FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
dct.c
Go to the documentation of this file.
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 /**
23  * @file
24  * DCT test (c) 2001 Fabrice Bellard
25  * Started from sample code by Juan J. Sierralta P.
26  */
27 
28 #include "config.h"
29 #include <stdlib.h>
30 #include <stdio.h>
31 #include <string.h>
32 #if HAVE_UNISTD_H
33 #include <unistd.h>
34 #endif
35 #include <math.h>
36 
37 #include "libavutil/cpu.h"
38 #include "libavutil/common.h"
39 #include "libavutil/internal.h"
40 #include "libavutil/lfg.h"
41 #include "libavutil/time.h"
42 
43 #include "libavcodec/dct.h"
44 #include "libavcodec/idctdsp.h"
45 #include "libavcodec/simple_idct.h"
46 #include "libavcodec/xvididct.h"
47 #include "libavcodec/aandcttab.h"
48 #include "libavcodec/faandct.h"
49 #include "libavcodec/faanidct.h"
50 #include "libavcodec/dctref.h"
51 
52 struct algo {
53  const char *name;
54  void (*func)(int16_t *block);
56  int cpu_flag;
57  int nonspec;
58 };
59 
60 static const struct algo fdct_tab[] = {
61  { "REF-DBL", ff_ref_fdct, FF_IDCT_PERM_NONE },
62  { "IJG-AAN-INT", ff_fdct_ifast, FF_IDCT_PERM_NONE },
63  { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, FF_IDCT_PERM_NONE },
64 #if CONFIG_FAANDCT
65  { "FAAN", ff_faandct, FF_IDCT_PERM_NONE },
66 #endif /* CONFIG_FAANDCT */
67 };
68 
69 static void ff_prores_idct_wrap(int16_t *dst){
70  LOCAL_ALIGNED(16, int16_t, qmat, [64]);
71  int i;
72 
73  for(i=0; i<64; i++){
74  qmat[i]=4;
75  }
76  ff_prores_idct(dst, qmat);
77  for(i=0; i<64; i++) {
78  dst[i] -= 512;
79  }
80 }
81 
82 static const struct algo idct_tab[] = {
83  { "REF-DBL", ff_ref_idct, FF_IDCT_PERM_NONE },
85  { "SIMPLE-C", ff_simple_idct_8, FF_IDCT_PERM_NONE },
86  { "SIMPLE-C10", ff_simple_idct_10, FF_IDCT_PERM_NONE },
87  { "SIMPLE-C12", ff_simple_idct_12, FF_IDCT_PERM_NONE, 0, 1 },
88  { "PR-C", ff_prores_idct_wrap, FF_IDCT_PERM_NONE, 0, 1 },
89 #if CONFIG_FAANIDCT
90  { "FAANI", ff_faanidct, FF_IDCT_PERM_NONE },
91 #endif /* CONFIG_FAANIDCT */
92 #if CONFIG_MPEG4_DECODER
93  { "XVID", ff_xvid_idct, FF_IDCT_PERM_NONE, 0, 1 },
94 #endif /* CONFIG_MPEG4_DECODER */
95 };
96 
97 #if ARCH_ARM
98 #include "arm/dct.c"
99 #elif ARCH_PPC
100 #include "ppc/dct.c"
101 #elif ARCH_X86
102 #include "x86/dct.c"
103 #else
104 static const struct algo fdct_tab_arch[] = { { 0 } };
105 static const struct algo idct_tab_arch[] = { { 0 } };
106 #endif
107 
108 #define AANSCALE_BITS 12
109 
110 #define NB_ITS 20000
111 #define NB_ITS_SPEED 50000
112 
113 DECLARE_ALIGNED(16, static int16_t, block)[64];
114 DECLARE_ALIGNED(8, static int16_t, block1)[64];
115 
116 static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng, int vals)
117 {
118  int i, j;
119 
120  memset(block, 0, 64 * sizeof(*block));
121 
122  switch (test) {
123  case 0:
124  for (i = 0; i < 64; i++)
125  block[i] = (av_lfg_get(prng) % (2*vals)) -vals;
126  if (is_idct) {
127  ff_ref_fdct(block);
128  for (i = 0; i < 64; i++)
129  block[i] >>= 3;
130  }
131  break;
132  case 1:
133  j = av_lfg_get(prng) % 10 + 1;
134  for (i = 0; i < j; i++) {
135  int idx = av_lfg_get(prng) % 64;
136  block[idx] = av_lfg_get(prng) % (2*vals) -vals;
137  }
138  break;
139  case 2:
