Parent Directory | Revision Log
Revision 1.9.2.8 - (view) (download)
1 : | edgomez | 1.9.2.2 | /***************************************************************************** |
2 : | Isibaar | 1.1 | * |
3 : | edgomez | 1.9.2.2 | * XVID MPEG-4 VIDEO CODEC |
4 : | * - Unit tests and benches - | ||
5 : | Isibaar | 1.1 | * |
6 : | edgomez | 1.9.2.2 | * Copyright(C) 2002 Pascal Massimino <skal@planet-d.net> |
7 : | Isibaar | 1.1 | * |
8 : | edgomez | 1.9.2.2 | * This program is free software; you can redistribute it and/or modify |
9 : | * it under the terms of the GNU General Public License as published by | ||
10 : | * the Free Software Foundation; either version 2 of the License, or | ||
11 : | * (at your option) any later version. | ||
12 : | Isibaar | 1.1 | * |
13 : | edgomez | 1.9.2.2 | * This program is distributed in the hope that it will be useful, |
14 : | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
15 : | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
16 : | * GNU General Public License for more details. | ||
17 : | Isibaar | 1.1 | * |
18 : | edgomez | 1.9.2.2 | * You should have received a copy of the GNU General Public License |
19 : | * along with this program; if not, write to the Free Software | ||
20 : | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
21 : | * | ||
22 : | edgomez | 1.9.2.4 | * $Id$ |
23 : | edgomez | 1.9.2.2 | * |
24 : | ****************************************************************************/ | ||
25 : | Isibaar | 1.1 | |
26 : | edgomez | 1.9.2.2 | /***************************************************************************** |
27 : | Isibaar | 1.1 | * |
28 : | * 'Reference' output is at the end of file. | ||
29 : | * Don't take the checksums and crc too seriouly, they aren't | ||
30 : | Isibaar | 1.3 | * bullet-proof (should plug some .md5 here)... |
31 : | Isibaar | 1.1 | * |
32 : | * compiles with something like: | ||
33 : | * gcc -o xvid_bench xvid_bench.c -I../src/ -lxvidcore -lm | ||
34 : | * | ||
35 : | edgomez | 1.9.2.2 | ****************************************************************************/ |
36 : | Isibaar | 1.1 | |
37 : | #include <stdio.h> | ||
38 : | #include <stdlib.h> | ||
39 : | edgomez | 1.9.2.1 | #include <string.h> /* for memset */ |
40 : | Isibaar | 1.1 | #include <assert.h> |
41 : | |||
42 : | suxen_drol | 1.9 | #ifndef WIN32 |
43 : | edgomez | 1.9.2.1 | #include <sys/time.h> /* for gettimeofday */ |
44 : | suxen_drol | 1.9 | #else |
45 : | #include <time.h> | ||
46 : | #endif | ||
47 : | |||
48 : | |||
49 : | Isibaar | 1.1 | #include "xvid.h" |
50 : | |||
51 : | edgomez | 1.8 | // inner guts |
52 : | Isibaar | 1.1 | #include "dct/idct.h" |
53 : | #include "dct/fdct.h" | ||
54 : | #include "image/colorspace.h" | ||
55 : | #include "image/interpolate8x8.h" | ||
56 : | #include "utils/mem_transfer.h" | ||
57 : | edgomez | 1.9.2.5 | #include "quant/quant.h" |
58 : | Isibaar | 1.1 | #include "motion/sad.h" |
59 : | #include "utils/emms.h" | ||
60 : | #include "utils/timer.h" | ||
61 : | #include "quant/quant_matrix.c" | ||
62 : | #include "bitstream/cbp.h" | ||
63 : | |||
64 : | Isibaar | 1.3 | #include <math.h> |
65 : | suxen_drol | 1.9 | |
66 : | #ifndef M_PI | ||
67 : | #define M_PI 3.14159265358979323846 | ||
68 : | #endif | ||
69 : | |||
70 : | edgomez | 1.9.2.1 | const int speed_ref = 100; /* on slow machines, decrease this value */ |
71 : | Isibaar | 1.1 | |
72 : | /********************************************************************* | ||
73 : | * misc | ||
74 : | *********************************************************************/ | ||
75 : | |||
76 : | edgomez | 1.9.2.3 | /* returns time in micro-s*/ |
77 : | Isibaar | 1.1 | double gettime_usec() |
78 : | { | ||
79 : | suxen_drol | 1.9 | #ifndef WIN32 |
80 : | edgomez | 1.9.2.3 | struct timeval tv; |
81 : | gettimeofday(&tv, 0); | ||
82 : | return tv.tv_sec*1.0e6 + tv.tv_usec; | ||
83 : | suxen_drol | 1.9 | #else |
84 : | clock_t clk; | ||
85 : | clk = clock(); | ||
86 : | return clk * 1000000 / CLOCKS_PER_SEC; | ||
87 : | #endif | ||
88 : | Isibaar | 1.1 | } |
89 : | |||
90 : | edgomez | 1.9.2.3 | /* returns squared deviates (mean(v*v)-mean(v)^2) of a 8x8 block */ |
91 : | Isibaar | 1.1 | double sqr_dev(uint8_t v[8*8]) |
92 : | { | ||
93 : | edgomez | 1.9.2.3 | double sum=0.; |
94 : | double sum2=0.; | ||
95 : | int n; | ||
96 : | for (n=0;n<8*8;n++) | ||
97 : | { | ||
98 : | sum += v[n]; | ||
99 : | sum2 += v[n]*v[n]; | ||
100 : | } | ||
101 : | sum2 /= n; | ||
102 : | sum /= n; | ||
103 : | return sum2-sum*sum; | ||
104 : | Isibaar | 1.1 | } |
105 : | |||
106 : | /********************************************************************* | ||
107 : | * cpu init | ||
108 : | *********************************************************************/ | ||
109 : | |||
110 : | typedef struct { | ||
111 : | edgomez | 1.9.2.3 | const char *name; |
112 : | unsigned int cpu; | ||
113 : | Isibaar | 1.1 | } CPU; |
114 : | |||
115 : | CPU cpu_list[] = | ||
116 : | { { "PLAINC", 0 } | ||
117 : | edgomez | 1.9.2.4 | #ifdef ARCH_IS_IA32 |
118 : | edgomez | 1.9.2.3 | , { "MMX ", XVID_CPU_MMX } |
119 : | , { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } | ||
120 : | , { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } | ||
121 : | , { "3DNOW ", XVID_CPU_3DNOW } | ||
122 : | , { "3DNOWE", XVID_CPU_3DNOWEXT } | ||
123 : | edgomez | 1.9.2.4 | #endif |
124 : | edgomez | 1.9.2.3 | //, { "IA64 ", XVID_CPU_IA64 } |
125 : | edgomez | 1.8 | //, { "TSC ", XVID_CPU_TSC } |
126 : | edgomez | 1.9.2.3 | , { 0, 0 } }; |
127 : | Isibaar | 1.1 | |
128 : | edgomez | 1.9.2.3 | CPU cpu_short_list[] = |
129 : | Isibaar | 1.1 | { { "PLAINC", 0 } |
130 : | edgomez | 1.9.2.4 | #ifdef ARCH_IS_IA32 |
131 : | edgomez | 1.9.2.3 | , { "MMX ", XVID_CPU_MMX } |
132 : | edgomez | 1.8 | //, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
133 : | edgomez | 1.9.2.4 | #endif |
134 : | edgomez | 1.9.2.3 | //, { "IA64 ", XVID_CPU_IA64 } |
135 : | , { 0, 0 } }; | ||
136 : | Isibaar | 1.1 | |
137 : | edgomez | 1.9.2.3 | CPU cpu_short_list2[] = |
138 : | Isibaar | 1.1 | { { "PLAINC", 0 } |
139 : | edgomez | 1.9.2.4 | #ifdef ARCH_IS_IA32 |
140 : | edgomez | 1.9.2.3 | , { "MMX ", XVID_CPU_MMX } |
141 : | , { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } | ||
142 : | edgomez | 1.9.2.4 | #endif |
143 : | edgomez | 1.9.2.3 | , { 0, 0 } }; |
144 : | Isibaar | 1.1 | |
145 : | |||
146 : | int init_cpu(CPU *cpu) | ||
147 : | { | ||
148 : | edgomez | 1.9.2.6 | xvid_gbl_info_t xinfo; |
149 : | Isibaar | 1.1 | |
150 : | edgomez | 1.9.2.6 | /* Get the available CPU flags */ |
151 : | memset(&xinfo, 0, sizeof(xinfo)); | ||
152 : | xinfo.version = XVID_VERSION; | ||
153 : | xvid_global(NULL, XVID_GBL_INFO, &xinfo, NULL); | ||
154 : | |||
155 : | /* Are we trying to test a subset of the host CPU features */ | ||
156 : | if ((xinfo.cpu_flags & cpu->cpu) == cpu->cpu) { | ||
157 : | int xerr; | ||
158 : | xvid_gbl_init_t xinit; | ||
159 : | memset(&xinit, 0, sizeof(xinit)); | ||
160 : | xinit.cpu_flags = cpu->cpu | XVID_CPU_FORCE; | ||
161 : | xinit.version = XVID_VERSION; | ||
162 : | xerr = xvid_global(NULL, XVID_GBL_INIT, &xinit, NULL); | ||
163 : | if (xerr==XVID_ERR_FAIL) { | ||
164 : | /* libxvidcore failed to init */ | ||
165 : | return 0; | ||
166 : | } | ||
167 : | } else { | ||
168 : | /* The host CPU doesn't support some required feature for this test */ | ||
169 : | return(0); | ||
170 : | edgomez | 1.9.2.3 | } |
171 : | return 1; | ||
172 : | Isibaar | 1.1 | } |
173 : | |||
174 : | /********************************************************************* | ||
175 : | * test DCT | ||
176 : | *********************************************************************/ | ||
177 : | |||
178 : | #define ABS(X) ((X)<0 ? -(X) : (X)) | ||
179 : | |||
180 : | void test_dct() | ||
181 : | { | ||
182 : | edgomez | 1.9.2.3 | const int nb_tests = 300*speed_ref; |
183 : | int tst; | ||
184 : | CPU *cpu; | ||
185 : | int i; | ||
186 : | edgomez | 1.9.2.7 | DECLARE_ALIGNED_MATRIX(iDst0, 8, 8, short, 16); |
