Parent Directory | Revision Log
Revision 262 - (view) (download)
1 : | Isibaar | 225 | /************************************************************************** |
2 : | * | ||
3 : | * XVID MPEG-4 VIDEO CODEC - Unit tests and benches | ||
4 : | * | ||
5 : | * This program is free software; you can redistribute it and/or modify | ||
6 : | * it under the terms of the GNU General Public License as published by | ||
7 : | * the Free Software Foundation; either version 2 of the License, or | ||
8 : | * (at your option) any later version. | ||
9 : | * | ||
10 : | * This program is distributed in the hope that it will be useful, | ||
11 : | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
12 : | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
13 : | * GNU General Public License for more details. | ||
14 : | * | ||
15 : | * You should have received a copy of the GNU General Public License | ||
16 : | * along with this program; if not, write to the Free Software | ||
17 : | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | ||
18 : | * | ||
19 : | *************************************************************************/ | ||
20 : | |||
21 : | /************************************************************************ | ||
22 : | * | ||
23 : | * 'Reference' output is at the end of file. | ||
24 : | * Don't take the checksums and crc too seriouly, they aren't | ||
25 : | Isibaar | 262 | * bullet-proof (should plug some .md5 here)... |
26 : | Isibaar | 225 | * |
27 : | * compiles with something like: | ||
28 : | * gcc -o xvid_bench xvid_bench.c -I../src/ -lxvidcore -lm | ||
29 : | * | ||
30 : | * History: | ||
31 : | * | ||
32 : | * 06.06.2002 initial coding -Skal- | ||
33 : | * | ||
34 : | *************************************************************************/ | ||
35 : | |||
36 : | #include <stdio.h> | ||
37 : | #include <stdlib.h> | ||
38 : | #include <sys/time.h> // for gettimeofday | ||
39 : | #include <string.h> // for memset | ||
40 : | #include <assert.h> | ||
41 : | |||
42 : | #include "xvid.h" | ||
43 : | |||
44 : | // inner guts | ||
45 : | #include "dct/idct.h" | ||
46 : | #include "dct/fdct.h" | ||
47 : | #include "image/colorspace.h" | ||
48 : | #include "image/interpolate8x8.h" | ||
49 : | #include "utils/mem_transfer.h" | ||
50 : | #include "quant/quant_h263.h" | ||
51 : | #include "quant/quant_mpeg4.h" | ||
52 : | #include "motion/sad.h" | ||
53 : | #include "utils/emms.h" | ||
54 : | #include "utils/timer.h" | ||
55 : | #include "quant/quant_matrix.c" | ||
56 : | #include "bitstream/cbp.h" | ||
57 : | |||
58 : | Isibaar | 262 | #include <math.h> |
59 : | Isibaar | 225 | const int speed_ref = 100; // on slow machines, decrease this value |
60 : | |||
61 : | /********************************************************************* | ||
62 : | * misc | ||
63 : | *********************************************************************/ | ||
64 : | |||
65 : | /* returns time in micro-s*/ | ||
66 : | double gettime_usec() | ||
67 : | { | ||
68 : | struct timeval tv; | ||
69 : | gettimeofday(&tv, 0); | ||
70 : | Isibaar | 262 | return tv.tv_sec*1.0e6f + tv.tv_usec; |
71 : | Isibaar | 225 | } |
72 : | |||
73 : | /* returns squared deviates (mean(v*v)-mean(v)^2) of a 8x8 block */ | ||
74 : | double sqr_dev(uint8_t v[8*8]) | ||
75 : | { | ||
76 : | double sum=0.; | ||
77 : | double sum2=0.; | ||
78 : | int n; | ||
79 : | for (n=0;n<8*8;n++) | ||
80 : | { | ||
81 : | sum += v[n]; | ||
82 : | sum2 += v[n]*v[n]; | ||
83 : | } | ||
84 : | sum2 /= n; | ||
85 : | sum /= n; | ||
86 : | return sum2-sum*sum; | ||
87 : | } | ||
88 : | |||
89 : | /********************************************************************* | ||
90 : | * cpu init | ||
91 : | *********************************************************************/ | ||
92 : | |||
93 : | typedef struct { | ||
94 : | const char *name; | ||
95 : | unsigned int cpu; | ||
96 : | } CPU; | ||
97 : | |||
98 : | CPU cpu_list[] = | ||
99 : | { { "PLAINC", 0 } | ||
100 : | , { "MMX ", XVID_CPU_MMX } | ||
101 : | , { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } | ||
102 : | , { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } | ||
103 : | , { "3DNOW ", XVID_CPU_3DNOW } | ||
104 : | , { "3DNOWE", XVID_CPU_3DNOWEXT } | ||
105 : | //, { "TSC ", XVID_CPU_TSC } | ||
106 : | , { 0, 0 } } | ||
107 : | |||
108 : | , cpu_short_list[] = | ||
109 : | { { "PLAINC", 0 } | ||
110 : | , { "MMX ", XVID_CPU_MMX } | ||
111 : | Isibaar | 262 | //, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
112 : | Isibaar | 225 | , { 0, 0 } } |
113 : | |||
114 : | , cpu_short_list2[] = | ||
115 : | { { "PLAINC", 0 } | ||
116 : | , { "MMX ", XVID_CPU_MMX } | ||
117 : | , { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } | ||
118 : | , { 0, 0 } }; | ||
119 : | |||
120 : | |||
121 : | int init_cpu(CPU *cpu) | ||
122 : | { | ||
123 : | int xerr, cpu_type; | ||
124 : | XVID_INIT_PARAM xinit; | ||
125 : | |||
126 : | cpu_type = check_cpu_features() & cpu->cpu; | ||
127 : | xinit.cpu_flags = cpu_type | XVID_CPU_FORCE; | ||
128 : | // xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; | ||
129 : | xerr = xvid_init(NULL, 0, &xinit, NULL); | ||
130 : | if (cpu->cpu>0 && (cpu_type==0 || xerr!=XVID_ERR_OK)) { | ||
131 : | printf( "%s - skipped...\n", cpu->name ); | ||
132 : | return 0; | ||
133 : | } | ||
134 : | return 1; | ||
135 : | } | ||
136 : | |||
137 : | /********************************************************************* | ||
138 : | * test DCT | ||
139 : | *********************************************************************/ | ||
140 : | |||
141 : | #define ABS(X) ((X)<0 ? -(X) : (X)) | ||
142 : | |||
143 : | void test_dct() | ||
144 : | { | ||
145 : | const int nb_tests = 300*speed_ref; | ||
146 : | int tst; | ||
147 : | CPU *cpu; | ||
148 : | int i; | ||
149 : | short iDst0[8*8], iDst[8*8], fDst[8*8]; | ||
150 : | double overhead; | ||
151 : | |||
152 : | printf( "\n ===== test fdct/idct =====\n" ); | ||
153 : | |||
154 : | for(i=0; i<8*8; ++i) iDst0[i] = (i*7-i*i) & 0x7f; | ||
155 : | overhead = gettime_usec(); | ||
156 : | for(tst=0; tst<nb_tests; ++tst) | ||
157 : | { | ||
158 : | for(i=0; i<8*8; ++i) fDst[i] = iDst0[i]; | ||
159 : | for(i=0; i<8*8; ++i) iDst[i] = fDst[i]; | ||
160 : | } | ||
161 : | overhead = gettime_usec() - overhead; | ||
162 : | |||
163 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
164 : | { | ||
165 : | Isibaar | 262 | double t, PSNR, MSE; |
166 : | Isibaar | 225 | |
167 : | if (!init_cpu(cpu)) | ||
168 : | continue; | ||
169 : | |||
170 : | t = gettime_usec(); | ||
171 : | emms(); | ||
172 : | for(tst=0; tst<nb_tests; ++tst) | ||
173 : | { | ||
174 : | for(i=0; i<8*8; ++i) fDst[i] = iDst0[i]; | ||
175 : | fdct(fDst); | ||
176 : | for(i=0; i<8*8; ++i) iDst[i] = fDst[i]; | ||
177 : | idct(iDst); | ||
178 : | } | ||
179 : | emms(); | ||
180 : | t = (gettime_usec() - t - overhead) / nb_tests; | ||
