You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

583 lines
16KB

  1. /*
  2. * (c) 2001 Fabrice Bellard
  3. * 2007 Marc Hoffman <marc.hoffman@analog.com>
  4. *
  5. * This file is part of FFmpeg.
  6. *
  7. * FFmpeg is free software; you can redistribute it and/or
  8. * modify it under the terms of the GNU Lesser General Public
  9. * License as published by the Free Software Foundation; either
  10. * version 2.1 of the License, or (at your option) any later version.
  11. *
  12. * FFmpeg is distributed in the hope that it will be useful,
  13. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15. * Lesser General Public License for more details.
  16. *
  17. * You should have received a copy of the GNU Lesser General Public
  18. * License along with FFmpeg; if not, write to the Free Software
  19. * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20. */
  21. /**
  22. * @file
  23. * DCT test (c) 2001 Fabrice Bellard
  24. * Started from sample code by Juan J. Sierralta P.
  25. */
  26. #include "config.h"
  27. #include <stdlib.h>
  28. #include <stdio.h>
  29. #include <string.h>
  30. #if HAVE_UNISTD_H
  31. #include <unistd.h>
  32. #endif
  33. #include <math.h>
  34. #include "libavutil/cpu.h"
  35. #include "libavutil/common.h"
  36. #include "libavutil/lfg.h"
  37. #include "libavutil/time.h"
  38. #include "simple_idct.h"
  39. #include "aandcttab.h"
  40. #include "faandct.h"
  41. #include "faanidct.h"
  42. #include "x86/idct_xvid.h"
  43. #include "dctref.h"
  44. #undef printf
  45. void ff_mmx_idct(DCTELEM *data);
  46. void ff_mmxext_idct(DCTELEM *data);
  47. // BFIN
  48. void ff_bfin_idct(DCTELEM *block);
  49. void ff_bfin_fdct(DCTELEM *block);
  50. // ALTIVEC
  51. void ff_fdct_altivec(DCTELEM *block);
  52. // ARM
  53. void ff_j_rev_dct_arm(DCTELEM *data);
  54. void ff_simple_idct_arm(DCTELEM *data);
  55. void ff_simple_idct_armv5te(DCTELEM *data);
  56. void ff_simple_idct_armv6(DCTELEM *data);
  57. void ff_simple_idct_neon(DCTELEM *data);
  58. void ff_simple_idct_axp(DCTELEM *data);
  59. struct algo {
  60. const char *name;
  61. void (*func)(DCTELEM *block);
  62. enum formattag { NO_PERM, MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM,
  63. SSE2_PERM, PARTTRANS_PERM, TRANSPOSE_PERM } format;
  64. int mm_support;
  65. int nonspec;
  66. };
  67. static int cpu_flags;
  68. static const struct algo fdct_tab[] = {
  69. { "REF-DBL", ff_ref_fdct, NO_PERM },
  70. { "FAAN", ff_faandct, NO_PERM },
  71. { "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM },
  72. { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM },
  73. #if HAVE_MMX_INLINE
  74. { "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX },
  75. { "MMXEXT", ff_fdct_mmx2, NO_PERM, AV_CPU_FLAG_MMXEXT },
  76. { "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 },
  77. #endif
  78. #if HAVE_ALTIVEC
  79. { "altivecfdct", ff_fdct_altivec, NO_PERM, AV_CPU_FLAG_ALTIVEC },
  80. #endif
  81. #if ARCH_BFIN
  82. { "BFINfdct", ff_bfin_fdct, NO_PERM },
  83. #endif
  84. { 0 }
  85. };
  86. #if ARCH_X86_64 && HAVE_MMX && HAVE_YASM
  87. void ff_prores_idct_put_10_sse2(uint16_t *dst, int linesize,
  88. DCTELEM *block, int16_t *qmat);
