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

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.45.2.24 - (view) (download)

1 : edgomez 1.16 /*****************************************************************************
2 : edgomez 1.17 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : edgomez 1.45.2.6 * 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 1.45.2.24 * $Id$
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 1.45.2.16 #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 : Isibaar 1.45.2.13 #include "image/qpel.h"
49 : Isibaar 1.45.2.23 #include "image/postprocessing.h"
50 : Isibaar 1.1
51 : suxen_drol 1.45.2.4 #if defined(_DEBUG)
52 :     unsigned int xvid_debug = 0; /* xvid debug mask */
53 :     #endif
54 :    
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 1.45.2.15 *
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 1.45.2.15 static
162 : suxen_drol 1.45.2.1 int xvid_gbl_init(xvid_gbl_init_t * init)
163 : Isibaar 1.1 {
164 : suxen_drol 1.45.2.1 unsigned int cpu_flags;
165 : suxen_drol 1.31
166 : edgomez 1.45.2.12 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
167 : suxen_drol 1.45.2.1 return XVID_ERR_VERSION;
168 : suxen_drol 1.31
169 : suxen_drol 1.45.2.1 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 : Isibaar 1.45.2.23 init_postproc();
175 : Isibaar 1.9
176 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
177 : Isibaar 1.1 fdct = fdct_int32;
178 :     idct = idct_int32;
179 :    
180 : edgomez 1.16 /* Only needed on PPC Altivec archs */
181 : canard 1.10 sadInit = 0;
182 : edgomez 1.15
183 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
184 : Isibaar 1.1 emms = emms_c;
185 :    
186 : Isibaar 1.45.2.13 /* Qpel stuff */
187 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
188 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
189 : edgomez 1.45.2.14 xvid_Init_QP();
190 : Isibaar 1.45.2.13
191 : edgomez 1.16 /* Quantization functions */
192 : edgomez 1.45.2.16 quant_h263_intra = quant_h263_intra_c;
193 :     quant_h263_inter = quant_h263_inter_c;
194 :     dequant_h263_intra = dequant_h263_intra_c;
195 :     dequant_h263_inter = dequant_h263_inter_c;
196 :    
197 :     quant_mpeg_intra = quant_mpeg_intra_c;
198 :     quant_mpeg_inter = quant_mpeg_inter_c;
199 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
200 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
201 : Isibaar 1.1
202 : edgomez 1.16 /* Block transfer related functions */
203 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
204 :     transfer_16to8copy = transfer_16to8copy_c;
205 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
206 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_c;
207 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
208 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
209 :     transfer8x8_copy = transfer8x8_copy_c;
210 : Isibaar 1.1
211 : edgomez 1.41 /* Interlacing functions */
212 :     MBFieldTest = MBFieldTest_c;
213 :    
214 : edgomez 1.16 /* Image interpolation related functions */
215 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
216 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
217 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
218 :    
219 : edgomez 1.41 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
220 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
221 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
222 :    
223 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
224 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
225 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
226 :    
227 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
228 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
229 :    
230 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
231 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
232 :    
233 : edgomez 1.45.2.16 /* reduced resolution */
234 : edgomez 1.41 copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_C;
235 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_C;
236 :     vfilter_31 = xvid_VFilter_31_C;
237 :     hfilter_31 = xvid_HFilter_31_C;
238 :     filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_C;
239 :     filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_C;
240 :    
241 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
242 : Isibaar 1.1 colorspace_init();
243 :    
244 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
245 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_c;
246 :     rgb555_to_yv12 = rgb555_to_yv12_c;
247 :     rgb565_to_yv12 = rgb565_to_yv12_c;
248 :     bgr_to_yv12 = bgr_to_yv12_c;
249 :     bgra_to_yv12 = bgra_to_yv12_c;
250 :     abgr_to_yv12 = abgr_to_yv12_c;
