Split global .gitignore file into per-directory files
[libav.git] / libavcodec / dct-test.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"
294eaa26 39#include "libavutil/lfg.h"
980f81d9 40#include "libavutil/time.h"
de6d9b64 41
5d3d39c7 42#include "dct.h"
e0a2e60c 43#include "idctdsp.h"
86748dbc 44#include "simple_idct.h"
7a1d6ddd 45#include "xvididct.h"
10ac3618 46#include "aandcttab.h"
65e4c8c9 47#include "faandct.h"
6f08c541 48#include "faanidct.h"
6a813295 49#include "dctref.h"
9e1586fc 50
3ac35bdb 51struct algo {
36fa9ef3 52 const char *name;
88bd7fdc 53 void (*func)(int16_t *block);
e0a2e60c 54 enum idct_permutation_type perm_type;
746ad4e0 55 int cpu_flag;
dbf396d4 56 int nonspec;
3ac35bdb
MH
57};
58
1274ea8d 59static const struct algo fdct_tab[] = {
e0a2e60c 60 { "REF-DBL", ff_ref_fdct, FF_IDCT_PERM_NONE },
e0a2e60c
DB
61 { "IJG-AAN-INT", ff_fdct_ifast, FF_IDCT_PERM_NONE },
62 { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, FF_IDCT_PERM_NONE },
1274ea8d
DB
63#if CONFIG_FAANDCT
64 { "FAAN", ff_faandct, FF_IDCT_PERM_NONE },
65#endif /* CONFIG_FAANDCT */
4b357756
MR
66};
67
7a1d6ddd 68static const struct algo idct_tab[] = {
e0a2e60c
DB
69 { "REF-DBL", ff_ref_idct, FF_IDCT_PERM_NONE },
70 { "INT", ff_j_rev_dct, FF_IDCT_PERM_LIBMPEG2 },
71 { "SIMPLE-C", ff_simple_idct_8, FF_IDCT_PERM_NONE },
1274ea8d
DB
72#if CONFIG_FAANIDCT
73 { "FAANI", ff_faanidct, FF_IDCT_PERM_NONE },
74#endif /* CONFIG_FAANIDCT */
7a1d6ddd
PM
75#if CONFIG_MPEG4_DECODER
76 { "XVID", ff_xvid_idct, FF_IDCT_PERM_NONE, 0, 1 },
77#endif /* CONFIG_MPEG4_DECODER */
81b9bf31 78};
3ac35bdb 79
b250f9c6 80#if ARCH_ARM
81b9bf31
DB
81#include "arm/dct-test.c"
82#elif ARCH_PPC
83#include "ppc/dct-test.c"
84#elif ARCH_X86
85#include "x86/dct-test.c"
86#else
d89e58f5
VG
87static const struct algo fdct_tab_arch[] = { { 0 } };
88static const struct algo idct_tab_arch[] = { { 0 } };
479044ce 89#endif
479044ce 90
de6d9b64 91#define AANSCALE_BITS 12
de6d9b64 92
de6d9b64
FB
93#define NB_ITS 20000
94#define NB_ITS_SPEED 50000
95
88bd7fdc
DB
96DECLARE_ALIGNED(16, static int16_t, block)[64];
97DECLARE_ALIGNED(8, static int16_t, block1)[64];
9e1586fc 98
88bd7fdc 99static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng)
ae2e8971
MR
100{
101 int i, j;
102
103 memset(block, 0, 64 * sizeof(*block));
104
105 switch (test) {
106 case 0:
107 for (i = 0; i < 64; i++)
108 block[i] = (av_lfg_get(prng) % 512) - 256;
109 if (is_idct) {
110 ff_ref_fdct(block);
111 for (i = 0; i < 64; i++)
112 block[i] >>= 3;
113 }
114 break;
115 case 1:
116 j = av_lfg_get(prng) % 10 + 1;
117 for (i = 0; i < j; i++)
118 block[av_lfg_get(prng) % 64] = av_lfg_get(prng) % 512 - 256;
119 break;
120 case 2:
121 block[ 0] = av_lfg_get(prng) % 4096 - 2048;
122 block[63] = (block[0] & 1) ^ 1;
123 break;
124 }
125}
126
e0a2e60c
DB
127static void permute(int16_t dst[64], const int16_t src[64],
128 enum idct_permutation_type perm_type)
ae2e8971
MR
129{
130 int i;
131
81b9bf31
DB
132#if ARCH_X86
133 if (permute_x86(dst, src, perm_type))
134 return;
135#endif
136
e0a2e60c
DB
137 switch (perm_type) {
138 case FF_IDCT_PERM_LIBMPEG2:
ae2e8971 139 for (i = 0; i < 64; i++)
913fa85a 140 dst[(i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2)] = src[i];
e0a2e60c 141 break;
