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.

575 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 <stdlib.h>
  27. #include <stdio.h>
  28. #include <string.h>
  29. #include <unistd.h>
  30. #include <math.h>
  31. #include "libavutil/cpu.h"
  32. #include "libavutil/common.h"
  33. #include "libavutil/lfg.h"
  34. #include "libavutil/time.h"
  35. #include "simple_idct.h"
  36. #include "aandcttab.h"
  37. #include "faandct.h"
  38. #include "faanidct.h"
  39. #include "x86/idct_xvid.h"
  40. #include "dctref.h"
  41. #undef printf
  42. void ff_mmx_idct(DCTELEM *data);
  43. void ff_mmxext_idct(DCTELEM *data);
  44. // BFIN
  45. void ff_bfin_idct(DCTELEM *block);
  46. void ff_bfin_fdct(DCTELEM *block);
  47. // ALTIVEC
  48. void ff_fdct_altivec(DCTELEM *block);
  49. //void ff_idct_altivec(DCTELEM *block);?? no routine
  50. // ARM
  51. void ff_j_rev_dct_arm(DCTELEM *data);
  52. void ff_simple_idct_arm(DCTELEM *data);
  53. void ff_simple_idct_armv5te(DCTELEM *data);
  54. void ff_simple_idct_armv6(DCTELEM *data);
  55. void ff_simple_idct_neon(DCTELEM *data);
  56. void ff_simple_idct_axp(DCTELEM *data);
  57. struct algo {
  58. const char *name;
  59. void (*func)(DCTELEM *block);
  60. enum formattag { NO_PERM, MMX_PERM, MMX_SIMPLE_PERM, SCALE_PERM,
  61. SSE2_PERM, PARTTRANS_PERM, TRANSPOSE_PERM } format;
  62. int mm_support;
  63. int nonspec;
  64. };
  65. static int cpu_flags;
  66. static const struct algo fdct_tab[] = {
  67. { "REF-DBL", ff_ref_fdct, NO_PERM },
  68. { "FAAN", ff_faandct, NO_PERM },
  69. { "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM },
  70. { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM },
  71. #if HAVE_MMX
  72. { "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX },
  73. { "MMX2", ff_fdct_mmx2, NO_PERM, AV_CPU_FLAG_MMX2 },
  74. { "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 },
  75. #endif
  76. #if HAVE_ALTIVEC
  77. { "altivecfdct", ff_fdct_altivec, NO_PERM, AV_CPU_FLAG_ALTIVEC },
  78. #endif
  79. #if ARCH_BFIN
  80. { "BFINfdct", ff_bfin_fdct, NO_PERM },
  81. #endif
  82. { 0 }
  83. };
  84. #if ARCH_X86_64 && HAVE_MMX && HAVE_YASM
  85. void ff_prores_idct_put_10_sse2(uint16_t *dst, int linesize,
  86. DCTELEM *block, int16_t *qmat);
  87. static void ff_prores_idct_put_10_sse2_wrap(DCTELEM *dst){
  88. DECLARE_ALIGNED(16, static int16_t, qmat)[64];
  89. DECLARE_ALIGNED(16, static int16_t, tmp)[64];
  90. int i;
  91. for(i=0; i<64; i++){
  92. qmat[i]=4;
  93. tmp[i]= dst[i];
  94. }
  95. ff_prores_idct_put_10_sse2(dst, 16, tmp, qmat);
  96. }
  97. #endif
  98. static const struct algo idct_tab[] = {
  99. { "FAANI", ff_faanidct, NO_PERM },
  100. { "REF-DBL", ff_ref_idct, NO_PERM },
  101. { "INT", ff_j_rev_dct, MMX_PERM },
  102. { "SIMPLE-C", ff_simple_idct_8, NO_PERM },
  103. #if HAVE_MMX
  104. #if CONFIG_GPL
  105. { "LIBMPEG2-MMX", ff_mmx_idct, MMX_PERM, AV_CPU_FLAG_MMX, 1 },
  106. { "LIBMPEG2-MMX2", ff_mmxext_idct, MMX_PERM, AV_CPU_FLAG_MMX2, 1 },
  107. #endif
  108. { "SIMPLE-MMX", ff_simple_idct_mmx, MMX_SIMPLE_PERM, AV_CPU_FLAG_MMX },
  109. { "XVID-MMX", ff_idct_xvid_mmx, NO_PERM, AV_CPU_FLAG_MMX, 1 },
  110. { "XVID-MMX2", ff_idct_xvid_mmx2, NO_PERM, AV_CPU_FLAG_MMX2, 1 },
  111. { "XVID-SSE2", ff_idct_xvid_sse2, SSE2_PERM, AV_CPU_FLAG_SSE2, 1 },
  112. #if ARCH_X86_64 && HAVE_YASM
  113. { "PR-SSE2", ff_prores_idct_put_10_sse2_wrap, TRANSPOSE_PERM, AV_CPU_FLAG_SSE2, 1 },
  114. #endif
  115. #endif
  116. #if ARCH_BFIN
  117. { "BFINidct", ff_bfin_idct, NO_PERM },
  118. #endif
  119. #if ARCH_ARM
  120. { "SIMPLE-ARM", ff_simple_idct_arm, NO_PERM },
  121. { "INT-ARM", ff_j_rev_dct_arm, MMX_PERM },
  122. #endif
  123. #if HAVE_ARMV5TE
  124. { "SIMPLE-ARMV5TE", ff_simple_idct_armv5te,NO_PERM },
