ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
[libav.git] / libavcodec / dct-test.c
CommitLineData
ff4ec49e 1/* DCT test. (c) 2001 Fabrice Bellard.
de6d9b64
FB
2 Started from sample code by Juan J. Sierralta P.
3*/
4#include <stdlib.h>
5#include <stdio.h>
6#include <string.h>
7#include <sys/time.h>
8#include <unistd.h>
9e1586fc 9#include <getopt.h>
de6d9b64
FB
10
11#include "dsputil.h"
12
9e1586fc 13#include "i386/mmx.h"
86748dbc 14#include "simple_idct.h"
9e1586fc
FB
15
16/* reference fdct/idct */
de6d9b64 17extern void fdct(DCTELEM *block);
9e1586fc 18extern void idct(DCTELEM *block);
de6d9b64
FB
19extern void init_fdct();
20
9e1586fc
FB
21extern void j_rev_dct(DCTELEM *data);
22extern void ff_mmx_idct(DCTELEM *data);
23extern void ff_mmxext_idct(DCTELEM *data);
24
86748dbc
MN
25extern void odivx_idct_c (short *block);
26
a46a3ce4
FB
27void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/,
28 UINT8 *pixels/*align 8*/, int line_size);
29
30void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/,
31 UINT8 *pixels/*align 8*/, int line_size);
32
de6d9b64
FB
33#define AANSCALE_BITS 12
34static const unsigned short aanscales[64] = {
35 /* precomputed values scaled up by 14 bits */
36 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
37 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
38 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
39 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
40 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
41 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
42 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
43 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
44};
45
86748dbc
MN
46UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
47
de6d9b64
FB
48INT64 gettime(void)
49{
50 struct timeval tv;
51 gettimeofday(&tv,NULL);
52 return (INT64)tv.tv_sec * 1000000 + tv.tv_usec;
53}
54
55#define NB_ITS 20000
56#define NB_ITS_SPEED 50000
57
9e1586fc
FB
58static short idct_mmx_perm[64];
59
86748dbc
MN
60static short idct_simple_mmx_perm[64]={
61 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
62 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
63 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
64 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
65 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
66 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
67 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
68 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
69};
70
9e1586fc
FB
71void idct_mmx_init(void)
72{
73 int i;
74
75 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
76 for (i = 0; i < 64; i++) {
77 idct_mmx_perm[i] = (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
86748dbc 78// idct_simple_mmx_perm[i] = simple_block_permute_op(i);
9e1586fc
FB
79 }
80}
81
82static DCTELEM block[64] __attribute__ ((aligned (8)));
83static DCTELEM block1[64] __attribute__ ((aligned (8)));
86748dbc 84static DCTELEM block_org[64] __attribute__ ((aligned (8)));
9e1586fc
FB
85
86void dct_error(const char *name, int is_idct,
87 void (*fdct_func)(DCTELEM *block),
86748dbc 88 void (*fdct_ref)(DCTELEM *block), int test)
de6d9b64
FB
89{
90 int it, i, scale;
de6d9b64
FB
91 int err_inf, v;
92 INT64 err2, ti, ti1, it1;
86748dbc
MN
93 INT64 sysErr[64], sysErrMax=0;
94 int maxout=0;
86748dbc 95 int blockSumErrMax=0, blockSumErr;
de6d9b64
FB
96
97 srandom(0);
98
99 err_inf = 0;
100 err2 = 0;
86748dbc 101 for(i=0; i<64; i++) sysErr[i]=0;
de6d9b64 102 for(it=0;it<NB_ITS;it++) {
86748dbc
MN
103 for(i=0;i<64;i++)
104 block1[i] = 0;
105 switch(test){
106 case 0:
107 for(i=0;i<64;i++)
108 block1[i] = (random() % 512) -256;
ad324c93 109 if (is_idct){
86748dbc 110 fdct(block1);
ad324c93
MN
111
112 for(i=0;i<64;i++)
113 block1[i]>>=3;
114 }
86748dbc
MN
115 break;
116 case 1:{
