[svn] / branches / dev-api-4 / xvidcore / src / xvid.c Repository:
ViewVC logotype

Annotation of /branches/dev-api-4/xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1174 - (view) (download)

1 : edgomez 200 /*****************************************************************************
2 : edgomez 201 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : edgomez 1054 * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org>
7 :     *
8 : edgomez 851 * 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 201 * (at your option) any later version.
12 :     *
13 :     * This program is distributed in the hope that it will be useful,
14 : edgomez 851 * but WITHOUT ANY WARRANTY ; without even the implied warranty of
15 : edgomez 201 * 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 851 * along with this program ; if not, write to the Free Software
20 : edgomez 201 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 :     *
22 : edgomez 1174 * $Id: xvid.c,v 1.45.2.16 2003-10-07 13:02:35 edgomez Exp $
23 : edgomez 410 *
24 : edgomez 201 ****************************************************************************/
25 : chenm001 274
26 : edgomez 851 #include <stdio.h>
27 :     #include <stdlib.h>
28 :     #include <string.h>
29 :     #include <time.h>
30 :    
31 : Isibaar 3 #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 851 #include "image/reduced.h"
40 : Isibaar 3 #include "utils/mem_transfer.h"
41 : edgomez 851 #include "utils/mbfunctions.h"
42 : edgomez 1174 #include "quant/quant.h"
43 : ia64p 299 #include "motion/motion.h"
44 : Isibaar 3 #include "motion/sad.h"
45 :     #include "utils/emms.h"
46 :     #include "utils/timer.h"
47 : Isibaar 100 #include "bitstream/mbcoding.h"
48 : Isibaar 1125 #include "image/qpel.h"
49 : Isibaar 3
50 : suxen_drol 1031 #if defined(_DEBUG)
51 :     unsigned int xvid_debug = 0; /* xvid debug mask */
52 :     #endif
53 :    
54 : edgomez 851 #if defined(ARCH_IS_IA32)
55 :     #if defined(_MSC_VER)
56 :     # include <windows.h>
57 : suxen_drol 311 #else
58 : edgomez 851 # include <signal.h>
59 :     # include <setjmp.h>
60 : suxen_drol 311
61 : edgomez 851 static jmp_buf mark;
62 : suxen_drol 311
63 : edgomez 851 static void
64 :     sigill_handler(int signal)
65 :     {
66 :     longjmp(mark, 1);
67 :     }
68 : suxen_drol 311 #endif
69 :    
70 :    
71 :     /*
72 : edgomez 874 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
73 :     * signalled
74 :     *
75 :     * Return values:
76 :     * -1 : could not determine
77 :     * 0 : SIGILL was *not* signalled
78 :     * 1 : SIGILL was signalled
79 :     */
80 : suxen_drol 311
81 :     int
82 :     sigill_check(void (*func)())
83 :     {
84 : edgomez 851 #if defined(_MSC_VER)
85 : suxen_drol 311 _try {
86 :     func();
87 :     }
88 :     _except(EXCEPTION_EXECUTE_HANDLER) {
89 :    
90 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
91 :     return 1;
92 :     }
93 :     return 0;
94 :     #else
95 :     void * old_handler;
96 :     int jmpret;
97 :    
98 :    
99 :     old_handler = signal(SIGILL, sigill_handler);
100 :     if (old_handler == SIG_ERR)
101 :     {
102 :     return -1;
103 :     }
104 :    
105 :     jmpret = setjmp(mark);
106 :     if (jmpret == 0)
107 :     {
108 :     func();
109 :     }
110 :    
111 :     signal(SIGILL, old_handler);
112 :    
113 :     return jmpret;
114 :     #endif
115 :     }
116 :     #endif
117 :    
118 : edgomez 851
119 :     /* detect cpu flags */
120 :     static unsigned int
121 :     detect_cpu_flags()
122 :     {
123 :     /* enable native assembly optimizations by default */
124 :     unsigned int cpu_flags = XVID_CPU_ASM;
125 :    
126 :     #if defined(ARCH_IS_IA32)
127 :     cpu_flags |= check_cpu_features();
128 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
129 :     cpu_flags &= ~XVID_CPU_SSE;
130 :    
131 :     if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger))
132 :     cpu_flags &= ~XVID_CPU_SSE2;
133 :     #endif
134 :    
135 :     #if defined(ARCH_IS_PPC)
136 :     #if defined(ARCH_IS_PPC_ALTIVEC)
137 :     cpu_flags |= XVID_CPU_ALTIVEC;
138 :     #endif
139 :     #endif
140 :    
141 :     return cpu_flags;
142 :     }
143 :    
144 :    
145 : edgomez 200 /*****************************************************************************
146 :     * XviD Init Entry point
147 :     *
148 :     * Well this function initialize all internal function pointers according
149 :     * to the CPU features forced by the library client or autodetected (depending
150 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
151 :     * image colorspace transformation tables.
