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

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.34 - (view) (download)

1 : edgomez 1.16 /*****************************************************************************
2 : edgomez 1.17 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : edgomez 1.34 * Copyright(C) 2001-2002 Peter Ross <pross@cs.rmit.edu.au>
7 :     *
8 : edgomez 1.17 * This program is an implementation of a part of one or more MPEG-4
9 :     * Video tools as specified in ISO/IEC 14496-2 standard. Those intending
10 :     * to use this software module in hardware or software products are
11 :     * advised that its use may infringe existing patents or copyrights, and
12 :     * any such use would be at such party's own risk. The original
13 :     * developer of this software module and his/her company, and subsequent
14 :     * editors and their companies, will have no liability for use of this
15 :     * software or modifications or derivatives thereof.
16 :     *
17 :     * This program is free software ; you can redistribute it and/or modify
18 :     * it under the terms of the GNU General Public License as published by
19 :     * the Free Software Foundation ; either version 2 of the License, or
20 :     * (at your option) any later version.
21 :     *
22 :     * This program is distributed in the hope that it will be useful,
23 :     * but WITHOUT ANY WARRANTY ; without even the implied warranty of
24 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 :     * GNU General Public License for more details.
26 :     *
27 :     * You should have received a copy of the GNU General Public License
28 :     * along with this program ; if not, write to the Free Software
29 :     * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
30 :     *
31 :     ****************************************************************************/
32 : chenm001 1.29
33 : Isibaar 1.1 #include "xvid.h"
34 :     #include "decoder.h"
35 :     #include "encoder.h"
36 :     #include "bitstream/cbp.h"
37 :     #include "dct/idct.h"
38 :     #include "dct/fdct.h"
39 :     #include "image/colorspace.h"
40 :     #include "image/interpolate8x8.h"
41 :     #include "utils/mem_transfer.h"
42 :     #include "quant/quant_h263.h"
43 :     #include "quant/quant_mpeg4.h"
44 : ia64p 1.30 #include "motion/motion.h"
45 : Isibaar 1.1 #include "motion/sad.h"
46 :     #include "utils/emms.h"
47 :     #include "utils/timer.h"
48 : Isibaar 1.9 #include "bitstream/mbcoding.h"
49 : Isibaar 1.1
50 : suxen_drol 1.31 #if defined(ARCH_X86) && defined(EXPERIMENTAL_SSE2_CODE)
51 :    
52 :     #ifdef WIN32
53 :     #include <windows.h>
54 :     #else
55 :     #include <signal.h>
56 :     #include <setjmp.h>
57 :     #endif
58 :    
59 :    
60 :     #ifndef WIN32
61 :    
62 :     static jmp_buf mark;
63 :    
64 :     static void
65 :     sigill_handler(int signal)
66 :     {
67 :     longjmp(mark, 1);
68 :     }
69 :     #endif
70 :    
71 :    
72 :     /*
73 : edgomez 1.34 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was signalled
74 :     * Return values:
75 :     * -1 : could not determine
76 :     * 0 : SIGILL was *not* signalled
77 :     * 1 : SIGILL was signalled
78 :     */
79 : suxen_drol 1.31
80 :     int
81 :     sigill_check(void (*func)())
82 :     {
83 :     #ifdef WIN32
84 :     _try {
85 :     func();
86 :     }
87 :     _except(EXCEPTION_EXECUTE_HANDLER) {
88 :    
89 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
90 :     return 1;
91 :     }
92 :     return 0;
93 :     #else
94 :     void * old_handler;
95 :     int jmpret;
96 :    
97 :    
98 :     old_handler = signal(SIGILL, sigill_handler);
99 :     if (old_handler == SIG_ERR)
100 :     {
101 :     return -1;
102 :     }
103 :    
104 :     jmpret = setjmp(mark);
105 :     if (jmpret == 0)
106 :     {
107 :     func();
108 :     }
109 :    
110 :     signal(SIGILL, old_handler);
111 :    
112 :     return jmpret;
113 :     #endif
114 :     }
115 :     #endif
116 :    
117 : edgomez 1.16 /*****************************************************************************
118 :     * XviD Init Entry point
119 :     *
120 :     * Well this function initialize all internal function pointers according
121 :     * to the CPU features forced by the library client or autodetected (depending
122 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
123 :     * image colorspace transformation tables.