251 :     rgba_to_yv12 = rgba_to_yv12_c;
252 :     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 :     yuyvi_to_yv12 = yuyvi_to_yv12_c;
262 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
263 :    
264 : Isibaar 1.1
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 :     yv12_to_yuyv = yv12_to_yuyv_c;
273 :     yv12_to_uyvy = yv12_to_uyvy_c;
274 : edgomez 1.45.2.15
275 : edgomez 1.41 yv12_to_rgb555i = yv12_to_rgb555i_c;
276 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
277 :     yv12_to_bgri = yv12_to_bgri_c;
278 :     yv12_to_bgrai = yv12_to_bgrai_c;
279 :     yv12_to_abgri = yv12_to_abgri_c;
280 :     yv12_to_rgbai = yv12_to_rgbai_c;
281 :     yv12_to_yuyvi = yv12_to_yuyvi_c;
282 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
283 : Isibaar 1.1
284 : edgomez 1.16 /* Functions used in motion estimation algorithms */
285 : edgomez 1.45.2.20 calc_cbp = calc_cbp_c;
286 :     sad16 = sad16_c;
287 :     sad8 = sad8_c;
288 :     sad16bi = sad16bi_c;
289 :     sad8bi = sad8bi_c;
290 :     dev16 = dev16_c;
291 :     sad16v = sad16v_c;
292 :     sse8_16bit = sse8_16bit_c;
293 : edgomez 1.45.2.15
294 : edgomez 1.41 #if defined(ARCH_IS_IA32)
295 :    
296 : edgomez 1.45.2.7 if ((cpu_flags & XVID_CPU_ASM)) {
297 : edgomez 1.41 vfilter_31 = xvid_VFilter_31_x86;
298 :     hfilter_31 = xvid_HFilter_31_x86;
299 :     }
300 :    
301 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
302 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
303 :     (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2))
304 :     {
305 :     /* Restore FPU context : emms_c is a nop functions */
306 :     emms = emms_mmx;
307 :     }
308 :    
309 :     if ((cpu_flags & XVID_CPU_MMX)) {
310 : edgomez 1.16
311 :     /* Forward and Inverse Discrete Cosine Transformation functions */
312 : edgomez 1.45.2.17 fdct = fdct_mmx_skal;
313 : Isibaar 1.45.2.8 idct = idct_mmx;
314 : Isibaar 1.1
315 : Isibaar 1.45.2.13 /* Qpel stuff */
316 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
317 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
318 :    
319 : edgomez 1.16 /* Quantization related functions */
320 : edgomez 1.45.2.16 quant_h263_intra = quant_h263_intra_mmx;
321 :     quant_h263_inter = quant_h263_inter_mmx;
322 :     dequant_h263_intra = dequant_h263_intra_mmx;
323 :     dequant_h263_inter = dequant_h263_inter_mmx;
324 :    
325 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
326 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
327 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
328 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
329 : Isibaar 1.1
330 : edgomez 1.16 /* Block related functions */
331 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
332 :     transfer_16to8copy = transfer_16to8copy_mmx;
333 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
334 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_mmx;
335 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
336 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
337 :     transfer8x8_copy = transfer8x8_copy_mmx;
338 : edgomez 1.22
339 : edgomez 1.41 /* Interlacing Functions */
340 :     MBFieldTest = MBFieldTest_mmx;
341 : edgomez 1.16
342 :     /* Image Interpolation related functions */
343 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
344 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
345 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
346 :    
347 : edgomez 1.41 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
348 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
349 :    
350 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
351 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
352 :    
353 :     /* reduced resolution */
354 :     copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_mmx;
355 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_mmx;
356 :     hfilter_31 = xvid_HFilter_31_mmx;
357 :     filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_mmx;
358 :     filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_mmx;
359 :    
360 :     /* image input xxx_to_yv12 related functions */
361 :     yv12_to_yv12 = yv12_to_yv12_mmx;
362 :     bgr_to_yv12 = bgr_to_yv12_mmx;
363 :     bgra_to_yv12 = bgra_to_yv12_mmx;
364 : edgomez 1.16 yuyv_to_yv12 = yuyv_to_yv12_mmx;
365 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
366 : Isibaar 1.1
367 : edgomez 1.41 /* image output yv12_to_xxx related functions */
368 :     yv12_to_bgr = yv12_to_bgr_mmx;
369 :     yv12_to_bgra = yv12_to_bgra_mmx;
370 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
371 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
372 : edgomez 1.45.2.15
373 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
