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 1.66 2005-11-22 10:53:10 suxen_drol Exp $ |
* $Id: xvid.c,v 1.74 2008-11-14 15:43:27 Isibaar 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" |
137 |
if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger)) |
if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger)) |
138 |
cpu_flags &= ~XVID_CPU_SSE; |
cpu_flags &= ~XVID_CPU_SSE; |
139 |
|
|
140 |
if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger)) |
if ((cpu_flags & (XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41)) && sigill_check(sse2_os_trigger)) |
141 |
cpu_flags &= ~XVID_CPU_SSE2; |
cpu_flags &= ~(XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41); |
142 |
#endif |
#endif |
143 |
|
|
144 |
#if defined(ARCH_IS_PPC) |
#if defined(ARCH_IS_PPC) |
257 |
yv12_to_yv12 = yv12_to_yv12_c; |
yv12_to_yv12 = yv12_to_yv12_c; |
258 |
rgb555_to_yv12 = rgb555_to_yv12_c; |
rgb555_to_yv12 = rgb555_to_yv12_c; |
259 |
rgb565_to_yv12 = rgb565_to_yv12_c; |
rgb565_to_yv12 = rgb565_to_yv12_c; |
260 |
|
rgb_to_yv12 = rgb_to_yv12_c; |
261 |
bgr_to_yv12 = bgr_to_yv12_c; |
bgr_to_yv12 = bgr_to_yv12_c; |
262 |
bgra_to_yv12 = bgra_to_yv12_c; |
bgra_to_yv12 = bgra_to_yv12_c; |
263 |
abgr_to_yv12 = abgr_to_yv12_c; |
abgr_to_yv12 = abgr_to_yv12_c; |
279 |
/* All colorspace transformation functions YV12->User format */ |
/* All colorspace transformation functions YV12->User format */ |
280 |
yv12_to_rgb555 = yv12_to_rgb555_c; |
yv12_to_rgb555 = yv12_to_rgb555_c; |
281 |
yv12_to_rgb565 = yv12_to_rgb565_c; |
yv12_to_rgb565 = yv12_to_rgb565_c; |
282 |
|
yv12_to_rgb = yv12_to_rgb_c; |
283 |
yv12_to_bgr = yv12_to_bgr_c; |
yv12_to_bgr = yv12_to_bgr_c; |
284 |
yv12_to_bgra = yv12_to_bgra_c; |
yv12_to_bgra = yv12_to_bgra_c; |
285 |
yv12_to_abgr = yv12_to_abgr_c; |
yv12_to_abgr = yv12_to_abgr_c; |
309 |
sse8_16bit = sse8_16bit_c; |
sse8_16bit = sse8_16bit_c; |
310 |
sse8_8bit = sse8_8bit_c; |
sse8_8bit = sse8_8bit_c; |
311 |
|
|
312 |
|
init_GMC(cpu_flags); |
313 |
|
|
314 |
#if defined(ARCH_IS_IA32) |
#if defined(ARCH_IS_IA32) |
315 |
|
|
316 |
if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || |
if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || |
317 |
(cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) || |
(cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) || |
318 |
(cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2)) |
(cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2) || |
319 |
|
(cpu_flags & XVID_CPU_SSE3) || (cpu_flags & XVID_CPU_SSE41)) |
320 |
{ |
{ |
321 |
/* Restore FPU context : emms_c is a nop functions */ |
/* Restore FPU context : emms_c is a nop functions */ |
322 |
emms = emms_mmx; |
emms = emms_mmx; |
382 |
/* image input xxx_to_yv12 related functions */ |
/* image input xxx_to_yv12 related functions */ |
383 |
yv12_to_yv12 = yv12_to_yv12_mmx; |
yv12_to_yv12 = yv12_to_yv12_mmx; |
384 |
bgr_to_yv12 = bgr_to_yv12_mmx; |
bgr_to_yv12 = bgr_to_yv12_mmx; |
385 |
|
rgb_to_yv12 = rgb_to_yv12_mmx; |
386 |
bgra_to_yv12 = bgra_to_yv12_mmx; |
bgra_to_yv12 = bgra_to_yv12_mmx; |
387 |
|
rgba_to_yv12 = rgba_to_yv12_mmx; |
388 |
yuyv_to_yv12 = yuyv_to_yv12_mmx; |
yuyv_to_yv12 = yuyv_to_yv12_mmx; |
389 |
uyvy_to_yv12 = uyvy_to_yv12_mmx; |
uyvy_to_yv12 = uyvy_to_yv12_mmx; |
390 |
|
|
444 |
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm; |
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm; |
445 |
|
|
446 |
/* Quantization */ |
/* Quantization */ |
|
quant_mpeg_intra = quant_mpeg_intra_xmm; |
|
447 |
quant_mpeg_inter = quant_mpeg_inter_xmm; |
quant_mpeg_inter = quant_mpeg_inter_xmm; |
448 |
|
|
449 |
dequant_h263_intra = dequant_h263_intra_xmm; |
dequant_h263_intra = dequant_h263_intra_xmm; |
541 |
|
|
542 |
/* DCT operators */ |
/* DCT operators */ |
543 |
fdct = fdct_sse2_skal; |
fdct = fdct_sse2_skal; |
544 |
/* idct = idct_sse2_skal; */ /* Is now IEEE1180 and Walken compliant. Disabled until fully tested. */ |
idct = idct_sse2_skal; /* Is now IEEE1180 and Walken compliant. */ |
545 |
|
|
546 |
/* postprocessing */ |
/* postprocessing */ |
547 |
image_brightness = image_brightness_sse2; |
image_brightness = image_brightness_sse2; |
548 |
} |
} |
549 |
|
|
550 |
|
if ((cpu_flags & XVID_CPU_SSE3)) { |
551 |
|
|
552 |
|
/* SAD operators */ |
553 |
|
sad16 = sad16_sse3; |
554 |
|
dev16 = dev16_sse3; |
555 |
|
} |
556 |
|
|
557 |
#endif /* ARCH_IS_IA32 */ |
#endif /* ARCH_IS_IA32 */ |
558 |
|
|
559 |
#if defined(ARCH_IS_IA64) |
#if defined(ARCH_IS_IA64) |
651 |
* features there really are. */ |
* features there really are. */ |
652 |
if (cpu_flags & XVID_CPU_ASM) { |
if (cpu_flags & XVID_CPU_ASM) { |
653 |
/* SIMD state flusher */ |
/* SIMD state flusher */ |
654 |
emms = emms_3dn; |
emms = emms_mmx; |
655 |
|
|
656 |
/* DCT operators */ |
/* DCT operators */ |
657 |
fdct = fdct_skal_x86_64; |
fdct = fdct_skal_x86_64; |
688 |
quant_h263_inter = quant_h263_inter_x86_64; |
quant_h263_inter = quant_h263_inter_x86_64; |
689 |
dequant_h263_intra = dequant_h263_intra_x86_64; |
dequant_h263_intra = dequant_h263_intra_x86_64; |
690 |
dequant_h263_inter = dequant_h263_inter_x86_64; |
dequant_h263_inter = dequant_h263_inter_x86_64; |
691 |
quant_mpeg_intra = quant_mpeg_intra_x86_64; |
/*quant_mpeg_intra = quant_mpeg_intra_x86_64; fix me! */ |
692 |
quant_mpeg_inter = quant_mpeg_inter_x86_64; |
quant_mpeg_inter = quant_mpeg_inter_x86_64; |
693 |
dequant_mpeg_intra = dequant_mpeg_intra_x86_64; |
dequant_mpeg_intra = dequant_mpeg_intra_x86_64; |
694 |
dequant_mpeg_inter = dequant_mpeg_inter_x86_64; |
dequant_mpeg_inter = dequant_mpeg_inter_x86_64; |
729 |
info->actual_version = XVID_VERSION; |
info->actual_version = XVID_VERSION; |
730 |
info->build = "xvid-1.2.0-dev"; |
info->build = "xvid-1.2.0-dev"; |
731 |
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 |
|
732 |
info->num_threads = 0; |
info->num_threads = 0; |
733 |
|
|
734 |
|
#if defined(WIN32) |
735 |
|
{ |
736 |
|
DWORD dwProcessAffinityMask, dwSystemAffinityMask; |
737 |
|
if (GetProcessAffinityMask(GetCurrentProcess(), &dwProcessAffinityMask, &dwSystemAffinityMask)) { |
738 |
|
int i; |
739 |
|
for(i=0; i<32; i++) { |
740 |
|
if ((dwProcessAffinityMask & (1<<i))) |
741 |
|
info->num_threads++; |
742 |
|
} |
743 |
|
} |
744 |
|
} |
745 |
#endif |
#endif |
746 |
|
|
747 |
return 0; |
return 0; |