36 |
#include "dct/fdct.h" |
#include "dct/fdct.h" |
37 |
#include "image/colorspace.h" |
#include "image/colorspace.h" |
38 |
#include "image/interpolate8x8.h" |
#include "image/interpolate8x8.h" |
|
#include "image/reduced.h" |
|
39 |
#include "utils/mem_transfer.h" |
#include "utils/mem_transfer.h" |
40 |
#include "utils/mbfunctions.h" |
#include "utils/mbfunctions.h" |
41 |
#include "quant/quant.h" |
#include "quant/quant.h" |
221 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; |
222 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
223 |
|
|
224 |
|
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c; |
225 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c; |
226 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c; |
227 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c; |
228 |
|
|
229 |
interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c; |
interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c; |
230 |
interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c; |
interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c; |
231 |
interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c; |
interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c; |
243 |
/* postprocessing */ |
/* postprocessing */ |
244 |
image_brightness = image_brightness_c; |
image_brightness = image_brightness_c; |
245 |
|
|
|
/* reduced resolution */ |
|
|
copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_C; |
|
|
add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_C; |
|
|
vfilter_31 = xvid_VFilter_31_C; |
|
|
hfilter_31 = xvid_HFilter_31_C; |
|
|
filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_C; |
|
|
filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_C; |
|
|
|
|
246 |
/* Initialize internal colorspace transformation tables */ |
/* Initialize internal colorspace transformation tables */ |
247 |
colorspace_init(); |
colorspace_init(); |
248 |
|
|
298 |
dev16 = dev16_c; |
dev16 = dev16_c; |
299 |
sad16v = sad16v_c; |
sad16v = sad16v_c; |
300 |
sse8_16bit = sse8_16bit_c; |
sse8_16bit = sse8_16bit_c; |
301 |
|
sse8_8bit = sse8_8bit_c; |
302 |
|
|
303 |
#if defined(ARCH_IS_IA32) |
#if defined(ARCH_IS_IA32) |
304 |
|
|
|
if ((cpu_flags & XVID_CPU_ASM)) { |
|
|
vfilter_31 = xvid_VFilter_31_x86; |
|
|
hfilter_31 = xvid_HFilter_31_x86; |
|
|
} |
|
|
|
|
305 |
if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || |
if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || |
306 |
(cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) || |
(cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) || |
307 |
(cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2)) |
(cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2)) |
348 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; |
349 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
350 |
|
|
351 |
|
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx; |
352 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx; |
353 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx; |
354 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx; |
355 |
|
|
356 |
interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx; |
interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx; |
357 |
interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx; |
interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx; |
358 |
|
|
362 |
/* postprocessing */ |
/* postprocessing */ |
363 |
image_brightness = image_brightness_mmx; |
image_brightness = image_brightness_mmx; |
364 |
|
|
|
/* reduced resolution */ |
|
|
copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_mmx; |
|
|
add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_mmx; |
|
|
hfilter_31 = xvid_HFilter_31_mmx; |
|
|
filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_mmx; |
|
|
filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_mmx; |
|
|
|
|
365 |
/* image input xxx_to_yv12 related functions */ |
/* image input xxx_to_yv12 related functions */ |
366 |
yv12_to_yv12 = yv12_to_yv12_mmx; |
yv12_to_yv12 = yv12_to_yv12_mmx; |
367 |
bgr_to_yv12 = bgr_to_yv12_mmx; |
bgr_to_yv12 = bgr_to_yv12_mmx; |
387 |
dev16 = dev16_mmx; |
dev16 = dev16_mmx; |
388 |
sad16v = sad16v_mmx; |
sad16v = sad16v_mmx; |
389 |
sse8_16bit = sse8_16bit_mmx; |
sse8_16bit = sse8_16bit_mmx; |
390 |
|
sse8_8bit = sse8_8bit_mmx; |
391 |
} |
} |
392 |
|
|
393 |
/* these 3dnow functions are faster than mmx, but slower than xmm. */ |
/* these 3dnow functions are faster than mmx, but slower than xmm. */ |
415 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; |
416 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
417 |
|
|
418 |
/* reduced resolution */ |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm; |
419 |
copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm; |
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm; |
420 |
add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm; |
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm; |
421 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm; |
422 |
|
|
423 |
/* Quantization */ |
/* Quantization */ |
424 |
quant_mpeg_intra = quant_mpeg_intra_xmm; |
quant_mpeg_intra = quant_mpeg_intra_xmm; |
454 |
|
|
455 |
if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
456 |
|
|
|
/* Inverse DCT */ |
|
|
idct = idct_3dne; |
|
|
|
|
457 |
/* Buffer transfer */ |
/* Buffer transfer */ |
458 |
transfer_8to16copy = transfer_8to16copy_3dne; |
transfer_8to16copy = transfer_8to16copy_3dne; |
459 |
transfer_16to8copy = transfer_16to8copy_3dne; |
transfer_16to8copy = transfer_16to8copy_3dne; |
460 |
transfer_8to16sub = transfer_8to16sub_3dne; |
transfer_8to16sub = transfer_8to16sub_3dne; |
461 |
transfer_8to16subro = transfer_8to16subro_3dne; |
transfer_8to16subro = transfer_8to16subro_3dne; |
|
transfer_8to16sub2 = transfer_8to16sub2_3dne; |
|
462 |
transfer_16to8add = transfer_16to8add_3dne; |
transfer_16to8add = transfer_16to8add_3dne; |
463 |
transfer8x8_copy = transfer8x8_copy_3dne; |
transfer8x8_copy = transfer8x8_copy_3dne; |
464 |
|
|
465 |
|
if ((cpu_flags & XVID_CPU_MMXEXT)) { |
466 |
|
/* Inverse DCT */ |
467 |
|
idct = idct_3dne; |
468 |
|
|
469 |
|
/* Buffer transfer */ |
470 |
|
transfer_8to16sub2 = transfer_8to16sub2_3dne; |
471 |
|
|
472 |
|
/* Interpolation */ |
473 |
|
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne; |
474 |
|
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; |
475 |
|
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; |
476 |
|
|
477 |
/* Quantization */ |
/* Quantization */ |
478 |
quant_h263_intra = quant_h263_intra_3dne; |
quant_h263_intra = quant_h263_intra_3dne; /* cmov only */ |
479 |
quant_h263_inter = quant_h263_inter_3dne; |
quant_h263_inter = quant_h263_inter_3dne; |
480 |
dequant_mpeg_intra = dequant_mpeg_intra_3dne; |
dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */ |
481 |
dequant_mpeg_inter = dequant_mpeg_inter_3dne; |
dequant_mpeg_inter = dequant_mpeg_inter_3dne; |
482 |
dequant_h263_intra = dequant_h263_intra_3dne; |
dequant_h263_intra = dequant_h263_intra_3dne; |
483 |
dequant_h263_inter = dequant_h263_inter_3dne; |
dequant_h263_inter = dequant_h263_inter_3dne; |
484 |
|
|
485 |
/* ME functions */ |
/* ME functions */ |
486 |
calc_cbp = calc_cbp_3dne; |
calc_cbp = calc_cbp_3dne; |
487 |
|
|
488 |
sad16 = sad16_3dne; |
sad16 = sad16_3dne; |
489 |
sad8 = sad8_3dne; |
sad8 = sad8_3dne; |
490 |
sad16bi = sad16bi_3dne; |
sad16bi = sad16bi_3dne; |
491 |
sad8bi = sad8bi_3dne; |
sad8bi = sad8bi_3dne; |
492 |
dev16 = dev16_3dne; |
dev16 = dev16_3dne; |
493 |
|
} |
|
/* Interpolation */ |
|
|
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne; |
|
|
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; |
|
|
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; |
|
494 |
} |
} |
495 |
|
|
496 |
if ((cpu_flags & XVID_CPU_SSE2)) { |
if ((cpu_flags & XVID_CPU_SSE2)) { |
510 |
/* DCT operators |
/* DCT operators |
511 |
* no iDCT because it's not "Walken matching" */ |
* no iDCT because it's not "Walken matching" */ |
512 |
fdct = fdct_sse2_skal; |
fdct = fdct_sse2_skal; |
513 |
|
|
514 |
|
/* postprocessing */ |
515 |
|
image_brightness = image_brightness_sse2; |
516 |
} |
} |
517 |
#endif /* ARCH_IS_IA32 */ |
#endif /* ARCH_IS_IA32 */ |
518 |
|
|
572 |
interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c; |
interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c; |
573 |
interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c; |
interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c; |
574 |
|
|
575 |
interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_altivec_c; |
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c; |
576 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c; |
577 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c; |
578 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c; |
579 |
|
|
580 |
/* Colorspace conversion */ |
/* Colorspace conversion */ |
581 |
bgra_to_yv12 = bgra_to_yv12_altivec_c; |
bgra_to_yv12 = bgra_to_yv12_altivec_c; |
594 |
quant_h263_inter = quant_h263_inter_altivec_c; |
quant_h263_inter = quant_h263_inter_altivec_c; |
595 |
dequant_h263_intra = dequant_h263_intra_altivec_c; |
dequant_h263_intra = dequant_h263_intra_altivec_c; |
596 |
dequant_h263_inter = dequant_h263_inter_altivec_c; |
dequant_h263_inter = dequant_h263_inter_altivec_c; |
597 |
|
|
598 |
|
/* Qpel stuff */ |
599 |
|
xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C; |
600 |
|
xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C; |
601 |
} |
} |
602 |
#endif |
#endif |
603 |
|
|
616 |
return XVID_ERR_VERSION; |
return XVID_ERR_VERSION; |
617 |
|
|
618 |
info->actual_version = XVID_VERSION; |
info->actual_version = XVID_VERSION; |
619 |
info->build = "xvid-1.0.0"; |
info->build = "xvid-1.1-cvshead"; |
620 |
info->cpu_flags = detect_cpu_flags(); |
info->cpu_flags = detect_cpu_flags(); |
621 |
|
|
622 |
#if defined(_SMP) && defined(WIN32) |
#if defined(_SMP) && defined(WIN32) |