Parent Directory
| 
Revision Log
Revision 1122 - (view) (download)
| 1 : | edgomez | 1054 | /***************************************************************************** |
| 2 : | Isibaar | 225 | * |
| 3 : | edgomez | 1054 | * XVID MPEG-4 VIDEO CODEC |
| 4 : | * - Unit tests and benches - | ||
| 5 : | Isibaar | 225 | * |
| 6 : | edgomez | 1054 | * Copyright(C) 2002 Pascal Massimino <skal@planet-d.net> |
| 7 : | Isibaar | 225 | * |
| 8 : | edgomez | 1054 | * 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 | 225 | * |
| 13 : | edgomez | 1054 | * 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 | 225 | * |
| 18 : | edgomez | 1054 | * 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 | 1122 | * $Id: xvid_bench.c,v 1.9.2.4 2003-08-13 11:43:45 edgomez Exp $ |
| 23 : | edgomez | 1054 | * |
| 24 : | ****************************************************************************/ | ||
| 25 : | Isibaar | 225 | |
| 26 : | edgomez | 1054 | /***************************************************************************** |
| 27 : | Isibaar | 225 | * |
| 28 : | * 'Reference' output is at the end of file. | ||
| 29 : | * Don't take the checksums and crc too seriouly, they aren't | ||
| 30 : | Isibaar | 262 | * bullet-proof (should plug some .md5 here)... |
| 31 : | Isibaar | 225 | * |
| 32 : | * compiles with something like: | ||
| 33 : | * gcc -o xvid_bench xvid_bench.c -I../src/ -lxvidcore -lm | ||
| 34 : | * | ||
| 35 : | edgomez | 1054 | ****************************************************************************/ |
| 36 : | Isibaar | 225 | |
| 37 : | #include <stdio.h> | ||
| 38 : | #include <stdlib.h> | ||
| 39 : | edgomez | 1053 | #include <string.h> /* for memset */ |
| 40 : | Isibaar | 225 | #include <assert.h> |
| 41 : | |||
| 42 : | suxen_drol | 860 | #ifndef WIN32 |
| 43 : | edgomez | 1053 | #include <sys/time.h> /* for gettimeofday */ |
| 44 : | suxen_drol | 860 | #else |
| 45 : | #include <time.h> | ||
| 46 : | #endif | ||
| 47 : | |||
| 48 : | |||
| 49 : | Isibaar | 225 | #include "xvid.h" |
| 50 : | |||
| 51 : | edgomez | 851 | // inner guts |
| 52 : | Isibaar | 225 | #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 : | #include "quant/quant_h263.h" | ||
| 58 : | #include "quant/quant_mpeg4.h" | ||
| 59 : | #include "motion/sad.h" | ||
| 60 : | #include "utils/emms.h" | ||
| 61 : | #include "utils/timer.h" | ||
| 62 : | #include "quant/quant_matrix.c" | ||
| 63 : | #include "bitstream/cbp.h" | ||
| 64 : | |||
| 65 : | Isibaar | 262 | #include <math.h> |
| 66 : | suxen_drol | 860 | |
| 67 : | #ifndef M_PI | ||
| 68 : | #define M_PI 3.14159265358979323846 | ||
| 69 : | #endif | ||
| 70 : | |||
| 71 : | edgomez | 1053 | const int speed_ref = 100; /* on slow machines, decrease this value */ |
| 72 : | Isibaar | 225 | |
| 73 : | /********************************************************************* | ||
| 74 : | * misc | ||
| 75 : | *********************************************************************/ | ||
| 76 : | |||
| 77 : | edgomez | 1060 | /* returns time in micro-s*/ |
| 78 : | Isibaar | 225 | double gettime_usec() |
| 79 : | { | ||
| 80 : | suxen_drol | 860 | #ifndef WIN32 |
| 81 : | edgomez | 1060 | struct timeval tv; |
| 82 : | gettimeofday(&tv, 0); | ||
| 83 : | return tv.tv_sec*1.0e6 + tv.tv_usec; | ||
| 84 : | suxen_drol | 860 | #else |
| 85 : | clock_t clk; | ||
| 86 : | clk = clock(); | ||
| 87 : | return clk * 1000000 / CLOCKS_PER_SEC; | ||
| 88 : | #endif | ||
| 89 : | Isibaar | 225 | } |
| 90 : | |||
| 91 : | edgomez | 1060 | /* returns squared deviates (mean(v*v)-mean(v)^2) of a 8x8 block */ |
| 92 : | Isibaar | 225 | double sqr_dev(uint8_t v[8*8]) |
| 93 : | { | ||
| 94 : | edgomez | 1060 | double sum=0.; |
| 95 : | double sum2=0.; | ||
| 96 : | int n; | ||
| 97 : | for (n=0;n<8*8;n++) | ||
| 98 : | { | ||
| 99 : | sum += v[n]; | ||
| 100 : | sum2 += v[n]*v[n]; | ||
| 101 : | } | ||
| 102 : | sum2 /= n; | ||
| 103 : | sum /= n; | ||
| 104 : | return sum2-sum*sum; | ||
| 105 : | Isibaar | 225 | } |
| 106 : | |||
| 107 : | /********************************************************************* | ||
| 108 : | * cpu init | ||
| 109 : | *********************************************************************/ | ||
| 110 : | |||
| 111 : | typedef struct { | ||
| 112 : | edgomez | 1060 | const char *name; |
| 113 : | unsigned int cpu; | ||
| 114 : | Isibaar | 225 | } CPU; |
| 115 : | |||
| 116 : | CPU cpu_list[] = | ||
| 117 : | { { "PLAINC", 0 } | ||
| 118 : | edgomez | 1122 | #ifdef ARCH_IS_IA32 |
| 119 : | edgomez | 1060 | , { "MMX ", XVID_CPU_MMX } |
| 120 : | , { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } | ||
| 121 : | , { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } | ||
| 122 : | , { "3DNOW ", XVID_CPU_3DNOW } | ||
| 123 : | , { "3DNOWE", XVID_CPU_3DNOWEXT } | ||
| 124 : | edgomez | 1122 | #endif |
| 125 : | edgomez | 1060 | //, { "IA64 ", XVID_CPU_IA64 } |
| 126 : | edgomez | 851 | //, { "TSC ", XVID_CPU_TSC } |
| 127 : | edgomez | 1060 | , { 0, 0 } }; |
| 128 : | Isibaar | 225 | |
| 129 : | edgomez | 1060 | CPU cpu_short_list[] = |
| 130 : | Isibaar | 225 | { { "PLAINC", 0 } |
| 131 : | edgomez | 1122 | #ifdef ARCH_IS_IA32 |
| 132 : | edgomez | 1060 | , { "MMX ", XVID_CPU_MMX } |
| 133 : | edgomez | 851 | //, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
| 134 : | edgomez | 1122 | #endif |
| 135 : | edgomez | 1060 | //, { "IA64 ", XVID_CPU_IA64 } |
| 136 : | , { 0, 0 } }; | ||
| 137 : | Isibaar | 225 | |
| 138 : | edgomez | 1060 | CPU cpu_short_list2[] = |
| 139 : | Isibaar | 225 | { { "PLAINC", 0 } |
| 140 : | edgomez | 1122 | #ifdef ARCH_IS_IA32 |
| 141 : | edgomez | 1060 | , { "MMX ", XVID_CPU_MMX } |
| 142 : | , { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } | ||
| 143 : | edgomez | 1122 | #endif |
| 144 : | edgomez | 1060 | , { 0, 0 } }; |
| 145 : | Isibaar | 225 | |
| 146 : | |||
| 147 : | int init_cpu(CPU *cpu) | ||
| 148 : | { | ||
| 149 : | edgomez | 1060 | int xerr, cpu_type; |
| 150 : | xvid_gbl_init_t xinit; | ||
| 151 : | Isibaar | 225 | |
| 152 : | edgomez | 1122 | #ifdef ARCH_IS_IA32 |
| 153 : | edgomez | 1060 | cpu_type = check_cpu_features() & cpu->cpu; |
| 154 : | edgomez | 1122 | #else |
| 155 : | cpu_type = XVID_CPU_ASM; | ||
| 156 : | #endif | ||
| 157 : | edgomez | 1060 | memset(&xinit, 0, sizeof(xinit)); |
| 158 : | xinit.cpu_flags = cpu_type | XVID_CPU_FORCE; | ||
| 159 : | xinit.version = XVID_VERSION; | ||
| 160 : | xerr = xvid_global(NULL, 0, &xinit, NULL); | ||
| 161 : | if (cpu->cpu>0 && (cpu_type==0 || xerr==XVID_ERR_FAIL)) { | ||
| 162 : | printf( "%s - skipped...\n", cpu->name ); | ||
| 163 : | return 0; | ||
| 164 : | } | ||
| 165 : | return 1; | ||
| 166 : | Isibaar | 225 | } |
| 167 : | |||
| 168 : | /********************************************************************* | ||
| 169 : | * test DCT | ||
| 170 : | *********************************************************************/ | ||
| 171 : | |||
| 172 : | #define ABS(X) ((X)<0 ? -(X) : (X)) | ||
| 173 : | |||
| 174 : | void test_dct() | ||
| 175 : | { | ||
| 176 : | edgomez | 1060 | const int nb_tests = 300*speed_ref; |
| 177 : | int tst; | ||
| 178 : | CPU *cpu; | ||
| 179 : | int i; | ||
| 180 : | short iDst0[8*8], iDst[8*8], fDst[8*8]; | ||
| 181 : | double overhead; | ||
| 182 : | Isibaar | 225 | |
| 183 : | edgomez | 1060 | printf( "\n ===== test fdct/idct =====\n" ); |
| 184 : | Isibaar | 225 | |
| 185 : | edgomez | 1060 | for(i=0; i<8*8; ++i) iDst0[i] = (i*7-i*i) & 0x7f; |
| 186 : | overhead = gettime_usec(); | ||
| 187 : | for(tst=0; tst<nb_tests; ++tst) | ||
| 188 : | { | ||
| 189 : | for(i=0; i<8*8; ++i) fDst[i] = iDst0[i]; | ||
| 190 : | for(i=0; i<8*8; ++i) iDst[i] = fDst[i]; | ||
| 191 : | } | ||
| 192 : | overhead = gettime_usec() - overhead; | ||
| 193 : | Isibaar | 225 | |
| 194 : | edgomez | 1060 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 195 : | { | ||
| 196 : | double t, PSNR, MSE; | ||
| 197 : | Isibaar | 225 | |
| 198 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 199 : | continue; | ||