e0a2e60c 142 case FF_IDCT_PERM_PARTTRANS:
ae2e8971
MR
143 for (i = 0; i < 64; i++)
144 dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
e0a2e60c
DB
145 break;
146 default:
ae2e8971
MR
147 for (i = 0; i < 64; i++)
148 dst[i] = src[i];
e0a2e60c 149 break;
ae2e8971
MR
150 }
151}
152
dbf396d4 153static int dct_error(const struct algo *dct, int test, int is_idct, int speed)
de6d9b64 154{
88bd7fdc 155 void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
de6d9b64 156 int it, i, scale;
de6d9b64 157 int err_inf, v;
dbf396d4 158 int64_t err2, ti, ti1, it1, err_sum = 0;
36fa9ef3
MR
159 int64_t sysErr[64], sysErrMax = 0;
160 int maxout = 0;
161 int blockSumErrMax = 0, blockSumErr;
64bde197 162 AVLFG prng;
dbf396d4
MR
163 double omse, ome;
164 int spec_err;
de6d9b64 165
64bde197 166 av_lfg_init(&prng, 1);
de6d9b64
FB
167
168 err_inf = 0;
169 err2 = 0;
36fa9ef3
MR
170 for (i = 0; i < 64; i++)
171 sysErr[i] = 0;
172 for (it = 0; it < NB_ITS; it++) {
ae2e8971 173 init_block(block1, test, is_idct, &prng);
e0a2e60c 174 permute(block, block1, dct->perm_type);
9e1586fc 175
4f905a65 176 dct->func(block);
db7d8fb4 177 emms_c();
9e1586fc 178
e0a2e60c 179 if (!strcmp(dct->name, "IJG-AAN-INT")) {
36fa9ef3
MR
180 for (i = 0; i < 64; i++) {
181 scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
182 block[i] = (block[i] * scale) >> AANSCALE_BITS;
86748dbc
MN
183 }
184 }
185
74965f26 186 ref(block1);
de6d9b64 187
36fa9ef3
MR
188 blockSumErr = 0;
189 for (i = 0; i < 64; i++) {
dbf396d4
MR
190 int err = block[i] - block1[i];
191 err_sum += err;
192 v = abs(err);
de6d9b64
FB
193 if (v > err_inf)
194 err_inf = v;
195 err2 += v * v;
bb270c08
DB
196 sysErr[i] += block[i] - block1[i];
197 blockSumErr += v;
36fa9ef3
MR
198 if (abs(block[i]) > maxout)
199 maxout = abs(block[i]);
de6d9b64 200 }
36fa9ef3
MR
201 if (blockSumErrMax < blockSumErr)
202 blockSumErrMax = blockSumErr;
86748dbc 203 }
36fa9ef3
MR
204 for (i = 0; i < 64; i++)
205 sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
115329f1 206
36fa9ef3
MR
207 for (i = 0; i < 64; i++) {
208 if (i % 8 == 0)
209 printf("\n");
210 printf("%7d ", (int) sysErr[i]);
de6d9b64 211 }
86748dbc 212 printf("\n");
115329f1 213
dbf396d4
MR
214 omse = (double) err2 / NB_ITS / 64;
215 ome = (double) err_sum / NB_ITS / 64;
216
217 spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
218
219 printf("%s %s: ppe=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
cf2b4f88 220 is_idct ? "IDCT" : "DCT", dct->name, err_inf,
dbf396d4 221 omse, ome, (double) sysErrMax / NB_ITS,
36fa9ef3 222 maxout, blockSumErrMax);
e6ff0648 223
dbf396d4
MR
224 if (spec_err && !dct->nonspec)
225 return 1;
226
7fd2c138 227 if (!speed)
dbf396d4 228 return 0;
7fd2c138 229
de6d9b64 230 /* speed test */
ae2e8971 231 init_block(block, test, is_idct, &prng);
e0a2e60c 232 permute(block1, block, dct->perm_type);
9e1586fc 233
6fdf35fe 234 ti = av_gettime_relative();
de6d9b64
FB
235 it1 = 0;
236 do {
36fa9ef3 237 for (it = 0; it < NB_ITS_SPEED; it++) {
ae2e8971 238 memcpy(block, block1, sizeof(block));
4f905a65 239 dct->func(block);
de6d9b64
FB
240 }
241 it1 += NB_ITS_SPEED;
6fdf35fe 242 ti1 = av_gettime_relative() - ti;
de6d9b64 243 } while (ti1 < 1000000);
db7d8fb4 244 emms_c();
de6d9b64 245
cf2b4f88 246 printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