140  block[ 0] = av_lfg_get(prng) % (16*vals) - (8*vals);
141  block[63] = (block[0] & 1) ^ 1;
142  break;
143  }
144 }
145 
146 static void permute(int16_t dst[64], const int16_t src[64],
148 {
149  int i;
150 
151 #if ARCH_X86
152  if (permute_x86(dst, src, perm_type))
153  return;
154 #endif
155 
156  switch (perm_type) {
158  for (i = 0; i < 64; i++)
159  dst[(i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2)] = src[i];
160  break;
162  for (i = 0; i < 64; i++)
163  dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
164  break;
166  for (i = 0; i < 64; i++)
167  dst[(i>>3) | ((i<<3)&0x38)] = src[i];
168  break;
169  default:
170  for (i = 0; i < 64; i++)
171  dst[i] = src[i];
172  break;
173  }
174 }
175 
176 static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits)
177 {
178  void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
179  int it, i, scale;
180  int err_inf, v;
181  int64_t err2, ti, ti1, it1, err_sum = 0;
182  int64_t sysErr[64], sysErrMax = 0;
183  int maxout = 0;
184  int blockSumErrMax = 0, blockSumErr;
185  AVLFG prng;
186  const int vals=1<<bits;
187  double omse, ome;
188  int spec_err;
189 
190  av_lfg_init(&prng, 1);
191 
192  err_inf = 0;
193  err2 = 0;
194  for (i = 0; i < 64; i++)
195  sysErr[i] = 0;
196  for (it = 0; it < NB_ITS; it++) {
197  init_block(block1, test, is_idct, &prng, vals);
198  permute(block, block1, dct->perm_type);
199 
200  dct->func(block);
201  emms_c();
202 
203  if (!strcmp(dct->name, "IJG-AAN-INT")) {
204  for (i = 0; i < 64; i++) {
205  scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
206  block[i] = (block[i] * scale) >> AANSCALE_BITS;
207  }
208  }
209 
210  ref(block1);
211  if (!strcmp(dct->name, "PR-SSE2"))
212  for (i = 0; i < 64; i++)
213  block1[i] = av_clip(block1[i], 4-512, 1019-512);
214 
215  blockSumErr = 0;
216  for (i = 0; i < 64; i++) {
217  int err = block[i] - block1[i];
218  err_sum += err;
219  v = abs(err);
220  if (v > err_inf)
221  err_inf = v;
222  err2 += v * v;
223  sysErr[i] += block[i] - block1[i];
224  blockSumErr += v;
225  if (abs(block[i]) > maxout)
226  maxout = abs(block[i]);
227  }
228  if (blockSumErrMax < blockSumErr)
229  blockSumErrMax = blockSumErr;
230  }
231  for (i = 0; i < 64; i++)
232  sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
233 
234  for (i = 0; i < 64; i++) {
235  if (i % 8 == 0)
236  printf("\n");
237  printf("%7d ", (int) sysErr[i]);
238  }
239  printf("\n");
240 
241  omse = (double) err2 / NB_ITS / 64;
242  ome = (double) err_sum / NB_ITS / 64;
243 
244  spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
245 
246  printf("%s %s: max_err=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
247  is_idct ? "IDCT" : "DCT", dct->name, err_inf,
248  omse, ome, (double) sysErrMax / NB_ITS,
249  maxout, blockSumErrMax);
250 
251  if (spec_err && !dct->nonspec) {
252  printf("Failed!\n");
253  return 1;
254  }
255 
256  if (!speed)
257  return 0;
258 
259  /* speed test */
260 
261  init_block(block, test, is_idct, &prng, vals);
262  permute(block1, block, dct->perm_type);
263 
264  ti = av_gettime_relative();
265  it1 = 0;
266  do {
267  for (it = 0; it < NB_ITS_SPEED; it++) {
268  memcpy(block, block1, sizeof(block));
269  dct->func(block);