181 : | Isibaar | 262 | MSE = 0.; |
182 : | Isibaar | 225 | for(i=0; i<8*8; ++i) { |
183 : | Isibaar | 262 | double delta = 1.0*(iDst[i] - iDst0[i]); |
184 : | MSE += delta*delta; | ||
185 : | Isibaar | 225 | } |
186 : | Isibaar | 262 | PSNR = (MSE==0.) ? 1.e6 : -4.3429448*log( MSE/64. ); |
187 : | printf( "%s - %.3f usec PSNR=%.3f MSE=%.3f\n", | ||
188 : | cpu->name, t, PSNR, MSE ); | ||
189 : | if (ABS(MSE)>=64) printf( "*** CRC ERROR! ***\n" ); | ||
190 : | Isibaar | 225 | } |
191 : | } | ||
192 : | |||
193 : | /********************************************************************* | ||
194 : | * test SAD | ||
195 : | *********************************************************************/ | ||
196 : | |||
197 : | void test_sad() | ||
198 : | { | ||
199 : | const int nb_tests = 2000*speed_ref; | ||
200 : | int tst; | ||
201 : | CPU *cpu; | ||
202 : | int i; | ||
203 : | uint8_t Cur[16*16], Ref1[16*16], Ref2[16*16]; | ||
204 : | |||
205 : | printf( "\n ====== test SAD ======\n" ); | ||
206 : | for(i=0; i<16*16;++i) { | ||
207 : | Cur[i] = (i/5) ^ 0x05; | ||
208 : | Ref1[i] = (i + 0x0b) & 0xff; | ||
209 : | Ref2[i] = i ^ 0x76; | ||
210 : | } | ||
211 : | |||
212 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
213 : | { | ||
214 : | double t; | ||
215 : | uint32_t s; | ||
216 : | if (!init_cpu(cpu)) | ||
217 : | continue; | ||
218 : | |||
219 : | t = gettime_usec(); | ||
220 : | emms(); | ||
221 : | for(tst=0; tst<nb_tests; ++tst) s = sad8(Cur, Ref1, 16); | ||
222 : | emms(); | ||
223 : | t = (gettime_usec() - t) / nb_tests; | ||
224 : | printf( "%s - sad8 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
225 : | if (s!=3776) printf( "*** CRC ERROR! ***\n" ); | ||
226 : | |||
227 : | t = gettime_usec(); | ||
228 : | emms(); | ||
229 : | for(tst=0; tst<nb_tests; ++tst) s = sad16(Cur, Ref1, 16, -1); | ||
230 : | emms(); | ||
231 : | t = (gettime_usec() - t) / nb_tests; | ||
232 : | printf( "%s - sad16 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
233 : | if (s!=27214) printf( "*** CRC ERROR! ***\n" ); | ||
234 : | |||
235 : | t = gettime_usec(); | ||
236 : | emms(); | ||
237 : | for(tst=0; tst<nb_tests; ++tst) s = sad16bi(Cur, Ref1, Ref2, 16); | ||
238 : | emms(); | ||
239 : | t = (gettime_usec() - t) / nb_tests; | ||
240 : | printf( "%s - sad16bi %.3f usec sad=%d\n", cpu->name, t, s ); | ||
241 : | if (s!=26274) printf( "*** CRC ERROR! ***\n" ); | ||
242 : | |||
243 : | t = gettime_usec(); | ||
244 : | emms(); | ||
245 : | for(tst=0; tst<nb_tests; ++tst) s = dev16(Cur, 16); | ||
246 : | emms(); | ||
247 : | t = (gettime_usec() - t) / nb_tests; | ||
248 : | printf( "%s - dev16 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
249 : | if (s!=3344) printf( "*** CRC ERROR! ***\n" ); | ||
250 : | |||
251 : | printf( " --- \n" ); | ||
252 : | } | ||
253 : | } | ||
254 : | |||
255 : | /********************************************************************* | ||
256 : | * test interpolation | ||
257 : | *********************************************************************/ | ||
258 : | |||
259 : | #define ENTER \ | ||
260 : | for(i=0; i<16*8; ++i) Dst[i] = 0; \ | ||
261 : | t = gettime_usec(); \ | ||
262 : | emms(); | ||
263 : | |||
264 : | #define LEAVE \ | ||
265 : | emms(); \ | ||
266 : | t = (gettime_usec() - t) / nb_tests; \ | ||
267 : | iCrc = 0; \ | ||
268 : | for(i=0; i<16*8; ++i) { iCrc += Dst[i]^i; } | ||
269 : | |||
270 : | #define TEST_MB(FUNC, R) \ | ||
271 : | ENTER \ | ||
272 : | for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst, Src0, 16, (R)); \ | ||
273 : | LEAVE | ||
274 : | |||
275 : | #define TEST_MB2(FUNC) \ | ||
276 : | ENTER \ | ||
277 : | for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst, Src0, 16); \ | ||
278 : | LEAVE | ||
279 : | |||
280 : | |||
281 : | void test_mb() | ||
282 : | { | ||
283 : | const int nb_tests = 2000*speed_ref; | ||
284 : | CPU *cpu; | ||
285 : | const uint8_t Src0[16*9] = { | ||
286 : | // try to have every possible combinaison of rounding... | ||
287 : | 0, 0, 1, 0, 2, 0, 3, 0, 4 ,0,0,0, 0,0,0,0 | ||
288 : | , 0, 1, 1, 1, 2, 1, 3, 1, 3 ,0,0,0, 0,0,0,0 | ||
289 : | , 0, 2, 1, 2, 2, 2, 3, 2, 2 ,0,0,0, 0,0,0,0 | ||
290 : | , 0, 3, 1, 3, 2, 3, 3, 3, 1 ,0,0,0, 0,0,0,0 | ||
291 : | , 1, 3, 0, 2, 1, 0, 2, 3, 4 ,0,0,0, 0,0,0,0 | ||
292 : | , 2, 2, 1, 2, 0, 1, 3, 5, 3 ,0,0,0, 0,0,0,0 | ||
293 : | , 3, 1, 2, 3, 1, 2, 2, 6, 2 ,0,0,0, 0,0,0,0 | ||
294 : | , 1, 0, 1, 3, 0, 3, 1, 6, 1 ,0,0,0, 0,0,0,0 | ||
295 : | , 4, 3, 2, 1, 2, 3, 4, 0, 3 ,0,0,0, 0,0,0,0 | ||
296 : | }; | ||
297 : | uint8_t Dst[16*8] = {0}; | ||
298 : | |||
299 : | printf( "\n === test block motion ===\n" ); | ||
300 : | |||
301 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
302 : | { | ||
303 : | double t; | ||
304 : | int tst, i, iCrc; | ||
305 : | |||
306 : | if (!init_cpu(cpu)) | ||
307 : | continue; | ||
308 : | |||
309 : | TEST_MB(interpolate8x8_halfpel_h, 0); | ||
310 : | printf( "%s - interp- h-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
311 : | if (iCrc!=8107) printf( "*** CRC ERROR! ***\n" ); | ||
312 : | |||
313 : | TEST_MB(interpolate8x8_halfpel_h, 1); | ||
314 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
315 : | if (iCrc!=8100) printf( "*** CRC ERROR! ***\n" ); | ||
316 : | |||
317 : | |||
318 : | TEST_MB(interpolate8x8_halfpel_v, 0); | ||
319 : | printf( "%s - interp- v-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
320 : | if (iCrc!=8108) printf( "*** CRC ERROR! ***\n" ); | ||
321 : | |||
322 : | TEST_MB(interpolate8x8_halfpel_v, 1); | ||
323 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
324 : | if (iCrc!=8105) printf( "*** CRC ERROR! ***\n" ); | ||
325 : | |||
326 : | |||
327 : | TEST_MB(interpolate8x8_halfpel_hv, 0); | ||
328 : | printf( "%s - interp-hv-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
329 : | if (iCrc!=8112) printf( "*** CRC ERROR! ***\n" ); | ||
330 : | |||
331 : | TEST_MB(interpolate8x8_halfpel_hv, 1); | ||
332 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
333 : | if (iCrc!=8103) printf( "*** CRC ERROR! ***\n" ); | ||
334 : | |||
335 : | Isibaar | 262 | |
336 : | // this is a new function, as of 06.06.2002 | ||
337 : | #if 0 | ||
338 : | TEST_MB2(interpolate8x8_avrg); | ||
339 : | printf( "%s - interpolate8x8_c %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
340 : | if (iCrc!=8107) printf( "*** CRC ERROR! ***\n" ); | ||
341 : | #endif | ||
342 : | |||
343 : | Isibaar | 225 | printf( " --- \n" ); |
344 : | } | ||
345 : | } | ||
346 : | |||
347 : | /********************************************************************* | ||
348 : | * test transfer | ||
349 : | *********************************************************************/ | ||
350 : | |||
351 : | #define INIT_TRANSFER \ | ||
352 : | for(i=0; i<8*32; ++i) { \ | ||
353 : | Src8[i] = i; Src16[i] = i; \ | ||
354 : | Dst8[i] = 0; Dst16[i] = 0; \ | ||
355 : | Ref1[i] = i^0x27; \ | ||
356 : | Ref2[i] = i^0x51; \ | ||