117 int num= (random()%10)+1;
118 for(i=0;i<num;i++)
119 block1[random()%64] = (random() % 512) -256;
120 }break;
121 case 2:
122 block1[0]= (random()%4096)-2048;
123 block1[63]= (block1[0]&1)^1;
124 break;
125 }
9e1586fc 126
86748dbc
MN
127#if 0 // simulate mismatch control
128{ int sum=0;
129 for(i=0;i<64;i++)
130 sum+=block1[i];
131
132 if((sum&1)==0) block1[63]^=1;
133}
134#endif
135
136 for(i=0; i<64; i++)
137 block_org[i]= block1[i];
9e1586fc
FB
138
139 if (fdct_func == ff_mmx_idct ||
86748dbc
MN
140 fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
141 for(i=0;i<64;i++)
9e1586fc 142 block[idct_mmx_perm[i]] = block1[i];
a46a3ce4 143 } else if(fdct_func == ff_simple_idct_mmx ) {
86748dbc
MN
144 for(i=0;i<64;i++)
145 block[idct_simple_mmx_perm[i]] = block1[i];
146
147 } else {
148 for(i=0; i<64; i++)
149 block[i]= block1[i];
9e1586fc 150 }
86748dbc
MN
151#if 0 // simulate mismatch control for tested IDCT but not the ref
152{ int sum=0;
153 for(i=0;i<64;i++)
154 sum+=block[i];
155
156 if((sum&1)==0) block[63]^=1;
157}
158#endif
9e1586fc 159
de6d9b64 160 fdct_func(block);
9e1586fc
FB
161 emms(); /* for ff_mmx_idct */
162
03c94ede 163 if (fdct_func == fdct_ifast) {
de6d9b64 164 for(i=0; i<64; i++) {
ad324c93 165 scale = 8*(1 << (AANSCALE_BITS + 11)) / aanscales[i];
86748dbc
MN
166 block[i] = (block[i] * scale /*+ (1<<(AANSCALE_BITS-1))*/) >> AANSCALE_BITS;
167 }
168 }
169
9e1586fc 170 fdct_ref(block1);
de6d9b64 171
86748dbc 172 blockSumErr=0;
de6d9b64
FB
173 for(i=0;i<64;i++) {
174 v = abs(block[i] - block1[i]);
175 if (v > err_inf)
176 err_inf = v;
177 err2 += v * v;
86748dbc
MN
178 sysErr[i] += block[i] - block1[i];
179 blockSumErr += v;
180 if( abs(block[i])>maxout) maxout=abs(block[i]);
de6d9b64 181 }
86748dbc
MN
182 if(blockSumErrMax < blockSumErr) blockSumErrMax= blockSumErr;
183#if 0 // print different matrix pairs
184 if(blockSumErr){
185 printf("\n");
186 for(i=0; i<64; i++){
187 if((i&7)==0) printf("\n");
188 printf("%4d ", block_org[i]);
189 }
190 for(i=0; i<64; i++){
191 if((i&7)==0) printf("\n");
192 printf("%4d ", block[i] - block1[i]);
193 }
194 }
195#endif
196 }
197 for(i=0; i<64; i++) sysErrMax= MAX(sysErrMax, ABS(sysErr[i]));
198
199#if 1 // dump systematic errors
200 for(i=0; i<64; i++){
201 if(i%8==0) printf("\n");
202 printf("%5d ", (int)sysErr[i]);
de6d9b64 203 }
86748dbc
MN
204 printf("\n");
205#endif
206
207 printf("%s %s: err_inf=%d err2=%0.8f syserr=%0.8f maxout=%d blockSumErr=%d\n",
9e1586fc 208 is_idct ? "IDCT" : "DCT",
86748dbc
MN
209 name, err_inf, (double)err2 / NB_ITS / 64.0, (double)sysErrMax / NB_ITS, maxout, blockSumErrMax);
210#if 1 //Speed test
de6d9b64 211 /* speed test */
86748dbc
MN
212 for(i=0;i<64;i++)
213 block1[i] = 0;
214 switch(test){
215 case 0:
216 for(i=0;i<64;i++)
217 block1[i] = (random() % 512) -256;
ad324c93 218 if (is_idct){
86748dbc 219 fdct(block1);
ad324c93
MN
220
221 for(i=0;i<64;i++)
222 block1[i]>>=3;
223 }
86748dbc
MN
224 break;
225 case 1:{
226 case 2:
227 block1[0] = (random() % 512) -256;
228 block1[1] = (random() % 512) -256;
229 block1[2] = (random() % 512) -256;
230 block1[3] = (random() % 512) -256;
231 }break;
232 }
de6d9b64 233
9e1586fc 234 if (fdct_func == ff_mmx_idct ||
86748dbc
MN
235 fdct_func == j_rev_dct || fdct_func == ff_mmxext_idct) {
236 for(i=0;i<64;i++)
9e1586fc 237 block[idct_mmx_perm[i]] = block1[i];
a46a3ce4 238 } else if(fdct_func == ff_simple_idct_mmx ) {
86748dbc
MN
239 for(i=0;i<64;i++)
240 block[idct_simple_mmx_perm[i]] = block1[i];
241 } else {
242 for(i=0; i<64; i++)
243 block[i]= block1[i];
9e1586fc
FB
244 }
245
de6d9b64
FB
246 ti = gettime();
247 it1 = 0;
248 do {
249 for(it=0;it<NB_ITS_SPEED;it++) {
86748dbc
MN
250 for(i=0; i<64; i++)
251 block[i]= block1[i];
252// memcpy(block, block1, sizeof(DCTELEM) * 64);
253// dont memcpy especially not fastmemcpy because it does movntq !!!
