19 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
20 |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
21 |
* |
* |
22 |
* $Id: xvid_bench.c,v 1.16 2005-04-18 08:31:42 Skal Exp $ |
* $Id: xvid_bench.c,v 1.19 2005-05-23 09:29:43 Skal Exp $ |
23 |
* |
* |
24 |
****************************************************************************/ |
****************************************************************************/ |
25 |
|
|
65 |
#define M_PI 3.14159265358979323846 |
#define M_PI 3.14159265358979323846 |
66 |
#endif |
#endif |
67 |
|
|
68 |
const int speed_ref = 100; /* on slow machines, decrease this value */ |
int speed_ref = 100; /* on slow machines, decrease this value */ |
69 |
|
int verbose = 0; |
70 |
|
unsigned int cpu_mask; |
71 |
|
|
72 |
/********************************************************************* |
/********************************************************************* |
73 |
* misc |
* misc |
259 |
}; |
}; |
260 |
|
|
261 |
uint32_t |
uint32_t |
262 |
calc_crc(uint8_t *mem, int len, uint32_t initial) |
calc_crc(uint8_t *mem, int len, uint32_t crc) |
263 |
{ |
{ |
|
|
|
|
register unsigned int crc; |
|
|
|
|
|
crc = initial; |
|
|
|
|
264 |
while( len >= 8) { |
while( len >= 8) { |
265 |
DO8(mem, crc); |
DO8(mem, crc); |
266 |
len -= 8; |
len -= 8; |
271 |
len--; |
len--; |
272 |
} |
} |
273 |
|
|
274 |
return(crc); |
return crc; |
|
|
|
275 |
} |
} |
276 |
|
|
277 |
/********************************************************************* |
/********************************************************************* |
1314 |
* measure raw decoding speed |
* measure raw decoding speed |
1315 |
*********************************************************************/ |
*********************************************************************/ |
1316 |
|
|
1317 |
void test_dec(const char *name, int width, int height, int with_chksum) |
void test_dec(const char *name, int width, int height, int ref_chksum) |
1318 |
{ |
{ |
1319 |
FILE *f = 0; |
FILE *f = 0; |
1320 |
void *dechandle = 0; |
void *dechandle = 0; |
1325 |
double t = 0.; |
double t = 0.; |
1326 |
int nb = 0; |
int nb = 0; |
1327 |
uint8_t *buf = 0; |
uint8_t *buf = 0; |
1328 |
uint8_t *rgb_out = 0; |
uint8_t *yuv_out = 0; |
1329 |
int buf_size, pos; |
int buf_size, pos; |
1330 |
uint32_t chksum = 0; |
uint32_t chksum = 0; |
1331 |
|
int bps = (width+31) & ~31; |
1332 |
|
|
1333 |
memset(&xinit, 0, sizeof(xinit)); |
memset(&xinit, 0, sizeof(xinit)); |
1334 |
xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; |
xinit.cpu_flags = cpu_mask; |
1335 |
xinit.version = XVID_VERSION; |
xinit.version = XVID_VERSION; |
1336 |
xvid_global(NULL, 0, &xinit, NULL); |
xvid_global(NULL, 0, &xinit, NULL); |
1337 |
|
|
1341 |
xparam.version = XVID_VERSION; |
xparam.version = XVID_VERSION; |
1342 |
xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
1343 |
if (xerr==XVID_ERR_FAIL) { |
if (xerr==XVID_ERR_FAIL) { |
1344 |
printf("can't init decoder (err=%d)\n", xerr); |
printf("ERROR: can't init decoder (err=%d)\n", xerr); |
1345 |
return; |
return; |
1346 |
} |
} |
1347 |
