2 * (c) 2001 Fabrice Bellard
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * DCT test. (c) 2001 Fabrice Bellard.
24 * Started from sample code by Juan J. Sierralta P.
35 #include "simple_idct.h"
39 #define MAX(a, b) (((a) > (b)) ? (a) : (b))
44 void *fast_memcpy(void *a, const void *b, size_t c){return memcpy(a,b,c);};
46 /* reference fdct/idct */
47 extern void fdct(DCTELEM *block);
48 extern void idct(DCTELEM *block);
49 extern void ff_idct_xvid_mmx(DCTELEM *block);
50 extern void ff_idct_xvid_mmx2(DCTELEM *block);
51 extern void init_fdct();
53 extern void j_rev_dct(DCTELEM *data);
54 extern void ff_mmx_idct(DCTELEM *data);
55 extern void ff_mmxext_idct(DCTELEM *data);
57 extern void odivx_idct_c (short *block);
59 #define AANSCALE_BITS 12
60 static const unsigned short aanscales[64] = {
61 /* precomputed values scaled up by 14 bits */
62 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
63 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
64 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
65 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
66 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
67 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
68 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
69 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
72 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
77 gettimeofday(&tv,NULL);
78 return (int64_t)tv.tv_sec * 1000000 + tv.tv_usec;
82 #define NB_ITS_SPEED 50000
84 static short idct_mmx_perm[64];
86 static short idct_simple_mmx_perm[64]={
87 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
88 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
89 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
90 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
91 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
92 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
93 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
94 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
97 void idct_mmx_init(void)
101 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
102 for (i = 0; i < 64; i++) {
103 idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
104 // idct_simple_mmx_perm[i] = simple_block_permute_op(i);
108 static DCTELEM block[64] __attribute__ ((aligned (8)));
109 static DCTELEM block1[64] __attribute__ ((aligned (8)));
110 static DCTELEM block_org[64] __attribute__ ((aligned (8)));
112 void dct_error(const char *name, int is_idct,
113 void (*fdct_func)(DCTELEM *block),
114 void (*fdct_ref)(DCTELEM *block), int test)
118 int64_t err2, ti, ti1, it1;
119 int64_t sysErr[64], sysErrMax=0;
121 int blockSumErrMax=0, blockSumErr;
127 for(i=0; i<64; i++) sysErr[i]=0;
128 for(it=0;it<NB_ITS;it++) {
134 block1[i] = (random() % 512) -256;
143 int num= (random()%10)+1;
145 block1[random()%64] = (random() % 512) -256;
148 block1[0]= (random()%4096)-2048;
149 block1[63]= (block1[0]&1)^1;
153 #if 0 // simulate mismatch control
158 if((sum&1)==0) block1[63]^=1;
163 block_org[i]= block1[i];
165 if (fdct_func == ff_mmx_idct ||
166 fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
168 block[idct_mmx_perm[i]] = block1[i];
169 } else if(fdct_func == ff_simple_idct_mmx ) {
171 block[idct_simple_mmx_perm[i]] = block1[i];
177 #if 0 // simulate mismatch control for tested IDCT but not the ref
182 if((sum&1)==0) block[63]^=1;
187 emms(); /* for ff_mmx_idct */
189 if (fdct_func == fdct_ifast
190 #ifndef FAAN_POSTSCALE
191 || fdct_func == ff_faandct
194 for(i=0; i<64; i++) {
195 scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i];
196 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS;
204 v = abs(block[i] - block1[i]);
208 sysErr[i] += block[i] - block1[i];
210 if( abs(block[i])>maxout) maxout=abs(block[i]);
212 if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr;
213 #if 0 // print different matrix pairs
217 if((i&7)==0) printf("\n");
218 printf("%4d ", block_org[i]);
221 if((i&7)==0) printf("\n");
222 printf("%4d ", block[i] - block1[i]);
227 for(i=0; i<64; i++) sysErrMax= MAX(sysErrMax, FFABS(sysErr[i]));
229 #if 1 // dump systematic errors
231 if(i%8==0) printf("\n");
232 printf("%5d ", (int)sysErr[i]);
237 printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
238 is_idct ? "IDCT" : "DCT",
239 name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax);
247 block1[i] = (random() % 512) -256;
257 block1[0] = (random() % 512) -256;
258 block1[1] = (random() % 512) -256;
259 block1[2] = (random() % 512) -256;
260 block1[3] = (random() % 512) -256;
264 if (fdct_func == ff_mmx_idct ||
265 fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
267 block[idct_mmx_perm[i]] = block1[i];
268 } else if(fdct_func == ff_simple_idct_mmx ) {
270 block[idct_simple_mmx_perm[i]] = block1[i];
279 for(it=0;it<NB_ITS_SPEED;it++) {
282 // memcpy(block, block1, sizeof(DCTELEM) * 64);
283 // dont memcpy especially not fastmemcpy because it does movntq !!!