  89. static void ff_prores_idct_put_10_sse2_wrap(DCTELEM *dst){
  90. DECLARE_ALIGNED(16, static int16_t, qmat)[64];
  91. DECLARE_ALIGNED(16, static int16_t, tmp)[64];
  92. int i;
  93. for(i=0; i<64; i++){
  94. qmat[i]=4;
  95. tmp[i]= dst[i];
  96. }
  97. ff_prores_idct_put_10_sse2(dst, 16, tmp, qmat);
  98. }
  99. #endif
  100. static const struct algo idct_tab[] = {
  101. { "FAANI", ff_faanidct, NO_PERM },
  102. { "REF-DBL", ff_ref_idct, NO_PERM },
  103. { "INT", ff_j_rev_dct, MMX_PERM },
  104. { "SIMPLE-C", ff_simple_idct_8, NO_PERM },
  105. #if HAVE_MMX_INLINE
  106. #if CONFIG_GPL
  107. { "LIBMPEG2-MMX", ff_mmx_idct, MMX_PERM, AV_CPU_FLAG_MMX, 1 },
  108. { "LIBMPEG2-MMX2", ff_mmxext_idct, MMX_PERM, AV_CPU_FLAG_MMX2, 1 },
  109. #endif
  110. { "SIMPLE-MMX", ff_simple_idct_mmx, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX },
  111. { "XVID-MMX", ff_idct_xvid_mmx, NO_PERM, AV_CPU_FLAG_MMX, 1 },
  112. { "XVID-MMXEXT", ff_idct_xvid_mmx2, NO_PERM, AV_CPU_FLAG_MMXEXT, 1 },
  113. { "XVID-SSE2", ff_idct_xvid_sse2, SSE2_PERM, AV_CPU_FLAG_SSE2, 1 },
  114. #if ARCH_X86_64 && HAVE_YASM
  115. { "PR-SSE2", ff_prores_idct_put_10_sse2_wrap, TRANSPOSE_PERM, AV_CPU_FLAG_SSE2, 1 },
  116. #endif
  117. #endif
  118. #if ARCH_BFIN
  119. { "BFINidct", ff_bfin_idct, NO_PERM },
  120. #endif
  121. #if ARCH_ARM
  122. { "SIMPLE-ARM", ff_simple_idct_arm, NO_PERM },
  123. { "INT-ARM", ff_j_rev_dct_arm, MMX_PERM },
  124. #endif
  125. #if HAVE_ARMV5TE
  126. { "SIMPLE-ARMV5TE", ff_simple_idct_armv5te,NO_PERM },
  127. #endif
  128. #if HAVE_ARMV6
  129. { "SIMPLE-ARMV6", ff_simple_idct_armv6, MMX_PERM },
  130. #endif
  131. #if HAVE_NEON
  132. { "SIMPLE-NEON", ff_simple_idct_neon, PARTTRANS_PERM },
  133. #endif
  134. #if ARCH_ALPHA
  135. { "SIMPLE-ALPHA", ff_simple_idct_axp, NO_PERM },
  136. #endif
  137. { 0 }
  138. };
  139. #define AANSCALE_BITS 12
  140. #define NB_ITS 20000
  141. #define NB_ITS_SPEED 50000
  142. static short idct_mmx_perm[64];
  143. static short idct_simple_mmx_perm[64] = {
  144. 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
  145. 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
  146. 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
  147. 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
  148. 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
  149. 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
  150. 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
  151. 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
  152. };
  153. static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };
  154. static void idct_mmx_init(void)
  155. {
  156. int i;
  157. /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
  158. for (i = 0; i < 64; i++) {
  159. idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
  160. }
  161. }
  162. DECLARE_ALIGNED(16, static DCTELEM, block)[64];
  163. DECLARE_ALIGNED(8, static DCTELEM, block1)[64];
  164. static void init_block(DCTELEM block[64], int test, int is_idct, AVLFG *prng, int vals)
  165. {
  166. int i, j;
  167. memset(block, 0, 64 * sizeof(*block));
  168. switch (test) {
  169. case 0:
  170. for (i = 0; i < 64; i++)