187 : | DECLARE_ALIGNED_MATRIX(iDst, 8, 8, short, 16); | ||
188 : | DECLARE_ALIGNED_MATRIX(fDst, 8, 8, short, 16); | ||
189 : | edgomez | 1.9.2.3 | double overhead; |
190 : | |||
191 : | printf( "\n ===== test fdct/idct =====\n" ); | ||
192 : | |||
193 : | for(i=0; i<8*8; ++i) iDst0[i] = (i*7-i*i) & 0x7f; | ||
194 : | overhead = gettime_usec(); | ||
195 : | for(tst=0; tst<nb_tests; ++tst) | ||
196 : | { | ||
197 : | for(i=0; i<8*8; ++i) fDst[i] = iDst0[i]; | ||
198 : | for(i=0; i<8*8; ++i) iDst[i] = fDst[i]; | ||
199 : | } | ||
200 : | overhead = gettime_usec() - overhead; | ||
201 : | |||
202 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
203 : | { | ||
204 : | double t, PSNR, MSE; | ||
205 : | |||
206 : | if (!init_cpu(cpu)) | ||
207 : | continue; | ||
208 : | |||
209 : | t = gettime_usec(); | ||
210 : | emms(); | ||
211 : | for(tst=0; tst<nb_tests; ++tst) | ||
212 : | { | ||
213 : | for(i=0; i<8*8; ++i) fDst[i] = iDst0[i]; | ||
214 : | fdct(fDst); | ||
215 : | for(i=0; i<8*8; ++i) iDst[i] = fDst[i]; | ||
216 : | idct(iDst); | ||
217 : | } | ||
218 : | emms(); | ||
219 : | t = (gettime_usec() - t - overhead) / nb_tests; | ||
220 : | MSE = 0.; | ||
221 : | for(i=0; i<8*8; ++i) { | ||
222 : | double delta = 1.0*(iDst[i] - iDst0[i]); | ||
223 : | MSE += delta*delta; | ||
224 : | } | ||
225 : | PSNR = (MSE==0.) ? 1.e6 : -4.3429448*log( MSE/64. ); | ||
226 : | printf( "%s - %.3f usec PSNR=%.3f MSE=%.3f\n", | ||
227 : | cpu->name, t, PSNR, MSE ); | ||
228 : | if (ABS(MSE)>=64) printf( "*** CRC ERROR! ***\n" ); | ||
229 : | } | ||
230 : | Isibaar | 1.1 | } |
231 : | |||
232 : | /********************************************************************* | ||
233 : | * test SAD | ||
234 : | *********************************************************************/ | ||
235 : | |||
236 : | void test_sad() | ||
237 : | { | ||
238 : | edgomez | 1.9.2.3 | const int nb_tests = 2000*speed_ref; |
239 : | int tst; | ||
240 : | CPU *cpu; | ||
241 : | int i; | ||
242 : | uint8_t Cur[16*16], Ref1[16*16], Ref2[16*16]; | ||
243 : | |||
244 : | printf( "\n ====== test SAD ======\n" ); | ||
245 : | for(i=0; i<16*16;++i) { | ||
246 : | Cur[i] = (i/5) ^ 0x05; | ||
247 : | Ref1[i] = (i + 0x0b) & 0xff; | ||
248 : | Ref2[i] = i ^ 0x76; | ||
249 : | } | ||
250 : | Isibaar | 1.1 | |
251 : | edgomez | 1.9.2.3 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
252 : | { | ||
253 : | double t; | ||
254 : | uint32_t s; | ||
255 : | if (!init_cpu(cpu)) | ||
256 : | continue; | ||
257 : | |||
258 : | t = gettime_usec(); | ||
259 : | emms(); | ||
260 : | for(tst=0; tst<nb_tests; ++tst) s = sad8(Cur, Ref1, 16); | ||
261 : | emms(); | ||
262 : | t = (gettime_usec() - t) / nb_tests; | ||
263 : | printf( "%s - sad8 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
264 : | if (s!=3776) printf( "*** CRC ERROR! ***\n" ); | ||
265 : | |||
266 : | t = gettime_usec(); | ||
267 : | emms(); | ||
268 : | for(tst=0; tst<nb_tests; ++tst) s = sad16(Cur, Ref1, 16, -1); | ||
269 : | emms(); | ||
270 : | t = (gettime_usec() - t) / nb_tests; | ||
271 : | printf( "%s - sad16 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
272 : | if (s!=27214) printf( "*** CRC ERROR! ***\n" ); | ||
273 : | |||
274 : | t = gettime_usec(); | ||
275 : | emms(); | ||
276 : | for(tst=0; tst<nb_tests; ++tst) s = sad16bi(Cur, Ref1, Ref2, 16); | ||
277 : | emms(); | ||
278 : | t = (gettime_usec() - t) / nb_tests; | ||
279 : | printf( "%s - sad16bi %.3f usec sad=%d\n", cpu->name, t, s ); | ||
280 : | if (s!=26274) printf( "*** CRC ERROR! ***\n" ); | ||
281 : | |||
282 : | t = gettime_usec(); | ||
283 : | emms(); | ||
284 : | for(tst=0; tst<nb_tests; ++tst) s = dev16(Cur, 16); | ||
285 : | emms(); | ||
286 : | t = (gettime_usec() - t) / nb_tests; | ||
287 : | printf( "%s - dev16 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
288 : | if (s!=3344) printf( "*** CRC ERROR! ***\n" ); | ||
289 : | |||
290 : | printf( " --- \n" ); | ||
291 : | } | ||
292 : | Isibaar | 1.1 | } |
293 : | |||
294 : | /********************************************************************* | ||
295 : | * test interpolation | ||
296 : | *********************************************************************/ | ||
297 : | |||
298 : | #define ENTER \ | ||
299 : | edgomez | 1.9.2.3 | for(i=0; i<16*8; ++i) Dst[i] = 0; \ |
300 : | t = gettime_usec(); \ | ||
301 : | emms(); | ||
302 : | Isibaar | 1.1 | |
303 : | #define LEAVE \ | ||
304 : | edgomez | 1.9.2.3 | emms(); \ |
305 : | t = (gettime_usec() - t) / nb_tests; \ | ||
306 : | iCrc = 0; \ | ||
307 : | for(i=0; i<16*8; ++i) { iCrc += Dst[i]^i; } | ||
308 : | Isibaar | 1.1 | |
309 : | #define TEST_MB(FUNC, R) \ | ||
310 : | edgomez | 1.9.2.3 | ENTER \ |
311 : | for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst, Src0, 16, (R)); \ | ||
312 : | LEAVE | ||
313 : | Isibaar | 1.1 | |
314 : | #define TEST_MB2(FUNC) \ | ||
315 : | edgomez | 1.9.2.3 | ENTER \ |
316 : | for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst, Src0, 16); \ | ||
317 : | LEAVE | ||
318 : | Isibaar | 1.1 | |
319 : | |||
320 : | void test_mb() | ||
321 : | { | ||
322 : | edgomez | 1.9.2.3 | const int nb_tests = 2000*speed_ref; |
323 : | CPU *cpu; | ||
324 : | const uint8_t Src0[16*9] = { | ||
325 : | /* try to have every possible combinaison of rounding... */ | ||
326 : | 0, 0, 1, 0, 2, 0, 3, 0, 4 ,0,0,0, 0,0,0,0 | ||
327 : | , 0, 1, 1, 1, 2, 1, 3, 1, 3 ,0,0,0, 0,0,0,0 | ||
328 : | , 0, 2, 1, 2, 2, 2, 3, 2, 2 ,0,0,0, 0,0,0,0 | ||
329 : | , 0, 3, 1, 3, 2, 3, 3, 3, 1 ,0,0,0, 0,0,0,0 | ||
330 : | , 1, 3, 0, 2, 1, 0, 2, 3, 4 ,0,0,0, 0,0,0,0 | ||
331 : | , 2, 2, 1, 2, 0, 1, 3, 5, 3 ,0,0,0, 0,0,0,0 | ||
332 : | , 3, 1, 2, 3, 1, 2, 2, 6, 2 ,0,0,0, 0,0,0,0 | ||
333 : | , 1, 0, 1, 3, 0, 3, 1, 6, 1 ,0,0,0, 0,0,0,0 | ||
334 : | , 4, 3, 2, 1, 2, 3, 4, 0, 3 ,0,0,0, 0,0,0,0 | ||
335 : | }; | ||
336 : | uint8_t Dst[16*8] = {0}; | ||
337 : | |||
338 : | printf( "\n === test block motion ===\n" ); | ||
339 : | |||
340 : | for(cpu = cpu_list; cpu->name!=0; ++cpu) | ||
341 : | { | ||
342 : | double t; | ||
343 : | int tst, i, iCrc; | ||
344 : | |||
345 : | if (!init_cpu(cpu)) | ||
346 : | continue; | ||
347 : | |||
348 : | TEST_MB(interpolate8x8_halfpel_h, 0); | ||
349 : | printf( "%s - interp- h-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
350 : | if (iCrc!=8107) printf( "*** CRC ERROR! ***\n" ); | ||
351 : | |||
352 : | TEST_MB(interpolate8x8_halfpel_h, 1); | ||
353 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
354 : | if (iCrc!=8100) printf( "*** CRC ERROR! ***\n" ); | ||
355 : | |||
356 : | |||
357 : | TEST_MB(interpolate8x8_halfpel_v, 0); | ||
358 : | printf( "%s - interp- v-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
359 : | if (iCrc!=8108) printf( "*** CRC ERROR! ***\n" ); | ||
360 : | |||
361 : | TEST_MB(interpolate8x8_halfpel_v, 1); | ||
362 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
363 : | if (iCrc!=8105) printf( "*** CRC ERROR! ***\n" ); | ||
364 : | |||
365 : | |||
366 : | TEST_MB(interpolate8x8_halfpel_hv, 0); | ||
367 : | printf( "%s - interp-hv-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
368 : | if (iCrc!=8112) printf( "*** CRC ERROR! ***\n" ); | ||
369 : | |||
370 : | TEST_MB(interpolate8x8_halfpel_hv, 1); | ||
371 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
372 : | if (iCrc!=8103) printf( "*** CRC ERROR! ***\n" ); | ||
373 : | Isibaar | 1.1 | |
374 : | Isibaar | 1.3 | |
375 : | edgomez | 1.9.2.3 | /* this is a new function, as of 06.06.2002 */ |
376 : | Isibaar | 1.3 | #if 0 |
377 : | edgomez | 1.9.2.3 | TEST_MB2(interpolate8x8_avrg); |
378 : | printf( "%s - interpolate8x8_c %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
379 : | if (iCrc!=8107) printf( "*** CRC ERROR! ***\n" ); | ||
380 : | Isibaar | 1.3 | #endif |
381 : | |||
382 : | edgomez | 1.9.2.3 | printf( " --- \n" ); |
383 : | } | ||
384 : | Isibaar | 1.1 | } |
385 : | |||
386 : | /********************************************************************* | ||
387 : | * test transfer | ||