124 :     *
125 :     * Returned value : XVID_ERR_OK
126 :     * + API_VERSION in the input XVID_INIT_PARAM structure
127 :     * + core build " " " " "
128 :     *
129 :     ****************************************************************************/
130 :    
131 : edgomez 1.15 int
132 :     xvid_init(void *handle,
133 :     int opt,
134 :     void *param1,
135 :     void *param2)
136 : Isibaar 1.1 {
137 :     int cpu_flags;
138 :     XVID_INIT_PARAM *init_param;
139 :    
140 :     init_param = (XVID_INIT_PARAM *) param1;
141 :    
142 : suxen_drol 1.21 /* Inform the client the API version */
143 :     init_param->api_version = API_VERSION;
144 :    
145 :     /* Inform the client the core build - unused because we're still alpha */
146 :     init_param->core_build = 1000;
147 :    
148 : suxen_drol 1.31 /* Do we have to force CPU features ? */
149 :     if ((init_param->cpu_flags & XVID_CPU_FORCE)) {
150 : suxen_drol 1.21
151 : Isibaar 1.1 cpu_flags = init_param->cpu_flags;
152 : suxen_drol 1.31
153 : edgomez 1.16 } else {
154 : Isibaar 1.1
155 : chenm001 1.29 cpu_flags = check_cpu_features();
156 : suxen_drol 1.31
157 :     #if defined(ARCH_X86) && defined(EXPERIMENTAL_SSE2_CODE)
158 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
159 :     cpu_flags &= ~XVID_CPU_SSE;
160 :    
161 :     if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger))
162 :     cpu_flags &= ~XVID_CPU_SSE2;
163 :     #endif
164 :     }
165 :    
166 :     if ((init_param->cpu_flags & XVID_CPU_CHKONLY))
167 :     {
168 : Isibaar 1.1 init_param->cpu_flags = cpu_flags;
169 : suxen_drol 1.31 return XVID_ERR_OK;
170 : Isibaar 1.1 }
171 : suxen_drol 1.31
172 :     init_param->cpu_flags = cpu_flags;
173 :    
174 : Isibaar 1.1
175 : edgomez 1.16 /* Initialize the function pointers */
176 : Isibaar 1.1 idct_int32_init();
177 : Isibaar 1.9 init_vlc_tables();
178 :    
179 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
180 : Isibaar 1.1 fdct = fdct_int32;
181 :     idct = idct_int32;
182 :    
183 : edgomez 1.16 /* Only needed on PPC Altivec archs */
184 : canard 1.10 sadInit = 0;
185 : edgomez 1.15
186 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
187 : Isibaar 1.1 emms = emms_c;
188 :    
189 : edgomez 1.16 /* Quantization functions */
190 :     quant_intra = quant_intra_c;
191 : Isibaar 1.1 dequant_intra = dequant_intra_c;
192 : edgomez 1.16 quant_inter = quant_inter_c;
193 : Isibaar 1.1 dequant_inter = dequant_inter_c;
194 :    
195 : edgomez 1.16 quant4_intra = quant4_intra_c;
196 : Isibaar 1.1 dequant4_intra = dequant4_intra_c;
197 : edgomez 1.16 quant4_inter = quant4_inter_c;
198 : Isibaar 1.1 dequant4_inter = dequant4_inter_c;
199 :    
200 : edgomez 1.16 /* Block transfer related functions */
201 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
202 :     transfer_16to8copy = transfer_16to8copy_c;
203 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
204 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
205 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
206 :     transfer8x8_copy = transfer8x8_copy_c;
207 : Isibaar 1.1
208 : edgomez 1.16 /* Image interpolation related functions */