357 : | } | ||
358 : | |||
359 : | #define TEST_TRANSFER_BEGIN(DST) \ | ||
360 : | INIT_TRANSFER \ | ||
361 : | overhead = -gettime_usec(); \ | ||
362 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
363 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
364 : | } \ | ||
365 : | overhead += gettime_usec(); \ | ||
366 : | t = gettime_usec(); \ | ||
367 : | emms(); \ | ||
368 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
369 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a; | ||
370 : | |||
371 : | |||
372 : | #define TEST_TRANSFER_END(DST) \ | ||
373 : | } \ | ||
374 : | emms(); \ | ||
375 : | t = (gettime_usec()-t -overhead) / nb_tests;\ | ||
376 : | s = 0; for(i=0; i<8*32; ++i) { s += (DST)[i]^i; } | ||
377 : | |||
378 : | #define TEST_TRANSFER(FUNC, DST, SRC) \ | ||
379 : | TEST_TRANSFER_BEGIN(DST); \ | ||
380 : | (FUNC)((DST), (SRC), 32); \ | ||
381 : | TEST_TRANSFER_END(DST) | ||
382 : | |||
383 : | |||
384 : | #define TEST_TRANSFER2_BEGIN(DST, SRC) \ | ||
385 : | INIT_TRANSFER \ | ||
386 : | overhead = -gettime_usec(); \ | ||
387 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
388 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
389 : | for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e;\ | ||
390 : | } \ | ||
391 : | overhead += gettime_usec(); \ | ||
392 : | t = gettime_usec(); \ | ||
393 : | emms(); \ | ||
394 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
395 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
396 : | for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e; | ||
397 : | |||
398 : | #define TEST_TRANSFER2_END(DST) \ | ||
399 : | } \ | ||
400 : | emms(); \ | ||
401 : | t = (gettime_usec()-t -overhead) / nb_tests;\ | ||
402 : | s = 0; for(i=0; i<8*32; ++i) { s += (DST)[i]; } | ||
403 : | |||
404 : | #define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ | ||
405 : | TEST_TRANSFER2_BEGIN(DST,SRC); \ | ||
406 : | (FUNC)((DST), (SRC), (R1), 32); \ | ||
407 : | TEST_TRANSFER2_END(DST) | ||
408 : | |||
409 : | #define TEST_TRANSFER3(FUNC, DST, SRC, R1, R2)\ | ||
410 : | TEST_TRANSFER_BEGIN(DST); \ | ||
411 : | (FUNC)((DST), (SRC), (R1), (R2), 32); \ | ||
412 : | TEST_TRANSFER_END(DST) | ||
413 : | |||
414 : | void test_transfer() | ||
415 : | { | ||
416 : | const int nb_tests = 4000*speed_ref; | ||
417 : | int i; | ||
418 : | CPU *cpu; | ||
419 : | uint8_t Src8[8*32], Dst8[8*32], Ref1[8*32], Ref2[8*32]; | ||
420 : | int16_t Src16[8*32], Dst16[8*32]; | ||
421 : | |||
422 : | printf( "\n === test transfer ===\n" ); | ||
423 : | |||
424 : | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) | ||
425 : | { | ||
426 : | double t, overhead; | ||
427 : | int tst, s; | ||
428 : | |||
429 : | if (!init_cpu(cpu)) | ||
430 : | continue; | ||
431 : | |||
432 : | TEST_TRANSFER(transfer_8to16copy, Dst16, Src8); | ||
433 : | printf( "%s - 8to16 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
434 : | if (s!=28288) printf( "*** CRC ERROR! ***\n" ); | ||
435 : | |||
436 : | TEST_TRANSFER(transfer_16to8copy, Dst8, Src16); | ||
437 : | printf( "%s - 16to8 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
438 : | if (s!=28288) printf( "*** CRC ERROR! ***\n" ); | ||
439 : | |||
440 : | TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); | ||
441 : | printf( "%s - 8to8 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
442 : | if (s!=20352) printf( "*** CRC ERROR! ***\n" ); | ||
443 : | |||
444 : | TEST_TRANSFER(transfer_16to8add, Dst8, Src16); | ||
445 : | printf( "%s - 16to8add %.3f usec crc=%d\n", cpu->name, t, s ); | ||
446 : | if (s!=25536) printf( "*** CRC ERROR! ***\n" ); | ||
447 : | |||
448 : | TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); | ||
449 : | printf( "%s - 8to16sub %.3f usec crc1=%d ", cpu->name, t, s ); | ||
450 : | if (s!=28064) printf( "*** CRC ERROR! ***\n" ); | ||
451 : | s = 0; for(i=0; i<8*32; ++i) { s += (Src8[i]-Ref1[i])&i; } | ||
452 : | printf( "crc2=%d\n", s); | ||
453 : | if (s!=16256) printf( "*** CRC ERROR! ***\n" ); | ||
454 : | Isibaar | 262 | #if 1 |
455 : | Isibaar | 225 | TEST_TRANSFER3(transfer_8to16sub2, Dst16, Src8, Ref1, Ref2); |
456 : | printf( "%s - 8to16sub2 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
457 : | if (s!=20384) printf( "*** CRC ERROR! ***\n" ); | ||
458 : | Isibaar | 262 | // for(i=0; i<64; ++i) printf( "[%d]", Dst16[i]); |
459 : | // printf("\n"); | ||
460 : | #endif | ||
461 : | Isibaar | 225 | printf( " --- \n" ); |
462 : | } | ||
463 : | } | ||
464 : | |||
465 : | /********************************************************************* | ||
466 : | * test quantization | ||
467 : | *********************************************************************/ | ||
468 : | |||
469 : | Isibaar | 262 | #define TEST_QUANT(FUNC, DST, SRC) \ |
470 : | t = gettime_usec(); \ | ||
471 : | for(s=0,qm=1; qm<=255; ++qm) { \ | ||
472 : | for(i=0; i<8*8; ++i) Quant[i] = qm; \ | ||
473 : | set_inter_matrix( Quant ); \ | ||
474 : | emms(); \ | ||
475 : | for(q=1; q<=max_Q; ++q) { \ | ||
476 : | for(tst=0; tst<nb_tests; ++tst) \ | ||
477 : | (FUNC)((DST), (SRC), q); \ | ||
478 : | for(i=0; i<64; ++i) s+=(DST)[i]^i^qm; \ | ||
479 : | } \ | ||
480 : | emms(); \ | ||
481 : | } \ | ||
482 : | t = (gettime_usec()-t-overhead)/nb_tests/qm;\ | ||
483 : | s = (s&0xffff)^(s>>16) | ||
484 : | Isibaar | 225 | |
485 : | Isibaar | 262 | #define TEST_QUANT2(FUNC, DST, SRC) \ |
486 : | t = gettime_usec(); \ | ||
487 : | for(s=0,qm=1; qm<=255; ++qm) { \ | ||
488 : | for(i=0; i<8*8; ++i) Quant[i] = qm; \ | ||
489 : | set_intra_matrix( Quant ); \ | ||
490 : | emms(); \ | ||
491 : | for(q=1; q<=max_Q; ++q) { \ | ||
492 : | for(tst=0; tst<nb_tests; ++tst) \ | ||
493 : | (FUNC)((DST), (SRC), q, q); \ | ||
494 : | for(i=0; i<64; ++i) s+=(DST)[i]^i^qm; \ | ||
495 : | } \ | ||
496 : | emms(); \ | ||
497 : | } \ | ||
498 : | t = (gettime_usec()-t-overhead)/nb_tests/qm;\ | ||
499 : | s = (s&0xffff)^(s>>16) | ||
500 : | Isibaar | 225 | |
501 : | void test_quant() | ||
502 : | { | ||
503 : | Isibaar | 262 | const int nb_tests = 1*speed_ref; |
504 : | Isibaar | 225 | const int max_Q = 31; |
505 : | Isibaar | 262 | int i, qm; |
506 : | Isibaar | 225 | CPU *cpu; |
507 : | int16_t Src[8*8], Dst[8*8]; | ||
508 : | Isibaar | 262 | uint8_t Quant[8*8]; |
509 : | Isibaar | 225 | |
510 : | printf( "\n ===== test quant =====\n" ); | ||
511 : | |||
512 : | Isibaar | 262 | // we deliberately enfringe the norm's specified range [-127,127], |
513 : | // to test the robustness of the iquant module | ||
514 : | Isibaar | 225 | for(i=0; i<64; ++i) { |
515 : | Isibaar | 262 | Src[i] = 1 + (i-32) * (i&6); |
516 : | Isibaar | 225 | Dst[i] = 0; |
517 : | } | ||
518 : | |||
519 : | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) | ||
520 : | { | ||
521 : | double t, overhead; | ||
522 : | Isibaar | 262 | int tst, q; |
523 : | uint32_t s; | ||
524 : | Isibaar | 225 | |