388 : | *********************************************************************/ | ||
389 : | |||
390 : | #define INIT_TRANSFER \ | ||
391 : | edgomez | 1.9.2.3 | for(i=0; i<8*32; ++i) { \ |
392 : | Src8[i] = i; Src16[i] = i; \ | ||
393 : | Dst8[i] = 0; Dst16[i] = 0; \ | ||
394 : | Ref1[i] = i^0x27; \ | ||
395 : | Ref2[i] = i^0x51; \ | ||
396 : | } | ||
397 : | Isibaar | 1.1 | |
398 : | #define TEST_TRANSFER_BEGIN(DST) \ | ||
399 : | edgomez | 1.9.2.3 | INIT_TRANSFER \ |
400 : | overhead = -gettime_usec(); \ | ||
401 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
402 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
403 : | } \ | ||
404 : | overhead += gettime_usec(); \ | ||
405 : | t = gettime_usec(); \ | ||
406 : | emms(); \ | ||
407 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
408 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a; | ||
409 : | Isibaar | 1.1 | |
410 : | |||
411 : | #define TEST_TRANSFER_END(DST) \ | ||
412 : | edgomez | 1.9.2.3 | } \ |
413 : | emms(); \ | ||
414 : | t = (gettime_usec()-t -overhead) / nb_tests;\ | ||
415 : | s = 0; for(i=0; i<8*32; ++i) { s += (DST)[i]^i; } | ||
416 : | Isibaar | 1.1 | |
417 : | #define TEST_TRANSFER(FUNC, DST, SRC) \ | ||
418 : | edgomez | 1.9.2.3 | TEST_TRANSFER_BEGIN(DST); \ |
419 : | (FUNC)((DST), (SRC), 32); \ | ||
420 : | TEST_TRANSFER_END(DST) | ||
421 : | Isibaar | 1.1 | |
422 : | |||
423 : | #define TEST_TRANSFER2_BEGIN(DST, SRC) \ | ||
424 : | edgomez | 1.9.2.3 | INIT_TRANSFER \ |
425 : | overhead = -gettime_usec(); \ | ||
426 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
427 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
428 : | for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e;\ | ||
429 : | } \ | ||
430 : | overhead += gettime_usec(); \ | ||
431 : | t = gettime_usec(); \ | ||
432 : | emms(); \ | ||
433 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
434 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
435 : | for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e; | ||
436 : | Isibaar | 1.1 | |
437 : | #define TEST_TRANSFER2_END(DST) \ | ||
438 : | edgomez | 1.9.2.3 | } \ |
439 : | emms(); \ | ||
440 : | t = (gettime_usec()-t -overhead) / nb_tests;\ | ||
441 : | s = 0; for(i=0; i<8*32; ++i) { s += (DST)[i]; } | ||
442 : | Isibaar | 1.1 | |
443 : | #define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ | ||
444 : | edgomez | 1.9.2.3 | TEST_TRANSFER2_BEGIN(DST,SRC); \ |
445 : | (FUNC)((DST), (SRC), (R1), 32); \ | ||
446 : | TEST_TRANSFER2_END(DST) | ||
447 : | Isibaar | 1.1 | |
448 : | #define TEST_TRANSFER3(FUNC, DST, SRC, R1, R2)\ | ||
449 : | edgomez | 1.9.2.3 | TEST_TRANSFER_BEGIN(DST); \ |
450 : | (FUNC)((DST), (SRC), (R1), (R2), 32); \ | ||
451 : | TEST_TRANSFER_END(DST) | ||
452 : | Isibaar | 1.1 | |
453 : | void test_transfer() | ||
454 : | { | ||
455 : | edgomez | 1.9.2.3 | const int nb_tests = 4000*speed_ref; |
456 : | int i; | ||
457 : | CPU *cpu; | ||
458 : | uint8_t Src8[8*32], Dst8[8*32], Ref1[8*32], Ref2[8*32]; | ||
459 : | int16_t Src16[8*32], Dst16[8*32]; | ||
460 : | |||
461 : | printf( "\n === test transfer ===\n" ); | ||
462 : | |||
463 : | edgomez | 1.9.2.5 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
464 : | edgomez | 1.9.2.3 | { |
465 : | double t, overhead; | ||
466 : | int tst, s; | ||
467 : | |||
468 : | if (!init_cpu(cpu)) | ||
469 : | continue; | ||
470 : | |||
471 : | TEST_TRANSFER(transfer_8to16copy, Dst16, Src8); | ||
472 : | printf( "%s - 8to16 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
473 : | if (s!=28288) printf( "*** CRC ERROR! ***\n" ); | ||
474 : | |||
475 : | TEST_TRANSFER(transfer_16to8copy, Dst8, Src16); | ||
476 : | printf( "%s - 16to8 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
477 : | if (s!=28288) printf( "*** CRC ERROR! ***\n" ); | ||
478 : | |||
479 : | TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); | ||
480 : | printf( "%s - 8to8 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
481 : | if (s!=20352) printf( "*** CRC ERROR! ***\n" ); | ||
482 : | |||
483 : | TEST_TRANSFER(transfer_16to8add, Dst8, Src16); | ||
484 : | printf( "%s - 16to8add %.3f usec crc=%d\n", cpu->name, t, s ); | ||
485 : | if (s!=25536) printf( "*** CRC ERROR! ***\n" ); | ||
486 : | |||
487 : | TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); | ||
488 : | printf( "%s - 8to16sub %.3f usec crc1=%d ", cpu->name, t, s ); | ||
489 : | if (s!=28064) printf( "*** CRC ERROR! ***\n" ); | ||
490 : | s = 0; for(i=0; i<8*32; ++i) { s += (Src8[i]-Ref1[i])&i; } | ||
491 : | printf( "crc2=%d\n", s); | ||
492 : | if (s!=16256) printf( "*** CRC ERROR! ***\n" ); | ||
493 : | edgomez | 1.9.2.8 | |
494 : | edgomez | 1.9.2.3 | TEST_TRANSFER3(transfer_8to16sub2, Dst16, Src8, Ref1, Ref2); |
495 : | printf( "%s - 8to16sub2 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
496 : | edgomez | 1.9.2.8 | if (s!=22368) printf( "*** CRC ERROR! ***\n" ); |
497 : | |||
498 : | edgomez | 1.9.2.3 | printf( " --- \n" ); |
499 : | } | ||
500 : | Isibaar | 1.1 | } |
501 : | |||
502 : | /********************************************************************* | ||
503 : | * test quantization | ||
504 : | *********************************************************************/ | ||
505 : | |||
506 : | Isibaar | 1.3 | #define TEST_QUANT(FUNC, DST, SRC) \ |
507 : | edgomez | 1.9.2.3 | t = gettime_usec(); \ |
508 : | for(s=0,qm=1; qm<=255; ++qm) { \ | ||
509 : | for(i=0; i<8*8; ++i) Quant[i] = qm; \ | ||
510 : | set_inter_matrix( Quant ); \ | ||
511 : | emms(); \ | ||
512 : | for(q=1; q<=max_Q; ++q) { \ | ||
513 : | for(tst=0; tst<nb_tests; ++tst) \ | ||
514 : | (FUNC)((DST), (SRC), q); \ | ||
515 : | for(i=0; i<64; ++i) s+=(DST)[i]^i^qm; \ | ||
516 : | } \ | ||
517 : | emms(); \ | ||
518 : | } \ | ||
519 : | t = (gettime_usec()-t-overhead)/nb_tests/qm;\ | ||
520 : | s = (s&0xffff)^(s>>16) | ||
521 : | Isibaar | 1.3 | |
522 : | #define TEST_QUANT2(FUNC, DST, SRC) \ | ||
523 : | edgomez | 1.9.2.3 | t = gettime_usec(); \ |
524 : | for(s=0,qm=1; qm<=255; ++qm) { \ | ||
525 : | for(i=0; i<8*8; ++i) Quant[i] = qm; \ | ||
526 : | set_intra_matrix( Quant ); \ | ||
527 : | emms(); \ | ||
528 : | for(q=1; q<=max_Q; ++q) { \ | ||
529 : | for(tst=0; tst<nb_tests; ++tst) \ | ||
530 : | (FUNC)((DST), (SRC), q, q); \ | ||
531 : | for(i=0; i<64; ++i) s+=(DST)[i]^i^qm; \ | ||
532 : | } \ | ||
533 : | emms(); \ | ||
534 : | } \ | ||
535 : | t = (gettime_usec()-t-overhead)/nb_tests/qm;\ | ||
536 : | s = (s&0xffff)^(s>>16) | ||
537 : | Isibaar | 1.1 | |
538 : | void test_quant() | ||
539 : | { | ||
540 : | edgomez | 1.9.2.3 | const int nb_tests = 1*speed_ref; |
541 : | const int max_Q = 31; | ||
542 : | int i, qm; | ||
543 : | CPU *cpu; | ||
544 : | int16_t Src[8*8], Dst[8*8]; | ||
545 : | uint8_t Quant[8*8]; | ||
546 : | |||
547 : | printf( "\n ===== test quant =====\n" ); | ||
548 : | |||
549 : | /* we deliberately enfringe the norm's specified range [-127,127], */ | ||
550 : | /* to test the robustness of the iquant module */ | ||
551 : | for(i=0; i<64; ++i) { | ||
552 : | Src[i] = 1 + (i-32) * (i&6); | ||
553 : | Dst[i] = 0; | ||
554 : | } | ||
555 : | |||
556 : | edgomez | 1.9.2.5 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
557 : | edgomez | 1.9.2.3 | { |
558 : | double t, overhead; | ||
559 : | int tst, q; | ||
560 : | uint32_t s; | ||
561 : | |||
562 : | if (!init_cpu(cpu)) | ||
563 : | continue; | ||
564 : | |||
565 : | overhead = -gettime_usec(); | ||
566 : | for(s=0,qm=1; qm<=255; ++qm) { | ||
567 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
568 : | set_inter_matrix( Quant ); | ||
569 : | for(q=1; q<=max_Q; ++q) | ||
570 : | for(i=0; i<64; ++i) s+=Dst[i]^i^qm; | ||
571 : | } | ||
572 : | overhead += gettime_usec(); | ||
573 : | Isibaar | 1.1 | |
574 : | Isibaar | 1.3 | #if 1 |
575 : | edgomez | 1.9.2.5 | TEST_QUANT2(quant_mpeg_intra, Dst, Src); |
576 : | printf( "%s - quant_mpeg_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
577 : | edgomez | 1.9.2.3 | if (s!=29809) printf( "*** CRC ERROR! ***\n" ); |