dechandle = xparam.handle; |
dechandle = xparam.handle; |
1349 |
|
|
1350 |
f = fopen(name, "rb"); |
f = fopen(name, "rb"); |
1351 |
if (f==0) { |
if (f==0) { |
1352 |
printf( "can't open file '%s'\n", name); |
printf( "ERROR: can't open file '%s'\n", name); |
1353 |
return; |
return; |
1354 |
} |
} |
1355 |
fseek(f, 0, SEEK_END); |
fseek(f, 0, SEEK_END); |
1356 |
buf_size = ftell(f); |
buf_size = ftell(f); |
1357 |
fseek(f, 0, SEEK_SET); |
fseek(f, 0, SEEK_SET); |
1358 |
if (buf_size<=0) { |
if (buf_size<=0) { |
1359 |
printf("error while stating file\n"); |
printf("ERROR: error while stating file\n"); |
1360 |
goto End; |
goto End; |
1361 |
} |
} |
|
else printf( "Input size: %d\n", buf_size); |
|
1362 |
|
|
1363 |
buf = malloc(buf_size); /* should be enuf' */ |
buf = malloc(buf_size); |
1364 |
rgb_out = calloc(4, width*height); /* <-room for _RGB24 */ |
yuv_out = calloc(1, bps*height*3/2 + 15); |
1365 |
if (buf==0 || rgb_out==0) { |
if (buf==0 || yuv_out==0) { |
1366 |
printf( "malloc failed!\n" ); |
printf( "ERROR: malloc failed!\n" ); |
1367 |
goto End; |
goto End; |
1368 |
} |
} |
1369 |
|
|
1370 |
if (fread(buf, buf_size, 1, f)!=1) { |
if (fread(buf, buf_size, 1, f)!=1) { |
1371 |
printf( "file-read failed\n" ); |
printf( "ERROR: file-read failed\n" ); |
1372 |
goto End; |
goto End; |
1373 |
} |
} |
1374 |
|
|
1376 |
pos = 0; |
pos = 0; |
1377 |
t = -gettime_usec(); |
t = -gettime_usec(); |
1378 |
while(1) { |
while(1) { |
1379 |
|
int y; |
1380 |
|
|
1381 |
memset(&xframe, 0, sizeof(xframe)); |
memset(&xframe, 0, sizeof(xframe)); |
1382 |
xframe.version = XVID_VERSION; |
xframe.version = XVID_VERSION; |
1383 |
xframe.bitstream = buf + pos; |
xframe.bitstream = buf + pos; |
1384 |
xframe.length = buf_size - pos; |
xframe.length = buf_size - pos; |
1385 |
xframe.output.plane[0] = rgb_out; |
xframe.output.plane[0] = (uint8_t*)(((size_t)yuv_out + 15) & ~15); |
1386 |
xframe.output.stride[0] = width; |
xframe.output.plane[1] = xframe.output.plane[0] + bps*height; |
1387 |
xframe.output.csp = XVID_CSP_BGR; |
xframe.output.plane[2] = xframe.output.plane[1] + bps/2; |
1388 |
|
xframe.output.stride[0] = bps; |
1389 |
|
xframe.output.stride[1] = bps; |
1390 |
|
xframe.output.stride[2] = bps; |
1391 |
|
xframe.output.csp = XVID_CSP_I420; |
1392 |
xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); |
xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); |
1393 |
|
if (xerr<0) { |
1394 |
|
printf("ERROR: decoding failed for frame #%d (err=%d)!\n", nb, xerr); |
1395 |
|
break; |
1396 |
|
} |
1397 |
|
else if (xerr==0) |
1398 |
|
break; |
1399 |
|
else if (verbose>0) printf("#%d %d\n", nb, xerr ); |
1400 |
|
|
1401 |
|
pos += xerr; |
1402 |
nb++; |
nb++; |
1403 |
pos += xframe.length; |
|
1404 |
if (with_chksum) { |
for(y=0; y<height/2; ++y) { |
1405 |
int k = width*height; |
chksum = calc_crc(xframe.output.plane[0] + (2*y+0)*bps, width, chksum); |
1406 |
uint32_t *ptr = (uint32_t *)rgb_out; |
chksum = calc_crc(xframe.output.plane[0] + (2*y+1)*bps, width, chksum); |
1407 |
while(k-->0) chksum += *ptr++; |
chksum = calc_crc(xframe.output.plane[1] + y*bps, width/2, chksum); |