152 : edgomez 1161 *
153 : edgomez 200 * Returned value : XVID_ERR_OK
154 :     * + API_VERSION in the input XVID_INIT_PARAM structure
155 :     * + core build " " " " "
156 :     *
157 :     ****************************************************************************/
158 :    
159 : edgomez 851
160 : edgomez 1161 static
161 : suxen_drol 890 int xvid_gbl_init(xvid_gbl_init_t * init)
162 : Isibaar 3 {
163 : suxen_drol 890 unsigned int cpu_flags;
164 : Isibaar 3
165 : edgomez 1107 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
166 : suxen_drol 890 return XVID_ERR_VERSION;
167 : suxen_drol 234
168 : suxen_drol 890 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
169 : suxen_drol 234
170 : edgomez 200 /* Initialize the function pointers */
171 : Isibaar 3 idct_int32_init();
172 : Isibaar 100 init_vlc_tables();
173 :    
174 : edgomez 200 /* Fixed Point Forward/Inverse DCT transformations */
175 : Isibaar 3 fdct = fdct_int32;
176 :     idct = idct_int32;
177 :    
178 : edgomez 200 /* Only needed on PPC Altivec archs */
179 : canard 115 sadInit = 0;
180 : edgomez 195
181 : edgomez 200 /* Restore FPU context : emms_c is a nop functions */
182 : Isibaar 3 emms = emms_c;
183 :    
184 : Isibaar 1125 /* Qpel stuff */
185 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
186 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
187 : edgomez 1127 xvid_Init_QP();
188 : Isibaar 1125
189 : edgomez 200 /* Quantization functions */
190 : edgomez 1174 quant_h263_intra = quant_h263_intra_c;
191 :     quant_h263_inter = quant_h263_inter_c;
192 :     dequant_h263_intra = dequant_h263_intra_c;
193 :     dequant_h263_inter = dequant_h263_inter_c;
194 : Isibaar 3
195 : edgomez 1174 quant_mpeg_intra = quant_mpeg_intra_c;
196 :     quant_mpeg_inter = quant_mpeg_inter_c;
197 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
198 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
199 : Isibaar 3
200 : edgomez 200 /* Block transfer related functions */
201 : Isibaar 3 transfer_8to16copy = transfer_8to16copy_c;
202 :     transfer_16to8copy = transfer_16to8copy_c;
203 : edgomez 200 transfer_8to16sub = transfer_8to16sub_c;
204 : edgomez 851 transfer_8to16subro = transfer_8to16subro_c;
205 : suxen_drol 118 transfer_8to16sub2 = transfer_8to16sub2_c;
206 : edgomez 200 transfer_16to8add = transfer_16to8add_c;
207 :     transfer8x8_copy = transfer8x8_copy_c;
208 : Isibaar 3
209 : edgomez 851 /* Interlacing functions */
210 :     MBFieldTest = MBFieldTest_c;
211 :    
212 : edgomez 200 /* Image interpolation related functions */
213 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
214 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
215 : Isibaar 3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
216 :    
217 : edgomez 851 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
218 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
219 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
220 :    
221 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
222 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
223 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
224 :    
225 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
226 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
227 :    
228 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
229 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
230 :    
231 : edgomez 1174 /* reduced resolution */
232 : edgomez 851 copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_C;
233 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_C;
234 :     vfilter_31 = xvid_VFilter_31_C;
235 :     hfilter_31 = xvid_HFilter_31_C;
236 :     filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_C;
237 :     filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_C;
238 :    