270  }
271  emms_c();
272  it1 += NB_ITS_SPEED;
273  ti1 = av_gettime_relative() - ti;
274  } while (ti1 < 1000000);
275 
276  printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
277  (double) it1 * 1000.0 / (double) ti1);
278 
279  return 0;
280 }
281 
284 
285 static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
286 {
287  static int init;
288  static double c8[8][8];
289  static double c4[4][4];
290  double block1[64], block2[64], block3[64];
291  double s, sum, v;
292  int i, j, k;
293 
294  if (!init) {
295  init = 1;
296 
297  for (i = 0; i < 8; i++) {
298  sum = 0;
299  for (j = 0; j < 8; j++) {
300  s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
301  c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
302  sum += c8[i][j] * c8[i][j];
303  }
304  }
305 
306  for (i = 0; i < 4; i++) {
307  sum = 0;
308  for (j = 0; j < 4; j++) {
309  s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
310  c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
311  sum += c4[i][j] * c4[i][j];
312  }
313  }
314  }
315 
316  /* butterfly */
317  s = 0.5 * sqrt(2.0);
318  for (i = 0; i < 4; i++) {
319  for (j = 0; j < 8; j++) {
320  block1[8 * (2 * i) + j] =
321  (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
322  block1[8 * (2 * i + 1) + j] =
323  (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
324  }
325  }
326 
327  /* idct8 on lines */
328  for (i = 0; i < 8; i++) {
329  for (j = 0; j < 8; j++) {
330  sum = 0;
331  for (k = 0; k < 8; k++)
332  sum += c8[k][j] * block1[8 * i + k];
333  block2[8 * i + j] = sum;
334  }
335  }
336 
337  /* idct4 */
338  for (i = 0; i < 8; i++) {
339  for (j = 0; j < 4; j++) {
340  /* top */
341  sum = 0;
342  for (k = 0; k < 4; k++)
343  sum += c4[k][j] * block2[8 * (2 * k) + i];
344  block3[8 * (2 * j) + i] = sum;
345 
346  /* bottom */
347  sum = 0;
348  for (k = 0; k < 4; k++)
349  sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
350  block3[8 * (2 * j + 1) + i] = sum;
351  }
352  }
353 
354  /* clamp and store the result */
355  for (i = 0; i < 8; i++) {
356  for (j = 0; j < 8; j++) {
357  v = block3[8 * i + j];
358  if (v < 0) v = 0;
359  else if (v > 255) v = 255;
360  dest[i * linesize + j] = (int) rint(v);
361  }
362  }
363 }
364 
365 static void idct248_error(const char *name,
366  void (*idct248_put)(uint8_t *dest, int line_size,
367  int16_t *block),
368  int speed)
369 {
370  int it, i, it1, ti, ti1, err_max, v;
371  AVLFG prng;
372 
373  av_lfg_init(&prng, 1);
374 
375  /* just one test to see if code is correct (precision is less
376  important here) */
377  err_max = 0;
378  for (it = 0; it < NB_ITS; it++) {
379  /* XXX: use forward transform to generate values */
380  for (i = 0; i < 64; i++)
381  block1[i] = av_lfg_get(&prng) % 256 - 128;
382  block1[0] += 1024;
383 
384  for (i = 0; i < 64; i++)
385  block[i] = block1[i];
386  idct248_ref(img_dest1, 8, block);
387 
388  for (i = 0; i < 64; i++)
389  block[i] = block1[i];
390  idct248_put(img_dest, 8, block);
391 
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 
409  printf("out=\n");
410  for(i=0;i<8;i++) {
411  int j;
412  for(j=0;j<8;j++) {
413  printf(" %3d", img_dest[i*8+j]);
414  }
415  printf("\n");
416  }
417 #endif
418  }