  171. block[i] = (av_lfg_get(prng) % (2*vals)) -vals;
  172. if (is_idct) {
  173. ff_ref_fdct(block);
  174. for (i = 0; i < 64; i++)
  175. block[i] >>= 3;
  176. }
  177. break;
  178. case 1:
  179. j = av_lfg_get(prng) % 10 + 1;
  180. for (i = 0; i < j; i++) {
  181. int idx = av_lfg_get(prng) % 64;
  182. block[idx] = av_lfg_get(prng) % (2*vals) -vals;
  183. }
  184. break;
  185. case 2:
  186. block[ 0] = av_lfg_get(prng) % (16*vals) - (8*vals);
  187. block[63] = (block[0] & 1) ^ 1;
  188. break;
  189. }
  190. }
  191. static void permute(DCTELEM dst[64], const DCTELEM src[64], int perm)
  192. {
  193. int i;
  194. if (perm == MMX_PERM) {
  195. for (i = 0; i < 64; i++)
  196. dst[idct_mmx_perm[i]] = src[i];
  197. } else if (perm == MMX_SIMPLE_PERM) {
  198. for (i = 0; i < 64; i++)
  199. dst[idct_simple_mmx_perm[i]] = src[i];
  200. } else if (perm == SSE2_PERM) {
  201. for (i = 0; i < 64; i++)
  202. dst[(i & 0x38) | idct_sse2_row_perm[i & 7]] = src[i];
  203. } else if (perm == PARTTRANS_PERM) {
  204. for (i = 0; i < 64; i++)
  205. dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
  206. } else if (perm == TRANSPOSE_PERM) {
  207. for (i = 0; i < 64; i++)
  208. dst[(i>>3) | ((i<<3)&0x38)] = src[i];
  209. } else {
  210. for (i = 0; i < 64; i++)
  211. dst[i] = src[i];
  212. }
  213. }
  214. static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits)
  215. {
  216. void (*ref)(DCTELEM *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
  217. int it, i, scale;
  218. int err_inf, v;
  219. int64_t err2, ti, ti1, it1, err_sum = 0;
  220. int64_t sysErr[64], sysErrMax = 0;
  221. int maxout = 0;
  222. int blockSumErrMax = 0, blockSumErr;
  223. AVLFG prng;
  224. const int vals=1<<bits;
  225. double omse, ome;
  226. int spec_err;
  227. av_lfg_init(&prng, 1);
  228. err_inf = 0;
  229. err2 = 0;
  230. for (i = 0; i < 64; i++)
  231. sysErr[i] = 0;
  232. for (it = 0; it < NB_ITS; it++) {
  233. init_block(block1, test, is_idct, &prng, vals);
  234. permute(block, block1, dct->format);
  235. dct->func(block);
  236. emms_c();
  237. if (dct->format == SCALE_PERM) {
  238. for (i = 0; i < 64; i++) {
  239. scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
  240. block[i] = (block[i] * scale) >> AANSCALE_BITS;
  241. }
  242. }
  243. ref(block1);
  244. blockSumErr = 0;
  245. for (i = 0; i < 64; i++) {
  246. int err = block[i] - block1[i];
  247. err_sum += err;
  248. v = abs(err);
  249. if (v > err_inf)
  250. err_inf = v;
  251. err2 += v * v;
  252. sysErr[i] += block[i] - block1[i];
  253. blockSumErr += v;
  254. if (abs(block[i]) > maxout)
  255. maxout = abs(block[i]);
  256. }
  257. if (blockSumErrMax < blockSumErr)
  258. blockSumErrMax = blockSumErr;
  259. }
  260. for (i = 0; i < 64; i++)
  261. sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
  262. for (i = 0; i < 64; i++) {
  263. if (i % 8 == 0)
  264. printf("\n");
  265. printf("%7d ", (int) sysErr[i]);
  266. }
  267. printf("\n");
  268. omse = (double) err2 / NB_ITS / 64;
  269. ome = (double) err_sum / NB_ITS / 64;
  270. spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
  271. printf("%s %s: max_err=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
