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 1.22 2005-06-15 06:07:28 Skal Exp $ |
* $Id: xvid_bench.c,v 1.27.2.1 2006-11-01 10:01:23 Isibaar Exp $ |
23 |
* |
* |
24 |
****************************************************************************/ |
****************************************************************************/ |
25 |
|
|
58 |
#include "utils/timer.h" |
#include "utils/timer.h" |
59 |
#include "quant/quant_matrix.c" |
#include "quant/quant_matrix.c" |
60 |
#include "bitstream/cbp.h" |
#include "bitstream/cbp.h" |
61 |
|
#include "bitstream/bitstream.h" |
62 |
|
|
63 |
#include <math.h> |
#include <math.h> |
64 |
|
|
519 |
(iCrc!=8107)?"| ERROR": "" ); |
(iCrc!=8107)?"| ERROR": "" ); |
520 |
#endif |
#endif |
521 |
|
|
522 |
|
/* New functions for field prediction by CK 1.10.2005 */ |
523 |
|
#pragma NEW8X4 |
524 |
|
TEST_MB(interpolate8x4_halfpel_h, 0); |
525 |
|
printf("%s - interpfield-h -round0 %.3f usec crc32=0x%08x %s\n", |
526 |
|
cpu->name, t, iCrc, |
527 |
|
(iCrc!=0x9538d6df)?"| ERROR": "" ); |
528 |
|
|
529 |
|
TEST_MB(interpolate8x4_halfpel_h, 1); |
530 |
|
printf("%s - round1 %.3f usec crc32=0x%08x %s\n", |
531 |
|
cpu->name, t, iCrc, |
532 |
|
(iCrc!=0xde5f1db4)?"| ERROR": "" ); |
533 |
|
|
534 |
|
|
535 |
|
TEST_MB(interpolate8x4_halfpel_v, 0); |
536 |
|
printf("%s - interpfield- v-round0 %.3f usec crc32=0x%08x %s\n", |
537 |
|
cpu->name, t, iCrc, |
538 |
|
(iCrc!=0xea5a69ef)?"| ERROR": "" ); |
539 |
|
|
540 |
|
TEST_MB(interpolate8x4_halfpel_v, 1); |
541 |
|
printf("%s - round1 %.3f usec crc32=0x%08x %s\n", |
542 |
|
cpu->name, t, iCrc, |
543 |
|
(iCrc!=0x4f10ec0f)?"| ERROR": "" ); |
544 |
|
|
545 |
|
|
546 |
|
TEST_MB(interpolate8x4_halfpel_hv, 0); |
547 |
|
printf("%s - interpfield-hv-round0 %.3f usec crc32=0x%08x %s\n", |
548 |
|
cpu->name, t, iCrc, |
549 |
|
(iCrc!=0xf97ee367)?"| ERROR": "" ); |
550 |
|
|
551 |
|
TEST_MB(interpolate8x4_halfpel_hv, 1); |
552 |
|
printf("%s - round1 %.3f usec crc32=0x%08x %s\n", |
553 |
|
cpu->name, t, iCrc, |
554 |
|
(iCrc!=0xb6a9f581)?"| ERROR": "" ); |
555 |
|
/* End of 8x4 functions */ |
556 |
|
|
557 |
printf( " --- \n" ); |
printf( " --- \n" ); |
558 |
} |
} |
559 |
} |
} |
661 |
cpu->name, t, s, |
cpu->name, t, s, |
662 |
(s!=0xee7ccbb4)?"| ERROR": ""); |
(s!=0xee7ccbb4)?"| ERROR": ""); |
663 |
|
|
664 |
|
/* New functions for field prediction by CK 1.10.2005 */ |
665 |
|
#pragma NEW8X4 |
666 |
|
TEST_TRANSFER(transfer8x4_copy, Dst8, Src8); |
667 |
|
printf("%s - 8to4 %.3f usec crc32=0x%08x %s\n", |
668 |
|
cpu->name, t, s, |
669 |
|
(s!=0xbb9c3db5)?"| ERROR": ""); |
670 |
|
/* End of new functions */ |
671 |
|
|
672 |
TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
673 |
printf("%s - 8to8 %.3f usec crc32=0x%08x %s\n", |
printf("%s - 8to8 %.3f usec crc32=0x%08x %s\n", |
674 |
cpu->name, t, s, |
cpu->name, t, s, |
949 |
* test non-zero AC counting |
* test non-zero AC counting |
950 |
*********************************************************************/ |
*********************************************************************/ |
951 |
|
|
952 |
#define TEST_CBP(FUNC, SRC) \ |
#define TEST_CBP(FUNC, SRC, NB) \ |
953 |
t = gettime_usec(); \ |