| 200 : | Isibaar | 225 | |
| 201 : | edgomez | 1060 | t = gettime_usec(); |
| 202 : | emms(); | ||
| 203 : | for(tst=0; tst<nb_tests; ++tst) | ||
| 204 : | { | ||
| 205 : | for(i=0; i<8*8; ++i) fDst[i] = iDst0[i]; | ||
| 206 : | fdct(fDst); | ||
| 207 : | for(i=0; i<8*8; ++i) iDst[i] = fDst[i]; | ||
| 208 : | idct(iDst); | ||
| 209 : | } | ||
| 210 : | emms(); | ||
| 211 : | t = (gettime_usec() - t - overhead) / nb_tests; | ||
| 212 : | MSE = 0.; | ||
| 213 : | for(i=0; i<8*8; ++i) { | ||
| 214 : | double delta = 1.0*(iDst[i] - iDst0[i]); | ||
| 215 : | MSE += delta*delta; | ||
| 216 : | } | ||
| 217 : | PSNR = (MSE==0.) ? 1.e6 : -4.3429448*log( MSE/64. ); | ||
| 218 : | printf( "%s - %.3f usec PSNR=%.3f MSE=%.3f\n", | ||
| 219 : | cpu->name, t, PSNR, MSE ); | ||
| 220 : | if (ABS(MSE)>=64) printf( "*** CRC ERROR! ***\n" ); | ||
| 221 : | } | ||
| 222 : | Isibaar | 225 | } |
| 223 : | |||
| 224 : | /********************************************************************* | ||
| 225 : | * test SAD | ||
| 226 : | *********************************************************************/ | ||
| 227 : | |||
| 228 : | void test_sad() | ||
| 229 : | { | ||
| 230 : | edgomez | 1060 | const int nb_tests = 2000*speed_ref; |
| 231 : | int tst; | ||
| 232 : | CPU *cpu; | ||
| 233 : | int i; | ||
| 234 : | uint8_t Cur[16*16], Ref1[16*16], Ref2[16*16]; | ||
| 235 : | Isibaar | 225 | |
| 236 : | edgomez | 1060 | printf( "\n ====== test SAD ======\n" ); |
| 237 : | for(i=0; i<16*16;++i) { | ||
| 238 : | Cur[i] = (i/5) ^ 0x05; | ||
| 239 : | Ref1[i] = (i + 0x0b) & 0xff; | ||
| 240 : | Ref2[i] = i ^ 0x76; | ||
| 241 : | } | ||
| 242 : | Isibaar | 225 | |
| 243 : | edgomez | 1060 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 244 : | { | ||
| 245 : | double t; | ||
| 246 : | uint32_t s; | ||
| 247 : | if (!init_cpu(cpu)) | ||
| 248 : | continue; | ||
| 249 : | Isibaar | 225 | |
| 250 : | edgomez | 1060 | t = gettime_usec(); |
| 251 : | emms(); | ||
| 252 : | for(tst=0; tst<nb_tests; ++tst) s = sad8(Cur, Ref1, 16); | ||
| 253 : | emms(); | ||
| 254 : | t = (gettime_usec() - t) / nb_tests; | ||
| 255 : | printf( "%s - sad8 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
| 256 : | if (s!=3776) printf( "*** CRC ERROR! ***\n" ); | ||
| 257 : | Isibaar | 225 | |
| 258 : | edgomez | 1060 | t = gettime_usec(); |
| 259 : | emms(); | ||
| 260 : | for(tst=0; tst<nb_tests; ++tst) s = sad16(Cur, Ref1, 16, -1); | ||
| 261 : | emms(); | ||
| 262 : | t = (gettime_usec() - t) / nb_tests; | ||
| 263 : | printf( "%s - sad16 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
| 264 : | if (s!=27214) printf( "*** CRC ERROR! ***\n" ); | ||
| 265 : | Isibaar | 225 | |
| 266 : | edgomez | 1060 | t = gettime_usec(); |
| 267 : | emms(); | ||
| 268 : | for(tst=0; tst<nb_tests; ++tst) s = sad16bi(Cur, Ref1, Ref2, 16); | ||
| 269 : | emms(); | ||
| 270 : | t = (gettime_usec() - t) / nb_tests; | ||
| 271 : | printf( "%s - sad16bi %.3f usec sad=%d\n", cpu->name, t, s ); | ||
| 272 : | if (s!=26274) printf( "*** CRC ERROR! ***\n" ); | ||
| 273 : | Isibaar | 225 | |
| 274 : | edgomez | 1060 | t = gettime_usec(); |
| 275 : | emms(); | ||
| 276 : | for(tst=0; tst<nb_tests; ++tst) s = dev16(Cur, 16); | ||
| 277 : | emms(); | ||
| 278 : | t = (gettime_usec() - t) / nb_tests; | ||
| 279 : | printf( "%s - dev16 %.3f usec sad=%d\n", cpu->name, t, s ); | ||
| 280 : | if (s!=3344) printf( "*** CRC ERROR! ***\n" ); | ||
| 281 : | Isibaar | 225 | |
| 282 : | edgomez | 1060 | printf( " --- \n" ); |
| 283 : | } | ||
| 284 : | Isibaar | 225 | } |
| 285 : | |||
| 286 : | /********************************************************************* | ||
| 287 : | * test interpolation | ||
| 288 : | *********************************************************************/ | ||
| 289 : | |||
| 290 : | #define ENTER \ | ||
| 291 : | edgomez | 1060 | for(i=0; i<16*8; ++i) Dst[i] = 0; \ |
| 292 : | t = gettime_usec(); \ | ||
| 293 : | emms(); | ||
| 294 : | Isibaar | 225 | |
| 295 : | #define LEAVE \ | ||
| 296 : | edgomez | 1060 | emms(); \ |
| 297 : | t = (gettime_usec() - t) / nb_tests; \ | ||
| 298 : | iCrc = 0; \ | ||
| 299 : | for(i=0; i<16*8; ++i) { iCrc += Dst[i]^i; } | ||
| 300 : | Isibaar | 225 | |
| 301 : | #define TEST_MB(FUNC, R) \ | ||
| 302 : | edgomez | 1060 | ENTER \ |
| 303 : | for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst, Src0, 16, (R)); \ | ||
| 304 : | LEAVE | ||
| 305 : | Isibaar | 225 | |
| 306 : | #define TEST_MB2(FUNC) \ | ||
| 307 : | edgomez | 1060 | ENTER \ |
| 308 : | for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst, Src0, 16); \ | ||
| 309 : | LEAVE | ||
| 310 : | Isibaar | 225 | |
| 311 : | |||
| 312 : | void test_mb() | ||
| 313 : | { | ||
| 314 : | edgomez | 1060 | const int nb_tests = 2000*speed_ref; |
| 315 : | CPU *cpu; | ||
| 316 : | const uint8_t Src0[16*9] = { | ||
| 317 : | /* try to have every possible combinaison of rounding... */ | ||
| 318 : | 0, 0, 1, 0, 2, 0, 3, 0, 4 ,0,0,0, 0,0,0,0 | ||
| 319 : | , 0, 1, 1, 1, 2, 1, 3, 1, 3 ,0,0,0, 0,0,0,0 | ||
| 320 : | , 0, 2, 1, 2, 2, 2, 3, 2, 2 ,0,0,0, 0,0,0,0 | ||
| 321 : | , 0, 3, 1, 3, 2, 3, 3, 3, 1 ,0,0,0, 0,0,0,0 | ||
| 322 : | , 1, 3, 0, 2, 1, 0, 2, 3, 4 ,0,0,0, 0,0,0,0 | ||
| 323 : | , 2, 2, 1, 2, 0, 1, 3, 5, 3 ,0,0,0, 0,0,0,0 | ||
| 324 : | , 3, 1, 2, 3, 1, 2, 2, 6, 2 ,0,0,0, 0,0,0,0 | ||
| 325 : | , 1, 0, 1, 3, 0, 3, 1, 6, 1 ,0,0,0, 0,0,0,0 | ||
| 326 : | , 4, 3, 2, 1, 2, 3, 4, 0, 3 ,0,0,0, 0,0,0,0 | ||
| 327 : | }; | ||
| 328 : | uint8_t Dst[16*8] = {0}; | ||
| 329 : | Isibaar | 225 | |
| 330 : | edgomez | 1060 | printf( "\n === test block motion ===\n" ); |
| 331 : | Isibaar | 225 | |
| 332 : | edgomez | 1060 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 333 : | { | ||
| 334 : | double t; | ||
| 335 : | int tst, i, iCrc; | ||
| 336 : | Isibaar | 225 | |
| 337 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 338 : | continue; | ||
| 339 : | Isibaar | 225 | |
| 340 : | edgomez | 1060 | TEST_MB(interpolate8x8_halfpel_h, 0); |
| 341 : | printf( "%s - interp- h-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 342 : | if (iCrc!=8107) printf( "*** CRC ERROR! ***\n" ); | ||
| 343 : | Isibaar | 225 | |
| 344 : | edgomez | 1060 | TEST_MB(interpolate8x8_halfpel_h, 1); |
| 345 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 346 : | if (iCrc!=8100) printf( "*** CRC ERROR! ***\n" ); | ||
| 347 : | Isibaar | 225 | |
| 348 : | |||
| 349 : | edgomez | 1060 | TEST_MB(interpolate8x8_halfpel_v, 0); |
| 350 : | printf( "%s - interp- v-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 351 : | if (iCrc!=8108) printf( "*** CRC ERROR! ***\n" ); | ||
| 352 : | Isibaar | 225 | |
| 353 : | edgomez | 1060 | TEST_MB(interpolate8x8_halfpel_v, 1); |
| 354 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 355 : | if (iCrc!=8105) printf( "*** CRC ERROR! ***\n" ); | ||
| 356 : | Isibaar | 225 | |
| 357 : | |||
| 358 : | edgomez | 1060 | TEST_MB(interpolate8x8_halfpel_hv, 0); |
| 359 : | printf( "%s - interp-hv-round0 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 360 : | if (iCrc!=8112) printf( "*** CRC ERROR! ***\n" ); | ||
| 361 : | Isibaar | 225 | |
| 362 : | edgomez | 1060 | TEST_MB(interpolate8x8_halfpel_hv, 1); |
| 363 : | printf( "%s - round1 %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 364 : | if (iCrc!=8103) printf( "*** CRC ERROR! ***\n" ); | ||
| 365 : | Isibaar | 225 | |
| 366 : | Isibaar | 262 | |
| 367 : | edgomez | 1060 | /* this is a new function, as of 06.06.2002 */ |
| 368 : | Isibaar | 262 | #if 0 |
| 369 : | edgomez | 1060 | TEST_MB2(interpolate8x8_avrg); |
| 370 : | printf( "%s - interpolate8x8_c %.3f usec iCrc=%d\n", cpu->name, t, iCrc ); | ||
| 371 : | if (iCrc!=8107) printf( "*** CRC ERROR! ***\n" ); | ||
| 372 : | Isibaar | 262 | #endif |
| 373 : | |||
| 374 : | edgomez | 1060 | printf( " --- \n" ); |
| 375 : | } | ||
| 376 : | Isibaar | 225 | } |
| 377 : | |||
| 378 : | /********************************************************************* | ||
| 379 : | * test transfer | ||