1408 |
|
chksum = calc_crc(xframe.output.plane[2] + y*bps, width/2, chksum); |
1409 |
} |
} |
1410 |
if (pos==buf_size) |
if (pos==buf_size) |
1411 |
break; |
break; |
|
if (xerr==XVID_ERR_FAIL) { |
|
|
printf("decoding failed for frame #%d (err=%d)!\n", nb, xerr); |
|
|
break; |
|
|
} |
|
1412 |
} |
} |
1413 |
t += gettime_usec(); |
t += gettime_usec(); |
1414 |
|
if (ref_chksum==0) { |
1415 |
if (t>0.) |
if (t>0.) |
1416 |
printf( "%d frames decoded in %.3f s -> %.1f FPS\n", nb, t*1.e-6f, (float)(nb*1.e6f/t) ); |
printf( "%d frames decoded in %.3f s -> %.1f FPS Checksum:0x%.8x\n", nb, t*1.e-6f, (float)(nb*1.e6f/t), chksum ); |
1417 |
if (with_chksum) |
} |
1418 |
printf("checksum: 0x%.8x\n", chksum); |
else { |
1419 |
|
printf("FPS:%.1f Checksum: 0x%.8x Expected:0x%.8x | %s\n", |
1420 |
|
t>0. ? (float)(nb*1.e6f/t) : 0.f, chksum, ref_chksum, (chksum==ref_chksum) ? "OK" : "ERROR"); |
1421 |
|
} |
1422 |
|
|
1423 |
End: |
End: |
1424 |
if (rgb_out!=0) free(rgb_out); |
if (yuv_out!=0) free(yuv_out); |
1425 |
if (buf!=0) free(buf); |
if (buf!=0) free(buf); |
1426 |
if (dechandle!=0) { |
if (dechandle!=0) { |
1427 |
xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); |
xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); |
1428 |
if (xerr==XVID_ERR_FAIL) |
if (xerr==XVID_ERR_FAIL) |
1429 |
printf("destroy-decoder failed (err=%d)!\n", xerr); |
printf("ERROR: destroy-decoder failed (err=%d)!\n", xerr); |
1430 |
} |
} |
1431 |
if (f!=0) fclose(f); |
if (f!=0) fclose(f); |
1432 |
} |
} |
1588 |
} |
} |
1589 |
#endif |
#endif |
1590 |
} |
} |
1591 |
|
/*********************************************************************/ |
1592 |
|
|
1593 |
|
static uint32_t __inline log2bin_v1(uint32_t value) |
1594 |
|
{ |
1595 |
|
int n = 0; |
1596 |
|
while (value) { |
1597 |
|
value >>= 1; |
1598 |
|
n++; |
1599 |
|
} |
1600 |
|
return n; |
1601 |
|
} |
1602 |
|
|
1603 |
|
static const uint8_t log2_tab_16[16] = { 0, 1, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4 }; |
1604 |
|
|
1605 |
|
static uint32_t __inline log2bin_v2(uint32_t value) |
1606 |
|
{ |
1607 |
|
int n = 0; |
1608 |
|
if (value & 0xffff0000) { |
1609 |
|
value >>= 16; |
1610 |
|
n += 16; |
1611 |
|
} |
1612 |
|
if (value & 0xff00) { |
1613 |
|
value >>= 8; |
1614 |
|
n += 8; |
1615 |
|
} |
1616 |
|
if (value & 0xf0) { |
1617 |
|
value >>= 4; |
1618 |
|
n += 4; |
1619 |
|
} |
1620 |
|
return n + log2_tab_16[value]; |
1621 |
|
} |
1622 |
|
|
1623 |
|
void test_log2bin() |
1624 |
|
{ |
1625 |
|
const int nb_tests = 3000*speed_ref; |
1626 |
|
int n, crc1=0, crc2=0; |
1627 |
|
uint32_t s, s0; |
1628 |
|
double t1, t2; |
1629 |
|
|
1630 |
|
t1 = gettime_usec(); |
1631 |
|
s0 = (int)(t1*31.241); |
1632 |
|
for(s=s0, n=0; n<nb_tests; ++n, s=(s*12363+31)&0x7fffffff) |
1633 |
|
crc1 += log2bin_v1(s); |
1634 |
|
t1 = (gettime_usec()-t1) / nb_tests; |
1635 |
|
|
1636 |
|
t2 = gettime_usec(); |
1637 |
|
for(s=s0, n=0; n<nb_tests; ++n, s=(s*12363+31)&0x7fffffff) |
1638 |
|
crc2 += log2bin_v2(s); |
1639 |
|
t2 = (gettime_usec() - t2) / nb_tests; |
1640 |
|
|
1641 |
|
