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_AARCH64
98 #include "aarch64/dct.c"
99 #elif ARCH_ARM
100 #include "arm/dct.c"
101 #elif ARCH_PPC
102 #include "ppc/dct.c"
103 #elif ARCH_X86
104 #include "x86/dct.c"
105 #else
106 static const struct algo fdct_tab_arch[] = { { 0 } };
107 static const struct algo idct_tab_arch[] = { { 0 } };
108 #endif
109 
110 #define AANSCALE_BITS 12
111 
112 #define NB_ITS 20000
113 #define NB_ITS_SPEED 50000
114 
115 DECLARE_ALIGNED(16, static int16_t, block)[64];
116 DECLARE_ALIGNED(8, static int16_t, block1)[64];
117 
118 static void init_block(int16_t block[64], int test, int is_idct, AVLFG *prng, int vals)
119 {
120  int i, j;
121 
122  memset(block, 0, 64 * sizeof(*block));
123 
124  switch (test) {
125  case 0:
126  for (i = 0; i < 64; i++)
127  block[i] = (av_lfg_get(prng) % (2*vals)) -vals;
128  if (is_idct) {
129  ff_ref_fdct(block);
130  for (i = 0; i < 64; i++)
131  block[i] >>= 3;
132  }
133  break;
134  case 1:
135  j = av_lfg_get(prng) % 10 + 1;
136  for (i = 0; i < j; i++) {
137  int idx = av_lfg_get(prng) % 64;
138  block[idx] = av_lfg_get(prng) % (2*vals) -vals;
139  }
140  break;
141  case 2:
142  block[ 0] = av_lfg_get(prng) % (16*vals) - (8*vals);
143  block[63] = (block[0] & 1) ^ 1;
144  break;
145  }
146 }
147 
148 static void permute(int16_t dst[64], const int16_t src[64],
150 {
151  int i;
152 
153 #if ARCH_X86
154  if (permute_x86(dst, src, perm_type))
155  return;
156 #endif
157 
158  switch (perm_type) {
160  for (i = 0; i < 64; i++)
161  dst[(i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2)] = src[i];
162  break;
164  for (i = 0; i < 64; i++)
165  dst[(i & 0x24) | ((i & 3) << 3) | ((i >> 3) & 3)] = src[i];
166  break;
168  for (i = 0; i < 64; i++)
169  dst[(i>>3) | ((i<<3)&0x38)] = src[i];
170  break;
171  default:
172  for (i = 0; i < 64; i++)
173  dst[i] = src[i];
174  break;
175  }
176 }
177 
178 static int dct_error(const struct algo *dct, int test, int is_idct, int speed, const int bits)
179 {
180  void (*ref)(int16_t *block) = is_idct ? ff_ref_idct : ff_ref_fdct;
181  int it, i, scale;
182  int err_inf, v;
183  int64_t err2, ti, ti1, it1, err_sum = 0;
184  int64_t sysErr[64], sysErrMax = 0;
185  int maxout = 0;
186  int blockSumErrMax = 0, blockSumErr;
187  AVLFG prng;
188  const int vals=1<<bits;
189  double omse, ome;
190  int spec_err;
191 
192  av_lfg_init(&prng, 1);
193 
194  err_inf = 0;
195  err2 = 0;
196  for (i = 0; i < 64; i++)
197  sysErr[i] = 0;
198  for (it = 0; it < NB_ITS; it++) {
199  init_block(block1, test, is_idct, &prng, vals);
200  permute(block, block1, dct->perm_type);
201 
202  dct->func(block);
203  emms_c();
204 
205  if (!strcmp(dct->name, "IJG-AAN-INT")) {
206  for (i = 0; i < 64; i++) {
207  scale = 8 * (1 << (AANSCALE_BITS + 11)) / ff_aanscales[i];
208  block[i] = (block[i] * scale) >> AANSCALE_BITS;
209  }
210  }
211 
212  ref(block1);
213  if (!strcmp(dct->name, "PR-SSE2"))
214  for (i = 0; i < 64; i++)
215  block1[i] = av_clip(block1[i], 4-512, 1019-512);
216 
217  blockSumErr = 0;
218  for (i = 0; i < 64; i++) {
219  int err = block[i] - block1[i];