239 : edgomez 200 /* Initialize internal colorspace transformation tables */
240 : Isibaar 3 colorspace_init();
241 :    
242 : edgomez 200 /* All colorspace transformation functions User Format->YV12 */
243 : edgomez 851 yv12_to_yv12 = yv12_to_yv12_c;
244 :     rgb555_to_yv12 = rgb555_to_yv12_c;
245 :     rgb565_to_yv12 = rgb565_to_yv12_c;
246 :     bgr_to_yv12 = bgr_to_yv12_c;
247 :     bgra_to_yv12 = bgra_to_yv12_c;
248 :     abgr_to_yv12 = abgr_to_yv12_c;
249 :     rgba_to_yv12 = rgba_to_yv12_c;
250 :     yuyv_to_yv12 = yuyv_to_yv12_c;
251 :     uyvy_to_yv12 = uyvy_to_yv12_c;
252 : Isibaar 3
253 : edgomez 851 rgb555i_to_yv12 = rgb555i_to_yv12_c;
254 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
255 :     bgri_to_yv12 = bgri_to_yv12_c;
256 :     bgrai_to_yv12 = bgrai_to_yv12_c;
257 :     abgri_to_yv12 = abgri_to_yv12_c;
258 :     rgbai_to_yv12 = rgbai_to_yv12_c;
259 :     yuyvi_to_yv12 = yuyvi_to_yv12_c;
260 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
261 :    
262 :    
263 : edgomez 200 /* All colorspace transformation functions YV12->User format */
264 : edgomez 851 yv12_to_rgb555 = yv12_to_rgb555_c;
265 :     yv12_to_rgb565 = yv12_to_rgb565_c;
266 :     yv12_to_bgr = yv12_to_bgr_c;
267 :     yv12_to_bgra = yv12_to_bgra_c;
268 :     yv12_to_abgr = yv12_to_abgr_c;
269 :     yv12_to_rgba = yv12_to_rgba_c;
270 :     yv12_to_yuyv = yv12_to_yuyv_c;
271 :     yv12_to_uyvy = yv12_to_uyvy_c;
272 : edgomez 1161
273 : edgomez 851 yv12_to_rgb555i = yv12_to_rgb555i_c;
274 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
275 :     yv12_to_bgri = yv12_to_bgri_c;
276 :     yv12_to_bgrai = yv12_to_bgrai_c;
277 :     yv12_to_abgri = yv12_to_abgri_c;
278 :     yv12_to_rgbai = yv12_to_rgbai_c;
279 :     yv12_to_yuyvi = yv12_to_yuyvi_c;
280 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
281 : Isibaar 3
282 : edgomez 200 /* Functions used in motion estimation algorithms */
283 : Isibaar 3 calc_cbp = calc_cbp_c;
284 : edgomez 200 sad16 = sad16_c;
285 : suxen_drol 329 sad8 = sad8_c;
286 : edgomez 200 sad16bi = sad16bi_c;
287 : suxen_drol 329 sad8bi = sad8bi_c;
288 : edgomez 200 dev16 = dev16_c;
289 : edgomez 851 sad16v = sad16v_c;
290 : edgomez 1161
291 : edgomez 874 /* Halfpel8_Refine = Halfpel8_Refine_c; */
292 : Isibaar 3
293 : edgomez 851 #if defined(ARCH_IS_IA32)
294 : edgomez 200
295 : edgomez 1058 if ((cpu_flags & XVID_CPU_ASM)) {
296 : edgomez 851 vfilter_31 = xvid_VFilter_31_x86;
297 :     hfilter_31 = xvid_HFilter_31_x86;
298 :     }
299 :    
300 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
301 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
302 :     (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2))
303 :     {
304 :     /* Restore FPU context : emms_c is a nop functions */
305 :     emms = emms_mmx;
306 :     }
307 :    
308 :     if ((cpu_flags & XVID_CPU_MMX)) {
309 :    
310 : edgomez 200 /* Forward and Inverse Discrete Cosine Transformation functions */
311 : Isibaar 3 fdct = fdct_mmx;
312 : Isibaar 1066 idct = idct_mmx;
313 : Isibaar 3
314 : Isibaar 1125 /* Qpel stuff */
315 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
316 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
317 :    
318 : edgomez 200 /* Quantization related functions */
319 : edgomez 1174 quant_h263_intra = quant_h263_intra_mmx;
320 :     quant_h263_inter = quant_h263_inter_mmx;
321 :     dequant_h263_intra = dequant_h263_intra_mmx;
322 :     dequant_h263_inter = dequant_h263_inter_mmx;
323 : Isibaar 3
324 : edgomez 1174 quant_mpeg_intra = quant_mpeg_intra_mmx;
325 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
326 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
327 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
328 : Isibaar 3
329 : edgomez 200 /* Block related functions */
330 : Isibaar 3 transfer_8to16copy = transfer_8to16copy_mmx;
