| 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 1.12 2004-04-02 21:29:21 edgomez Exp $ |
* $Id: xvid_bench.c,v 1.34 2006-11-01 07:12:26 Skal Exp $ |
| 23 |
* |
* |
| 24 |
****************************************************************************/ |
****************************************************************************/ |
| 25 |
|
|
| 26 |
/***************************************************************************** |
/***************************************************************************** |
| 27 |
* |
* |
| 28 |
* 'Reference' output is at the end of file. |
* 'Reference' output is at the end of file. |
|
* Don't take the checksums and crc too seriouly, they aren't |
|
|
* bullet-proof (should plug some .md5 here)... |
|
| 29 |
* |
* |
| 30 |
* compiles with something like: |
* compiles with something like: |
| 31 |
* gcc -o xvid_bench xvid_bench.c -I../src/ -lxvidcore -lm |
* gcc -o xvid_bench xvid_bench.c -I../src/ -lxvidcore -lm |
| 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 |
|
|
| 66 |
#define M_PI 3.14159265358979323846 |
#define M_PI 3.14159265358979323846 |
| 67 |
#endif |
#endif |
| 68 |
|
|
| 69 |
const int speed_ref = 100; /* on slow machines, decrease this value */ |
int speed_ref = 100; /* on slow machines, decrease this value */ |
| 70 |
|
int verbose = 0; |
| 71 |
|
unsigned int cpu_mask; |
| 72 |
|
|
| 73 |
/********************************************************************* |
/********************************************************************* |
| 74 |
* misc |
* misc |
| 84 |
#else |
#else |
| 85 |
clock_t clk; |
clock_t clk; |
| 86 |
clk = clock(); |
clk = clock(); |
| 87 |
return clk * 1000000 / CLOCKS_PER_SEC; |
return clk * 1000. / CLOCKS_PER_SEC; /* clock() returns time in Milliseconds */ |
| 88 |
#endif |
#endif |
| 89 |
} |
} |
| 90 |
|
|
| 113 |
unsigned int cpu; |
unsigned int cpu; |
| 114 |
} CPU; |
} CPU; |
| 115 |
|
|
| 116 |
CPU cpu_list[] = |
CPU cpu_list[] = { |
| 117 |
{ { "PLAINC", 0 } |
{ "PLAINC ", 0 }, |
| 118 |
#ifdef ARCH_IS_IA32 |
#ifdef ARCH_IS_IA32 |
| 119 |
, { "MMX ", XVID_CPU_MMX } |
{ "MMX ", XVID_CPU_MMX }, |
| 120 |
, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
{ "MMXEXT ", XVID_CPU_MMXEXT | XVID_CPU_MMX }, |
| 121 |
, { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } |
{ "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX }, |
| 122 |
, { "3DNOW ", XVID_CPU_3DNOW } |
{ "3DNOW ", XVID_CPU_3DNOW }, |
| 123 |
, { "3DNOWE", XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT } |
{ "3DNOWE ", XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT }, |
| 124 |
#endif |
#endif |
| 125 |
//, { "IA64 ", XVID_CPU_IA64 } |
#ifdef ARCH_IS_PPC |
| 126 |
//, { "TSC ", XVID_CPU_TSC } |
{ "ALTIVEC", XVID_CPU_ALTIVEC }, |
|
, { 0, 0 } }; |
|
|
|
|
|
CPU cpu_short_list[] = |
|
|
{ { "PLAINC", 0 } |
|
|
#ifdef ARCH_IS_IA32 |
|
|
, { "MMX ", XVID_CPU_MMX } |
|
|
//, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
|
| 127 |
#endif |
#endif |
| 128 |
//, { "IA64 ", XVID_CPU_IA64 } |
#ifdef ARCH_IS_X86_64 |
| 129 |
, { 0, 0 } }; |
{ "X86_64 ", XVID_CPU_ASM}, |
| 130 |
|
#endif |
| 131 |
CPU cpu_short_list2[] = |
#ifdef ARCH_IS_IA64 |
| 132 |
{ { "PLAINC", 0 } |
// { "IA64 ", XVID_CPU_IA64 }, |
|
#ifdef ARCH_IS_IA32 |
|
|
, { "MMX ", XVID_CPU_MMX } |
|
|
, { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } |
|
| 133 |
#endif |
#endif |
| 134 |
, { 0, 0 } }; |
// { "TSC ", XVID_CPU_TSC }, |
| 135 |
|
{ 0, 0 } |
| 136 |
|
}; |
| 137 |
|
|
| 138 |
|
|
| 139 |
int init_cpu(CPU *cpu) |
int init_cpu(CPU *cpu) |
| 246 |
}; |
}; |
| 247 |
|
|
| 248 |
uint32_t |
uint32_t |
| 249 |
calc_crc(uint8_t *mem, int len, uint32_t initial) |
calc_crc(uint8_t *mem, int len, uint32_t crc) |
| 250 |
{ |
{ |
|
|
|
|
register unsigned int crc; |
|
|
|
|
|
crc = initial; |
|
|
|
|
| 251 |
while( len >= 8) { |
while( len >= 8) { |
| 252 |
DO8(mem, crc); |
DO8(mem, crc); |
| 253 |
len -= 8; |
len -= 8; |
| 258 |
len--; |
len--; |
| 259 |
} |
} |
| 260 |
|
|
| 261 |
return(crc); |
return crc; |
| 262 |
|
} |
| 263 |
|
|
| 264 |
|
void byte_swap(uint8_t *mem, int len, int element_size) { |
| 265 |
|
#ifdef ARCH_IS_BIG_ENDIAN |
| 266 |
|
int i; |
| 267 |
|
|
| 268 |
|
if(element_size == 1) { |
| 269 |
|
/* No need to swap */ |
| 270 |
|
} else if(element_size == 2) { |
| 271 |
|
uint8_t temp[2]; |
| 272 |
|
|
| 273 |
|
for(i=0; i < (len/2); i++ ) { |
| 274 |
|
temp[0] = mem[0]; |
| 275 |
|
temp[1] = mem[1]; |
| 276 |
|
mem[0] = temp[1]; |
| 277 |
|
mem[1] = temp[0]; |
| 278 |
|
|
| 279 |
|
mem += 2; |
| 280 |
|
} |
| 281 |
|
} else if(element_size == 4) { |
| 282 |
|
uint8_t temp[4]; |
| 283 |
|
|
| 284 |
|
for(i=0; i < (len/4); i++ ) { |
| 285 |
|
temp[0] = mem[0]; |
| 286 |
|
temp[1] = mem[1]; |
| 287 |
|
temp[2] = mem[2]; |
| 288 |
|
temp[3] = mem[3]; |
| 289 |
|
mem[0] = temp[3]; |
| 290 |
|
mem[1] = temp[2]; |
| 291 |
|
mem[2] = temp[1]; |
| 292 |
|
mem[3] = temp[0]; |
| 293 |
|
|
| 294 |
|
mem += 4; |
| 295 |
|
} |
| 296 |
|
} else { |
| 297 |
|
printf("ERROR: byte_swap unsupported element_size(%u)\n", element_size); |
| 298 |
|
} |
| 299 |
|
#endif |
| 300 |
} |
} |
| 301 |
|
|
| 302 |
/********************************************************************* |
/********************************************************************* |
| 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 |
} |
} |
| 560 |
|
|
| 561 |
|
#undef ENTER |
| 562 |
|
#undef LEAVE |
| 563 |
|
#undef TEST_MB |
| 564 |
|
#undef TEST_MB2 |
| 565 |
|
|
| 566 |
/********************************************************************* |
/********************************************************************* |
| 567 |
* test transfer |
* test transfer |
| 568 |
*********************************************************************/ |
*********************************************************************/ |
| 592 |