t = gettime_usec(); \ |
954 |
emms(); \ |
emms(); \ |
955 |
for(tst=0; tst<nb_tests; ++tst) { \ |
for(tst=0; tst<NB; ++tst) { \ |
956 |
cbp = (FUNC)((SRC)); \ |
cbp = (FUNC)((SRC)); \ |
957 |
} \ |
} \ |
958 |
emms(); \ |
emms(); \ |
961 |
void test_cbp() |
void test_cbp() |
962 |
{ |
{ |
963 |
const int nb_tests = 10000*speed_ref; |
const int nb_tests = 10000*speed_ref; |
964 |
int i; |
int i, n, m; |
965 |
CPU *cpu; |
CPU *cpu; |
966 |
DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16); |
967 |
DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16); |
968 |
DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16); |
969 |
DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16); |
970 |
|
DECLARE_ALIGNED_MATRIX(Src5, 6, 64, int16_t, 16); |
971 |
|
|
972 |
printf( "\n ===== test cbp =====\n" ); |
printf( "\n ===== test cbp =====\n" ); |
973 |
|
|
976 |
Src2[i] = (i<3*64); /* half-full */ |
Src2[i] = (i<3*64); /* half-full */ |
977 |
Src3[i] = ((i+32)>3*64); |
Src3[i] = ((i+32)>3*64); |
978 |
Src4[i] = (i==(3*64+2) || i==(5*64+9)); |
Src4[i] = (i==(3*64+2) || i==(5*64+9)); |
979 |
|
Src5[i] = ieee_rand(0,1) ? -1 : 1; /* +/- test */ |
980 |
} |
} |
981 |
|
|
982 |
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
987 |
if (!init_cpu(cpu)) |
if (!init_cpu(cpu)) |
988 |
continue; |
continue; |
989 |
|
|
990 |
TEST_CBP(calc_cbp, Src1); |
TEST_CBP(calc_cbp, Src1, nb_tests); |
991 |
printf("%s - calc_cbp#1 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#1 %.3f usec cbp=0x%02x %s\n", |
992 |
cpu->name, t, cbp, (cbp!=0x15)?"| ERROR": ""); |
cpu->name, t, cbp, (cbp!=0x15)?"| ERROR": ""); |
993 |
TEST_CBP(calc_cbp, Src2); |
TEST_CBP(calc_cbp, Src2, nb_tests); |
994 |
printf("%s - calc_cbp#2 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#2 %.3f usec cbp=0x%02x %s\n", |
995 |
cpu->name, t, cbp, (cbp!=0x38)?"| ERROR": ""); |
cpu->name, t, cbp, (cbp!=0x38)?"| ERROR": ""); |
996 |
TEST_CBP(calc_cbp, Src3); |
TEST_CBP(calc_cbp, Src3, nb_tests); |
997 |
printf("%s - calc_cbp#3 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#3 %.3f usec cbp=0x%02x %s\n", |
998 |
cpu->name, t, cbp, (cbp!=0x0f)?"| ERROR": "" ); |
cpu->name, t, cbp, (cbp!=0x0f)?"| ERROR": "" ); |
999 |
TEST_CBP(calc_cbp, Src4); |
TEST_CBP(calc_cbp, Src4, nb_tests); |
1000 |
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
1001 |
cpu->name, t, cbp, (cbp!=0x05)?"| ERROR": "" ); |
cpu->name, t, cbp, (cbp!=0x05)?"| ERROR": "" ); |
1002 |
|
TEST_CBP(calc_cbp, Src5, nb_tests); |
1003 |
|
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
1004 |
|
cpu->name, t, cbp, (cbp!=0x3f)?"| ERROR": "" ); |
1005 |
printf( " --- \n" ); |
printf( " --- \n" ); |
1006 |
} |
} |
1007 |
|
|
1008 |
|
for(cpu = cpu_list; cpu->name!=0; ++cpu) /* bench suggested by Carlo (carlo dot bramix at libero dot it) */ |
1009 |
|
{ |
1010 |
|
double t; |
1011 |
|
int tst, cbp, err; |
1012 |
|
|
1013 |
|
if (!init_cpu(cpu)) |
1014 |
|
continue; |
1015 |
|
|
1016 |
|
err = 0; |
1017 |
|
for(n=0; n<6; ++n) |
1018 |
|
{ |
1019 |
|
for(m=0; m<64; ++m) |
1020 |
|
{ |
1021 |
|
for(i=0; i<6*64; ++i) |
1022 |
|
Src1[i] = (i== (m + n*64)); |
1023 |
|
|
1024 |
|
TEST_CBP(calc_cbp, Src1, 1); |
1025 |
|
if (cbp!= (((m!=0)<<(5-n)))) |
1026 |
|
{ |
1027 |
|
printf( "%s - calc_cbp#5: ERROR at pos %d / %d!\n", cpu->name, n, m); |
1028 |
|
err = 1; |
1029 |
|
break; |
1030 |
|
} |
1031 |
|
} |
1032 |
|
} |
1033 |
|
if (!err) |
1034 |
|
printf( " %s - calc_cbp#5 : OK\n", cpu->name ); |
1035 |
|
|
1036 |
|
} |
1037 |
} |
} |
1038 |
|
|
1039 |
/********************************************************************* |
/********************************************************************* |
1496 |
xframe.bitstream = buf + pos; |
xframe.bitstream = buf + pos; |
1497 |
xframe.length = buf_size - pos; |
xframe.length = buf_size - pos; |
1498 |
xframe.output.plane[0] = (uint8_t*)(((size_t)yuv_out + 15) & ~15); |
xframe.output.plane[0] = (uint8_t*)(((size_t)yuv_out + 15) & ~15); |
1499 |
xframe.output.plane[1] = xframe.output.plane[0] + bps*height; |
xframe.output.plane[1] = (uint8_t*)xframe.output.plane[0] + bps*height; |
1500 |
xframe.output.plane[2] = xframe.output.plane[1] + bps/2; |
xframe.output.plane[2] = (uint8_t*)xframe.output.plane[1] + bps/2; |
1501 |
xframe.output.stride[0] = bps; |
xframe.output.stride[0] = bps; |
1502 |
xframe.output.stride[1] = bps; |
xframe.output.stride[1] = bps; |
1503 |
xframe.output.stride[2] = bps; |
xframe.output.stride[2] = bps; |
1515 |
nb++; |
nb++; |
1516 |
|
|
1517 |
for(y=0; y<height/2; ++y) { |
for(y=0; y<height/2; ++y) { |
1518 |
chksum = calc_crc(xframe.output.plane[0] + (2*y+0)*bps, width, chksum); |
chksum = calc_crc((uint8_t*)xframe.output.plane[0] + (2*y+0)*bps, width, chksum); |
1519 |
chksum = calc_crc(xframe.output.plane[0] + (2*y+1)*bps, width, chksum); |
chksum = calc_crc((uint8_t*)xframe.output.plane[0] + (2*y+1)*bps, width, chksum); |
1520 |
chksum = calc_crc(xframe.output.plane[1] + y*bps, width/2, chksum); |
chksum = calc_crc((uint8_t*)xframe.output.plane[1] + y*bps, width/2, chksum); |
1521 |
chksum = calc_crc(xframe.output.plane[2] + y*bps, width/2, chksum); |
chksum = calc_crc((uint8_t*)xframe.output.plane[2] + y*bps, width/2, chksum); |
1522 |
} |
} |
1523 |
if (pos==buf_size) |
if (pos==buf_size) |
1524 |
break; |
break; |
1823 |
} |
} |
1824 |
|
|
1825 |
/********************************************************************* |
/********************************************************************* |
1826 |
|
* test compiler |
1827 |
|
*********************************************************************/ |
1828 |
|
|
1829 |
|
void test_compiler() { |
1830 |
|
int nb_err = 0; |
1831 |
|
int32_t v; |
1832 |
|
if (sizeof(uint16_t)<2) { |
1833 |
|
printf( "ERROR: sizeof(uint16_t)<2 !!\n" ); |
1834 |
|
nb_err++; |
1835 |
|
} |
1836 |
|
if (sizeof(int16_t)<2) { |
1837 |
|
printf( "ERROR: sizeof(int16_t)<2 !!\n" ); |
1838 |
|
nb_err++; |
1839 |
|
} |
1840 |
|
if (sizeof(uint8_t)!=1) { |
1841 |
|
printf( "ERROR: sizeof(uint8_t)!=1 !!\n" ); |
1842 |
|
nb_err++; |
1843 |
|
} |
1844 |
|
if (sizeof(int8_t)!=1) { |
1845 |
|
printf( "ERROR: sizeof(int8_t)!=1 !!\n" ); |
1846 |
|
nb_err++; |
1847 |
|
} |
1848 |
|
if (sizeof(uint32_t)<4) { |
1849 |
|
printf( "ERROR: sizeof(uint32_t)<4 !!\n" ); |
1850 |
|
nb_err++; |
1851 |
|
} |
1852 |
|
if (sizeof(int32_t)<4) { |
1853 |
|
printf( "ERROR: sizeof(int32_t)<4 !!\n" ); |
1854 |
|
nb_err++; |
1855 |
|
} |
1856 |
|
/* yes, i know, this test is silly. But better be safe than sorry. :) */ |
1857 |
|
for(v=1000; v>=0; v--) { |
1858 |
|
if ( (v>>2) != v/4) |
1859 |
|
nb_err++; |
1860 |
|
} |
1861 |
|
for(v=-1000; v!=-1; v++) { |
1862 |
|
if ( (v>>2) != (v/4)-!!(v%4)) |
1863 |
|
nb_err++; |
1864 |
|
} |
1865 |
|
if (nb_err!=0) { |
1866 |
|
printf( "ERROR! please post your platform/compiler specs to xvid-devel@xvid.org !\n" ); |
1867 |
|
} |
1868 |
|
} |
1869 |
|
|
1870 |
|
/********************************************************************* |
1871 |
|
* test bitstream functions |
1872 |
|
*********************************************************************/ |
1873 |
|
|
1874 |
|
#define BIT_BUF_SIZE 2000 |
1875 |
|
|
1876 |
|
static void test_bits() |
1877 |
|
{ |
1878 |
|
const int nb_tests = 50*speed_ref; |
1879 |
|
int tst; |
1880 |
|
uint32_t Crc; |
1881 |
|
uint8_t Buf[BIT_BUF_SIZE]; |
1882 |
|
uint32_t Extracted[BIT_BUF_SIZE*8]; /* worst case: bits read 1 by 1 */ |
1883 |
|
int Lens[BIT_BUF_SIZE*8]; |
1884 |
|
double t1; |
1885 |
|
|
1886 |
|
|
1887 |
|
printf( "\n === test bitstream ===\n" ); |
1888 |
|
ieee_reseed(1); |
1889 |
|
Crc = 0; |
1890 |
|
|
1891 |
|
t1 = gettime_usec(); |
1892 |
|
for(tst=0; tst<nb_tests; ++tst) { |
1893 |
|
Bitstream bs; |
1894 |
|
int m, m2, l, l2; |
1895 |
|
|
1896 |
|
for(l=0; l<BIT_BUF_SIZE; ++l) |
1897 |
|
Buf[l] = (uint8_t)ieee_rand(0,255); |
1898 |
|
|
1899 |
|
l = BIT_BUF_SIZE - ieee_rand(1,BIT_BUF_SIZE/10); |
1900 |
|
BitstreamInit(&bs, (void*)(Buf+BIT_BUF_SIZE-l), l); |
1901 |
|
|
1902 |
|
|
1903 |
|
BitstreamReset(&bs); |
1904 |
|
for(l2=l*8, m=0; l2>0; m++) { |
1905 |
|
const int b = ieee_rand(1,32); |
1906 |
|
Lens[m] = b; |
1907 |
|
l2 -= b; |
1908 |
|
if (l2<0) break; |
1909 |
|
Extracted[m] = BitstreamShowBits(&bs, b); |
1910 |
|
BitstreamSkip(&bs, b); |
1911 |
|
// printf( "<= %d: %d 0x%x\n", m, b, Extracted[m]); |
1912 |
|
} |
1913 |
|
|
1914 |
|
BitstreamReset(&bs); |
1915 |
|
for(m2=0; m2<m; ++m2) { |
1916 |
|
const int b = Lens[m2]; |
1917 |
|
const uint32_t v = BitstreamGetBits(&bs, b); |
1918 |
|
Crc |= (v!=Extracted[m2]); |
1919 |
|
// printf( "=> %d: %d 0x%x %c\n", m2, b, v, " *"[Crc]); |
1920 |
|
} |
1921 |
|
} |
1922 |
|
t1 = (gettime_usec() - t1) / nb_tests; |
1923 |
|
printf(" test_bits %.3f usec %s\n", t1, (Crc!=0)?"| ERROR": "" ); |
1924 |
|
} |
1925 |
|
|
1926 |
|
/********************************************************************* |
1927 |
* main |
* main |
1928 |
*********************************************************************/ |
*********************************************************************/ |
1929 |
|
|
1991 |
if (what==0 || what==10) test_sse(); |
if (what==0 || what==10) test_sse(); |
1992 |
if (what==0 || what==11) test_log2bin(); |
if (what==0 || what==11) test_log2bin(); |
1993 |
if (what==0 || what==12) test_gcd(); |
if (what==0 || what==12) test_gcd(); |
1994 |
|
if (what==0 || what==13) test_compiler(); |
1995 |
|
if (what==0 || what==17) test_bits(); |
1996 |
|
|
1997 |
|
|
1998 |
if (what==7) { |
if (what==7) { |