419  printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
420 
421  if (!speed)
422  return;
423 
424  ti = av_gettime_relative();
425  it1 = 0;
426  do {
427  for (it = 0; it < NB_ITS_SPEED; it++) {
428  for (i = 0; i < 64; i++)
429  block[i] = block1[i];
430  idct248_put(img_dest, 8, block);
431  }
432  emms_c();
433  it1 += NB_ITS_SPEED;
434  ti1 = av_gettime_relative() - ti;
435  } while (ti1 < 1000000);
436 
437  printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
438  (double) it1 * 1000.0 / (double) ti1);
439 }
440 
441 static void help(void)
442 {
443  printf("dct-test [-i] [<test-number>] [<bits>]\n"
444  "test-number 0 -> test with random matrixes\n"
445  " 1 -> test with random sparse matrixes\n"
446  " 2 -> do 3. test from MPEG-4 std\n"
447  "bits Number of time domain bits to use, 8 is default\n"
448  "-i test IDCT implementations\n"
449  "-4 test IDCT248 implementations\n"
450  "-t speed test\n");
451 }
452 
453 #if !HAVE_GETOPT
454 #include "compat/getopt.c"
455 #endif
456 
457 int main(int argc, char **argv)
458 {
459  int test_idct = 0, test_248_dct = 0;
460  int c, i;
461  int test = 1;
462  int speed = 0;
463  int err = 0;
464  int bits=8;
465 
466  ff_ref_dct_init();
467 
468  for (;;) {
469  c = getopt(argc, argv, "ih4t");
470  if (c == -1)
471  break;
472  switch (c) {
473  case 'i':
474  test_idct = 1;
475  break;
476  case '4':
477  test_248_dct = 1;
478  break;
479  case 't':
480  speed = 1;
481  break;
482  default:
483  case 'h':
484  help();
485  return 0;
486  }
487  }
488 
489  if (optind < argc)
490  test = atoi(argv[optind]);
491  if(optind+1 < argc) bits= atoi(argv[optind+1]);
492 
493  printf("ffmpeg DCT/IDCT test\n");
494 
495  if (test_248_dct) {
496  idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
497  } else {
498  const int cpu_flags = av_get_cpu_flags();
499  if (test_idct) {
500  for (i = 0; i < FF_ARRAY_ELEMS(idct_tab); i++)
501  err |= dct_error(&idct_tab[i], test, test_idct, speed, bits);
502 
503  for (i = 0; idct_tab_arch[i].name; i++)
504  if (!(~cpu_flags & idct_tab_arch[i].cpu_flag))
505  err |= dct_error(&idct_tab_arch[i], test, test_idct, speed, bits);
506  }
507 #if CONFIG_FDCTDSP
508  else {
509  for (i = 0; i < FF_ARRAY_ELEMS(fdct_tab); i++)
510  err |= dct_error(&fdct_tab[i], test, test_idct, speed, bits);
511 
512  for (i = 0; fdct_tab_arch[i].name; i++)
513  if (!(~cpu_flags & fdct_tab_arch[i].cpu_flag))
514  err |= dct_error(&fdct_tab_arch[i], test, test_idct, speed, bits);
515  }
516 #endif /* CONFIG_FDCTDSP */
517  }
518 
519  if (err)
520  printf("Error: %d.\n", err);
521 
522  return !!err;
523 }
static const struct algo idct_tab_arch[]
Definition: dct.c:105
Definition: lfg.h:25
idct_permutation_type
Definition: idctdsp.h:35
const char * s
Definition: avisynth_c.h:768
#define AANSCALE_BITS
Definition: dct.c:108
void ff_fdct_ifast(int16_t *data)
Definition: jfdctfst.c:208
static int optind
Definition: getopt.c:37
static av_cold int init(AVCodecContext *avctx)
Definition: avrndec.c:35
const char * name
Definition: dct.c:53
static void ff_prores_idct_wrap(int16_t *dst)
Definition: dct.c:69
av_cold void ff_ref_dct_init(void)
Initialize the double precision discrete cosine transform functions fdct & idct.