331 :     transfer_16to8copy = transfer_16to8copy_mmx;
332 : edgomez 200 transfer_8to16sub = transfer_8to16sub_mmx;
333 : edgomez 851 transfer_8to16subro = transfer_8to16subro_mmx;
334 : edgomez 236 transfer_8to16sub2 = transfer_8to16sub2_mmx;
335 : edgomez 200 transfer_16to8add = transfer_16to8add_mmx;
336 :     transfer8x8_copy = transfer8x8_copy_mmx;
337 : Isibaar 3
338 : edgomez 851 /* Interlacing Functions */
339 :     MBFieldTest = MBFieldTest_mmx;
340 : edgomez 236
341 : edgomez 200 /* Image Interpolation related functions */
342 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
343 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
344 : Isibaar 3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
345 :    
346 : edgomez 851 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
347 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
348 :    
349 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
350 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
351 :    
352 :     /* reduced resolution */
353 :     copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_mmx;
354 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_mmx;
355 :     hfilter_31 = xvid_HFilter_31_mmx;
356 :     filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_mmx;
357 :     filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_mmx;
358 :    
359 :     /* image input xxx_to_yv12 related functions */
360 :     yv12_to_yv12 = yv12_to_yv12_mmx;
361 :     bgr_to_yv12 = bgr_to_yv12_mmx;
362 :     bgra_to_yv12 = bgra_to_yv12_mmx;
363 : edgomez 200 yuyv_to_yv12 = yuyv_to_yv12_mmx;
364 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
365 : Isibaar 3
366 : edgomez 851 /* image output yv12_to_xxx related functions */
367 :     yv12_to_bgr = yv12_to_bgr_mmx;
368 :     yv12_to_bgra = yv12_to_bgra_mmx;
369 : edgomez 200 yv12_to_yuyv = yv12_to_yuyv_mmx;
370 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
371 : edgomez 1161
372 : edgomez 851 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
373 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
374 : Isibaar 3
375 : edgomez 200 /* Motion estimation related functions */
376 : Isibaar 3 calc_cbp = calc_cbp_mmx;
377 : edgomez 200 sad16 = sad16_mmx;
378 :     sad8 = sad8_mmx;
379 : suxen_drol 329 sad16bi = sad16bi_mmx;
380 :     sad8bi = sad8bi_mmx;
381 : edgomez 200 dev16 = dev16_mmx;
382 : edgomez 851 sad16v = sad16v_mmx;
383 : Isibaar 3 }
384 :    
385 : suxen_drol 329 /* these 3dnow functions are faster than mmx, but slower than xmm. */
386 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOW)) {
387 : suxen_drol 329
388 : edgomez 851 emms = emms_3dn;
389 :    
390 : suxen_drol 329 /* ME functions */
391 :     sad16bi = sad16bi_3dn;
392 :     sad8bi = sad8bi_3dn;
393 : edgomez 851
394 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
395 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
396 : suxen_drol 329 }
397 :    
398 :    
399 : edgomez 851 if ((cpu_flags & XVID_CPU_MMXEXT)) {
400 : edgomez 200
401 :     /* Inverse DCT */
402 : Isibaar 3 idct = idct_xmm;
403 : edgomez 200
404 :     /* Interpolation */
405 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
406 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
407 : h 38 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
408 : edgomez 200
409 : edgomez 851 /* reduced resolution */
410 :     copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm;
411 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm;
412 :    
413 : chenm001 274 /* Quantization */
414 : edgomez 1174 quant_mpeg_intra = quant_mpeg_intra_xmm;
415 :     quant_mpeg_inter = quant_mpeg_inter_xmm;
416 : edgomez 851
417 : edgomez 1174 dequant_h263_intra = dequant_h263_intra_xmm;
418 :     dequant_h263_inter = dequant_h263_inter_xmm;
419 : chenm001 274