printf( "log2bin_v1: %.3f sec crc=%d\n", t1, crc1 ); |
1642 |
|
printf( "log2bin_v2: %.3f sec crc=%d\n", t2, crc2 ); |
1643 |
|
if (crc1!=crc2) printf( " CRC ERROR !\n" ); |
1644 |
|
} |
1645 |
|
|
1646 |
|
/*********************************************************************/ |
1647 |
|
|
1648 |
|
static void __inline old_gcd(int *num, int *den) |
1649 |
|
{ |
1650 |
|
int i = *num; |
1651 |
|
while (i > 1) { |
1652 |
|
if (*num % i == 0 && *den % i == 0) { |
1653 |
|
*num /= i; |
1654 |
|
*den /= i; |
1655 |
|
i = *num; |
1656 |
|
continue; |
1657 |
|
} |
1658 |
|
i--; |
1659 |
|
} |
1660 |
|
} |
1661 |
|
|
1662 |
|
static uint32_t gcd(int num, int den) |
1663 |
|
{ |
1664 |
|
int tmp; |
1665 |
|
while( (tmp=num%den) ) { num = den; den = tmp; } |
1666 |
|
return den; |
1667 |
|
} |
1668 |
|
static void __inline new_gcd(int *num, int *den) |
1669 |
|
{ |
1670 |
|
const int div = gcd(*num, *den); |
1671 |
|
if (num) { |
1672 |
|
*num /= div; |
1673 |
|
*den /= div; |
1674 |
|
} |
1675 |
|
} |
1676 |
|
|
1677 |
|
void test_gcd() |
1678 |
|
{ |
1679 |
|
const int nb_tests = 10*speed_ref; |
1680 |
|
int i; |
1681 |
|
uint32_t crc1=0, crc2=0; |
1682 |
|
uint32_t n0, n, d0, d; |
1683 |
|
double t1, t2; |
1684 |
|
|
1685 |
|
t1 = gettime_usec(); |
1686 |
|
n0 = 0xfffff & (int)(t1*31.241); |
1687 |
|
d0 = 0xfffff & (int)( ((n0*4123)%17) | 1 ); |
1688 |
|
for(n=n0, d=d0, i=0; i<nb_tests; ++i) { |
1689 |
|
old_gcd(&n, &d); |
1690 |
|
crc1 = (((crc1>>4)^d) + ((crc1<<2)^n) ) & 0xffffff; |
1691 |
|
n = d; |
1692 |
|
d = (d*12363+31) & 0xffff; |
1693 |
|
d |= !d; |
1694 |
|
} |
1695 |
|
t1 = (gettime_usec()-t1) / nb_tests; |
1696 |
|
|
1697 |
|
t2 = gettime_usec(); |
1698 |
|
for(n=n0, d=d0, i=0; i<nb_tests; ++i) { |
1699 |
|
new_gcd(&n, &d); |
1700 |
|
crc2 = (((crc2>>4)^d) + ((crc2<<2)^n) ) & 0xffffff; |
1701 |
|
n = d; |
1702 |
|
d = (d*12363+31) & 0xffff; |
1703 |
|
d |= !d; |
1704 |
|
} |
1705 |
|
t2 = (gettime_usec() - t2) / nb_tests; |
1706 |
|
|
1707 |
|
printf( "old_gcd: %.3f sec crc=%d\n", t1, crc1 ); |
1708 |
|
printf( "new_gcd: %.3f sec crc=%d\n", t2, crc2 ); |
1709 |
|
if (crc1!=crc2) printf( " CRC ERROR !\n" ); |
1710 |
|
} |
1711 |
|
|
1712 |
/********************************************************************* |
/********************************************************************* |
1713 |
* main |
* main |
1714 |
*********************************************************************/ |
*********************************************************************/ |
1715 |
|
|
1716 |
int main(int argc, char *argv[]) |
static void arg_missing(const char *opt) |
1717 |
{ |
{ |
1718 |
int what = 0; |
printf( "missing argument after option '%s'\n", opt); |
1719 |
if (argc>1) what = atoi(argv[1]); |
exit(-1); |
1720 |
|
} |
1721 |
|
|
1722 |
|
int main(int argc, const char *argv[]) |
1723 |
|
{ |
1724 |
|
int c, what = 0; |
1725 |
|
int width, height; |
1726 |
|
uint32_t chksum = 0; |
1727 |
|
const char * test_bitstream = 0; |
1728 |
|
|
1729 |
|
cpu_mask = 0; // default => will use autodectect |
1730 |
|
for(c=1; c<argc; ++c) |
1731 |
|
{ |
1732 |
|