374 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
375 : Isibaar 1.1
376 : edgomez 1.16 /* Motion estimation related functions */
377 : Isibaar 1.1 calc_cbp = calc_cbp_mmx;
378 : edgomez 1.16 sad16 = sad16_mmx;
379 :     sad8 = sad8_mmx;
380 : suxen_drol 1.33 sad16bi = sad16bi_mmx;
381 :     sad8bi = sad8bi_mmx;
382 : edgomez 1.16 dev16 = dev16_mmx;
383 : edgomez 1.41 sad16v = sad16v_mmx;
384 : edgomez 1.45.2.20 sse8_16bit = sse8_16bit_mmx;
385 : Isibaar 1.1 }
386 :    
387 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
388 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
389 :    
390 :     emms = emms_3dn;
391 : suxen_drol 1.33
392 :     /* ME functions */
393 :     sad16bi = sad16bi_3dn;
394 :     sad8bi = sad8bi_3dn;
395 : edgomez 1.41
396 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
397 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
398 : suxen_drol 1.33 }
399 :    
400 :    
401 : edgomez 1.41 if ((cpu_flags & XVID_CPU_MMXEXT)) {
402 : edgomez 1.16
403 : edgomez 1.45.2.17 /* DCT */
404 :     fdct = fdct_xmm_skal;
405 : Isibaar 1.1 idct = idct_xmm;
406 : edgomez 1.16
407 :     /* Interpolation */
408 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
409 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
410 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
411 : Isibaar 1.25
412 : edgomez 1.41 /* reduced resolution */
413 :     copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm;
414 :     add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm;
415 :    
416 : chenm001 1.29 /* Quantization */
417 : edgomez 1.45.2.16 quant_mpeg_intra = quant_mpeg_intra_xmm;
418 :     quant_mpeg_inter = quant_mpeg_inter_xmm;
419 : edgomez 1.41
420 : edgomez 1.45.2.16 dequant_h263_intra = dequant_h263_intra_xmm;
421 :     dequant_h263_inter = dequant_h263_inter_xmm;
422 : chenm001 1.29
423 : edgomez 1.19 /* Buffer transfer */
424 :     transfer_8to16sub2 = transfer_8to16sub2_xmm;
425 : edgomez 1.16
426 :     /* Colorspace transformation */
427 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_xmm;
428 :     yuyv_to_yv12 = yuyv_to_yv12_xmm;
429 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
430 : Isibaar 1.1
431 : edgomez 1.16 /* ME functions */
432 : Isibaar 1.1 sad16 = sad16_xmm;
433 : suxen_drol 1.33 sad8 = sad8_xmm;
434 : chenm001 1.29 sad16bi = sad16bi_xmm;
435 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
436 : Isibaar 1.1 dev16 = dev16_xmm;
437 : edgomez 1.41 sad16v = sad16v_xmm;
438 : Isibaar 1.1 }
439 :    
440 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
441 : edgomez 1.16
442 :     /* Interpolation */
443 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
444 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
445 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
446 : Isibaar 1.1 }
447 :    
448 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
449 :    
450 :     /* Inverse DCT */
451 :     idct = idct_3dne;
452 :    
453 :     /* Buffer transfer */
454 :     transfer_8to16copy = transfer_8to16copy_3dne;
455 :     transfer_16to8copy = transfer_16to8copy_3dne;
456 :     transfer_8to16sub = transfer_8to16sub_3dne;
457 :     transfer_8to16subro = transfer_8to16subro_3dne;
458 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
459 :     transfer_16to8add = transfer_16to8add_3dne;
460 :     transfer8x8_copy = transfer8x8_copy_3dne;
461 :    
462 :     /* Quantization */
463 : edgomez 1.45.2.16 quant_h263_intra = quant_h263_intra_3dne;
464 :     quant_h263_inter = quant_h263_inter_3dne;
465 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne;
466 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
467 :     dequant_h263_intra = dequant_h263_intra_3dne;
468 :     dequant_h263_inter = dequant_h263_inter_3dne;
469 : edgomez 1.41
470 :     /* ME functions */
471 :     calc_cbp = calc_cbp_3dne;
472 :     sad16 = sad16_3dne;
473 :     sad8 = sad8_3dne;
474 :     sad16bi = sad16bi_3dne;
475 :     sad8bi = sad8bi_3dne;
476 :     dev16 = dev16_3dne;
477 :    
478 :     /* Interpolation */
479 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
480 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
481 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
482 :     }
483 :    
484 :     if ((cpu_flags & XVID_CPU_SSE2)) {
485 : edgomez 1.16
486 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
487 :    
488 : edgomez 1.16 /* Quantization */
489 : edgomez 1.45.2.16 quant_h263_intra = quant_h263_intra_sse2;
490 :     quant_h263_inter = quant_h263_inter_sse2;
491 :     dequant_h263_intra = dequant_h263_intra_sse2;
492 :     dequant_h263_inter = dequant_h263_inter_sse2;
493 : h 1.13
494 : edgomez 1.45.2.18 /* SAD operators */
495 : edgomez 1.16 sad16 = sad16_sse2;