525 : | if (!init_cpu(cpu)) | ||
526 : | continue; | ||
527 : | |||
528 : | overhead = -gettime_usec(); | ||
529 : | Isibaar | 262 | for(s=0,qm=1; qm<=255; ++qm) { |
530 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
531 : | set_inter_matrix( Quant ); | ||
532 : | for(q=1; q<=max_Q; ++q) | ||
533 : | for(i=0; i<64; ++i) s+=Dst[i]^i^qm; | ||
534 : | } | ||
535 : | Isibaar | 225 | overhead += gettime_usec(); |
536 : | |||
537 : | Isibaar | 262 | #if 1 |
538 : | TEST_QUANT2(quant4_intra, Dst, Src); | ||
539 : | Isibaar | 225 | printf( "%s - quant4_intra %.3f usec crc=%d\n", cpu->name, t, s ); |
540 : | Isibaar | 262 | if (s!=29809) printf( "*** CRC ERROR! ***\n" ); |
541 : | Isibaar | 225 | |
542 : | TEST_QUANT(quant4_inter, Dst, Src); | ||
543 : | printf( "%s - quant4_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
544 : | Isibaar | 262 | if (s!=12574) printf( "*** CRC ERROR! ***\n" ); |
545 : | #endif | ||
546 : | #if 1 | ||
547 : | TEST_QUANT2(dequant4_intra, Dst, Src); | ||
548 : | Isibaar | 225 | printf( "%s - dequant4_intra %.3f usec crc=%d\n", cpu->name, t, s ); |
549 : | Isibaar | 262 | if (s!=24052) printf( "*** CRC ERROR! ***\n" ); |
550 : | Isibaar | 225 | |
551 : | TEST_QUANT(dequant4_inter, Dst, Src); | ||
552 : | printf( "%s - dequant4_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
553 : | Isibaar | 262 | if (s!=63847) printf( "*** CRC ERROR! ***\n" ); |
554 : | #endif | ||
555 : | #if 1 | ||
556 : | TEST_QUANT2(quant_intra, Dst, Src); | ||
557 : | Isibaar | 225 | printf( "%s - quant_intra %.3f usec crc=%d\n", cpu->name, t, s ); |
558 : | Isibaar | 262 | if (s!=25662) printf( "*** CRC ERROR! ***\n" ); |
559 : | Isibaar | 225 | |
560 : | TEST_QUANT(quant_inter, Dst, Src); | ||
561 : | printf( "%s - quant_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
562 : | Isibaar | 262 | if (s!=23972) printf( "*** CRC ERROR! ***\n" ); |
563 : | #endif | ||
564 : | #if 1 | ||
565 : | TEST_QUANT2(dequant_intra, Dst, Src); | ||
566 : | Isibaar | 225 | printf( "%s - dequant_intra %.3f usec crc=%d\n", cpu->name, t, s ); |
567 : | Isibaar | 262 | if (s!=49900) printf( "*** CRC ERROR! ***\n" ); |
568 : | Isibaar | 225 | |
569 : | TEST_QUANT(dequant_inter, Dst, Src); | ||
570 : | printf( "%s - dequant_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
571 : | Isibaar | 262 | if (s!=48899) printf( "*** CRC ERROR! ***\n" ); |
572 : | #endif | ||
573 : | Isibaar | 225 | printf( " --- \n" ); |
574 : | } | ||
575 : | } | ||
576 : | |||
577 : | /********************************************************************* | ||
578 : | * test non-zero AC counting | ||
579 : | *********************************************************************/ | ||
580 : | |||
581 : | #define TEST_CBP(FUNC, SRC) \ | ||
582 : | t = gettime_usec(); \ | ||
583 : | emms(); \ | ||
584 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
585 : | cbp = (FUNC)((SRC)); \ | ||
586 : | } \ | ||
587 : | emms(); \ | ||
588 : | t = (gettime_usec()-t ) / nb_tests; | ||
589 : | |||
590 : | void test_cbp() | ||
591 : | { | ||
592 : | const int nb_tests = 10000*speed_ref; | ||
593 : | int i; | ||
594 : | CPU *cpu; | ||
595 : | int16_t Src1[6*64], Src2[6*64], Src3[6*64], Src4[6*64]; | ||
596 : | |||
597 : | printf( "\n ===== test cbp =====\n" ); | ||
598 : | |||
599 : | for(i=0; i<6*64; ++i) { | ||
600 : | Src1[i] = (i*i*3/8192)&(i/64)&1; // 'random' | ||
601 : | Src2[i] = (i<3*64); // half-full | ||
602 : | Src3[i] = ((i+32)>3*64); | ||
603 : | Src4[i] = (i==(3*64+2) || i==(5*64+9)); | ||
604 : | } | ||
605 : | |||
606 : | for(cpu = cpu_short_list2; cpu->name!=0; ++cpu) | ||
607 : | { | ||
608 : | double t; | ||
609 : | int tst, cbp; | ||
610 : | |||
611 : | if (!init_cpu(cpu)) | ||
612 : | continue; | ||
613 : | |||
614 : | TEST_CBP(calc_cbp, Src1); | ||
615 : | printf( "%s - calc_cbp#1 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
616 : | if (cbp!=0x15) printf( "*** CRC ERROR! ***\n" ); | ||
617 : | TEST_CBP(calc_cbp, Src2); | ||
618 : | printf( "%s - calc_cbp#2 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
619 : | if (cbp!=0x38) printf( "*** CRC ERROR! ***\n" ); | ||
620 : | TEST_CBP(calc_cbp, Src3); | ||
621 : | printf( "%s - calc_cbp#3 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
622 : | if (cbp!=0x0f) printf( "*** CRC ERROR! ***\n" ); | ||
623 : | TEST_CBP(calc_cbp, Src4); | ||
624 : | printf( "%s - calc_cbp#4 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
625 : | if (cbp!=0x05) printf( "*** CRC ERROR! ***\n" ); | ||
626 : | printf( " --- \n" ); | ||
627 : | } | ||
628 : | } | ||
629 : | |||
630 : | /********************************************************************* | ||
631 : | Isibaar | 262 | * fdct/idct IEEE1180 compliance |
632 : | *********************************************************************/ | ||
633 : | |||
634 : | typedef struct { | ||
635 : | long Errors[64]; | ||
636 : | long Sqr_Errors[64]; | ||
637 : | long Max_Errors[64]; | ||
638 : | long Nb; | ||
639 : | } STATS_8x8; | ||
640 : | |||
641 : | void init_stats(STATS_8x8 *S) | ||
642 : | { | ||
643 : | int i; | ||
644 : | for(i=0; i<64; ++i) { | ||
645 : | S->Errors[i] = 0; | ||
646 : | S->Sqr_Errors[i] = 0; | ||
647 : | S->Max_Errors[i] = 0; | ||
648 : | } | ||
649 : | S->Nb = 0; | ||
650 : | } | ||
651 : | |||
652 : | void store_stats(STATS_8x8 *S, short Blk[64], short Ref[64]) | ||
653 : | { | ||
654 : | int i; | ||
655 : | for(i=0; i<64; ++i) | ||
656 : | { | ||
657 : | short Err = Blk[i] - Ref[i]; | ||
658 : | S->Errors[i] += Err; | ||
659 : | S->Sqr_Errors[i] += Err * Err; | ||
660 : | if (Err<0) Err = -Err; | ||
661 : | if (S->Max_Errors[i]<Err) | ||
662 : | S->Max_Errors[i] = Err; | ||
663 : | } | ||
664 : | S->Nb++; | ||
665 : | } | ||
666 : | |||
667 : | void print_stats(STATS_8x8 *S) | ||
668 : | { | ||
669 : | int i; | ||
670 : | double Norm; | ||
671 : | |||
672 : | assert(S->Nb>0); | ||
673 : | Norm = 1. / (double)S->Nb; | ||
674 : | printf("\n== Max absolute values of errors ==\n"); | ||
675 : | for(i=0; i<64; i++) { | ||
676 : | printf(" %4ld", S->Max_Errors[i]); | ||
677 : | if ((i&7)==7) printf("\n"); | ||
678 : | } | ||
679 : | |||
680 : | printf("\n== Mean square errors ==\n"); | ||
681 : | for(i=0; i<64; i++) | ||
682 : | { | ||
683 : | double Err = Norm * (double)S->Sqr_Errors[i]; | ||
684 : | printf(" %.3f", Err); | ||
685 : | if ((i&7)==7) printf("\n"); | ||
686 : | } | ||
687 : | |||
688 : | printf("\n== Mean errors ==\n"); | ||
689 : | for(i=0; i<64; i++) | ||
690 : | { | ||
691 : | double Err = Norm * (double)S->Errors[i]; | ||
692 : | printf(" %.3f", Err); | ||
693 : | if ((i&7)==7) printf("\n"); | ||
694 : | } | ||
695 : | printf("\n"); | ||
696 : | } | ||
697 : | |||
698 : | static const char *CHECK(double v, double l) { | ||
699 : | if (fabs(v)<=l) return "ok"; | ||
700 : | else return "FAIL!"; | ||
701 : | } | ||
702 : | |||
703 : | void report_stats(STATS_8x8 *S, const double *Limits) | ||