  125. #endif
  126. #if HAVE_ARMV6
  127. { "SIMPLE-ARMV6", ff_simple_idct_armv6, MMX_PERM },
  128. #endif
  129. #if HAVE_NEON
  130. { "SIMPLE-NEON", ff_simple_idct_neon, PARTTRANS_PERM },
  131. #endif
  132. #if ARCH_ALPHA
  133. { "SIMPLE-ALPHA", ff_simple_idct_axp, NO_PERM },
  134. #endif
  135. { 0 }
  136. };
  137. #define AANSCALE_BITS 12
  138. #define NB_ITS 20000
  139. #define NB_ITS_SPEED 50000
  140. static short idct_mmx_perm[64];
  141. static short idct_simple_mmx_perm[64] = {
  142. 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
  143. 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
  144. 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
  145. 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
  146. 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
  147. 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
  148. 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
  149. 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
  150. };
  151. static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };
  152. static void idct_mmx_init(void)
  153. {
  154. int i;
  155. /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
  156. for (i = 0; i < 64; i++) {
  157. idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
  158. }
  159. }
  160. DECLARE_ALIGNED(16, static DCTELEM, block)[64];
  161. DECLARE_ALIGNED(8, static DCTELEM, block1)[64];
  162. static void init_block(DCTELEM block[64], int test, int is_idct, AVLFG *prng, int vals)
  163. {
  164. int i, j;
  165. memset(block, 0, 64 * sizeof(*block));
  166. switch (test) {
  167. case 0:
  168. for (i = 0; i < 64; i++)
  169. block[i] = (av_lfg_get(prng) % (2*vals)) -vals;
  170. if (is_idct) {
  171. ff_ref_fdct(block);
  172. for (i = 0; i < 64; i++)
  173. block[i] >>= 3;
  174. }
  175. break;
  176. case 1:
  177. j = av_lfg_get(prng) % 10 + 1;
  178. for (i = 0; i < j; i++)
  179. block[av_lfg_get(prng) % 64] = av_lfg_get(prng) % (2*vals) -vals;
  180. break;
  181. case 2:
  182. block[ 0] = av_lfg_get(prng) % (16*vals) - (8*vals);
  183. block[63] = (block[0] & 1) ^ 1;
  184. break;
  185. }
  186. }
  187. static void permute(DCTELEM dst[64], const DCTELEM src[64], int perm)
  188. {
  189. int i;
  190. if (perm == MMX_PERM) {
  191. for (i = 0; i < 64; i++)
  192. dst[idct_mmx_perm[i]] = src[i];
  193. } else if (perm == MMX_SIMPLE_PERM) {
  194. for (i = 0; i < 64; i++)
  195. dst[idct_simple_mmx_perm[i]] = src[i];
  196. } else if (perm == SSE2_PERM) {
  197. for (i = 0; i < 64; i++)
  198. dst[(i & 0x38) | idct_sse2_row_perm[i & 7]] = src[i];
  199. } else if (perm == PARTTRANS_PERM) {
  200. for (i = 0; i < 64; i++)
  201. dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
  202. } else if (perm == TRANSPOSE_PERM) {
  203. for (i = 0; i < 64; i++)
  204. dst[(i>>3) | ((i<<3)&0x38)] = src[i];
  205. } else {
  206. for (i = 0; i < 64; i++)
  207. dst[i] = src[i];
  208. }
  209. }
  210. static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits)
  211. {
  212. void (*ref)(DCTELEM *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
  213. int it, i, scale;
  214. int err_inf, v;
  215. int64_t err2, ti, ti1, it1, err_sum = 0;
  216. int64_t sysErr[64], sysErrMax = 0;
  217. int maxout = 0;
  218. int blockSumErrMax = 0, blockSumErr;
  219. AVLFG prng;
  220. const int vals=1<<bits;
  221. double omse, ome;
  222. int spec_err;
  223. av_lfg_init(&prng, 1);
  224. err_inf = 0;
  225. err2 = 0;
  226. for (i = 0; i < 64; i++)
  227. sysErr[i] = 0;
  228. for (it = 0; it < NB_ITS; it++) {
  229. init_block(block1, test, is_idct, &prng, vals);
  230. permute(block, block1, dct->format);
  231. dct->func(block);
  232. emms_c();
  233. if (dct->format == SCALE_PERM) {