36fa9ef3 247 (double) it1 * 1000.0 / (double) ti1);
dbf396d4
MR
248
249 return 0;
de6d9b64
FB
250}
251
c6727809
MR
252DECLARE_ALIGNED(8, static uint8_t, img_dest)[64];
253DECLARE_ALIGNED(8, static uint8_t, img_dest1)[64];
a46a3ce4 254
504ffed1 255static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
a46a3ce4
FB
256{
257 static int init;
258 static double c8[8][8];
259 static double c4[4][4];
260 double block1[64], block2[64], block3[64];
261 double s, sum, v;
262 int i, j, k;
263
264 if (!init) {
265 init = 1;
266
36fa9ef3 267 for (i = 0; i < 8; i++) {
a46a3ce4 268 sum = 0;
36fa9ef3
MR
269 for (j = 0; j < 8; j++) {
270 s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
a46a3ce4
FB
271 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
272 sum += c8[i][j] * c8[i][j];
273 }
274 }
115329f1 275
36fa9ef3 276 for (i = 0; i < 4; i++) {
a46a3ce4 277 sum = 0;
36fa9ef3
MR
278 for (j = 0; j < 4; j++) {
279 s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
a46a3ce4
FB
280 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
281 sum += c4[i][j] * c4[i][j];
282 }
283 }
284 }
285
286 /* butterfly */
652f0197 287 s = 0.5 * sqrt(2.0);
36fa9ef3
MR
288 for (i = 0; i < 4; i++) {
289 for (j = 0; j < 8; j++) {
290 block1[8 * (2 * i) + j] =
291 (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
292 block1[8 * (2 * i + 1) + j] =
293 (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
a46a3ce4
FB
294 }
295 }
296
297 /* idct8 on lines */
36fa9ef3
MR
298 for (i = 0; i < 8; i++) {
299 for (j = 0; j < 8; j++) {
a46a3ce4 300 sum = 0;
36fa9ef3
MR
301 for (k = 0; k < 8; k++)
302 sum += c8[k][j] * block1[8 * i + k];
303 block2[8 * i + j] = sum;
a46a3ce4
FB
304 }
305 }
306
307 /* idct4 */
36fa9ef3
MR
308 for (i = 0; i < 8; i++) {
309 for (j = 0; j < 4; j++) {
a46a3ce4
FB
310 /* top */
311 sum = 0;
36fa9ef3
MR
312 for (k = 0; k < 4; k++)
313 sum += c4[k][j] * block2[8 * (2 * k) + i];
314 block3[8 * (2 * j) + i] = sum;
a46a3ce4
FB
315
316 /* bottom */
317 sum = 0;
36fa9ef3
MR
318 for (k = 0; k < 4; k++)
319 sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
320 block3[8 * (2 * j + 1) + i] = sum;
a46a3ce4
FB
321 }
322 }
323
324 /* clamp and store the result */
36fa9ef3
MR
325 for (i = 0; i < 8; i++) {
326 for (j = 0; j < 8; j++) {
327 v = block3[8 * i + j];
328 if (v < 0) v = 0;
329 else if (v > 255) v = 255;
330 dest[i * linesize + j] = (int) rint(v);
a46a3ce4
FB
331 }
332 }
333}
334
504ffed1 335static void idct248_error(const char *name,
36fa9ef3 336 void (*idct248_put)(uint8_t *dest, int line_size,
7fd2c138
MR
337 int16_t *block),
338 int speed)
a46a3ce4
FB
339{
340 int it, i, it1, ti, ti1, err_max, v;
64bde197 341 AVLFG prng;
294eaa26 342
64bde197 343 av_lfg_init(&prng, 1);
115329f1 344
a46a3ce4
FB
345 /* just one test to see if code is correct (precision is less
346 important here) */
347 err_max = 0;
36fa9ef3 348 for (it = 0; it < NB_ITS; it++) {
652f0197 349 /* XXX: use forward transform to generate values */
36fa9ef3 350 for (i = 0; i < 64; i++)
64bde197 351 block1[i] = av_lfg_get(&prng) % 256 - 128;
652f0197
FB
352 block1[0] += 1024;
353
36fa9ef3
MR
354 for (i = 0; i < 64; i++)
355 block[i] = block1[i];
a46a3ce4 356 idct248_ref(img_dest1, 8, block);
115329f1 357
36fa9ef3
MR