  272. is_idct ? "IDCT" : "DCT", dct->name, err_inf,
  273. omse, ome, (double) sysErrMax / NB_ITS,
  274. maxout, blockSumErrMax);
  275. if (spec_err && !dct->nonspec)
  276. return 1;
  277. if (!speed)
  278. return 0;
  279. /* speed test */
  280. init_block(block, test, is_idct, &prng, vals);
  281. permute(block1, block, dct->format);
  282. ti = av_gettime();
  283. it1 = 0;
  284. do {
  285. for (it = 0; it < NB_ITS_SPEED; it++) {
  286. memcpy(block, block1, sizeof(block));
  287. dct->func(block);
  288. }
  289. emms_c();
  290. it1 += NB_ITS_SPEED;
  291. ti1 = av_gettime() - ti;
  292. } while (ti1 < 1000000);
  293. printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
  294. (double) it1 * 1000.0 / (double) ti1);
  295. return 0;
  296. }
  297. DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
  298. DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64];
  299. static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
  300. {
  301. static int init;
  302. static double c8[8][8];
  303. static double c4[4][4];
  304. double block1[64], block2[64], block3[64];
  305. double s, sum, v;
  306. int i, j, k;
  307. if (!init) {
  308. init = 1;
  309. for (i = 0; i < 8; i++) {
  310. sum = 0;
  311. for (j = 0; j < 8; j++) {
  312. s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
  313. c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
  314. sum += c8[i][j] * c8[i][j];
  315. }
  316. }
  317. for (i = 0; i < 4; i++) {
  318. sum = 0;
  319. for (j = 0; j < 4; j++) {
  320. s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
  321. c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
  322. sum += c4[i][j] * c4[i][j];
  323. }
  324. }
  325. }
  326. /* butterfly */
  327. s = 0.5 * sqrt(2.0);
  328. for (i = 0; i < 4; i++) {
  329. for (j = 0; j < 8; j++) {
  330. block1[8 * (2 * i) + j] =
  331. (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
  332. block1[8 * (2 * i + 1) + j] =
  333. (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
  334. }
  335. }
  336. /* idct8 on lines */
  337. for (i = 0; i < 8; i++) {
  338. for (j = 0; j < 8; j++) {
  339. sum = 0;
  340. for (k = 0; k < 8; k++)
  341. sum += c8[k][j] * block1[8 * i + k];
  342. block2[8 * i + j] = sum;
  343. }
  344. }
  345. /* idct4 */
  346. for (i = 0; i < 8; i++) {
  347. for (j = 0; j < 4; j++) {
  348. /* top */
  349. sum = 0;
  350. for (k = 0; k < 4; k++)
  351. sum += c4[k][j] * block2[8 * (2 * k) + i];
  352. block3[8 * (2 * j) + i] = sum;
  353. /* bottom */
  354. sum = 0;
  355. for (k = 0; k < 4; k++)
  356. sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
  357. block3[8 * (2 * j + 1) + i] = sum;
  358. }
  359. }
  360. /* clamp and store the result */
  361. for (i = 0; i < 8; i++) {
  362. for (j = 0; j < 8; j++) {
  363. v = block3[8 * i + j];
  364. if (v < 0) v = 0;
  365. else if (v > 255) v = 255;
  366. dest[i * linesize + j] = (int) rint(v);
  367. }
  368. }
  369. }
  370. static void idct248_error(const char *name,
  371. void (*idct248_put)(uint8_t *dest, int line_size,
  372. int16_t *block),
  373. int speed)
  374. {
  375. int it, i, it1, ti, ti1, err_max, v;
  376. AVLFG prng;
  377. av_lfg_init(&prng, 1);