704 : | { | ||
705 : | int i; | ||
706 : | double Norm, PE, PMSE, OMSE, PME, OME; | ||
707 : | |||
708 : | assert(S->Nb>0); | ||
709 : | Norm = 1. / (double)S->Nb; | ||
710 : | PE = 0.; | ||
711 : | for(i=0; i<64; i++) { | ||
712 : | if (PE<S->Max_Errors[i]) | ||
713 : | PE = S->Max_Errors[i]; | ||
714 : | } | ||
715 : | |||
716 : | PMSE = 0.; | ||
717 : | OMSE = 0.; | ||
718 : | for(i=0; i<64; i++) | ||
719 : | { | ||
720 : | double Err = Norm * (double)S->Sqr_Errors[i]; | ||
721 : | OMSE += Err; | ||
722 : | if (PMSE < Err) PMSE = Err; | ||
723 : | } | ||
724 : | OMSE /= 64.; | ||
725 : | |||
726 : | PME = 0.; | ||
727 : | OME = 0.; | ||
728 : | for(i=0; i<64; i++) | ||
729 : | { | ||
730 : | double Err = Norm * (double)S->Errors[i]; | ||
731 : | OME += Err; | ||
732 : | Err = fabs(Err); | ||
733 : | if (PME < Err) PME = Err; | ||
734 : | } | ||
735 : | OME /= 64.; | ||
736 : | |||
737 : | printf( "Peak error: %4.4f\n", PE ); | ||
738 : | printf( "Peak MSE: %4.4f\n", PMSE ); | ||
739 : | printf( "Overall MSE: %4.4f\n", OMSE ); | ||
740 : | printf( "Peak ME: %4.4f\n", PME ); | ||
741 : | printf( "Overall ME: %4.4f\n", OME ); | ||
742 : | |||
743 : | if (Limits!=0) | ||
744 : | { | ||
745 : | printf( "[PE<=%.4f %s] ", Limits[0], CHECK(PE, Limits[0]) ); | ||
746 : | printf( "\n" ); | ||
747 : | printf( "[PMSE<=%.4f %s]", Limits[1], CHECK(PMSE, Limits[1]) ); | ||
748 : | printf( "[OMSE<=%.4f %s]", Limits[2], CHECK(OMSE, Limits[2]) ); | ||
749 : | printf( "\n" ); | ||
750 : | printf( "[PME<=%.4f %s] ", Limits[3], CHECK(PME , Limits[3]) ); | ||
751 : | printf( "[OME<=%.4f %s] ", Limits[4], CHECK(OME , Limits[4]) ); | ||
752 : | printf( "\n" ); | ||
753 : | } | ||
754 : | } | ||
755 : | |||
756 : | ////////////////////////////////////////////////////////// | ||
757 : | /* Pseudo-random generator specified by IEEE 1180 */ | ||
758 : | |||
759 : | static long ieee_seed = 1; | ||
760 : | static void ieee_reseed(long s) { | ||
761 : | ieee_seed = s; | ||
762 : | } | ||
763 : | static long ieee_rand(int Min, int Max) | ||
764 : | { | ||
765 : | static double z = (double) 0x7fffffff; | ||
766 : | |||
767 : | long i,j; | ||
768 : | double x; | ||
769 : | |||
770 : | ieee_seed = (ieee_seed * 1103515245) + 12345; | ||
771 : | i = ieee_seed & 0x7ffffffe; | ||
772 : | x = ((double) i) / z; | ||
773 : | x *= (Max-Min+1); | ||
774 : | j = (long)x; | ||
775 : | j = j + Min; | ||
776 : | assert(j>=Min && j<=Max); | ||
777 : | return (short)j; | ||
778 : | } | ||
779 : | |||
780 : | #define CLAMP(x, M) (x) = ((x)<-(M)) ? (-(M)) : ((x)>=(M) ? ((M)-1) : (x)) | ||
781 : | |||
782 : | static double Cos[8][8]; | ||
783 : | static void init_ref_dct() | ||
784 : | { | ||
785 : | int i, j; | ||
786 : | for(i=0; i<8; i++) | ||
787 : | { | ||
788 : | double scale = (i == 0) ? sqrt(0.125) : 0.5; | ||
789 : | for (j=0; j<8; j++) | ||
790 : | Cos[i][j] = scale*cos( (M_PI/8.0)*i*(j + 0.5) ); | ||
791 : | } | ||
792 : | } | ||
793 : | |||
794 : | void ref_idct(short *M) | ||
795 : | { | ||
796 : | int i, j, k; | ||
797 : | double Tmp[8][8]; | ||
798 : | |||
799 : | for(i=0; i<8; i++) { | ||
800 : | for(j=0; j<8; j++) | ||
801 : | { | ||
802 : | double Sum = 0.0; | ||
803 : | for (k=0; k<8; k++) Sum += Cos[k][j]*M[8*i+k]; | ||
804 : | Tmp[i][j] = Sum; | ||
805 : | } | ||
806 : | } | ||
807 : | for(i=0; i<8; i++) { | ||
808 : | for(j=0; j<8; j++) { | ||
809 : | double Sum = 0.0; | ||
810 : | for (k=0; k<8; k++) Sum += Cos[k][i]*Tmp[k][j]; | ||
811 : | M[8*i+j] = (short)floor(Sum + .5); | ||
812 : | } | ||
813 : | } | ||
814 : | } | ||
815 : | |||
816 : | void ref_fdct(short *M) | ||
817 : | { | ||
818 : | int i, j, k; | ||
819 : | double Tmp[8][8]; | ||
820 : | |||
821 : | for(i=0; i<8; i++) { | ||
822 : | for(j=0; j<8; j++) | ||
823 : | { | ||
824 : | double Sum = 0.0; | ||
825 : | for (k=0; k<8; k++) Sum += Cos[j][k]*M[8*i+k]; | ||
826 : | Tmp[i][j] = Sum; | ||
827 : | } | ||
828 : | } | ||
829 : | for(i=0; i<8; i++) { | ||
830 : | for(j=0; j<8; j++) { | ||
831 : | double Sum = 0.0; | ||
832 : | for (k=0; k<8; k++) Sum += Cos[i][k]*Tmp[k][j]; | ||
833 : | M[8*i+j] = (short)floor(Sum + 0.5); | ||
834 : | } | ||
835 : | } | ||
836 : | } | ||
837 : | |||
838 : | void test_IEEE1180_compliance(int Min, int Max, int Sign) | ||
839 : | { | ||
840 : | static const double ILimits[5] = { 1., 0.06, 0.02, 0.015, 0.0015 }; | ||
841 : | int Loops = 10000; | ||
842 : | int i, m, n; | ||
843 : | short Blk0[64]; // reference | ||
844 : | short Blk[64], iBlk[64]; | ||
845 : | short Ref_FDCT[64]; | ||
846 : | short Ref_IDCT[64]; | ||
847 : | |||
848 : | STATS_8x8 FStats; // forward dct stats | ||
849 : | STATS_8x8 IStats; // inverse dct stats | ||
850 : | |||
851 : | CPU *cpu; | ||
852 : | |||
853 : | init_ref_dct(); | ||
854 : | |||
855 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
856 : | { | ||
857 : | if (!init_cpu(cpu)) | ||
858 : | continue; | ||
859 : | |||
860 : | printf( "\n===== IEEE test for %s ==== (Min=%d Max=%d Sign=%d Loops=%d)\n", | ||
861 : | cpu->name, Min, Max, Sign, Loops); | ||
862 : | |||
863 : | init_stats(&IStats); | ||
864 : | init_stats(&FStats); | ||
865 : | |||
866 : | ieee_reseed(1); | ||
867 : | for(n=0; n<Loops; ++n) | ||
868 : | { | ||
869 : | for(i=0; i<64; ++i) | ||
870 : | Blk0[i] = (short)ieee_rand(Min,Max) * Sign; | ||
871 : | |||
872 : | // hmm, I'm not quite sure this is exactly | ||
873 : | // the tests described in the norm. check... | ||
874 : | |||
875 : | memcpy(Ref_FDCT, Blk0, 64*sizeof(short)); | ||
876 : | ref_fdct(Ref_FDCT); | ||
877 : | for(i=0; i<64; i++) CLAMP( Ref_FDCT[i], 2048 ); | ||
878 : | |||
879 : | memcpy(Blk, Blk0, 64*sizeof(short)); | ||
880 : | emms(); fdct(Blk); emms(); | ||
881 : | for(i=0; i<64; i++) CLAMP( Blk[i], 2048 ); | ||
882 : | |||
883 : | store_stats(&FStats, Blk, Ref_FDCT); | ||
884 : | |||
885 : | |||
886 : | memcpy(Ref_IDCT, Ref_FDCT, 64*sizeof(short)); | ||
887 : | ref_idct(Ref_IDCT); | ||
888 : | for (i=0; i<64; i++) CLAMP( Ref_IDCT[i], 256 ); | ||
889 : | |||
890 : | memcpy(iBlk, Ref_FDCT, 64*sizeof(short)); | ||
891 : | emms(); idct(iBlk); emms(); | ||
892 : | for(i=0; i<64; i++) CLAMP( iBlk[i], 256 ); | ||
893 : | |||
894 : | store_stats(&IStats, iBlk, Ref_IDCT); | ||
895 : | } | ||
896 : | |||
897 : | |||
898 : | printf( "\n -- FDCT report --\n" ); | ||
899 : | // print_stats(&FStats); | ||
900 : | report_stats(&FStats, 0); // so far I know, IEEE1180 says nothing for fdct | ||
901 : | |||
902 : | for(i=0; i<64; i++) Blk[i] = 0; | ||
903 : | emms(); fdct(Blk); emms(); | ||
904 : | for(m=i=0; i<64; i++) if (Blk[i]!=0) m++; | ||
905 : | printf( "FDCT(0) == 0 ? %s\n", (m!=0) ? "NOPE!" : "yup." ); | ||
906 : | |||
907 : | printf( "\n -- IDCT report --\n" ); | ||
908 : | // print_stats(&IStats); | ||
909 : | report_stats(&IStats, ILimits); | ||
910 : | |||
911 : | |||
912 : | for(i=0; i<64; i++) Blk[i] = 0; | ||