420 : edgomez 218 /* Buffer transfer */
421 :     transfer_8to16sub2 = transfer_8to16sub2_xmm;
422 :    
423 : edgomez 200 /* Colorspace transformation */
424 : edgomez 851 yv12_to_yv12 = yv12_to_yv12_xmm;
425 :     yuyv_to_yv12 = yuyv_to_yv12_xmm;
426 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
427 : Isibaar 3
428 : edgomez 200 /* ME functions */
429 : Isibaar 3 sad16 = sad16_xmm;
430 : suxen_drol 329 sad8 = sad8_xmm;
431 : chenm001 274 sad16bi = sad16bi_xmm;
432 : suxen_drol 329 sad8bi = sad8bi_xmm;
433 : Isibaar 3 dev16 = dev16_xmm;
434 : edgomez 851 sad16v = sad16v_xmm;
435 : Isibaar 3 }
436 :    
437 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOW)) {
438 : edgomez 200
439 :     /* Interpolation */
440 : Isibaar 3 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
441 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
442 : h 40 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
443 : Isibaar 3 }
444 :    
445 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
446 : edgomez 200
447 : edgomez 851 /* Inverse DCT */
448 :     idct = idct_3dne;
449 :    
450 :     /* Buffer transfer */
451 :     transfer_8to16copy = transfer_8to16copy_3dne;
452 :     transfer_16to8copy = transfer_16to8copy_3dne;
453 :     transfer_8to16sub = transfer_8to16sub_3dne;
454 :     transfer_8to16subro = transfer_8to16subro_3dne;
455 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
456 :     transfer_16to8add = transfer_16to8add_3dne;
457 :     transfer8x8_copy = transfer8x8_copy_3dne;
458 :    
459 :     /* Quantization */
460 : edgomez 1174 quant_h263_intra = quant_h263_intra_3dne;
461 :     quant_h263_inter = quant_h263_inter_3dne;
462 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne;
463 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
464 :     dequant_h263_intra = dequant_h263_intra_3dne;
465 :     dequant_h263_inter = dequant_h263_inter_3dne;
466 : edgomez 851
467 :     /* ME functions */
468 :     calc_cbp = calc_cbp_3dne;
469 :     sad16 = sad16_3dne;
470 :     sad8 = sad8_3dne;
471 :     sad16bi = sad16bi_3dne;
472 :     sad8bi = sad8bi_3dne;
473 :     dev16 = dev16_3dne;
474 :    
475 :     /* Interpolation */
476 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
477 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
478 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
479 :     }
480 :    
481 : Isibaar 1125 #if defined(EXPERIMENTAL_SSE2_CODE) /* mark the whole SSE2 stuff as experimental. At least on
482 :     my P4, it crashes... */
483 : edgomez 851 if ((cpu_flags & XVID_CPU_SSE2)) {
484 :    
485 : chenm001 274 calc_cbp = calc_cbp_sse2;
486 :    
487 : edgomez 200 /* Quantization */
488 : edgomez 1174 quant_h263_intra = quant_h263_intra_sse2;
489 :     quant_h263_inter = quant_h263_inter_sse2;
490 :     dequant_h263_intra = dequant_h263_intra_sse2;
491 :     dequant_h263_inter = dequant_h263_inter_sse2;
492 : h 135
493 : edgomez 851 /* ME; slower than xmm */
494 : edgomez 200 sad16 = sad16_sse2;
495 :     dev16 = dev16_sse2;
496 :     /* Forward and Inverse DCT */
497 : edgomez 1058 #if 0 /* Both function are known to be unprecise, better keep them deactivated */
498 : edgomez 200 idct = idct_sse2;
499 : Isibaar 154 fdct = fdct_sse2;
500 : edgomez 1058 #endif
501 : h 126 }
502 : Isibaar 3 #endif
503 : Isibaar 1125 #endif
504 : edgomez 200
505 : edgomez 851 #if defined(ARCH_IS_IA64)
506 : edgomez 874 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
507 : Isibaar 209 idct_ia64_init();
508 :     fdct = fdct_ia64;
509 : edgomez 874 idct = idct_ia64; /*not yet working, crashes */
510 : Isibaar 209 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
511 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
512 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
513 :     sad16 = sad16_ia64;
514 :     sad16bi = sad16bi_ia64;
