tests: Move all test programs to a subdirectory
[libav.git] / libavcodec / tests / dct.c
CommitLineData
04d7f601
DB
1/*
2 * (c) 2001 Fabrice Bellard
3ac35bdb 3 * 2007 Marc Hoffman <marc.hoffman@analog.com>
04d7f601 4 *
2912e87a 5 * This file is part of Libav.
b78e7197 6 *
2912e87a 7 * Libav is free software; you can redistribute it and/or
04d7f601
DB
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
b78e7197 10 * version 2.1 of the License, or (at your option) any later version.
04d7f601 11 *
2912e87a 12 * Libav is distributed in the hope that it will be useful,
04d7f601
DB
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
2912e87a 18 * License along with Libav; if not, write to the Free Software
04d7f601
DB
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
983e3246 22/**
ba87f080 23 * @file
94f694a4 24 * DCT test (c) 2001 Fabrice Bellard
983e3246
MN
25 * Started from sample code by Juan J. Sierralta P.
26 */
27
667fb97a 28#include "config.h"
de6d9b64
FB
29#include <stdlib.h>
30#include <stdio.h>
31#include <string.h>
667fb97a 32#if HAVE_UNISTD_H
de6d9b64 33#include <unistd.h>
667fb97a 34#endif
12807c8d 35#include <math.h>
de6d9b64 36
c6c98d08 37#include "libavutil/cpu.h"
ae32e509 38#include "libavutil/common.h"
a6a750c7 39#include "libavutil/internal.h"
294eaa26 40#include "libavutil/lfg.h"
980f81d9 41#include "libavutil/time.h"
de6d9b64 42
a6a750c7
DB
43#include "libavcodec/aandcttab.h"
44#include "libavcodec/dct.h"
45#include "libavcodec/dctref.h"
46#include "libavcodec/faandct.h"
47#include "libavcodec/faanidct.h"
48#include "libavcodec/idctdsp.h"
49#include "libavcodec/simple_idct.h"
50#include "libavcodec/xvididct.h"
9e1586fc 51
3ac35bdb 52struct algo {
36fa9ef3 53 const char *name;
88bd7fdc 54 void (*func)(int16_t *block);
e0a2e60c 55 enum idct_permutation_type perm_type;
746ad4e0 56 int cpu_flag;
dbf396d4 57 int nonspec;
3ac35bdb
MH
58};
59
1274ea8d 60static const struct algo fdct_tab[] = {
e0a2e60c 61 { "REF-DBL", ff_ref_fdct, FF_IDCT_PERM_NONE },
e0a2e60c
DB
62 { "IJG-AAN-INT", ff_fdct_ifast, FF_IDCT_PERM_NONE },
63 { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, FF_IDCT_PERM_NONE },
1274ea8d
DB
64#if CONFIG_FAANDCT
65 { "FAAN", ff_faandct, FF_IDCT_PERM_NONE },
66#endif /* CONFIG_FAANDCT */
4b357756
MR
67};
68
7a1d6ddd 69static const struct algo idct_tab[] = {
e0a2e60c
DB
70 { "REF-DBL", ff_ref_idct, FF_IDCT_PERM_NONE },
71 { "INT", ff_j_rev_dct, FF_IDCT_PERM_LIBMPEG2 },
72 { "SIMPLE-C", ff_simple_idct_8, FF_IDCT_PERM_NONE },
1274ea8d
DB
73#if CONFIG_FAANIDCT
74 { "FAANI", ff_faanidct, FF_IDCT_PERM_NONE },
75#endif /* CONFIG_FAANIDCT */
7a1d6ddd
PM
76#if CONFIG_MPEG4_DECODER
77 { "XVID", ff_xvid_idct, FF_IDCT_PERM_NONE, 0, 1 },
78#endif /* CONFIG_MPEG4_DECODER */
81b9bf31 79};
3ac35bdb 80
b250f9c6 81#if ARCH_ARM
a6a750c7 82#include "arm/dct.c"
81b9bf31 83#elif ARCH_PPC