913 : | emms(); idct(Blk); emms(); | ||
914 : | for(m=i=0; i<64; i++) if (Blk[i]!=0) m++; | ||
915 : | printf( "IDCT(0) == 0 ? %s\n", (m!=0) ? "NOPE!" : "yup." ); | ||
916 : | } | ||
917 : | } | ||
918 : | |||
919 : | |||
920 : | void test_dct_saturation(int Min, int Max) | ||
921 : | { | ||
922 : | // test behaviour on input range fringe | ||
923 : | |||
924 : | int i, n, p; | ||
925 : | CPU *cpu; | ||
926 : | // const short IDCT_MAX = 2047; // 12bits input | ||
927 : | // const short IDCT_MIN = -2048; | ||
928 : | // const short IDCT_OUT = 256; // 9bits ouput | ||
929 : | const int Partitions = 4; | ||
930 : | const int Loops = 10000 / Partitions; | ||
931 : | |||
932 : | init_ref_dct(); | ||
933 : | |||
934 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
935 : | { | ||
936 : | short Blk0[64], Blk[64]; | ||
937 : | STATS_8x8 Stats; | ||
938 : | |||
939 : | if (!init_cpu(cpu)) | ||
940 : | continue; | ||
941 : | |||
942 : | printf( "\n===== IEEE test for %s Min=%d Max=%d =====\n", | ||
943 : | cpu->name, Min, Max ); | ||
944 : | |||
945 : | // FDCT tests // | ||
946 : | |||
947 : | init_stats(&Stats); | ||
948 : | |||
949 : | // test each computation channels separately | ||
950 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? Max : 0; | ||
951 : | ref_fdct(Blk0); | ||
952 : | emms(); fdct(Blk); emms(); | ||
953 : | store_stats(&Stats, Blk, Blk0); | ||
954 : | |||
955 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? Min : 0; | ||
956 : | ref_fdct(Blk0); | ||
957 : | emms(); fdct(Blk); emms(); | ||
958 : | store_stats(&Stats, Blk, Blk0); | ||
959 : | |||
960 : | // randomly saturated inputs | ||
961 : | for(p=0; p<Partitions; ++p) | ||
962 : | { | ||
963 : | for(n=0; n<Loops; ++n) | ||
964 : | { | ||
965 : | for(i=0; i<64; ++i) | ||
966 : | Blk0[i] = Blk[i] = (ieee_rand(0,Partitions)>=p)? Max : Min; | ||
967 : | ref_fdct(Blk0); | ||
968 : | emms(); fdct(Blk); emms(); | ||
969 : | store_stats(&Stats, Blk, Blk0); | ||
970 : | } | ||
971 : | } | ||
972 : | printf( "\n -- FDCT saturation report --\n" ); | ||
973 : | report_stats(&Stats, 0); | ||
974 : | |||
975 : | |||
976 : | // IDCT tests // | ||
977 : | #if 0 | ||
978 : | // no finished yet | ||
979 : | |||
980 : | init_stats(&Stats); | ||
981 : | |||
982 : | // test each computation channel separately | ||
983 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? IDCT_MAX : 0; | ||
984 : | ref_idct(Blk0); | ||
985 : | emms(); idct(Blk); emms(); | ||
986 : | for(i=0; i<64; i++) { CLAMP(Blk0[i], IDCT_OUT); CLAMP(Blk[i], IDCT_OUT); } | ||
987 : | store_stats(&Stats, Blk, Blk0); | ||
988 : | |||
989 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? IDCT_MIN : 0; | ||
990 : | ref_idct(Blk0); | ||
991 : | emms(); idct(Blk); emms(); | ||
992 : | for(i=0; i<64; i++) { CLAMP(Blk0[i], IDCT_OUT); CLAMP(Blk[i], IDCT_OUT); } | ||
993 : | store_stats(&Stats, Blk, Blk0); | ||
994 : | |||
995 : | // randomly saturated inputs | ||
996 : | for(p=0; p<Partitions; ++p) | ||
997 : | { | ||
998 : | for(n=0; n<Loops; ++n) | ||
999 : | { | ||
1000 : | for(i=0; i<64; ++i) | ||
1001 : | Blk0[i] = Blk[i] = (ieee_rand(0,Partitions)>=p)? IDCT_MAX : IDCT_MIN; | ||
1002 : | ref_idct(Blk0); | ||
1003 : | emms(); idct(Blk); emms(); | ||
1004 : | for(i=0; i<64; i++) { CLAMP(Blk0[i],IDCT_OUT); CLAMP(Blk[i],IDCT_OUT); } | ||
1005 : | store_stats(&Stats, Blk, Blk0); | ||
1006 : | } | ||
1007 : | } | ||
1008 : | |||
1009 : | printf( "\n -- IDCT saturation report --\n" ); | ||
1010 : | print_stats(&Stats); | ||
1011 : | report_stats(&Stats, 0); | ||
1012 : | #endif | ||
1013 : | } | ||
1014 : | } | ||
1015 : | |||
1016 : | /********************************************************************* | ||
1017 : | Isibaar | 225 | * measure raw decoding speed |
1018 : | *********************************************************************/ | ||
1019 : | |||
1020 : | void test_dec(const char *name, int width, int height, int with_chksum) | ||
1021 : | { | ||
1022 : | FILE *f = 0; | ||
1023 : | void *dechandle = 0; | ||
1024 : | int xerr; | ||
1025 : | XVID_INIT_PARAM xinit; | ||
1026 : | XVID_DEC_PARAM xparam; | ||
1027 : | XVID_DEC_FRAME xframe; | ||
1028 : | double t = 0.; | ||
1029 : | int nb = 0; | ||
1030 : | uint8_t *buf = 0; | ||
1031 : | uint8_t *rgb_out = 0; | ||
1032 : | int buf_size, pos; | ||
1033 : | uint32_t chksum = 0; | ||
1034 : | |||
1035 : | Isibaar | 262 | xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; |
1036 : | Isibaar | 225 | xvid_init(NULL, 0, &xinit, NULL); |
1037 : | printf( "API version: %d, core build:%d\n", xinit.api_version, xinit.core_build); | ||
1038 : | |||
1039 : | |||
1040 : | xparam.width = width; | ||
1041 : | xparam.height = height; | ||
1042 : | xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); | ||
1043 : | if (xerr!=XVID_ERR_OK) { | ||
1044 : | printf("can't init decoder (err=%d)\n", xerr); | ||
1045 : | return; | ||
1046 : | } | ||
1047 : | dechandle = xparam.handle; | ||
1048 : | |||
1049 : | |||
1050 : | f = fopen(name, "rb"); | ||
1051 : | if (f==0) { | ||
1052 : | printf( "can't open file '%s'\n", name); | ||
1053 : | return; | ||
1054 : | } | ||
1055 : | fseek(f, 0, SEEK_END); | ||
1056 : | buf_size = ftell(f); | ||
1057 : | fseek(f, 0, SEEK_SET); | ||
1058 : | if (buf_size<=0) { | ||
1059 : | printf("error while stating file\n"); | ||
1060 : | goto End; | ||
1061 : | } | ||
1062 : | else printf( "Input size: %d\n", buf_size); | ||
1063 : | |||
1064 : | buf = malloc(buf_size); // should be enuf' | ||
1065 : | rgb_out = calloc(4, width*height); // <-room for _RGB24 | ||
1066 : | if (buf==0 || rgb_out==0) { | ||
1067 : | printf( "malloc failed!\n" ); | ||
1068 : | goto End; | ||
1069 : | } | ||
1070 : | |||
1071 : | if (fread(buf, buf_size, 1, f)!=1) { | ||
1072 : | printf( "file-read failed\n" ); | ||
1073 : | goto End; | ||
1074 : | } | ||
1075 : | |||
1076 : | nb = 0; | ||
1077 : | pos = 0; | ||
1078 : | t = -gettime_usec(); | ||
1079 : | while(1) { | ||
1080 : | xframe.bitstream = buf + pos; | ||
1081 : | xframe.length = buf_size - pos; | ||
1082 : | xframe.image = rgb_out; | ||
1083 : | xframe.stride = width; | ||
1084 : | xframe.colorspace = XVID_CSP_RGB24; | ||
1085 : | xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); | ||
1086 : | nb++; | ||
1087 : | pos += xframe.length; | ||
1088 : | if (with_chksum) { | ||
1089 : | int k = width*height; | ||
1090 : | uint32_t *ptr = (uint32_t *)rgb_out; | ||
1091 : | while(k-->0) chksum += *ptr++; | ||
1092 : | } | ||
1093 : | if (pos==buf_size) | ||
1094 : | break; | ||
1095 : | if (xerr!=XVID_ERR_OK) { | ||
1096 : | printf("decoding failed for frame #%d (err=%d)!\n", nb, xerr); | ||
1097 : | break; | ||
1098 : | } | ||
1099 : | } | ||
1100 : | t += gettime_usec(); | ||
1101 : | if (t>0.) | ||
1102 : | printf( "%d frames decoded in %.3f s -> %.1f FPS\n", nb, t*1.e-6f, (float)(nb*1.e6f/t) ); | ||
1103 : | if (with_chksum) | ||
