);
282
        phase = get_phase(src_pos);
283
        filter = filters + phase * NB_TAPS;
284
        sum = 0;
285
        for(j=0;j<NB_TAPS;j++) {
286
            if (s < src)
287
                v = src[0];
288
            else if (s >= src_end)
289
                v = src_end[-1];
290
            else
291
                v = s[0];
292
            sum += v * filter[j];
293
            s++;
294
        }
295
        sum = sum >> FILTER_BITS;
296
        if (sum < 0)
297
            sum = 0;
298
        else if (sum > 255)
299
            sum = 255;
300
        dst[0] = sum;
301
        src_pos += src_incr;
302
        dst++;
303
    }
304
}
305
306
static void h_resample(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
307
                       int src_start, int src_incr, INT16 *filters)
308
{
309
    int n, src_end;
310
311
    if (src_start < 0) {
312
        n = (0 - src_start + src_incr - 1) / src_incr;
313
        h_resample_slow(dst, n, src, src_width, src_start, src_incr, filters);
314
        dst += n;
315
        dst_width -= n;
316
        src_start += n * src_incr;
317
    }
318
    src_end = src_start + dst_width * src_incr;
319
    if (src_end > ((src_width - NB_TAPS) << POS_FRAC_BITS)) {
320
        n = (((src_width - NB_TAPS + 1) << POS_FRAC_BITS) - 1 - src_start) / 
321
            src_incr;
322
    } else {
323
        n = dst_width;
324
    }
325 980fc7b8 Fabrice Bellard
#ifdef HAVE_MMX
326 de6d9b64 Fabrice Bellard
    if ((mm_flags & MM_MMX) && NB_TAPS == 4)
327
        h_resample_fast4_mmx(dst, n, 
328
                             src, src_width, src_start, src_incr, filters);
329
    else
330
#endif
331
        h_resample_fast(dst, n, 
332
                        src, src_width, src_start, src_incr, filters);
333
    if (n < dst_width) {
334
        dst += n;
335
        dst_width -= n;
336
        src_start += n * src_incr;
337
        h_resample_slow(dst, dst_width, 
338
                        src, src_width, src_start, src_incr, filters);
339
    }
340
}
341
342
static void component_resample(ImgReSampleContext *s, 
343
                               UINT8 *output, int owrap, int owidth, int oheight,
344
                               UINT8 *input, int iwrap, int iwidth, int iheight)
345
{
346
    int src_y, src_y1, last_src_y, ring_y, phase_y, y1, y;
347
    UINT8 *new_line, *src_line;
348
349
    last_src_y = - FCENTER - 1;
350
    /* position of the bottom of the filter in the source image */
351
    src_y = (last_src_y + NB_TAPS) * POS_FRAC; 
352
    ring_y = NB_TAPS; /* position in ring buffer */
353
    for(y=0;y<oheight;y++) {
354
        /* apply horizontal filter on new lines from input if needed */
355
        src_y1 = src_y >> POS_FRAC_BITS;
356
        while (last_src_y < src_y1) {
357
            if (++ring_y >= LINE_BUF_HEIGHT + NB_TAPS)
358
                ring_y = NB_TAPS;
359
            last_src_y++;
360
            /* handle limit conditions : replicate line (slighly
361
               inefficient because we filter multiple times */
362
            y1 = last_src_y;
363
            if (y1 < 0) {
364
                y1 = 0;
365
            } else if (y1 >= iheight) {
366
                y1 = iheight - 1;
367
            }
368
            src_line = input + y1 * iwrap;
369
            new_line = s->line_buf + ring_y * owidth;
370
            /* apply filter and handle limit cases correctly */
371
            h_resample(new_line, owidth, 
372
                       src_line, iwidth, - FCENTER * POS_FRAC, s->h_incr, 
373
                       &s->h_filters[0][0]);
374
            /* handle ring buffer wraping */
375
            if (ring_y >= LINE_BUF_HEIGHT) {
376
                memcpy(s->line_buf + (ring_y - LINE_BUF_HEIGHT) * owidth,
377
                       new_line, owidth);
378
            }
379
        }
380
        /* apply vertical filter */
381
        phase_y = get_phase(src_y);
382 980fc7b8 Fabrice Bellard
#ifdef HAVE_MMX
383 de6d9b64 Fabrice Bellard
        /* desactivated MMX because loss of precision */
384
        if ((mm_flags & MM_MMX) && NB_TAPS == 4 && 0)
385
            v_resample4_mmx(output, owidth, 
386
                            s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth, 
387
                            &s->v_filters[phase_y][0]);
388
        else
389
#endif
390
            v_resample(output, owidth, 
391
                       s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth, 
392
                       &s->v_filters[phase_y][0]);
393
            
394
        src_y += s->v_incr;
395
        output += owrap;
396
    }
397
}
398
399
/* XXX: the following filter is quite naive, but it seems to suffice
400
   for 4 taps */
401
static void build_filter(INT16 *filter, float factor)
402
{
403
    int ph, i, v;
404
    float x, y, tab[NB_TAPS], norm, mult;
405
406
    /* if upsampling, only need to interpolate, no filter */
407
    if (factor > 1.0)
408
        factor = 1.0;
409
410
    for(ph=0;ph<NB_PHASES;ph++) {
411
        norm = 0;
412
        for(i=0;i<NB_TAPS;i++) {
413
            
414
            x = M_PI * ((float)(i - FCENTER) - (float)ph / NB_PHASES) * factor;
415
            if (x == 0)
416
                y = 1.0;
417
            else
418
                y = sin(x) / x;
419
            tab[i] = y;
420
            norm += y;
421
        }
422
423
        /* normalize so that an uniform color remains the same */
424
        mult = (float)(1 << FILTER_BITS) / norm;
425
        for(i=0;i<NB_TAPS;i++) {
426
            v = (int)(tab[i] * mult);
427
            filter[ph * NB_TAPS + i] = v;
428
        }
429
    }
430
}
431
432
ImgReSampleContext *img_resample_init(int owidth, int oheight,
433
                                      int iwidth, int iheight)
434
{
435
    ImgReSampleContext *s;
436
437
    s = av_mallocz(sizeof(ImgReSampleContext));
438
    if (!s)
439
        return NULL;
440
    s->line_buf = av_mallocz(owidth * (LINE_BUF_HEIGHT + NB_TAPS));
441
    if (!s->line_buf) 
442
        goto fail;
443
    
444
    s->owidth = owidth;
445
    s->oheight = oheight;
446
    s->iwidth = iwidth;
447
    s->iheight = iheight;
448
    
449
    s->h_incr = (iwidth * POS_FRAC) / owidth;
450
    s->v_incr = (iheight * POS_FRAC) / oheight;
451
    
