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.c,v 1.62 2005-04-04 23:49:37 edgomez Exp $ |
* $Id: xvid.c,v 1.69 2006-06-14 21:44:07 Skal Exp $ |
23 |
* |
* |
24 |
****************************************************************************/ |
****************************************************************************/ |
25 |
|
|
40 |
#include "utils/mbfunctions.h" |
#include "utils/mbfunctions.h" |
41 |
#include "quant/quant.h" |
#include "quant/quant.h" |
42 |
#include "motion/motion.h" |
#include "motion/motion.h" |
43 |
|
#include "motion/gmc.h" |
44 |
#include "motion/sad.h" |
#include "motion/sad.h" |
45 |
#include "utils/emms.h" |
#include "utils/emms.h" |
46 |
#include "utils/timer.h" |
#include "utils/timer.h" |
127 |
|
|
128 |
/* detect cpu flags */ |
/* detect cpu flags */ |
129 |
static unsigned int |
static unsigned int |
130 |
detect_cpu_flags() |
detect_cpu_flags(void) |
131 |
{ |
{ |
132 |
/* enable native assembly optimizations by default */ |
/* enable native assembly optimizations by default */ |
133 |
unsigned int cpu_flags = XVID_CPU_ASM; |
unsigned int cpu_flags = XVID_CPU_ASM; |
184 |
idct = idct_int32; |
idct = idct_int32; |
185 |
|
|
186 |
/* Only needed on PPC Altivec archs */ |
/* Only needed on PPC Altivec archs */ |
187 |
sadInit = 0; |
sadInit = NULL; |
188 |
|
|
189 |
/* Restore FPU context : emms_c is a nop functions */ |
/* Restore FPU context : emms_c is a nop functions */ |
190 |
emms = emms_c; |
emms = emms_c; |
214 |
transfer_8to16sub2ro = transfer_8to16sub2ro_c; |
transfer_8to16sub2ro = transfer_8to16sub2ro_c; |
215 |
transfer_16to8add = transfer_16to8add_c; |
transfer_16to8add = transfer_16to8add_c; |
216 |
transfer8x8_copy = transfer8x8_copy_c; |
transfer8x8_copy = transfer8x8_copy_c; |
217 |
|
transfer8x4_copy = transfer8x4_copy_c; |
218 |
|
|
219 |
/* Interlacing functions */ |
/* Interlacing functions */ |
220 |
MBFieldTest = MBFieldTest_c; |
MBFieldTest = MBFieldTest_c; |
224 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; |
225 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
226 |
|
|
227 |
|
interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_c; |
228 |
|
interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_c; |
229 |
|
interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_c; |
230 |
|
|
231 |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c; |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c; |
232 |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c; |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c; |
233 |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c; |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c; |
307 |
sse8_16bit = sse8_16bit_c; |
sse8_16bit = sse8_16bit_c; |
308 |
sse8_8bit = sse8_8bit_c; |
sse8_8bit = sse8_8bit_c; |
309 |
|
|
310 |
|
init_GMC(cpu_flags); |
311 |
|
|
312 |
#if defined(ARCH_IS_IA32) |
#if defined(ARCH_IS_IA32) |
313 |
|
|
314 |
if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || |
if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || |
348 |
transfer_8to16sub2 = transfer_8to16sub2_mmx; |
transfer_8to16sub2 = transfer_8to16sub2_mmx; |
349 |
transfer_16to8add = transfer_16to8add_mmx; |
transfer_16to8add = transfer_16to8add_mmx; |
350 |
transfer8x8_copy = transfer8x8_copy_mmx; |
transfer8x8_copy = transfer8x8_copy_mmx; |
351 |
|
transfer8x4_copy = transfer8x4_copy_mmx; |
352 |
|
|
353 |
/* Interlacing Functions */ |
/* Interlacing Functions */ |
354 |
MBFieldTest = MBFieldTest_mmx; |
MBFieldTest = MBFieldTest_mmx; |
358 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; |
359 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
360 |
|
|
361 |
|
interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_mmx; |
362 |
|
interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_mmx; |
363 |
|
interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_mmx; |
364 |
|
|
365 |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx; |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx; |
366 |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx; |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx; |
367 |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx; |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx; |
429 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; |
430 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
431 |
|
|
432 |
|
interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_xmm; |
433 |
|
interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_xmm; |
434 |
|
interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_xmm; |
435 |
|
|
436 |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm; |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm; |
437 |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm; |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm; |
438 |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm; |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm; |
469 |
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn; |
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn; |
470 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn; |
471 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn; |
472 |
|
|
473 |
|
interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dn; |
474 |
|
interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dn; |
475 |
|
interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dn; |
476 |
} |
} |
477 |
|
|
478 |
if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
484 |
transfer_8to16subro = transfer_8to16subro_3dne; |
transfer_8to16subro = transfer_8to16subro_3dne; |
485 |
transfer_16to8add = transfer_16to8add_3dne; |
transfer_16to8add = transfer_16to8add_3dne; |
486 |
transfer8x8_copy = transfer8x8_copy_3dne; |
transfer8x8_copy = transfer8x8_copy_3dne; |
487 |
|
transfer8x4_copy = transfer8x4_copy_3dne; |
488 |
|
|
489 |
if ((cpu_flags & XVID_CPU_MMXEXT)) { |
if ((cpu_flags & XVID_CPU_MMXEXT)) { |
490 |
/* Inverse DCT */ |
/* Inverse DCT */ |
498 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; |
499 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; |
500 |
|
|
501 |
|
interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dne; |
502 |
|
interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dne; |
503 |
|
interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dne; |
504 |
|
|
505 |
/* Quantization */ |
/* Quantization */ |
506 |
quant_h263_intra = quant_h263_intra_3dne; /* cmov only */ |
quant_h263_intra = quant_h263_intra_3dne; /* cmov only */ |
507 |
quant_h263_inter = quant_h263_inter_3dne; |
quant_h263_inter = quant_h263_inter_3dne; |
535 |
sad16 = sad16_sse2; |
sad16 = sad16_sse2; |
536 |
dev16 = dev16_sse2; |
dev16 = dev16_sse2; |
537 |
|
|
538 |
/* DCT operators |
/* DCT operators */ |
|
* no iDCT because it's not "Walken matching" */ |
|
539 |
fdct = fdct_sse2_skal; |
fdct = fdct_sse2_skal; |
540 |
|
/* idct = idct_sse2_skal; */ /* Is now IEEE1180 and Walken compliant. Disabled until fully tested. */ |
541 |
|
|
542 |
/* postprocessing */ |
/* postprocessing */ |
543 |
image_brightness = image_brightness_sse2; |
image_brightness = image_brightness_sse2; |
639 |
* features there really are. */ |
* features there really are. */ |
640 |
if (cpu_flags & XVID_CPU_ASM) { |
if (cpu_flags & XVID_CPU_ASM) { |
641 |
/* SIMD state flusher */ |
/* SIMD state flusher */ |
642 |
emms = emms_3dn; |
emms = emms_mmx; |
643 |
|
|
644 |
/* DCT operators */ |
/* DCT operators */ |
645 |
fdct = fdct_skal_x86_64; |
fdct = fdct_skal_x86_64; |
715 |
return XVID_ERR_VERSION; |
return XVID_ERR_VERSION; |
716 |
|
|
717 |
info->actual_version = XVID_VERSION; |
info->actual_version = XVID_VERSION; |
718 |
info->build = "xvid-1.1.0-beta2"; |
info->build = "xvid-1.2.0-dev"; |
719 |
info->cpu_flags = detect_cpu_flags(); |
info->cpu_flags = detect_cpu_flags(); |
|
|
|
|
#if defined(_SMP) && defined(WIN32) |
|
|
info->num_threads = pthread_num_processors_np();; |
|
|
#else |
|
720 |
info->num_threads = 0; |
info->num_threads = 0; |
721 |
|
|
722 |
|
#if defined(WIN32) |
723 |
|
{ |
724 |
|
DWORD dwProcessAffinityMask, dwSystemAffinityMask; |
725 |
|
if (GetProcessAffinityMask(GetCurrentProcess(), &dwProcessAffinityMask, &dwSystemAffinityMask)) { |
726 |
|
int i; |
727 |
|
for(i=0; i<32; i++) { |
728 |
|
if ((dwProcessAffinityMask & (1<<i))) |
729 |
|
info->num_threads++; |
730 |
|
} |
731 |
|
} |
732 |
|
} |
733 |
#endif |
#endif |
734 |
|
|
735 |
return 0; |
return 0; |