578 : | |||
579 : | edgomez | 1.9.2.5 | TEST_QUANT(quant_mpeg_inter, Dst, Src); |
580 : | printf( "%s - quant_mpeg_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
581 : | edgomez | 1.9.2.3 | if (s!=12574) printf( "*** CRC ERROR! ***\n" ); |
582 : | Isibaar | 1.3 | #endif |
583 : | #if 1 | ||
584 : | edgomez | 1.9.2.5 | TEST_QUANT2(dequant_mpeg_intra, Dst, Src); |
585 : | printf( "%s - dequant_mpeg_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
586 : | edgomez | 1.9.2.3 | if (s!=24052) printf( "*** CRC ERROR! ***\n" ); |
587 : | |||
588 : | edgomez | 1.9.2.5 | TEST_QUANT(dequant_mpeg_inter, Dst, Src); |
589 : | printf( "%s - dequant_mpeg_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
590 : | edgomez | 1.9.2.3 | if (s!=63847) printf( "*** CRC ERROR! ***\n" ); |
591 : | Isibaar | 1.3 | #endif |
592 : | #if 1 | ||
593 : | edgomez | 1.9.2.5 | TEST_QUANT2(quant_h263_intra, Dst, Src); |
594 : | printf( "%s - quant_h263_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
595 : | edgomez | 1.9.2.3 | if (s!=25662) printf( "*** CRC ERROR! ***\n" ); |
596 : | |||
597 : | edgomez | 1.9.2.5 | TEST_QUANT(quant_h263_inter, Dst, Src); |
598 : | printf( "%s - quant_h263_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
599 : | edgomez | 1.9.2.3 | if (s!=23972) printf( "*** CRC ERROR! ***\n" ); |
600 : | Isibaar | 1.3 | #endif |
601 : | #if 1 | ||
602 : | edgomez | 1.9.2.5 | TEST_QUANT2(dequant_h263_intra, Dst, Src); |
603 : | printf( "%s - dequant_h263_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
604 : | edgomez | 1.9.2.3 | if (s!=49900) printf( "*** CRC ERROR! ***\n" ); |
605 : | |||
606 : | edgomez | 1.9.2.5 | TEST_QUANT(dequant_h263_inter, Dst, Src); |
607 : | printf( "%s - dequant_h263_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
608 : | edgomez | 1.9.2.3 | if (s!=48899) printf( "*** CRC ERROR! ***\n" ); |
609 : | Isibaar | 1.3 | #endif |
610 : | edgomez | 1.9.2.3 | printf( " --- \n" ); |
611 : | } | ||
612 : | Isibaar | 1.1 | } |
613 : | |||
614 : | /********************************************************************* | ||
615 : | * test non-zero AC counting | ||
616 : | *********************************************************************/ | ||
617 : | |||
618 : | #define TEST_CBP(FUNC, SRC) \ | ||
619 : | edgomez | 1.9.2.3 | t = gettime_usec(); \ |
620 : | emms(); \ | ||
621 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
622 : | cbp = (FUNC)((SRC)); \ | ||
623 : | } \ | ||
624 : | emms(); \ | ||
625 : | t = (gettime_usec()-t ) / nb_tests; | ||
626 : | Isibaar | 1.1 | |
627 : | void test_cbp() | ||
628 : | { | ||
629 : | edgomez | 1.9.2.3 | const int nb_tests = 10000*speed_ref; |
630 : | int i; | ||
631 : | CPU *cpu; | ||
632 : | int16_t Src1[6*64], Src2[6*64], Src3[6*64], Src4[6*64]; | ||
633 : | |||
634 : | printf( "\n ===== test cbp =====\n" ); | ||
635 : | |||
636 : | for(i=0; i<6*64; ++i) { | ||
637 : | Src1[i] = (i*i*3/8192)&(i/64)&1; /* 'random' */ | ||
638 : | Src2[i] = (i<3*64); /* half-full */ | ||
639 : | Src3[i] = ((i+32)>3*64); | ||
640 : | Src4[i] = (i==(3*64+2) || i==(5*64+9)); | ||
641 : | } | ||
642 : | |||
643 : | edgomez | 1.9.2.5 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
644 : | edgomez | 1.9.2.3 | { |
645 : | double t; | ||
646 : | int tst, cbp; | ||
647 : | |||
648 : | if (!init_cpu(cpu)) | ||
649 : | continue; | ||
650 : | |||
651 : | TEST_CBP(calc_cbp, Src1); | ||
652 : | printf( "%s - calc_cbp#1 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
653 : | if (cbp!=0x15) printf( "*** CRC ERROR! ***\n" ); | ||
654 : | TEST_CBP(calc_cbp, Src2); | ||
655 : | printf( "%s - calc_cbp#2 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
656 : | if (cbp!=0x38) printf( "*** CRC ERROR! ***\n" ); | ||
657 : | TEST_CBP(calc_cbp, Src3); | ||
658 : | printf( "%s - calc_cbp#3 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
659 : | if (cbp!=0x0f) printf( "*** CRC ERROR! ***\n" ); | ||
660 : | TEST_CBP(calc_cbp, Src4); | ||
661 : | printf( "%s - calc_cbp#4 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
662 : | if (cbp!=0x05) printf( "*** CRC ERROR! ***\n" ); | ||
663 : | printf( " --- \n" ); | ||
664 : | } | ||
665 : | Isibaar | 1.1 | } |
666 : | |||
667 : | /********************************************************************* | ||
668 : | Isibaar | 1.3 | * fdct/idct IEEE1180 compliance |
669 : | *********************************************************************/ | ||
670 : | |||
671 : | typedef struct { | ||
672 : | edgomez | 1.9.2.3 | long Errors[64]; |
673 : | long Sqr_Errors[64]; | ||
674 : | long Max_Errors[64]; | ||
675 : | long Nb; | ||
676 : | Isibaar | 1.3 | } STATS_8x8; |
677 : | |||
678 : | void init_stats(STATS_8x8 *S) | ||
679 : | { | ||
680 : | edgomez | 1.9.2.3 | int i; |
681 : | for(i=0; i<64; ++i) { | ||
682 : | S->Errors[i] = 0; | ||
683 : | S->Sqr_Errors[i] = 0; | ||
684 : | S->Max_Errors[i] = 0; | ||
685 : | } | ||
686 : | S->Nb = 0; | ||
687 : | Isibaar | 1.3 | } |
688 : | |||
689 : | void store_stats(STATS_8x8 *S, short Blk[64], short Ref[64]) | ||
690 : | { | ||
691 : | edgomez | 1.9.2.3 | int i; |
692 : | for(i=0; i<64; ++i) | ||
693 : | { | ||
694 : | short Err = Blk[i] - Ref[i]; | ||
695 : | S->Errors[i] += Err; | ||
696 : | S->Sqr_Errors[i] += Err * Err; | ||
697 : | if (Err<0) Err = -Err; | ||
698 : | if (S->Max_Errors[i]<Err) | ||
699 : | S->Max_Errors[i] = Err; | ||
700 : | } | ||
701 : | S->Nb++; | ||
702 : | Isibaar | 1.3 | } |
703 : | |||
704 : | void print_stats(STATS_8x8 *S) | ||
705 : | { | ||
706 : | edgomez | 1.9.2.3 | int i; |
707 : | double Norm; | ||
708 : | |||
709 : | assert(S->Nb>0); | ||
710 : | Norm = 1. / (double)S->Nb; | ||
711 : | printf("\n== Max absolute values of errors ==\n"); | ||
712 : | for(i=0; i<64; i++) { | ||
713 : | printf(" %4ld", S->Max_Errors[i]); | ||
714 : | if ((i&7)==7) printf("\n"); | ||
715 : | } | ||
716 : | |||
717 : | printf("\n== Mean square errors ==\n"); | ||
718 : | for(i=0; i<64; i++) | ||
719 : | { | ||
720 : | double Err = Norm * (double)S->Sqr_Errors[i]; | ||
721 : | printf(" %.3f", Err); | ||
722 : | if ((i&7)==7) printf("\n"); | ||
723 : | } | ||
724 : | Isibaar | 1.3 | |
725 : | edgomez | 1.9.2.3 | printf("\n== Mean errors ==\n"); |
726 : | for(i=0; i<64; i++) | ||
727 : | { | ||
728 : | double Err = Norm * (double)S->Errors[i]; | ||
729 : | printf(" %.3f", Err); | ||
730 : | if ((i&7)==7) printf("\n"); | ||
731 : | } | ||
732 : | printf("\n"); | ||
733 : | Isibaar | 1.3 | } |
734 : | |||
735 : | static const char *CHECK(double v, double l) { | ||
736 : | edgomez | 1.9.2.3 | if (fabs(v)<=l) return "ok"; |
737 : | else return "FAIL!"; | ||
738 : | Isibaar | 1.3 | } |
739 : | |||
740 : | void report_stats(STATS_8x8 *S, const double *Limits) | ||
741 : | { | ||
742 : | edgomez | 1.9.2.3 | int i; |
743 : | double Norm, PE, PMSE, OMSE, PME, OME; | ||
744 : | |||
745 : | assert(S->Nb>0); | ||
746 : | Norm = 1. / (double)S->Nb; | ||
747 : | PE = 0.; | ||
748 : | for(i=0; i<64; i++) { | ||
749 : | if (PE<S->Max_Errors[i]) | ||
750 : | PE = S->Max_Errors[i]; | ||
751 : | } | ||
752 : | |||
753 : | PMSE = 0.; | ||
754 : | OMSE = 0.; | ||
755 : | for(i=0; i<64; i++) | ||
756 : | { | ||
757 : | double Err = Norm * (double)S->Sqr_Errors[i]; | ||
758 : | OMSE += Err; | ||
759 : | if (PMSE < Err) PMSE = Err; | ||
760 : | } | ||
761 : | OMSE /= 64.; | ||
762 : | |||
763 : | PME = 0.; | ||
764 : | OME = 0.; | ||
765 : | for(i=0; i<64; i++) | ||
766 : | { | ||
767 : | double Err = Norm * (double)S->Errors[i]; | ||
768 : | OME += Err; | ||
769 : | Err = fabs(Err); | ||
770 : | if (PME < Err) PME = Err; | ||
771 : | } | ||
772 : | OME /= 64.; | ||
773 : | Isibaar | 1.3 | |
774 : | edgomez | 1.9.2.3 | printf( "Peak error: %4.4f\n", PE ); |
775 : | printf( "Peak MSE: %4.4f\n", PMSE ); | ||
776 : | printf( "Overall MSE: %4.4f\n", OMSE ); | ||
777 : | printf( "Peak ME: %4.4f\n", PME ); | ||
778 : | printf( "Overall ME: %4.4f\n", OME ); | ||
779 : | |||
780 : | if (Limits!=0) | ||