209 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
210 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
211 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
212 :    
213 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
214 : Isibaar 1.1 colorspace_init();
215 :    
216 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
217 : Isibaar 1.1 rgb555_to_yv12 = rgb555_to_yv12_c;
218 :     rgb565_to_yv12 = rgb565_to_yv12_c;
219 : edgomez 1.16 rgb24_to_yv12 = rgb24_to_yv12_c;
220 :     rgb32_to_yv12 = rgb32_to_yv12_c;
221 :     yuv_to_yv12 = yuv_to_yv12_c;
222 :     yuyv_to_yv12 = yuyv_to_yv12_c;
223 :     uyvy_to_yv12 = uyvy_to_yv12_c;
224 : Isibaar 1.1
225 : edgomez 1.16 /* All colorspace transformation functions YV12->User format */
226 : Isibaar 1.1 yv12_to_rgb555 = yv12_to_rgb555_c;
227 :     yv12_to_rgb565 = yv12_to_rgb565_c;
228 : edgomez 1.16 yv12_to_rgb24 = yv12_to_rgb24_c;
229 :     yv12_to_rgb32 = yv12_to_rgb32_c;
230 :     yv12_to_yuv = yv12_to_yuv_c;
231 :     yv12_to_yuyv = yv12_to_yuyv_c;
232 :     yv12_to_uyvy = yv12_to_uyvy_c;
233 : Isibaar 1.1
234 : edgomez 1.16 /* Functions used in motion estimation algorithms */
235 : Isibaar 1.1 calc_cbp = calc_cbp_c;
236 : edgomez 1.16 sad16 = sad16_c;
237 : suxen_drol 1.33 sad8 = sad8_c;
238 : edgomez 1.16 sad16bi = sad16bi_c;
239 : suxen_drol 1.33 sad8bi = sad8bi_c;
240 : edgomez 1.16 dev16 = dev16_c;
241 : suxen_drol 1.33
242 : ia64p 1.30 Halfpel8_Refine = Halfpel8_Refine_c;
243 : Isibaar 1.1
244 :     #ifdef ARCH_X86
245 : edgomez 1.15 if ((cpu_flags & XVID_CPU_MMX) > 0) {
246 : edgomez 1.16
247 :     /* Forward and Inverse Discrete Cosine Transformation functions */
248 : Isibaar 1.1 fdct = fdct_mmx;
249 :     idct = idct_mmx;
250 :    
251 : edgomez 1.16 /* To restore FPU context after mmx use */
252 : Isibaar 1.1 emms = emms_mmx;
253 :    
254 : edgomez 1.16 /* Quantization related functions */
255 :     quant_intra = quant_intra_mmx;
256 : Isibaar 1.1 dequant_intra = dequant_intra_mmx;
257 : edgomez 1.16 quant_inter = quant_inter_mmx;
258 : Isibaar 1.1 dequant_inter = dequant_inter_mmx;
259 :    
260 : edgomez 1.16 quant4_intra = quant4_intra_mmx;
261 : Isibaar 1.1 dequant4_intra = dequant4_intra_mmx;
262 : edgomez 1.16 quant4_inter = quant4_inter_mmx;
263 : Isibaar 1.1 dequant4_inter = dequant4_inter_mmx;
264 :    
265 : edgomez 1.16 /* Block related functions */
266 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
267 :     transfer_16to8copy = transfer_16to8copy_mmx;
268 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
269 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
270 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
271 :     transfer8x8_copy = transfer8x8_copy_mmx;
272 : edgomez 1.22
273 : edgomez 1.16
274 :     /* Image Interpolation related functions */
275 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
276 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
277 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
278 :    
279 : edgomez 1.16 /* Image RGB->YV12 related functions */
280 : Isibaar 1.1 rgb24_to_yv12 = rgb24_to_yv12_mmx;
281 :     rgb32_to_yv12 = rgb32_to_yv12_mmx;
282 : edgomez 1.16 yuv_to_yv12 = yuv_to_yv12_mmx;