if (!strcmp(argv[c], "-v")) verbose++; |
1733 |
|
else if (!strcmp(argv[c], "-c")) cpu_mask = 0 /* PLAIN_C */ | XVID_CPU_FORCE; |
1734 |
|
else if (!strcmp(argv[c], "-mmx")) cpu_mask = XVID_CPU_MMX | XVID_CPU_FORCE; |
1735 |
|
else if (!strcmp(argv[c], "-mmxext")) cpu_mask = XVID_CPU_MMXEXT | XVID_CPU_MMX | XVID_CPU_FORCE; |
1736 |
|
else if (!strcmp(argv[c], "-sse2")) cpu_mask = XVID_CPU_SSE2 | XVID_CPU_MMX | XVID_CPU_FORCE; |
1737 |
|
else if (!strcmp(argv[c], "-3dnow")) cpu_mask = XVID_CPU_3DNOW | XVID_CPU_FORCE; |
1738 |
|
else if (!strcmp(argv[c], "-3dnowe")) cpu_mask = XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT | XVID_CPU_FORCE; |
1739 |
|
else if (!strcmp(argv[c], "-altivec")) cpu_mask = XVID_CPU_ALTIVEC | XVID_CPU_FORCE; |
1740 |
|
else if (!strcmp(argv[c], "-spd")) { |
1741 |
|
if (++c==argc) arg_missing( argv[argc-1] ); |
1742 |
|
speed_ref = atoi(argv[c]); |
1743 |
|
} |
1744 |
|
else if (argv[c][0]!='-') { |
1745 |
|
what = atoi(argv[c]); |
1746 |
|
if (what==9) { |
1747 |
|
if (c+4>argc) { |
1748 |
|
printf("usage: %s %d bitstream width height (checksum)\n", argv[0], what); |
1749 |
|
exit(-1); |
1750 |
|
} |
1751 |
|
test_bitstream = argv[++c]; |
1752 |
|
width = atoi(argv[++c]); |
1753 |
|
height = atoi(argv[++c]); |
1754 |
|
if (c+1<argc && argv[c+1][0]!='-') { |
1755 |
|
if (sscanf(argv[c+1], "0x%x", &chksum)!=1) { |
1756 |
|
printf( "can't read checksum value.\n" ); |
1757 |
|
exit(-1); |
1758 |
|
} |
1759 |
|
else c++; |
1760 |
|
} |
1761 |
|
// printf( "[%s] %dx%d (0x%.8x)\n", test_bitstream, width, height, chksum); |
1762 |
|
} |
1763 |
|
} |
1764 |
|
else { |
1765 |
|
printf( "unrecognized option '%s'\n", argv[c]); |
1766 |
|
exit(-1); |
1767 |
|
} |
1768 |
|
} |
1769 |
|
|
1770 |
|
|
1771 |
if (what==0 || what==1) test_dct(); |
if (what==0 || what==1) test_dct(); |
1772 |
if (what==0 || what==2) test_mb(); |
if (what==0 || what==2) test_mb(); |
1773 |
if (what==0 || what==3) test_sad(); |
if (what==0 || what==3) test_sad(); |
1775 |
if (what==0 || what==5) test_quant(); |
if (what==0 || what==5) test_quant(); |
1776 |
if (what==0 || what==6) test_cbp(); |
if (what==0 || what==6) test_cbp(); |
1777 |
if (what==0 || what==10) test_sse(); |
if (what==0 || what==10) test_sse(); |
1778 |
|
if (what==0 || what==11) test_log2bin(); |
1779 |
|
if (what==0 || what==12) test_gcd(); |
1780 |
|
|
1781 |
|
|
1782 |
if (what==7) { |
if (what==7) { |
1783 |
test_IEEE1180_compliance(-256, 255, 1); |
test_IEEE1180_compliance(-256, 255, 1); |
1789 |
} |
} |
1790 |
if (what==8) test_dct_saturation(-256, 255); |
if (what==8) test_dct_saturation(-256, 255); |
1791 |
|
|
1792 |
if (what==9) { |
if (test_bitstream) |
1793 |
int width, height; |
test_dec(test_bitstream, width, height, chksum); |
|
if (argc<5) { |
|
|
printf("usage: %s %d [bitstream] [width] [height]\n", argv[0], what); |
|
|
return 1; |
|
|
} |
|
|
width = atoi(argv[3]); |
|
|
height = atoi(argv[4]); |
|
|
test_dec(argv[2], width, height, (argc>5)); |
|
|
} |
|
|
|
|
1794 |
if (what==-1) { |
if (what==-1) { |
1795 |
test_dct_precision_diffs(); |
test_dct_precision_diffs(); |