781 : | { | ||
782 : | printf( "[PE<=%.4f %s] ", Limits[0], CHECK(PE, Limits[0]) ); | ||
783 : | printf( "\n" ); | ||
784 : | printf( "[PMSE<=%.4f %s]", Limits[1], CHECK(PMSE, Limits[1]) ); | ||
785 : | printf( "[OMSE<=%.4f %s]", Limits[2], CHECK(OMSE, Limits[2]) ); | ||
786 : | printf( "\n" ); | ||
787 : | printf( "[PME<=%.4f %s] ", Limits[3], CHECK(PME , Limits[3]) ); | ||
788 : | printf( "[OME<=%.4f %s] ", Limits[4], CHECK(OME , Limits[4]) ); | ||
789 : | printf( "\n" ); | ||
790 : | } | ||
791 : | Isibaar | 1.3 | } |
792 : | |||
793 : | edgomez | 1.9.2.1 | ///* ////////////////////////////////////////////////////// */ |
794 : | Isibaar | 1.3 | /* Pseudo-random generator specified by IEEE 1180 */ |
795 : | |||
796 : | static long ieee_seed = 1; | ||
797 : | static void ieee_reseed(long s) { | ||
798 : | edgomez | 1.9.2.3 | ieee_seed = s; |
799 : | Isibaar | 1.3 | } |
800 : | static long ieee_rand(int Min, int Max) | ||
801 : | { | ||
802 : | edgomez | 1.9.2.3 | static double z = (double) 0x7fffffff; |
803 : | Isibaar | 1.3 | |
804 : | edgomez | 1.9.2.3 | long i,j; |
805 : | double x; | ||
806 : | Isibaar | 1.3 | |
807 : | edgomez | 1.9.2.3 | ieee_seed = (ieee_seed * 1103515245) + 12345; |
808 : | i = ieee_seed & 0x7ffffffe; | ||
809 : | x = ((double) i) / z; | ||
810 : | x *= (Max-Min+1); | ||
811 : | j = (long)x; | ||
812 : | j = j + Min; | ||
813 : | assert(j>=Min && j<=Max); | ||
814 : | return (short)j; | ||
815 : | Isibaar | 1.3 | } |
816 : | |||
817 : | #define CLAMP(x, M) (x) = ((x)<-(M)) ? (-(M)) : ((x)>=(M) ? ((M)-1) : (x)) | ||
818 : | |||
819 : | static double Cos[8][8]; | ||
820 : | static void init_ref_dct() | ||
821 : | { | ||
822 : | edgomez | 1.9.2.3 | int i, j; |
823 : | for(i=0; i<8; i++) | ||
824 : | { | ||
825 : | double scale = (i == 0) ? sqrt(0.125) : 0.5; | ||
826 : | for (j=0; j<8; j++) | ||
827 : | Cos[i][j] = scale*cos( (M_PI/8.0)*i*(j + 0.5) ); | ||
828 : | } | ||
829 : | Isibaar | 1.3 | } |
830 : | |||
831 : | void ref_idct(short *M) | ||
832 : | { | ||
833 : | edgomez | 1.9.2.3 | int i, j, k; |
834 : | double Tmp[8][8]; | ||
835 : | Isibaar | 1.3 | |
836 : | edgomez | 1.9.2.3 | for(i=0; i<8; i++) { |
837 : | for(j=0; j<8; j++) | ||
838 : | { | ||
839 : | double Sum = 0.0; | ||
840 : | for (k=0; k<8; k++) Sum += Cos[k][j]*M[8*i+k]; | ||
841 : | Tmp[i][j] = Sum; | ||
842 : | } | ||
843 : | } | ||
844 : | for(i=0; i<8; i++) { | ||
845 : | for(j=0; j<8; j++) { | ||
846 : | double Sum = 0.0; | ||
847 : | for (k=0; k<8; k++) Sum += Cos[k][i]*Tmp[k][j]; | ||
848 : | M[8*i+j] = (short)floor(Sum + .5); | ||
849 : | } | ||
850 : | } | ||
851 : | Isibaar | 1.3 | } |
852 : | |||
853 : | void ref_fdct(short *M) | ||
854 : | { | ||
855 : | edgomez | 1.9.2.3 | int i, j, k; |
856 : | double Tmp[8][8]; | ||
857 : | Isibaar | 1.3 | |
858 : | edgomez | 1.9.2.3 | for(i=0; i<8; i++) { |
859 : | for(j=0; j<8; j++) | ||
860 : | { | ||
861 : | double Sum = 0.0; | ||
862 : | for (k=0; k<8; k++) Sum += Cos[j][k]*M[8*i+k]; | ||
863 : | Tmp[i][j] = Sum; | ||
864 : | } | ||
865 : | } | ||
866 : | for(i=0; i<8; i++) { | ||
867 : | for(j=0; j<8; j++) { | ||
868 : | double Sum = 0.0; | ||
869 : | for (k=0; k<8; k++) Sum += Cos[i][k]*Tmp[k][j]; | ||
870 : | M[8*i+j] = (short)floor(Sum + 0.5); | ||
871 : | } | ||
872 : | } | ||
873 : | Isibaar | 1.3 | } |
874 : | |||
875 : | void test_IEEE1180_compliance(int Min, int Max, int Sign) | ||
876 : | { | ||
877 : | edgomez | 1.9.2.3 | static const double ILimits[5] = { 1., 0.06, 0.02, 0.015, 0.0015 }; |
878 : | int Loops = 10000; | ||
879 : | int i, m, n; | ||
880 : | edgomez | 1.9.2.7 | DECLARE_ALIGNED_MATRIX(Blk0, 8, 8, short, 16); /* reference */ |
881 : | DECLARE_ALIGNED_MATRIX(Blk, 8, 8, short, 16); | ||
882 : | DECLARE_ALIGNED_MATRIX(iBlk, 8, 8, short, 16); | ||
883 : | DECLARE_ALIGNED_MATRIX(Ref_FDCT, 8, 8, short, 16); | ||
884 : | DECLARE_ALIGNED_MATRIX(Ref_IDCT, 8, 8, short, 16); | ||
885 : | Isibaar | 1.3 | |
886 : | edgomez | 1.9.2.3 | STATS_8x8 FStats; /* forward dct stats */ |
887 : | STATS_8x8 IStats; /* inverse dct stats */ | ||
888 : | Isibaar | 1.3 | |
889 : | edgomez | 1.9.2.3 | CPU *cpu; |
890 : | Isibaar | 1.3 | |
891 : | edgomez | 1.9.2.3 | init_ref_dct(); |
892 : | Isibaar | 1.3 | |
893 : | edgomez | 1.9.2.3 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
894 : | { | ||
895 : | if (!init_cpu(cpu)) | ||
896 : | continue; | ||
897 : | Isibaar | 1.3 | |
898 : | edgomez | 1.9.2.3 | printf( "\n===== IEEE test for %s ==== (Min=%d Max=%d Sign=%d Loops=%d)\n", |
899 : | cpu->name, Min, Max, Sign, Loops); | ||
900 : | Isibaar | 1.3 | |
901 : | edgomez | 1.9.2.3 | init_stats(&IStats); |
902 : | init_stats(&FStats); | ||
903 : | Isibaar | 1.3 | |
904 : | edgomez | 1.9.2.3 | ieee_reseed(1); |
905 : | for(n=0; n<Loops; ++n) | ||
906 : | { | ||
907 : | for(i=0; i<64; ++i) | ||
908 : | Blk0[i] = (short)ieee_rand(Min,Max) * Sign; | ||
909 : | Isibaar | 1.3 | |
910 : | edgomez | 1.9.2.3 | /* hmm, I'm not quite sure this is exactly */ |
911 : | /* the tests described in the norm. check... */ | ||
912 : | Isibaar | 1.3 | |
913 : | edgomez | 1.9.2.3 | memcpy(Ref_FDCT, Blk0, 64*sizeof(short)); |
914 : | ref_fdct(Ref_FDCT); | ||
915 : | for(i=0; i<64; i++) CLAMP( Ref_FDCT[i], 2048 ); | ||
916 : | Isibaar | 1.3 | |
917 : | edgomez | 1.9.2.3 | memcpy(Blk, Blk0, 64*sizeof(short)); |
918 : | emms(); fdct(Blk); emms(); | ||
919 : | for(i=0; i<64; i++) CLAMP( Blk[i], 2048 ); | ||
920 : | Isibaar | 1.3 | |
921 : | edgomez | 1.9.2.3 | store_stats(&FStats, Blk, Ref_FDCT); |
922 : | Isibaar | 1.3 | |
923 : | |||
924 : | edgomez | 1.9.2.3 | memcpy(Ref_IDCT, Ref_FDCT, 64*sizeof(short)); |
925 : | ref_idct(Ref_IDCT); | ||
926 : | for (i=0; i<64; i++) CLAMP( Ref_IDCT[i], 256 ); | ||
927 : | Isibaar | 1.3 | |
928 : | edgomez | 1.9.2.3 | memcpy(iBlk, Ref_FDCT, 64*sizeof(short)); |
929 : | emms(); idct(iBlk); emms(); | ||
930 : | for(i=0; i<64; i++) CLAMP( iBlk[i], 256 ); | ||
931 : | Isibaar | 1.3 | |
932 : | edgomez | 1.9.2.3 | store_stats(&IStats, iBlk, Ref_IDCT); |
933 : | } | ||
934 : | Isibaar | 1.3 | |
935 : | |||
936 : | edgomez | 1.9.2.3 | printf( "\n -- FDCT report --\n" ); |
937 : | edgomez | 1.8 | // print_stats(&FStats); |
938 : | edgomez | 1.9.2.3 | report_stats(&FStats, 0); /* so far I know, IEEE1180 says nothing for fdct */ |
939 : | Isibaar | 1.3 | |
940 : | edgomez | 1.9.2.3 | for(i=0; i<64; i++) Blk[i] = 0; |
941 : | emms(); fdct(Blk); emms(); | ||
942 : | for(m=i=0; i<64; i++) if (Blk[i]!=0) m++; | ||
943 : | printf( "FDCT(0) == 0 ? %s\n", (m!=0) ? "NOPE!" : "yup." ); | ||
944 : | Isibaar | 1.3 | |
945 : | edgomez | 1.9.2.3 | printf( "\n -- IDCT report --\n" ); |
946 : | edgomez | 1.8 | // print_stats(&IStats); |
947 : | edgomez | 1.9.2.3 | report_stats(&IStats, ILimits); |
948 : | Isibaar | 1.3 | |
949 : | |||
950 : | edgomez | 1.9.2.3 | for(i=0; i<64; i++) Blk[i] = 0; |
951 : | emms(); idct(Blk); emms(); | ||
952 : | for(m=i=0; i<64; i++) if (Blk[i]!=0) m++; | ||
953 : | printf( "IDCT(0) == 0 ? %s\n", (m!=0) ? "NOPE!" : "yup." ); | ||
954 : | } | ||
955 : | Isibaar | 1.3 | } |
956 : | |||
957 : | |||
958 : | void test_dct_saturation(int Min, int Max) | ||
959 : | { | ||
960 : | edgomez | 1.9.2.3 | /* test behaviour on input range fringe */ |
961 : | Isibaar | 1.3 | |
962 : | edgomez | 1.9.2.3 | int i, n, p; |
963 : | CPU *cpu; | ||
964 : | edgomez | 1.9.2.1 | // const short IDCT_MAX = 2047; /* 12bits input */ |
965 : | edgomez | 1.8 | // const short IDCT_MIN = -2048; |
966 : | edgomez | 1.9.2.1 | // const short IDCT_OUT = 256; /* 9bits ouput */ |
967 : | edgomez | 1.9.2.3 | const int Partitions = 4; |
968 : | const int Loops = 10000 / Partitions; | ||
969 : | Isibaar | 1.3 | |
970 : | edgomez | 1.9.2.3 | init_ref_dct(); |
971 : | Isibaar | 1.3 | |
972 : | edgomez | 1.9.2.3 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
973 : | { | ||
974 : | short Blk0[64], Blk[64]; | ||
975 : | STATS_8x8 Stats; | ||
976 : | |||
977 : | if (!init_cpu(cpu)) | ||
978 : | continue; | ||
979 : | |||