283 :     yuyv_to_yv12 = yuyv_to_yv12_mmx;
284 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
285 : Isibaar 1.1
286 : edgomez 1.16 /* Image YV12->RGB related functions */
287 : Isibaar 1.1 yv12_to_rgb24 = yv12_to_rgb24_mmx;
288 :     yv12_to_rgb32 = yv12_to_rgb32_mmx;
289 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
290 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
291 : Isibaar 1.1
292 : edgomez 1.16 /* Motion estimation related functions */
293 : Isibaar 1.1 calc_cbp = calc_cbp_mmx;
294 : edgomez 1.16 sad16 = sad16_mmx;
295 :     sad8 = sad8_mmx;
296 : suxen_drol 1.33 sad16bi = sad16bi_mmx;
297 :     sad8bi = sad8bi_mmx;
298 : edgomez 1.16 dev16 = dev16_mmx;
299 : Isibaar 1.1
300 :     }
301 :    
302 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
303 :     if ((cpu_flags & XVID_CPU_3DNOW) > 0) {
304 :    
305 :     /* ME functions */
306 :     sad16bi = sad16bi_3dn;
307 :     sad8bi = sad8bi_3dn;
308 :     }
309 :    
310 :    
311 : edgomez 1.15 if ((cpu_flags & XVID_CPU_MMXEXT) > 0) {
312 : edgomez 1.16
313 :     /* Inverse DCT */
314 : Isibaar 1.1 idct = idct_xmm;
315 : edgomez 1.16
316 :     /* Interpolation */
317 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
318 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
319 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
320 : Isibaar 1.25
321 : chenm001 1.29 /* Quantization */
322 :     dequant_intra = dequant_intra_xmm;
323 :     dequant_inter = dequant_inter_xmm;
324 :    
325 : edgomez 1.19 /* Buffer transfer */
326 :     transfer_8to16sub2 = transfer_8to16sub2_xmm;
327 : edgomez 1.16
328 :     /* Colorspace transformation */
329 : Isibaar 1.1 yuv_to_yv12 = yuv_to_yv12_xmm;
330 :    
331 : edgomez 1.16 /* ME functions */
332 : Isibaar 1.1 sad16 = sad16_xmm;
333 : suxen_drol 1.33 sad8 = sad8_xmm;
334 : chenm001 1.29 sad16bi = sad16bi_xmm;
335 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
336 : Isibaar 1.1 dev16 = dev16_xmm;
337 :    
338 :     }
339 :    
340 : edgomez 1.15 if ((cpu_flags & XVID_CPU_3DNOW) > 0) {
341 : edgomez 1.16
342 :     /* Interpolation */
343 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
344 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
345 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
346 : Isibaar 1.1 }
347 :    
348 : edgomez 1.15 if ((cpu_flags & XVID_CPU_SSE2) > 0) {
349 : Isibaar 1.14 #ifdef EXPERIMENTAL_SSE2_CODE
350 : edgomez 1.16
351 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
352 :    
353 : edgomez 1.16 /* Quantization */
354 :     quant_intra = quant_intra_sse2;
355 : Isibaar 1.14 dequant_intra = dequant_intra_sse2;
356 : edgomez 1.16 quant_inter = quant_inter_sse2;
357 : Isibaar 1.14 dequant_inter = dequant_inter_sse2;
358 : h 1.13
359 : edgomez 1.16 /* ME */
360 :     sad16 = sad16_sse2;
361 :     dev16 = dev16_sse2;
362 :    
363 :     /* Forward and Inverse DCT */
364 :     idct = idct_sse2;
365 : Isibaar 1.14 fdct = fdct_sse2;
366 :     #endif
367 : h 1.12 }
368 : edgomez 1.16
369 : Isibaar 1.1 #endif
370 : Isibaar 1.18
371 :     #ifdef ARCH_IA64
372 :     if ((cpu_flags & XVID_CPU_IA64) > 0) { //use assembler routines?
373 :     idct_ia64_init();
374 :     fdct = fdct_ia64;
375 :     idct = idct_ia64; //not yet working, crashes
376 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
377 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
378 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
379 :     sad16 = sad16_ia64;
380 :     sad16bi = sad16bi_ia64;
381 :     sad8 = sad8_ia64;
382 :     dev16 = dev16_ia64;
383 : ia64p 1.30 Halfpel8_Refine = Halfpel8_Refine_ia64;
384 : Isibaar 1.18 quant_intra = quant_intra_ia64;
385 :     dequant_intra = dequant_intra_ia64;
386 :     quant_inter = quant_inter_ia64;
387 :     dequant_inter = dequant_inter_ia64;
388 :     transfer_8to16copy = transfer_8to16copy_ia64;
389 :     transfer_16to8copy = transfer_16to8copy_ia64;
390 :     transfer_8to16sub = transfer_8to16sub_ia64;
391 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
392 :     transfer_16to8add = transfer_16to8add_ia64;
393 :     transfer8x8_copy = transfer8x8_copy_ia64;
394 :     DEBUG("Using IA-64 assembler routines.\n");
395 :     }
396 :     #endif
397 : edgomez 1.16
398 : canard 1.5 #ifdef ARCH_PPC
399 : canard 1.6 #ifdef ARCH_PPC_ALTIVEC
400 :     calc_cbp = calc_cbp_altivec;
401 : canard 1.7 fdct = fdct_altivec;
402 :     idct = idct_altivec;
403 : canard 1.10 sadInit = sadInit_altivec;
404 : canard 1.8 sad16 = sad16_altivec;
405 :     sad8 = sad8_altivec;
406 :     dev16 = dev16_altivec;
407 : canard 1.6 #else
408 : canard 1.5 calc_cbp = calc_cbp_ppc;
409 : canard 1.6 #endif
410 : canard 1.5 #endif
411 : Isibaar 1.1
412 :     return XVID_ERR_OK;
413 :     }
414 :    
415 : edgomez 1.16 /*****************************************************************************
416 :     * XviD Native decoder entry point
417 :     *
418 :     * This function is just a wrapper to all the option cases.
419 :     *
420 :     * Returned values : XVID_ERR_FAIL when opt is invalid
421 :     * else returns the wrapped function result
422 :     *
423 :     ****************************************************************************/
424 :    
425 : edgomez 1.15 int
426 :     xvid_decore(void *handle,
427 :     int opt,
428 :     void *param1,
429 :     void *param2)
430 : Isibaar 1.1 {
431 : edgomez 1.15 switch (opt) {
432 :     case XVID_DEC_DECODE:
433 :     return decoder_decode((DECODER *) handle, (XVID_DEC_FRAME *) param1);
434 :    
435 :     case XVID_DEC_CREATE:
436 : chenm001 1.29 return decoder_create((XVID_DEC_PARAM *) param1);
437 : edgomez 1.15
438 :     case XVID_DEC_DESTROY:
439 :     return decoder_destroy((DECODER *) handle);
440 : Isibaar 1.1
441 :     default:
442 : edgomez 1.15 return XVID_ERR_FAIL;
443 :     }
444 : Isibaar 1.1 }
445 :    
446 : edgomez 1.16
447 :     /*****************************************************************************
448 :     * XviD Native encoder entry point
449 :     *
450 :     * This function is just a wrapper to all the option cases.
451 :     *
452 :     * Returned values : XVID_ERR_FAIL when opt is invalid
453 :     * else returns the wrapped function result
454 :     *
455 :     ****************************************************************************/
456 : Isibaar 1.1
457 : edgomez 1.15 int
458 :     xvid_encore(void *handle,
459 :     int opt,
460 :     void *param1,
461 :     void *param2)
462 : Isibaar 1.1 {
463 : edgomez 1.15 switch (opt) {
464 :     case XVID_ENC_ENCODE:
465 : suxen_drol 1.20 #ifdef BFRAMES
466 :     if (((Encoder *) handle)->mbParam.max_bframes >= 0)
467 :     return encoder_encode_bframes((Encoder *) handle, (XVID_ENC_FRAME *) param1,
468 :     (XVID_ENC_STATS *) param2);
469 :     else
470 :     #endif
471 : edgomez 1.15 return encoder_encode((Encoder *) handle, (XVID_ENC_FRAME *) param1,
472 :     (XVID_ENC_STATS *) param2);
473 :    
474 :     case XVID_ENC_CREATE:
475 :     return encoder_create((XVID_ENC_PARAM *) param1);
476 :    
477 :     case XVID_ENC_DESTROY:
478 :     return encoder_destroy((Encoder *) handle);
479 : Isibaar 1.1
480 :     default:
481 : edgomez 1.15 return XVID_ERR_FAIL;
482 :     }
483 : Isibaar 1.1 }

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