| 380 : | *********************************************************************/ | ||
| 381 : | |||
| 382 : | #define INIT_TRANSFER \ | ||
| 383 : | edgomez | 1060 | for(i=0; i<8*32; ++i) { \ |
| 384 : | Src8[i] = i; Src16[i] = i; \ | ||
| 385 : | Dst8[i] = 0; Dst16[i] = 0; \ | ||
| 386 : | Ref1[i] = i^0x27; \ | ||
| 387 : | Ref2[i] = i^0x51; \ | ||
| 388 : | } | ||
| 389 : | Isibaar | 225 | |
| 390 : | #define TEST_TRANSFER_BEGIN(DST) \ | ||
| 391 : | edgomez | 1060 | INIT_TRANSFER \ |
| 392 : | overhead = -gettime_usec(); \ | ||
| 393 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
| 394 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
| 395 : | } \ | ||
| 396 : | overhead += gettime_usec(); \ | ||
| 397 : | t = gettime_usec(); \ | ||
| 398 : | emms(); \ | ||
| 399 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
| 400 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a; | ||
| 401 : | Isibaar | 225 | |
| 402 : | |||
| 403 : | #define TEST_TRANSFER_END(DST) \ | ||
| 404 : | edgomez | 1060 | } \ |
| 405 : | emms(); \ | ||
| 406 : | t = (gettime_usec()-t -overhead) / nb_tests;\ | ||
| 407 : | s = 0; for(i=0; i<8*32; ++i) { s += (DST)[i]^i; } | ||
| 408 : | Isibaar | 225 | |
| 409 : | #define TEST_TRANSFER(FUNC, DST, SRC) \ | ||
| 410 : | edgomez | 1060 | TEST_TRANSFER_BEGIN(DST); \ |
| 411 : | (FUNC)((DST), (SRC), 32); \ | ||
| 412 : | TEST_TRANSFER_END(DST) | ||
| 413 : | Isibaar | 225 | |
| 414 : | |||
| 415 : | #define TEST_TRANSFER2_BEGIN(DST, SRC) \ | ||
| 416 : | edgomez | 1060 | INIT_TRANSFER \ |
| 417 : | overhead = -gettime_usec(); \ | ||
| 418 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
| 419 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
| 420 : | for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e;\ | ||
| 421 : | } \ | ||
| 422 : | overhead += gettime_usec(); \ | ||
| 423 : | t = gettime_usec(); \ | ||
| 424 : | emms(); \ | ||
| 425 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
| 426 : | for(i=0; i<8*32; ++i) (DST)[i] = i^0x6a;\ | ||
| 427 : | for(i=0; i<8*32; ++i) (SRC)[i] = i^0x3e; | ||
| 428 : | Isibaar | 225 | |
| 429 : | #define TEST_TRANSFER2_END(DST) \ | ||
| 430 : | edgomez | 1060 | } \ |
| 431 : | emms(); \ | ||
| 432 : | t = (gettime_usec()-t -overhead) / nb_tests;\ | ||
| 433 : | s = 0; for(i=0; i<8*32; ++i) { s += (DST)[i]; } | ||
| 434 : | Isibaar | 225 | |
| 435 : | #define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ | ||
| 436 : | edgomez | 1060 | TEST_TRANSFER2_BEGIN(DST,SRC); \ |
| 437 : | (FUNC)((DST), (SRC), (R1), 32); \ | ||
| 438 : | TEST_TRANSFER2_END(DST) | ||
| 439 : | Isibaar | 225 | |
| 440 : | #define TEST_TRANSFER3(FUNC, DST, SRC, R1, R2)\ | ||
| 441 : | edgomez | 1060 | TEST_TRANSFER_BEGIN(DST); \ |
| 442 : | (FUNC)((DST), (SRC), (R1), (R2), 32); \ | ||
| 443 : | TEST_TRANSFER_END(DST) | ||
| 444 : | Isibaar | 225 | |
| 445 : | void test_transfer() | ||
| 446 : | { | ||
| 447 : | edgomez | 1060 | const int nb_tests = 4000*speed_ref; |
| 448 : | int i; | ||
| 449 : | CPU *cpu; | ||
| 450 : | uint8_t Src8[8*32], Dst8[8*32], Ref1[8*32], Ref2[8*32]; | ||
| 451 : | int16_t Src16[8*32], Dst16[8*32]; | ||
| 452 : | Isibaar | 225 | |
| 453 : | edgomez | 1060 | printf( "\n === test transfer ===\n" ); |
| 454 : | Isibaar | 225 | |
| 455 : | edgomez | 1060 | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) |
| 456 : | { | ||
| 457 : | double t, overhead; | ||
| 458 : | int tst, s; | ||
| 459 : | Isibaar | 225 | |
| 460 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 461 : | continue; | ||
| 462 : | Isibaar | 225 | |
| 463 : | edgomez | 1060 | TEST_TRANSFER(transfer_8to16copy, Dst16, Src8); |
| 464 : | printf( "%s - 8to16 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 465 : | if (s!=28288) printf( "*** CRC ERROR! ***\n" ); | ||
| 466 : | Isibaar | 225 | |
| 467 : | edgomez | 1060 | TEST_TRANSFER(transfer_16to8copy, Dst8, Src16); |
| 468 : | printf( "%s - 16to8 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 469 : | if (s!=28288) printf( "*** CRC ERROR! ***\n" ); | ||
| 470 : | Isibaar | 225 | |
| 471 : | edgomez | 1060 | TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
| 472 : | printf( "%s - 8to8 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 473 : | if (s!=20352) printf( "*** CRC ERROR! ***\n" ); | ||
| 474 : | Isibaar | 225 | |
| 475 : | edgomez | 1060 | TEST_TRANSFER(transfer_16to8add, Dst8, Src16); |
| 476 : | printf( "%s - 16to8add %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 477 : | if (s!=25536) printf( "*** CRC ERROR! ***\n" ); | ||
| 478 : | Isibaar | 225 | |
| 479 : | edgomez | 1060 | TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); |
| 480 : | printf( "%s - 8to16sub %.3f usec crc1=%d ", cpu->name, t, s ); | ||
| 481 : | if (s!=28064) printf( "*** CRC ERROR! ***\n" ); | ||
| 482 : | s = 0; for(i=0; i<8*32; ++i) { s += (Src8[i]-Ref1[i])&i; } | ||
| 483 : | printf( "crc2=%d\n", s); | ||
| 484 : | if (s!=16256) printf( "*** CRC ERROR! ***\n" ); | ||
| 485 : | Isibaar | 262 | #if 1 |
| 486 : | edgomez | 1060 | TEST_TRANSFER3(transfer_8to16sub2, Dst16, Src8, Ref1, Ref2); |
| 487 : | printf( "%s - 8to16sub2 %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 488 : | if (s!=20384) printf( "*** CRC ERROR! ***\n" ); | ||
| 489 : | edgomez | 851 | // for(i=0; i<64; ++i) printf( "[%d]", Dst16[i]); |
| 490 : | // printf("\n"); | ||
| 491 : | Isibaar | 262 | #endif |
| 492 : | edgomez | 1060 | printf( " --- \n" ); |
| 493 : | } | ||
| 494 : | Isibaar | 225 | } |
| 495 : | |||
| 496 : | /********************************************************************* | ||
| 497 : | * test quantization | ||
| 498 : | *********************************************************************/ | ||
| 499 : | |||
| 500 : | Isibaar | 262 | #define TEST_QUANT(FUNC, DST, SRC) \ |
| 501 : | edgomez | 1060 | t = gettime_usec(); \ |
| 502 : | for(s=0,qm=1; qm<=255; ++qm) { \ | ||
| 503 : | for(i=0; i<8*8; ++i) Quant[i] = qm; \ | ||
| 504 : | set_inter_matrix( Quant ); \ | ||
| 505 : | emms(); \ | ||
| 506 : | for(q=1; q<=max_Q; ++q) { \ | ||
| 507 : | for(tst=0; tst<nb_tests; ++tst) \ | ||
| 508 : | (FUNC)((DST), (SRC), q); \ | ||
| 509 : | for(i=0; i<64; ++i) s+=(DST)[i]^i^qm; \ | ||
| 510 : | } \ | ||
| 511 : | emms(); \ | ||
| 512 : | } \ | ||
| 513 : | t = (gettime_usec()-t-overhead)/nb_tests/qm;\ | ||
| 514 : | s = (s&0xffff)^(s>>16) | ||
| 515 : | Isibaar | 225 | |
| 516 : | Isibaar | 262 | #define TEST_QUANT2(FUNC, DST, SRC) \ |
| 517 : | edgomez | 1060 | t = gettime_usec(); \ |
| 518 : | for(s=0,qm=1; qm<=255; ++qm) { \ | ||
| 519 : | for(i=0; i<8*8; ++i) Quant[i] = qm; \ | ||
| 520 : | set_intra_matrix( Quant ); \ | ||
| 521 : | emms(); \ | ||
| 522 : | for(q=1; q<=max_Q; ++q) { \ | ||
| 523 : | for(tst=0; tst<nb_tests; ++tst) \ | ||
| 524 : | (FUNC)((DST), (SRC), q, q); \ | ||
| 525 : | for(i=0; i<64; ++i) s+=(DST)[i]^i^qm; \ | ||
| 526 : | } \ | ||
| 527 : | emms(); \ | ||
| 528 : | } \ | ||
| 529 : | t = (gettime_usec()-t-overhead)/nb_tests/qm;\ | ||
| 530 : | s = (s&0xffff)^(s>>16) | ||
| 531 : | Isibaar | 225 | |
| 532 : | void test_quant() | ||
| 533 : | { | ||
| 534 : | edgomez | 1060 | const int nb_tests = 1*speed_ref; |
| 535 : | const int max_Q = 31; | ||
| 536 : | int i, qm; | ||
| 537 : | CPU *cpu; | ||
| 538 : | int16_t Src[8*8], Dst[8*8]; | ||
| 539 : | uint8_t Quant[8*8]; | ||
| 540 : | Isibaar | 225 | |
| 541 : | edgomez | 1060 | printf( "\n ===== test quant =====\n" ); |
| 542 : | Isibaar | 225 | |
| 543 : | edgomez | 1060 | /* we deliberately enfringe the norm's specified range [-127,127], */ |
| 544 : | /* to test the robustness of the iquant module */ | ||
| 545 : | for(i=0; i<64; ++i) { | ||
| 546 : | Src[i] = 1 + (i-32) * (i&6); | ||
| 547 : | Dst[i] = 0; | ||
| 548 : | } | ||
| 549 : | Isibaar | 225 | |
| 550 : | edgomez | 1060 | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) |
| 551 : | { | ||
| 552 : | double t, overhead; | ||
| 553 : | int tst, q; | ||
| 554 : | uint32_t s; | ||
| 555 : | Isibaar | 225 | |