980 : | printf( "\n===== IEEE test for %s Min=%d Max=%d =====\n", | ||
981 : | cpu->name, Min, Max ); | ||
982 : | |||
983 : | /* FDCT tests // */ | ||
984 : | |||
985 : | init_stats(&Stats); | ||
986 : | |||
987 : | /* test each computation channels separately */ | ||
988 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? Max : 0; | ||
989 : | ref_fdct(Blk0); | ||
990 : | emms(); fdct(Blk); emms(); | ||
991 : | store_stats(&Stats, Blk, Blk0); | ||
992 : | |||
993 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? Min : 0; | ||
994 : | ref_fdct(Blk0); | ||
995 : | emms(); fdct(Blk); emms(); | ||
996 : | store_stats(&Stats, Blk, Blk0); | ||
997 : | |||
998 : | /* randomly saturated inputs */ | ||
999 : | for(p=0; p<Partitions; ++p) | ||
1000 : | { | ||
1001 : | for(n=0; n<Loops; ++n) | ||
1002 : | { | ||
1003 : | for(i=0; i<64; ++i) | ||
1004 : | Blk0[i] = Blk[i] = (ieee_rand(0,Partitions)>=p)? Max : Min; | ||
1005 : | ref_fdct(Blk0); | ||
1006 : | emms(); fdct(Blk); emms(); | ||
1007 : | store_stats(&Stats, Blk, Blk0); | ||
1008 : | } | ||
1009 : | } | ||
1010 : | printf( "\n -- FDCT saturation report --\n" ); | ||
1011 : | report_stats(&Stats, 0); | ||
1012 : | Isibaar | 1.3 | |
1013 : | |||
1014 : | edgomez | 1.9.2.3 | /* IDCT tests // */ |
1015 : | Isibaar | 1.3 | #if 0 |
1016 : | edgomez | 1.9.2.3 | /* no finished yet */ |
1017 : | Isibaar | 1.3 | |
1018 : | edgomez | 1.9.2.3 | init_stats(&Stats); |
1019 : | Isibaar | 1.3 | |
1020 : | edgomez | 1.9.2.3 | /* test each computation channel separately */ |
1021 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? IDCT_MAX : 0; | ||
1022 : | ref_idct(Blk0); | ||
1023 : | emms(); idct(Blk); emms(); | ||
1024 : | for(i=0; i<64; i++) { CLAMP(Blk0[i], IDCT_OUT); CLAMP(Blk[i], IDCT_OUT); } | ||
1025 : | store_stats(&Stats, Blk, Blk0); | ||
1026 : | |||
1027 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? IDCT_MIN : 0; | ||
1028 : | ref_idct(Blk0); | ||
1029 : | emms(); idct(Blk); emms(); | ||
1030 : | for(i=0; i<64; i++) { CLAMP(Blk0[i], IDCT_OUT); CLAMP(Blk[i], IDCT_OUT); } | ||
1031 : | store_stats(&Stats, Blk, Blk0); | ||
1032 : | |||
1033 : | /* randomly saturated inputs */ | ||
1034 : | for(p=0; p<Partitions; ++p) | ||
1035 : | { | ||
1036 : | for(n=0; n<Loops; ++n) | ||
1037 : | { | ||
1038 : | for(i=0; i<64; ++i) | ||
1039 : | Blk0[i] = Blk[i] = (ieee_rand(0,Partitions)>=p)? IDCT_MAX : IDCT_MIN; | ||
1040 : | ref_idct(Blk0); | ||
1041 : | emms(); idct(Blk); emms(); | ||
1042 : | for(i=0; i<64; i++) { CLAMP(Blk0[i],IDCT_OUT); CLAMP(Blk[i],IDCT_OUT); } | ||
1043 : | store_stats(&Stats, Blk, Blk0); | ||
1044 : | } | ||
1045 : | } | ||
1046 : | |||
1047 : | printf( "\n -- IDCT saturation report --\n" ); | ||
1048 : | print_stats(&Stats); | ||
1049 : | report_stats(&Stats, 0); | ||
1050 : | Isibaar | 1.3 | #endif |
1051 : | edgomez | 1.9.2.3 | } |
1052 : | Isibaar | 1.3 | } |
1053 : | |||
1054 : | /********************************************************************* | ||
1055 : | Isibaar | 1.1 | * measure raw decoding speed |
1056 : | *********************************************************************/ | ||
1057 : | |||
1058 : | void test_dec(const char *name, int width, int height, int with_chksum) | ||
1059 : | { | ||
1060 : | edgomez | 1.9.2.3 | FILE *f = 0; |
1061 : | void *dechandle = 0; | ||
1062 : | int xerr; | ||
1063 : | xvid_gbl_init_t xinit; | ||
1064 : | xvid_dec_create_t xparam; | ||
1065 : | xvid_dec_frame_t xframe; | ||
1066 : | Isibaar | 1.1 | double t = 0.; |
1067 : | int nb = 0; | ||
1068 : | edgomez | 1.9.2.3 | uint8_t *buf = 0; |
1069 : | uint8_t *rgb_out = 0; | ||
1070 : | int buf_size, pos; | ||
1071 : | uint32_t chksum = 0; | ||
1072 : | Isibaar | 1.1 | |
1073 : | edgomez | 1.9.2.3 | memset(&xinit, 0, sizeof(xinit)); |
1074 : | Isibaar | 1.3 | xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; |
1075 : | edgomez | 1.9.2.3 | xinit.version = XVID_VERSION; |
1076 : | xvid_global(NULL, 0, &xinit, NULL); | ||
1077 : | Isibaar | 1.1 | |
1078 : | edgomez | 1.9.2.3 | memset(&xparam, 0, sizeof(xparam)); |
1079 : | Isibaar | 1.1 | xparam.width = width; |
1080 : | xparam.height = height; | ||
1081 : | edgomez | 1.9.2.3 | xparam.version = XVID_VERSION; |
1082 : | Isibaar | 1.1 | xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
1083 : | edgomez | 1.9.2.3 | if (xerr==XVID_ERR_FAIL) { |
1084 : | printf("can't init decoder (err=%d)\n", xerr); | ||
1085 : | return; | ||
1086 : | Isibaar | 1.1 | } |
1087 : | dechandle = xparam.handle; | ||
1088 : | |||
1089 : | |||
1090 : | f = fopen(name, "rb"); | ||
1091 : | edgomez | 1.9.2.3 | if (f==0) { |
1092 : | printf( "can't open file '%s'\n", name); | ||
1093 : | return; | ||
1094 : | } | ||
1095 : | fseek(f, 0, SEEK_END); | ||
1096 : | buf_size = ftell(f); | ||
1097 : | fseek(f, 0, SEEK_SET); | ||
1098 : | if (buf_size<=0) { | ||
1099 : | printf("error while stating file\n"); | ||
1100 : | goto End; | ||
1101 : | } | ||
1102 : | else printf( "Input size: %d\n", buf_size); | ||
1103 : | |||
1104 : | buf = malloc(buf_size); /* should be enuf' */ | ||
1105 : | rgb_out = calloc(4, width*height); /* <-room for _RGB24 */ | ||
1106 : | if (buf==0 || rgb_out==0) { | ||
1107 : | printf( "malloc failed!\n" ); | ||
1108 : | goto End; | ||
1109 : | } | ||
1110 : | |||
1111 : | if (fread(buf, buf_size, 1, f)!=1) { | ||
1112 : | printf( "file-read failed\n" ); | ||
1113 : | goto End; | ||
1114 : | } | ||
1115 : | |||
1116 : | nb = 0; | ||
1117 : | pos = 0; | ||
1118 : | t = -gettime_usec(); | ||
1119 : | while(1) { | ||
1120 : | memset(&xframe, 0, sizeof(xframe)); | ||
1121 : | xframe.version = XVID_VERSION; | ||
1122 : | xframe.bitstream = buf + pos; | ||
1123 : | xframe.length = buf_size - pos; | ||
1124 : | xframe.output.plane[0] = rgb_out; | ||
1125 : | xframe.output.stride[0] = width; | ||
1126 : | xframe.output.csp = XVID_CSP_BGR; | ||
1127 : | xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); | ||
1128 : | nb++; | ||
1129 : | pos += xframe.length; | ||
1130 : | if (with_chksum) { | ||
1131 : | int k = width*height; | ||
1132 : | uint32_t *ptr = (uint32_t *)rgb_out; | ||
1133 : | while(k-->0) chksum += *ptr++; | ||
1134 : | } | ||
1135 : | if (pos==buf_size) | ||
1136 : | break; | ||
1137 : | if (xerr==XVID_ERR_FAIL) { | ||
1138 : | printf("decoding failed for frame #%d (err=%d)!\n", nb, xerr); | ||
1139 : | break; | ||
1140 : | } | ||
1141 : | } | ||
1142 : | t += gettime_usec(); | ||
1143 : | if (t>0.) | ||
1144 : | printf( "%d frames decoded in %.3f s -> %.1f FPS\n", nb, t*1.e-6f, (float)(nb*1.e6f/t) ); | ||
1145 : | if (with_chksum) | ||
1146 : | printf("checksum: 0x%.8x\n", chksum); | ||
1147 : | |||
1148 : | End: | ||
1149 : | if (rgb_out!=0) free(rgb_out); | ||
1150 : | if (buf!=0) free(buf); | ||
1151 : | if (dechandle!=0) { | ||
1152 : | xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); | ||
1153 : | if (xerr==XVID_ERR_FAIL) | ||
1154 : | printf("destroy-decoder failed (err=%d)!\n", xerr); | ||
1155 : | } | ||
1156 : | if (f!=0) fclose(f); | ||
1157 : | Isibaar | 1.1 | } |
1158 : | |||
1159 : | /********************************************************************* | ||
1160 : | * non-regression tests | ||
1161 : | *********************************************************************/ | ||
1162 : | |||
1163 : | void test_bugs1() | ||
1164 : | { | ||
1165 : | edgomez | 1.9.2.3 | CPU *cpu; |
1166 : | Isibaar | 1.1 | |
1167 : | edgomez | 1.9.2.3 | printf( "\n ===== (de)quant4_intra saturation bug? =====\n" ); |
1168 : | Isibaar | 1.1 | |
1169 : | edgomez | 1.9.2.5 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
1170 : | edgomez | 1.9.2.3 | { |
1171 : | int i; | ||
1172 : | int16_t Src[8*8], Dst[8*8]; | ||
1173 : | |||
1174 : | if (!init_cpu(cpu)) | ||
1175 : | continue; | ||
1176 : | |||
1177 : | for(i=0; i<64; ++i) Src[i] = i-32; | ||
1178 : | set_intra_matrix( get_default_intra_matrix() ); | ||
1179 : | edgomez | 1.9.2.5 | dequant_mpeg_intra(Dst, Src, 31, 5); |
1180 : | printf( "dequant_mpeg_intra with CPU=%s: ", cpu->name); | ||