220  err_sum += err;
221  v = abs(err);
222  if (v > err_inf)
223  err_inf = v;
224  err2 += v * v;
225  sysErr[i] += block[i] - block1[i];
226  blockSumErr += v;
227  if (abs(block[i]) > maxout)
228  maxout = abs(block[i]);
229  }
230  if (blockSumErrMax < blockSumErr)
231  blockSumErrMax = blockSumErr;
232  }
233  for (i = 0; i < 64; i++)
234  sysErrMax = FFMAX(sysErrMax, FFABS(sysErr[i]));
235 
236  for (i = 0; i < 64; i++) {
237  if (i % 8 == 0)
238  printf("\n");
239  printf("%7d ", (int) sysErr[i]);
240  }
241  printf("\n");
242 
243  omse = (double) err2 / NB_ITS / 64;
244  ome = (double) err_sum / NB_ITS / 64;
245 
246  spec_err = is_idct && (err_inf > 1 || omse > 0.02 || fabs(ome) > 0.0015);
247 
248  printf("%s %s: max_err=%d omse=%0.8f ome=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
249  is_idct ? "IDCT" : "DCT", dct->name, err_inf,
250  omse, ome, (double) sysErrMax / NB_ITS,
251  maxout, blockSumErrMax);
252 
253  if (spec_err && !dct->nonspec) {
254  printf("Failed!\n");
255  return 1;
256  }
257 
258  if (!speed)
259  return 0;
260 
261  /* speed test */
262 
263  init_block(block, test, is_idct, &prng, vals);
264  permute(block1, block, dct->perm_type);
265 
266  ti = av_gettime_relative();
267  it1 = 0;
268  do {
269  for (it = 0; it < NB_ITS_SPEED; it++) {
270  memcpy(block, block1, sizeof(block));
271  dct->func(block);
272  }
273  emms_c();
274  it1 += NB_ITS_SPEED;
275  ti1 = av_gettime_relative() - ti;
276  } while (ti1 < 1000000);
277 
278  printf("%s %s: %0.1f kdct/s\n", is_idct ? "IDCT" : "DCT", dct->name,
279  (double) it1 * 1000.0 / (double) ti1);
280 
281  return 0;
282 }
283 
286 
287 static void idct248_ref(uint8_t *dest, ptrdiff_t linesize, int16_t *block)
288 {
289  static int init;
290  static double c8[8][8];
291  static double c4[4][4];
292  double block1[64], block2[64], block3[64];
293  double s, sum, v;
294  int i, j, k;
295 
296  if (!init) {
297  init = 1;
298 
299  for (i = 0; i < 8; i++) {
300  sum = 0;
301  for (j = 0; j < 8; j++) {
302  s = (i == 0) ? sqrt(1.0 / 8.0) : sqrt(1.0 / 4.0);
303  c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
304  sum += c8[i][j] * c8[i][j];
305  }
306  }
307 
308  for (i = 0; i < 4; i++) {
309  sum = 0;
310  for (j = 0; j < 4; j++) {
311  s = (i == 0) ? sqrt(1.0 / 4.0) : sqrt(1.0 / 2.0);
312  c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
313  sum += c4[i][j] * c4[i][j];
314  }
315  }
316  }
317 
318  /* butterfly */
319  s = 0.5 * sqrt(2.0);
320  for (i = 0; i < 4; i++) {
321  for (j = 0; j < 8; j++) {
322  block1[8 * (2 * i) + j] =
323  (block[8 * (2 * i) + j] + block[8 * (2 * i + 1) + j]) * s;
324  block1[8 * (2 * i + 1) + j] =
325  (block[8 * (2 * i) + j] - block[8 * (2 * i + 1) + j]) * s;
326  }
327  }
328 
329  /* idct8 on lines */
330  for (i = 0; i < 8; i++) {
331  for (j = 0; j < 8; j++) {
332  sum = 0;
333  for (k = 0; k < 8; k++)
334  sum += c8[k][j] * block1[8 * i + k];
335  block2[8 * i + j] = sum;
336  }
337  }
338 
339  /* idct4 */
340  for (i = 0; i < 8; i++) {
341  for (j = 0; j < 4; j++) {
342  /* top */
343  sum = 0;
344  for (k = 0; k < 4; k++)