| 556 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 557 : | continue; | ||
| 558 : | Isibaar | 225 | |
| 559 : | edgomez | 1060 | overhead = -gettime_usec(); |
| 560 : | for(s=0,qm=1; qm<=255; ++qm) { | ||
| 561 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
| 562 : | set_inter_matrix( Quant ); | ||
| 563 : | for(q=1; q<=max_Q; ++q) | ||
| 564 : | for(i=0; i<64; ++i) s+=Dst[i]^i^qm; | ||
| 565 : | } | ||
| 566 : | overhead += gettime_usec(); | ||
| 567 : | Isibaar | 225 | |
| 568 : | Isibaar | 262 | #if 1 |
| 569 : | edgomez | 1060 | TEST_QUANT2(quant4_intra, Dst, Src); |
| 570 : | printf( "%s - quant4_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 571 : | if (s!=29809) printf( "*** CRC ERROR! ***\n" ); | ||
| 572 : | Isibaar | 225 | |
| 573 : | edgomez | 1060 | TEST_QUANT(quant4_inter, Dst, Src); |
| 574 : | printf( "%s - quant4_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 575 : | if (s!=12574) printf( "*** CRC ERROR! ***\n" ); | ||
| 576 : | Isibaar | 262 | #endif |
| 577 : | #if 1 | ||
| 578 : | edgomez | 1060 | TEST_QUANT2(dequant4_intra, Dst, Src); |
| 579 : | printf( "%s - dequant4_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 580 : | if (s!=24052) printf( "*** CRC ERROR! ***\n" ); | ||
| 581 : | Isibaar | 225 | |
| 582 : | edgomez | 1060 | TEST_QUANT(dequant4_inter, Dst, Src); |
| 583 : | printf( "%s - dequant4_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 584 : | if (s!=63847) printf( "*** CRC ERROR! ***\n" ); | ||
| 585 : | Isibaar | 262 | #endif |
| 586 : | #if 1 | ||
| 587 : | edgomez | 1060 | TEST_QUANT2(quant_intra, Dst, Src); |
| 588 : | printf( "%s - quant_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 589 : | if (s!=25662) printf( "*** CRC ERROR! ***\n" ); | ||
| 590 : | Isibaar | 225 | |
| 591 : | edgomez | 1060 | TEST_QUANT(quant_inter, Dst, Src); |
| 592 : | printf( "%s - quant_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 593 : | if (s!=23972) printf( "*** CRC ERROR! ***\n" ); | ||
| 594 : | Isibaar | 262 | #endif |
| 595 : | #if 1 | ||
| 596 : | edgomez | 1060 | TEST_QUANT2(dequant_intra, Dst, Src); |
| 597 : | printf( "%s - dequant_intra %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 598 : | if (s!=49900) printf( "*** CRC ERROR! ***\n" ); | ||
| 599 : | Isibaar | 225 | |
| 600 : | edgomez | 1060 | TEST_QUANT(dequant_inter, Dst, Src); |
| 601 : | printf( "%s - dequant_inter %.3f usec crc=%d\n", cpu->name, t, s ); | ||
| 602 : | if (s!=48899) printf( "*** CRC ERROR! ***\n" ); | ||
| 603 : | Isibaar | 262 | #endif |
| 604 : | edgomez | 1060 | printf( " --- \n" ); |
| 605 : | } | ||
| 606 : | Isibaar | 225 | } |
| 607 : | |||
| 608 : | /********************************************************************* | ||
| 609 : | * test non-zero AC counting | ||
| 610 : | *********************************************************************/ | ||
| 611 : | |||
| 612 : | #define TEST_CBP(FUNC, SRC) \ | ||
| 613 : | edgomez | 1060 | t = gettime_usec(); \ |
| 614 : | emms(); \ | ||
| 615 : | for(tst=0; tst<nb_tests; ++tst) { \ | ||
| 616 : | cbp = (FUNC)((SRC)); \ | ||
| 617 : | } \ | ||
| 618 : | emms(); \ | ||
| 619 : | t = (gettime_usec()-t ) / nb_tests; | ||
| 620 : | Isibaar | 225 | |
| 621 : | void test_cbp() | ||
| 622 : | { | ||
| 623 : | edgomez | 1060 | const int nb_tests = 10000*speed_ref; |
| 624 : | int i; | ||
| 625 : | CPU *cpu; | ||
| 626 : | int16_t Src1[6*64], Src2[6*64], Src3[6*64], Src4[6*64]; | ||
| 627 : | Isibaar | 225 | |
| 628 : | edgomez | 1060 | printf( "\n ===== test cbp =====\n" ); |
| 629 : | Isibaar | 225 | |
| 630 : | edgomez | 1060 | for(i=0; i<6*64; ++i) { |
| 631 : | Src1[i] = (i*i*3/8192)&(i/64)&1; /* 'random' */ | ||
| 632 : | Src2[i] = (i<3*64); /* half-full */ | ||
| 633 : | Src3[i] = ((i+32)>3*64); | ||
| 634 : | Src4[i] = (i==(3*64+2) || i==(5*64+9)); | ||
| 635 : | } | ||
| 636 : | Isibaar | 225 | |
| 637 : | edgomez | 1060 | for(cpu = cpu_short_list2; cpu->name!=0; ++cpu) |
| 638 : | { | ||
| 639 : | double t; | ||
| 640 : | int tst, cbp; | ||
| 641 : | Isibaar | 225 | |
| 642 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 643 : | continue; | ||
| 644 : | |||
| 645 : | TEST_CBP(calc_cbp, Src1); | ||
| 646 : | printf( "%s - calc_cbp#1 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
| 647 : | if (cbp!=0x15) printf( "*** CRC ERROR! ***\n" ); | ||
| 648 : | TEST_CBP(calc_cbp, Src2); | ||
| 649 : | printf( "%s - calc_cbp#2 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
| 650 : | if (cbp!=0x38) printf( "*** CRC ERROR! ***\n" ); | ||
| 651 : | TEST_CBP(calc_cbp, Src3); | ||
| 652 : | printf( "%s - calc_cbp#3 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
| 653 : | if (cbp!=0x0f) printf( "*** CRC ERROR! ***\n" ); | ||
| 654 : | TEST_CBP(calc_cbp, Src4); | ||
| 655 : | printf( "%s - calc_cbp#4 %.3f usec cbp=0x%x\n", cpu->name, t, cbp ); | ||
| 656 : | if (cbp!=0x05) printf( "*** CRC ERROR! ***\n" ); | ||
| 657 : | printf( " --- \n" ); | ||
| 658 : | } | ||
| 659 : | Isibaar | 225 | } |
| 660 : | |||
| 661 : | /********************************************************************* | ||
| 662 : | Isibaar | 262 | * fdct/idct IEEE1180 compliance |
| 663 : | *********************************************************************/ | ||
| 664 : | |||
| 665 : | typedef struct { | ||
| 666 : | edgomez | 1060 | long Errors[64]; |
| 667 : | long Sqr_Errors[64]; | ||
| 668 : | long Max_Errors[64]; | ||
| 669 : | long Nb; | ||
| 670 : | Isibaar | 262 | } STATS_8x8; |
| 671 : | |||
| 672 : | void init_stats(STATS_8x8 *S) | ||
| 673 : | { | ||
| 674 : | edgomez | 1060 | int i; |
| 675 : | for(i=0; i<64; ++i) { | ||
| 676 : | S->Errors[i] = 0; | ||
| 677 : | S->Sqr_Errors[i] = 0; | ||
| 678 : | S->Max_Errors[i] = 0; | ||
| 679 : | } | ||
| 680 : | S->Nb = 0; | ||
| 681 : | Isibaar | 262 | } |
| 682 : | |||
| 683 : | void store_stats(STATS_8x8 *S, short Blk[64], short Ref[64]) | ||
| 684 : | { | ||
| 685 : | edgomez | 1060 | int i; |
| 686 : | for(i=0; i<64; ++i) | ||
| 687 : | { | ||
| 688 : | short Err = Blk[i] - Ref[i]; | ||
| 689 : | S->Errors[i] += Err; | ||
| 690 : | S->Sqr_Errors[i] += Err * Err; | ||
| 691 : | if (Err<0) Err = -Err; | ||
| 692 : | if (S->Max_Errors[i]<Err) | ||
| 693 : | S->Max_Errors[i] = Err; | ||
| 694 : | } | ||
| 695 : | S->Nb++; | ||
| 696 : | Isibaar | 262 | } |
| 697 : | |||
| 698 : | void print_stats(STATS_8x8 *S) | ||
| 699 : | { | ||
| 700 : | edgomez | 1060 | int i; |
| 701 : | double Norm; | ||
| 702 : | Isibaar | 262 | |
| 703 : | edgomez | 1060 | assert(S->Nb>0); |
| 704 : | Norm = 1. / (double)S->Nb; | ||
| 705 : | printf("\n== Max absolute values of errors ==\n"); | ||
| 706 : | for(i=0; i<64; i++) { | ||
| 707 : | printf(" %4ld", S->Max_Errors[i]); | ||
| 708 : | if ((i&7)==7) printf("\n"); | ||
| 709 : | } | ||
| 710 : | Isibaar | 262 | |
| 711 : | edgomez | 1060 | printf("\n== Mean square errors ==\n"); |
| 712 : | for(i=0; i<64; i++) | ||
| 713 : | { | ||
| 714 : | double Err = Norm * (double)S->Sqr_Errors[i]; | ||
| 715 : | printf(" %.3f", Err); | ||
| 716 : | if ((i&7)==7) printf("\n"); | ||
| 717 : | } | ||
| 718 : | Isibaar | 262 | |
| 719 : | edgomez | 1060 | printf("\n== Mean errors ==\n"); |
| 720 : | for(i=0; i<64; i++) | ||
| 721 : | { | ||
| 722 : | double Err = Norm * (double)S->Errors[i]; | ||
| 723 : | printf(" %.3f", Err); | ||
| 724 : | if ((i&7)==7) printf("\n"); | ||
| 725 : | } | ||
| 726 : | printf("\n"); | ||
| 727 : | Isibaar | 262 | } |
| 728 : | |||
| 729 : | static const char *CHECK(double v, double l) { | ||
| 730 : | edgomez | 1060 | if (fabs(v)<=l) return "ok"; |
| 731 : | else return "FAIL!"; | ||
| 732 : | Isibaar | 262 | } |
| 733 : | |||
| 734 : | void report_stats(STATS_8x8 *S, const double *Limits) | ||
| 735 : | { | ||
| 736 : | edgomez | 1060 | int i; |
| 737 : | double Norm, PE, PMSE, OMSE, PME, OME; | ||
| 738 : | Isibaar | 262 | |
| 739 : | edgomez | 1060 | assert(S->Nb>0); |
| 740 : | Norm = 1. / (double)S->Nb; | ||
| 741 : | PE = 0.; | ||
| 742 : | for(i=0; i<64; i++) { | ||