de6d9b64
FB
254 fdct_func(block);
255 }
256 it1 += NB_ITS_SPEED;
257 ti1 = gettime() - ti;
258 } while (ti1 < 1000000);
9e1586fc 259 emms();
de6d9b64 260
86748dbc 261 printf("%s %s: %0.1f kdct/s\n",
9e1586fc 262 is_idct ? "IDCT" : "DCT",
de6d9b64 263 name, (double)it1 * 1000.0 / (double)ti1);
86748dbc 264#endif
de6d9b64
FB
265}
266
a46a3ce4
FB
267static UINT8 img_dest[64] __attribute__ ((aligned (8)));
268static UINT8 img_dest1[64] __attribute__ ((aligned (8)));
269
270void idct248_ref(UINT8 *dest, int linesize, INT16 *block)
271{
272 static int init;
273 static double c8[8][8];
274 static double c4[4][4];
275 double block1[64], block2[64], block3[64];
276 double s, sum, v;
277 int i, j, k;
278
279 if (!init) {
280 init = 1;
281
282 for(i=0;i<8;i++) {
283 sum = 0;
284 for(j=0;j<8;j++) {
285 s = (i==0) ? sqrt(1.0/8.0) : sqrt(1.0/4.0);
286 c8[i][j] = s * cos(M_PI * i * (j + 0.5) / 8.0);
287 sum += c8[i][j] * c8[i][j];
288 }
289 }
290
291 for(i=0;i<4;i++) {
292 sum = 0;
293 for(j=0;j<4;j++) {
294 s = (i==0) ? sqrt(1.0/4.0) : sqrt(1.0/2.0);
295 c4[i][j] = s * cos(M_PI * i * (j + 0.5) / 4.0);
296 sum += c4[i][j] * c4[i][j];
297 }
298 }
299 }
300
301 /* butterfly */
302 for(i=0;i<4;i++) {
303 for(j=0;j<8;j++) {
304 block1[8*(2*i)+j] = (block[8*(2*i)+j] + block[8*(2*i+1)+j]) * 0.5;
305 block1[8*(2*i+1)+j] = (block[8*(2*i)+j] - block[8*(2*i+1)+j]) * 0.5;
306 }
307 }
308
309 /* idct8 on lines */
310 for(i=0;i<8;i++) {
311 for(j=0;j<8;j++) {
312 sum = 0;
313 for(k=0;k<8;k++)
314 sum += c8[k][j] * block1[8*i+k];
315 block2[8*i+j] = sum;
316 }
317 }
318
319 /* idct4 */
320 for(i=0;i<8;i++) {
321 for(j=0;j<4;j++) {
322 /* top */
323 sum = 0;
324 for(k=0;k<4;k++)
325 sum += c4[k][j] * block2[8*(2*k)+i];
326 block3[8*(2*j)+i] = sum;
327
328 /* bottom */
329 sum = 0;
330 for(k=0;k<4;k++)
331 sum += c4[k][j] * block2[8*(2*k+1)+i];
332 block3[8*(2*j+1)+i] = sum;
333 }
334 }
335
336 /* clamp and store the result */
337 for(i=0;i<8;i++) {
338 for(j=0;j<8;j++) {
339 v = block3[8*i+j] + 128.0;
340 if (v < 0)
341 v = 0;
342 else if (v > 255)
343 v = 255;
344 dest[i * linesize + j] = (int)rint(v);
345 }
346 }
347}
348
349void idct248_error(const char *name,
350 void (*idct248_put)(UINT8 *dest, int line_size, INT16 *block))
351{
352 int it, i, it1, ti, ti1, err_max, v;
353
354 srandom(0);
355
356 /* just one test to see if code is correct (precision is less
357 important here) */
358 err_max = 0;
359 for(it=0;it<NB_ITS;it++) {
360 for(i=0;i<64;i++)
361 block1[i] = (random() % 512) - 256;
362
363 for(i=0; i<64; i++)
364 block[i]= block1[i];
365 idct248_ref(img_dest1, 8, block);
366
367#if 0
368 printf("ref=\n");
369 for(i=0;i<8;i++) {
370 int j;
371 for(j=0;j<8;j++) {
372 printf(" %3d", img_dest1[i*8+j]);
373 }
374 printf("\n");
375 }
376#endif
377
378 for(i=0; i<64; i++)
379 block[i]= block1[i];
380 idct248_put(img_dest, 8, block);
381
382#if 0
383 printf("out=\n");
384 for(i=0;i<8;i++) {
385 int j;
386 for(j=0;j<8;j++) {
387 printf(" %3d", img_dest[i*8+j]);
388 }
389 printf("\n");
390 }
391#endif
392 for(i=0;i<64;i++) {
393 v = abs(img_dest[i] - img_dest1[i]);
394 if (v > err_max)
395 err_max = v;
396 }
397 }
398 printf("%s %s: err_inf=%d\n",
399 1 ? "IDCT248" : "DCT248",
400 name, err_max);
401
402 ti = gettime();
403 it1 = 0;
404 do {
405 for(it=0;it<NB_ITS_SPEED;it++) {
406 for(i=0; i<64; i++)
407 block[i]= block1[i];
408// memcpy(block, block1, sizeof(DCTELEM) * 64);
409// dont memcpy especially not fastmemcpy because it does movntq !!!