a6a750c7 84#include "ppc/dct.c"
81b9bf31 85#elif ARCH_X86
a6a750c7 86#include "x86/dct.c"
81b9bf31 87#else
d89e58f5
VG
88static const struct algo fdct_tab_arch[] = { { 0 } };
89static const struct algo idct_tab_arch[] = { { 0 } };
479044ce 90#endif
479044ce 91
de6d9b64 92#define AANSCALE_BITS 12
de6d9b64 93
de6d9b64
FB
94#define NB_ITS 20000
95#define NB_ITS_SPEED 50000
96
88bd7fdc
DB
97DECLARE_ALIGNED(16, static int16_t, block)[64];
98DECLARE_ALIGNED(8, static int16_t, block1)[64];
9e1586fc 99
88bd7fdc 100static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng)
ae2e8971
MR
101{
102 int i, j;
103
104 memset(block, 0, 64 * sizeof(*block));
105
106 switch (test) {
107 case 0:
108 for (i = 0; i < 64; i++)
109 block[i] = (av_lfg_get(prng) % 512) - 256;
110 if (is_idct) {
111 ff_ref_fdct(block);
112 for (i = 0; i < 64; i++)
113 block[i] >>= 3;
114 }
115 break;
116 case 1:
117 j = av_lfg_get(prng) % 10 + 1;
118 for (i = 0; i < j; i++)
119 block[av_lfg_get(prng) % 64] = av_lfg_get(prng) % 512 - 256;
120 break;
121 case 2:
122 block[ 0] = av_lfg_get(prng) % 4096 - 2048;
123 block[63] = (block[0] & 1) ^ 1;
124 break;
125 }
126}
127
e0a2e60c
DB
128static void permute(int16_t dst[64], const int16_t src[64],
129 enum idct_permutation_type perm_type)
ae2e8971
MR
130{
131 int i;
132
81b9bf31
DB
133#if ARCH_X86
134 if (permute_x86(dst, src, perm_type))
135 return;
136#endif
137
e0a2e60c
DB
138 switch (perm_type) {
139 case FF_IDCT_PERM_LIBMPEG2:
ae2e8971 140 for (i = 0; i < 64; i++)
913fa85a 141 dst[(i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2)] = src[i];
e0a2e60c 142 break;
e0a2e60c 143 case FF_IDCT_PERM_PARTTRANS:
ae2e8971
MR
144 for (i = 0; i < 64; i++)
145 dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
e0a2e60c
DB
146 break;
147 default:
ae2e8971
MR
148 for (i = 0; i < 64; i++)
149 dst[i] = src[i];
e0a2e60c 150 break;
ae2e8971
MR
151 }
152}
153
dbf396d4 154static int dct_error(const struct algo *dct, int test, int is_idct, int speed)
de6d9b64 155{
88bd7fdc 156 void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
de6d9b64 157 int it, i, scale;
de6d9b64 158 int err_inf, v;
dbf396d4 159 int64_t err2, ti, ti1, it1, err_sum = 0;
36fa9ef3
MR
160 int64_t sysErr[64], sysErrMax = 0;
161 int maxout = 0;
162 int blockSumErrMax = 0, blockSumErr;
64bde197 163 AVLFG prng;
dbf396d4
MR
164 double omse, ome;
165 int spec_err;
de6d9b64 166
64bde197 167 av_lfg_init(&prng, 1);
de6d9b64
FB
168
169 err_inf = 0;
170 err2 = 0;
36fa9ef3
MR
171 for (i = 0; i < 64; i++)
172 sysErr[i] = 0;
173 for (it = 0; it < NB_ITS; it++) {
ae2e8971 174 init_block(block1, test, is_idct, &prng);
e0a2e60c 175 permute(block, block1, dct->perm_type);
9e1586fc 176
4f905a65 177 dct->func(block);
db7d8fb4 178 emms_c();
9e1586fc 179
e0a2e60c 180 if (!strcmp(dct->name, "IJG-AAN-INT")) {
36fa9ef3
MR
181 for (i = 0; i < 64; i++) {
182 scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
183 block[i] = (block[i] * scale) >> AANSCALE_BITS;
86748dbc
MN
184 }
185 }
186
74965f26 187 ref(block1);
de6d9b64 188
36fa9ef3
MR
189 blockSumErr = 0;
190 for (i = 0; i < 64; i++) {
dbf396d4
MR
191 int err = block[i] - block1[i];
192 err_sum += err;
193 v = abs(err);
de6d9b64
FB
194 if (v > err_inf)
195 err_inf = v;
196 err2 += v * v;
bb270c08
DB
197 sysErr[i] += block[i] - block1[i];
198 blockSumErr += v;
36fa9ef3
MR
199 if (abs(block[i]) > maxout)
200 maxout = abs(block[i]);
de6d9b64 201 }
36fa9ef3
MR
202 if (blockSumErrMax < blockSumErr)
203 blockSumErrMax = blockSumErr;
86748dbc 204 }
36fa9ef3
MR
205 for (i = 0; i < 64; i++)
206 sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
115329f1 207
36fa9ef3
MR
208 for (i = 0; i < 64; i++) {
209 if (i % 8 == 0)
210 printf("\n");
211 printf("%7d ", (int) sysErr[i]);
de6d9b64 212 }
86748dbc 213 printf("\n");
115329f1 214
dbf396d4
MR
215 omse = (double) err2 / NB_ITS / 64;
216 ome = (double) err_sum / NB_ITS / 64;
217
218 spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
219
220 printf("%s %s: ppe=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
cf2b4f88 221 is_idct ? "IDCT" : "DCT", dct->name, err_inf,
dbf396d4 222 omse, ome, (double) sysErrMax / NB_ITS,
36fa9ef3 223 maxout, blockSumErrMax);
e6ff0648 224
dbf396d4
MR
225 if (spec_err && !dct->nonspec)
226 return 1;
227
7fd2c138 228 if (!speed)
dbf396d4 229 return 0;
7fd2c138 230
de6d9b64 231 /* speed test */
ae2e8971 232 init_block(block, test, is_idct, &prng);
e0a2e60c 233 permute(block1, block, dct->perm_type);
9e1586fc 234
6fdf35fe 235 ti = av_gettime_relative();
de6d9b64
FB
236 it1 = 0;
237 do {
36fa9ef3 238 for (it = 0; it < NB_ITS_SPEED; it++) {
ae2e8971 239 memcpy(block, block1, sizeof(block));
4f905a65 240 dct->func(block);
de6d9b64
FB
241 }
242 it1 += NB_ITS_SPEED;
6fdf35fe 243 ti1 = av_gettime_relative() - ti;
de6d9b64 244 } while (ti1 < 1000000);
db7d8fb4 245 emms_c();
de6d9b64 246
cf2b4f88 247 printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
36fa9ef3 248 (double) it1 * 1000.0 / (double) ti1);
dbf396d4
MR
249
250 return 0;
de6d9b64
FB
251}
252
c6727809
MR
253DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
254DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64];
a46a3ce4 255
504ffed1 256static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
a46a3ce4
FB
257{
258 static int init;
259 static double c8[8][8];
260 static double c4[4][4];
261 double block1[64], block2[64], block3[64];
262 double s, sum, v;
263 int i, j, k;
264
265 if (!init) {
266 init = 1;
267
36fa9ef3 268 for (i = 0; i < 8; i++) {
a46a3ce4 269 sum = 0;
36fa9ef3
MR
270 for (j = 0; j < 8; j++) {
271 s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
a46a3ce4
FB
272 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
273 sum += c8[i][j] * c8[i][j];
274 }
275 }
115329f1 276
36fa9ef3 277 for (i = 0; i < 4; i++) {
a46a3ce4 278 sum = 0;
36fa9ef3
MR
279 for (j = 0; j < 4; j++) {
280 s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
a46a3ce4
FB
281 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
282 sum += c4[i][j] * c4[i][j];
283 }
284 }
285 }
286
287 /* butterfly */
652f0197 288 s = 0.5 * sqrt(2.0);
36fa9ef3
MR
289 for (i = 0; i < 4; i++) {
290 for (j = 0; j < 8; j++) {
291 block1[8 * (2 * i) + j] =
292 (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
293 block1[8 * (2 * i + 1) + j] =
294 (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
a46a3ce4
FB
295 }
296 }
297
298 /* idct8 on lines */
36fa9ef3
MR
299 for (i = 0; i < 8; i++) {
300 for (j = 0; j < 8; j++) {
a46a3ce4 301 sum = 0;
36fa9ef3
MR
302 for (k = 0; k < 8; k++)
303 sum += c8[k][j] * block1[8 * i + k];
304 block2[8 * i + j] = sum;
a46a3ce4
FB
305 }
306 }
307
308 /* idct4 */
36fa9ef3
MR
309 for (i = 0; i < 8; i++) {
310 for (j = 0; j < 4; j++) {
a46a3ce4
FB
311 /* top */
312 sum = 0;
36fa9ef3
MR
313 for (k = 0; k < 4; k++)
314 sum += c4[k][j] * block2[8 * (2 * k) + i];
315 block3[8 * (2 * j) + i] = sum;
a46a3ce4
FB
316
317 /* bottom */
318 sum = 0;
36fa9ef3
MR
319 for (k = 0; k < 4; k++)
320 sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
321 block3[8 * (2 * j + 1) + i] = sum;
a46a3ce4
FB
322 }
323 }
324
325 /* clamp and store the result */
36fa9ef3
MR
326 for (i = 0; i < 8; i++) {
327 for (j = 0; j < 8; j++) {
328 v = block3[8 * i + j];
329 if (v < 0) v = 0;
330 else if (v > 255) v = 255;
331 dest[i * linesize + j] = (int) rint(v);
a46a3ce4
FB
332 }
333 }
334}
335
504ffed1 336static void idct248_error(const char *name,
36fa9ef3 337 void (*idct248_put)(uint8_t *dest, int line_size,
7fd2c138
MR
338 int16_t *block),
339 int speed)
a46a3ce4
FB
340{
341 int it, i, it1, ti, ti1, err_max, v;
64bde197 342 AVLFG prng;
294eaa26 343
64bde197 344 av_lfg_init(&prng, 1);
115329f1 345
a46a3ce4
FB
346 /* just one test to see if code is correct (precision is less
347 important here) */
348 err_max = 0;
36fa9ef3 349 for (it = 0; it < NB_ITS; it++) {
652f0197 350 /* XXX: use forward transform to generate values */
36fa9ef3 351 for (i = 0; i < 64; i++)
64bde197 352 block1[i] = av_lfg_get(&prng) % 256 - 128;
652f0197
FB
353 block1[0] += 1024;
354
36fa9ef3
MR
355 for (i = 0; i < 64; i++)
356 block[i] = block1[i];
a46a3ce4 357 idct248_ref(img_dest1, 8, block);
115329f1 358
36fa9ef3
MR
359 for (i = 0; i < 64; i++)
360 block[i] = block1[i];
652f0197 361 idct248_put(img_dest, 8, block);
115329f1 362
36fa9ef3
MR
363 for (i = 0; i < 64; i++) {
364 v = abs((int) img_dest[i] - (int) img_dest1[i]);
652f0197
FB
365 if (v == 255)
366 printf("%d %d\n", img_dest[i], img_dest1[i]);
367 if (v > err_max)
368 err_max = v;
369 }
a46a3ce4 370 }
36fa9ef3 371 printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
a46a3ce4 372
7fd2c138
MR
373 if (!speed)
374 return;
375
6fdf35fe 376 ti = av_gettime_relative();
a46a3ce4
FB
377 it1 = 0;
378 do {
36fa9ef3
MR
379 for (it = 0; it < NB_ITS_SPEED; it++) {
380 for (i = 0; i < 64; i++)
381 block[i] = block1[i];
a46a3ce4
FB
382 idct248_put(img_dest, 8, block);
383 }
384 it1 += NB_ITS_SPEED;
6fdf35fe 385 ti1 = av_gettime_relative() - ti;
a46a3ce4 386 } while (ti1 < 1000000);
db7d8fb4 387 emms_c();
a46a3ce4 388
36fa9ef3
MR
389 printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
390 (double) it1 * 1000.0 / (double) ti1);
a46a3ce4
FB
391}
392
504ffed1 393static void help(void)
9e1586fc 394{
86748dbc
MN
395 printf("dct-test [-i] [<test-number>]\n"
396 "test-number 0 -> test with random matrixes\n"
397 " 1 -> test with random sparse matrixes\n"
41ed7ab4 398 " 2 -> do 3. test from MPEG-4 std\n"
a46a3ce4 399 "-i test IDCT implementations\n"
7fd2c138
MR
400 "-4 test IDCT248 implementations\n"
401 "-t speed test\n");
9e1586fc
FB
402}
403
667fb97a
RB
404#if !HAVE_GETOPT
405#include "compat/getopt.c"
406#endif
407
de6d9b64
FB
408int main(int argc, char **argv)
409{
a46a3ce4 410 int test_idct = 0, test_248_dct = 0;
36fa9ef3
MR
411 int c, i;
412 int test = 1;
7fd2c138 413 int speed = 0;
dbf396d4 414 int err = 0;
36fa9ef3 415
0de74546 416 ff_ref_dct_init();
f67a10cd 417
36fa9ef3 418 for (;;) {
7fd2c138 419 c = getopt(argc, argv, "ih4t");
9e1586fc
FB
420 if (c == -1)
421 break;
36fa9ef3 422 switch (c) {
9e1586fc
FB
423 case 'i':
424 test_idct = 1;
425 break;
a46a3ce4
FB
426 case '4':
427 test_248_dct = 1;
428 break;
7fd2c138
MR
429 case 't':
430 speed = 1;
431 break;
36fa9ef3 432 default:
9e1586fc
FB
433 case 'h':
434 help();
c6bdc908 435 return 0;
9e1586fc
FB
436 }
437 }
115329f1 438
36fa9ef3
MR
439 if (optind < argc)
440 test = atoi(argv[optind]);
115329f1 441
f36b3902 442 printf("Libav DCT/IDCT test\n");
9e1586fc 443
a46a3ce4 444 if (test_248_dct) {
7fd2c138 445 idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
9e1586fc 446 } else {
cb44b21d 447 const int cpu_flags = av_get_cpu_flags();
81b9bf31
DB
448 if (test_idct) {
449 for (i = 0; i < FF_ARRAY_ELEMS(idct_tab); i++)
450 err |= dct_error(&idct_tab[i], test, test_idct, speed);
451
452 for (i = 0; idct_tab_arch[i].name; i++)
453 if (!(~cpu_flags & idct_tab_arch[i].cpu_flag))
454 err |= dct_error(&idct_tab_arch[i], test, test_idct, speed);
16b73280
DB
455 }
456#if CONFIG_FDCTDSP
457 else {
81b9bf31
DB
458 for (i = 0; i < FF_ARRAY_ELEMS(fdct_tab); i++)
459 err |= dct_error(&fdct_tab[i], test, test_idct, speed);
460
461 for (i = 0; fdct_tab_arch[i].name; i++)
462 if (!(~cpu_flags & fdct_tab_arch[i].cpu_flag))
463 err |= dct_error(&fdct_tab_arch[i], test, test_idct, speed);
464 }
16b73280 465#endif /* CONFIG_FDCTDSP */
9e1586fc 466 }
dbf396d4 467
5331d2b9
DB
468 if (err)
469 printf("Error: %d.\n", err);
470
471 return !!err;
de6d9b64 472}