| 743 : | if (PE<S->Max_Errors[i]) | ||
| 744 : | PE = S->Max_Errors[i]; | ||
| 745 : | } | ||
| 746 : | Isibaar | 262 | |
| 747 : | edgomez | 1060 | PMSE = 0.; |
| 748 : | OMSE = 0.; | ||
| 749 : | for(i=0; i<64; i++) | ||
| 750 : | { | ||
| 751 : | double Err = Norm * (double)S->Sqr_Errors[i]; | ||
| 752 : | OMSE += Err; | ||
| 753 : | if (PMSE < Err) PMSE = Err; | ||
| 754 : | } | ||
| 755 : | OMSE /= 64.; | ||
| 756 : | Isibaar | 262 | |
| 757 : | edgomez | 1060 | PME = 0.; |
| 758 : | OME = 0.; | ||
| 759 : | for(i=0; i<64; i++) | ||
| 760 : | { | ||
| 761 : | double Err = Norm * (double)S->Errors[i]; | ||
| 762 : | OME += Err; | ||
| 763 : | Err = fabs(Err); | ||
| 764 : | if (PME < Err) PME = Err; | ||
| 765 : | } | ||
| 766 : | OME /= 64.; | ||
| 767 : | Isibaar | 262 | |
| 768 : | edgomez | 1060 | printf( "Peak error: %4.4f\n", PE ); |
| 769 : | printf( "Peak MSE: %4.4f\n", PMSE ); | ||
| 770 : | printf( "Overall MSE: %4.4f\n", OMSE ); | ||
| 771 : | printf( "Peak ME: %4.4f\n", PME ); | ||
| 772 : | printf( "Overall ME: %4.4f\n", OME ); | ||
| 773 : | |||
| 774 : | if (Limits!=0) | ||
| 775 : | { | ||
| 776 : | printf( "[PE<=%.4f %s] ", Limits[0], CHECK(PE, Limits[0]) ); | ||
| 777 : | printf( "\n" ); | ||
| 778 : | printf( "[PMSE<=%.4f %s]", Limits[1], CHECK(PMSE, Limits[1]) ); | ||
| 779 : | printf( "[OMSE<=%.4f %s]", Limits[2], CHECK(OMSE, Limits[2]) ); | ||
| 780 : | printf( "\n" ); | ||
| 781 : | printf( "[PME<=%.4f %s] ", Limits[3], CHECK(PME , Limits[3]) ); | ||
| 782 : | printf( "[OME<=%.4f %s] ", Limits[4], CHECK(OME , Limits[4]) ); | ||
| 783 : | printf( "\n" ); | ||
| 784 : | } | ||
| 785 : | Isibaar | 262 | } |
| 786 : | |||
| 787 : | edgomez | 1053 | ///* ////////////////////////////////////////////////////// */ |
| 788 : | Isibaar | 262 | /* Pseudo-random generator specified by IEEE 1180 */ |
| 789 : | |||
| 790 : | static long ieee_seed = 1; | ||
| 791 : | static void ieee_reseed(long s) { | ||
| 792 : | edgomez | 1060 | ieee_seed = s; |
| 793 : | Isibaar | 262 | } |
| 794 : | static long ieee_rand(int Min, int Max) | ||
| 795 : | { | ||
| 796 : | edgomez | 1060 | static double z = (double) 0x7fffffff; |
| 797 : | Isibaar | 262 | |
| 798 : | edgomez | 1060 | long i,j; |
| 799 : | double x; | ||
| 800 : | Isibaar | 262 | |
| 801 : | edgomez | 1060 | ieee_seed = (ieee_seed * 1103515245) + 12345; |
| 802 : | i = ieee_seed & 0x7ffffffe; | ||
| 803 : | x = ((double) i) / z; | ||
| 804 : | x *= (Max-Min+1); | ||
| 805 : | j = (long)x; | ||
| 806 : | j = j + Min; | ||
| 807 : | assert(j>=Min && j<=Max); | ||
| 808 : | return (short)j; | ||
| 809 : | Isibaar | 262 | } |
| 810 : | |||
| 811 : | #define CLAMP(x, M) (x) = ((x)<-(M)) ? (-(M)) : ((x)>=(M) ? ((M)-1) : (x)) | ||
| 812 : | |||
| 813 : | static double Cos[8][8]; | ||
| 814 : | static void init_ref_dct() | ||
| 815 : | { | ||
| 816 : | edgomez | 1060 | int i, j; |
| 817 : | for(i=0; i<8; i++) | ||
| 818 : | { | ||
| 819 : | double scale = (i == 0) ? sqrt(0.125) : 0.5; | ||
| 820 : | for (j=0; j<8; j++) | ||
| 821 : | Cos[i][j] = scale*cos( (M_PI/8.0)*i*(j + 0.5) ); | ||
| 822 : | } | ||
| 823 : | Isibaar | 262 | } |
| 824 : | |||
| 825 : | void ref_idct(short *M) | ||
| 826 : | { | ||
| 827 : | edgomez | 1060 | int i, j, k; |
| 828 : | double Tmp[8][8]; | ||
| 829 : | Isibaar | 262 | |
| 830 : | edgomez | 1060 | for(i=0; i<8; i++) { |
| 831 : | for(j=0; j<8; j++) | ||
| 832 : | { | ||
| 833 : | double Sum = 0.0; | ||
| 834 : | for (k=0; k<8; k++) Sum += Cos[k][j]*M[8*i+k]; | ||
| 835 : | Tmp[i][j] = Sum; | ||
| 836 : | } | ||
| 837 : | } | ||
| 838 : | for(i=0; i<8; i++) { | ||
| 839 : | for(j=0; j<8; j++) { | ||
| 840 : | double Sum = 0.0; | ||
| 841 : | for (k=0; k<8; k++) Sum += Cos[k][i]*Tmp[k][j]; | ||
| 842 : | M[8*i+j] = (short)floor(Sum + .5); | ||
| 843 : | } | ||
| 844 : | } | ||
| 845 : | Isibaar | 262 | } |
| 846 : | |||
| 847 : | void ref_fdct(short *M) | ||
| 848 : | { | ||
| 849 : | edgomez | 1060 | int i, j, k; |
| 850 : | double Tmp[8][8]; | ||
| 851 : | Isibaar | 262 | |
| 852 : | edgomez | 1060 | for(i=0; i<8; i++) { |
| 853 : | for(j=0; j<8; j++) | ||
| 854 : | { | ||
| 855 : | double Sum = 0.0; | ||
| 856 : | for (k=0; k<8; k++) Sum += Cos[j][k]*M[8*i+k]; | ||
| 857 : | Tmp[i][j] = Sum; | ||
| 858 : | } | ||
| 859 : | } | ||
| 860 : | for(i=0; i<8; i++) { | ||
| 861 : | for(j=0; j<8; j++) { | ||
| 862 : | double Sum = 0.0; | ||
| 863 : | for (k=0; k<8; k++) Sum += Cos[i][k]*Tmp[k][j]; | ||
| 864 : | M[8*i+j] = (short)floor(Sum + 0.5); | ||
| 865 : | } | ||
| 866 : | } | ||
| 867 : | Isibaar | 262 | } |
| 868 : | |||
| 869 : | void test_IEEE1180_compliance(int Min, int Max, int Sign) | ||
| 870 : | { | ||
| 871 : | edgomez | 1060 | static const double ILimits[5] = { 1., 0.06, 0.02, 0.015, 0.0015 }; |
| 872 : | int Loops = 10000; | ||
| 873 : | int i, m, n; | ||
| 874 : | short Blk0[64]; /* reference */ | ||
| 875 : | short Blk[64], iBlk[64]; | ||
| 876 : | short Ref_FDCT[64]; | ||
| 877 : | short Ref_IDCT[64]; | ||
| 878 : | Isibaar | 262 | |
| 879 : | edgomez | 1060 | STATS_8x8 FStats; /* forward dct stats */ |
| 880 : | STATS_8x8 IStats; /* inverse dct stats */ | ||
| 881 : | Isibaar | 262 | |
| 882 : | edgomez | 1060 | CPU *cpu; |
| 883 : | Isibaar | 262 | |
| 884 : | edgomez | 1060 | init_ref_dct(); |
| 885 : | Isibaar | 262 | |
| 886 : | edgomez | 1060 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 887 : | { | ||
| 888 : | if (!init_cpu(cpu)) | ||
| 889 : | continue; | ||
| 890 : | Isibaar | 262 | |
| 891 : | edgomez | 1060 | printf( "\n===== IEEE test for %s ==== (Min=%d Max=%d Sign=%d Loops=%d)\n", |
| 892 : | cpu->name, Min, Max, Sign, Loops); | ||
| 893 : | Isibaar | 262 | |
| 894 : | edgomez | 1060 | init_stats(&IStats); |
| 895 : | init_stats(&FStats); | ||
| 896 : | Isibaar | 262 | |
| 897 : | edgomez | 1060 | ieee_reseed(1); |
| 898 : | for(n=0; n<Loops; ++n) | ||
| 899 : | { | ||
| 900 : | for(i=0; i<64; ++i) | ||
| 901 : | Blk0[i] = (short)ieee_rand(Min,Max) * Sign; | ||
| 902 : | Isibaar | 262 | |
| 903 : | edgomez | 1060 | /* hmm, I'm not quite sure this is exactly */ |
| 904 : | /* the tests described in the norm. check... */ | ||
| 905 : | Isibaar | 262 | |
| 906 : | edgomez | 1060 | memcpy(Ref_FDCT, Blk0, 64*sizeof(short)); |
| 907 : | ref_fdct(Ref_FDCT); | ||
| 908 : | for(i=0; i<64; i++) CLAMP( Ref_FDCT[i], 2048 ); | ||
| 909 : | Isibaar | 262 | |
| 910 : | edgomez | 1060 | memcpy(Blk, Blk0, 64*sizeof(short)); |
| 911 : | emms(); fdct(Blk); emms(); | ||
| 912 : | for(i=0; i<64; i++) CLAMP( Blk[i], 2048 ); | ||
| 913 : | Isibaar | 262 | |
| 914 : | edgomez | 1060 | store_stats(&FStats, Blk, Ref_FDCT); |
| 915 : | Isibaar | 262 | |
| 916 : | |||
| 917 : | edgomez | 1060 | memcpy(Ref_IDCT, Ref_FDCT, 64*sizeof(short)); |
| 918 : | ref_idct(Ref_IDCT); | ||
| 919 : | for (i=0; i<64; i++) CLAMP( Ref_IDCT[i], 256 ); | ||
| 920 : | Isibaar | 262 | |
| 921 : | edgomez | 1060 | memcpy(iBlk, Ref_FDCT, 64*sizeof(short)); |
| 922 : | emms(); idct(iBlk); emms(); | ||
| 923 : | for(i=0; i<64; i++) CLAMP( iBlk[i], 256 ); | ||
| 924 : | Isibaar | 262 | |
| 925 : | edgomez | 1060 | store_stats(&IStats, iBlk, Ref_IDCT); |
| 926 : | } | ||
| 927 : | Isibaar | 262 | |
| 928 : | |||
| 929 : | edgomez | 1060 | printf( "\n -- FDCT report --\n" ); |
| 930 : | edgomez | 851 | // print_stats(&FStats); |
| 931 : | edgomez | 1060 | report_stats(&FStats, 0); /* so far I know, IEEE1180 says nothing for fdct */ |
| 932 : | Isibaar | 262 | |
| 933 : | edgomez | 1060 | for(i=0; i<64; i++) Blk[i] = 0; |
| 934 : | emms(); fdct(Blk); emms(); | ||
| 935 : | for(m=i=0; i<64; i++) if (Blk[i]!=0) m++; | ||
| 936 : | printf( "FDCT(0) == 0 ? %s\n", (m!=0) ? "NOPE!" : "yup." ); | ||
| 937 : | Isibaar | 262 | |
| 938 : | edgomez | 1060 | printf( "\n -- IDCT report --\n" ); |
| 939 : | edgomez | 851 | // print_stats(&IStats); |
| 940 : | edgomez | 1060 | report_stats(&IStats, ILimits); |
| 941 : | Isibaar | 262 | |
| 942 : | |||
| 943 : | edgomez | 1060 | for(i=0; i<64; i++) Blk[i] = 0; |
