quicktime/ffmpeg/libavcodec/dct-test.c

   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 "i386/mmx.h"
  16 #include "simple_idct.h"
  17 #include "faandct.h"
  18
  19 #ifndef MAX
  20 #define MAX(a, b)  (((a) > (b)) ? (a) : (b))
  21 #endif
  22
  23 /* reference fdct/idct */
  24 extern void fdct(DCTELEM *block);
  25 extern void idct(DCTELEM *block);
  26 extern void init_fdct();
  27
  28 extern void j_rev_dct(DCTELEM *data);
  29 extern void ff_mmx_idct(DCTELEM *data);
  30 extern void ff_mmxext_idct(DCTELEM *data);
  31
  32 extern void odivx_idct_c (short *block);
  33
  34 #define AANSCALE_BITS 12
  35 static const unsigned short aanscales[64] = {
  36     /* precomputed values scaled up by 14 bits */
  37     16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
  38     22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
  39     21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
  40     19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
  41     16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
  42     12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
  43     8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
  44     4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
  45 };
  46
  47 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
  48
  49 int64_t gettime(void)
  50 {
  51     struct timeval tv;
  52     gettimeofday(&tv,NULL);
  53     return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
  54 }
  55
  56 #define NB_ITS 20000
  57 #define NB_ITS_SPEED 50000
  58
  59 static short idct_mmx_perm[64];
  60
  61 static short idct_simple_mmx_perm[64]={
  62         0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
  63         0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
  64         0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
  65         0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
  66         0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
  67         0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
  68         0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
  69         0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
  70 };
  71
  72 void idct_mmx_init(void)
  73 {
  74     int i;
  75
  76     /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
  77     for (i = 0; i < 64; i++) {
  78         idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
  79 //      idct_simple_mmx_perm[i] = simple_block_permute_op(i);
  80     }
  81 }
  82
  83 static DCTELEM block[64] __attribute__ ((aligned (8)));
  84 static DCTELEM block1[64] __attribute__ ((aligned (8)));
  85 static DCTELEM block_org[64] __attribute__ ((aligned (8)));
  86
  87 void dct_error(const char *name, int is_idct,
  88                void (*fdct_func)(DCTELEM *block),
  89                void (*fdct_ref)(DCTELEM *block), int test)
  90 {
  91     int it, i, scale;
  92     int err_inf, v;
  93     int64_t err2, ti, ti1, it1;
  94     int64_t sysErr[64], sysErrMax=0;
  95     int maxout=0;
  96     int blockSumErrMax=0, blockSumErr;
  97
  98     srandom(0);
  99
 100     err_inf = 0;
 101     err2 = 0;
 102     for(i=0; i<64; i++) sysErr[i]=0;
 103     for(it=0;it<NB_ITS;it++) {
 104         for(i=0;i<64;i++)
 105             block1[i] = 0;
 106         switch(test){
 107         case 0:
 108             for(i=0;i<64;i++)
 109                 block1[i] = (random() % 512) -256;
 110             if (is_idct){
 111                 fdct(block1);
 112
 113                 for(i=0;i<64;i++)
 114                     block1[i]>>=3;
 115             }
 116         break;
 117         case 1:{
 118             int num= (random()%10)+1;
 119             for(i=0;i<num;i++)
 120                 block1[random()%64] = (random() % 512) -256;
 121         }break;
 122         case 2:
 123             block1[0]= (random()%4096)-2048;
 124             block1[63]= (block1[0]&1)^1;
 125         break;
 126         }
 127
 128 #if 0 // simulate mismatch control
 129 { int sum=0;
 130         for(i=0;i<64;i++)
 131            sum+=block1[i];
 132
 133         if((sum&1)==0) block1[63]^=1;