  378. /* just one test to see if code is correct (precision is less
  379. important here) */
  380. err_max = 0;
  381. for (it = 0; it < NB_ITS; it++) {
  382. /* XXX: use forward transform to generate values */
  383. for (i = 0; i < 64; i++)
  384. block1[i] = av_lfg_get(&prng) % 256 - 128;
  385. block1[0] += 1024;
  386. for (i = 0; i < 64; i++)
  387. block[i] = block1[i];
  388. idct248_ref(img_dest1, 8, block);
  389. for (i = 0; i < 64; i++)
  390. block[i] = block1[i];
  391. idct248_put(img_dest, 8, block);
  392. for (i = 0; i < 64; i++) {
  393. v = abs((int) img_dest[i] - (int) img_dest1[i]);
  394. if (v == 255)
  395. printf("%d %d\n", img_dest[i], img_dest1[i]);
  396. if (v > err_max)
  397. err_max = v;
  398. }
  399. #if 0
  400. printf("ref=\n");
  401. for(i=0;i<8;i++) {
  402. int j;
  403. for(j=0;j<8;j++) {
  404. printf(" %3d", img_dest1[i*8+j]);
  405. }
  406. printf("\n");
  407. }
  408. printf("out=\n");
  409. for(i=0;i<8;i++) {
  410. int j;
  411. for(j=0;j<8;j++) {
  412. printf(" %3d", img_dest[i*8+j]);
  413. }
  414. printf("\n");
  415. }
  416. #endif
  417. }
  418. printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
  419. if (!speed)
  420. return;
  421. ti = av_gettime();
  422. it1 = 0;
  423. do {
  424. for (it = 0; it < NB_ITS_SPEED; it++) {
  425. for (i = 0; i < 64; i++)
  426. block[i] = block1[i];
  427. idct248_put(img_dest, 8, block);
  428. }
  429. emms_c();
  430. it1 += NB_ITS_SPEED;
  431. ti1 = av_gettime() - ti;
  432. } while (ti1 < 1000000);
  433. printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
  434. (double) it1 * 1000.0 / (double) ti1);
  435. }
  436. static void help(void)
  437. {
  438. printf("dct-test [-i] [<test-number>] [<bits>]\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. "bits Number of time domain bits to use, 8 is default\n"
  443. "-i test IDCT implementations\n"
  444. "-4 test IDCT248 implementations\n"
  445. "-t speed test\n");
  446. }
  447. #if !HAVE_GETOPT
  448. #include "compat/getopt.c"
  449. #endif
  450. int main(int argc, char **argv)
  451. {
  452. int test_idct = 0, test_248_dct = 0;
  453. int c, i;
  454. int test = 1;
  455. int speed = 0;
  456. int err = 0;
  457. int bits=8;
  458. cpu_flags = av_get_cpu_flags();
  459. ff_ref_dct_init();
  460. idct_mmx_init();
  461. for (;;) {
  462. c = getopt(argc, argv, "ih4t");
  463. if (c == -1)
  464. break;
  465. switch (c) {
  466. case 'i':
  467. test_idct = 1;
  468. break;
  469. case '4':
  470. test_248_dct = 1;
  471. break;
  472. case 't':
  473. speed = 1;
  474. break;
  475. default:
  476. case 'h':
  477. help();
  478. return 0;
  479. }
  480. }
  481. if (optind < argc)
  482. test = atoi(argv[optind]);
  483. if(optind+1 < argc) bits= atoi(argv[optind+1]);
  484. printf("ffmpeg DCT/IDCT test\n");
  485. if (test_248_dct) {
  486. idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
  487. } else {
  488. const struct algo *algos = test_idct ? idct_tab : fdct_tab;
  489. for (i = 0; algos[i].name; i++)
  490. if (!(~cpu_flags & algos[i].mm_support)) {
  491. err |= dct_error(&algos[i], test, test_idct, speed, bits);
  492. }
  493. }
  494. return err;
  495. }