  234. for (i = 0; i < 64; i++) {
  235. scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
  236. block[i] = (block[i] * scale) >> AANSCALE_BITS;
  237. }
  238. }
  239. ref(block1);
  240. blockSumErr = 0;
  241. for (i = 0; i < 64; i++) {
  242. int err = block[i] - block1[i];
  243. err_sum += err;
  244. v = abs(err);
  245. if (v > err_inf)
  246. err_inf = v;
  247. err2 += v * v;
  248. sysErr[i] += block[i] - block1[i];
  249. blockSumErr += v;
  250. if (abs(block[i]) > maxout)
  251. maxout = abs(block[i]);
  252. }
  253. if (blockSumErrMax < blockSumErr)
  254. blockSumErrMax = blockSumErr;
  255. }
  256. for (i = 0; i < 64; i++)
  257. sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
  258. for (i = 0; i < 64; i++) {
  259. if (i % 8 == 0)
  260. printf("\n");
  261. printf("%7d ", (int) sysErr[i]);
  262. }
  263. printf("\n");
  264. omse = (double) err2 / NB_ITS / 64;
  265. ome = (double) err_sum / NB_ITS / 64;
  266. spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
  267. printf("%s %s: max_err=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
  268. is_idct ? "IDCT" : "DCT", dct->name, err_inf,
  269. omse, ome, (double) sysErrMax / NB_ITS,
  270. maxout, blockSumErrMax);
  271. if (spec_err && !dct->nonspec)
  272. return 1;
  273. if (!speed)
  274. return 0;
  275. /* speed test */
  276. init_block(block, test, is_idct, &prng, vals);
  277. permute(block1, block, dct->format);
  278. ti = av_gettime();
  279. it1 = 0;
  280. do {
  281. for (it = 0; it < NB_ITS_SPEED; it++) {
  282. memcpy(block, block1, sizeof(block));
  283. dct->func(block);
  284. }
  285. emms_c();
  286. it1 += NB_ITS_SPEED;
  287. ti1 = av_gettime() - ti;
  288. } while (ti1 < 1000000);
  289. printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
  290. (double) it1 * 1000.0 / (double) ti1);
  291. return 0;
  292. }
  293. DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
  294. DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64];
  295. static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
  296. {
  297. static int init;
  298. static double c8[8][8];
  299. static double c4[4][4];
  300. double block1[64], block2[64], block3[64];
  301. double s, sum, v;
  302. int i, j, k;
  303. if (!init) {
  304. init = 1;
  305. for (i = 0; i < 8; i++) {
  306. sum = 0;
  307. for (j = 0; j < 8; j++) {
  308. s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
  309. c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
  310. sum += c8[i][j] * c8[i][j];
  311. }
  312. }
  313. for (i = 0; i < 4; i++) {
  314. sum = 0;
  315. for (j = 0; j < 4; j++) {
  316. s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
  317. c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
  318. sum += c4[i][j] * c4[i][j];
  319. }
  320. }
  321. }
  322. /* butterfly */
  323. s = 0.5 * sqrt(2.0);
  324. for (i = 0; i < 4; i++) {
  325. for (j = 0; j < 8; j++) {
  326. block1[8 * (2 * i) + j] =
  327. (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
  328. block1[8 * (2 * i + 1) + j] =
  329. (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
  330. }
  331. }
  332. /* idct8 on lines */
  333. for (i = 0; i < 8; i++) {
  334. for (j = 0; j < 8; j++) {
  335. sum = 0;
  336. for (k = 0; k < 8; k++)
  337. sum += c8[k][j] * block1[8 * i + k];
  338. block2[8 * i + j] = sum;
  339. }
  340. }
  341. /* idct4 */
  342. for (i = 0; i < 8; i++) {
  343. for (j = 0; j < 4; j++) {
  344. /* top */
  345. sum = 0;
  346. for (k = 0; k < 4; k++)
  347. sum += c4[k][j] * block2[8 * (2 * k) + i];
  348. block3[8 * (2 * j) + i] = sum;
  349. /* bottom */
  350. sum = 0;
  351. for (k = 0; k < 4; k++)
  352. sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
  353. block3[8 * (2 * j + 1) + i] = sum;