 134 }
 135 #endif
 136
 137         for(i=0; i<64; i++)
 138             block_org[i]= block1[i];
 139
 140         if (fdct_func == ff_mmx_idct ||
 141             fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
 142             for(i=0;i<64;i++)
 143                 block[idct_mmx_perm[i]] = block1[i];
 144         } else if(fdct_func == ff_simple_idct_mmx ) {
 145             for(i=0;i<64;i++)
 146                 block[idct_simple_mmx_perm[i]] = block1[i];
 147
 148         } else {
 149             for(i=0; i<64; i++)
 150                 block[i]= block1[i];
 151         }
 152 #if 0 // simulate mismatch control for tested IDCT but not the ref
 153 { int sum=0;
 154         for(i=0;i<64;i++)
 155            sum+=block[i];
 156
 157         if((sum&1)==0) block[63]^=1;
 158 }
 159 #endif
 160
 161         fdct_func(block);
 162         emms(); /* for ff_mmx_idct */
 163
 164         if (fdct_func == fdct_ifast
 165 #ifndef FAAN_POSTSCALE
 166             || fdct_func == ff_faandct
 167 #endif
 168             ) {
 169             for(i=0; i<64; i++) {
 170                 scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i];
 171                 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS;
 172             }
 173         }
 174
 175         fdct_ref(block1);
 176
 177         blockSumErr=0;
 178         for(i=0;i<64;i++) {
 179             v = abs(block[i] - block1[i]);
 180             if (v > err_inf)
 181                 err_inf = v;
 182             err2 += v * v;
 183             sysErr[i] += block[i] - block1[i];
 184             blockSumErr += v;
 185             if( abs(block[i])>maxout) maxout=abs(block[i]);
 186         }
 187         if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr;
 188 #if 0 // print different matrix pairs
 189         if(blockSumErr){
 190             printf("\n");
 191             for(i=0; i<64; i++){
 192                 if((i&7)==0) printf("\n");
 193                 printf("%4d ", block_org[i]);
 194             }
 195             for(i=0; i<64; i++){
 196                 if((i&7)==0) printf("\n");
 197                 printf("%4d ", block[i] - block1[i]);
 198             }
 199         }
 200 #endif
 201     }
 202     for(i=0; i<64; i++) sysErrMax= MAX(sysErrMax, ABS(sysErr[i]));
 203
 204 #if 1 // dump systematic errors
 205     for(i=0; i<64; i++){
 206         if(i%8==0) printf("\n");
 207         printf("%5d ", (int)sysErr[i]);
 208     }
 209     printf("\n");
 210 #endif
 211
 212     printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
 213            is_idct ? "IDCT" : "DCT",
 214            name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax);
 215 #if 1 //Speed test
 216     /* speed test */
 217     for(i=0;i<64;i++)
 218         block1[i] = 0;
 219     switch(test){
 220     case 0:
 221         for(i=0;i<64;i++)
 222             block1[i] = (random() % 512) -256;
 223         if (is_idct){
 224             fdct(block1);
 225
 226             for(i=0;i<64;i++)
 227                 block1[i]>>=3;
 228         }
 229     break;
 230     case 1:{
 231     case 2:
 232         block1[0] = (random() % 512) -256;
 233         block1[1] = (random() % 512) -256;
 234         block1[2] = (random() % 512) -256;
 235         block1[3] = (random() % 512) -256;
 236     }break;
 237     }
 238
 239     if (fdct_func == ff_mmx_idct ||
 240         fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
 241         for(i=0;i<64;i++)
 242             block[idct_mmx_perm[i]] = block1[i];
 243     } else if(fdct_func == ff_simple_idct_mmx ) {
 244         for(i=0;i<64;i++)
 245             block[idct_simple_mmx_perm[i]] = block1[i];
 246     } else {
 247         for(i=0; i<64; i++)
 248             block[i]= block1[i];
 249     }
 250
 251     ti = gettime();
 252     it1 = 0;
 253     do {
 254         for(it=0;it<NB_ITS_SPEED;it++) {
 255             for(i=0; i<64; i++)
 256                 block[i]= block1[i];
 257 //            memcpy(block, block1, sizeof(DCTELEM) * 64);
 258 // dont memcpy especially not fastmemcpy because it does movntq !!!