| 944 : | emms(); idct(Blk); emms(); | ||
| 945 : | for(m=i=0; i<64; i++) if (Blk[i]!=0) m++; | ||
| 946 : | printf( "IDCT(0) == 0 ? %s\n", (m!=0) ? "NOPE!" : "yup." ); | ||
| 947 : | } | ||
| 948 : | Isibaar | 262 | } |
| 949 : | |||
| 950 : | |||
| 951 : | void test_dct_saturation(int Min, int Max) | ||
| 952 : | { | ||
| 953 : | edgomez | 1060 | /* test behaviour on input range fringe */ |
| 954 : | Isibaar | 262 | |
| 955 : | edgomez | 1060 | int i, n, p; |
| 956 : | CPU *cpu; | ||
| 957 : | edgomez | 1053 | // const short IDCT_MAX = 2047; /* 12bits input */ |
| 958 : | edgomez | 851 | // const short IDCT_MIN = -2048; |
| 959 : | edgomez | 1053 | // const short IDCT_OUT = 256; /* 9bits ouput */ |
| 960 : | edgomez | 1060 | const int Partitions = 4; |
| 961 : | const int Loops = 10000 / Partitions; | ||
| 962 : | Isibaar | 262 | |
| 963 : | edgomez | 1060 | init_ref_dct(); |
| 964 : | Isibaar | 262 | |
| 965 : | edgomez | 1060 | for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 966 : | { | ||
| 967 : | short Blk0[64], Blk[64]; | ||
| 968 : | STATS_8x8 Stats; | ||
| 969 : | Isibaar | 262 | |
| 970 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 971 : | continue; | ||
| 972 : | Isibaar | 262 | |
| 973 : | edgomez | 1060 | printf( "\n===== IEEE test for %s Min=%d Max=%d =====\n", |
| 974 : | cpu->name, Min, Max ); | ||
| 975 : | Isibaar | 262 | |
| 976 : | edgomez | 1060 | /* FDCT tests // */ |
| 977 : | Isibaar | 262 | |
| 978 : | edgomez | 1060 | init_stats(&Stats); |
| 979 : | Isibaar | 262 | |
| 980 : | edgomez | 1060 | /* test each computation channels separately */ |
| 981 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? Max : 0; | ||
| 982 : | ref_fdct(Blk0); | ||
| 983 : | emms(); fdct(Blk); emms(); | ||
| 984 : | store_stats(&Stats, Blk, Blk0); | ||
| 985 : | Isibaar | 262 | |
| 986 : | edgomez | 1060 | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? Min : 0; |
| 987 : | ref_fdct(Blk0); | ||
| 988 : | emms(); fdct(Blk); emms(); | ||
| 989 : | store_stats(&Stats, Blk, Blk0); | ||
| 990 : | Isibaar | 262 | |
| 991 : | edgomez | 1060 | /* randomly saturated inputs */ |
| 992 : | for(p=0; p<Partitions; ++p) | ||
| 993 : | { | ||
| 994 : | for(n=0; n<Loops; ++n) | ||
| 995 : | { | ||
| 996 : | for(i=0; i<64; ++i) | ||
| 997 : | Blk0[i] = Blk[i] = (ieee_rand(0,Partitions)>=p)? Max : Min; | ||
| 998 : | ref_fdct(Blk0); | ||
| 999 : | emms(); fdct(Blk); emms(); | ||
| 1000 : | store_stats(&Stats, Blk, Blk0); | ||
| 1001 : | } | ||
| 1002 : | } | ||
| 1003 : | printf( "\n -- FDCT saturation report --\n" ); | ||
| 1004 : | report_stats(&Stats, 0); | ||
| 1005 : | Isibaar | 262 | |
| 1006 : | |||
| 1007 : | edgomez | 1060 | /* IDCT tests // */ |
| 1008 : | Isibaar | 262 | #if 0 |
| 1009 : | edgomez | 1060 | /* no finished yet */ |
| 1010 : | Isibaar | 262 | |
| 1011 : | edgomez | 1060 | init_stats(&Stats); |
| 1012 : | Isibaar | 262 | |
| 1013 : | edgomez | 1060 | /* test each computation channel separately */ |
| 1014 : | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? IDCT_MAX : 0; | ||
| 1015 : | ref_idct(Blk0); | ||
| 1016 : | emms(); idct(Blk); emms(); | ||
| 1017 : | for(i=0; i<64; i++) { CLAMP(Blk0[i], IDCT_OUT); CLAMP(Blk[i], IDCT_OUT); } | ||
| 1018 : | store_stats(&Stats, Blk, Blk0); | ||
| 1019 : | Isibaar | 262 | |
| 1020 : | edgomez | 1060 | for(i=0; i<64; i++) Blk[i] = Blk0[i] = ((i/8)==(i%8)) ? IDCT_MIN : 0; |
| 1021 : | ref_idct(Blk0); | ||
| 1022 : | emms(); idct(Blk); emms(); | ||
| 1023 : | for(i=0; i<64; i++) { CLAMP(Blk0[i], IDCT_OUT); CLAMP(Blk[i], IDCT_OUT); } | ||
| 1024 : | store_stats(&Stats, Blk, Blk0); | ||
| 1025 : | Isibaar | 262 | |
| 1026 : | edgomez | 1060 | /* randomly saturated inputs */ |
| 1027 : | for(p=0; p<Partitions; ++p) | ||
| 1028 : | { | ||
| 1029 : | for(n=0; n<Loops; ++n) | ||
| 1030 : | { | ||
| 1031 : | for(i=0; i<64; ++i) | ||
| 1032 : | Blk0[i] = Blk[i] = (ieee_rand(0,Partitions)>=p)? IDCT_MAX : IDCT_MIN; | ||
| 1033 : | ref_idct(Blk0); | ||
| 1034 : | emms(); idct(Blk); emms(); | ||
| 1035 : | for(i=0; i<64; i++) { CLAMP(Blk0[i],IDCT_OUT); CLAMP(Blk[i],IDCT_OUT); } | ||
| 1036 : | store_stats(&Stats, Blk, Blk0); | ||
| 1037 : | } | ||
| 1038 : | } | ||
| 1039 : | Isibaar | 262 | |
| 1040 : | edgomez | 1060 | printf( "\n -- IDCT saturation report --\n" ); |
| 1041 : | print_stats(&Stats); | ||
| 1042 : | report_stats(&Stats, 0); | ||
| 1043 : | Isibaar | 262 | #endif |
| 1044 : | edgomez | 1060 | } |
| 1045 : | Isibaar | 262 | } |
| 1046 : | |||
| 1047 : | /********************************************************************* | ||
| 1048 : | Isibaar | 225 | * measure raw decoding speed |
| 1049 : | *********************************************************************/ | ||
| 1050 : | |||
| 1051 : | void test_dec(const char *name, int width, int height, int with_chksum) | ||
| 1052 : | { | ||
| 1053 : | edgomez | 1060 | FILE *f = 0; |
| 1054 : | void *dechandle = 0; | ||
| 1055 : | int xerr; | ||
| 1056 : | xvid_gbl_init_t xinit; | ||
| 1057 : | xvid_dec_create_t xparam; | ||
| 1058 : | xvid_dec_frame_t xframe; | ||
| 1059 : | Isibaar | 225 | double t = 0.; |
| 1060 : | int nb = 0; | ||
| 1061 : | edgomez | 1060 | uint8_t *buf = 0; |
| 1062 : | uint8_t *rgb_out = 0; | ||
| 1063 : | int buf_size, pos; | ||
| 1064 : | uint32_t chksum = 0; | ||
| 1065 : | Isibaar | 225 | |
| 1066 : | edgomez | 1060 | memset(&xinit, 0, sizeof(xinit)); |
| 1067 : | Isibaar | 262 | xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; |
| 1068 : | edgomez | 1060 | xinit.version = XVID_VERSION; |
| 1069 : | xvid_global(NULL, 0, &xinit, NULL); | ||
| 1070 : | Isibaar | 225 | |
| 1071 : | edgomez | 1060 | memset(&xparam, 0, sizeof(xparam)); |
| 1072 : | Isibaar | 225 | xparam.width = width; |
| 1073 : | xparam.height = height; | ||
| 1074 : | edgomez | 1060 | xparam.version = XVID_VERSION; |
| 1075 : | Isibaar | 225 | xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
| 1076 : | edgomez | 1060 | if (xerr==XVID_ERR_FAIL) { |
| 1077 : | printf("can't init decoder (err=%d)\n", xerr); | ||
| 1078 : | return; | ||
| 1079 : | Isibaar | 225 | } |
| 1080 : | dechandle = xparam.handle; | ||
| 1081 : | |||
| 1082 : | |||
| 1083 : | f = fopen(name, "rb"); | ||
| 1084 : | edgomez | 1060 | if (f==0) { |
| 1085 : | printf( "can't open file '%s'\n", name); | ||
| 1086 : | return; | ||
| 1087 : | } | ||
| 1088 : | fseek(f, 0, SEEK_END); | ||
| 1089 : | buf_size = ftell(f); | ||
| 1090 : | fseek(f, 0, SEEK_SET); | ||
| 1091 : | if (buf_size<=0) { | ||
| 1092 : | printf("error while stating file\n"); | ||
| 1093 : | goto End; | ||
| 1094 : | } | ||
| 1095 : | else printf( "Input size: %d\n", buf_size); | ||
| 1096 : | Isibaar | 225 | |
| 1097 : | edgomez | 1060 | buf = malloc(buf_size); /* should be enuf' */ |
| 1098 : | rgb_out = calloc(4, width*height); /* <-room for _RGB24 */ | ||
| 1099 : | if (buf==0 || rgb_out==0) { | ||
| 1100 : | printf( "malloc failed!\n" ); | ||
| 1101 : | goto End; | ||
| 1102 : | } | ||
| 1103 : | Isibaar | 225 | |
| 1104 : | edgomez | 1060 | if (fread(buf, buf_size, 1, f)!=1) { |
| 1105 : | printf( "file-read failed\n" ); | ||
| 1106 : | goto End; | ||
| 1107 : | } | ||
| 1108 : | Isibaar | 225 | |
| 1109 : | edgomez | 1060 | nb = 0; |
| 1110 : | pos = 0; | ||
| 1111 : | t = -gettime_usec(); | ||
| 1112 : | while(1) { | ||
| 1113 : | memset(&xframe, 0, sizeof(xframe)); | ||
| 1114 : | xframe.version = XVID_VERSION; | ||
| 1115 : | xframe.bitstream = buf + pos; | ||
| 1116 : | xframe.length = buf_size - pos; | ||
| 1117 : | xframe.output.plane[0] = rgb_out; | ||
| 1118 : | xframe.output.stride[0] = width; | ||
| 1119 : | xframe.output.csp = XVID_CSP_BGR; | ||
| 1120 : | xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); | ||
| 1121 : | nb++; | ||
| 1122 : | pos += xframe.length; | ||
| 1123 : | if (with_chksum) { | ||
| 1124 : | int k = width*height; | ||
| 1125 : | uint32_t *ptr = (uint32_t *)rgb_out; | ||
| 1126 : | while(k-->0) chksum += *ptr++; | ||
| 1127 : | } | ||