410 idct248_put(img_dest, 8, block);
411 }
412 it1 += NB_ITS_SPEED;
413 ti1 = gettime() - ti;
414 } while (ti1 < 1000000);
415 emms();
416
417 printf("%s %s: %0.1f kdct/s\n",
418 1 ? "IDCT248" : "DCT248",
419 name, (double)it1 * 1000.0 / (double)ti1);
420}
421
9e1586fc
FB
422void help(void)
423{
86748dbc
MN
424 printf("dct-test [-i] [<test-number>]\n"
425 "test-number 0 -> test with random matrixes\n"
426 " 1 -> test with random sparse matrixes\n"
427 " 2 -> do 3. test from mpeg4 std\n"
a46a3ce4
FB
428 "-i test IDCT implementations\n"
429 "-4 test IDCT248 implementations\n");
9e1586fc
FB
430 exit(1);
431}
432
de6d9b64
FB
433int main(int argc, char **argv)
434{
a46a3ce4 435 int test_idct = 0, test_248_dct = 0;
86748dbc
MN
436 int c,i;
437 int test=1;
9e1586fc 438
de6d9b64 439 init_fdct();
9e1586fc 440 idct_mmx_init();
de6d9b64 441
86748dbc
MN
442 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
443 for(i=0;i<MAX_NEG_CROP;i++) {
444 cropTbl[i] = 0;
445 cropTbl[i + MAX_NEG_CROP + 256] = 255;
446 }
447
9e1586fc 448 for(;;) {
a46a3ce4 449 c = getopt(argc, argv, "ih4");
9e1586fc
FB
450 if (c == -1)
451 break;
452 switch(c) {
453 case 'i':
454 test_idct = 1;
455 break;
a46a3ce4
FB
456 case '4':
457 test_248_dct = 1;
458 break;
86748dbc 459 default :
9e1586fc
FB
460 case 'h':
461 help();
462 break;
463 }
464 }
86748dbc
MN
465
466 if(optind <argc) test= atoi(argv[optind]);
9e1586fc
FB
467
468 printf("ffmpeg DCT/IDCT test\n");
469
a46a3ce4
FB
470 if (test_248_dct) {
471 idct248_error("SIMPLE-C", simple_idct248_put);
9e1586fc 472 } else {
a46a3ce4
FB
473 if (!test_idct) {
474 dct_error("REF-DBL", 0, fdct, fdct, test); /* only to verify code ! */
475 dct_error("IJG-AAN-INT", 0, fdct_ifast, fdct, test);
476 dct_error("IJG-LLM-INT", 0, ff_jpeg_fdct_islow, fdct, test);
477 dct_error("MMX", 0, ff_fdct_mmx, fdct, test);
478 } else {
479 dct_error("REF-DBL", 1, idct, idct, test);
480 dct_error("INT", 1, j_rev_dct, idct, test);
481 dct_error("LIBMPEG2-MMX", 1, ff_mmx_idct, idct, test);
482 dct_error("LIBMPEG2-MMXEXT", 1, ff_mmxext_idct, idct, test);
483 dct_error("SIMPLE-C", 1, simple_idct, idct, test);
484 dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, idct, test);
485 // dct_error("ODIVX-C", 1, odivx_idct_c, idct);
486 //printf(" test against odivx idct\n");
487 // dct_error("REF", 1, idct, odivx_idct_c);
488 // dct_error("INT", 1, j_rev_dct, odivx_idct_c);
489 // dct_error("MMX", 1, ff_mmx_idct, odivx_idct_c);
490 // dct_error("MMXEXT", 1, ff_mmxext_idct, odivx_idct_c);
491 // dct_error("SIMPLE-C", 1, simple_idct, odivx_idct_c);
492 // dct_error("SIMPLE-MMX", 1, ff_simple_idct_mmx, odivx_idct_c);
493 // dct_error("ODIVX-C", 1, odivx_idct_c, odivx_idct_c);
494 }
9e1586fc 495 }
de6d9b64
FB
496 return 0;
497}