 259             fdct_func(block);
 260         }
 261         it1 += NB_ITS_SPEED;
 262         ti1 = gettime() - ti;
 263     } while (ti1 < 1000000);
 264     emms();
 265
 266     printf("%s %s: %0.1f kdct/s\n",
 267            is_idct ? "IDCT" : "DCT",
 268            name, (double)it1 * 1000.0 / (double)ti1);
 269 #endif
 270 }
 271
 272 static uint8_t img_dest[64] __attribute__ ((aligned (8)));
 273 static uint8_t img_dest1[64] __attribute__ ((aligned (8)));
 274
 275 void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
 276 {
 277     static int init;
 278     static double c8[8][8];
 279     static double c4[4][4];
 280     double block1[64], block2[64], block3[64];
 281     double s, sum, v;
 282     int i, j, k;
 283
 284     if (!init) {
 285         init = 1;
 286
 287         for(i=0;i<8;i++) {
 288             sum = 0;
 289             for(j=0;j<8;j++) {
 290                 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0);
 291                 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
 292                 sum += c8[i][j] * c8[i][j];
 293             }
 294         }
 295
 296         for(i=0;i<4;i++) {
 297             sum = 0;
 298             for(j=0;j<4;j++) {
 299                 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0);
 300                 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
 301                 sum += c4[i][j] * c4[i][j];
 302             }
 303         }
 304     }
 305
 306     /* butterfly */
 307     s = 0.5 * sqrt(2.0);
 308     for(i=0;i<4;i++) {
 309         for(j=0;j<8;j++) {
 310             block1[8*(2*i)+j] = (block[8*(2*i)+j] + block[8*(2*i+1)+j]) * s;
 311             block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s;
 312         }
 313     }
 314
 315     /* idct8 on lines */
 316     for(i=0;i<8;i++) {
 317         for(j=0;j<8;j++) {
 318             sum = 0;
 319             for(k=0;k<8;k++)
 320                 sum += c8[k][j] * block1[8*i+k];
 321             block2[8*i+j] = sum;
 322         }
 323     }
 324
 325     /* idct4 */
 326     for(i=0;i<8;i++) {
 327         for(j=0;j<4;j++) {
 328             /* top */
 329             sum = 0;
 330             for(k=0;k<4;k++)
 331                 sum += c4[k][j] * block2[8*(2*k)+i];
 332             block3[8*(2*j)+i] = sum;
 333
 334             /* bottom */
 335             sum = 0;
 336             for(k=0;k<4;k++)
 337                 sum += c4[k][j] * block2[8*(2*k+1)+i];
 338             block3[8*(2*j+1)+i] = sum;
 339         }
 340     }
 341
 342     /* clamp and store the result */
 343     for(i=0;i<8;i++) {
 344         for(j=0;j<8;j++) {
 345             v = block3[8*i+j];
 346             if (v < 0)
 347                 v = 0;
 348             else if (v > 255)
 349                 v = 255;
 350             dest[i * linesize + j] = (int)rint(v);
 351         }
 352     }
 353 }
 354
 355 void idct248_error(const char *name,
 356                     void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block))
 357 {
 358     int it, i, it1, ti, ti1, err_max, v;
 359
 360     srandom(0);
 361
 362     /* just one test to see if code is correct (precision is less
 363        important here) */
 364     err_max = 0;
 365     for(it=0;it<NB_ITS;it++) {
 366
 367         /* XXX: use forward transform to generate values */
 368         for(i=0;i<64;i++)
 369             block1[i] = (random() % 256) - 128;
 370         block1[0] += 1024;
 371
 372         for(i=0; i<64; i++)
 373             block[i]= block1[i];
 374         idct248_ref(img_dest1, 8, block);
 375
 376         for(i=0; i<64; i++)
 377             block[i]= block1[i];
 378         idct248_put(img_dest, 8, block);
 379
 380         for(i=0;i<64;i++) {
 381             v = abs((int)img_dest[i] - (int)img_dest1[i]);
 382             if (v == 255)
 383                 printf("%d %d\n", img_dest[i], img_dest1[i]);
 384             if (v > err_max)
 385                 err_max = v;
 386         }
 387 #if 0
 388         printf("ref=\n");
 389         for(i=0;i<8;i++) {
 390             int j;
 391             for(j=0;j<8;j++) {
 392                 printf(" %3d", img_dest1[i*8+j]);
 393             }
 394             printf("\n");
 395         }
 396
 397         printf("out=\n");
 398         for(i=0;i<8;i++) {
 399             int j;
 400             for(j=0;j<8;j++) {
 401                 printf(" %3d", img_dest[i*8+j]);
 402             }
 403             printf("\n");
 404         }
 405 #endif
 406     }
 407     printf("%s %s: err_inf=%d\n",
 408            1 ? "IDCT248" : "DCT248",
 409            name, err_max);
 410
 411     ti = gettime();
 412     it1 = 0;
 413     do {
 414         for(it=0;it<NB_ITS_SPEED;it++) {
 415             for(i=0; i<64; i++)
 416                 block[i]= block1[i];
 417 //            memcpy(block, block1, sizeof(DCTELEM) * 64);
 418 // dont memcpy especially not fastmemcpy because it does movntq !!!