| 1128 : | if (pos==buf_size) | ||
| 1129 : | break; | ||
| 1130 : | if (xerr==XVID_ERR_FAIL) { | ||
| 1131 : | printf("decoding failed for frame #%d (err=%d)!\n", nb, xerr); | ||
| 1132 : | break; | ||
| 1133 : | } | ||
| 1134 : | } | ||
| 1135 : | t += gettime_usec(); | ||
| 1136 : | if (t>0.) | ||
| 1137 : | printf( "%d frames decoded in %.3f s -> %.1f FPS\n", nb, t*1.e-6f, (float)(nb*1.e6f/t) ); | ||
| 1138 : | if (with_chksum) | ||
| 1139 : | printf("checksum: 0x%.8x\n", chksum); | ||
| 1140 : | Isibaar | 225 | |
| 1141 : | edgomez | 1060 | End: |
| 1142 : | if (rgb_out!=0) free(rgb_out); | ||
| 1143 : | if (buf!=0) free(buf); | ||
| 1144 : | if (dechandle!=0) { | ||
| 1145 : | xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); | ||
| 1146 : | if (xerr==XVID_ERR_FAIL) | ||
| 1147 : | printf("destroy-decoder failed (err=%d)!\n", xerr); | ||
| 1148 : | } | ||
| 1149 : | if (f!=0) fclose(f); | ||
| 1150 : | Isibaar | 225 | } |
| 1151 : | |||
| 1152 : | /********************************************************************* | ||
| 1153 : | * non-regression tests | ||
| 1154 : | *********************************************************************/ | ||
| 1155 : | |||
| 1156 : | void test_bugs1() | ||
| 1157 : | { | ||
| 1158 : | edgomez | 1060 | CPU *cpu; |
| 1159 : | Isibaar | 225 | |
| 1160 : | edgomez | 1060 | printf( "\n ===== (de)quant4_intra saturation bug? =====\n" ); |
| 1161 : | Isibaar | 225 | |
| 1162 : | edgomez | 1060 | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) |
| 1163 : | { | ||
| 1164 : | int i; | ||
| 1165 : | int16_t Src[8*8], Dst[8*8]; | ||
| 1166 : | Isibaar | 225 | |
| 1167 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 1168 : | continue; | ||
| 1169 : | Isibaar | 225 | |
| 1170 : | edgomez | 1060 | for(i=0; i<64; ++i) Src[i] = i-32; |
| 1171 : | set_intra_matrix( get_default_intra_matrix() ); | ||
| 1172 : | dequant4_intra(Dst, Src, 31, 5); | ||
| 1173 : | printf( "dequant4_intra with CPU=%s: ", cpu->name); | ||
| 1174 : | printf( " Out[]= " ); | ||
| 1175 : | for(i=0; i<64; ++i) printf( "[%d]", Dst[i]); | ||
| 1176 : | printf( "\n" ); | ||
| 1177 : | } | ||
| 1178 : | Isibaar | 225 | |
| 1179 : | edgomez | 1060 | printf( "\n ===== (de)quant4_inter saturation bug? =====\n" ); |
| 1180 : | Isibaar | 225 | |
| 1181 : | edgomez | 1060 | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) |
| 1182 : | { | ||
| 1183 : | int i; | ||
| 1184 : | int16_t Src[8*8], Dst[8*8]; | ||
| 1185 : | Isibaar | 225 | |
| 1186 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 1187 : | continue; | ||
| 1188 : | Isibaar | 225 | |
| 1189 : | edgomez | 1060 | for(i=0; i<64; ++i) Src[i] = i-32; |
| 1190 : | set_inter_matrix( get_default_inter_matrix() ); | ||
| 1191 : | dequant4_inter(Dst, Src, 31); | ||
| 1192 : | printf( "dequant4_inter with CPU=%s: ", cpu->name); | ||
| 1193 : | printf( " Out[]= " ); | ||
| 1194 : | for(i=0; i<64; ++i) printf( "[%d]", Dst[i]); | ||
| 1195 : | printf( "\n" ); | ||
| 1196 : | } | ||
| 1197 : | Isibaar | 225 | } |
| 1198 : | |||
| 1199 : | void test_dct_precision_diffs() | ||
| 1200 : | { | ||
| 1201 : | edgomez | 1060 | CPU *cpu; |
| 1202 : | short Blk[8*8], Blk0[8*8]; | ||
| 1203 : | Isibaar | 225 | |
| 1204 : | edgomez | 1060 | printf( "\n ===== fdct/idct precision diffs =====\n" ); |
| 1205 : | Isibaar | 225 | |
| 1206 : | edgomez | 1060 | for(cpu = cpu_short_list; cpu->name!=0; ++cpu) |
| 1207 : | { | ||
| 1208 : | int i; | ||
| 1209 : | Isibaar | 225 | |
| 1210 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 1211 : | continue; | ||
| 1212 : | Isibaar | 225 | |
| 1213 : | edgomez | 1060 | for(i=0; i<8*8; ++i) { |
| 1214 : | Blk0[i] = (i*7-i*i) & 0x7f; | ||
| 1215 : | Blk[i] = Blk0[i]; | ||
| 1216 : | } | ||
| 1217 : | Isibaar | 225 | |
| 1218 : | edgomez | 1060 | fdct(Blk); |
| 1219 : | idct(Blk); | ||
| 1220 : | printf( " fdct+idct diffs with CPU=%s: \n", cpu->name ); | ||
| 1221 : | for(i=0; i<8; ++i) { | ||
| 1222 : | int j; | ||
| 1223 : | for(j=0; j<8; ++j) printf( " %d ", Blk[i*8+j]-Blk0[i*8+j]); | ||
| 1224 : | printf("\n"); | ||
| 1225 : | } | ||
| 1226 : | printf("\n"); | ||
| 1227 : | } | ||
| 1228 : | Isibaar | 225 | } |
| 1229 : | |||
| 1230 : | Isibaar | 262 | void test_quant_bug() |
| 1231 : | { | ||
| 1232 : | edgomez | 1060 | const int max_Q = 31; |
| 1233 : | int i, n, qm, q; | ||
| 1234 : | CPU *cpu; | ||
| 1235 : | int16_t Src[8*8], Dst[8*8]; | ||
| 1236 : | uint8_t Quant[8*8]; | ||
| 1237 : | CPU cpu_bug_list[] = { { "PLAINC", 0 }, { "MMX ", XVID_CPU_MMX }, {0,0} }; | ||
| 1238 : | uint16_t Crcs_Inter[2][32]; | ||
| 1239 : | uint16_t Crcs_Intra[2][32]; | ||
| 1240 : | printf( "\n ===== test MPEG4-quantize bug =====\n" ); | ||
| 1241 : | Isibaar | 225 | |
| 1242 : | edgomez | 1060 | for(i=0; i<64; ++i) Src[i] = 2048*(i-32)/32; |
| 1243 : | Isibaar | 262 | |
| 1244 : | #if 1 | ||
| 1245 : | edgomez | 1060 | for(qm=1; qm<=255; ++qm) |
| 1246 : | { | ||
| 1247 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
| 1248 : | set_inter_matrix( Quant ); | ||
| 1249 : | Isibaar | 262 | |
| 1250 : | edgomez | 1060 | for(n=0, cpu = cpu_bug_list; cpu->name!=0; ++cpu, ++n) |
| 1251 : | { | ||
| 1252 : | uint16_t s; | ||
| 1253 : | Isibaar | 262 | |
| 1254 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 1255 : | continue; | ||
| 1256 : | Isibaar | 262 | |
| 1257 : | edgomez | 1060 | for(q=1; q<=max_Q; ++q) { |
| 1258 : | emms(); | ||
| 1259 : | quant4_inter( Dst, Src, q ); | ||
| 1260 : | emms(); | ||
| 1261 : | for(s=0, i=0; i<64; ++i) s+=((uint16_t)Dst[i])^i; | ||
| 1262 : | Crcs_Inter[n][q] = s; | ||
| 1263 : | } | ||
| 1264 : | } | ||
| 1265 : | Isibaar | 262 | |
| 1266 : | edgomez | 1060 | for(q=1; q<=max_Q; ++q) |
| 1267 : | for(i=0; i<n-1; ++i) | ||
| 1268 : | if (Crcs_Inter[i][q]!=Crcs_Inter[i+1][q]) | ||
| 1269 : | printf( "Discrepancy Inter: qm=%d, q=%d -> %d/%d !\n", | ||
| 1270 : | qm, q, Crcs_Inter[i][q], Crcs_Inter[i+1][q]); | ||
| 1271 : | } | ||
| 1272 : | Isibaar | 262 | #endif |
| 1273 : | |||
| 1274 : | #if 1 | ||
| 1275 : | edgomez | 1060 | for(qm=1; qm<=255; ++qm) |
| 1276 : | { | ||
| 1277 : | for(i=0; i<8*8; ++i) Quant[i] = qm; | ||
| 1278 : | set_intra_matrix( Quant ); | ||
| 1279 : | Isibaar | 262 | |
| 1280 : | edgomez | 1060 | for(n=0, cpu = cpu_bug_list; cpu->name!=0; ++cpu, ++n) |
| 1281 : | { | ||
| 1282 : | uint16_t s; | ||
| 1283 : | Isibaar | 262 | |
| 1284 : | edgomez | 1060 | if (!init_cpu(cpu)) |
| 1285 : | continue; | ||
| 1286 : | Isibaar | 262 | |
| 1287 : | edgomez | 1060 | for(q=1; q<=max_Q; ++q) { |
| 1288 : | emms(); | ||
| 1289 : | quant4_intra( Dst, Src, q, q); | ||
| 1290 : | emms(); | ||
| 1291 : | for(s=0, i=0; i<64; ++i) s+=((uint16_t)Dst[i])^i; | ||
| 1292 : | Crcs_Intra[n][q] = s; | ||
| 1293 : | } | ||
| 1294 : | } | ||
| 1295 : | Isibaar | 262 | |
| 1296 : | edgomez | 1060 | for(q=1; q<=max_Q; ++q) |
| 1297 : | for(i=0; i<n-1; ++i) | ||
| 1298 : | if (Crcs_Intra[i][q]!=Crcs_Intra[i+1][q]) | ||
| 1299 : | printf( "Discrepancy Intra: qm=%d, q=%d -> %d/%d!\n", | ||
| 1300 : | qm, q, Crcs_Inter[i][q], Crcs_Inter[i+1][q]); | ||
| 1301 : | } | ||
| 1302 : | Isibaar | 262 | #endif |
| 1303 : | } | ||
| 1304 : | |||
| 1305 : | Isibaar | 225 | /********************************************************************* |
| 1306 : | * main | ||
| 1307 : | *********************************************************************/ | ||
| 1308 : | |||
| 1309 : | int main(int argc, char *argv[]) | ||
| 1310 : | { | ||
| 1311 : | edgomez | 1060 | int what = 0; |
| 1312 : | if (argc>1) what = atoi(argv[1]); | ||
| 1313 : | if (what==0 || what==1) test_dct(); | ||
| 1314 : | if (what==0 || what==2) test_mb(); | ||
| 1315 : | if (what==0 || what==3) test_sad(); | ||
| 1316 : | if (what==0 || what==4) test_transfer(); | ||
| 1317 : | if (what==0 || what==5) test_quant(); | ||
| 1318 : | if (what==0 || what==6) test_cbp(); | ||
| 1319 : | Isibaar | 225 | |
| 1320 : | edgomez | 1060 | if (what==7) { |
| 1321 : | test_IEEE1180_compliance(-256, 255, 1); | ||
| 1322 : | Isibaar | 262 | #if 0 |