496 :     dev16 = dev16_sse2;
497 : edgomez 1.45.2.18
498 : edgomez 1.45.2.24 /* DCT operators
499 :     * no iDCT because it's not "Walken matching" */
500 : edgomez 1.45.2.18 fdct = fdct_sse2_skal;
501 : h 1.12 }
502 : edgomez 1.45.2.24 #endif /* ARCH_IS_IA32 */
503 : Isibaar 1.18
504 : edgomez 1.41 #if defined(ARCH_IS_IA64)
505 : edgomez 1.43 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
506 : Isibaar 1.18 idct_ia64_init();
507 :     fdct = fdct_ia64;
508 : edgomez 1.43 idct = idct_ia64; /*not yet working, crashes */
509 : Isibaar 1.18 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
510 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
511 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
512 :     sad16 = sad16_ia64;
513 :     sad16bi = sad16bi_ia64;
514 :     sad8 = sad8_ia64;
515 :     dev16 = dev16_ia64;
516 : edgomez 1.43 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
517 : edgomez 1.45.2.16 quant_h263_intra = quant_h263_intra_ia64;
518 :     quant_h263_inter = quant_h263_inter_ia64;
519 :     dequant_h263_intra = dequant_h263_intra_ia64;
520 :     dequant_h263_inter = dequant_h263_inter_ia64;
521 : Isibaar 1.18 transfer_8to16copy = transfer_8to16copy_ia64;
522 :     transfer_16to8copy = transfer_16to8copy_ia64;
523 :     transfer_8to16sub = transfer_8to16sub_ia64;
524 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
525 :     transfer_16to8add = transfer_16to8add_ia64;
526 :     transfer8x8_copy = transfer8x8_copy_ia64;
527 :     }
528 : edgomez 1.45.2.15 #endif
529 : edgomez 1.16
530 : edgomez 1.41 #if defined(ARCH_IS_PPC)
531 :     if ((cpu_flags & XVID_CPU_ASM))
532 :     {
533 :     calc_cbp = calc_cbp_ppc;
534 :     }
535 :    
536 :     if ((cpu_flags & XVID_CPU_ALTIVEC))
537 :     {
538 :     calc_cbp = calc_cbp_altivec;
539 :     fdct = fdct_altivec;
540 :     idct = idct_altivec;
541 :     sadInit = sadInit_altivec;
542 :     sad16 = sad16_altivec;
543 :     sad8 = sad8_altivec;
544 :     dev16 = dev16_altivec;
545 :     }
546 : canard 1.6 #endif
547 : edgomez 1.41
548 : suxen_drol 1.45.2.4 #if defined(_DEBUG)
549 :     xvid_debug = init->debug;
550 :     #endif
551 :    
552 :     return 0;
553 : edgomez 1.41 }
554 :    
555 :    
556 : suxen_drol 1.45.2.1 static int
557 :     xvid_gbl_info(xvid_gbl_info_t * info)
558 :     {
559 : edgomez 1.45.2.12 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
560 : suxen_drol 1.45.2.1 return XVID_ERR_VERSION;
561 :    
562 :     info->actual_version = XVID_VERSION;
563 : edgomez 1.45.2.21 info->build = "xvid-1.0.0";
564 : suxen_drol 1.45.2.1 info->cpu_flags = detect_cpu_flags();
565 :    
566 :     #if defined(_SMP) && defined(WIN32)
567 :     info->num_threads = pthread_num_processors_np();;
568 :     #else
569 :     info->num_threads = 0;
570 :     #endif
571 :    
572 :     return 0;
573 :     }
574 :    
575 : edgomez 1.41
576 :     static int
577 : suxen_drol 1.45.2.1 xvid_gbl_convert(xvid_gbl_convert_t* convert)
578 : edgomez 1.41 {
579 : suxen_drol 1.45.2.1 int width;
580 :     int height;
581 :     int width2;
582 :     int height2;
583 : edgomez 1.41 IMAGE img;
584 :    
585 : edgomez 1.45.2.12 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
586 : suxen_drol 1.45.2.1 return XVID_ERR_VERSION;
587 :    
588 : edgomez 1.45.2.5 #if 0
589 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
590 :     #endif
591 : suxen_drol 1.45.2.1 width = convert->width;
592 :     height = convert->height;
593 :     width2 = convert->width/2;
594 :     height2 = convert->height/2;
595 :    
596 :     switch (convert->input.csp & ~XVID_CSP_VFLIP)
597 : edgomez 1.41 {
598 :     case XVID_CSP_YV12 :
599 : suxen_drol 1.45.2.1 img.y = convert->input.plane[0];
600 : edgomez 1.45.2.15 img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
601 : suxen_drol 1.45.2.1 img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
602 : edgomez 1.41 image_output(&img, width, height, width,
603 : suxen_drol 1.45.2.1 (uint8_t**)convert->output.plane, convert->output.stride,
604 :     convert->output.csp, convert->interlacing);
605 : edgomez 1.41 break;
606 :    
607 :     default :
608 :     return XVID_ERR_FORMAT;
609 :     }
610 :    
611 :    
612 :     emms();
613 : suxen_drol 1.45.2.1 return 0;
614 : edgomez 1.41 }
615 :    
616 : suxen_drol 1.45.2.1 /*****************************************************************************
617 :     * XviD Global Entry point
618 :     *
619 :     * Well this function initialize all internal function pointers according
620 :     * to the CPU features forced by the library client or autodetected (depending
621 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
622 :     * image colorspace transformation tables.