345  sum += c4[k][j] * block2[8 * (2 * k) + i];
346  block3[8 * (2 * j) + i] = sum;
347 
348  /* bottom */
349  sum = 0;
350  for (k = 0; k < 4; k++)
351  sum += c4[k][j] * block2[8 * (2 * k + 1) + i];
352  block3[8 * (2 * j + 1) + i] = sum;
353  }
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 
367 static void idct248_error(const char *name,
368  void (*idct248_put)(uint8_t *dest,
369  ptrdiff_t line_size,
370  int16_t *block),
371  int speed)
372 {
373  int it, i, it1, ti, ti1, err_max, v;
374  AVLFG prng;
375 
376  av_lfg_init(&prng, 1);
377 
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 
387  for (i = 0; i < 64; i++)
388  block[i] = block1[i];
389  idct248_ref(img_dest1, 8, block);
390 
391  for (i = 0; i < 64; i++)
392  block[i] = block1[i];
393  idct248_put(img_dest, 8, block);
394 
395  for (i = 0; i < 64; i++) {
396  v = abs((int) img_dest[i] - (int) img_dest1[i]);
397  if (v == 255)
398  printf("%d %d\n", img_dest[i], img_dest1[i]);
399  if (v > err_max)
400  err_max = v;
401  }
402 #if 0
403  printf("ref=\n");
404  for(i=0;i<8;i++) {
405  int j;
406  for(j=0;j<8;j++) {
407  printf(" %3d", img_dest1[i*8+j]);
408  }
409  printf("\n");
410  }
411 
412  printf("out=\n");
413  for(i=0;i<8;i++) {
414  int j;
415  for(j=0;j<8;j++) {
416  printf(" %3d", img_dest[i*8+j]);
417  }
418  printf("\n");
419  }
420 #endif
421  }
422  printf("%s %s: err_inf=%d\n", 1 ? "IDCT248" : "DCT248", name, err_max);
423 
424  if (!speed)
425  return;
426 
427  ti = av_gettime_relative();
428  it1 = 0;
429  do {
430  for (it = 0; it < NB_ITS_SPEED; it++) {
431  for (i = 0; i < 64; i++)
432  block[i] = block1[i];
433  idct248_put(img_dest, 8, block);
434  }
435  emms_c();
436  it1 += NB_ITS_SPEED;
437  ti1 = av_gettime_relative() - ti;
438  } while (ti1 < 1000000);
439 
440  printf("%s %s: %0.1f kdct/s\n", 1 ? "IDCT248" : "DCT248", name,
441  (double) it1 * 1000.0 / (double) ti1);
442 }
443 
444 static void help(void)
445 {
446  printf("dct-test [-i] [<test-number>] [<bits>]\n"
447  "test-number 0 -> test with random matrixes\n"
448  " 1 -> test with random sparse matrixes\n"
449  " 2 -> do 3. test from MPEG-4 std\n"
450  "bits Number of time domain bits to use, 8 is default\n"
451  "-i test IDCT implementations\n"
452  "-4 test IDCT248 implementations\n"
453  "-t speed test\n");
454 }
455 
456 #if !HAVE_GETOPT
457 #include "compat/getopt.c"
458 #endif
459 
460 int main(int argc, char **argv)
461 {
462  int test_idct = 0, test_248_dct = 0;
463  int c, i;
464  int test = 1;
465  int speed = 0;
466  int err = 0;
467  int bits=8;
468 
469  ff_ref_dct_init();
470 
471  for (;;) {
472  c = getopt(argc, argv, "ih4t");
473  if (c == -1)
474  break;
475  switch (c) {
476  case 'i':
477  test_idct = 1;
478  break;
479  case '4':
480  test_248_dct = 1;
481  break;
482  case 't':
483  speed = 1;
484  break;
485  default:
486  case 'h':
487  help();
488  return 0;
489  }
490  }
491 
492  if (optind < argc)
493  test = atoi(argv[optind]);
494  if(optind+1 < argc) bits= atoi(argv[optind+1]);
495 
496  printf("ffmpeg DCT/IDCT test\n");
497 
498  if (test_248_dct) {
499  idct248_error("SIMPLE-C", ff_simple_idct248_put, speed);