1104 : | printf("checksum: 0x%.8x\n", chksum); | ||
1105 : | |||
1106 : | End: | ||
1107 : | if (rgb_out!=0) free(rgb_out); | ||
1108 : | if (buf!=0) free(buf); | ||
1109 : | if (dechandle!=0) { | ||
1110 : | xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); | ||
1111 : | if (xerr!=XVID_ERR_OK) | ||
1112 : | printf("destroy-decoder failed (err=%d)!\n", xerr); | ||
1113 : | } | ||
1114 : | if (f!=0) fclose(f); | ||
1115 : | } | ||
1116 : | |||
1117 : | /********************************************************************* | ||
1118 : | * non-regression tests | ||
1119 : | *********************************************************************/ | ||
1120 : | |||
1121 : | void test_bugs1() | ||
1122 : | { | ||
1123 : | CPU *cpu; | ||
1124 : | |||
1125 : | printf( "\n ===== (de)quant4_intra saturation bug? =====\n" ); | ||
1126 : | |||
1127 : | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) | ||
1128 : | { | ||
1129 : | int i; | ||
1130 : | int16_t Src[8*8], Dst[8*8]; | ||
1131 : | |||
1132 : | if (!init_cpu(cpu)) | ||
1133 : | continue; | ||
1134 : | |||
1135 : | for(i=0; i<64; ++i) Src[i] = i-32; | ||
1136 : | set_intra_matrix( get_default_intra_matrix() ); | ||
1137 : | Isibaar | 262 | dequant4_intra(Dst, Src, 31, 5); |
1138 : | Isibaar | 225 | printf( "dequant4_intra with CPU=%s: ", cpu->name); |
1139 : | printf( " Out[]= " ); | ||
1140 : | for(i=0; i<64; ++i) printf( "[%d]", Dst[i]); | ||
1141 : | printf( "\n" ); | ||
1142 : | } | ||
1143 : | |||
1144 : | printf( "\n ===== (de)quant4_inter saturation bug? =====\n" ); | ||
1145 : | |||
1146 : | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) | ||
1147 : | { | ||
1148 : | int i; | ||
1149 : | int16_t Src[8*8], Dst[8*8]; | ||
1150 : | |||
1151 : | if (!init_cpu(cpu)) | ||
1152 : | continue; | ||
1153 : | |||
1154 : | for(i=0; i<64; ++i) Src[i] = i-32; | ||
1155 : | set_inter_matrix( get_default_inter_matrix() ); | ||
1156 : | Isibaar | 262 | dequant4_inter(Dst, Src, 31); |
1157 : | Isibaar | 225 | printf( "dequant4_inter with CPU=%s: ", cpu->name); |
1158 : | printf( " Out[]= " ); | ||
1159 : | for(i=0; i<64; ++i) printf( "[%d]", Dst[i]); | ||
1160 : | printf( "\n" ); | ||
1161 : | } | ||
1162 : | } | ||
1163 : | |||
1164 : | void test_dct_precision_diffs() | ||
1165 : | { | ||
1166 : | CPU *cpu; | ||
1167 : | short Blk[8*8], Blk0[8*8]; | ||
1168 : | |||
1169 : | Isibaar | 262 | printf( "\n ===== fdct/idct precision diffs =====\n" ); |
1170 : | Isibaar | 225 | |
1171 : | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) | ||
1172 : | { | ||
1173 : | int i; | ||
1174 : | |||
1175 : | if (!init_cpu(cpu)) | ||
1176 : | continue; | ||
1177 : | |||
1178 : | for(i=0; i<8*8; ++i) { | ||
1179 : | Blk0[i] = (i*7-i*i) & 0x7f; | ||
1180 : | Blk[i] = Blk0[i]; | ||
1181 : | } | ||
1182 : | |||
1183 : | fdct(Blk); | ||
1184 : | idct(Blk); | ||
1185 : | printf( " fdct+idct diffs with CPU=%s: \n", cpu->name ); | ||
1186 : | for(i=0; i<8; ++i) { | ||
1187 : | int j; | ||
1188 : | for(j=0; j<8; ++j) printf( " %d ", Blk[i*8+j]-Blk0[i*8+j]); | ||
1189 : | printf("\n"); | ||
1190 : | } | ||
1191 : | printf("\n"); | ||
1192 : | } | ||
1193 : | } | ||
1194 : | |||
1195 : | Isibaar | 262 | void test_quant_bug() |
1196 : | { | ||
1197 : | const int max_Q = 31; | ||
1198 : | int i, n, qm, q; | ||
1199 : | CPU *cpu; | ||
1200 : | int16_t Src[8*8], Dst[8*8]; | ||
1201 : | uint8_t Quant[8*8]; | ||
1202 : | CPU cpu_bug_list[] = { { "PLAINC", 0 }, { "MMX ", XVID_CPU_MMX }, {0,0} }; | ||
1203 : | uint16_t Crcs_Inter[2][32]; | ||
1204 : | uint16_t Crcs_Intra[2][32]; | ||
1205 : | printf( "\n ===== test MPEG4-quantize bug =====\n" ); | ||
1206 : | Isibaar | 225 | |
1207 : | Isibaar | 262 | for(i=0; i<64; ++i) Src[i] = 2048*(i-32)/32; |
1208 : | |||
1209 : | #if 1 | ||
1210 : | for(qm=1; qm<=255; ++qm) | ||
1211 : | { | ||
1212 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
1213 : | set_inter_matrix( Quant ); | ||
1214 : | |||
1215 : | for(n=0, cpu = cpu_bug_list; cpu->name!=0; ++cpu, ++n) | ||
1216 : | { | ||
1217 : | uint16_t s; | ||
1218 : | |||
1219 : | if (!init_cpu(cpu)) | ||
1220 : | continue; | ||
1221 : | |||
1222 : | for(q=1; q<=max_Q; ++q) { | ||
1223 : | emms(); | ||
1224 : | quant4_inter( Dst, Src, q ); | ||
1225 : | emms(); | ||
1226 : | for(s=0, i=0; i<64; ++i) s+=((uint16_t)Dst[i])^i; | ||
1227 : | Crcs_Inter[n][q] = s; | ||
1228 : | } | ||
1229 : | } | ||
1230 : | |||
1231 : | for(q=1; q<=max_Q; ++q) | ||
1232 : | for(i=0; i<n-1; ++i) | ||
1233 : | if (Crcs_Inter[i][q]!=Crcs_Inter[i+1][q]) | ||
1234 : | printf( "Discrepancy Inter: qm=%d, q=%d -> %d/%d !\n", | ||
1235 : | qm, q, Crcs_Inter[i][q], Crcs_Inter[i+1][q]); | ||
1236 : | } | ||
1237 : | #endif | ||
1238 : | |||
1239 : | #if 1 | ||
1240 : | for(qm=1; qm<=255; ++qm) | ||
1241 : | { | ||
1242 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
1243 : | set_intra_matrix( Quant ); | ||
1244 : | |||
1245 : | for(n=0, cpu = cpu_bug_list; cpu->name!=0; ++cpu, ++n) | ||
1246 : | { | ||
1247 : | uint16_t s; | ||
1248 : | |||
1249 : | if (!init_cpu(cpu)) | ||
1250 : | continue; | ||
1251 : | |||
1252 : | for(q=1; q<=max_Q; ++q) { | ||
1253 : | emms(); | ||
1254 : | quant4_intra( Dst, Src, q, q); | ||
1255 : | emms(); | ||
1256 : | for(s=0, i=0; i<64; ++i) s+=((uint16_t)Dst[i])^i; | ||
1257 : | Crcs_Intra[n][q] = s; | ||
1258 : | } | ||
1259 : | } | ||
1260 : | |||
1261 : | for(q=1; q<=max_Q; ++q) | ||
1262 : | for(i=0; i<n-1; ++i) | ||
1263 : | if (Crcs_Intra[i][q]!=Crcs_Intra[i+1][q]) | ||
1264 : | printf( "Discrepancy Intra: qm=%d, q=%d -> %d/%d!\n", | ||
1265 : | qm, q, Crcs_Inter[i][q], Crcs_Inter[i+1][q]); | ||
1266 : | } | ||
1267 : | #endif | ||
1268 : | } | ||
1269 : | |||
1270 : | Isibaar | 225 | /********************************************************************* |
1271 : | * main | ||
1272 : | *********************************************************************/ | ||
1273 : | |||
1274 : | int main(int argc, char *argv[]) | ||
1275 : | { | ||
1276 : | int what = 0; | ||
1277 : | if (argc>1) what = atoi(argv[1]); | ||
1278 : | if (what==0 || what==1) test_dct(); | ||
1279 : | if (what==0 || what==2) test_mb(); | ||
1280 : | if (what==0 || what==3) test_sad(); | ||
1281 : | if (what==0 || what==4) test_transfer(); | ||
1282 : | if (what==0 || what==5) test_quant(); | ||
1283 : | if (what==0 || what==6) test_cbp(); | ||
1284 : | |||
1285 : | Isibaar | 262 | if (what==7) { |
1286 : | test_IEEE1180_compliance(-256, 255, 1); | ||
1287 : | #if 0 | ||
1288 : | test_IEEE1180_compliance(-256, 255,-1); | ||
1289 : | test_IEEE1180_compliance( -5, 5, 1); | ||
1290 : | test_IEEE1180_compliance( -5, 5,-1); | ||
1291 : | test_IEEE1180_compliance(-300, 300, 1); | ||
1292 : | test_IEEE1180_compliance(-300, 300,-1); | ||
1293 : | #endif | ||
1294 : | } | ||
1295 : | if (what==8) test_dct_saturation(-256, 255); | ||
1296 : | |||
1297 : | if (what==9) { | ||
1298 : | Isibaar | 225 | int width, height; |