} \ |
} \ |
| 593 |
emms(); \ |
emms(); \ |
| 594 |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
| 595 |
s = calc_crc((uint8_t*)(DST), sizeof((DST)), CRC32_INITIAL) |
byte_swap((uint8_t*)(DST), 8*32*sizeof((DST)[0]), sizeof((DST)[0])); \ |
| 596 |
|
s = calc_crc((uint8_t*)(DST), 8*32*sizeof((DST)[0]), CRC32_INITIAL) |
| 597 |
|
|
| 598 |
#define TEST_TRANSFER(FUNC, DST, SRC) \ |
#define TEST_TRANSFER(FUNC, DST, SRC) \ |
| 599 |
TEST_TRANSFER_BEGIN(DST); \ |
TEST_TRANSFER_BEGIN(DST); \ |
| 619 |
} \ |
} \ |
| 620 |
emms(); \ |
emms(); \ |
| 621 |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
| 622 |
s = calc_crc((uint8_t*)(DST), sizeof((DST)), CRC32_INITIAL) |
byte_swap((uint8_t*)(DST), 8*32*sizeof((DST)[0]), sizeof((DST)[0])); \ |
| 623 |
|
s = calc_crc((uint8_t*)(DST), 8*32*sizeof((DST)[0]), CRC32_INITIAL) |
| 624 |
|
|
| 625 |
#define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ |
#define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ |
| 626 |
TEST_TRANSFER2_BEGIN(DST,SRC); \ |
TEST_TRANSFER2_BEGIN(DST,SRC); \ |
| 637 |
const int nb_tests = 4000*speed_ref; |
const int nb_tests = 4000*speed_ref; |
| 638 |
int i; |
int i; |
| 639 |
CPU *cpu; |
CPU *cpu; |
| 640 |
uint8_t Src8[8*32], Dst8[8*32], Ref1[8*32], Ref2[8*32]; |
// uint8_t Src8[8*32], Dst8[8*32], Ref1[8*32], Ref2[8*32]; |
| 641 |
int16_t Src16[8*32], Dst16[8*32]; |
// int16_t Src16[8*32], Dst16[8*32]; |
| 642 |
|
DECLARE_ALIGNED_MATRIX(Src8, 8, 32, uint8_t, CACHE_LINE); |
| 643 |
|
DECLARE_ALIGNED_MATRIX(Dst8, 8, 32, uint8_t, CACHE_LINE); |
| 644 |
|
DECLARE_ALIGNED_MATRIX(Ref1, 8, 32, uint8_t, CACHE_LINE); |
| 645 |
|
DECLARE_ALIGNED_MATRIX(Ref2, 8, 32, uint8_t, CACHE_LINE); |
| 646 |
|
DECLARE_ALIGNED_MATRIX(Src16, 8, 32, uint16_t, CACHE_LINE); |
| 647 |
|
DECLARE_ALIGNED_MATRIX(Dst16, 8, 32, uint16_t, CACHE_LINE); |
| 648 |
|
|
| 649 |
printf( "\n === test transfer ===\n" ); |
printf( "\n === test transfer ===\n" ); |
| 650 |
|
|
| 666 |
cpu->name, t, s, |
cpu->name, t, s, |
| 667 |
(s!=0xee7ccbb4)?"| ERROR": ""); |
(s!=0xee7ccbb4)?"| ERROR": ""); |
| 668 |
|
|
| 669 |
|
/* New functions for field prediction by CK 1.10.2005 */ |
| 670 |
|
#pragma NEW8X4 |
| 671 |
|
TEST_TRANSFER(transfer8x4_copy, Dst8, Src8); |
| 672 |
|
printf("%s - 8to4 %.3f usec crc32=0x%08x %s\n", |
| 673 |
|
cpu->name, t, s, |
| 674 |
|
(s!=0xbb9c3db5)?"| ERROR": ""); |
| 675 |
|
/* End of new functions */ |
| 676 |
|
|
| 677 |
TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
| 678 |
printf("%s - 8to8 %.3f usec crc32=0x%08x %s\n", |
printf("%s - 8to8 %.3f usec crc32=0x%08x %s\n", |
| 679 |
cpu->name, t, s, |
cpu->name, t, s, |
| 687 |
TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); |
TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); |
| 688 |
{ |
{ |
| 689 |
int s1, s2; |
int s1, s2; |
| 690 |
s1 = calc_crc((uint8_t*)Dst16, sizeof(Dst16), CRC32_INITIAL); |
s1 = calc_crc((uint8_t*)Dst16, 8*32*sizeof(Dst16[0]), CRC32_INITIAL); |
| 691 |
s2 = calc_crc((uint8_t*)Src8, sizeof(Src8), CRC32_INITIAL); |
s2 = calc_crc((uint8_t*)Src8, 8*32*sizeof(Src8[0]), CRC32_INITIAL); |
| 692 |
printf("%s - 8to16sub %.3f usec crc32(1)=0x%08x crc32(2)=0x%08x %s %s\n", |
printf("%s - 8to16sub %.3f usec crc32(1)=0x%08x crc32(2)=0x%08x %s %s\n", |
| 693 |
cpu->name, t, s1, s2, |
cpu->name, t, s1, s2, |
| 694 |
(s1!=0xa1e07163)?"| ERROR1": "", |
(s1!=0xa1e07163)?"| ERROR1": "", |
| 717 |
for(q=1; q<=max_Q; ++q) { \ |
for(q=1; q<=max_Q; ++q) { \ |
| 718 |
for(tst=0; tst<nb_tests; ++tst) \ |
for(tst=0; tst<nb_tests; ++tst) \ |
| 719 |
(FUNC)((DST), (SRC), q, mpeg_quant_matrices); \ |
(FUNC)((DST), (SRC), q, mpeg_quant_matrices); \ |
| 720 |
s = calc_crc((uint8_t*)(DST), 64*sizeof(int16_t), s); \ |
byte_swap((uint8_t*)(DST), 64*sizeof((DST)[0]), sizeof((DST)[0])); \ |
| 721 |
|
s = calc_crc((uint8_t*)(DST), 64*sizeof((DST)[0]), s); \ |
| 722 |
} \ |
} \ |
| 723 |
emms(); \ |
emms(); \ |
| 724 |
} \ |
} \ |
| 733 |
for(q=1; q<=max_Q; ++q) { \ |
for(q=1; q<=max_Q; ++q) { \ |
| 734 |
for(tst=0; tst<nb_tests; ++tst) \ |
for(tst=0; tst<nb_tests; ++tst) \ |
| 735 |
(FUNC)((DST), (SRC), q, q, mpeg_quant_matrices); \ |
(FUNC)((DST), (SRC), q, q, mpeg_quant_matrices); \ |
| 736 |
s = calc_crc((uint8_t*)(DST), 64*sizeof(int16_t), s); \ |
byte_swap((uint8_t*)(DST), 64*sizeof((DST)[0]), sizeof((DST)[0])); \ |
| 737 |
|
s = calc_crc((uint8_t*)(DST), 64*sizeof((DST)[0]), s); \ |
| 738 |
} \ |
} \ |
| 739 |
emms(); \ |
emms(); \ |
| 740 |
} \ |
} \ |
| 862 |
} |
} |
| 863 |
|
|
| 864 |
/********************************************************************* |
/********************************************************************* |
| 865 |
|
* test distortion operators |
| 866 |
|
*********************************************************************/ |
| 867 |
|
|
| 868 |
|
static void ieee_reseed(long s); |
| 869 |
|
static long ieee_rand(int Min, int Max); |
| 870 |
|
|
| 871 |
|
#define TEST_SSE(FUNCTION, SRC1, SRC2, STRIDE) \ |
| 872 |
|
do { \ |
| 873 |
|
t = gettime_usec(); \ |
| 874 |
|
tst = nb_tests; \ |
| 875 |
|
while((tst--)>0) sse = (FUNCTION)((SRC1), (SRC2), (STRIDE)); \ |
| 876 |
|
emms(); \ |
| 877 |
|
t = (gettime_usec() - t)/(double)nb_tests; \ |
| 878 |
|
} while(0) |
| 879 |
|
|
| 880 |
|
|
| 881 |
|
void test_sse() |
| 882 |
|
{ |
| 883 |
|
const int nb_tests = 100000*speed_ref; |
| 884 |
|
int i; |
| 885 |
|
CPU *cpu; |
| 886 |
|
DECLARE_ALIGNED_MATRIX(Src1, 8, 8, int16_t, 16); |
| 887 |
|
DECLARE_ALIGNED_MATRIX(Src2, 8, 8, int16_t, 16); |
| 888 |
|
DECLARE_ALIGNED_MATRIX(Src3, 8, 8, int16_t, 16); |
| 889 |
|
DECLARE_ALIGNED_MATRIX(Src4, 8, 8, int16_t, 16); |
| 890 |
|
|
| 891 |
|
printf( "\n ===== test sse =====\n" ); |
| 892 |
|
|
| 893 |
|
ieee_reseed(1); |
| 894 |
|
for(i=0; i<64; ++i) { |
| 895 |
|
Src1[i] = ieee_rand(-2048, 2047); |
| 896 |
|
Src2[i] = ieee_rand(-2048, 2047); |