500  } else {
501  const int cpu_flags = av_get_cpu_flags();
502  if (test_idct) {
503  for (i = 0; i < FF_ARRAY_ELEMS(idct_tab); i++)
504  err |= dct_error(&idct_tab[i], test, test_idct, speed, bits);
505 
506  for (i = 0; idct_tab_arch[i].name; i++)
507  if (!(~cpu_flags & idct_tab_arch[i].cpu_flag))
508  err |= dct_error(&idct_tab_arch[i], test, test_idct, speed, bits);
509  }
510 #if CONFIG_FDCTDSP
511  else {
512  for (i = 0; i < FF_ARRAY_ELEMS(fdct_tab); i++)
513  err |= dct_error(&fdct_tab[i], test, test_idct, speed, bits);
514 
515  for (i = 0; fdct_tab_arch[i].name; i++)
516  if (!(~cpu_flags & fdct_tab_arch[i].cpu_flag))
517  err |= dct_error(&fdct_tab_arch[i], test, test_idct, speed, bits);
518  }
519 #endif /* CONFIG_FDCTDSP */
520  }
521 
522  if (err)
523  printf("Error: %d.\n", err);
524 
525  return !!err;
526 }
static const struct algo idct_tab_arch[]
Definition: dct.c:107
Definition: lfg.h:27
idct_permutation_type
Definition: idctdsp.h:37
const char * s
Definition: avisynth_c.h:768
#define AANSCALE_BITS
Definition: dct.c:110
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
static atomic_int cpu_flags
Definition: cpu.c:48
#define src
Definition: vp8dsp.c:254
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:115
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:127
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
#define DECLARE_ALIGNED(n, t, v)
Declare a variable that is aligned in memory.
Definition: mem.h:104
const uint16_t ff_aanscales[64]
Definition: aandcttab.c:26
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 idct248_error(const char *name, void(*idct248_put)(uint8_t *dest, ptrdiff_t line_size, int16_t *block), int speed)
Definition: dct.c:367
static void help(void)
Definition: dct.c:444
static uint8_t img_dest1[64]
Definition: dct.c:285
#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:460
static void permute(int16_t dst[64], const int16_t src[64], enum idct_permutation_type perm_type)
Definition: dct.c:148
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:118
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:178
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:205
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:114
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:47
static const struct algo fdct_tab_arch[]
Definition: dct.c:106
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:32
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:89
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:284
AAN (Arai, Agui and Nakajima) (I)DCT tables.
static void idct248_ref(uint8_t *dest, ptrdiff_t linesize, int16_t *block)
Definition: dct.c:287
int
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:113
void ff_simple_idct248_put(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
Definition: simple_idct.c:97
static double c[64]
#define NB_ITS_SPEED
Definition: dct.c:113
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:116
#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:112
void ff_simple_idct_8(int16_t *block)
const char * name
Definition: opengl_enc.c:103