1796 |
test_bugs1(); |
test_bugs1(); |
1818 |
return 0; |
return 0; |
1819 |
} |
} |
1820 |
|
|
1821 |
/********************************************************************* |
/*********************************************************************/ |
|
* 'Reference' output (except for timing) on an Athlon XP 2200+ |
|
|
*********************************************************************/ |
|
|
|
|
|
/* as of 2002-01-07, there's a problem with MMX mpeg4-quantization */ |
|
|
/* as of 2003-11-30, the problem is still here */ |
|
|
|
|
|
/********************************************************************* |
|
|
|
|
|
|
|
|
===== test fdct/idct ===== |
|
|
PLAINC - 2.867 usec PSNR=13.291 MSE=3.000 |
|
|
MMX - -0.211 usec PSNR=9.611 MSE=7.000 |
|
|
MMXEXT - -0.256 usec PSNR=9.611 MSE=7.000 |
|
|
3DNOW - 2.855 usec PSNR=13.291 MSE=3.000 |
|
|
3DNOWE - 1.429 usec PSNR=13.291 MSE=3.000 |
|
|
|
|
|
=== test block motion === |
|
|
PLAINC - interp- h-round0 0.538 usec crc32=0x115381ba |
|
|
PLAINC - round1 0.527 usec crc32=0x2b1f528f |
|
|
PLAINC - interp- v-round0 0.554 usec crc32=0x423cdcc7 |
|
|
PLAINC - round1 0.551 usec crc32=0x42202efe |
|
|
PLAINC - interp-hv-round0 1.041 usec crc32=0xd198d387 |
|
|
PLAINC - round1 1.038 usec crc32=0x9ecfd921 |
|
|
--- |
|
|
MMX - interp- h-round0 0.051 usec crc32=0x115381ba |
|
|
MMX - round1 0.053 usec crc32=0x2b1f528f |
|
|
MMX - interp- v-round0 0.048 usec crc32=0x423cdcc7 |
|
|
MMX - round1 0.048 usec crc32=0x42202efe |
|
|
MMX - interp-hv-round0 0.074 usec crc32=0xd198d387 |
|
|
MMX - round1 0.073 usec crc32=0x9ecfd921 |
|
|
--- |
|
|
MMXEXT - interp- h-round0 0.020 usec crc32=0x115381ba |
|
|
MMXEXT - round1 0.025 usec crc32=0x2b1f528f |
|
|
MMXEXT - interp- v-round0 0.016 usec crc32=0x423cdcc7 |
|
|
MMXEXT - round1 0.024 usec crc32=0x42202efe |
|
|
MMXEXT - interp-hv-round0 0.037 usec crc32=0xd198d387 |
|
|
MMXEXT - round1 0.037 usec crc32=0x9ecfd921 |
|
|
--- |
|
|
3DNOW - interp- h-round0 0.020 usec crc32=0x115381ba |
|
|
3DNOW - round1 0.029 usec crc32=0x2b1f528f |
|
|
3DNOW - interp- v-round0 0.016 usec crc32=0x423cdcc7 |
|
|
3DNOW - round1 0.024 usec crc32=0x42202efe |
|
|
3DNOW - interp-hv-round0 0.038 usec crc32=0xd198d387 |
|
|
3DNOW - round1 0.039 usec crc32=0x9ecfd921 |
|
|
--- |
|
|
3DNOWE - interp- h-round0 0.020 usec crc32=0x115381ba |
|
|
3DNOWE - round1 0.024 usec crc32=0x2b1f528f |
|
|
3DNOWE - interp- v-round0 0.016 usec crc32=0x423cdcc7 |
|
|
3DNOWE - round1 0.021 usec crc32=0x42202efe |
|
|
3DNOWE - interp-hv-round0 0.037 usec crc32=0xd198d387 |
|
|
3DNOWE - round1 0.036 usec crc32=0x9ecfd921 |
|
|
--- |
|
|
|
|
|
====== test SAD ====== |
|
|
PLAINC - sad8 0.505 usec sad=3776 |
|
|
PLAINC - sad16 1.941 usec sad=27214 |
|
|
PLAINC - sad16bi 4.925 usec sad=26274 |
|
|
PLAINC - dev16 4.254 usec sad=3344 |
|
|
--- |
|
|
MMX - sad8 0.036 usec sad=3776 |
|
|
MMX - sad16 0.107 usec sad=27214 |
|
|
MMX - sad16bi 0.259 usec sad=26274 |
|
|
MMX - dev16 0.187 usec sad=3344 |
|
|
--- |
|
|
MMXEXT - sad8 0.016 usec sad=3776 |
|
|
MMXEXT - sad16 0.050 usec sad=27214 |
|
|
