Statistics
| Branch: | Revision:

ffmpeg / libavcodec / dct-test.c @ 9c39071d

History | View | Annotate | Download (14.3 KB)

1
/**
2
 * @file dct-test.c
3
 * DCT test. (c) 2001 Fabrice Bellard.
4
 * Started from sample code by Juan J. Sierralta P.
5
 */
6

    
7
#include <stdlib.h>
8
#include <stdio.h>
9
#include <string.h>
10
#include <sys/time.h>
11
#include <unistd.h>
12

    
13
#include "dsputil.h"
14

    
15
#include "simple_idct.h"
16
#include "faandct.h"
17

    
18
#ifndef MAX
19
#define MAX(a, b)  (((a) > (b)) ? (a) : (b))
20
#endif
21

    
22
#undef printf
23

    
24
void *fast_memcpy(void *a, const void *b, size_t c){return memcpy(a,b,c);};
25

    
26
/* reference fdct/idct */
27
extern void fdct(DCTELEM *block);
28
extern void idct(DCTELEM *block);
29
extern void ff_idct_xvid_mmx(DCTELEM *block);
30
extern void ff_idct_xvid_mmx2(DCTELEM *block);
31
extern void init_fdct();
32

    
33
extern void j_rev_dct(DCTELEM *data);
34
extern void ff_mmx_idct(DCTELEM *data);
35
extern void ff_mmxext_idct(DCTELEM *data);
36

    
37
extern void odivx_idct_c (short *block);
38

    
39
#define AANSCALE_BITS 12
40
static const unsigned short aanscales[64] = {
41
    /* precomputed values scaled up by 14 bits */
42
    16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
43
    22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
44
    21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
45
    19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
46
    16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
47
    12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
48
    8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
49
    4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
50
};
51

    
52
uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
53

    
54
int64_t gettime(void)
55
{
56
    struct timeval tv;
57
    gettimeofday(&tv,NULL);
58
    return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
59
}
60

    
61
#define NB_ITS 20000
62
#define NB_ITS_SPEED 50000
63

    
64
static short idct_mmx_perm[64];
65

    
66
static short idct_simple_mmx_perm[64]={
67
        0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
68
        0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
69
        0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
70
        0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
71
        0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
72
        0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
73
        0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
74
        0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
75
};
76

    
77
void idct_mmx_init(void)
78
{
79
    int i;
80

    
81
    /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
82
    for (i = 0; i < 64; i++) {
83
        idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
84
//        idct_simple_mmx_perm[i] = simple_block_permute_op(i);
85
    }
86
}
87

    
88
static DCTELEM block[64] __attribute__ ((aligned (8)));
89
static DCTELEM block1[64] __attribute__ ((aligned (8)));
90
static DCTELEM block_org[64] __attribute__ ((aligned (8)));
91

    
92
void dct_error(const char *name, int is_idct,
93
               void (*fdct_func)(DCTELEM *block),
94
               void (*fdct_ref)(DCTELEM *block), int test)
95
{
96
    int it, i, scale;
97
    int err_inf, v;
98
    int64_t err2, ti, ti1, it1;
99
    int64_t sysErr[64], sysErrMax=0;
100
    int maxout=0;
101
    int blockSumErrMax=0, blockSumErr;
102

    
103
    srandom(0);
104

    
105
    err_inf = 0;
106
    err2 = 0;
107
    for(i=0; i<64; i++) sysErr[i]=0;
108
    for(it=0;it<NB_ITS;it++) {
109
        for(i=0;i<64;i++)
110
            block1[i] = 0;
111
        switch(test){
112
        case 0:
113
            for(i=0;i<64;i++)
114
                block1[i] = (random() % 512) -256;
115
            if (is_idct){
116
                fdct(block1);
117

    
118
                for(i=0;i<64;i++)
119
                    block1[i]>>=3;
120
            }
121
        break;
122
        case 1:{
123
            int num= (random()%10)+1;
124
            for(i=0;i<num;i++)
125
                block1[random()%64] = (random() % 512) -256;
126
        }break;
127
        case 2:
128
            block1[0]= (random()%4096)-2048;
129
            block1[63]= (block1[0]&1)^1;
130
        break;
131
        }
132

    
133
#if 0 // simulate mismatch control
134
{ int sum=0;
135
        for(i=0;i<64;i++)
136
           sum+=block1[i];
137