1181 : | edgomez | 1.9.2.3 | printf( " Out[]= " ); |
1182 : | for(i=0; i<64; ++i) printf( "[%d]", Dst[i]); | ||
1183 : | printf( "\n" ); | ||
1184 : | } | ||
1185 : | |||
1186 : | printf( "\n ===== (de)quant4_inter saturation bug? =====\n" ); | ||
1187 : | |||
1188 : | edgomez | 1.9.2.5 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
1189 : | edgomez | 1.9.2.3 | { |
1190 : | int i; | ||
1191 : | int16_t Src[8*8], Dst[8*8]; | ||
1192 : | |||
1193 : | if (!init_cpu(cpu)) | ||
1194 : | continue; | ||
1195 : | |||
1196 : | for(i=0; i<64; ++i) Src[i] = i-32; | ||
1197 : | set_inter_matrix( get_default_inter_matrix() ); | ||
1198 : | edgomez | 1.9.2.5 | dequant_mpeg_inter(Dst, Src, 31); |
1199 : | printf( "dequant_mpeg_inter with CPU=%s: ", cpu->name); | ||
1200 : | edgomez | 1.9.2.3 | printf( " Out[]= " ); |
1201 : | for(i=0; i<64; ++i) printf( "[%d]", Dst[i]); | ||
1202 : | printf( "\n" ); | ||
1203 : | } | ||
1204 : | Isibaar | 1.1 | } |
1205 : | |||
1206 : | void test_dct_precision_diffs() | ||
1207 : | { | ||
1208 : | edgomez | 1.9.2.3 | CPU *cpu; |
1209 : | short Blk[8*8], Blk0[8*8]; | ||
1210 : | Isibaar | 1.1 | |
1211 : | edgomez | 1.9.2.3 | printf( "\n ===== fdct/idct precision diffs =====\n" ); |
1212 : | Isibaar | 1.1 | |
1213 : | edgomez | 1.9.2.5 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
1214 : | edgomez | 1.9.2.3 | { |
1215 : | int i; | ||
1216 : | |||
1217 : | if (!init_cpu(cpu)) | ||
1218 : | continue; | ||
1219 : | |||
1220 : | for(i=0; i<8*8; ++i) { | ||
1221 : | Blk0[i] = (i*7-i*i) & 0x7f; | ||
1222 : | Blk[i] = Blk0[i]; | ||
1223 : | } | ||
1224 : | |||
1225 : | fdct(Blk); | ||
1226 : | idct(Blk); | ||
1227 : | printf( " fdct+idct diffs with CPU=%s: \n", cpu->name ); | ||
1228 : | for(i=0; i<8; ++i) { | ||
1229 : | int j; | ||
1230 : | for(j=0; j<8; ++j) printf( " %d ", Blk[i*8+j]-Blk0[i*8+j]); | ||
1231 : | printf("\n"); | ||
1232 : | } | ||
1233 : | printf("\n"); | ||
1234 : | } | ||
1235 : | Isibaar | 1.1 | } |
1236 : | |||
1237 : | Isibaar | 1.3 | void test_quant_bug() |
1238 : | { | ||
1239 : | edgomez | 1.9.2.3 | const int max_Q = 31; |
1240 : | int i, n, qm, q; | ||
1241 : | CPU *cpu; | ||
1242 : | int16_t Src[8*8], Dst[8*8]; | ||
1243 : | uint8_t Quant[8*8]; | ||
1244 : | CPU cpu_bug_list[] = { { "PLAINC", 0 }, { "MMX ", XVID_CPU_MMX }, {0,0} }; | ||
1245 : | uint16_t Crcs_Inter[2][32]; | ||
1246 : | uint16_t Crcs_Intra[2][32]; | ||
1247 : | printf( "\n ===== test MPEG4-quantize bug =====\n" ); | ||
1248 : | Isibaar | 1.3 | |
1249 : | edgomez | 1.9.2.3 | for(i=0; i<64; ++i) Src[i] = 2048*(i-32)/32; |
1250 : | Isibaar | 1.3 | |
1251 : | #if 1 | ||
1252 : | edgomez | 1.9.2.3 | for(qm=1; qm<=255; ++qm) |
1253 : | { | ||
1254 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
1255 : | set_inter_matrix( Quant ); | ||
1256 : | |||
1257 : | for(n=0, cpu = cpu_bug_list; cpu->name!=0; ++cpu, ++n) | ||
1258 : | { | ||
1259 : | uint16_t s; | ||
1260 : | |||
1261 : | if (!init_cpu(cpu)) | ||
1262 : | continue; | ||
1263 : | |||
1264 : | for(q=1; q<=max_Q; ++q) { | ||
1265 : | emms(); | ||
1266 : | edgomez | 1.9.2.5 | quant_mpeg_inter( Dst, Src, q ); |
1267 : | edgomez | 1.9.2.3 | emms(); |
1268 : | for(s=0, i=0; i<64; ++i) s+=((uint16_t)Dst[i])^i; | ||
1269 : | Crcs_Inter[n][q] = s; | ||
1270 : | } | ||
1271 : | } | ||
1272 : | |||
1273 : | for(q=1; q<=max_Q; ++q) | ||
1274 : | for(i=0; i<n-1; ++i) | ||
1275 : | if (Crcs_Inter[i][q]!=Crcs_Inter[i+1][q]) | ||
1276 : | printf( "Discrepancy Inter: qm=%d, q=%d -> %d/%d !\n", | ||
1277 : | qm, q, Crcs_Inter[i][q], Crcs_Inter[i+1][q]); | ||
1278 : | } | ||
1279 : | Isibaar | 1.3 | #endif |
1280 : | |||
1281 : | #if 1 | ||
1282 : | edgomez | 1.9.2.3 | for(qm=1; qm<=255; ++qm) |
1283 : | { | ||
1284 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
1285 : | set_intra_matrix( Quant ); | ||
1286 : | |||
1287 : | for(n=0, cpu = cpu_bug_list; cpu->name!=0; ++cpu, ++n) | ||
1288 : | { | ||
1289 : | uint16_t s; | ||
1290 : | |||
1291 : | if (!init_cpu(cpu)) | ||
1292 : | continue; | ||
1293 : | |||
1294 : | for(q=1; q<=max_Q; ++q) { | ||
1295 : | emms(); | ||
1296 : | edgomez | 1.9.2.5 | quant_mpeg_intra( Dst, Src, q, q); |
1297 : | edgomez | 1.9.2.3 | emms(); |
1298 : | for(s=0, i=0; i<64; ++i) s+=((uint16_t)Dst[i])^i; | ||
1299 : | Crcs_Intra[n][q] = s; | ||
1300 : | } | ||
1301 : | } | ||
1302 : | |||
1303 : | for(q=1; q<=max_Q; ++q) | ||
1304 : | for(i=0; i<n-1; ++i) | ||
1305 : | if (Crcs_Intra[i][q]!=Crcs_Intra[i+1][q]) | ||
1306 : | printf( "Discrepancy Intra: qm=%d, q=%d -> %d/%d!\n", | ||
1307 : | qm, q, Crcs_Inter[i][q], Crcs_Inter[i+1][q]); | ||
1308 : | } | ||
1309 : | Isibaar | 1.3 | #endif |
1310 : | } | ||
1311 : | Isibaar | 1.1 | |
1312 : | /********************************************************************* | ||
1313 : | * main | ||
1314 : | *********************************************************************/ | ||
1315 : | |||
1316 : | int main(int argc, char *argv[]) | ||
1317 : | { | ||
1318 : | edgomez | 1.9.2.3 | int what = 0; |
1319 : | if (argc>1) what = atoi(argv[1]); | ||
1320 : | if (what==0 || what==1) test_dct(); | ||
1321 : | if (what==0 || what==2) test_mb(); | ||
1322 : | if (what==0 || what==3) test_sad(); | ||
1323 : | if (what==0 || what==4) test_transfer(); | ||
1324 : | if (what==0 || what==5) test_quant(); | ||
1325 : | if (what==0 || what==6) test_cbp(); | ||
1326 : | Isibaar | 1.1 | |
1327 : | edgomez | 1.9.2.3 | if (what==7) { |
1328 : | test_IEEE1180_compliance(-256, 255, 1); | ||
1329 : | test_IEEE1180_compliance(-256, 255,-1); | ||
1330 : | test_IEEE1180_compliance( -5, 5, 1); | ||
1331 : | test_IEEE1180_compliance( -5, 5,-1); | ||
1332 : | test_IEEE1180_compliance(-300, 300, 1); | ||
1333 : | test_IEEE1180_compliance(-300, 300,-1); | ||
1334 : | } | ||
1335 : | if (what==8) test_dct_saturation(-256, 255); | ||
1336 : | Isibaar | 1.3 | |
1337 : | edgomez | 1.9.2.3 | if (what==9) { |
1338 : | int width, height; | ||
1339 : | if (argc<5) { | ||
1340 : | printf("usage: %s %d [bitstream] [width] [height]\n", argv[0], what); | ||
1341 : | return 1; | ||
1342 : | } | ||
1343 : | width = atoi(argv[3]); | ||
1344 : | height = atoi(argv[4]); | ||
1345 : | test_dec(argv[2], width, height, (argc>5)); | ||
1346 : | } | ||
1347 : | |||
1348 : | if (what==-1) { | ||
1349 : | test_dct_precision_diffs(); | ||
1350 : | test_bugs1(); | ||
1351 : | } | ||
1352 : | if (what==-2) | ||
1353 : | test_quant_bug(); | ||
1354 : | Isibaar | 1.3 | |
1355 : | edgomez | 1.9.2.3 | return 0; |
1356 : | Isibaar | 1.1 | } |
1357 : | |||
1358 : | /********************************************************************* | ||
1359 : | edgomez | 1.9.2.8 | * 'Reference' output (except for timing) on an Athlon XP 2200+ |
1360 : | Isibaar | 1.1 | *********************************************************************/ |
1361 : | Isibaar | 1.3 | |
1362 : | edgomez | 1.9.2.8 | /* as of 07/01/2002, there's a problem with MMX mpeg4-quantization */ |
1363 : | Isibaar | 1.1 | /* |
1364 : | |||
1365 : | edgomez | 1.9.2.8 | ===== test fdct/idct ===== |
1366 : | PLAINC - 1.446 usec PSNR=13.291 MSE=3.000 | ||
1367 : | MMX - -0.260 usec PSNR=9.611 MSE=7.000 | ||
1368 : | MMXEXT - -0.293 usec PSNR=9.611 MSE=7.000 | ||
1369 : | 3DNOW - 1.535 usec PSNR=13.291 MSE=3.000 | ||
1370 : | 3DNOWE - 0.376 usec PSNR=13.291 MSE=3.000 | ||
1371 : | |||
1372 : | === test block motion === | ||
1373 : | PLAINC - interp- h-round0 0.126 usec iCrc=8107 | ||
1374 : | PLAINC - round1 0.136 usec iCrc=8100 | ||
1375 : | PLAINC - interp- v-round0 0.121 usec iCrc=8108 | ||
1376 : | PLAINC - round1 0.127 usec iCrc=8105 | ||
1377 : | PLAINC - interp-hv-round0 0.192 usec iCrc=8112 | ||
1378 : | PLAINC - round1 0.213 usec iCrc=8103 | ||
1379 : | --- | ||
1380 : | MMX - interp- h-round0 0.048 usec iCrc=8107 | ||
1381 : | MMX - round1 0.048 usec iCrc=8100 | ||
1382 : | MMX - interp- v-round0 0.046 usec iCrc=8108 | ||
1383 : | MMX - round1 0.047 usec iCrc=8105 | ||
1384 : | MMX - interp-hv-round0 0.074 usec iCrc=8112 | ||