MMXEXT - sad16bi 0.060 usec sad=26274 |
|
|
MMXEXT - dev16 0.086 usec sad=3344 |
|
|
--- |
|
|
3DNOW - sad8 0.506 usec sad=3776 |
|
|
3DNOW - sad16 1.954 usec sad=27214 |
|
|
3DNOW - sad16bi 0.119 usec sad=26274 |
|
|
3DNOW - dev16 4.252 usec sad=3344 |
|
|
--- |
|
|
3DNOWE - sad8 0.017 usec sad=3776 |
|
|
3DNOWE - sad16 0.038 usec sad=27214 |
|
|
3DNOWE - sad16bi 0.052 usec sad=26274 |
|
|
3DNOWE - dev16 0.067 usec sad=3344 |
|
|
--- |
|
|
|
|
|
=== test transfer === |
|
|
PLAINC - 8to16 0.603 usec crc32=0x115814bb |
|
|
PLAINC - 16to8 1.077 usec crc32=0xee7ccbb4 |
|
|
PLAINC - 8to8 0.679 usec crc32=0xd37b3295 |
|
|
PLAINC - 16to8add 1.341 usec crc32=0xdd817bf4 |
|
|
PLAINC - 8to16sub 1.566 usec crc32(1)=0xa1e07163 crc32(2)=0xd86c5d23 |
|
|
PLAINC - 8to16sub2 2.206 usec crc32=0x99b6c4c7 |
|
|
--- |
|
|
MMX - 8to16 -0.025 usec crc32=0x115814bb |
|
|
MMX - 16to8 -0.049 usec crc32=0xee7ccbb4 |
|
|
MMX - 8to8 0.014 usec crc32=0xd37b3295 |
|
|
MMX - 16to8add 0.011 usec crc32=0xdd817bf4 |
|
|
MMX - 8to16sub 0.108 usec crc32(1)=0xa1e07163 crc32(2)=0xd86c5d23 |
|
|
MMX - 8to16sub2 0.164 usec crc32=0x99b6c4c7 |
|
|
--- |
|
|
MMXEXT - 8to16 -0.054 usec crc32=0x115814bb |
|
|
MMXEXT - 16to8 0.010 usec crc32=0xee7ccbb4 |
|
|
MMXEXT - 8to8 0.015 usec crc32=0xd37b3295 |
|
|
MMXEXT - 16to8add 0.008 usec crc32=0xdd817bf4 |
|
|
MMXEXT - 8to16sub 0.263 usec crc32(1)=0xa1e07163 crc32(2)=0xd86c5d23 |
|
|
MMXEXT - 8to16sub2 0.178 usec crc32=0x99b6c4c7 |
|
|
--- |
|
|
3DNOW - 8to16 0.666 usec crc32=0x115814bb |
|
|
3DNOW - 16to8 1.078 usec crc32=0xee7ccbb4 |
|
|
3DNOW - 8to8 0.665 usec crc32=0xd37b3295 |
|
|
3DNOW - 16to8add 1.365 usec crc32=0xdd817bf4 |
|
|
3DNOW - 8to16sub 1.356 usec crc32(1)=0xa1e07163 crc32(2)=0xd86c5d23 |
|
|
3DNOW - 8to16sub2 2.098 usec crc32=0x99b6c4c7 |
|
|
--- |
|
|
3DNOWE - 8to16 -0.024 usec crc32=0x115814bb |
|
|
3DNOWE - 16to8 0.010 usec crc32=0xee7ccbb4 |
|
|
3DNOWE - 8to8 0.014 usec crc32=0xd37b3295 |
|
|
3DNOWE - 16to8add 0.016 usec crc32=0xdd817bf4 |
|
|
3DNOWE - 8to16sub -0.000 usec crc32(1)=0xa1e07163 crc32(2)=0xd86c5d23 |
|
|
3DNOWE - 8to16sub2 -0.031 usec crc32=0x99b6c4c7 |
|
|
--- |
|
|
|
|
|
===== test quant ===== |
|
|
PLAINC - quant_mpeg_intra 98.631 usec crc32=0xfd6a21a4 |
|
|
PLAINC - quant_mpeg_inter 104.876 usec crc32=0xf6de7757 |
|
|
PLAINC - dequant_mpeg_intra 50.285 usec crc32=0x2def7bc7 |
|
|
PLAINC - dequant_mpeg_inter 58.316 usec crc32=0xd878c722 |
|
|
PLAINC - quant_h263_intra 33.803 usec crc32=0x2eba9d43 |
|
|
PLAINC - quant_h263_inter 45.411 usec crc32=0xbd315a7e |
|
|
PLAINC - dequant_h263_intra 39.302 usec crc32=0x9841212a |
|
|
PLAINC - dequant_h263_inter 44.124 usec crc32=0xe7df8fba |
|
|
--- |
|
|
MMX - quant_mpeg_intra 4.273 usec crc32=0xdacabdb6 | ERROR |
|
|
MMX - quant_mpeg_inter 3.576 usec crc32=0x72883ab6 | ERROR |
|
|
MMX - dequant_mpeg_intra 3.793 usec crc32=0x2def7bc7 |
|
|
MMX - dequant_mpeg_inter 4.808 usec crc32=0xd878c722 |
|
|
MMX - quant_h263_intra 2.881 usec crc32=0x2eba9d43 |
|
|
MMX - quant_h263_inter 2.550 usec crc32=0xbd315a7e |
|
|
MMX - dequant_h263_intra 2.974 usec crc32=0x9841212a |
|
|