138
        if((sum&1)==0) block1[63]^=1;
139
}
140
#endif
141

    
142
        for(i=0; i<64; i++)
143
            block_org[i]= block1[i];
144

    
145
        if (fdct_func == ff_mmx_idct ||
146
            fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
147
            for(i=0;i<64;i++)
148
                block[idct_mmx_perm[i]] = block1[i];
149
        } else if(fdct_func == ff_simple_idct_mmx ) {
150
            for(i=0;i<64;i++)
151
                block[idct_simple_mmx_perm[i]] = block1[i];
152

    
153
        } else {
154
            for(i=0; i<64; i++)
155
                block[i]= block1[i];
156
        }
157
#if 0 // simulate mismatch control for tested IDCT but not the ref
158
{ int sum=0;
159
        for(i=0;i<64;i++)
160
           sum+=block[i];
161

162
        if((sum&1)==0) block[63]^=1;
163
}
164
#endif
165

    
166
        fdct_func(block);
167
        emms(); /* for ff_mmx_idct */
168

    
169
        if (fdct_func == fdct_ifast
170
#ifndef FAAN_POSTSCALE
171
            || fdct_func == ff_faandct
172
#endif
173
            ) {
174
            for(i=0; i<64; i++) {
175
                scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i];
176
                block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS;
177
            }
178
        }
179

    
180
        fdct_ref(block1);
181

    
182
        blockSumErr=0;
183
        for(i=0;i<64;i++) {
184
            v = abs(block[i] - block1[i]);
185
            if (v > err_inf)
186
                err_inf = v;
187
            err2 += v * v;
188
            sysErr[i] += block[i] - block1[i];
189
            blockSumErr += v;
190
            if( abs(block[i])>maxout) maxout=abs(block[i]);
191
        }
192
        if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr;
193
#if 0 // print different matrix pairs
194
        if(blockSumErr){
195
            printf("\n");
196
            for(i=0; i<64; i++){
197
                if((i&7)==0) printf("\n");
198
                printf("%4d ", block_org[i]);
199
            }
200
            for(i=0; i<64; i++){
201
                if((i&7)==0) printf("\n");
202
                printf("%4d ", block[i] - block1[i]);
203
            }
204
        }
205
#endif
206
    }
207
    for(i=0; i<64; i++) sysErrMax= MAX(sysErrMax, ABS(sysErr[i]));
208

    
209
#if 1 // dump systematic errors
210
    for(i=0; i<64; i++){
211
        if(i%8==0) printf("\n");
212
        printf("%5d ", (int)sysErr[i]);
213
    }
214
    printf("\n");
215
#endif
216

    
217
    printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
218
           is_idct ? "IDCT" : "DCT",
219
           name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax);
220
#if 1 //Speed test
221
    /* speed test */
222
    for(i=0;i<64;i++)
223
        block1[i] = 0;
224
    switch(test){
225
    case 0:
226
        for(i=0;i<64;i++)
227
            block1[i] = (random() % 512) -256;
228
        if (is_idct){
229
            fdct(block1);
230

    
231
            for(i=0;i<64;i++)
232
                block1[i]>>=3;
233
        }
234
    break;
235
    case 1:{
236
    case 2:
237
        block1[0] = (random() % 512) -256;
238
        block1[1] = (random() % 512) -256;
239
        block1[2] = (random() % 512) -256;
240
        block1[3] = (random() % 512) -256;
241
    }break;
242
    }
243

    
244
    if (fdct_func == ff_mmx_idct ||
245
        fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
246
        for(i=0;i<64;i++)
247
            block[idct_mmx_perm[i]] = block1[i];
248
    } else if(fdct_func == ff_simple_idct_mmx ) {
249
        for(i=0;i<64;i++)
250
            block[idct_simple_mmx_perm[i]] = block1[i];
251
    } else {
252
        for(i=0; i<64; i++)
253
            block[i]= block1[i];
254
    }
255

    
256
    ti = gettime();
257
    it1 = 0;
258
    do {
259
        for(it=0;it<NB_ITS_SPEED;it++) {
260
            for(i=0; i<64; i++)
261
                block[i]= block1[i];
262
//            memcpy(block, block1, sizeof(DCTELEM) * 64);
263
// dont memcpy especially not fastmemcpy because it does movntq !!!
264
            fdct_func(block);
265
        }
266
        it1 += NB_ITS_SPEED;
267
        ti1 = gettime() - ti;
268
    } while (ti1 < 1000000);
269
    emms();
270

    
271
    printf("%s %s: %0.1f kdct/s\n",
272
           is_idct ? "IDCT" : "DCT",
273
           name, (double)it1 * 1000.0 / (double)ti1);
274
#endif
275
}
276

    
277
static uint8_t img_dest[64] __attribute__ ((aligned (8)));
278
static uint8_t img_dest1[64] __attribute__ ((aligned (8)));
279

    
280
void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
281
{
282
    static int init;
283
    static double c8[8][8];
284
    static double c4[4][4];
285
    double block1[64], block2[64], block3[64];
286
    double s, sum, v;
287
    int i, j, k;
288

    
289
    if (!init) {
290
        init = 1;
291

    
292
        for(i=0;i<8;i++) {
293
            sum = 0;
294
            for(j=0;j<8;j++) {
295
                s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0);
296
                c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
297
                sum += c8[i][j] * c8[i][j];
298
            }
299
        }
300

    
301
        for(i=0;i<4;i++) {
302
            sum = 0;
303
            for(j=0;j<4;j++) {
304
                s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0);
305
                c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
306
                sum += c4[i][j] * c4[i][j];
307
            }
308
        }
309
    }
310

    
311
    /* butterfly */
312
    s = 0.5 * sqrt(2.0);
313
    for(i=0;i<4;i++) {
314
        for(j=0;j<8;j++) {
315
            block1[8*(2*i)+j] = (block[8*(2*i)+j] + block[8*(2*i+1)+j]) * s;
316
            block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s;
317
        }
318
    }
319

    
320
    /* idct8 on lines */
321
    for(i=0;i<8;i++) {
322
        for(j=0;j<8;j++) {
323
            sum = 0;
324
            for(k=0;k<8;k++)
325
                sum += c8[k][j] * block1[8*i+k];
326
            block2[8*i+j] = sum;
327
        }
328
    }
329

    
330
    /* idct4 */
331
    for(i=0;i<8;i++) {
332
        for(j=0;j<4;j++) {
333
            /* top */
334
            sum = 0;
335
            for(k=0;k<4;k++)
336
                sum += c4[k][j] * block2[8*(2*k)+i];
337
            block3[8*(2*j)+i] = sum;
338

    
339
            /* bottom */
340
            sum = 0;
341
            for(k=0;k<4;k++)
342
                sum += c4[k][j] * block2[8*(2*k+1)+i];
343
            block3[8*(2*j+1)+i] = sum;
344
        }
345
    }
346

    
347
    /* clamp and store the result */
348
    for(i=0;i<8;i++) {
349
        for(j=0;j<8;j++) {
350
            v = block3[8*i+j];
351
            if (v < 0)
352
                v = 0;
353
            else if (v > 255)
354
                v = 255;
355
            dest[i * linesize + j] = (int)rint(v);
356
        }
357
    }
358
}
359

    
360
void idct248_error(const char *name,
361
                    void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block))
362
{
363
    int it, i, it1, ti, ti1, err_max, v;
364

    
365
    srandom(0);
366

    
367
    /* just one test to see if code is correct (precision is less
368
       important here) */
369
    err_max = 0;
370
    for(it=0;it<NB_ITS;it++) {
371

    
372
        /* XXX: use forward transform to generate values */
373
        for(i=0;i<64;i++)
374
            block1[i] = (random() % 256) - 128;
375
        block1[0] += 1024;
376

    
377
        for(i=0; i<64; i++)
378
            block[i]= block1[i];
379
        idct248_ref(img_dest1, 8, block);
380

    
381
        for(i=0; i<64; i++)
382
            block[i]= block1[i];
383
        idct248_put(img_dest, 8, block);
384

    
385
        for(i=0;i<64;i++) {
386
            v = abs((int)img_dest[i] - (int)img_dest1[i]);
387
            if (v == 255)
388
                printf("%d %d\n", img_dest[i], img_dest1[i]);
389
            if (v > err_max)
390
                err_max = v;
391
        }
392
#if 0
393
        printf("ref=\n");
394
        for(i=0;i<8;i++) {
395
            int j;
396
            for(j=0;j<8;j++) {
397
                printf(" %3d", img_dest1[i*8+j]);
398
            }
399
            printf("\n");
400
        }
401

402
        printf("out=\n");
403
        for(i=0;i<8;i++) {
404
            int j;
405
            for(j=0;j<8;j++) {
406
                printf(" %3d", img_dest[i*8+j]);
407
            }
408
            printf("\n");
409
        }
410
#endif
411
    }
412
    printf("%s %s: err_inf=%d\n",
413
           1 ? "IDCT248" : "DCT248",
414
           name, err_max);
415