  354. }
  355. }
  356. /* clamp and store the result */
  357. for (i = 0; i < 8; i++) {
  358. for (j = 0; j < 8; j++) {
  359. v = block3[8 * i + j];
  360. if (v < 0) v = 0;
  361. else if (v > 255) v = 255;
  362. dest[i * linesize + j] = (int) rint(v);
  363. }
  364. }
  365. }
  366. static void idct248_error(const char *name,
  367. void (*idct248_put)(uint8_t *dest, int line_size,
  368. int16_t *block),
  369. int speed)
  370. {
  371. int it, i, it1, ti, ti1, err_max, v;
  372. AVLFG prng;
  373. av_lfg_init(&prng, 1);
  374. /* just one test to see if code is correct (precision is less
  375. important here) */
  376. err_max = 0;
  377. for (it = 0; it < NB_ITS; it++) {
  378. /* XXX: use forward transform to generate values */
  379. for (i = 0; i < 64; i++)
  380. block1[i] = av_lfg_get(&prng) % 256 - 128;
  381. block1[0] += 1024;
  382. for (i = 0; i < 64; i++)
  383. block[i] = block1[i];
  384. idct248_ref(img_dest1, 8, block);
  385. for (i = 0; i < 64; i++)
  386. block[i] = block1[i];
  387. idct248_put(img_dest, 8, block);
  388. for (i = 0; i < 64; i++) {
  389. v = abs((int) img_dest[i] - (int) img_dest1[i]);
  390. if (v == 255)
  391. printf("%d %d\n", img_dest[i], img_dest1[i]);
  392. if (v > err_max)
  393. err_max = v;
  394. }
  395. #if 0
  396. printf("ref=\n");
  397. for(i=0;i<8;i++) {
  398. int j;
  399. for(j=0;j<8;j++) {
  400. printf(" %3d", img_dest1[i*8+j]);
  401. }
  402. printf("\n");
  403. }
  404. printf("out=\n");
  405. for(i=0;i<8;i++) {
  406. int j;
  407. for(j=0;j<8;j++) {
  408. printf(" %3d", img_dest[i*8+j]);
  409. }
  410. printf("\n");
  411. }
  412. #endif
  413. }
  414. printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
  415. if (!speed)
  416. return;
  417. ti = av_gettime();
  418. it1 = 0;
  419. do {
  420. for (it = 0; it < NB_ITS_SPEED; it++) {
  421. for (i = 0; i < 64; i++)
  422. block[i] = block1[i];
  423. idct248_put(img_dest, 8, block);
  424. }
  425. emms_c();
  426. it1 += NB_ITS_SPEED;
  427. ti1 = av_gettime() - ti;
  428. } while (ti1 < 1000000);
  429. printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
  430. (double) it1 * 1000.0 / (double) ti1);
  431. }
  432. static void help(void)
  433. {
  434. printf("dct-test [-i] [<test-number>] [<bits>]\n"
  435. "test-number 0 -> test with random matrixes\n"
  436. " 1 -> test with random sparse matrixes\n"
  437. " 2 -> do 3. test from mpeg4 std\n"
  438. "bits Number of time domain bits to use, 8 is default\n"
  439. "-i test IDCT implementations\n"
  440. "-4 test IDCT248 implementations\n"
  441. "-t speed test\n");
  442. }
  443. int main(int argc, char **argv)
  444. {
  445. int test_idct = 0, test_248_dct = 0;
  446. int c, i;
  447. int test = 1;
  448. int speed = 0;
  449. int err = 0;
  450. int bits=8;
  451. cpu_flags = av_get_cpu_flags();
  452. ff_ref_dct_init();
  453. idct_mmx_init();
  454. for (;;) {
  455. c = getopt(argc, argv, "ih4t");
  456. if (c == -1)
  457. break;
  458. switch (c) {
  459. case 'i':
  460. test_idct = 1;
  461. break;
  462. case '4':
  463. test_248_dct = 1;
  464. break;
  465. case 't':
  466. speed = 1;
  467. break;
  468. default:
  469. case 'h':
  470. help();
  471. return 0;
  472. }
  473. }
  474. if (optind < argc)
  475. test = atoi(argv[optind]);
  476. if(optind+1 < argc) bits= atoi(argv[optind+1]);
  477. printf("ffmpeg DCT/IDCT test\n");
  478. if (test_248_dct) {
  479. idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
  480. } else {
  481. const struct algo *algos = test_idct ? idct_tab : fdct_tab;
  482. for (i = 0; algos[i].name; i++)
  483. if (!(~cpu_flags & algos[i].mm_support)) {
  484. err |= dct_error(&algos[i], test, test_idct, speed, bits);
  485. }
  486. }
  487. return err;
  488. }