452
    build_filter(&s->h_filters[0][0], (float)owidth / (float)iwidth);
453
    build_filter(&s->v_filters[0][0], (float)oheight / (float)iheight);
454
455
    return s;
456
 fail:
457
    free(s);
458
    return NULL;
459
}
460
461
void img_resample(ImgReSampleContext *s, 
462
                  AVPicture *output, AVPicture *input)
463
{
464
    int i, shift;
465
466
    for(i=0;i<3;i++) {
467
        shift = (i == 0) ? 0 : 1;
468
        component_resample(s, output->data[i], output->linesize[i], 
469
                           s->owidth >> shift, s->oheight >> shift,
470
                           input->data[i], input->linesize[i], 
471
                           s->iwidth >> shift, s->iheight >> shift);
472
    }
473
}
474
475
void img_resample_close(ImgReSampleContext *s)
476
{
477
    free(s->line_buf);
478
    free(s);
479
}
480
481
#ifdef TEST
482
483
void *av_mallocz(int size)
484
{
485
    void *ptr;
486
    ptr = malloc(size);
487
    memset(ptr, 0, size);
488
    return ptr;
489
}
490
491
/* input */
492
#define XSIZE 256
493
#define YSIZE 256
494
UINT8 img[XSIZE * YSIZE];
495
496
/* output */
497
#define XSIZE1 512
498
#define YSIZE1 512
499
UINT8 img1[XSIZE1 * YSIZE1];
500
UINT8 img2[XSIZE1 * YSIZE1];
501
502
void save_pgm(const char *filename, UINT8 *img, int xsize, int ysize)
503
{
504
    FILE *f;
505
    f=fopen(filename,"w");
506
    fprintf(f,"P5\n%d %d\n%d\n", xsize, ysize, 255);
507
    fwrite(img,1, xsize * ysize,f);
508
    fclose(f);
509
}
510
511
static void dump_filter(INT16 *filter)
512
{
513
    int i, ph;
514
515
    for(ph=0;ph<NB_PHASES;ph++) {
516
        printf("%2d: ", ph);
517
        for(i=0;i<NB_TAPS;i++) {
518
            printf(" %5.2f", filter[ph * NB_TAPS + i] / 256.0);
519
        }
520
        printf("\n");
521
    }
522
}
523
524 980fc7b8 Fabrice Bellard
#ifdef HAVE_MMX
525 de6d9b64 Fabrice Bellard
int mm_flags;
526
#endif
527
528
int main(int argc, char **argv)
529
{
530
    int x, y, v, i, xsize, ysize;
531
    ImgReSampleContext *s;
532
    float fact, factors[] = { 1/2.0, 3.0/4.0, 1.0, 4.0/3.0, 16.0/9.0, 2.0 };
533
    char buf[256];
534
535
    /* build test image */
536
    for(y=0;y<YSIZE;y++) {
537
        for(x=0;x<XSIZE;x++) {
538
            if (x < XSIZE/2 && y < YSIZE/2) {
539
                if (x < XSIZE/4 && y < YSIZE/4) {
540
                    if ((x % 10) <= 6 &&
541
                        (y % 10) <= 6)
542
                        v = 0xff;
543
                    else
544
                        v = 0x00;
545
                } else if (x < XSIZE/4) {
546
                    if (x & 1) 
547
                        v = 0xff;
548
                    else 
549
                        v = 0;
550
                } else if (y < XSIZE/4) {
551
                    if (y & 1) 
552
                        v = 0xff;
553
                    else 
554
                        v = 0;
555
                } else {
556
                    if (y < YSIZE*3/8) {
557
                        if ((y+x) & 1) 
558
                            v = 0xff;
559
                        else 
560
                            v = 0;
561
                    } else {
562
                        if (((x+3) % 4) <= 1 &&
563
                            ((y+3) % 4) <= 1)
564
                            v = 0xff;
565
                        else
566
                            v = 0x00;
567
                    }
568
                }
569
            } else if (x < XSIZE/2) {
570
                v = ((x - (XSIZE/2)) * 255) / (XSIZE/2);
571
            } else if (y < XSIZE/2) {
572
                v = ((y - (XSIZE/2)) * 255) / (XSIZE/2);
573
            } else {
574
                v = ((x + y - XSIZE) * 255) / XSIZE;
575
            }
576
            img[y * XSIZE + x] = v;
577
        }
578
    }
579
    save_pgm("/tmp/in.pgm", img, XSIZE, YSIZE);
580
    for(i=0;i<sizeof(factors)/sizeof(float);i++) {
581
        fact = factors[i];
582
        xsize = (int)(XSIZE * fact);
583
        ysize = (int)(YSIZE * fact);
584
        s = img_resample_init(xsize, ysize, XSIZE, YSIZE);
585
        printf("Factor=%0.2f\n", fact);
586
        dump_filter(&s->h_filters[0][0]);
587
        component_resample(s, img1, xsize, xsize, ysize,
588
                           img, XSIZE, XSIZE, YSIZE);
589
        img_resample_close(s);
590
591
        sprintf(buf, "/tmp/out%d.pgm", i);
592
        save_pgm(buf, img1, xsize, ysize);
593
    }
594
595
    /* mmx test */
596 980fc7b8 Fabrice Bellard
#ifdef HAVE_MMX
597 de6d9b64 Fabrice Bellard
    printf("MMX test\n");
598
    fact = 0.72;
599
    xsize = (int)(XSIZE * fact);
600
    ysize = (int)(YSIZE * fact);
601
    mm_flags = MM_MMX;
602
    s = img_resample_init(xsize, ysize, XSIZE, YSIZE);
603
    component_resample(s, img1, xsize, xsize, ysize,
604
                       img, XSIZE, XSIZE, YSIZE);
605
606
    mm_flags = 0;
607
    s = img_resample_init(xsize, ysize, XSIZE, YSIZE);
608
    component_resample(s, img2, xsize, xsize, ysize,
609
                       img, XSIZE, XSIZE, YSIZE);
610
    if (memcmp(img1, img2, xsize * ysize) != 0) {
611
        fprintf(stderr, "mmx error\n");
612
        exit(1);
613
    }
614
    printf("MMX OK\n");
615
#endif
616
    return 0;
617
}
618
619
#endif