| 897 |
|
Src3[i] = ieee_rand(-2048, 2047); |
| 898 |
|
Src4[i] = ieee_rand(-2048, 2047); |
| 899 |
|
} |
| 900 |
|
|
| 901 |
|
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 902 |
|
{ |
| 903 |
|
double t; |
| 904 |
|
int tst, sse; |
| 905 |
|
|
| 906 |
|
if (!init_cpu(cpu)) |
| 907 |
|
continue; |
| 908 |
|
|
| 909 |
|
/* 16 bit element blocks */ |
| 910 |
|
TEST_SSE(sse8_16bit, Src1, Src2, 16); |
| 911 |
|
printf("%s - sse8_16bit#1 %.3f usec sse=%d %s\n", |
| 912 |
|
cpu->name, t, sse, (sse!=182013834)?"| ERROR": ""); |
| 913 |
|
TEST_SSE(sse8_16bit, Src1, Src3, 16); |
| 914 |
|
printf("%s - sse8_16bit#2 %.3f usec sse=%d %s\n", |
| 915 |
|
cpu->name, t, sse, (sse!=142545203)?"| ERROR": ""); |
| 916 |
|
TEST_SSE(sse8_16bit, Src1, Src4, 16); |
| 917 |
|
printf("%s - sse8_16bit#3 %.3f usec sse=%d %s\n", |
| 918 |
|
cpu->name, t, sse, (sse!=146340935)?"| ERROR": ""); |
| 919 |
|
TEST_SSE(sse8_16bit, Src2, Src3, 16); |
| 920 |
|
printf("%s - sse8_16bit#4 %.3f usec sse=%d %s\n", |
| 921 |
|
cpu->name, t, sse, (sse!=130136661)?"| ERROR": ""); |
| 922 |
|
TEST_SSE(sse8_16bit, Src2, Src4, 16); |
| 923 |
|
printf("%s - sse8_16bit#5 %.3f usec sse=%d %s\n", |
| 924 |
|
cpu->name, t, sse, (sse!=136870353)?"| ERROR": ""); |
| 925 |
|
TEST_SSE(sse8_16bit, Src3, Src4, 16); |
| 926 |
|
printf("%s - sse8_16bit#6 %.3f usec sse=%d %s\n", |
| 927 |
|
cpu->name, t, sse, (sse!=164107772)?"| ERROR": ""); |
| 928 |
|
|
| 929 |
|
/* 8 bit element blocks */ |
| 930 |
|
TEST_SSE(sse8_8bit, (int8_t*)Src1, (int8_t*)Src2, 8); |
| 931 |
|
printf("%s - sse8_8bit#1 %.3f usec sse=%d %s\n", |
| 932 |
|
cpu->name, t, sse, (sse!=1356423)?"| ERROR": ""); |
| 933 |
|
TEST_SSE(sse8_8bit, (int8_t*)Src1, (int8_t*)Src3, 8); |
| 934 |
|
printf("%s - sse8_8bit#2 %.3f usec sse=%d %s\n", |
| 935 |
|
cpu->name, t, sse, (sse!=1173074)?"| ERROR": ""); |
| 936 |
|
TEST_SSE(sse8_8bit, (int8_t*)Src1, (int8_t*)Src4, 8); |
| 937 |
|
printf("%s - sse8_8bit#3 %.3f usec sse=%d %s\n", |
| 938 |
|
cpu->name, t, sse, (sse!=1092357)?"| ERROR": ""); |
| 939 |
|
TEST_SSE(sse8_8bit, (int8_t*)Src2, (int8_t*)Src3, 8); |
| 940 |
|
printf("%s - sse8_8bit#4 %.3f usec sse=%d %s\n", |
| 941 |
|
cpu->name, t, sse, (sse!=1360239)?"| ERROR": ""); |
| 942 |
|
TEST_SSE(sse8_8bit, (int8_t*)Src2, (int8_t*)Src4, 8); |
| 943 |
|
printf("%s - sse8_8bit#5 %.3f usec sse=%d %s\n", |
| 944 |
|
cpu->name, t, sse, (sse!=1208414)?"| ERROR": ""); |
| 945 |
|
TEST_SSE(sse8_8bit, (int8_t*)Src3, (int8_t*)Src4, 8); |
| 946 |
|
printf("%s - sse8_8bit#6 %.3f usec sse=%d %s\n", |
| 947 |
|
cpu->name, t, sse, (sse!=1099285)?"| ERROR": ""); |
| 948 |
|
|
| 949 |
|
printf(" ---\n"); |
| 950 |
|
} |
| 951 |
|
} |
| 952 |
|
|
| 953 |
|
/********************************************************************* |
| 954 |
* test non-zero AC counting |
* test non-zero AC counting |
| 955 |
*********************************************************************/ |
*********************************************************************/ |
| 956 |
|
|
| 957 |
#define TEST_CBP(FUNC, SRC) \ |
#define TEST_CBP(FUNC, SRC, NB) \ |
| 958 |
t = gettime_usec(); \ |
t = gettime_usec(); \ |
| 959 |
emms(); \ |
emms(); \ |
| 960 |
for(tst=0; tst<nb_tests; ++tst) { \ |
for(tst=0; tst<NB; ++tst) { \ |
| 961 |
cbp = (FUNC)((SRC)); \ |
cbp = (FUNC)((SRC)); \ |
| 962 |
} \ |
} \ |
| 963 |
emms(); \ |
emms(); \ |
| 966 |
void test_cbp() |
void test_cbp() |
| 967 |
{ |
{ |
| 968 |
const int nb_tests = 10000*speed_ref; |
const int nb_tests = 10000*speed_ref; |
| 969 |
int i; |
int i, n, m; |
| 970 |
CPU *cpu; |
CPU *cpu; |
| 971 |
DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16); |
| 972 |
DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16); |
| 973 |
DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16); |
| 974 |
DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16); |
| 975 |
|
DECLARE_ALIGNED_MATRIX(Src5, 6, 64, int16_t, 16); |
| 976 |
|
|
| 977 |
printf( "\n ===== test cbp =====\n" ); |
printf( "\n ===== test cbp =====\n" ); |
| 978 |
|
|
| 981 |
Src2[i] = (i<3*64); /* half-full */ |
Src2[i] = (i<3*64); /* half-full */ |
| 982 |
Src3[i] = ((i+32)>3*64); |
Src3[i] = ((i+32)>3*64); |
| 983 |
Src4[i] = (i==(3*64+2) || i==(5*64+9)); |
Src4[i] = (i==(3*64+2) || i==(5*64+9)); |
| 984 |
|
Src5[i] = ieee_rand(0,1) ? -1 : 1; /* +/- test */ |
| 985 |
} |
} |
| 986 |
|
|
| 987 |
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 992 |
if (!init_cpu(cpu)) |
if (!init_cpu(cpu)) |
| 993 |
continue; |
continue; |
| 994 |
|
|
| 995 |
TEST_CBP(calc_cbp, Src1); |
TEST_CBP(calc_cbp, Src1, nb_tests); |
| 996 |
printf("%s - calc_cbp#1 %.3f usec cbp=0x%02x\n", |
printf("%s - calc_cbp#1 %.3f usec cbp=0x%02x %s\n", |
| 997 |
cpu->name, t, cbp, (cbp!=0x15)?"| ERROR": ""); |
cpu->name, t, cbp, (cbp!=0x15)?"| ERROR": ""); |
| 998 |
TEST_CBP(calc_cbp, Src2); |
TEST_CBP(calc_cbp, Src2, nb_tests); |
| 999 |
printf("%s - calc_cbp#2 %.3f usec cbp=0x%02x\n", |
printf("%s - calc_cbp#2 %.3f usec cbp=0x%02x %s\n", |
| 1000 |
cpu->name, t, cbp, (cbp!=0x38)?"| ERROR": ""); |
cpu->name, t, cbp, (cbp!=0x38)?"| ERROR": ""); |
| 1001 |
TEST_CBP(calc_cbp, Src3); |
TEST_CBP(calc_cbp, Src3, nb_tests); |
| 1002 |
printf("%s - calc_cbp#3 %.3f usec cbp=0x%02x\n", |
printf("%s - calc_cbp#3 %.3f usec cbp=0x%02x %s\n", |
| 1003 |
cpu->name, t, cbp, (cbp!=0x0f)?"| ERROR": "" ); |
cpu->name, t, cbp, (cbp!=0x0f)?"| ERROR": "" ); |
| 1004 |
TEST_CBP(calc_cbp, Src4); |
TEST_CBP(calc_cbp, Src4, nb_tests); |
| 1005 |
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x\n", |
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
| 1006 |
cpu->name, t, cbp, (cbp!=0x05)?"| ERROR": "" ); |
cpu->name, t, cbp, (cbp!=0x05)?"| ERROR": "" ); |
| 1007 |
|
TEST_CBP(calc_cbp, Src5, nb_tests); |
| 1008 |
|
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
| 1009 |
|