 419             idct248_put(img_dest, 8, block);
 420         }
 421         it1 += NB_ITS_SPEED;
 422         ti1 = gettime() - ti;
 423     } while (ti1 < 1000000);
 424     emms();
 425
 426     printf("%s %s: %0.1f kdct/s\n",
 427            1 ? "IDCT248" : "DCT248",
 428            name, (double)it1 * 1000.0 / (double)ti1);
 429 }
 430
 431 void help(void)
 432 {
 433     printf("dct-test [-i] [<test-number>]\n"
 434            "test-number 0 -> test with random matrixes\n"
 435            "            1 -> test with random sparse matrixes\n"
 436            "            2 -> do 3. test from mpeg4 std\n"
 437            "-i          test IDCT implementations\n"
 438            "-4          test IDCT248 implementations\n");
 439     exit(1);
 440 }
 441
 442 int main(int argc, char **argv)
 443 {
 444     int test_idct = 0, test_248_dct = 0;
 445     int c,i;
 446     int test=1;
 447
 448     init_fdct();
 449     idct_mmx_init();
 450
 451     for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
 452     for(i=0;i<MAX_NEG_CROP;i++) {
 453         cropTbl[i] = 0;
 454         cropTbl[i + MAX_NEG_CROP + 256] = 255;
 455     }
 456
 457     for(;;) {
 458         c = getopt(argc, argv, "ih4");
 459         if (c == -1)
 460             break;
 461         switch(c) {
 462         case 'i':
 463             test_idct = 1;
 464             break;
 465         case '4':
 466             test_248_dct = 1;
 467             break;
 468         default :
 469         case 'h':
 470             help();
 471             break;
 472         }
 473     }
 474
 475     if(optind <argc) test= atoi(argv[optind]);
 476
 477     printf("ffmpeg DCT/IDCT test\n");
 478
 479     if (test_248_dct) {
 480         idct248_error("SIMPLE-C", simple_idct248_put);
 481     } else {
 482         if (!test_idct) {
 483             dct_error("REF-DBL", 0, fdct, fdct, test); /* only to verify code ! */
 484             dct_error("IJG-AAN-INT", 0, fdct_ifast, fdct, test);
 485             dct_error("IJG-LLM-INT", 0, ff_jpeg_fdct_islow, fdct, test);
 486             dct_error("MMX", 0, ff_fdct_mmx, fdct, test);
 487             dct_error("MMX2", 0, ff_fdct_mmx2, fdct, test);
 488             dct_error("FAAN", 0, ff_faandct, fdct, test);
 489         } else {
 490             dct_error("REF-DBL", 1, idct, idct, test);
 491             dct_error("INT", 1, j_rev_dct, idct, test);
 492             dct_error("LIBMPEG2-MMX", 1, ff_mmx_idct, idct, test);
 493             dct_error("LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, idct, test);
 494             dct_error("SIMPLE-C", 1, simple_idct, idct, test);
 495             dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, idct, test);
 496             //        dct_error("ODIVX-C", 1, odivx_idct_c, idct);
 497             //printf(" test against odivx idct\n");
 498             //  dct_error("REF", 1, idct, odivx_idct_c);
 499             //        dct_error("INT", 1, j_rev_dct, odivx_idct_c);
 500             //        dct_error("MMX", 1, ff_mmx_idct, odivx_idct_c);
 501             //        dct_error("MMXEXT", 1, ff_mmxext_idct, odivx_idct_c);
 502             //        dct_error("SIMPLE-C", 1, simple_idct, odivx_idct_c);
 503             //        dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, odivx_idct_c);
 504             //        dct_error("ODIVX-C", 1, odivx_idct_c, odivx_idct_c);
 505         }
 506     }
 507     return 0;
 508 }