515 :     sad8 = sad8_ia64;
516 :     dev16 = dev16_ia64;
517 : edgomez 874 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
518 : edgomez 1174 quant_h263_intra = quant_h263_intra_ia64;
519 :     quant_h263_inter = quant_h263_inter_ia64;
520 :     dequant_h263_intra = dequant_h263_intra_ia64;
521 :     dequant_h263_inter = dequant_h263_inter_ia64;
522 : Isibaar 209 transfer_8to16copy = transfer_8to16copy_ia64;
523 :     transfer_16to8copy = transfer_16to8copy_ia64;
524 :     transfer_8to16sub = transfer_8to16sub_ia64;
525 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
526 :     transfer_16to8add = transfer_16to8add_ia64;
527 :     transfer8x8_copy = transfer8x8_copy_ia64;
528 :     }
529 : edgomez 1161 #endif
530 : Isibaar 209
531 : edgomez 851 #if defined(ARCH_IS_PPC)
532 :     if ((cpu_flags & XVID_CPU_ASM))
533 :     {
534 :     calc_cbp = calc_cbp_ppc;
535 :     }
536 :    
537 :     if ((cpu_flags & XVID_CPU_ALTIVEC))
538 :     {
539 :     calc_cbp = calc_cbp_altivec;
540 :     fdct = fdct_altivec;
541 :     idct = idct_altivec;
542 :     sadInit = sadInit_altivec;
543 :     sad16 = sad16_altivec;
544 :     sad8 = sad8_altivec;
545 :     dev16 = dev16_altivec;
546 :     }
547 : canard 52 #endif
548 : edgomez 851
549 : suxen_drol 1031 #if defined(_DEBUG)
550 :     xvid_debug = init->debug;
551 :     #endif
552 :    
553 :     return 0;
554 : edgomez 851 }
555 :    
556 :    
557 : suxen_drol 890 static int
558 :     xvid_gbl_info(xvid_gbl_info_t * info)
559 :     {
560 : edgomez 1107 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
561 : suxen_drol 890 return XVID_ERR_VERSION;
562 : edgomez 851
563 : suxen_drol 890 info->actual_version = XVID_VERSION;
564 :     info->build = "dev-api-4";
565 :     info->cpu_flags = detect_cpu_flags();
566 :    
567 :     #if defined(_SMP) && defined(WIN32)
568 :     info->num_threads = pthread_num_processors_np();;
569 :     #else
570 :     info->num_threads = 0;
571 :     #endif
572 :    
573 :     return 0;
574 :     }
575 :    
576 :    
577 : edgomez 851 static int
578 : suxen_drol 890 xvid_gbl_convert(xvid_gbl_convert_t* convert)
579 : edgomez 851 {
580 : suxen_drol 890 int width;
581 :     int height;
582 :     int width2;
583 :     int height2;
584 : edgomez 851 IMAGE img;
585 :    
586 : edgomez 1107 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
587 : suxen_drol 890 return XVID_ERR_VERSION;
588 :    
589 : edgomez 1053 #if 0
590 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
591 :     #endif
592 : suxen_drol 890 width = convert->width;
593 :     height = convert->height;
594 :     width2 = convert->width/2;
595 :     height2 = convert->height/2;
596 :    
597 :     switch (convert->input.csp & ~XVID_CSP_VFLIP)
598 : edgomez 851 {
599 :     case XVID_CSP_YV12 :
600 : suxen_drol 890 img.y = convert->input.plane[0];
601 : edgomez 1161 img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
602 : suxen_drol 890 img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
603 : edgomez 851 image_output(&img, width, height, width,
604 : suxen_drol 890 (uint8_t**)convert->output.plane, convert->output.stride,
605 :     convert->output.csp, convert->interlacing);
606 : edgomez 851 break;
607 :    
608 :     default :
609 :     return XVID_ERR_FORMAT;
610 :     }
611 :    
612 :    
613 :     emms();
614 : suxen_drol 890 return 0;
615 : edgomez 851 }
616 :    
617 : suxen_drol 890 /*****************************************************************************
618 :     * XviD Global Entry point
619 :     *
620 :     * Well this function initialize all internal function pointers according
621 :     * to the CPU features forced by the library client or autodetected (depending
622 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
623 :     * image colorspace transformation tables.
624 : edgomez 1161 *
625 : suxen_drol 890 ****************************************************************************/
626 :    
627 :    
628 : edgomez 851 int
629 : suxen_drol 890 xvid_global(void *handle,
630 : edgomez 851 int opt,
631 :     void *param1,
632 :     void *param2)
633 :     {
634 :     switch(opt)
635 :     {
636 : suxen_drol 890 case XVID_GBL_INIT :
637 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
638 : edgomez 851
639 : suxen_drol 890 case XVID_GBL_INFO :
640 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
641 : edgomez 851
642 : suxen_drol 890 case XVID_GBL_CONVERT :
643 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
644 :    
645 : edgomez 851 default :
646 :     return XVID_ERR_FAIL;
647 :     }
648 :     }
649 :    
650 : edgomez 200 /*****************************************************************************
651 :     * XviD Native decoder entry point
652 :     *
653 :     * This function is just a wrapper to all the option cases.
654 :     *
655 :     * Returned values : XVID_ERR_FAIL when opt is invalid
656 :     * else returns the wrapped function result
657 :     *
658 :     ****************************************************************************/
659 :    
660 : edgomez 195 int
661 :     xvid_decore(void *handle,
662 :     int opt,
663 :     void *param1,
664 :     void *param2)
665 : Isibaar 3 {
666 : edgomez 195 switch (opt) {
667 :     case XVID_DEC_CREATE:
668 : suxen_drol 890 return decoder_create((xvid_dec_create_t *) param1);
669 : Isibaar 3
670 : edgomez 195 case XVID_DEC_DESTROY:
671 :     return decoder_destroy((DECODER *) handle);
672 :    
673 : suxen_drol 890 case XVID_DEC_DECODE:
674 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
675 :    
676 : Isibaar 3 default:
677 : edgomez 195 return XVID_ERR_FAIL;
678 :     }
679 : Isibaar 3 }
680 :    
681 :    
682 : edgomez 200 /*****************************************************************************
683 :     * XviD Native encoder entry point
684 :     *
685 :     * This function is just a wrapper to all the option cases.
686 :     *
687 :     * Returned values : XVID_ERR_FAIL when opt is invalid
688 :     * else returns the wrapped function result
689 :     *
690 :     ****************************************************************************/
691 :    
692 : edgomez 195 int
693 :     xvid_encore(void *handle,
694 :     int opt,
695 :     void *param1,
696 :     void *param2)
697 : Isibaar 3 {
698 : edgomez 195 switch (opt) {
699 :     case XVID_ENC_ENCODE:
700 : edgomez 851
701 : suxen_drol 890 return enc_encode((Encoder *) handle,
702 :     (xvid_enc_frame_t *) param1,
703 :     (xvid_enc_stats_t *) param2);
704 : Isibaar 3
705 : edgomez 195 case XVID_ENC_CREATE:
706 : suxen_drol 948 return enc_create((xvid_enc_create_t *) param1);
707 : Isibaar 3
708 : edgomez 195 case XVID_ENC_DESTROY:
709 : suxen_drol 890 return enc_destroy((Encoder *) handle);
710 : edgomez 195
711 : Isibaar 3 default:
712 : edgomez 195 return XVID_ERR_FAIL;
713 :     }
714 : Isibaar 3 }

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