| 1323 : | edgomez | 1060 | test_IEEE1180_compliance(-256, 255,-1); |
| 1324 : | test_IEEE1180_compliance( -5, 5, 1); | ||
| 1325 : | test_IEEE1180_compliance( -5, 5,-1); | ||
| 1326 : | test_IEEE1180_compliance(-300, 300, 1); | ||
| 1327 : | test_IEEE1180_compliance(-300, 300,-1); | ||
| 1328 : | Isibaar | 262 | #endif |
| 1329 : | edgomez | 1060 | } |
| 1330 : | if (what==8) test_dct_saturation(-256, 255); | ||
| 1331 : | Isibaar | 262 | |
| 1332 : | edgomez | 1060 | if (what==9) { |
| 1333 : | int width, height; | ||
| 1334 : | if (argc<5) { | ||
| 1335 : | printf("usage: %s %d [bitstream] [width] [height]\n", argv[0], what); | ||
| 1336 : | return 1; | ||
| 1337 : | } | ||
| 1338 : | width = atoi(argv[3]); | ||
| 1339 : | height = atoi(argv[4]); | ||
| 1340 : | test_dec(argv[2], width, height, (argc>5)); | ||
| 1341 : | } | ||
| 1342 : | Isibaar | 225 | |
| 1343 : | edgomez | 1060 | if (what==-1) { |
| 1344 : | test_dct_precision_diffs(); | ||
| 1345 : | test_bugs1(); | ||
| 1346 : | } | ||
| 1347 : | if (what==-2) | ||
| 1348 : | test_quant_bug(); | ||
| 1349 : | Isibaar | 262 | |
| 1350 : | edgomez | 1060 | return 0; |
| 1351 : | Isibaar | 225 | } |
| 1352 : | |||
| 1353 : | /********************************************************************* | ||
| 1354 : | * 'Reference' output (except for timing) on a PIII 1.13Ghz/linux | ||
| 1355 : | *********************************************************************/ | ||
| 1356 : | Isibaar | 262 | |
| 1357 : | edgomez | 1060 | /* as of 07/01/2002, there's a problem with mpeg4-quantization */ |
| 1358 : | Isibaar | 225 | /* |
| 1359 : | |||
| 1360 : | edgomez | 1060 | ===== test fdct/idct ===== |
| 1361 : | Isibaar | 262 | PLAINC - 3.312 usec PSNR=13.291 MSE=3.000 |
| 1362 : | MMX - 0.591 usec PSNR=13.291 MSE=3.000 | ||
| 1363 : | MMXEXT - 0.577 usec PSNR=13.291 MSE=3.000 | ||
| 1364 : | SSE2 - 0.588 usec PSNR=13.291 MSE=3.000 | ||
| 1365 : | Isibaar | 225 | 3DNOW - skipped... |
| 1366 : | 3DNOWE - skipped... | ||
| 1367 : | |||
| 1368 : | edgomez | 1060 | === test block motion === |
| 1369 : | Isibaar | 262 | PLAINC - interp- h-round0 0.911 usec iCrc=8107 |
| 1370 : | PLAINC - round1 0.863 usec iCrc=8100 | ||
| 1371 : | PLAINC - interp- v-round0 0.860 usec iCrc=8108 | ||
| 1372 : | PLAINC - round1 0.857 usec iCrc=8105 | ||
| 1373 : | PLAINC - interp-hv-round0 2.103 usec iCrc=8112 | ||
| 1374 : | PLAINC - round1 2.050 usec iCrc=8103 | ||
| 1375 : | edgomez | 1060 | --- |
| 1376 : | Isibaar | 225 | MMX - interp- h-round0 0.105 usec iCrc=8107 |
| 1377 : | Isibaar | 262 | MMX - round1 0.106 usec iCrc=8100 |
| 1378 : | Isibaar | 225 | MMX - interp- v-round0 0.106 usec iCrc=8108 |
| 1379 : | Isibaar | 262 | MMX - round1 0.106 usec iCrc=8105 |
| 1380 : | Isibaar | 225 | MMX - interp-hv-round0 0.145 usec iCrc=8112 |
| 1381 : | MMX - round1 0.145 usec iCrc=8103 | ||
| 1382 : | edgomez | 1060 | --- |
| 1383 : | Isibaar | 262 | MMXEXT - interp- h-round0 0.028 usec iCrc=8107 |
| 1384 : | Isibaar | 225 | MMXEXT - round1 0.041 usec iCrc=8100 |
| 1385 : | MMXEXT - interp- v-round0 0.027 usec iCrc=8108 | ||
| 1386 : | Isibaar | 262 | MMXEXT - round1 0.041 usec iCrc=8105 |
| 1387 : | MMXEXT - interp-hv-round0 0.066 usec iCrc=8112 | ||
| 1388 : | MMXEXT - round1 0.065 usec iCrc=8103 | ||
| 1389 : | edgomez | 1060 | --- |
| 1390 : | Isibaar | 262 | SSE2 - interp- h-round0 0.109 usec iCrc=8107 |
| 1391 : | Isibaar | 225 | SSE2 - round1 0.105 usec iCrc=8100 |
| 1392 : | SSE2 - interp- v-round0 0.106 usec iCrc=8108 | ||
| 1393 : | Isibaar | 262 | SSE2 - round1 0.109 usec iCrc=8105 |
| 1394 : | Isibaar | 225 | SSE2 - interp-hv-round0 0.145 usec iCrc=8112 |
| 1395 : | SSE2 - round1 0.145 usec iCrc=8103 | ||
| 1396 : | edgomez | 1060 | --- |
| 1397 : | Isibaar | 225 | 3DNOW - skipped... |
| 1398 : | 3DNOWE - skipped... | ||
| 1399 : | |||
| 1400 : | edgomez | 1060 | ====== test SAD ====== |
| 1401 : | Isibaar | 262 | PLAINC - sad8 0.251 usec sad=3776 |
| 1402 : | PLAINC - sad16 1.601 usec sad=27214 | ||
| 1403 : | PLAINC - sad16bi 2.371 usec sad=26274 | ||
| 1404 : | PLAINC - dev16 1.564 usec sad=3344 | ||
| 1405 : | edgomez | 1060 | --- |
| 1406 : | Isibaar | 225 | MMX - sad8 0.057 usec sad=3776 |
| 1407 : | Isibaar | 262 | MMX - sad16 0.182 usec sad=27214 |
| 1408 : | MMX - sad16bi 2.462 usec sad=26274 | ||
| 1409 : | MMX - dev16 0.311 usec sad=3344 | ||
| 1410 : | edgomez | 1060 | --- |
| 1411 : | Isibaar | 225 | MMXEXT - sad8 0.036 usec sad=3776 |
| 1412 : | Isibaar | 262 | MMXEXT - sad16 0.109 usec sad=27214 |
| 1413 : | MMXEXT - sad16bi 0.143 usec sad=26274 | ||
| 1414 : | MMXEXT - dev16 0.192 usec sad=3344 | ||
| 1415 : | edgomez | 1060 | --- |
| 1416 : | Isibaar | 225 | SSE2 - sad8 0.057 usec sad=3776 |
| 1417 : | Isibaar | 262 | SSE2 - sad16 0.179 usec sad=27214 |
| 1418 : | SSE2 - sad16bi 2.456 usec sad=26274 | ||
| 1419 : | SSE2 - dev16 0.321 usec sad=3344 | ||
| 1420 : | edgomez | 1060 | --- |
| 1421 : | Isibaar | 225 | 3DNOW - skipped... |
| 1422 : | 3DNOWE - skipped... | ||
| 1423 : | |||
| 1424 : | edgomez | 1060 | === test transfer === |
| 1425 : | Isibaar | 262 | PLAINC - 8to16 0.151 usec crc=28288 |
| 1426 : | PLAINC - 16to8 1.113 usec crc=28288 | ||
| 1427 : | PLAINC - 8to8 0.043 usec crc=20352 | ||
| 1428 : | PLAINC - 16to8add 1.069 usec crc=25536 | ||
| 1429 : | PLAINC - 8to16sub 0.631 usec crc1=28064 crc2=16256 | ||
| 1430 : | PLAINC - 8to16sub2 0.597 usec crc=20384 | ||
| 1431 : | edgomez | 1060 | --- |
| 1432 : | Isibaar | 262 | MMX - 8to16 0.032 usec crc=28288 |
| 1433 : | MMX - 16to8 0.024 usec crc=28288 | ||
| 1434 : | MMX - 8to8 0.020 usec crc=20352 | ||
| 1435 : | MMX - 16to8add 0.043 usec crc=25536 | ||
| 1436 : | MMX - 8to16sub 0.066 usec crc1=28064 crc2=16256 | ||
| 1437 : | MMX - 8to16sub2 0.111 usec crc=20384 | ||
| 1438 : | edgomez | 1060 | --- |
| 1439 : | Isibaar | 225 | |
| 1440 : | edgomez | 1060 | ===== test quant ===== |
| 1441 : | Isibaar | 262 | PLAINC - quant4_intra 74.248 usec crc=29809 |
| 1442 : | PLAINC - quant4_inter 70.850 usec crc=12574 | ||
| 1443 : | PLAINC - dequant4_intra 40.628 usec crc=24052 | ||
| 1444 : | PLAINC - dequant4_inter 45.691 usec crc=63847 | ||
| 1445 : | PLAINC - quant_intra 43.357 usec crc=25662 | ||
| 1446 : | PLAINC - quant_inter 33.410 usec crc=23972 | ||
| 1447 : | PLAINC - dequant_intra 36.384 usec crc=49900 | ||
| 1448 : | PLAINC - dequant_inter 48.930 usec crc=48899 | ||
| 1449 : | edgomez | 1060 | --- |
| 1450 : | Isibaar | 262 | MMX - quant4_intra 7.445 usec crc=3459 |
| 1451 : | *** CRC ERROR! *** | ||
| 1452 : | MMX - quant4_inter 5.384 usec crc=51072 | ||
| 1453 : | *** CRC ERROR! *** | ||
| 1454 : | MMX - dequant4_intra 5.515 usec crc=24052 | ||
| 1455 : | MMX - dequant4_inter 7.745 usec crc=63847 | ||
| 1456 : | MMX - quant_intra 4.661 usec crc=25662 | ||
| 1457 : | MMX - quant_inter 4.406 usec crc=23972 | ||
| 1458 : | MMX - dequant_intra 4.928 usec crc=49900 | ||
| 1459 : | MMX - dequant_inter 4.532 usec crc=48899 | ||
| 1460 : | edgomez | 1060 | --- |
| 1461 : | Isibaar | 225 | |
| 1462 : | edgomez | 1060 | ===== test cbp ===== |
| 1463 : | Isibaar | 262 | PLAINC - calc_cbp#1 0.371 usec cbp=0x15 |
| 1464 : | PLAINC - calc_cbp#2 0.432 usec cbp=0x38 | ||
| 1465 : | PLAINC - calc_cbp#3 0.339 usec cbp=0xf | ||
| 1466 : | PLAINC - calc_cbp#4 0.506 usec cbp=0x5 | ||
| 1467 : | edgomez | 1060 | --- |
| 1468 : | Isibaar | 225 | MMX - calc_cbp#1 0.136 usec cbp=0x15 |
| 1469 : | Isibaar | 262 | MMX - calc_cbp#2 0.134 usec cbp=0x38 |
| 1470 : | MMX - calc_cbp#3 0.138 usec cbp=0xf | ||
| 1471 : | Isibaar | 225 | MMX - calc_cbp#4 0.135 usec cbp=0x5 |
| 1472 : | edgomez | 1060 | --- |
| 1473 : | Isibaar | 262 | SSE2 - calc_cbp#1 0.136 usec cbp=0x15 |
| 1474 : | SSE2 - calc_cbp#2 0.133 usec cbp=0x38 | ||
| 1475 : | SSE2 - calc_cbp#3 0.133 usec cbp=0xf | ||
| 1476 : | SSE2 - calc_cbp#4 0.141 usec cbp=0x5 | ||
| 1477 : | edgomez | 1060 | --- |
| 1478 : | Isibaar | 262 | |
| 1479 : | Isibaar | 225 | */ |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |