[cvs] / xvidcore / src / xvid.c Repository:
ViewVC logotype

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.48 - (view) (download)

1 : edgomez 1.16 /*****************************************************************************
2 : edgomez 1.17 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : edgomez 1.48 * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org>
7 :     *
8 : edgomez 1.41 * This program is free software ; you can redistribute it and/or modify
9 :     * it under the terms of the GNU General Public License as published by
10 :     * the Free Software Foundation ; either version 2 of the License, or
11 : edgomez 1.17 * (at your option) any later version.
12 :     *
13 :     * This program is distributed in the hope that it will be useful,
14 : edgomez 1.41 * but WITHOUT ANY WARRANTY ; without even the implied warranty of
15 : edgomez 1.17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 :     * GNU General Public License for more details.
17 :     *
18 :     * You should have received a copy of the GNU General Public License
19 : edgomez 1.41 * along with this program ; if not, write to the Free Software
20 : edgomez 1.17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 :     *
22 : edgomez 1.48 * $Id$
23 : edgomez 1.36 *
24 : edgomez 1.17 ****************************************************************************/
25 : chenm001 1.29
26 : edgomez 1.41 #include <stdio.h>
27 :     #include <stdlib.h>
28 :     #include <string.h>
29 :     #include <time.h>
30 :    
31 : Isibaar 1.1 #include "xvid.h"
32 :     #include "decoder.h"
33 :     #include "encoder.h"
34 :     #include "bitstream/cbp.h"
35 :     #include "dct/idct.h"
36 :     #include "dct/fdct.h"
37 :     #include "image/colorspace.h"
38 :     #include "image/interpolate8x8.h"
39 : edgomez 1.41 #include "image/reduced.h"
40 : Isibaar 1.1 #include "utils/mem_transfer.h"
41 : edgomez 1.41 #include "utils/mbfunctions.h"
42 : edgomez 1.48 #include "quant/quant.h"
43 : ia64p 1.30 #include "motion/motion.h"
44 : Isibaar 1.1 #include "motion/sad.h"
45 :     #include "utils/emms.h"
46 :     #include "utils/timer.h"
47 : Isibaar 1.9 #include "bitstream/mbcoding.h"
48 : edgomez 1.48 #include "image/qpel.h"
49 :     #include "image/postprocessing.h"
50 :    
51 :     #if defined(_DEBUG)
52 :     unsigned int xvid_debug = 0; /* xvid debug mask */
53 :     #endif
54 : Isibaar 1.1
55 : edgomez 1.41 #if defined(ARCH_IS_IA32)
56 :     #if defined(_MSC_VER)
57 :     # include <windows.h>
58 : suxen_drol 1.31 #else
59 : edgomez 1.41 # include <signal.h>
60 :     # include <setjmp.h>
61 : suxen_drol 1.31
62 : edgomez 1.41 static jmp_buf mark;
63 : suxen_drol 1.31
64 : edgomez 1.41 static void
65 :     sigill_handler(int signal)
66 :     {
67 :     longjmp(mark, 1);
68 :     }
69 : suxen_drol 1.31 #endif
70 :    
71 :    
72 :     /*
73 : edgomez 1.43 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
74 :     * signalled
75 :     *
76 :     * Return values:
77 :     * -1 : could not determine
78 :     * 0 : SIGILL was *not* signalled
79 :     * 1 : SIGILL was signalled
80 :     */
81 : suxen_drol 1.31
82 :     int
83 :     sigill_check(void (*func)())
84 :     {
85 : edgomez 1.41 #if defined(_MSC_VER)
86 : suxen_drol 1.31 _try {
87 :     func();
88 :     }
89 :     _except(EXCEPTION_EXECUTE_HANDLER) {
90 :    
91 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
92 :     return 1;
93 :     }
94 :     return 0;
95 :     #else
96 :     void * old_handler;
97 :     int jmpret;
98 :    
99 :    
100 :     old_handler = signal(SIGILL, sigill_handler);
101 :     if (old_handler == SIG_ERR)
102 :     {
103 :     return -1;
104 :     }
105 :    
106 :     jmpret = setjmp(mark);
107 :     if (jmpret == 0)
108 :     {
109 :     func();
110 :     }
111 :    
112 :     signal(SIGILL, old_handler);
113 :    
114 :     return jmpret;
115 :     #endif
116 :     }
117 :     #endif
118 :    
119 : edgomez 1.41
120 :     /* detect cpu flags */
121 :     static unsigned int
122 :     detect_cpu_flags()
123 :     {
124 :     /* enable native assembly optimizations by default */
125 :     unsigned int cpu_flags = XVID_CPU_ASM;
126 :    
127 :     #if defined(ARCH_IS_IA32)
128 :     cpu_flags |= check_cpu_features();
129 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
130 :     cpu_flags &= ~XVID_CPU_SSE;
131 :    
132 :     if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger))
133 :     cpu_flags &= ~XVID_CPU_SSE2;
134 :     #endif
135 :    
136 :     #if defined(ARCH_IS_PPC)
137 :     #if defined(ARCH_IS_PPC_ALTIVEC)
138 :     cpu_flags |= XVID_CPU_ALTIVEC;
139 :     #endif
140 :     #endif
141 :    
142 :     return cpu_flags;
143 :     }
144 :    
145 :    
146 : edgomez 1.16 /*****************************************************************************
147 :     * XviD Init Entry point
148 :     *
149 :     * Well this function initialize all internal function pointers according
150 :     * to the CPU features forced by the library client or autodetected (depending
151 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
152 :     * image colorspace transformation tables.
153 : edgomez 1.48 *
154 : edgomez 1.16 * Returned value : XVID_ERR_OK
155 :     * + API_VERSION in the input XVID_INIT_PARAM structure
156 :     * + core build " " " " "
157 :     *
158 :     ****************************************************************************/
159 :    
160 : edgomez 1.41
161 : edgomez 1.48 static
162 :     int xvid_gbl_init(xvid_gbl_init_t * init)
163 : Isibaar 1.1 {
164 : edgomez 1.48 unsigned int cpu_flags;
165 : suxen_drol 1.21
166 : edgomez 1.48 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
167 :     return XVID_ERR_VERSION;
168 : suxen_drol 1.31
169 : edgomez 1.48 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
170 : Isibaar 1.1
171 : edgomez 1.16 /* Initialize the function pointers */
172 : Isibaar 1.1 idct_int32_init();
173 : Isibaar 1.9 init_vlc_tables();
174 :    
175 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
176 : Isibaar 1.1 fdct = fdct_int32;
177 :     idct = idct_int32;
178 :    
179 : edgomez 1.16 /* Only needed on PPC Altivec archs */
180 : canard 1.10 sadInit = 0;
181 : edgomez 1.15
182 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
183 : Isibaar 1.1 emms = emms_c;
184 :    
185 : edgomez 1.48 /* Qpel stuff */
186 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
187 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
188 :     xvid_Init_QP();
189 :    
190 : edgomez 1.16 /* Quantization functions */
191 : edgomez 1.48 quant_h263_intra = quant_h263_intra_c;
192 :     quant_h263_inter = quant_h263_inter_c;
193 :     dequant_h263_intra = dequant_h263_intra_c;
194 :     dequant_h263_inter = dequant_h263_inter_c;
195 :    
196 :     quant_mpeg_intra = quant_mpeg_intra_c;
197 :     quant_mpeg_inter = quant_mpeg_inter_c;
198 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
199 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
200 : Isibaar 1.1
201 : edgomez 1.16 /* Block transfer related functions */
202 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
203 :     transfer_16to8copy = transfer_16to8copy_c;
204 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
205 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_c;
206 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
207 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
208 :     transfer8x8_copy = transfer8x8_copy_c;
209 : Isibaar 1.1
210 : edgomez 1.41 /* Interlacing functions */
211 :     MBFieldTest = MBFieldTest_c;
212 :    
213 : edgomez 1.16 /* Image interpolation related functions */
214 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
215 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
216 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
217 :    
218 : edgomez 1.41 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
219 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
220 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
221 :    
222 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
223 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
224 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
225 :    
226 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
227 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
228 :    
229 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
230 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
231 :    
232 : edgomez 1.48 /* reduced resolution */
233 : edgomez 1.41 copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_C;
234 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_C;
235 :     vfilter_31 = xvid_VFilter_31_C;
236 :     hfilter_31 = xvid_HFilter_31_C;
237 :     filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_C;
238 :     filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_C;
239 :    
240 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
241 : Isibaar 1.1 colorspace_init();
242 :    
243 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
244 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_c;
245 :     rgb555_to_yv12 = rgb555_to_yv12_c;
246 :     rgb565_to_yv12 = rgb565_to_yv12_c;
247 :     bgr_to_yv12 = bgr_to_yv12_c;
248 :     bgra_to_yv12 = bgra_to_yv12_c;
249 :     abgr_to_yv12 = abgr_to_yv12_c;
250 :     rgba_to_yv12 = rgba_to_yv12_c;
251 : edgomez 1.48 argb_to_yv12 = argb_to_yv12_c;
252 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_c;
253 :     uyvy_to_yv12 = uyvy_to_yv12_c;
254 :    
255 :     rgb555i_to_yv12 = rgb555i_to_yv12_c;
256 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
257 :     bgri_to_yv12 = bgri_to_yv12_c;
258 :     bgrai_to_yv12 = bgrai_to_yv12_c;
259 :     abgri_to_yv12 = abgri_to_yv12_c;
260 :     rgbai_to_yv12 = rgbai_to_yv12_c;
261 : edgomez 1.48 argbi_to_yv12 = argbi_to_yv12_c;
262 : edgomez 1.41 yuyvi_to_yv12 = yuyvi_to_yv12_c;
263 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
264 :    
265 : edgomez 1.16 /* All colorspace transformation functions YV12->User format */
266 : edgomez 1.41 yv12_to_rgb555 = yv12_to_rgb555_c;
267 :     yv12_to_rgb565 = yv12_to_rgb565_c;
268 :     yv12_to_bgr = yv12_to_bgr_c;
269 :     yv12_to_bgra = yv12_to_bgra_c;
270 :     yv12_to_abgr = yv12_to_abgr_c;
271 :     yv12_to_rgba = yv12_to_rgba_c;
272 : edgomez 1.48 yv12_to_argb = yv12_to_argb_c;
273 : edgomez 1.41 yv12_to_yuyv = yv12_to_yuyv_c;
274 :     yv12_to_uyvy = yv12_to_uyvy_c;
275 : edgomez 1.48
276 : edgomez 1.41 yv12_to_rgb555i = yv12_to_rgb555i_c;
277 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
278 :     yv12_to_bgri = yv12_to_bgri_c;
279 :     yv12_to_bgrai = yv12_to_bgrai_c;
280 :     yv12_to_abgri = yv12_to_abgri_c;
281 :     yv12_to_rgbai = yv12_to_rgbai_c;
282 : edgomez 1.48 yv12_to_argbi = yv12_to_argbi_c;
283 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_c;
284 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
285 : Isibaar 1.1
286 : edgomez 1.16 /* Functions used in motion estimation algorithms */
287 : edgomez 1.48 calc_cbp = calc_cbp_c;
288 :     sad16 = sad16_c;
289 :     sad8 = sad8_c;
290 :     sad16bi = sad16bi_c;
291 :     sad8bi = sad8bi_c;
292 :     dev16 = dev16_c;
293 :     sad16v = sad16v_c;
294 :     sse8_16bit = sse8_16bit_c;
295 : Isibaar 1.1
296 : edgomez 1.41 #if defined(ARCH_IS_IA32)
297 :    
298 : edgomez 1.48 if ((cpu_flags & XVID_CPU_ASM)) {
299 : edgomez 1.41 vfilter_31 = xvid_VFilter_31_x86;
300 :     hfilter_31 = xvid_HFilter_31_x86;
301 :     }
302 :    
303 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
304 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
305 :     (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2))
306 :     {
307 :     /* Restore FPU context : emms_c is a nop functions */
308 :     emms = emms_mmx;
309 :     }
310 :    
311 :     if ((cpu_flags & XVID_CPU_MMX)) {
312 : edgomez 1.16
313 :     /* Forward and Inverse Discrete Cosine Transformation functions */
314 : edgomez 1.48 fdct = fdct_mmx_skal;
315 : Isibaar 1.47 idct = idct_mmx;
316 : Isibaar 1.1
317 : edgomez 1.48 /* Qpel stuff */
318 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
319 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
320 :    
321 : edgomez 1.16 /* Quantization related functions */
322 : edgomez 1.48 quant_h263_intra = quant_h263_intra_mmx;
323 :     quant_h263_inter = quant_h263_inter_mmx;
324 :     dequant_h263_intra = dequant_h263_intra_mmx;
325 :     dequant_h263_inter = dequant_h263_inter_mmx;
326 :    
327 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
328 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
329 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
330 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
331 : Isibaar 1.1
332 : edgomez 1.16 /* Block related functions */
333 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
334 :     transfer_16to8copy = transfer_16to8copy_mmx;
335 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
336 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_mmx;
337 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
338 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
339 :     transfer8x8_copy = transfer8x8_copy_mmx;
340 : edgomez 1.22
341 : edgomez 1.41 /* Interlacing Functions */
342 :     MBFieldTest = MBFieldTest_mmx;
343 : edgomez 1.16
344 :     /* Image Interpolation related functions */
345 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
346 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
347 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
348 :    
349 : edgomez 1.41 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
350 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
351 :    
352 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
353 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
354 :    
355 :     /* reduced resolution */
356 :     copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_mmx;
357 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_mmx;
358 :     hfilter_31 = xvid_HFilter_31_mmx;
359 :     filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_mmx;
360 :     filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_mmx;
361 :    
362 :     /* image input xxx_to_yv12 related functions */
363 :     yv12_to_yv12 = yv12_to_yv12_mmx;
364 :     bgr_to_yv12 = bgr_to_yv12_mmx;
365 :     bgra_to_yv12 = bgra_to_yv12_mmx;
366 : edgomez 1.16 yuyv_to_yv12 = yuyv_to_yv12_mmx;
367 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
368 : Isibaar 1.1
369 : edgomez 1.41 /* image output yv12_to_xxx related functions */
370 :     yv12_to_bgr = yv12_to_bgr_mmx;
371 :     yv12_to_bgra = yv12_to_bgra_mmx;
372 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
373 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
374 : edgomez 1.48
375 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
376 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
377 : Isibaar 1.1
378 : edgomez 1.16 /* Motion estimation related functions */
379 : Isibaar 1.1 calc_cbp = calc_cbp_mmx;
380 : edgomez 1.16 sad16 = sad16_mmx;
381 :     sad8 = sad8_mmx;
382 : suxen_drol 1.33 sad16bi = sad16bi_mmx;
383 :     sad8bi = sad8bi_mmx;
384 : edgomez 1.16 dev16 = dev16_mmx;
385 : edgomez 1.41 sad16v = sad16v_mmx;
386 : edgomez 1.48 sse8_16bit = sse8_16bit_mmx;
387 : Isibaar 1.1 }
388 :    
389 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
390 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
391 :    
392 :     emms = emms_3dn;
393 : suxen_drol 1.33
394 :     /* ME functions */
395 :     sad16bi = sad16bi_3dn;
396 :     sad8bi = sad8bi_3dn;
397 : edgomez 1.41
398 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
399 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
400 : suxen_drol 1.33 }
401 :    
402 :    
403 : edgomez 1.41 if ((cpu_flags & XVID_CPU_MMXEXT)) {
404 : edgomez 1.16
405 : edgomez 1.48 /* DCT */
406 :     fdct = fdct_xmm_skal;
407 : Isibaar 1.47 idct = idct_xmm;
408 : edgomez 1.16
409 :     /* Interpolation */
410 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
411 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
412 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
413 : Isibaar 1.25
414 : edgomez 1.41 /* reduced resolution */
415 :     copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm;
416 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm;
417 :    
418 : chenm001 1.29 /* Quantization */
419 : edgomez 1.48 quant_mpeg_intra = quant_mpeg_intra_xmm;
420 :     quant_mpeg_inter = quant_mpeg_inter_xmm;
421 : edgomez 1.41
422 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_xmm;
423 :     dequant_h263_inter = dequant_h263_inter_xmm;
424 : chenm001 1.29
425 : edgomez 1.19 /* Buffer transfer */
426 :     transfer_8to16sub2 = transfer_8to16sub2_xmm;
427 : edgomez 1.16
428 :     /* Colorspace transformation */
429 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_xmm;
430 :     yuyv_to_yv12 = yuyv_to_yv12_xmm;
431 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
432 : Isibaar 1.1
433 : edgomez 1.16 /* ME functions */
434 : Isibaar 1.1 sad16 = sad16_xmm;
435 : suxen_drol 1.33 sad8 = sad8_xmm;
436 : chenm001 1.29 sad16bi = sad16bi_xmm;
437 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
438 : Isibaar 1.1 dev16 = dev16_xmm;
439 : edgomez 1.41 sad16v = sad16v_xmm;
440 : Isibaar 1.1 }
441 :    
442 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
443 : edgomez 1.16
444 :     /* Interpolation */
445 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
446 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
447 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
448 : Isibaar 1.1 }
449 :    
450 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
451 :    
452 :     /* Inverse DCT */
453 : Isibaar 1.47 idct = idct_3dne;
454 : edgomez 1.41
455 :     /* Buffer transfer */
456 :     transfer_8to16copy = transfer_8to16copy_3dne;
457 :     transfer_16to8copy = transfer_16to8copy_3dne;
458 :     transfer_8to16sub = transfer_8to16sub_3dne;
459 :     transfer_8to16subro = transfer_8to16subro_3dne;
460 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
461 :     transfer_16to8add = transfer_16to8add_3dne;
462 :     transfer8x8_copy = transfer8x8_copy_3dne;
463 :    
464 :     /* Quantization */
465 : edgomez 1.48 quant_h263_intra = quant_h263_intra_3dne;
466 :     quant_h263_inter = quant_h263_inter_3dne;
467 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne;
468 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
469 :     dequant_h263_intra = dequant_h263_intra_3dne;
470 :     dequant_h263_inter = dequant_h263_inter_3dne;
471 : edgomez 1.41
472 :     /* ME functions */
473 :     calc_cbp = calc_cbp_3dne;
474 :     sad16 = sad16_3dne;
475 :     sad8 = sad8_3dne;
476 :     sad16bi = sad16bi_3dne;
477 :     sad8bi = sad8bi_3dne;
478 :     dev16 = dev16_3dne;
479 :    
480 :     /* Interpolation */
481 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
482 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
483 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
484 :     }
485 :    
486 :     if ((cpu_flags & XVID_CPU_SSE2)) {
487 : edgomez 1.16
488 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
489 :    
490 : edgomez 1.16 /* Quantization */
491 : edgomez 1.48 quant_h263_intra = quant_h263_intra_sse2;
492 :     quant_h263_inter = quant_h263_inter_sse2;
493 :     dequant_h263_intra = dequant_h263_intra_sse2;
494 :     dequant_h263_inter = dequant_h263_inter_sse2;
495 : h 1.13
496 : edgomez 1.48 /* SAD operators */
497 : edgomez 1.16 sad16 = sad16_sse2;
498 :     dev16 = dev16_sse2;
499 : edgomez 1.48
500 :     /* DCT operators
501 :     * no iDCT because it's not "Walken matching" */
502 :     fdct = fdct_sse2_skal;
503 : h 1.12 }
504 : edgomez 1.48 #endif /* ARCH_IS_IA32 */
505 : Isibaar 1.18
506 : edgomez 1.41 #if defined(ARCH_IS_IA64)
507 : edgomez 1.43 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
508 : Isibaar 1.18 idct_ia64_init();
509 :     fdct = fdct_ia64;
510 : edgomez 1.43 idct = idct_ia64; /*not yet working, crashes */
511 : Isibaar 1.18 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
512 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
513 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
514 :     sad16 = sad16_ia64;
515 :     sad16bi = sad16bi_ia64;
516 :     sad8 = sad8_ia64;
517 :     dev16 = dev16_ia64;
518 : edgomez 1.43 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
519 : edgomez 1.48 quant_h263_intra = quant_h263_intra_ia64;
520 :     quant_h263_inter = quant_h263_inter_ia64;
521 :     dequant_h263_intra = dequant_h263_intra_ia64;
522 :     dequant_h263_inter = dequant_h263_inter_ia64;
523 : Isibaar 1.18 transfer_8to16copy = transfer_8to16copy_ia64;
524 :     transfer_16to8copy = transfer_16to8copy_ia64;
525 :     transfer_8to16sub = transfer_8to16sub_ia64;
526 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
527 :     transfer_16to8add = transfer_16to8add_ia64;
528 :     transfer8x8_copy = transfer8x8_copy_ia64;
529 :     }
530 : edgomez 1.48 #endif
531 : edgomez 1.16
532 : edgomez 1.41 #if defined(ARCH_IS_PPC)
533 :     if ((cpu_flags & XVID_CPU_ASM))
534 :     {
535 :     calc_cbp = calc_cbp_ppc;
536 :     }
537 :    
538 :     if ((cpu_flags & XVID_CPU_ALTIVEC))
539 :     {
540 :     calc_cbp = calc_cbp_altivec;
541 :     fdct = fdct_altivec;
542 :     idct = idct_altivec;
543 :     sadInit = sadInit_altivec;
544 :     sad16 = sad16_altivec;
545 :     sad8 = sad8_altivec;
546 :     dev16 = dev16_altivec;
547 :     }
548 : canard 1.6 #endif
549 : edgomez 1.41
550 : edgomez 1.48 #if defined(_DEBUG)
551 :     xvid_debug = init->debug;
552 :     #endif
553 :    
554 :     return 0;
555 : edgomez 1.41 }
556 :    
557 :    
558 :     static int
559 : edgomez 1.48 xvid_gbl_info(xvid_gbl_info_t * info)
560 : edgomez 1.41 {
561 : edgomez 1.48 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
562 :     return XVID_ERR_VERSION;
563 : edgomez 1.41
564 : edgomez 1.48 info->actual_version = XVID_VERSION;
565 :     info->build = "xvid-1.0.0";
566 :     info->cpu_flags = detect_cpu_flags();
567 : edgomez 1.41
568 : edgomez 1.48 #if defined(_SMP) && defined(WIN32)
569 :     info->num_threads = pthread_num_processors_np();;
570 :     #else
571 :     info->num_threads = 0;
572 :     #endif
573 : edgomez 1.41
574 :     return 0;
575 :     }
576 :    
577 :    
578 : edgomez 1.48 static int
579 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
580 : edgomez 1.41 {
581 : edgomez 1.48 int width;
582 :     int height;
583 :     int width2;
584 :     int height2;
585 :     IMAGE img;
586 : edgomez 1.41
587 : edgomez 1.48 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
588 :     return XVID_ERR_VERSION;
589 : edgomez 1.41
590 : edgomez 1.48 #if 0
591 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
592 :     #endif
593 :     width = convert->width;
594 :     height = convert->height;
595 :     width2 = convert->width/2;
596 :     height2 = convert->height/2;
597 : edgomez 1.41
598 : edgomez 1.48 switch (convert->input.csp & ~XVID_CSP_VFLIP)
599 : edgomez 1.41 {
600 : edgomez 1.48 case XVID_CSP_YV12 :
601 :     img.y = convert->input.plane[0];
602 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
603 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
604 :     image_output(&img, width, height, width,
605 :     (uint8_t**)convert->output.plane, convert->output.stride,
606 :     convert->output.csp, convert->interlacing);
607 :     break;
608 : edgomez 1.41
609 : edgomez 1.48 default :
610 :     return XVID_ERR_FORMAT;
611 : edgomez 1.41 }
612 :    
613 :    
614 : edgomez 1.48 emms();
615 : edgomez 1.41 return 0;
616 :     }
617 :    
618 : edgomez 1.48 /*****************************************************************************
619 :     * XviD Global Entry point
620 :     *
621 :     * Well this function initialize all internal function pointers according
622 :     * to the CPU features forced by the library client or autodetected (depending
623 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
624 :     * image colorspace transformation tables.
625 :     *
626 :     ****************************************************************************/
627 : Isibaar 1.1
628 : edgomez 1.41
629 :     int
630 : edgomez 1.48 xvid_global(void *handle,
631 : edgomez 1.41 int opt,
632 :     void *param1,
633 :     void *param2)
634 :     {
635 :     switch(opt)
636 :     {
637 : edgomez 1.48 case XVID_GBL_INIT :
638 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
639 : edgomez 1.41
640 : edgomez 1.48 case XVID_GBL_INFO :
641 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
642 :    
643 :     case XVID_GBL_CONVERT :
644 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
645 : edgomez 1.41
646 :     default :
647 :     return XVID_ERR_FAIL;
648 :     }
649 :     }
650 :    
651 : edgomez 1.16 /*****************************************************************************
652 :     * XviD Native decoder entry point
653 :     *
654 :     * This function is just a wrapper to all the option cases.
655 :     *
656 :     * Returned values : XVID_ERR_FAIL when opt is invalid
657 :     * else returns the wrapped function result
658 :     *
659 :     ****************************************************************************/
660 :    
661 : edgomez 1.15 int
662 :     xvid_decore(void *handle,
663 :     int opt,
664 :     void *param1,
665 :     void *param2)
666 : Isibaar 1.1 {
667 : edgomez 1.15 switch (opt) {
668 :     case XVID_DEC_CREATE:
669 : edgomez 1.48 return decoder_create((xvid_dec_create_t *) param1);
670 : edgomez 1.15
671 :     case XVID_DEC_DESTROY:
672 :     return decoder_destroy((DECODER *) handle);
673 : Isibaar 1.1
674 : edgomez 1.48 case XVID_DEC_DECODE:
675 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
676 :    
677 : Isibaar 1.1 default:
678 : edgomez 1.15 return XVID_ERR_FAIL;
679 :     }
680 : Isibaar 1.1 }
681 :    
682 : edgomez 1.16
683 :     /*****************************************************************************
684 :     * XviD Native encoder entry point
685 :     *
686 :     * This function is just a wrapper to all the option cases.
687 :     *
688 :     * Returned values : XVID_ERR_FAIL when opt is invalid
689 :     * else returns the wrapped function result
690 :     *
691 :     ****************************************************************************/
692 : Isibaar 1.1
693 : edgomez 1.15 int
694 :     xvid_encore(void *handle,
695 :     int opt,
696 :     void *param1,
697 :     void *param2)
698 : Isibaar 1.1 {
699 : edgomez 1.15 switch (opt) {
700 :     case XVID_ENC_ENCODE:
701 : edgomez 1.41
702 : edgomez 1.48 return enc_encode((Encoder *) handle,
703 :     (xvid_enc_frame_t *) param1,
704 :     (xvid_enc_stats_t *) param2);
705 : edgomez 1.15
706 :     case XVID_ENC_CREATE:
707 : edgomez 1.48 return enc_create((xvid_enc_create_t *) param1);
708 : edgomez 1.15
709 :     case XVID_ENC_DESTROY:
710 : edgomez 1.48 return enc_destroy((Encoder *) handle);
711 : Isibaar 1.1
712 :     default:
713 : edgomez 1.15 return XVID_ERR_FAIL;
714 :     }
715 : Isibaar 1.1 }

No admin address has been configured
ViewVC Help
Powered by ViewVC 1.0.4