Definition: dctref.c:41
static int16_t block[64]
Definition: dct.c:113
int nonspec
Definition: dct.c:57
uint8_t bits
Definition: crc.c:296
uint8_t
void ff_faanidct(int16_t block[64])
Definition: faanidct.c:134
static int permute_x86(int16_t dst[64], const int16_t src[64], enum idct_permutation_type perm_type)
Definition: dct.c:116
int cpu_flag
Definition: dct.c:56
void ff_simple_idct248_put(uint8_t *dest, int line_size, int16_t *block)
Definition: simple_idct.c:97
#define DECLARE_ALIGNED(n, t, v)
Declare a variable that is aligned in memory.
Definition: mem.h:101
const uint16_t ff_aanscales[64]
Definition: aandcttab.c:26
static void idct248_error(const char *name, void(*idct248_put)(uint8_t *dest, int line_size, int16_t *block), int speed)
Definition: dct.c:365
void ff_prores_idct(int16_t *block, const int16_t *qmat)
Special version of ff_simple_idct_10() which does dequantization and scales by a factor of 2 more bet...
Definition: simple_idct.c:230
static void help(void)
Definition: dct.c:441
static uint8_t img_dest1[64]
Definition: dct.c:283
#define FFMAX(a, b)
Definition: common.h:94
common internal API header
Definition: dct.c:52
int main(int argc, char **argv)
Definition: dct.c:457
static void permute(int16_t dst[64], const int16_t src[64], enum idct_permutation_type perm_type)
Definition: dct.c:146
typedef void(APIENTRY *FF_PFNGLACTIVETEXTUREPROC)(GLenum texture)
static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng, int vals)
Definition: dct.c:116
static const struct algo idct_tab[]
Definition: dct.c:82
#define FFABS(a)
Absolute value, Note, INT_MIN / INT64_MIN result in undefined behavior as they are not representable ...
Definition: common.h:72
static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits)
Definition: dct.c:176
#define src
Definition: vp9dsp.c:530
static void idct248_ref(uint8_t *dest, int linesize, int16_t *block)
Definition: dct.c:285
void ff_simple_idct_10(int16_t *block)
#define FF_ARRAY_ELEMS(a)
#define rint
Definition: tablegen.h:41
int cpu_flag
Definition: checkasm.c:189
void ff_jpeg_fdct_islow_8(int16_t *data)
void ff_j_rev_dct(int16_t *data)
void ff_faandct(int16_t *data)
Definition: faandct.c:122
static void test(const char *pattern, const char *host)
Definition: noproxy.c:23
void ff_simple_idct_12(int16_t *block)
static int getopt(int argc, char *argv[], char *opts)
Definition: getopt.c:41
static unsigned int av_lfg_get(AVLFG *c)
Get the next random unsigned 32-bit number using an ALFG.
Definition: lfg.h:38
static const struct algo fdct_tab_arch[]
Definition: dct.c:104
void ff_ref_fdct(short *block)
Transform 8x8 block of data with a double precision forward DCT This is a reference implementation...
Definition: dctref.c:59
av_cold void av_lfg_init(AVLFG *c, unsigned int seed)
Definition: lfg.c:30
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:76
void ff_xvid_idct(int16_t *const in)
Definition: xvididct.c:291
int64_t av_gettime_relative(void)
Get the current time in microseconds since some unspecified starting point.
Definition: time.c:56
static uint8_t img_dest[64]
Definition: dct.c:282
AAN (Arai, Agui and Nakajima) (I)DCT tables.
common internal and external API header
static int ref[MAX_W *MAX_W]
Definition: jpeg2000dwt.c:107
#define LOCAL_ALIGNED(a, t, v,...)
Definition: internal.h:110
static double c[64]
#define NB_ITS_SPEED
Definition: dct.c:111
enum idct_permutation_type perm_type
Definition: dct.c:55
void ff_ref_idct(short *block)
Transform 8x8 block of data with a double precision inverse DCT This is a reference implementation...
Definition: dctref.c:95
simple idct header.
static int16_t block1[64]
Definition: dct.c:114
#define M_PI
Definition: mathematics.h:52
Floating point AAN DCT
void(* func)(int16_t *block)
Definition: dct.c:54
static const struct algo fdct_tab[]
Definition: dct.c:60
#define NB_ITS
Definition: dct.c:110
void ff_simple_idct_8(int16_t *block)
const char * name
Definition: opengl_enc.c:103