cpu->name, t, cbp, (cbp!=0x3f)?"| ERROR": "" ); |
| 1010 |
printf( " --- \n" ); |
printf( " --- \n" ); |
| 1011 |
} |
} |
| 1012 |
|
|
| 1013 |
|
for(cpu = cpu_list; cpu->name!=0; ++cpu) /* bench suggested by Carlo (carlo dot bramix at libero dot it) */ |
| 1014 |
|
{ |
| 1015 |
|
double t; |
| 1016 |
|
int tst, cbp, err; |
| 1017 |
|
|
| 1018 |
|
if (!init_cpu(cpu)) |
| 1019 |
|
continue; |
| 1020 |
|
|
| 1021 |
|
err = 0; |
| 1022 |
|
for(n=0; n<6; ++n) |
| 1023 |
|
{ |
| 1024 |
|
for(m=0; m<64; ++m) |
| 1025 |
|
{ |
| 1026 |
|
for(i=0; i<6*64; ++i) |
| 1027 |
|
Src1[i] = (i== (m + n*64)); |
| 1028 |
|
|
| 1029 |
|
TEST_CBP(calc_cbp, Src1, 1); |
| 1030 |
|
if (cbp!= (((m!=0)<<(5-n)))) |
| 1031 |
|
{ |
| 1032 |
|
printf( "%s - calc_cbp#5: ERROR at pos %d / %d!\n", cpu->name, n, m); |
| 1033 |
|
err = 1; |
| 1034 |
|
break; |
| 1035 |
|
} |
| 1036 |
|
} |
| 1037 |
|
} |
| 1038 |
|
if (!err) |
| 1039 |
|
printf( " %s - calc_cbp#5 : OK\n", cpu->name ); |
| 1040 |
|
|
| 1041 |
|
} |
| 1042 |
} |
} |
| 1043 |
|
|
| 1044 |
/********************************************************************* |
/********************************************************************* |
| 1432 |
* measure raw decoding speed |
* measure raw decoding speed |
| 1433 |
*********************************************************************/ |
*********************************************************************/ |
| 1434 |
|
|
| 1435 |
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) |
| 1436 |
{ |
{ |
| 1437 |
FILE *f = 0; |
FILE *f = 0; |
| 1438 |
void *dechandle = 0; |
void *dechandle = 0; |
| 1443 |
double t = 0.; |
double t = 0.; |
| 1444 |
int nb = 0; |
int nb = 0; |
| 1445 |
uint8_t *buf = 0; |
uint8_t *buf = 0; |
| 1446 |
uint8_t *rgb_out = 0; |
uint8_t *yuv_out = 0; |
| 1447 |
int buf_size, pos; |
int buf_size, pos; |
| 1448 |
uint32_t chksum = 0; |
uint32_t chksum = 0; |
| 1449 |
|
int bps = (width+31) & ~31; |
| 1450 |
|
|
| 1451 |
memset(&xinit, 0, sizeof(xinit)); |
memset(&xinit, 0, sizeof(xinit)); |
| 1452 |
xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; |
xinit.cpu_flags = cpu_mask; |
| 1453 |
xinit.version = XVID_VERSION; |
xinit.version = XVID_VERSION; |
| 1454 |
xvid_global(NULL, 0, &xinit, NULL); |
xvid_global(NULL, 0, &xinit, NULL); |
| 1455 |
|
|
| 1459 |
xparam.version = XVID_VERSION; |
xparam.version = XVID_VERSION; |
| 1460 |
xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
| 1461 |
if (xerr==XVID_ERR_FAIL) { |
if (xerr==XVID_ERR_FAIL) { |
| 1462 |
printf("can't init decoder (err=%d)\n", xerr); |
printf("ERROR: can't init decoder (err=%d)\n", xerr); |
| 1463 |
return; |
return; |
| 1464 |
} |
} |
| 1465 |
dechandle = xparam.handle; |
dechandle = xparam.handle; |
| 1467 |
|
|
| 1468 |
f = fopen(name, "rb"); |
f = fopen(name, "rb"); |
| 1469 |
if (f==0) { |
if (f==0) { |
| 1470 |
printf( "can't open file '%s'\n", name); |
printf( "ERROR: can't open file '%s'\n", name); |
| 1471 |
return; |
return; |
| 1472 |
} |
} |
| 1473 |
fseek(f, 0, SEEK_END); |
fseek(f, 0, SEEK_END); |
| 1474 |
buf_size = ftell(f); |
buf_size = ftell(f); |
| 1475 |
fseek(f, 0, SEEK_SET); |
fseek(f, 0, SEEK_SET); |
| 1476 |
if (buf_size<=0) { |
if (buf_size<=0) { |
| 1477 |
printf("error while stating file\n"); |
printf("ERROR: error while stating file\n"); |
| 1478 |
goto End; |
goto End; |
| 1479 |
} |
} |
|
else printf( "Input size: %d\n", buf_size); |
|
| 1480 |
|
|
| 1481 |
buf = malloc(buf_size); /* should be enuf' */ |
buf = malloc(buf_size); |
| 1482 |
rgb_out = calloc(4, width*height); /* <-room for _RGB24 */ |
yuv_out = calloc(1, bps*height*3/2 + 15); |
| 1483 |
if (buf==0 || rgb_out==0) { |
if (buf==0 || yuv_out==0) { |
| 1484 |
printf( "malloc failed!\n" ); |
printf( "ERROR: malloc failed!\n" ); |
| 1485 |
goto End; |
goto End; |
| 1486 |
} |
} |
| 1487 |
|
|
| 1488 |
if (fread(buf, buf_size, 1, f)!=1) { |
if (fread(buf, buf_size, 1, f)!=1) { |
| 1489 |
printf( "file-read failed\n" ); |
printf( "ERROR: file-read failed\n" ); |
| 1490 |
goto End; |
goto End; |
| 1491 |
} |
} |
| 1492 |
|
|
| 1494 |
pos = 0; |
pos = 0; |
| 1495 |
t = -gettime_usec(); |
t = -gettime_usec(); |
| 1496 |
while(1) { |
while(1) { |
| 1497 |
|
int y; |
| 1498 |
|
|
| 1499 |
memset(&xframe, 0, sizeof(xframe)); |
memset(&xframe, 0, sizeof(xframe)); |
| 1500 |
xframe.version = XVID_VERSION; |
xframe.version = XVID_VERSION; |
| 1501 |
xframe.bitstream = buf + pos; |
xframe.bitstream = buf + pos; |
| 1502 |
xframe.length = buf_size - pos; |
xframe.length = buf_size - pos; |
| 1503 |
xframe.output.plane[0] = rgb_out; |
xframe.output.plane[0] = (uint8_t*)(((size_t)yuv_out + 15) & ~15); |
| 1504 |
xframe.output.stride[0] = width; |
xframe.output.plane[1] = (uint8_t*)xframe.output.plane[0] + bps*height; |
| 1505 |
xframe.output.csp = XVID_CSP_BGR; |
xframe.output.plane[2] = (uint8_t*)xframe.output.plane[1] + bps/2; |
| 1506 |
|
xframe.output.stride[0] = bps; |
| 1507 |
|
xframe.output.stride[1] = bps; |
| 1508 |
|
xframe.output.stride[2] = bps; |
| 1509 |
|
xframe.output.csp = XVID_CSP_I420; |
| 1510 |
xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); |
xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); |
| 1511 |
|
if (xerr<0) { |
| 1512 |
|
printf("ERROR: decoding failed for frame #%d (err=%d)!\n", nb, xerr); |
| 1513 |
|
break; |
| 1514 |
|
} |
| 1515 |
|
else if (xerr==0) |
| 1516 |
|
break; |
| 1517 |
|
else if (verbose>0) printf("#%d %d\n", nb, xerr ); |
| 1518 |
|
|
| 1519 |
|
pos += xerr; |
| 1520 |
nb++; |
nb++; |
| 1521 |
pos += xframe.length; |
|
| 1522 |
if (with_chksum) { |
for(y=0; y<height/2; ++y) { |
| 1523 |
int k = width*height; |
chksum = calc_crc((uint8_t*)xframe.output.plane[0] + (2*y+0)*bps, width, chksum); |
| 1524 |
uint32_t *ptr = (uint32_t *)rgb_out; |
chksum = calc_crc((uint8_t*)xframe.output.plane[0] + (2*y+1)*bps, width, chksum); |
| 1525 |
while(k-->0) chksum += *ptr++; |
chksum = calc_crc((uint8_t*)xframe.output.plane[1] + y*bps, width/2, chksum); |
| 1526 |
|
chksum = calc_crc((uint8_t*)xframe.output.plane[2] + y*bps, width/2, chksum); |
| 1527 |
} |
} |
| 1528 |
if (pos==buf_size) |
if (pos==buf_size) |
| 1529 |
break; |
break; |
|
if (xerr==XVID_ERR_FAIL) { |
|
|
printf("decoding failed for frame #%d (err=%d)!\n", nb, xerr); |
|
|
break; |
|
|
} |
|
| 1530 |
} |
} |
| 1531 |
t += gettime_usec(); |
t += gettime_usec(); |
| 1532 |
|
if (ref_chksum==0) { |
| 1533 |
if (t>0.) |
if (t>0.) |
| 1534 |
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 ); |
| 1535 |
if (with_chksum) |
} |
| 1536 |
printf("checksum: 0x%.8x\n", chksum); |
else { |
| 1537 |
|
printf("FPS:%.1f Checksum: 0x%.8x Expected:0x%.8x | %s\n", |
| 1538 |
|
t>0. ? (float)(nb*1.e6f/t) : 0.f, chksum, ref_chksum, (chksum==ref_chksum) ? "OK" : "ERROR"); |
| 1539 |
|
} |
| 1540 |
|
|
| 1541 |
End: |
End: |
| 1542 |
if (rgb_out!=0) free(rgb_out); |
if (yuv_out!=0) free(yuv_out); |
| 1543 |
if (buf!=0) free(buf); |
if (buf!=0) free(buf); |
| 1544 |
if (dechandle!=0) { |
if (dechandle!=0) { |
| 1545 |
xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); |
xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); |
| 1546 |
if (xerr==XVID_ERR_FAIL) |
if (xerr==XVID_ERR_FAIL) |
| 1547 |
printf("destroy-decoder failed (err=%d)!\n", xerr); |
printf("ERROR: destroy-decoder failed (err=%d)!\n", xerr); |
| 1548 |
} |
} |
| 1549 |
if (f!=0) fclose(f); |
if (f!=0) fclose(f); |
| 1550 |
} |
} |
| 1708 |
} |
} |
| 1709 |
|
|
| 1710 |
/********************************************************************* |
/********************************************************************* |
| 1711 |
|
* test some YUV func |
| 1712 |
|
*********************************************************************/ |
| 1713 |
|
|
| 1714 |
|
#define ENTER \ |
| 1715 |
|
for(i=0; i<(int)sizeof(Dst0); ++i) Dst0[0][i] = 0; \ |
| 1716 |
|
t = gettime_usec(); \ |
| 1717 |
|
emms(); |
| 1718 |
|
|
| 1719 |
|
#define LEAVE \ |
| 1720 |
|
emms(); \ |
| 1721 |
|
t = (gettime_usec() - t) / nb_tests; \ |
| 1722 |
|
iCrc = calc_crc((uint8_t*)Dst0, sizeof(Dst0), CRC32_INITIAL) |
| 1723 |
|
|
| 1724 |
|
#define TEST_YUYV(FUNC, S) \ |
| 1725 |
|
ENTER \ |
| 1726 |
|
for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst0[0], S*WIDTH, Src0[0], Src0[1], Src0[2], WIDTH, WIDTH/2, WIDTH, HEIGHT, 0); \ |
| 1727 |
|
LEAVE |
| 1728 |
|
|
| 1729 |
|
#define WIDTH 128 |
| 1730 |
|
#define HEIGHT 32 |
| 1731 |
|
void test_yuv() |
| 1732 |
|
{ |
| 1733 |
|
const int nb_tests = 200*speed_ref; |
| 1734 |
|
CPU *cpu; |
| 1735 |
|
uint8_t Src0[3][WIDTH*HEIGHT]; |
| 1736 |
|
uint8_t Dst0[4][WIDTH*HEIGHT]; |
| 1737 |
|
int i, j; |
| 1738 |
|
double t; |
| 1739 |
|
int tst, iCrc; |
| 1740 |
|
|
| 1741 |
|
colorspace_init(); |
| 1742 |
|
ieee_reseed(1); |
| 1743 |
|
for(i=0; i<(int)sizeof(Src0); ++i) Src0[0][i] = ieee_rand(0,255); |
| 1744 |
|
for(i=0; i<(int)sizeof(Dst0); ++i) Dst0[0][i] = 0x5a; |
| 1745 |
|
|
| 1746 |
|
printf( "\n === test YUV ===\n" ); |
| 1747 |
|
|
| 1748 |
|
init_cpu(&cpu_list[0]); |
| 1749 |
|
TEST_YUYV(yv12_to_yuyv_c, 4); |
| 1750 |
|
printf(" yv12_to_yuyv_c %.3f usec crc32=0x%08x %s\n", |
| 1751 |
|
t, iCrc, (iCrc!=0xeb1a0b0a)?"| ERROR": "" ); |
| 1752 |
|
TEST_YUYV(yv12_to_uyvy_c, 4); |
| 1753 |
|
printf(" yv12_to_uyvy_c %.3f usec crc32=0x%08x %s\n", |
| 1754 |
|
t, iCrc, (iCrc!=0x6e82f55b)?"| ERROR": "" ); |
| 1755 |
|
|
| 1756 |
|
#ifdef ARCH_IS_IA32 |
| 1757 |
|
init_cpu(&cpu_list[1]); |
| 1758 |
|
TEST_YUYV(yv12_to_yuyv_mmx, 4); |
| 1759 |
|
printf(" yv12_to_yuyv_mmx %.3f usec crc32=0x%08x %s\n", |
| 1760 |
|
t, iCrc, (iCrc!=0xeb1a0b0a)?"| ERROR": "" ); |
| 1761 |
|
|
| 1762 |
|
TEST_YUYV(yv12_to_uyvy_mmx, 4); |
| 1763 |
|
printf(" yv12_to_uyvy_mmx %.3f usec crc32=0x%08x %s\n", |
| 1764 |
|
t, iCrc, (iCrc!=0x6e82f55b)?"| ERROR": "" ); |
| 1765 |
|
#endif |
| 1766 |
|
|
| 1767 |
|
#ifdef ARCH_IS_PPC |
| 1768 |
|
init_cpu(&cpu_list[1]); |
| 1769 |
|
TEST_YUYV(yv12_to_yuyv_altivec_c, 4); |
| 1770 |
|
printf(" yv12_to_yuyv_altivec_c %.3f usec crc32=0x%08x %s\n", |
| 1771 |
|
t, iCrc, (iCrc!=0xeb1a0b0a)?"| ERROR": "" ); |
| 1772 |
|
|
| 1773 |
|
TEST_YUYV(yv12_to_uyvy_altivec_c, 4); |
| 1774 |
|
printf(" yv12_to_uyvy_altivec_c %.3f usec crc32=0x%08x %s\n", |
| 1775 |
|
t, iCrc, (iCrc!=0x6e82f55b)?"| ERROR": "" ); |
| 1776 |
|
#endif |
| 1777 |
|
printf( " --- \n" ); |
| 1778 |
|
} |
| 1779 |
|
|
| 1780 |
|
#define TEST_YV2(FUNC, WITH_UV, WITH_FLIP) \ |
| 1781 |
|
ENTER \ |
| 1782 |
|
for(tst=0; tst<nb_tests; ++tst) (FUNC)(Dst0[0], Dst0[1], Dst0[2], WIDTH, WIDTH, \ |
| 1783 |
|
Src0[0], (WITH_UV) ? Src0[1] : 0, (WITH_UV) ? Src0[2] : 0, WIDTH, WIDTH, \ |
| 1784 |
|
WIDTH-2, HEIGHT-2, WITH_FLIP); \ |
| 1785 |
|
LEAVE |
| 1786 |
|
|
| 1787 |
|
#define PRINT_NxN(DATA,W,H,STR) { \ |
| 1788 |
|
int i,j; \ |
| 1789 |
|
for(j=0; j<(H); ++j) { \ |
| 1790 |
|
for(i=0; i<(W); ++i) printf( "0x%.2x ", (DATA)[i+j*(STR)] );\ |
| 1791 |
|
printf("\n"); \ |
| 1792 |
|
} \ |
| 1793 |
|
printf("---\n"); \ |
| 1794 |
|
} |
| 1795 |
|
|
| 1796 |
|
static const int yv12_CRCs[2][2] = { |
| 1797 |
|
{0x5cab7cf0,0xdab46541} |
| 1798 |
|
, {0xe8bae865,0x1faf77b7} |
| 1799 |
|
}; |
| 1800 |
|
|
| 1801 |
|
void test_yuv2() |
| 1802 |
|
{ |
| 1803 |
|
const int nb_tests = 800*speed_ref; |
| 1804 |
|
CPU *cpu; |
| 1805 |
|
uint8_t Src0[3][WIDTH*HEIGHT]; |
| 1806 |
|
uint8_t Dst0[3][WIDTH*HEIGHT]; |
| 1807 |
|
int with_uv, with_flip; |
| 1808 |
|
int i, j; |
| 1809 |
|
double t; |
| 1810 |
|
int tst, iCrc; |
| 1811 |
|
|
| 1812 |
|
colorspace_init(); |
| 1813 |
|
ieee_reseed(1); |
| 1814 |
|
for(i=0; i<(int)sizeof(Src0); ++i) Src0[0][i] = ieee_rand(0,255); |
| 1815 |
|
|
| 1816 |
|
printf( "\n === test YV2 ===\n" ); |
| 1817 |
|
for(with_flip=0; with_flip<=1; ++with_flip) { |
| 1818 |
|
for(with_uv=0; with_uv<=1; ++with_uv) { |
| 1819 |
|
init_cpu(&cpu_list[0]); |
| 1820 |
|
TEST_YV2(yv12_to_yv12_c, with_uv, with_flip); |
| 1821 |
|
printf(" yv12_to_yv12_c %.3f usec \tcrc32=0x%08x %s\n", |
| 1822 |
|
t, iCrc, (iCrc!=yv12_CRCs[with_flip][with_uv])?"| ERROR": "" ); |
| 1823 |
|
/* if (!with_uv) PRINT_NxN(Dst0[1], WIDTH/2, HEIGHT/2, WIDTH ); */ |
| 1824 |
|
|
| 1825 |
|
#ifdef ARCH_IS_IA32 |
| 1826 |
|
init_cpu(&cpu_list[1]); |
| 1827 |
|
TEST_YV2(yv12_to_yv12_mmx, with_uv, with_flip); |
| 1828 |
|
printf(" yv12_to_yv12_mmx %.3f usec \tcrc32=0x%08x %s\n", |
| 1829 |
|
t, iCrc, (iCrc!=yv12_CRCs[with_flip][with_uv])?"| ERROR": "" ); |
| 1830 |
|
/* if (!with_uv) PRINT_NxN(Dst0[1], WIDTH/2, HEIGHT/2, WIDTH ); */ |
| 1831 |
|
|
| 1832 |
|
TEST_YV2(yv12_to_yv12_xmm, with_uv, with_flip); |
| 1833 |
|
printf(" yv12_to_yv12_xmm %.3f usec \tcrc32=0x%08x %s\n", |
| 1834 |
|
t, iCrc, (iCrc!=yv12_CRCs[with_flip][with_uv])?"| ERROR": "" ); |
| 1835 |
|
#endif |
| 1836 |
|
} |
| 1837 |
|
|
| 1838 |
|
printf( " --- \n" ); |
| 1839 |
|
} |
| 1840 |
|
printf( " ===== \n" ); |
| 1841 |
|
} |
| 1842 |
|
|
| 1843 |
|
#undef WIDTH |
| 1844 |
|
#undef HEIGHT |
| 1845 |
|
#undef ENTER |
| 1846 |
|
#undef LEAVE |
| 1847 |
|
|
| 1848 |
|
/*********************************************************************/ |
| 1849 |
|
|
| 1850 |
|
static uint32_t __inline log2bin_v1(uint32_t value) |
| 1851 |
|
{ |
| 1852 |
|
int n = 0; |
| 1853 |
|
while (value) { |
| 1854 |
|
value >>= 1; |
| 1855 |
|
n++; |
| 1856 |
|
} |
| 1857 |
|
return n; |
| 1858 |
|
} |
| 1859 |
|
|
| 1860 |
|
static const uint8_t log2_tab_16[16] = { 0, 1, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4 }; |
| 1861 |
|
|
| 1862 |
|
static uint32_t __inline log2bin_v2(uint32_t value) |
| 1863 |
|
{ |
| 1864 |
|
int n = 0; |
| 1865 |
|
if (value & 0xffff0000) { |
| 1866 |
|
value >>= 16; |
| 1867 |
|
n += 16; |
| 1868 |
|
} |
| 1869 |
|
if (value & 0xff00) { |
| 1870 |
|
value >>= 8; |
| 1871 |
|
n += 8; |
| 1872 |
|
} |
| 1873 |
|
if (value & 0xf0) { |
| 1874 |
|
value >>= 4; |
| 1875 |
|
n += 4; |
| 1876 |
|
} |
| 1877 |
|
return n + log2_tab_16[value]; |
| 1878 |
|
} |
| 1879 |
|
|
| 1880 |
|
void test_log2bin() |
| 1881 |
|
{ |
| 1882 |
|
const int nb_tests = 3000*speed_ref; |
| 1883 |
|
int n, crc1=0, crc2=0; |
| 1884 |
|
uint32_t s, s0; |
| 1885 |
|
double t1, t2; |
| 1886 |
|
|
| 1887 |
|
t1 = gettime_usec(); |
| 1888 |
|
s0 = (int)(t1*31.241); |
| 1889 |
|
for(s=s0, n=0; n<nb_tests; ++n, s=(s*12363+31)&0x7fffffff) |
| 1890 |
|
crc1 += log2bin_v1(s); |
| 1891 |
|
t1 = (gettime_usec()-t1) / nb_tests; |
| 1892 |
|
|
| 1893 |
|
t2 = gettime_usec(); |
| 1894 |
|
for(s=s0, n=0; n<nb_tests; ++n, s=(s*12363+31)&0x7fffffff) |
| 1895 |
|
crc2 += log2bin_v2(s); |
| 1896 |
|
t2 = (gettime_usec() - t2) / nb_tests; |
| 1897 |
|
|
| 1898 |
|
printf( "log2bin_v1: %.3f sec crc=%d\n", t1, crc1 ); |
| 1899 |
|
printf( "log2bin_v2: %.3f sec crc=%d\n", t2, crc2 ); |
| 1900 |
|
if (crc1!=crc2) printf( " CRC ERROR !\n" ); |
| 1901 |
|
} |
| 1902 |
|
|
| 1903 |
|
/*********************************************************************/ |
| 1904 |
|
|
| 1905 |
|
static void __inline old_gcd(int *num, int *den) |
| 1906 |
|
{ |
| 1907 |
|
int i = *num; |
| 1908 |
|
while (i > 1) { |
| 1909 |
|
if (*num % i == 0 && *den % i == 0) { |
| 1910 |
|
*num /= i; |
| 1911 |
|
*den /= i; |
| 1912 |
|
i = *num; |
| 1913 |
|
continue; |
| 1914 |
|
} |
| 1915 |
|
i--; |
| 1916 |
|
} |
| 1917 |
|
} |
| 1918 |
|
|
| 1919 |
|
static uint32_t gcd(int num, int den) |
| 1920 |
|
{ |
| 1921 |
|
int tmp; |
| 1922 |
|
while( (tmp=num%den) ) { num = den; den = tmp; } |
| 1923 |
|
return den; |
| 1924 |
|
} |
| 1925 |
|
static void __inline new_gcd(int *num, int *den) |
| 1926 |
|
{ |
| 1927 |
|
const int div = gcd(*num, *den); |
| 1928 |
|
if (num) { |
| 1929 |
|
*num /= div; |
| 1930 |
|
*den /= div; |
| 1931 |
|
} |
| 1932 |
|
} |
| 1933 |
|
|
| 1934 |
|
void test_gcd() |
| 1935 |
|
{ |
| 1936 |
|
const int nb_tests = 10*speed_ref; |
| 1937 |
|
int i; |
| 1938 |
|
uint32_t crc1=0, crc2=0; |
| 1939 |
|
uint32_t n0, n, d0, d; |
| 1940 |
|
double t1, t2; |
| 1941 |
|
|
| 1942 |
|
t1 = gettime_usec(); |
| 1943 |
|
n0 = 0xfffff & (int)(t1*31.241); |
| 1944 |
|
d0 = 0xfffff & (int)( ((n0*4123)%17) | 1 ); |
| 1945 |
|
for(n=n0, d=d0, i=0; i<nb_tests; ++i) { |
| 1946 |
|
old_gcd(&n, &d); |
| 1947 |
|
crc1 = (((crc1>>4)^d) + ((crc1<<2)^n) ) & 0xffffff; |
| 1948 |
|
n = d; |
| 1949 |
|
d = (d*12363+31) & 0xffff; |
| 1950 |
|
d |= !d; |
| 1951 |
|
} |
| 1952 |
|
t1 = (gettime_usec()-t1) / nb_tests; |
| 1953 |
|
|
| 1954 |
|
t2 = gettime_usec(); |
| 1955 |
|
for(n=n0, d=d0, i=0; i<nb_tests; ++i) { |
| 1956 |
|
new_gcd(&n, &d); |
| 1957 |
|
crc2 = (((crc2>>4)^d) + ((crc2<<2)^n) ) & 0xffffff; |
| 1958 |
|
n = d; |
| 1959 |
|
d = (d*12363+31) & 0xffff; |
| 1960 |
|
d |= !d; |
| 1961 |
|
} |
| 1962 |
|
t2 = (gettime_usec() - t2) / nb_tests; |
| 1963 |
|
|
| 1964 |
|
printf( "old_gcd: %.3f sec crc=%d\n", t1, crc1 ); |
| 1965 |
|
printf( "new_gcd: %.3f sec crc=%d\n", t2, crc2 ); |
| 1966 |
|
if (crc1!=crc2) printf( " CRC ERROR !\n" ); |
| 1967 |
|
} |
| 1968 |
|
|
| 1969 |
|
/********************************************************************* |
| 1970 |
|
* test compiler |
| 1971 |
|
*********************************************************************/ |
| 1972 |
|
|
| 1973 |
|
void test_compiler() { |
| 1974 |
|
int nb_err = 0; |
| 1975 |
|
int32_t v; |
| 1976 |
|
if (sizeof(uint16_t)<2) { |
| 1977 |
|
printf( "ERROR: sizeof(uint16_t)<2 !!\n" ); |
| 1978 |
|
nb_err++; |
| 1979 |
|
} |
| 1980 |
|
if (sizeof(int16_t)<2) { |
| 1981 |
|
printf( "ERROR: sizeof(int16_t)<2 !!\n" ); |
| 1982 |
|
nb_err++; |
| 1983 |
|
} |
| 1984 |
|
if (sizeof(uint8_t)!=1) { |
| 1985 |
|
printf( "ERROR: sizeof(uint8_t)!=1 !!\n" ); |
| 1986 |
|
nb_err++; |
| 1987 |
|
} |
| 1988 |
|
if (sizeof(int8_t)!=1) { |
| 1989 |
|
printf( "ERROR: sizeof(int8_t)!=1 !!\n" ); |
| 1990 |
|
nb_err++; |
| 1991 |
|
} |
| 1992 |
|
if (sizeof(uint32_t)<4) { |
| 1993 |
|
printf( "ERROR: sizeof(uint32_t)<4 !!\n" ); |
| 1994 |
|
nb_err++; |
| 1995 |
|
} |
| 1996 |
|
if (sizeof(int32_t)<4) { |
| 1997 |
|
printf( "ERROR: sizeof(int32_t)<4 !!\n" ); |
| 1998 |
|
nb_err++; |
| 1999 |
|
} |
| 2000 |
|
/* yes, i know, this test is silly. But better be safe than sorry. :) */ |
| 2001 |
|
for(v=1000; v>=0; v--) { |
| 2002 |
|
if ( (v>>2) != v/4) |
| 2003 |
|
nb_err++; |
| 2004 |
|
} |
| 2005 |
|
for(v=-1000; v!=-1; v++) { |
| 2006 |
|
if ( (v>>2) != (v/4)-!!(v%4)) |
| 2007 |
|
nb_err++; |
| 2008 |
|
} |
| 2009 |
|
if (nb_err!=0) { |
| 2010 |
|
printf( "ERROR! please post your platform/compiler specs to xvid-devel@xvid.org !\n" ); |
| 2011 |
|
} |
| 2012 |
|
} |
| 2013 |
|
|
| 2014 |
|
/********************************************************************* |
| 2015 |
|
* test SSIM functions |
| 2016 |
|
*********************************************************************/ |
| 2017 |
|
|
| 2018 |
|
typedef int (*lumfunc)(uint8_t* ptr, int stride); |
| 2019 |
|
typedef void (*csfunc)(uint8_t* ptro, uint8_t* ptrc, int stride, int lumo, int lumc, int* pdevo, int* pdevc, int* pcorr); |
| 2020 |
|
|
| 2021 |
|
extern int lum_8x8_c(uint8_t* ptr, int stride); |
| 2022 |
|
extern int lum_8x8_mmx(uint8_t* ptr, int stride); |
| 2023 |
|
extern int lum_2x8_c(uint8_t* ptr, int stride); |
| 2024 |
|
extern void consim_c(uint8_t* ptro, uint8_t* ptrc, int stride, int lumo, int lumc, int* pdevo, int* pdevc, int* pcorr); |
| 2025 |
|
extern void consim_mmx(uint8_t* ptro, uint8_t* ptrc, int stride, int lumo, int lumc, int* pdevo, int* pdevc, int* pcorr); |
| 2026 |
|
extern void consim_sse2(uint8_t* ptro, uint8_t* ptrc, int stride, int lumo, int lumc, int* pdevo, int* pdevc, int* pcorr); |
| 2027 |
|
|
| 2028 |
|
void test_SSIM() |
| 2029 |
|
{ |
| 2030 |
|
const int nb_tests = 3000*speed_ref; |
| 2031 |
|
int tst; |
| 2032 |
|
CPU *cpu; |
| 2033 |
|
int i; |
| 2034 |
|
int devs[3]; |
| 2035 |
|
long lumo, lumc; |
| 2036 |
|
DECLARE_ALIGNED_MATRIX(Ref1, 16, 16, uint8_t, 16); |
| 2037 |
|
DECLARE_ALIGNED_MATRIX(Ref2, 16, 16, uint8_t, 16); |
| 2038 |
|
lumfunc lum8x8; |
| 2039 |
|
lumfunc lum2x8; |
| 2040 |
|
csfunc csim; |
| 2041 |
|
|
| 2042 |
|
ieee_reseed(1); |
| 2043 |
|
printf( "\n ====== test SSIM ======\n" ); |
| 2044 |
|
for(i=0; i<16*16;++i) { |
| 2045 |
|
long v1, v2; |
| 2046 |
|
v1 = ieee_rand(-256, 511); |
| 2047 |
|
v2 = ieee_rand(-256, 511); |
| 2048 |
|
Ref1[i] = (v1<0) ? 0 : (v1>255) ? 255 : v1; |
| 2049 |
|
Ref2[i] = (v2<0) ? 0 : (v2>255) ? 255 : v2; |
| 2050 |
|
} |
| 2051 |
|
lumc = ieee_rand(0, 255); |
| 2052 |
|
lumo = ieee_rand(0, 255); |
| 2053 |
|
|
| 2054 |
|
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
| 2055 |
|
{ |
| 2056 |
|
double t; |
| 2057 |
|
int m; |
| 2058 |
|
if (!init_cpu(cpu)) |
| 2059 |
|
continue; |
| 2060 |
|
lum8x8 = lum_8x8_c; |
| 2061 |
|
lum2x8 = lum_2x8_c; |
| 2062 |
|
csim = consim_c; |
| 2063 |
|
if (cpu->cpu & XVID_CPU_MMX){ |
| 2064 |
|
lum8x8 = lum_8x8_mmx; |
| 2065 |
|
csim = consim_mmx; |
| 2066 |
|
} |
| 2067 |
|
if (cpu->cpu & XVID_CPU_MMX){ |
| 2068 |
|
csim = consim_sse2; |
| 2069 |
|
} |
| 2070 |
|
|
| 2071 |
|
t = gettime_usec(); |
| 2072 |
|
emms(); |
| 2073 |
|
for(tst=0; tst<nb_tests; ++tst) m = lum8x8(Ref1, 16); |
| 2074 |
|
emms(); |
| 2075 |
|
t = (gettime_usec() - t) / nb_tests; |
| 2076 |
|
printf("%s - ssim-lum8x8 %.3f usec m=%d %s\n", |
| 2077 |
|
cpu->name, t, m, |
| 2078 |
|
(m!=8230)?"| ERROR": "" ); |
| 2079 |
|
|
| 2080 |
|
t = gettime_usec(); |
| 2081 |
|
emms(); |
| 2082 |
|
for(tst=0; tst<nb_tests; ++tst) m = lum2x8(Ref1+8, 16); |
| 2083 |
|
emms(); |
| 2084 |
|
t = (gettime_usec() - t) / nb_tests; |
| 2085 |
|
printf("%s - ssim-lum2x8 %.3f usec m=%d %s\n", |
| 2086 |
|
cpu->name, t, m, |
| 2087 |
|
(m!=681)?"| ERROR": "" ); |
| 2088 |
|
|
| 2089 |
|
t = gettime_usec(); |
| 2090 |
|
emms(); |
| 2091 |
|
for(tst=0; tst<nb_tests; ++tst) csim(Ref1, Ref2, 16, lumo, lumc, devs+0, devs+1, devs+2); |
| 2092 |
|
emms(); |
| 2093 |
|
t = (gettime_usec() - t) / nb_tests; |
| 2094 |
|
printf("%s - ssim-consim %.3f usec devs=[0x%x 0x%x 0x%x] %s\n", |
| 2095 |
|
cpu->name, t, devs[0], devs[1], devs[2], |
| 2096 |
|
(devs[0]!=0x1bdf0f || devs[1]!=0x137258 || devs[2]!=0xcdb13)?"| ERROR": "" ); |
| 2097 |
|
printf( " --- \n" ); |
| 2098 |
|
} |
| 2099 |
|
} |
| 2100 |
|
|
| 2101 |
|
/********************************************************************* |
| 2102 |
|
* test bitstream functions |
| 2103 |
|
*********************************************************************/ |
| 2104 |
|
|
| 2105 |
|
#define BIT_BUF_SIZE 2000 |
| 2106 |
|
|
| 2107 |
|
static void test_bits() |
| 2108 |
|
{ |
| 2109 |
|
const int nb_tests = 50*speed_ref; |
| 2110 |
|
int tst; |
| 2111 |
|
uint32_t Crc; |
| 2112 |
|
uint8_t Buf[BIT_BUF_SIZE]; |
| 2113 |
|
uint32_t Extracted[BIT_BUF_SIZE*8]; /* worst case: bits read 1 by 1 */ |
| 2114 |
|
int Lens[BIT_BUF_SIZE*8]; |
| 2115 |
|
double t1; |
| 2116 |
|
|
| 2117 |
|
|
| 2118 |
|
printf( "\n === test bitstream ===\n" ); |
| 2119 |
|
ieee_reseed(1); |
| 2120 |
|
Crc = 0; |
| 2121 |
|
|
| 2122 |
|
t1 = gettime_usec(); |
| 2123 |
|
for(tst=0; tst<nb_tests; ++tst) { |
| 2124 |
|
Bitstream bs; |
| 2125 |
|
int m, m2, l, l2; |
| 2126 |
|
|
| 2127 |
|
for(l=0; l<BIT_BUF_SIZE; ++l) |
| 2128 |
|
Buf[l] = (uint8_t)ieee_rand(0,255); |
| 2129 |
|
|
| 2130 |
|
l = BIT_BUF_SIZE - ieee_rand(1,BIT_BUF_SIZE/10); |
| 2131 |
|
BitstreamInit(&bs, (void*)(Buf+BIT_BUF_SIZE-l), l); |
| 2132 |
|
|
| 2133 |
|
|
| 2134 |
|
BitstreamReset(&bs); |
| 2135 |
|
for(l2=l*8, m=0; l2>0; m++) { |
| 2136 |
|
const int b = ieee_rand(1,32); |
| 2137 |
|
Lens[m] = b; |
| 2138 |
|
l2 -= b; |
| 2139 |
|
if (l2<0) break; |
| 2140 |
|
Extracted[m] = BitstreamShowBits(&bs, b); |
| 2141 |
|
BitstreamSkip(&bs, b); |
| 2142 |
|
// printf( "<= %d: %d 0x%x\n", m, b, Extracted[m]); |
| 2143 |
|
} |
| 2144 |
|
|
| 2145 |
|
BitstreamReset(&bs); |
| 2146 |
|
for(m2=0; m2<m; ++m2) { |
| 2147 |
|
const int b = Lens[m2]; |
| 2148 |
|
const uint32_t v = BitstreamGetBits(&bs, b); |
| 2149 |
|
Crc |= (v!=Extracted[m2]); |
| 2150 |
|
// printf( "=> %d: %d 0x%x %c\n", m2, b, v, " *"[Crc]); |
| 2151 |
|
} |
| 2152 |
|
} |
| 2153 |
|
t1 = (gettime_usec() - t1) / nb_tests; |
| 2154 |
|
printf(" test_bits %.3f usec %s\n", t1, (Crc!=0)?"| ERROR": "" ); |
| 2155 |
|
} |
| 2156 |
|
|
| 2157 |
|
/********************************************************************* |
| 2158 |
* main |
* main |
| 2159 |
*********************************************************************/ |
*********************************************************************/ |
| 2160 |
|
|
| 2161 |
int main(int argc, char *argv[]) |
static void arg_missing(const char *opt) |
| 2162 |
|
{ |
| 2163 |
|
printf( "missing argument after option '%s'\n", opt); |
| 2164 |
|
exit(-1); |
| 2165 |
|
} |
| 2166 |
|
|
| 2167 |
|
int main(int argc, const char *argv[]) |
| 2168 |
|
{ |
| 2169 |
|
int c, what = 0; |
| 2170 |
|
int width, height; |
| 2171 |
|
uint32_t chksum = 0; |
| 2172 |
|
const char * test_bitstream = 0; |
| 2173 |
|
|
| 2174 |
|
cpu_mask = 0; // default => will use autodectect |
| 2175 |
|
for(c=1; c<argc; ++c) |
| 2176 |
{ |
{ |
| 2177 |
int what = 0; |
if (!strcmp(argv[c], "-v")) verbose++; |
| 2178 |
if (argc>1) what = atoi(argv[1]); |
else if (!strcmp(argv[c], "-c")) cpu_mask = 0 /* PLAIN_C */ | XVID_CPU_FORCE; |
| 2179 |
|
else if (!strcmp(argv[c], "-mmx")) cpu_mask = XVID_CPU_MMX | XVID_CPU_FORCE; |
| 2180 |
|
else if (!strcmp(argv[c], "-mmxext")) cpu_mask = XVID_CPU_MMXEXT | XVID_CPU_MMX | XVID_CPU_FORCE; |
| 2181 |
|
else if (!strcmp(argv[c], "-sse2")) cpu_mask = XVID_CPU_SSE2 | XVID_CPU_MMXEXT | XVID_CPU_MMX | XVID_CPU_FORCE; |
| 2182 |
|
else if (!strcmp(argv[c], "-3dnow")) cpu_mask = XVID_CPU_3DNOW | XVID_CPU_FORCE; |
| 2183 |
|
else if (!strcmp(argv[c], "-3dnowe")) cpu_mask = XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT | XVID_CPU_FORCE; |
| 2184 |
|
else if (!strcmp(argv[c], "-altivec")) cpu_mask = XVID_CPU_ALTIVEC | XVID_CPU_FORCE; |
| 2185 |
|
else if (!strcmp(argv[c], "-spd")) { |
| 2186 |
|
if (++c==argc) arg_missing( argv[argc-1] ); |
| 2187 |
|
speed_ref = atoi(argv[c]); |
| 2188 |
|
} |
| 2189 |
|
else if (argv[c][0]!='-') { |
| 2190 |
|
what = atoi(argv[c]); |
| 2191 |
|
if (what==9) { |
| 2192 |
|
if (c+4>argc) { |
| 2193 |
|
printf("usage: %s %d bitstream width height (checksum)\n", argv[0], what); |
| 2194 |
|
exit(-1); |
| 2195 |
|
} |
| 2196 |
|
test_bitstream = argv[++c]; |
| 2197 |
|
width = atoi(argv[++c]); |
| 2198 |
|
height = atoi(argv[++c]); |
| 2199 |
|
if (c+1<argc && argv[c+1][0]!='-') { |
| 2200 |
|
if (sscanf(argv[c+1], "0x%x", &chksum)!=1) { |
| 2201 |
|
printf( "can't read checksum value.\n" ); |
| 2202 |
|
exit(-1); |
| 2203 |
|
} |
| 2204 |
|
else c++; |
| 2205 |
|
} |
| 2206 |
|
// printf( "[%s] %dx%d (0x%.8x)\n", test_bitstream, width, height, chksum); |
| 2207 |
|
} |
| 2208 |
|
} |
| 2209 |
|
else { |
| 2210 |
|
printf( "unrecognized option '%s'\n", argv[c]); |
| 2211 |
|
exit(-1); |
| 2212 |
|
} |
| 2213 |
|
} |
| 2214 |
|
|
| 2215 |
|
|
| 2216 |
if (what==0 || what==1) test_dct(); |
if (what==0 || what==1) test_dct(); |
| 2217 |
if (what==0 || what==2) test_mb(); |
if (what==0 || what==2) test_mb(); |
| 2218 |
if (what==0 || what==3) test_sad(); |
if (what==0 || what==3) test_sad(); |
| 2219 |
if (what==0 || what==4) test_transfer(); |
if (what==0 || what==4) test_transfer(); |
| 2220 |
if (what==0 || what==5) test_quant(); |
if (what==0 || what==5) test_quant(); |
| 2221 |
if (what==0 || what==6) test_cbp(); |
if (what==0 || what==6) test_cbp(); |
| 2222 |
|
if (what==0 || what==10) test_sse(); |
| 2223 |
|
if (what==0 || what==11) test_log2bin(); |
| 2224 |
|
if (what==0 || what==12) test_gcd(); |
| 2225 |
|
if (what==0 || what==13) test_compiler(); |
| 2226 |
|
if (what==0 || what==14) test_yuv(); |
| 2227 |
|
if (what==0 || what==15) test_SSIM(); |
| 2228 |
|
if (what==0 || what==16) test_yuv2(); |
| 2229 |
|
if (what==0 || what==17) test_bits(); |
| 2230 |
|
|
| 2231 |
if (what==7) { |
if (what==7) { |
| 2232 |
test_IEEE1180_compliance(-256, 255, 1); |
test_IEEE1180_compliance(-256, 255, 1); |
| 2238 |
} |
} |
| 2239 |
if (what==8) test_dct_saturation(-256, 255); |
if (what==8) test_dct_saturation(-256, 255); |
| 2240 |
|
|
| 2241 |
if (what==9) { |
if (test_bitstream) |
| 2242 |
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)); |
|
|
} |
|
|
|
|
| 2243 |
if (what==-1) { |
if (what==-1) { |
| 2244 |
test_dct_precision_diffs(); |
test_dct_precision_diffs(); |
| 2245 |
test_bugs1(); |
test_bugs1(); |
| 2247 |
if (what==-2) |
if (what==-2) |
| 2248 |
test_quant_bug(); |
test_quant_bug(); |
| 2249 |
|
|
| 2250 |
if (what >= 0 && what <= 6) { |
if ((what >= 0 && what <= 6) || what == 10) { |
| 2251 |
printf("\n\n" |
printf("\n\n" |
| 2252 |
"NB: If a function isn't optimised for a specific set of intructions,\n" |
"NB: If a function isn't optimised for a specific set of intructions,\n" |
| 2253 |
" a C function is used instead. So don't panic if some functions\n" |
" a C function is used instead. So don't panic if some functions\n" |
| 2267 |
return 0; |
return 0; |
| 2268 |
} |
} |
| 2269 |
|
|
| 2270 |
/********************************************************************* |
/*********************************************************************/ |
|
* '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 |
|
|
|
|
|
*********************************************************************/ |
|
Parent Directory
| 
Revision Log
| 
Patch