MMX - dequant_h263_inter 2.906 usec crc32=0xe7df8fba |
|
|
--- |
|
|
MMXEXT - quant_mpeg_intra 4.221 usec crc32=0xfd6a21a4 |
|
|
MMXEXT - quant_mpeg_inter 4.339 usec crc32=0xf6de7757 |
|
|
MMXEXT - dequant_mpeg_intra 3.802 usec crc32=0x2def7bc7 |
|
|
MMXEXT - dequant_mpeg_inter 4.821 usec crc32=0xd878c722 |
|
|
MMXEXT - quant_h263_intra 2.884 usec crc32=0x2eba9d43 |
|
|
MMXEXT - quant_h263_inter 2.554 usec crc32=0xbd315a7e |
|
|
MMXEXT - dequant_h263_intra 2.728 usec crc32=0x9841212a |
|
|
MMXEXT - dequant_h263_inter 2.611 usec crc32=0xe7df8fba |
|
|
--- |
|
|
3DNOW - quant_mpeg_intra 98.512 usec crc32=0xfd6a21a4 |
|
|
3DNOW - quant_mpeg_inter 104.873 usec crc32=0xf6de7757 |
|
|
3DNOW - dequant_mpeg_intra 50.219 usec crc32=0x2def7bc7 |
|
|
3DNOW - dequant_mpeg_inter 58.254 usec crc32=0xd878c722 |
|
|
3DNOW - quant_h263_intra 33.778 usec crc32=0x2eba9d43 |
|
|
3DNOW - quant_h263_inter 41.998 usec crc32=0xbd315a7e |
|
|
3DNOW - dequant_h263_intra 39.344 usec crc32=0x9841212a |
|
|
3DNOW - dequant_h263_inter 43.607 usec crc32=0xe7df8fba |
|
|
--- |
|
|
3DNOWE - quant_mpeg_intra 98.490 usec crc32=0xfd6a21a4 |
|
|
3DNOWE - quant_mpeg_inter 104.889 usec crc32=0xf6de7757 |
|
|
3DNOWE - dequant_mpeg_intra 3.277 usec crc32=0x2def7bc7 |
|
|
3DNOWE - dequant_mpeg_inter 4.485 usec crc32=0xd878c722 |
|
|
3DNOWE - quant_h263_intra 1.882 usec crc32=0x2eba9d43 |
|
|
3DNOWE - quant_h263_inter 2.246 usec crc32=0xbd315a7e |
|
|
3DNOWE - dequant_h263_intra 3.457 usec crc32=0x9841212a |
|
|
3DNOWE - dequant_h263_inter 3.275 usec crc32=0xe7df8fba |
|
|
--- |
|
|
|
|
|
===== test cbp ===== |
|
|
PLAINC - calc_cbp#1 0.168 usec cbp=0x15 |
|
|
PLAINC - calc_cbp#2 0.168 usec cbp=0x38 |
|
|
PLAINC - calc_cbp#3 0.157 usec cbp=0x0f |
|
|
PLAINC - calc_cbp#4 0.235 usec cbp=0x05 |
|
|
--- |
|
|
MMX - calc_cbp#1 0.070 usec cbp=0x15 |
|
|
MMX - calc_cbp#2 0.062 usec cbp=0x38 |
|
|
MMX - calc_cbp#3 0.062 usec cbp=0x0f |
|
|
MMX - calc_cbp#4 0.061 usec cbp=0x05 |
|
|
--- |
|
|
MMXEXT - calc_cbp#1 0.062 usec cbp=0x15 |
|
|
MMXEXT - calc_cbp#2 0.061 usec cbp=0x38 |
|
|
MMXEXT - calc_cbp#3 0.061 usec cbp=0x0f |
|
|
MMXEXT - calc_cbp#4 0.061 usec cbp=0x05 |
|
|
--- |
|
|
3DNOW - calc_cbp#1 0.168 usec cbp=0x15 |
|
|
3DNOW - calc_cbp#2 0.168 usec cbp=0x38 |
|
|
3DNOW - calc_cbp#3 0.157 usec cbp=0x0f |
|
|
3DNOW - calc_cbp#4 0.238 usec cbp=0x05 |
|
|
--- |
|
|
3DNOWE - calc_cbp#1 0.049 usec cbp=0x15 |
|
|
3DNOWE - calc_cbp#2 0.049 usec cbp=0x38 |
|
|
3DNOWE - calc_cbp#3 0.049 usec cbp=0x0f |
|
|
3DNOWE - calc_cbp#4 0.049 usec cbp=0x05 |
|
|
--- |
|
|
|
|
|
|
|
|
NB: If a function isn't optimised for a specific set of intructions, |
|
|
a C function is used instead. So don't panic if some functions |
|
|
may appear to be slow. |
|
|
|
|
|
NB: MMX mpeg4 quantization is known to have very small errors (+/-1 magnitude) |
|
|
for 1 or 2 coefficients a block. This is mainly caused by the fact the unit |
|
|
test goes far behind the usual limits of real encoding. Please do not report |
|
|
this error to the developers |
|
|
|
|
|
*********************************************************************/ |
|