358 for (i = 0; i < 64; i++)
359 block[i] = block1[i];
652f0197 360 idct248_put(img_dest, 8, block);
115329f1 361
36fa9ef3
MR
362 for (i = 0; i < 64; i++) {
363 v = abs((int) img_dest[i] - (int) img_dest1[i]);
652f0197
FB
364 if (v == 255)
365 printf("%d %d\n", img_dest[i], img_dest1[i]);
366 if (v > err_max)
367 err_max = v;
368 }
a46a3ce4 369 }
36fa9ef3 370 printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
a46a3ce4 371
7fd2c138
MR
372 if (!speed)
373 return;
374
6fdf35fe 375 ti = av_gettime_relative();
a46a3ce4
FB
376 it1 = 0;
377 do {
36fa9ef3
MR
378 for (it = 0; it < NB_ITS_SPEED; it++) {
379 for (i = 0; i < 64; i++)
380 block[i] = block1[i];
a46a3ce4
FB
381 idct248_put(img_dest, 8, block);
382 }
383 it1 += NB_ITS_SPEED;
6fdf35fe 384 ti1 = av_gettime_relative() - ti;
a46a3ce4 385 } while (ti1 < 1000000);
db7d8fb4 386 emms_c();
a46a3ce4 387
36fa9ef3
MR
388 printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
389 (double) it1 * 1000.0 / (double) ti1);
a46a3ce4
FB
390}
391
504ffed1 392static void help(void)
9e1586fc 393{
86748dbc
MN
394 printf("dct-test [-i] [<test-number>]\n"
395 "test-number 0 -> test with random matrixes\n"
396 " 1 -> test with random sparse matrixes\n"
41ed7ab4 397 " 2 -> do 3. test from MPEG-4 std\n"
a46a3ce4 398 "-i test IDCT implementations\n"
7fd2c138
MR
399 "-4 test IDCT248 implementations\n"
400 "-t speed test\n");
9e1586fc
FB
401}
402
667fb97a
RB
403#if !HAVE_GETOPT
404#include "compat/getopt.c"
405#endif
406
de6d9b64
FB
407int main(int argc, char **argv)
408{
a46a3ce4 409 int test_idct = 0, test_248_dct = 0;
36fa9ef3
MR
410 int c, i;
411 int test = 1;
7fd2c138 412 int speed = 0;
dbf396d4 413 int err = 0;
36fa9ef3 414
0de74546 415 ff_ref_dct_init();
f67a10cd 416
36fa9ef3 417 for (;;) {
7fd2c138 418 c = getopt(argc, argv, "ih4t");
9e1586fc
FB
419 if (c == -1)
420 break;
36fa9ef3 421 switch (c) {
9e1586fc
FB
422 case 'i':
423 test_idct = 1;
424 break;
a46a3ce4
FB
425 case '4':
426 test_248_dct = 1;
427 break;
7fd2c138
MR
428 case 't':
429 speed = 1;
430 break;
36fa9ef3 431 default:
9e1586fc
FB
432 case 'h':
433 help();
c6bdc908 434 return 0;
9e1586fc
FB
435 }
436 }
115329f1 437
36fa9ef3
MR
438 if (optind < argc)
439 test = atoi(argv[optind]);
115329f1 440
f36b3902 441 printf("Libav DCT/IDCT test\n");
9e1586fc 442
a46a3ce4 443 if (test_248_dct) {
7fd2c138 444 idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
9e1586fc 445 } else {
cb44b21d 446 const int cpu_flags = av_get_cpu_flags();
81b9bf31
DB
447 if (test_idct) {
448 for (i = 0; i < FF_ARRAY_ELEMS(idct_tab); i++)
449 err |= dct_error(&idct_tab[i], test, test_idct, speed);
450
451 for (i = 0; idct_tab_arch[i].name; i++)
452 if (!(~cpu_flags & idct_tab_arch[i].cpu_flag))
453 err |= dct_error(&idct_tab_arch[i], test, test_idct, speed);
16b73280
DB
454 }
455#if CONFIG_FDCTDSP
456 else {
81b9bf31
DB
457 for (i = 0; i < FF_ARRAY_ELEMS(fdct_tab); i++)
458 err |= dct_error(&fdct_tab[i], test, test_idct, speed);
459
460 for (i = 0; fdct_tab_arch[i].name; i++)
461 if (!(~cpu_flags & fdct_tab_arch[i].cpu_flag))
462 err |= dct_error(&fdct_tab_arch[i], test, test_idct, speed);
463 }
16b73280 464#endif /* CONFIG_FDCTDSP */
9e1586fc 465 }
dbf396d4 466
5331d2b9
DB
467 if (err)
468 printf("Error: %d.\n", err);
469
470 return !!err;
de6d9b64 471}