1385 : | MMX - round1 0.074 usec iCrc=8103 | ||
1386 : | --- | ||
1387 : | MMXEXT - interp- h-round0 0.019 usec iCrc=8107 | ||
1388 : | MMXEXT - round1 0.025 usec iCrc=8100 | ||
1389 : | MMXEXT - interp- v-round0 0.015 usec iCrc=8108 | ||
1390 : | MMXEXT - round1 0.024 usec iCrc=8105 | ||
1391 : | MMXEXT - interp-hv-round0 0.039 usec iCrc=8112 | ||
1392 : | MMXEXT - round1 0.037 usec iCrc=8103 | ||
1393 : | --- | ||
1394 : | 3DNOW - interp- h-round0 0.019 usec iCrc=8107 | ||
1395 : | 3DNOW - round1 0.023 usec iCrc=8100 | ||
1396 : | 3DNOW - interp- v-round0 0.015 usec iCrc=8108 | ||
1397 : | 3DNOW - round1 0.024 usec iCrc=8105 | ||
1398 : | 3DNOW - interp-hv-round0 0.037 usec iCrc=8112 | ||
1399 : | 3DNOW - round1 0.038 usec iCrc=8103 | ||
1400 : | --- | ||
1401 : | 3DNOWE - interp- h-round0 0.022 usec iCrc=8107 | ||
1402 : | 3DNOWE - round1 0.023 usec iCrc=8100 | ||
1403 : | 3DNOWE - interp- v-round0 0.016 usec iCrc=8108 | ||
1404 : | 3DNOWE - round1 0.021 usec iCrc=8105 | ||
1405 : | 3DNOWE - interp-hv-round0 0.036 usec iCrc=8112 | ||
1406 : | 3DNOWE - round1 0.036 usec iCrc=8103 | ||
1407 : | --- | ||
1408 : | |||
1409 : | ====== test SAD ====== | ||
1410 : | PLAINC - sad8 0.165 usec sad=3776 | ||
1411 : | PLAINC - sad16 0.587 usec sad=27214 | ||
1412 : | PLAINC - sad16bi 1.290 usec sad=26274 | ||
1413 : | PLAINC - dev16 1.535 usec sad=3344 | ||
1414 : | --- | ||
1415 : | MMX - sad8 0.036 usec sad=3776 | ||
1416 : | MMX - sad16 0.113 usec sad=27214 | ||
1417 : | MMX - sad16bi 0.250 usec sad=26274 | ||
1418 : | MMX - dev16 0.187 usec sad=3344 | ||
1419 : | --- | ||
1420 : | MMXEXT - sad8 0.015 usec sad=3776 | ||
1421 : | MMXEXT - sad16 0.046 usec sad=27214 | ||
1422 : | MMXEXT - sad16bi 0.059 usec sad=26274 | ||
1423 : | MMXEXT - dev16 0.088 usec sad=3344 | ||
1424 : | --- | ||
1425 : | 3DNOW - sad8 0.165 usec sad=3776 | ||
1426 : | 3DNOW - sad16 0.589 usec sad=27214 | ||
1427 : | 3DNOW - sad16bi 0.119 usec sad=26274 | ||
1428 : | 3DNOW - dev16 1.541 usec sad=3344 | ||
1429 : | --- | ||
1430 : | 3DNOWE - sad8 0.018 usec sad=3776 | ||
1431 : | 3DNOWE - sad16 0.039 usec sad=27214 | ||
1432 : | 3DNOWE - sad16bi 0.051 usec sad=26274 | ||
1433 : | 3DNOWE - dev16 0.070 usec sad=3344 | ||
1434 : | --- | ||
1435 : | |||
1436 : | === test transfer === | ||
1437 : | PLAINC - 8to16 0.207 usec crc=28288 | ||
1438 : | PLAINC - 16to8 0.357 usec crc=28288 | ||
1439 : | PLAINC - 8to8 0.154 usec crc=20352 | ||
1440 : | PLAINC - 16to8add 0.391 usec crc=25536 | ||
1441 : | PLAINC - 8to16sub 0.562 usec crc1=28064 crc2=16256 | ||
1442 : | PLAINC - 8to16sub2 0.519 usec crc=22368 | ||
1443 : | --- | ||
1444 : | MMX - 8to16 0.048 usec crc=28288 | ||
1445 : | MMX - 16to8 0.205 usec crc=28288 | ||
1446 : | MMX - 8to8 -0.158 usec crc=20352 | ||
1447 : | MMX - 16to8add 0.015 usec crc=25536 | ||
1448 : | MMX - 8to16sub 0.203 usec crc1=28064 crc2=16256 | ||
1449 : | MMX - 8to16sub2 0.087 usec crc=22368 | ||
1450 : | --- | ||
1451 : | MMXEXT - 8to16 0.013 usec crc=28288 | ||
1452 : | MMXEXT - 16to8 0.011 usec crc=28288 | ||
1453 : | MMXEXT - 8to8 -0.023 usec crc=20352 | ||
1454 : | MMXEXT - 16to8add 0.023 usec crc=25536 | ||
1455 : | MMXEXT - 8to16sub 0.072 usec crc1=28064 crc2=16256 | ||
1456 : | MMXEXT - 8to16sub2 0.093 usec crc=22368 | ||
1457 : | --- | ||
1458 : | 3DNOW - 8to16 0.192 usec crc=28288 | ||
1459 : | 3DNOW - 16to8 0.367 usec crc=28288 | ||
1460 : | 3DNOW - 8to8 0.132 usec crc=20352 | ||
1461 : | 3DNOW - 16to8add 0.440 usec crc=25536 | ||
1462 : | 3DNOW - 8to16sub 0.557 usec crc1=28064 crc2=16256 | ||
1463 : | 3DNOW - 8to16sub2 0.691 usec crc=22368 | ||
1464 : | --- | ||
1465 : | 3DNOWE - 8to16 0.004 usec crc=28288 | ||
1466 : | 3DNOWE - 16to8 0.019 usec crc=28288 | ||
1467 : | 3DNOWE - 8to8 -0.294 usec crc=20352 | ||
1468 : | 3DNOWE - 16to8add 0.028 usec crc=25536 | ||
1469 : | 3DNOWE - 8to16sub 0.065 usec crc1=28064 crc2=16256 | ||
1470 : | 3DNOWE - 8to16sub2 0.027 usec crc=22368 | ||
1471 : | --- | ||
1472 : | |||
1473 : | ===== test quant ===== | ||
1474 : | PLAINC - quant_mpeg_intra 67.757 usec crc=29809 | ||
1475 : | PLAINC - quant_mpeg_inter 68.482 usec crc=12574 | ||
1476 : | PLAINC - dequant_mpeg_intra 20.764 usec crc=24052 | ||
1477 : | PLAINC - dequant_mpeg_inter 24.413 usec crc=63847 | ||
1478 : | PLAINC - quant_h263_intra 16.446 usec crc=25662 | ||
1479 : | PLAINC - quant_h263_inter 14.499 usec crc=23972 | ||
1480 : | PLAINC - dequant_h263_intra 16.419 usec crc=49900 | ||
1481 : | PLAINC - dequant_h263_inter 18.679 usec crc=48899 | ||
1482 : | --- | ||
1483 : | MMX - quant_mpeg_intra 8.299 usec crc=3459 | ||
1484 : | Isibaar | 1.3 | *** CRC ERROR! *** |
1485 : | edgomez | 1.9.2.8 | MMX - quant_mpeg_inter 7.078 usec crc=13247 |
1486 : | Isibaar | 1.3 | *** CRC ERROR! *** |
1487 : | edgomez | 1.9.2.8 | MMX - dequant_mpeg_intra 3.455 usec crc=24052 |
1488 : | MMX - dequant_mpeg_inter 4.483 usec crc=63847 | ||
1489 : | MMX - quant_h263_intra 2.597 usec crc=25662 | ||
1490 : | MMX - quant_h263_inter 2.151 usec crc=23972 | ||
1491 : | MMX - dequant_h263_intra 2.684 usec crc=49900 | ||
1492 : | MMX - dequant_h263_inter 2.647 usec crc=48899 | ||
1493 : | --- | ||
1494 : | MMXEXT - quant_mpeg_intra 3.878 usec crc=29809 | ||
1495 : | MMXEXT - quant_mpeg_inter 4.112 usec crc=12574 | ||
1496 : | MMXEXT - dequant_mpeg_intra 3.452 usec crc=24052 | ||
1497 : | MMXEXT - dequant_mpeg_inter 4.473 usec crc=63847 | ||
1498 : | MMXEXT - quant_h263_intra 2.608 usec crc=25662 | ||
1499 : | MMXEXT - quant_h263_inter 2.145 usec crc=23972 | ||
1500 : | MMXEXT - dequant_h263_intra 2.478 usec crc=49900 | ||
1501 : | MMXEXT - dequant_h263_inter 2.450 usec crc=48899 | ||
1502 : | --- | ||
1503 : | 3DNOW - quant_mpeg_intra 66.051 usec crc=29809 | ||
1504 : | 3DNOW - quant_mpeg_inter 73.494 usec crc=12574 | ||
1505 : | 3DNOW - dequant_mpeg_intra 20.374 usec crc=24052 | ||
1506 : | 3DNOW - dequant_mpeg_inter 23.645 usec crc=63847 | ||
1507 : | 3DNOW - quant_h263_intra 16.292 usec crc=25662 | ||
1508 : | 3DNOW - quant_h263_inter 14.322 usec crc=23972 | ||
1509 : | 3DNOW - dequant_h263_intra 16.613 usec crc=49900 | ||
1510 : | 3DNOW - dequant_h263_inter 18.382 usec crc=48899 | ||
1511 : | --- | ||
1512 : | 3DNOWE - quant_mpeg_intra 66.140 usec crc=29809 | ||
1513 : | 3DNOWE - quant_mpeg_inter 68.454 usec crc=12574 | ||
1514 : | 3DNOWE - dequant_mpeg_intra 2.881 usec crc=24052 | ||
1515 : | 3DNOWE - dequant_mpeg_inter 4.155 usec crc=63847 | ||
1516 : | 3DNOWE - quant_h263_intra 1.451 usec crc=25662 | ||
1517 : | 3DNOWE - quant_h263_inter 1.849 usec crc=23972 | ||
1518 : | 3DNOWE - dequant_h263_intra 2.101 usec crc=49900 | ||
1519 : | 3DNOWE - dequant_h263_inter 2.109 usec crc=48899 | ||
1520 : | --- | ||
1521 : | |||
1522 : | ===== test cbp ===== | ||
1523 : | PLAINC - calc_cbp#1 0.090 usec cbp=0x15 | ||
1524 : | PLAINC - calc_cbp#2 0.086 usec cbp=0x38 | ||
1525 : | PLAINC - calc_cbp#3 0.087 usec cbp=0xf | ||
1526 : | PLAINC - calc_cbp#4 0.114 usec cbp=0x5 | ||
1527 : | --- | ||
1528 : | MMX - calc_cbp#1 0.061 usec cbp=0x15 | ||
1529 : | MMX - calc_cbp#2 0.063 usec cbp=0x38 | ||
1530 : | MMX - calc_cbp#3 0.061 usec cbp=0xf | ||
1531 : | MMX - calc_cbp#4 0.060 usec cbp=0x5 | ||
1532 : | --- | ||
1533 : | MMXEXT - calc_cbp#1 0.062 usec cbp=0x15 | ||
1534 : | MMXEXT - calc_cbp#2 0.060 usec cbp=0x38 | ||
1535 : | MMXEXT - calc_cbp#3 0.062 usec cbp=0xf | ||
1536 : | MMXEXT - calc_cbp#4 0.061 usec cbp=0x5 | ||
1537 : | --- | ||
1538 : | 3DNOW - calc_cbp#1 0.089 usec cbp=0x15 | ||
1539 : | 3DNOW - calc_cbp#2 0.087 usec cbp=0x38 | ||
1540 : | 3DNOW - calc_cbp#3 0.087 usec cbp=0xf | ||
1541 : | 3DNOW - calc_cbp#4 0.116 usec cbp=0x5 | ||
1542 : | --- | ||
1543 : | 3DNOWE - calc_cbp#1 0.050 usec cbp=0x15 | ||
1544 : | 3DNOWE - calc_cbp#2 0.051 usec cbp=0x38 | ||
1545 : | 3DNOWE - calc_cbp#3 0.050 usec cbp=0xf | ||
1546 : | 3DNOWE - calc_cbp#4 0.049 usec cbp=0x5 | ||
1547 : | --- | ||
1548 : | Isibaar | 1.3 | |
1549 : | Isibaar | 1.1 | */ |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |