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 |
83 |
#else |
#else |
84 |
clock_t clk; |
clock_t clk; |
85 |
clk = clock(); |
clk = clock(); |
86 |
return clk * 1000000 / CLOCKS_PER_SEC; |
return clk * 1000. / CLOCKS_PER_SEC; /* clock() returns time in Milliseconds */ |
87 |
#endif |
#endif |
88 |
} |
} |
89 |
|
|
112 |
unsigned int cpu; |
unsigned int cpu; |
113 |
} CPU; |
} CPU; |
114 |
|
|
115 |
CPU cpu_list[] = |
CPU cpu_list[] = { |
116 |
{ { "PLAINC", 0 } |
{ "PLAINC ", 0 }, |
117 |
#ifdef ARCH_IS_IA32 |
#ifdef ARCH_IS_IA32 |
118 |
, { "MMX ", XVID_CPU_MMX } |
{ "MMX ", XVID_CPU_MMX }, |
119 |
, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
{ "MMXEXT ", XVID_CPU_MMXEXT | XVID_CPU_MMX }, |
120 |
, { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } |
{ "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX }, |
121 |
, { "3DNOW ", XVID_CPU_3DNOW } |
{ "3DNOW ", XVID_CPU_3DNOW }, |
122 |
, { "3DNOWE", XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT } |
{ "3DNOWE ", XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT }, |
123 |
#endif |
#endif |
124 |
#ifdef ARCH_IS_PPC |
#ifdef ARCH_IS_PPC |
125 |
, { "ALTIVEC", XVID_CPU_ALTIVEC } |
{ "ALTIVEC", XVID_CPU_ALTIVEC }, |
126 |
#endif |
#endif |
127 |
#ifdef ARCH_IS_X86_64 |
#ifdef ARCH_IS_X86_64 |
128 |
, { "X86_64", XVID_CPU_ASM} |
{ "X86_64 ", XVID_CPU_ASM}, |
129 |
#endif |
#endif |
130 |
//, { "IA64 ", XVID_CPU_IA64 } |
#ifdef ARCH_IS_IA64 |
131 |
//, { "TSC ", XVID_CPU_TSC } |
// { "IA64 ", XVID_CPU_IA64 }, |
|
, { 0, 0 } }; |
|
|
|
|
|
CPU cpu_short_list[] = |
|
|
{ { "PLAINC", 0 } |
|
|
#ifdef ARCH_IS_IA32 |
|
|
, { "MMX ", XVID_CPU_MMX } |
|
|
//, { "MMXEXT", XVID_CPU_MMXEXT | XVID_CPU_MMX } |
|
|
#endif |
|
|
//, { "IA64 ", XVID_CPU_IA64 } |
|
|
, { 0, 0 } }; |
|
|
|
|
|
CPU cpu_short_list2[] = |
|
|
{ { "PLAINC", 0 } |
|
|
#ifdef ARCH_IS_IA32 |
|
|
, { "MMX ", XVID_CPU_MMX } |
|
|
, { "SSE2 ", XVID_CPU_SSE2 | XVID_CPU_MMX } |
|
132 |
#endif |
#endif |
133 |
, { 0, 0 } }; |
// { "TSC ", XVID_CPU_TSC }, |
134 |
|
{ 0, 0 } |
135 |
|
}; |
136 |
|
|
137 |
|
|
138 |
int init_cpu(CPU *cpu) |
int init_cpu(CPU *cpu) |
245 |
}; |
}; |
246 |
|
|
247 |
uint32_t |
uint32_t |
248 |
calc_crc(uint8_t *mem, int len, uint32_t initial) |
calc_crc(uint8_t *mem, int len, uint32_t crc) |
249 |
{ |
{ |
|
|
|
|
register unsigned int crc; |
|
|
|
|
|
crc = initial; |
|
|
|
|
250 |
while( len >= 8) { |
while( len >= 8) { |
251 |
DO8(mem, crc); |
DO8(mem, crc); |
252 |
len -= 8; |
len -= 8; |
257 |
len--; |
len--; |
258 |
} |
} |
259 |
|
|
260 |
return(crc); |
return crc; |
261 |
|
} |
262 |
|
|
263 |
|
void byte_swap(uint8_t *mem, int len, int element_size) { |
264 |
|
#ifdef ARCH_IS_BIG_ENDIAN |
265 |
|
int i; |
266 |
|
|
267 |
|
if(element_size == 1) { |
268 |
|
/* No need to swap */ |
269 |
|
} else if(element_size == 2) { |
270 |
|
uint8_t temp[2]; |
271 |
|
|
272 |
|
for(i=0; i < (len/2); i++ ) { |
273 |
|
temp[0] = mem[0]; |
274 |
|
temp[1] = mem[1]; |
275 |
|
mem[0] = temp[1]; |
276 |
|
mem[1] = temp[0]; |
277 |
|
|
278 |
|
mem += 2; |
279 |
|
} |
280 |
|
} else if(element_size == 4) { |
281 |
|
uint8_t temp[4]; |
282 |
|
|
283 |
|
for(i=0; i < (len/4); i++ ) { |
284 |
|
temp[0] = mem[0]; |
285 |
|
temp[1] = mem[1]; |
286 |
|
temp[2] = mem[2]; |
287 |
|
temp[3] = mem[3]; |
288 |
|
mem[0] = temp[3]; |
289 |
|
mem[1] = temp[2]; |
290 |
|
mem[2] = temp[1]; |
291 |
|
mem[3] = temp[0]; |
292 |
|
|
293 |
|
mem += 4; |
294 |
|
} |
295 |
|
} else { |
296 |
|
printf("ERROR: byte_swap unsupported element_size(%u)\n", element_size); |
297 |
|
} |
298 |
|
#endif |
299 |
} |
} |
300 |
|
|
301 |
/********************************************************************* |
/********************************************************************* |
518 |
(iCrc!=8107)?"| ERROR": "" ); |
(iCrc!=8107)?"| ERROR": "" ); |
519 |
#endif |
#endif |
520 |
|
|
521 |
|
/* New functions for field prediction by CK 1.10.2005 */ |
522 |
|
#pragma NEW8X4 |
523 |
|
TEST_MB(interpolate8x4_halfpel_h, 0); |
524 |
|
printf("%s - interpfield-h -round0 %.3f usec crc32=0x%08x %s\n", |
525 |
|
cpu->name, t, iCrc, |
526 |
|
(iCrc!=0x9538d6df)?"| ERROR": "" ); |
527 |
|
|
528 |
|
TEST_MB(interpolate8x4_halfpel_h, 1); |
529 |
|
printf("%s - round1 %.3f usec crc32=0x%08x %s\n", |
530 |
|
cpu->name, t, iCrc, |
531 |
|
(iCrc!=0xde5f1db4)?"| ERROR": "" ); |
532 |
|
|
533 |
|
|
534 |
|
TEST_MB(interpolate8x4_halfpel_v, 0); |
535 |
|
printf("%s - interpfield- v-round0 %.3f usec crc32=0x%08x %s\n", |
536 |
|
cpu->name, t, iCrc, |
537 |
|
(iCrc!=0xea5a69ef)?"| ERROR": "" ); |
538 |
|
|
539 |
|
TEST_MB(interpolate8x4_halfpel_v, 1); |
540 |
|
printf("%s - round1 %.3f usec crc32=0x%08x %s\n", |
541 |
|
cpu->name, t, iCrc, |
542 |
|
(iCrc!=0x4f10ec0f)?"| ERROR": "" ); |
543 |
|
|
544 |
|
|
545 |
|
TEST_MB(interpolate8x4_halfpel_hv, 0); |
546 |
|
printf("%s - interpfield-hv-round0 %.3f usec crc32=0x%08x %s\n", |
547 |
|
cpu->name, t, iCrc, |
548 |
|
(iCrc!=0xf97ee367)?"| ERROR": "" ); |
549 |
|
|
550 |
|
TEST_MB(interpolate8x4_halfpel_hv, 1); |
551 |
|
printf("%s - round1 %.3f usec crc32=0x%08x %s\n", |
552 |
|
cpu->name, t, iCrc, |
553 |
|
(iCrc!=0xb6a9f581)?"| ERROR": "" ); |
554 |
|
/* End of 8x4 functions */ |
555 |
|
|
556 |
printf( " --- \n" ); |
printf( " --- \n" ); |
557 |
} |
} |
558 |
} |
} |
586 |
} \ |
} \ |
587 |
emms(); \ |
emms(); \ |
588 |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
589 |
s = calc_crc((uint8_t*)(DST), sizeof((DST)), CRC32_INITIAL) |
byte_swap((uint8_t*)(DST), 8*32*sizeof((DST)[0]), sizeof((DST)[0])); \ |
590 |
|
s = calc_crc((uint8_t*)(DST), 8*32*sizeof((DST)[0]), CRC32_INITIAL) |
591 |
|
|
592 |
#define TEST_TRANSFER(FUNC, DST, SRC) \ |
#define TEST_TRANSFER(FUNC, DST, SRC) \ |
593 |
TEST_TRANSFER_BEGIN(DST); \ |
TEST_TRANSFER_BEGIN(DST); \ |
613 |
} \ |
} \ |
614 |
emms(); \ |
emms(); \ |
615 |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
t = (gettime_usec()-t -overhead) / nb_tests;\ |
616 |
s = calc_crc((uint8_t*)(DST), sizeof((DST)), CRC32_INITIAL) |
byte_swap((uint8_t*)(DST), 8*32*sizeof((DST)[0]), sizeof((DST)[0])); \ |
617 |
|
s = calc_crc((uint8_t*)(DST), 8*32*sizeof((DST)[0]), CRC32_INITIAL) |
618 |
|
|
619 |
#define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ |
#define TEST_TRANSFER2(FUNC, DST, SRC, R1) \ |
620 |
TEST_TRANSFER2_BEGIN(DST,SRC); \ |
TEST_TRANSFER2_BEGIN(DST,SRC); \ |
631 |
const int nb_tests = 4000*speed_ref; |
const int nb_tests = 4000*speed_ref; |
632 |
int i; |
int i; |
633 |
CPU *cpu; |
CPU *cpu; |
634 |
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]; |
635 |
int16_t Src16[8*32], Dst16[8*32]; |
// int16_t Src16[8*32], Dst16[8*32]; |
636 |
|
DECLARE_ALIGNED_MATRIX(Src8, 8, 32, uint8_t, CACHE_LINE); |
637 |
|
DECLARE_ALIGNED_MATRIX(Dst8, 8, 32, uint8_t, CACHE_LINE); |
638 |
|
DECLARE_ALIGNED_MATRIX(Ref1, 8, 32, uint8_t, CACHE_LINE); |
639 |
|
DECLARE_ALIGNED_MATRIX(Ref2, 8, 32, uint8_t, CACHE_LINE); |
640 |
|
DECLARE_ALIGNED_MATRIX(Src16, 8, 32, uint16_t, CACHE_LINE); |
641 |
|
DECLARE_ALIGNED_MATRIX(Dst16, 8, 32, uint16_t, CACHE_LINE); |
642 |
|
|
643 |
printf( "\n === test transfer ===\n" ); |
printf( "\n === test transfer ===\n" ); |
644 |
|
|
660 |
cpu->name, t, s, |
cpu->name, t, s, |
661 |
(s!=0xee7ccbb4)?"| ERROR": ""); |
(s!=0xee7ccbb4)?"| ERROR": ""); |
662 |
|
|
663 |
|
/* New functions for field prediction by CK 1.10.2005 */ |
664 |
|
#pragma NEW8X4 |
665 |
|
TEST_TRANSFER(transfer8x4_copy, Dst8, Src8); |
666 |
|
printf("%s - 8to4 %.3f usec crc32=0x%08x %s\n", |
667 |
|
cpu->name, t, s, |
668 |
|
(s!=0xbb9c3db5)?"| ERROR": ""); |
669 |
|
/* End of new functions */ |
670 |
|
|
671 |
TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
TEST_TRANSFER(transfer8x8_copy, Dst8, Src8); |
672 |
printf("%s - 8to8 %.3f usec crc32=0x%08x %s\n", |
printf("%s - 8to8 %.3f usec crc32=0x%08x %s\n", |
673 |
cpu->name, t, s, |
cpu->name, t, s, |
681 |
TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); |
TEST_TRANSFER2(transfer_8to16sub, Dst16, Src8, Ref1); |
682 |
{ |
{ |
683 |
int s1, s2; |
int s1, s2; |
684 |
s1 = calc_crc((uint8_t*)Dst16, sizeof(Dst16), CRC32_INITIAL); |
s1 = calc_crc((uint8_t*)Dst16, 8*32*sizeof(Dst16[0]), CRC32_INITIAL); |
685 |
s2 = calc_crc((uint8_t*)Src8, sizeof(Src8), CRC32_INITIAL); |
s2 = calc_crc((uint8_t*)Src8, 8*32*sizeof(Src8[0]), CRC32_INITIAL); |
686 |
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", |
687 |
cpu->name, t, s1, s2, |
cpu->name, t, s1, s2, |
688 |
(s1!=0xa1e07163)?"| ERROR1": "", |
(s1!=0xa1e07163)?"| ERROR1": "", |
711 |
for(q=1; q<=max_Q; ++q) { \ |
for(q=1; q<=max_Q; ++q) { \ |
712 |
for(tst=0; tst<nb_tests; ++tst) \ |
for(tst=0; tst<nb_tests; ++tst) \ |
713 |
(FUNC)((DST), (SRC), q, mpeg_quant_matrices); \ |
(FUNC)((DST), (SRC), q, mpeg_quant_matrices); \ |
714 |
s = calc_crc((uint8_t*)(DST), 64*sizeof(int16_t), s); \ |
byte_swap((uint8_t*)(DST), 64*sizeof((DST)[0]), sizeof((DST)[0])); \ |
715 |
|
s = calc_crc((uint8_t*)(DST), 64*sizeof((DST)[0]), s); \ |
716 |
} \ |
} \ |
717 |
emms(); \ |
emms(); \ |
718 |
} \ |
} \ |
727 |
for(q=1; q<=max_Q; ++q) { \ |
for(q=1; q<=max_Q; ++q) { \ |
728 |
for(tst=0; tst<nb_tests; ++tst) \ |
for(tst=0; tst<nb_tests; ++tst) \ |
729 |
(FUNC)((DST), (SRC), q, q, mpeg_quant_matrices); \ |
(FUNC)((DST), (SRC), q, q, mpeg_quant_matrices); \ |
730 |
s = calc_crc((uint8_t*)(DST), 64*sizeof(int16_t), s); \ |
byte_swap((uint8_t*)(DST), 64*sizeof((DST)[0]), sizeof((DST)[0])); \ |
731 |
|
s = calc_crc((uint8_t*)(DST), 64*sizeof((DST)[0]), s); \ |
732 |
} \ |
} \ |
733 |
emms(); \ |
emms(); \ |
734 |
} \ |
} \ |
948 |
* test non-zero AC counting |
* test non-zero AC counting |
949 |
*********************************************************************/ |
*********************************************************************/ |
950 |
|
|
951 |
#define TEST_CBP(FUNC, SRC) \ |
#define TEST_CBP(FUNC, SRC, NB) \ |
952 |
t = gettime_usec(); \ |
t = gettime_usec(); \ |
953 |
emms(); \ |
emms(); \ |
954 |
for(tst=0; tst<nb_tests; ++tst) { \ |
for(tst=0; tst<NB; ++tst) { \ |
955 |
cbp = (FUNC)((SRC)); \ |
cbp = (FUNC)((SRC)); \ |
956 |
} \ |
} \ |
957 |
emms(); \ |
emms(); \ |
960 |
void test_cbp() |
void test_cbp() |
961 |
{ |
{ |
962 |
const int nb_tests = 10000*speed_ref; |
const int nb_tests = 10000*speed_ref; |
963 |
int i; |
int i, n, m; |
964 |
CPU *cpu; |
CPU *cpu; |
965 |
DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src1, 6, 64, int16_t, 16); |
966 |
DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src2, 6, 64, int16_t, 16); |
967 |
DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src3, 6, 64, int16_t, 16); |
968 |
DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16); |
DECLARE_ALIGNED_MATRIX(Src4, 6, 64, int16_t, 16); |
969 |
|
DECLARE_ALIGNED_MATRIX(Src5, 6, 64, int16_t, 16); |
970 |
|
|
971 |
printf( "\n ===== test cbp =====\n" ); |
printf( "\n ===== test cbp =====\n" ); |
972 |
|
|
975 |
Src2[i] = (i<3*64); /* half-full */ |
Src2[i] = (i<3*64); /* half-full */ |
976 |
Src3[i] = ((i+32)>3*64); |
Src3[i] = ((i+32)>3*64); |
977 |
Src4[i] = (i==(3*64+2) || i==(5*64+9)); |
Src4[i] = (i==(3*64+2) || i==(5*64+9)); |
978 |
|
Src5[i] = ieee_rand(0,1) ? -1 : 1; /* +/- test */ |
979 |
} |
} |
980 |
|
|
981 |
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
for(cpu = cpu_list; cpu->name!=0; ++cpu) |
986 |
if (!init_cpu(cpu)) |
if (!init_cpu(cpu)) |
987 |
continue; |
continue; |
988 |
|
|
989 |
TEST_CBP(calc_cbp, Src1); |
TEST_CBP(calc_cbp, Src1, nb_tests); |
990 |
printf("%s - calc_cbp#1 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#1 %.3f usec cbp=0x%02x %s\n", |
991 |
cpu->name, t, cbp, (cbp!=0x15)?"| ERROR": ""); |
cpu->name, t, cbp, (cbp!=0x15)?"| ERROR": ""); |
992 |
TEST_CBP(calc_cbp, Src2); |
TEST_CBP(calc_cbp, Src2, nb_tests); |
993 |
printf("%s - calc_cbp#2 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#2 %.3f usec cbp=0x%02x %s\n", |
994 |
cpu->name, t, cbp, (cbp!=0x38)?"| ERROR": ""); |
cpu->name, t, cbp, (cbp!=0x38)?"| ERROR": ""); |
995 |
TEST_CBP(calc_cbp, Src3); |
TEST_CBP(calc_cbp, Src3, nb_tests); |
996 |
printf("%s - calc_cbp#3 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#3 %.3f usec cbp=0x%02x %s\n", |
997 |
cpu->name, t, cbp, (cbp!=0x0f)?"| ERROR": "" ); |
cpu->name, t, cbp, (cbp!=0x0f)?"| ERROR": "" ); |
998 |
TEST_CBP(calc_cbp, Src4); |
TEST_CBP(calc_cbp, Src4, nb_tests); |
999 |
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
1000 |
cpu->name, t, cbp, (cbp!=0x05)?"| ERROR": "" ); |
cpu->name, t, cbp, (cbp!=0x05)?"| ERROR": "" ); |
1001 |
|
TEST_CBP(calc_cbp, Src5, nb_tests); |
1002 |
|
printf("%s - calc_cbp#4 %.3f usec cbp=0x%02x %s\n", |
1003 |
|
cpu->name, t, cbp, (cbp!=0x3f)?"| ERROR": "" ); |
1004 |
printf( " --- \n" ); |
printf( " --- \n" ); |
1005 |
} |
} |
1006 |
|
|
1007 |
|
for(cpu = cpu_list; cpu->name!=0; ++cpu) /* bench suggested by Carlo (carlo dot bramix at libero dot it) */ |
1008 |
|
{ |
1009 |
|
double t; |
1010 |
|
int tst, cbp, err; |
1011 |
|
|
1012 |
|
if (!init_cpu(cpu)) |
1013 |
|
continue; |
1014 |
|
|
1015 |
|
err = 0; |
1016 |
|
for(n=0; n<6; ++n) |
1017 |
|
{ |
1018 |
|
for(m=0; m<64; ++m) |
1019 |
|
{ |
1020 |
|
for(i=0; i<6*64; ++i) |
1021 |
|
Src1[i] = (i== (m + n*64)); |
1022 |
|
|
1023 |
|
TEST_CBP(calc_cbp, Src1, 1); |
1024 |
|
if (cbp!= (((m!=0)<<(5-n)))) |
1025 |
|
{ |
1026 |
|
printf( "%s - calc_cbp#5: ERROR at pos %d / %d!\n", cpu->name, n, m); |
1027 |
|
err = 1; |
1028 |
|
break; |
1029 |
|
} |
1030 |
|
} |
1031 |
|
} |
1032 |
|
if (!err) |
1033 |
|
printf( " %s - calc_cbp#5 : OK\n", cpu->name ); |
1034 |
|
|
1035 |
|
} |
1036 |
} |
} |
1037 |
|
|
1038 |
/********************************************************************* |
/********************************************************************* |
1426 |
* measure raw decoding speed |
* measure raw decoding speed |
1427 |
*********************************************************************/ |
*********************************************************************/ |
1428 |
|
|
1429 |
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) |
1430 |
{ |
{ |
1431 |
FILE *f = 0; |
FILE *f = 0; |
1432 |
void *dechandle = 0; |
void *dechandle = 0; |
1437 |
double t = 0.; |
double t = 0.; |
1438 |
int nb = 0; |
int nb = 0; |
1439 |
uint8_t *buf = 0; |
uint8_t *buf = 0; |
1440 |
uint8_t *rgb_out = 0; |
uint8_t *yuv_out = 0; |
1441 |
int buf_size, pos; |
int buf_size, pos; |
1442 |
uint32_t chksum = 0; |
uint32_t chksum = 0; |
1443 |
|
int bps = (width+31) & ~31; |
1444 |
|
|
1445 |
memset(&xinit, 0, sizeof(xinit)); |
memset(&xinit, 0, sizeof(xinit)); |
1446 |
xinit.cpu_flags = XVID_CPU_MMX | XVID_CPU_FORCE; |
xinit.cpu_flags = cpu_mask; |
1447 |
xinit.version = XVID_VERSION; |
xinit.version = XVID_VERSION; |
1448 |
xvid_global(NULL, 0, &xinit, NULL); |
xvid_global(NULL, 0, &xinit, NULL); |
1449 |
|
|
1453 |
xparam.version = XVID_VERSION; |
xparam.version = XVID_VERSION; |
1454 |
xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
xerr = xvid_decore(NULL, XVID_DEC_CREATE, &xparam, NULL); |
1455 |
if (xerr==XVID_ERR_FAIL) { |
if (xerr==XVID_ERR_FAIL) { |
1456 |
printf("can't init decoder (err=%d)\n", xerr); |
printf("ERROR: can't init decoder (err=%d)\n", xerr); |
1457 |
return; |
return; |
1458 |
} |
} |
1459 |
dechandle = xparam.handle; |
dechandle = xparam.handle; |
1461 |
|
|
1462 |
f = fopen(name, "rb"); |
f = fopen(name, "rb"); |
1463 |
if (f==0) { |
if (f==0) { |
1464 |
printf( "can't open file '%s'\n", name); |
printf( "ERROR: can't open file '%s'\n", name); |
1465 |
return; |
return; |
1466 |
} |
} |
1467 |
fseek(f, 0, SEEK_END); |
fseek(f, 0, SEEK_END); |
1468 |
buf_size = ftell(f); |
buf_size = ftell(f); |
1469 |
fseek(f, 0, SEEK_SET); |
fseek(f, 0, SEEK_SET); |
1470 |
if (buf_size<=0) { |
if (buf_size<=0) { |
1471 |
printf("error while stating file\n"); |
printf("ERROR: error while stating file\n"); |
1472 |
goto End; |
goto End; |
1473 |
} |
} |
|
else printf( "Input size: %d\n", buf_size); |
|
1474 |
|
|
1475 |
buf = malloc(buf_size); /* should be enuf' */ |
buf = malloc(buf_size); |
1476 |
rgb_out = calloc(4, width*height); /* <-room for _RGB24 */ |
yuv_out = calloc(1, bps*height*3/2 + 15); |
1477 |
if (buf==0 || rgb_out==0) { |
if (buf==0 || yuv_out==0) { |
1478 |
printf( "malloc failed!\n" ); |
printf( "ERROR: malloc failed!\n" ); |
1479 |
goto End; |
goto End; |
1480 |
} |
} |
1481 |
|
|
1482 |
if (fread(buf, buf_size, 1, f)!=1) { |
if (fread(buf, buf_size, 1, f)!=1) { |
1483 |
printf( "file-read failed\n" ); |
printf( "ERROR: file-read failed\n" ); |
1484 |
goto End; |
goto End; |
1485 |
} |
} |
1486 |
|
|
1488 |
pos = 0; |
pos = 0; |
1489 |
t = -gettime_usec(); |
t = -gettime_usec(); |
1490 |
while(1) { |
while(1) { |
1491 |
|
int y; |
1492 |
|
|
1493 |
memset(&xframe, 0, sizeof(xframe)); |
memset(&xframe, 0, sizeof(xframe)); |
1494 |
xframe.version = XVID_VERSION; |
xframe.version = XVID_VERSION; |
1495 |
xframe.bitstream = buf + pos; |
xframe.bitstream = buf + pos; |
1496 |
xframe.length = buf_size - pos; |
xframe.length = buf_size - pos; |
1497 |
xframe.output.plane[0] = rgb_out; |
xframe.output.plane[0] = (uint8_t*)(((size_t)yuv_out + 15) & ~15); |
1498 |
xframe.output.stride[0] = width; |
xframe.output.plane[1] = (uint8_t*)xframe.output.plane[0] + bps*height; |
1499 |
xframe.output.csp = XVID_CSP_BGR; |
xframe.output.plane[2] = (uint8_t*)xframe.output.plane[1] + bps/2; |
1500 |
|
xframe.output.stride[0] = bps; |
1501 |
|
xframe.output.stride[1] = bps; |
1502 |
|
xframe.output.stride[2] = bps; |
1503 |
|
xframe.output.csp = XVID_CSP_I420; |
1504 |
xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); |
xerr = xvid_decore(dechandle, XVID_DEC_DECODE, &xframe, 0); |
1505 |
|
if (xerr<0) { |
1506 |
|
printf("ERROR: decoding failed for frame #%d (err=%d)!\n", nb, xerr); |
1507 |
|
break; |
1508 |
|
} |
1509 |
|
else if (xerr==0) |
1510 |
|
break; |
1511 |
|
else if (verbose>0) printf("#%d %d\n", nb, xerr ); |
1512 |
|
|
1513 |
|
pos += xerr; |
1514 |
nb++; |
nb++; |
1515 |
pos += xframe.length; |
|
1516 |
if (with_chksum) { |
for(y=0; y<height/2; ++y) { |
1517 |
int k = width*height; |
chksum = calc_crc((uint8_t*)xframe.output.plane[0] + (2*y+0)*bps, width, chksum); |
1518 |
uint32_t *ptr = (uint32_t *)rgb_out; |
chksum = calc_crc((uint8_t*)xframe.output.plane[0] + (2*y+1)*bps, width, chksum); |
1519 |
while(k-->0) chksum += *ptr++; |
chksum = calc_crc((uint8_t*)xframe.output.plane[1] + y*bps, width/2, chksum); |
1520 |
|
chksum = calc_crc((uint8_t*)xframe.output.plane[2] + y*bps, width/2, chksum); |
1521 |
} |
} |
1522 |
if (pos==buf_size) |
if (pos==buf_size) |
1523 |
break; |
break; |
|
if (xerr==XVID_ERR_FAIL) { |
|
|
printf("decoding failed for frame #%d (err=%d)!\n", nb, xerr); |
|
|
break; |
|
|
} |
|
1524 |
} |
} |
1525 |
t += gettime_usec(); |
t += gettime_usec(); |
1526 |
|
if (ref_chksum==0) { |
1527 |
if (t>0.) |
if (t>0.) |
1528 |
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 ); |
1529 |
if (with_chksum) |
} |
1530 |
printf("checksum: 0x%.8x\n", chksum); |
else { |
1531 |
|
printf("FPS:%.1f Checksum: 0x%.8x Expected:0x%.8x | %s\n", |
1532 |
|
t>0. ? (float)(nb*1.e6f/t) : 0.f, chksum, ref_chksum, (chksum==ref_chksum) ? "OK" : "ERROR"); |
1533 |
|
} |
1534 |
|
|
1535 |
End: |
End: |
1536 |
if (rgb_out!=0) free(rgb_out); |
if (yuv_out!=0) free(yuv_out); |
1537 |
if (buf!=0) free(buf); |
if (buf!=0) free(buf); |
1538 |
if (dechandle!=0) { |
if (dechandle!=0) { |
1539 |
xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); |
xerr= xvid_decore(dechandle, XVID_DEC_DESTROY, NULL, NULL); |
1540 |
if (xerr==XVID_ERR_FAIL) |
if (xerr==XVID_ERR_FAIL) |
1541 |
printf("destroy-decoder failed (err=%d)!\n", xerr); |
printf("ERROR: destroy-decoder failed (err=%d)!\n", xerr); |
1542 |
} |
} |
1543 |
if (f!=0) fclose(f); |
if (f!=0) fclose(f); |
1544 |
} |
} |
1700 |
} |
} |
1701 |
#endif |
#endif |
1702 |
} |
} |
1703 |
|
/*********************************************************************/ |
1704 |
|
|
1705 |
|
static uint32_t __inline log2bin_v1(uint32_t value) |
1706 |
|
{ |
1707 |
|
int n = 0; |
1708 |
|
while (value) { |
1709 |
|
value >>= 1; |
1710 |
|
n++; |
1711 |
|
} |
1712 |
|
return n; |
1713 |
|
} |
1714 |
|
|
1715 |
|
static const uint8_t log2_tab_16[16] = { 0, 1, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4 }; |
1716 |
|
|
1717 |
|
static uint32_t __inline log2bin_v2(uint32_t value) |
1718 |
|
{ |
1719 |
|
int n = 0; |
1720 |
|
if (value & 0xffff0000) { |
1721 |
|
value >>= 16; |
1722 |
|
n += 16; |
1723 |
|
} |
1724 |
|
if (value & 0xff00) { |
1725 |
|
value >>= 8; |
1726 |
|
n += 8; |
1727 |
|
} |
1728 |
|
if (value & 0xf0) { |
1729 |
|
value >>= 4; |
1730 |
|
n += 4; |
1731 |
|
} |
1732 |
|
return n + log2_tab_16[value]; |
1733 |
|
} |
1734 |
|
|
1735 |
|
void test_log2bin() |
1736 |
|
{ |
1737 |
|
const int nb_tests = 3000*speed_ref; |
1738 |
|
int n, crc1=0, crc2=0; |
1739 |
|
uint32_t s, s0; |
1740 |
|
double t1, t2; |
1741 |
|
|
1742 |
|
t1 = gettime_usec(); |
1743 |
|
s0 = (int)(t1*31.241); |
1744 |
|
for(s=s0, n=0; n<nb_tests; ++n, s=(s*12363+31)&0x7fffffff) |
1745 |
|
crc1 += log2bin_v1(s); |
1746 |
|
t1 = (gettime_usec()-t1) / nb_tests; |
1747 |
|
|
1748 |
|
t2 = gettime_usec(); |
1749 |
|
for(s=s0, n=0; n<nb_tests; ++n, s=(s*12363+31)&0x7fffffff) |
1750 |
|
crc2 += log2bin_v2(s); |
1751 |
|
t2 = (gettime_usec() - t2) / nb_tests; |
1752 |
|
|
1753 |
|
printf( "log2bin_v1: %.3f sec crc=%d\n", t1, crc1 ); |
1754 |
|
printf( "log2bin_v2: %.3f sec crc=%d\n", t2, crc2 ); |
1755 |
|
if (crc1!=crc2) printf( " CRC ERROR !\n" ); |
1756 |
|
} |
1757 |
|
|
1758 |
|
/*********************************************************************/ |
1759 |
|
|
1760 |
|
static void __inline old_gcd(int *num, int *den) |
1761 |
|
{ |
1762 |
|
int i = *num; |
1763 |
|
while (i > 1) { |
1764 |
|
if (*num % i == 0 && *den % i == 0) { |
1765 |
|
*num /= i; |
1766 |
|
*den /= i; |
1767 |
|
i = *num; |
1768 |
|
continue; |
1769 |
|
} |
1770 |
|
i--; |
1771 |
|
} |
1772 |
|
} |
1773 |
|
|
1774 |
|
static uint32_t gcd(int num, int den) |
1775 |
|
{ |
1776 |
|
int tmp; |
1777 |
|
while( (tmp=num%den) ) { num = den; den = tmp; } |
1778 |
|
return den; |
1779 |
|
} |
1780 |
|
static void __inline new_gcd(int *num, int *den) |
1781 |
|
{ |
1782 |
|
const int div = gcd(*num, *den); |
1783 |
|
if (num) { |
1784 |
|
*num /= div; |
1785 |
|
*den /= div; |
1786 |
|
} |
1787 |
|
} |
1788 |
|
|
1789 |
|
void test_gcd() |
1790 |
|
{ |
1791 |
|
const int nb_tests = 10*speed_ref; |
1792 |
|
int i; |
1793 |
|
uint32_t crc1=0, crc2=0; |
1794 |
|
uint32_t n0, n, d0, d; |
1795 |
|
double t1, t2; |
1796 |
|
|
1797 |
|
t1 = gettime_usec(); |
1798 |
|
n0 = 0xfffff & (int)(t1*31.241); |
1799 |
|
d0 = 0xfffff & (int)( ((n0*4123)%17) | 1 ); |
1800 |
|
for(n=n0, d=d0, i=0; i<nb_tests; ++i) { |
1801 |
|
old_gcd(&n, &d); |
1802 |
|
crc1 = (((crc1>>4)^d) + ((crc1<<2)^n) ) & 0xffffff; |
1803 |
|
n = d; |
1804 |
|
d = (d*12363+31) & 0xffff; |
1805 |
|
d |= !d; |
1806 |
|
} |
1807 |
|
t1 = (gettime_usec()-t1) / nb_tests; |
1808 |
|
|
1809 |
|
t2 = gettime_usec(); |
1810 |
|
for(n=n0, d=d0, i=0; i<nb_tests; ++i) { |
1811 |
|
new_gcd(&n, &d); |
1812 |
|
crc2 = (((crc2>>4)^d) + ((crc2<<2)^n) ) & 0xffffff; |
1813 |
|
n = d; |
1814 |
|
d = (d*12363+31) & 0xffff; |
1815 |
|
d |= !d; |
1816 |
|
} |
1817 |
|
t2 = (gettime_usec() - t2) / nb_tests; |
1818 |
|
|
1819 |
|
printf( "old_gcd: %.3f sec crc=%d\n", t1, crc1 ); |
1820 |
|
printf( "new_gcd: %.3f sec crc=%d\n", t2, crc2 ); |
1821 |
|
if (crc1!=crc2) printf( " CRC ERROR !\n" ); |
1822 |
|
} |
1823 |
|
|
1824 |
|
/********************************************************************* |
1825 |
|
* test compiler |
1826 |
|
*********************************************************************/ |
1827 |
|
|
1828 |
|
void test_compiler() { |
1829 |
|
int nb_err = 0; |
1830 |
|
int32_t v; |
1831 |
|
if (sizeof(uint16_t)<2) { |
1832 |
|
printf( "ERROR: sizeof(uint16_t)<2!\n" ); |
1833 |
|
nb_err++; |
1834 |
|
} |
1835 |
|
if (sizeof(int16_t)<2) { |
1836 |
|
printf( "ERROR: sizeof(int16_t)<2!\n" ); |
1837 |
|
nb_err++; |
1838 |
|
} |
1839 |
|
if (sizeof(uint8_t)!=1) { |
1840 |
|
printf( "ERROR: sizeof(uint8_t)!=1!\n" ); |
1841 |
|
nb_err++; |
1842 |
|
} |
1843 |
|
if (sizeof(int8_t)!=1) { |
1844 |
|
printf( "ERROR: sizeof(uint8_t)!=1!\n" ); |
1845 |
|
nb_err++; |
1846 |
|
} |
1847 |
|
if (sizeof(uint32_t)<4) { |
1848 |
|
printf( "ERROR: sizeof(uint8_t)<4!\n" ); |
1849 |
|
nb_err++; |
1850 |
|
} |
1851 |
|
if (sizeof(int32_t)<4) { |
1852 |
|
printf( "ERROR: sizeof(uint32_t)<4!\n" ); |
1853 |
|
nb_err++; |
1854 |
|
} |
1855 |
|
/* yes, i know, this test is silly. But better be safe than sorry. :) */ |
1856 |
|
for(v=1000; v>=0; v--) { |
1857 |
|
if ( (v>>2) != v/4) |
1858 |
|
nb_err++; |
1859 |
|
} |
1860 |
|
for(v=-1000; v!=-1; v++) { |
1861 |
|
if ( (v>>2) != (v/4)-!!(v%4)) |
1862 |
|
nb_err++; |
1863 |
|
} |
1864 |
|
if (nb_err!=0) { |
1865 |
|
printf( "ERROR! please post your platform/compiler specs to xvid-devel@xvid.org !\n" ); |
1866 |
|
} |
1867 |
|
} |
1868 |
|
|
1869 |
/********************************************************************* |
/********************************************************************* |
1870 |
* main |
* main |
1871 |
*********************************************************************/ |
*********************************************************************/ |
1872 |
|
|
1873 |
int main(int argc, char *argv[]) |
static void arg_missing(const char *opt) |
1874 |
{ |
{ |
1875 |
int what = 0; |
printf( "missing argument after option '%s'\n", opt); |
1876 |
if (argc>1) what = atoi(argv[1]); |
exit(-1); |
1877 |
|
} |
1878 |
|
|
1879 |
|
int main(int argc, const char *argv[]) |
1880 |
|
{ |
1881 |
|
int c, what = 0; |
1882 |
|
int width, height; |
1883 |
|
uint32_t chksum = 0; |
1884 |
|
const char * test_bitstream = 0; |
1885 |
|
|
1886 |
|
cpu_mask = 0; // default => will use autodectect |
1887 |
|
for(c=1; c<argc; ++c) |
1888 |
|
{ |
1889 |
|
if (!strcmp(argv[c], "-v")) verbose++; |
1890 |
|
else if (!strcmp(argv[c], "-c")) cpu_mask = 0 /* PLAIN_C */ | XVID_CPU_FORCE; |
1891 |
|
else if (!strcmp(argv[c], "-mmx")) cpu_mask = XVID_CPU_MMX | XVID_CPU_FORCE; |
1892 |
|
else if (!strcmp(argv[c], "-mmxext")) cpu_mask = XVID_CPU_MMXEXT | XVID_CPU_MMX | XVID_CPU_FORCE; |
1893 |
|
else if (!strcmp(argv[c], "-sse2")) cpu_mask = XVID_CPU_SSE2 | XVID_CPU_MMXEXT | XVID_CPU_MMX | XVID_CPU_FORCE; |
1894 |
|
else if (!strcmp(argv[c], "-3dnow")) cpu_mask = XVID_CPU_3DNOW | XVID_CPU_FORCE; |
1895 |
|
else if (!strcmp(argv[c], "-3dnowe")) cpu_mask = XVID_CPU_3DNOW | XVID_CPU_3DNOWEXT | XVID_CPU_FORCE; |
1896 |
|
else if (!strcmp(argv[c], "-altivec")) cpu_mask = XVID_CPU_ALTIVEC | XVID_CPU_FORCE; |
1897 |
|
else if (!strcmp(argv[c], "-spd")) { |
1898 |
|
if (++c==argc) arg_missing( argv[argc-1] ); |
1899 |
|
speed_ref = atoi(argv[c]); |
1900 |
|
} |
1901 |
|
else if (argv[c][0]!='-') { |
1902 |
|
what = atoi(argv[c]); |
1903 |
|
if (what==9) { |
1904 |
|
if (c+4>argc) { |
1905 |
|
printf("usage: %s %d bitstream width height (checksum)\n", argv[0], what); |
1906 |
|
exit(-1); |
1907 |
|
} |
1908 |
|
test_bitstream = argv[++c]; |
1909 |
|
width = atoi(argv[++c]); |
1910 |
|
height = atoi(argv[++c]); |
1911 |
|
if (c+1<argc && argv[c+1][0]!='-') { |
1912 |
|
if (sscanf(argv[c+1], "0x%x", &chksum)!=1) { |
1913 |
|
printf( "can't read checksum value.\n" ); |
1914 |
|
exit(-1); |
1915 |
|
} |
1916 |
|
else c++; |
1917 |
|
} |
1918 |
|
// printf( "[%s] %dx%d (0x%.8x)\n", test_bitstream, width, height, chksum); |
1919 |
|
} |
1920 |
|
} |
1921 |
|
else { |
1922 |
|
printf( "unrecognized option '%s'\n", argv[c]); |
1923 |
|
exit(-1); |
1924 |
|
} |
1925 |
|
} |
1926 |
|
|
1927 |
|
|
1928 |
if (what==0 || what==1) test_dct(); |
if (what==0 || what==1) test_dct(); |
1929 |
if (what==0 || what==2) test_mb(); |
if (what==0 || what==2) test_mb(); |
1930 |
if (what==0 || what==3) test_sad(); |
if (what==0 || what==3) test_sad(); |
1932 |
if (what==0 || what==5) test_quant(); |
if (what==0 || what==5) test_quant(); |
1933 |
if (what==0 || what==6) test_cbp(); |
if (what==0 || what==6) test_cbp(); |
1934 |
if (what==0 || what==10) test_sse(); |
if (what==0 || what==10) test_sse(); |
1935 |
|
if (what==0 || what==11) test_log2bin(); |
1936 |
|
if (what==0 || what==12) test_gcd(); |
1937 |
|
if (what==0 || what==13) test_compiler(); |
1938 |
|
|
1939 |
|
|
1940 |
if (what==7) { |
if (what==7) { |
1941 |
test_IEEE1180_compliance(-256, 255, 1); |
test_IEEE1180_compliance(-256, 255, 1); |
1947 |
} |
} |
1948 |
if (what==8) test_dct_saturation(-256, 255); |
if (what==8) test_dct_saturation(-256, 255); |
1949 |
|
|
1950 |
if (what==9) { |
if (test_bitstream) |
1951 |
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)); |
|
|
} |
|
|
|
|
1952 |
if (what==-1) { |
if (what==-1) { |
1953 |
test_dct_precision_diffs(); |
test_dct_precision_diffs(); |
1954 |
test_bugs1(); |
test_bugs1(); |
1976 |
return 0; |
return 0; |
1977 |
} |
} |
1978 |
|
|
1979 |
/********************************************************************* |
/*********************************************************************/ |
|
* '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 |
|
|
|
|
|
*********************************************************************/ |
|