    
416
    ti = gettime();
417
    it1 = 0;
418
    do {
419
        for(it=0;it<NB_ITS_SPEED;it++) {
420
            for(i=0; i<64; i++)
421
                block[i]= block1[i];
422
//            memcpy(block, block1, sizeof(DCTELEM) * 64);
423
// dont memcpy especially not fastmemcpy because it does movntq !!!
424
            idct248_put(img_dest, 8, block);
425
        }
426
        it1 += NB_ITS_SPEED;
427
        ti1 = gettime() - ti;
428
    } while (ti1 < 1000000);
429
    emms();
430

    
431
    printf("%s %s: %0.1f kdct/s\n",
432
           1 ? "IDCT248" : "DCT248",
433
           name, (double)it1 * 1000.0 / (double)ti1);
434
}
435

    
436
void help(void)
437
{
438
    printf("dct-test [-i] [<test-number>]\n"
439
           "test-number 0 -> test with random matrixes\n"
440
           "            1 -> test with random sparse matrixes\n"
441
           "            2 -> do 3. test from mpeg4 std\n"
442
           "-i          test IDCT implementations\n"
443
           "-4          test IDCT248 implementations\n");
444
    exit(1);
445
}
446

    
447
int main(int argc, char **argv)
448
{
449
    int test_idct = 0, test_248_dct = 0;
450
    int c,i;
451
    int test=1;
452

    
453
    init_fdct();
454
    idct_mmx_init();
455

    
456
    for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
457
    for(i=0;i<MAX_NEG_CROP;i++) {
458
        cropTbl[i] = 0;
459
        cropTbl[i + MAX_NEG_CROP + 256] = 255;
460
    }
461

    
462
    for(;;) {
463
        c = getopt(argc, argv, "ih4");
464
        if (c == -1)
465
            break;
466
        switch(c) {
467
        case 'i':
468
            test_idct = 1;
469
            break;
470
        case '4':
471
            test_248_dct = 1;
472
            break;
473
        default :
474
        case 'h':
475
            help();
476
            break;
477
        }
478
    }
479

    
480
    if(optind <argc) test= atoi(argv[optind]);
481

    
482
    printf("ffmpeg DCT/IDCT test\n");
483

    
484
    if (test_248_dct) {
485
        idct248_error("SIMPLE-C", simple_idct248_put);
486
    } else {
487
        if (!test_idct) {
488
            dct_error("REF-DBL", 0, fdct, fdct, test); /* only to verify code ! */
489
            dct_error("IJG-AAN-INT", 0, fdct_ifast, fdct, test);
490
            dct_error("IJG-LLM-INT", 0, ff_jpeg_fdct_islow, fdct, test);
491
            dct_error("MMX", 0, ff_fdct_mmx, fdct, test);
492
            dct_error("MMX2", 0, ff_fdct_mmx2, fdct, test);
493
            dct_error("FAAN", 0, ff_faandct, fdct, test);
494
        } else {
495
            dct_error("REF-DBL", 1, idct, idct, test);
496
            dct_error("INT", 1, j_rev_dct, idct, test);
497
            dct_error("LIBMPEG2-MMX", 1, ff_mmx_idct, idct, test);
498
            dct_error("LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, idct, test);
499
            dct_error("SIMPLE-C", 1, simple_idct, idct, test);
500
            dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, idct, test);
501
            dct_error("XVID-MMX", 1, ff_idct_xvid_mmx, idct, test);
502
            dct_error("XVID-MMX2", 1, ff_idct_xvid_mmx2, idct, test);
503
            //        dct_error("ODIVX-C", 1, odivx_idct_c, idct);
504
            //printf(" test against odivx idct\n");
505
            //        dct_error("REF", 1, idct, odivx_idct_c);
506
            //        dct_error("INT", 1, j_rev_dct, odivx_idct_c);
507
            //        dct_error("MMX", 1, ff_mmx_idct, odivx_idct_c);
508
            //        dct_error("MMXEXT", 1, ff_mmxext_idct, odivx_idct_c);
509
            //        dct_error("SIMPLE-C", 1, simple_idct, odivx_idct_c);
510
            //        dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, odivx_idct_c);
511
            //        dct_error("ODIVX-C", 1, odivx_idct_c, odivx_idct_c);
512
        }
513
    }
514
    return 0;
515
}