623 : edgomez 1.45.2.15 *
624 : suxen_drol 1.45.2.1 ****************************************************************************/
625 :    
626 :    
627 : edgomez 1.41 int
628 : suxen_drol 1.45.2.1 xvid_global(void *handle,
629 : edgomez 1.41 int opt,
630 :     void *param1,
631 :     void *param2)
632 :     {
633 :     switch(opt)
634 :     {
635 : suxen_drol 1.45.2.1 case XVID_GBL_INIT :
636 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
637 : edgomez 1.41
638 : suxen_drol 1.45.2.1 case XVID_GBL_INFO :
639 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
640 : edgomez 1.41
641 : suxen_drol 1.45.2.1 case XVID_GBL_CONVERT :
642 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
643 :    
644 : edgomez 1.41 default :
645 :     return XVID_ERR_FAIL;
646 :     }
647 :     }
648 :    
649 : edgomez 1.16 /*****************************************************************************
650 :     * XviD Native decoder entry point
651 :     *
652 :     * This function is just a wrapper to all the option cases.
653 :     *
654 :     * Returned values : XVID_ERR_FAIL when opt is invalid
655 :     * else returns the wrapped function result
656 :     *
657 :     ****************************************************************************/
658 :    
659 : edgomez 1.15 int
660 :     xvid_decore(void *handle,
661 :     int opt,
662 :     void *param1,
663 :     void *param2)
664 : Isibaar 1.1 {
665 : edgomez 1.15 switch (opt) {
666 :     case XVID_DEC_CREATE:
667 : suxen_drol 1.45.2.1 return decoder_create((xvid_dec_create_t *) param1);
668 : edgomez 1.15
669 :     case XVID_DEC_DESTROY:
670 :     return decoder_destroy((DECODER *) handle);
671 : Isibaar 1.1
672 : suxen_drol 1.45.2.1 case XVID_DEC_DECODE:
673 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
674 :    
675 : Isibaar 1.1 default:
676 : edgomez 1.15 return XVID_ERR_FAIL;
677 :     }
678 : Isibaar 1.1 }
679 :    
680 : edgomez 1.16
681 :     /*****************************************************************************
682 :     * XviD Native encoder entry point
683 :     *
684 :     * This function is just a wrapper to all the option cases.
685 :     *
686 :     * Returned values : XVID_ERR_FAIL when opt is invalid
687 :     * else returns the wrapped function result
688 :     *
689 :     ****************************************************************************/
690 : Isibaar 1.1
691 : edgomez 1.15 int
692 :     xvid_encore(void *handle,
693 :     int opt,
694 :     void *param1,
695 :     void *param2)
696 : Isibaar 1.1 {
697 : edgomez 1.15 switch (opt) {
698 :     case XVID_ENC_ENCODE:
699 : edgomez 1.41
700 : suxen_drol 1.45.2.1 return enc_encode((Encoder *) handle,
701 :     (xvid_enc_frame_t *) param1,
702 :     (xvid_enc_stats_t *) param2);
703 : edgomez 1.15
704 :     case XVID_ENC_CREATE:
705 : suxen_drol 1.45.2.2 return enc_create((xvid_enc_create_t *) param1);
706 : edgomez 1.15
707 :     case XVID_ENC_DESTROY:
708 : suxen_drol 1.45.2.1 return enc_destroy((Encoder *) handle);
709 : Isibaar 1.1
710 :     default:
711 : edgomez 1.15 return XVID_ERR_FAIL;
712 :     }
713 : Isibaar 1.1 }

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