287 ti1 = gettime() - ti;
288 } while (ti1 < 1000000);
291 printf("%s %s: %0.1f kdct/s\n",
292 is_idct ? "IDCT" : "DCT",
293 name, (double)it1 * 1000.0 / (double)ti1);
297 static uint8_t img_dest[64] __attribute__ ((aligned (8)));
298 static uint8_t img_dest1[64] __attribute__ ((aligned (8)));
300 void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
303 static double c8[8][8];
304 static double c4[4][4];
305 double block1[64], block2[64], block3[64];
315 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0);
316 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
317 sum += c8[i][j] * c8[i][j];
324 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0);
325 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
326 sum += c4[i][j] * c4[i][j];
335 block1[8*(2*i)+j] = (block[8*(2*i)+j] + block[8*(2*i+1)+j]) * s;
336 block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * s;
345 sum += c8[k][j] * block1[8*i+k];
356 sum += c4[k][j] * block2[8*(2*k)+i];
357 block3[8*(2*j)+i] = sum;
362 sum += c4[k][j] * block2[8*(2*k+1)+i];
363 block3[8*(2*j+1)+i] = sum;
367 /* clamp and store the result */
375 dest[i * linesize + j] = (int)rint(v);
380 void idct248_error(const char *name,
381 void (*idct248_put)(uint8_t *dest, int line_size, int16_t *block))
383 int it, i, it1, ti, ti1, err_max, v;
387 /* just one test to see if code is correct (precision is less
390 for(it=0;it<NB_ITS;it++) {
392 /* XXX: use forward transform to generate values */
394 block1[i] = (random() % 256) - 128;
399 idct248_ref(img_dest1, 8, block);
403 idct248_put(img_dest, 8, block);
406 v = abs((int)img_dest[i] - (int)img_dest1[i]);
408 printf("%d %d\n", img_dest[i], img_dest1[i]);
417 printf(" %3d", img_dest1[i*8+j]);
426 printf(" %3d", img_dest[i*8+j]);
432 printf("%s %s: err_inf=%d\n",
433 1 ? "IDCT248" : "DCT248",
439 for(it=0;it<NB_ITS_SPEED;it++) {
442 // memcpy(block, block1, sizeof(DCTELEM) * 64);
443 // dont memcpy especially not fastmemcpy because it does movntq !!!
444 idct248_put(img_dest, 8, block);
447 ti1 = gettime() - ti;
448 } while (ti1 < 1000000);
451 printf("%s %s: %0.1f kdct/s\n",
452 1 ? "IDCT248" : "DCT248",
453 name, (double)it1 * 1000.0 / (double)ti1);
458 printf("dct-test [-i] [<test-number>]\n"
459 "test-number 0 -> test with random matrixes\n"
460 " 1 -> test with random sparse matrixes\n"
461 " 2 -> do 3. test from mpeg4 std\n"
462 "-i test IDCT implementations\n"
463 "-4 test IDCT248 implementations\n");
467 int main(int argc, char **argv)
469 int test_idct = 0, test_248_dct = 0;
476 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
477 for(i=0;i<MAX_NEG_CROP;i++) {
479 cropTbl[i + MAX_NEG_CROP + 256] = 255;
483 c = getopt(argc, argv, "ih4");
500 if(optind <argc) test= atoi(argv[optind]);
502 printf("ffmpeg DCT/IDCT test\n");
505 idct248_error("SIMPLE-C", simple_idct248_put);
508 dct_error("REF-DBL", 0, fdct, fdct, test); /* only to verify code ! */
509 dct_error("IJG-AAN-INT", 0, fdct_ifast, fdct, test);
510 dct_error("IJG-LLM-INT", 0, ff_jpeg_fdct_islow, fdct, test);
511 dct_error("MMX", 0, ff_fdct_mmx, fdct, test);
512 dct_error("MMX2", 0, ff_fdct_mmx2, fdct, test);
513 dct_error("FAAN", 0, ff_faandct, fdct, test);
515 dct_error("REF-DBL", 1, idct, idct, test);
516 dct_error("INT", 1, j_rev_dct, idct, test);
517 dct_error("LIBMPEG2-MMX", 1, ff_mmx_idct, idct, test);
518 dct_error("LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, idct, test);
519 dct_error("SIMPLE-C", 1, simple_idct, idct, test);
520 dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, idct, test);
521 dct_error("XVID-MMX", 1, ff_idct_xvid_mmx, idct, test);
522 dct_error("XVID-MMX2", 1, ff_idct_xvid_mmx2, idct, test);
523 // dct_error("ODIVX-C", 1, odivx_idct_c, idct);
524 //printf(" test against odivx idct\n");
525 // dct_error("REF", 1, idct, odivx_idct_c);
526 // dct_error("INT", 1, j_rev_dct, odivx_idct_c);
527 // dct_error("MMX", 1, ff_mmx_idct, odivx_idct_c);
528 // dct_error("MMXEXT", 1, ff_mmxext_idct, odivx_idct_c);
529 // dct_error("SIMPLE-C", 1, simple_idct, odivx_idct_c);
530 // dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, odivx_idct_c);
531 // dct_error("ODIVX-C", 1, odivx_idct_c, odivx_idct_c);