1299 : | if (argc<5) { | ||
1300 : | printf("usage: %s %d [bitstream] [width] [height]\n", argv[0], what); | ||
1301 : | return 1; | ||
1302 : | } | ||
1303 : | width = atoi(argv[3]); | ||
1304 : | height = atoi(argv[4]); | ||
1305 : | test_dec(argv[2], width, height, (argc>5)); | ||
1306 : | } | ||
1307 : | |||
1308 : | if (what==-1) { | ||
1309 : | Isibaar | 262 | test_dct_precision_diffs(); |
1310 : | Isibaar | 225 | test_bugs1(); |
1311 : | } | ||
1312 : | Isibaar | 262 | if (what==-2) |
1313 : | test_quant_bug(); | ||
1314 : | |||
1315 : | Isibaar | 225 | return 0; |
1316 : | } | ||
1317 : | |||
1318 : | /********************************************************************* | ||
1319 : | * 'Reference' output (except for timing) on a PIII 1.13Ghz/linux | ||
1320 : | *********************************************************************/ | ||
1321 : | Isibaar | 262 | |
1322 : | /* as of 07/01/2002, there's a problem with mpeg4-quantization */ | ||
1323 : | Isibaar | 225 | /* |
1324 : | |||
1325 : | ===== test fdct/idct ===== | ||
1326 : | Isibaar | 262 | PLAINC - 3.312 usec PSNR=13.291 MSE=3.000 |
1327 : | MMX - 0.591 usec PSNR=13.291 MSE=3.000 | ||
1328 : | MMXEXT - 0.577 usec PSNR=13.291 MSE=3.000 | ||
1329 : | SSE2 - 0.588 usec PSNR=13.291 MSE=3.000 | ||
1330 : | Isibaar | 225 | 3DNOW - skipped... |
1331 : | 3DNOWE - skipped... | ||
1332 : | |||
1333 : | === test block motion === | ||
1334 : | Isibaar | 262 | PLAINC - interp- h-round0 0.911 usec iCrc=8107 |
1335 : | PLAINC - round1 0.863 usec iCrc=8100 | ||
1336 : | PLAINC - interp- v-round0 0.860 usec iCrc=8108 | ||
1337 : | PLAINC - round1 0.857 usec iCrc=8105 | ||
1338 : | PLAINC - interp-hv-round0 2.103 usec iCrc=8112 | ||
1339 : | PLAINC - round1 2.050 usec iCrc=8103 | ||
1340 : | Isibaar | 225 | --- |
1341 : | MMX - interp- h-round0 0.105 usec iCrc=8107 | ||
1342 : | Isibaar | 262 | MMX - round1 0.106 usec iCrc=8100 |
1343 : | Isibaar | 225 | MMX - interp- v-round0 0.106 usec iCrc=8108 |
1344 : | Isibaar | 262 | MMX - round1 0.106 usec iCrc=8105 |
1345 : | Isibaar | 225 | MMX - interp-hv-round0 0.145 usec iCrc=8112 |
1346 : | MMX - round1 0.145 usec iCrc=8103 | ||
1347 : | --- | ||
1348 : | Isibaar | 262 | MMXEXT - interp- h-round0 0.028 usec iCrc=8107 |
1349 : | Isibaar | 225 | MMXEXT - round1 0.041 usec iCrc=8100 |
1350 : | MMXEXT - interp- v-round0 0.027 usec iCrc=8108 | ||
1351 : | Isibaar | 262 | MMXEXT - round1 0.041 usec iCrc=8105 |
1352 : | MMXEXT - interp-hv-round0 0.066 usec iCrc=8112 | ||
1353 : | MMXEXT - round1 0.065 usec iCrc=8103 | ||
1354 : | Isibaar | 225 | --- |
1355 : | Isibaar | 262 | SSE2 - interp- h-round0 0.109 usec iCrc=8107 |
1356 : | Isibaar | 225 | SSE2 - round1 0.105 usec iCrc=8100 |
1357 : | SSE2 - interp- v-round0 0.106 usec iCrc=8108 | ||
1358 : | Isibaar | 262 | SSE2 - round1 0.109 usec iCrc=8105 |
1359 : | Isibaar | 225 | SSE2 - interp-hv-round0 0.145 usec iCrc=8112 |
1360 : | SSE2 - round1 0.145 usec iCrc=8103 | ||
1361 : | --- | ||
1362 : | 3DNOW - skipped... | ||
1363 : | 3DNOWE - skipped... | ||
1364 : | |||
1365 : | ====== test SAD ====== | ||
1366 : | Isibaar | 262 | PLAINC - sad8 0.251 usec sad=3776 |
1367 : | PLAINC - sad16 1.601 usec sad=27214 | ||
1368 : | PLAINC - sad16bi 2.371 usec sad=26274 | ||
1369 : | PLAINC - dev16 1.564 usec sad=3344 | ||
1370 : | Isibaar | 225 | --- |
1371 : | MMX - sad8 0.057 usec sad=3776 | ||
1372 : | Isibaar | 262 | MMX - sad16 0.182 usec sad=27214 |
1373 : | MMX - sad16bi 2.462 usec sad=26274 | ||
1374 : | MMX - dev16 0.311 usec sad=3344 | ||
1375 : | Isibaar | 225 | --- |
1376 : | MMXEXT - sad8 0.036 usec sad=3776 | ||
1377 : | Isibaar | 262 | MMXEXT - sad16 0.109 usec sad=27214 |
1378 : | MMXEXT - sad16bi 0.143 usec sad=26274 | ||
1379 : | MMXEXT - dev16 0.192 usec sad=3344 | ||
1380 : | Isibaar | 225 | --- |
1381 : | SSE2 - sad8 0.057 usec sad=3776 | ||
1382 : | Isibaar | 262 | SSE2 - sad16 0.179 usec sad=27214 |
1383 : | SSE2 - sad16bi 2.456 usec sad=26274 | ||
1384 : | SSE2 - dev16 0.321 usec sad=3344 | ||
1385 : | Isibaar | 225 | --- |
1386 : | 3DNOW - skipped... | ||
1387 : | 3DNOWE - skipped... | ||
1388 : | |||
1389 : | === test transfer === | ||
1390 : | Isibaar | 262 | PLAINC - 8to16 0.151 usec crc=28288 |
1391 : | PLAINC - 16to8 1.113 usec crc=28288 | ||
1392 : | PLAINC - 8to8 0.043 usec crc=20352 | ||
1393 : | PLAINC - 16to8add 1.069 usec crc=25536 | ||
1394 : | PLAINC - 8to16sub 0.631 usec crc1=28064 crc2=16256 | ||
1395 : | PLAINC - 8to16sub2 0.597 usec crc=20384 | ||
1396 : | Isibaar | 225 | --- |
1397 : | Isibaar | 262 | MMX - 8to16 0.032 usec crc=28288 |
1398 : | MMX - 16to8 0.024 usec crc=28288 | ||
1399 : | MMX - 8to8 0.020 usec crc=20352 | ||
1400 : | MMX - 16to8add 0.043 usec crc=25536 | ||
1401 : | MMX - 8to16sub 0.066 usec crc1=28064 crc2=16256 | ||
1402 : | MMX - 8to16sub2 0.111 usec crc=20384 | ||
1403 : | Isibaar | 225 | --- |
1404 : | |||
1405 : | ===== test quant ===== | ||
1406 : | Isibaar | 262 | PLAINC - quant4_intra 74.248 usec crc=29809 |
1407 : | PLAINC - quant4_inter 70.850 usec crc=12574 | ||
1408 : | PLAINC - dequant4_intra 40.628 usec crc=24052 | ||
1409 : | PLAINC - dequant4_inter 45.691 usec crc=63847 | ||
1410 : | PLAINC - quant_intra 43.357 usec crc=25662 | ||
1411 : | PLAINC - quant_inter 33.410 usec crc=23972 | ||
1412 : | PLAINC - dequant_intra 36.384 usec crc=49900 | ||
1413 : | PLAINC - dequant_inter 48.930 usec crc=48899 | ||
1414 : | Isibaar | 225 | --- |
1415 : | Isibaar | 262 | MMX - quant4_intra 7.445 usec crc=3459 |
1416 : | *** CRC ERROR! *** | ||
1417 : | MMX - quant4_inter 5.384 usec crc=51072 | ||
1418 : | *** CRC ERROR! *** | ||
1419 : | MMX - dequant4_intra 5.515 usec crc=24052 | ||
1420 : | MMX - dequant4_inter 7.745 usec crc=63847 | ||
1421 : | MMX - quant_intra 4.661 usec crc=25662 | ||
1422 : | MMX - quant_inter 4.406 usec crc=23972 | ||
1423 : | MMX - dequant_intra 4.928 usec crc=49900 | ||
1424 : | MMX - dequant_inter 4.532 usec crc=48899 | ||
1425 : | Isibaar | 225 | --- |
1426 : | |||
1427 : | ===== test cbp ===== | ||
1428 : | Isibaar | 262 | PLAINC - calc_cbp#1 0.371 usec cbp=0x15 |
1429 : | PLAINC - calc_cbp#2 0.432 usec cbp=0x38 | ||
1430 : | PLAINC - calc_cbp#3 0.339 usec cbp=0xf | ||
1431 : | PLAINC - calc_cbp#4 0.506 usec cbp=0x5 | ||
1432 : | Isibaar | 225 | --- |
1433 : | MMX - calc_cbp#1 0.136 usec cbp=0x15 | ||
1434 : | Isibaar | 262 | MMX - calc_cbp#2 0.134 usec cbp=0x38 |
1435 : | MMX - calc_cbp#3 0.138 usec cbp=0xf | ||
1436 : | Isibaar | 225 | MMX - calc_cbp#4 0.135 usec cbp=0x5 |
1437 : | --- | ||
1438 : | Isibaar | 262 | SSE2 - calc_cbp#1 0.136 usec cbp=0x15 |
1439 : | SSE2 - calc_cbp#2 0.133 usec cbp=0x38 | ||
1440 : | SSE2 - calc_cbp#3 0.133 usec cbp=0xf | ||
1441 : | SSE2 - calc_cbp#4 0.141 usec cbp=0x5 | ||
1442 : | Isibaar | 225 | --- |
1443 : | Isibaar | 262 | |
1444 : | Isibaar | 225 | */ |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |