[svn] / branches / release-1_3-branch / xvidcore / src / xvid.c Repository:
ViewVC logotype

Annotation of /branches/release-1_3-branch/xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1835 - (view) (download)
Original Path: trunk/xvidcore/src/xvid.c

1 : edgomez 200 /*****************************************************************************
2 : edgomez 201 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : suxen_drol 1397 * Copyright(C) 2001-2004 Peter Ross <pross@xvid.org>
7 : edgomez 1382 *
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 : Isibaar 1835 * $Id: xvid.c,v 1.80 2008-12-01 15:06:48 Isibaar 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 1835 #if !defined(_WIN32)
32 :     #include <unistd.h>
33 :     #endif
34 :    
35 : Isibaar 3 #include "xvid.h"
36 :     #include "decoder.h"
37 :     #include "encoder.h"
38 :     #include "bitstream/cbp.h"
39 :     #include "dct/idct.h"
40 :     #include "dct/fdct.h"
41 :     #include "image/colorspace.h"
42 :     #include "image/interpolate8x8.h"
43 :     #include "utils/mem_transfer.h"
44 : edgomez 851 #include "utils/mbfunctions.h"
45 : edgomez 1382 #include "quant/quant.h"
46 : ia64p 299 #include "motion/motion.h"
47 : Skal 1709 #include "motion/gmc.h"
48 : Isibaar 3 #include "motion/sad.h"
49 :     #include "utils/emms.h"
50 :     #include "utils/timer.h"
51 : Isibaar 100 #include "bitstream/mbcoding.h"
52 : edgomez 1382 #include "image/qpel.h"
53 :     #include "image/postprocessing.h"
54 : Isibaar 3
55 : edgomez 1382 #if defined(_DEBUG)
56 :     unsigned int xvid_debug = 0; /* xvid debug mask */
57 :     #endif
58 : suxen_drol 311
59 : edgomez 1586 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
60 : edgomez 851 # include <windows.h>
61 : edgomez 1586 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
62 : edgomez 851 # include <signal.h>
63 :     # include <setjmp.h>
64 : suxen_drol 311
65 : edgomez 851 static jmp_buf mark;
66 : suxen_drol 311
67 : edgomez 851 static void
68 :     sigill_handler(int signal)
69 :     {
70 :     longjmp(mark, 1);
71 :     }
72 : suxen_drol 311 #endif
73 :    
74 :    
75 :     /*
76 : edgomez 874 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
77 :     * signalled
78 :     *
79 :     * Return values:
80 :     * -1 : could not determine
81 :     * 0 : SIGILL was *not* signalled
82 :     * 1 : SIGILL was signalled
83 :     */
84 : edgomez 1586 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
85 : edgomez 1412 static int
86 : suxen_drol 311 sigill_check(void (*func)())
87 :     {
88 :     _try {
89 :     func();
90 : edgomez 1412 } _except(EXCEPTION_EXECUTE_HANDLER) {
91 : suxen_drol 311
92 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
93 : edgomez 1412 return(1);
94 : suxen_drol 311 }
95 : edgomez 1412 return(0);
96 :     }
97 : edgomez 1586 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
98 : edgomez 1412 static int
99 :     sigill_check(void (*func)())
100 :     {
101 :     void *old_handler;
102 : suxen_drol 311 int jmpret;
103 :    
104 : edgomez 1412 /* Set our SIGILL handler */
105 :     old_handler = signal(SIGILL, sigill_handler);
106 : suxen_drol 311
107 : edgomez 1412 /* Check for error */
108 :     if (old_handler == SIG_ERR) {
109 :     return(-1);
110 : suxen_drol 311 }
111 :    
112 : edgomez 1412 /* Save stack context, so if func triggers a SIGILL, we can still roll
113 :     * back to a valid CPU state */
114 :     jmpret = setjmp(mark);
115 :    
116 :     /* If setjmp returned directly, then its returned value is 0, and we still
117 :     * have to test the passed func. Otherwise it means the stack context has
118 :     * been restored by a longjmp() call, which in our case happens only in the
119 :     * signal handler */
120 :     if (jmpret == 0) {
121 : suxen_drol 311 func();
122 :     }
123 :    
124 : edgomez 1412 /* Restore old signal handler */
125 : suxen_drol 311 signal(SIGILL, old_handler);
126 :    
127 : edgomez 1412 return(jmpret);
128 : suxen_drol 311 }
129 :     #endif
130 :    
131 : edgomez 851
132 :     /* detect cpu flags */
133 :     static unsigned int
134 : suxen_drol 1653 detect_cpu_flags(void)
135 : edgomez 851 {
136 :     /* enable native assembly optimizations by default */
137 :     unsigned int cpu_flags = XVID_CPU_ASM;
138 :    
139 : edgomez 1586 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
140 : edgomez 851 cpu_flags |= check_cpu_features();
141 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
142 :     cpu_flags &= ~XVID_CPU_SSE;
143 :    
144 : Isibaar 1794 if ((cpu_flags & (XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41)) && sigill_check(sse2_os_trigger))
145 :     cpu_flags &= ~(XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41);
146 : edgomez 851 #endif
147 :    
148 :     #if defined(ARCH_IS_PPC)
149 : edgomez 1412 if (!sigill_check(altivec_trigger))
150 :     cpu_flags |= XVID_CPU_ALTIVEC;
151 : edgomez 851 #endif
152 :    
153 :     return cpu_flags;
154 :     }
155 :    
156 :    
157 : edgomez 200 /*****************************************************************************
158 :     * XviD Init Entry point
159 :     *
160 :     * Well this function initialize all internal function pointers according
161 :     * to the CPU features forced by the library client or autodetected (depending
162 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
163 :     * image colorspace transformation tables.
164 : edgomez 1382 *
165 : edgomez 200 * Returned value : XVID_ERR_OK
166 :     * + API_VERSION in the input XVID_INIT_PARAM structure
167 :     * + core build " " " " "
168 :     *
169 :     ****************************************************************************/
170 :    
171 : edgomez 851
172 : edgomez 1382 static
173 :     int xvid_gbl_init(xvid_gbl_init_t * init)
174 : Isibaar 3 {
175 : edgomez 1382 unsigned int cpu_flags;
176 : Isibaar 3
177 : edgomez 1382 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
178 :     return XVID_ERR_VERSION;
179 : suxen_drol 234
180 : edgomez 1382 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
181 : suxen_drol 234
182 : edgomez 200 /* Initialize the function pointers */
183 : Isibaar 3 idct_int32_init();
184 : Isibaar 100 init_vlc_tables();
185 :    
186 : edgomez 200 /* Fixed Point Forward/Inverse DCT transformations */
187 : Isibaar 3 fdct = fdct_int32;
188 :     idct = idct_int32;
189 :    
190 : edgomez 200 /* Only needed on PPC Altivec archs */
191 : suxen_drol 1653 sadInit = NULL;
192 : edgomez 195
193 : edgomez 200 /* Restore FPU context : emms_c is a nop functions */
194 : Isibaar 3 emms = emms_c;
195 :    
196 : edgomez 1382 /* Qpel stuff */
197 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
198 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
199 :     xvid_Init_QP();
200 :    
201 : edgomez 200 /* Quantization functions */
202 : edgomez 1382 quant_h263_intra = quant_h263_intra_c;
203 :     quant_h263_inter = quant_h263_inter_c;
204 :     dequant_h263_intra = dequant_h263_intra_c;
205 :     dequant_h263_inter = dequant_h263_inter_c;
206 : Isibaar 3
207 : edgomez 1382 quant_mpeg_intra = quant_mpeg_intra_c;
208 :     quant_mpeg_inter = quant_mpeg_inter_c;
209 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
210 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
211 : Isibaar 3
212 : edgomez 200 /* Block transfer related functions */
213 : Isibaar 3 transfer_8to16copy = transfer_8to16copy_c;
214 :     transfer_16to8copy = transfer_16to8copy_c;
215 : edgomez 200 transfer_8to16sub = transfer_8to16sub_c;
216 : edgomez 851 transfer_8to16subro = transfer_8to16subro_c;
217 : suxen_drol 118 transfer_8to16sub2 = transfer_8to16sub2_c;
218 : syskin 1583 transfer_8to16sub2ro = transfer_8to16sub2ro_c;
219 : edgomez 200 transfer_16to8add = transfer_16to8add_c;
220 :     transfer8x8_copy = transfer8x8_copy_c;
221 : suxen_drol 1632 transfer8x4_copy = transfer8x4_copy_c;
222 : Isibaar 3
223 : edgomez 851 /* Interlacing functions */
224 :     MBFieldTest = MBFieldTest_c;
225 :    
226 : edgomez 200 /* Image interpolation related functions */
227 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
228 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
229 : Isibaar 3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
230 :    
231 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_c;
232 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_c;
233 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_c;
234 :    
235 : edgomez 1530 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c;
236 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c;
237 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c;
238 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c;
239 :    
240 : edgomez 851 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
241 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
242 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
243 :    
244 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
245 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
246 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
247 :    
248 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
249 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
250 :    
251 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
252 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
253 :    
254 : suxen_drol 1397 /* postprocessing */
255 :     image_brightness = image_brightness_c;
256 :    
257 : edgomez 200 /* Initialize internal colorspace transformation tables */
258 : Isibaar 3 colorspace_init();
259 :    
260 : edgomez 200 /* All colorspace transformation functions User Format->YV12 */
261 : edgomez 851 yv12_to_yv12 = yv12_to_yv12_c;
262 :     rgb555_to_yv12 = rgb555_to_yv12_c;
263 :     rgb565_to_yv12 = rgb565_to_yv12_c;
264 : chl 1762 rgb_to_yv12 = rgb_to_yv12_c;
265 : edgomez 851 bgr_to_yv12 = bgr_to_yv12_c;
266 :     bgra_to_yv12 = bgra_to_yv12_c;
267 :     abgr_to_yv12 = abgr_to_yv12_c;
268 :     rgba_to_yv12 = rgba_to_yv12_c;
269 : edgomez 1382 argb_to_yv12 = argb_to_yv12_c;
270 : edgomez 851 yuyv_to_yv12 = yuyv_to_yv12_c;
271 :     uyvy_to_yv12 = uyvy_to_yv12_c;
272 : Isibaar 3
273 : edgomez 851 rgb555i_to_yv12 = rgb555i_to_yv12_c;
274 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
275 :     bgri_to_yv12 = bgri_to_yv12_c;
276 :     bgrai_to_yv12 = bgrai_to_yv12_c;
277 :     abgri_to_yv12 = abgri_to_yv12_c;
278 :     rgbai_to_yv12 = rgbai_to_yv12_c;
279 : edgomez 1382 argbi_to_yv12 = argbi_to_yv12_c;
280 : edgomez 851 yuyvi_to_yv12 = yuyvi_to_yv12_c;
281 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
282 :    
283 : edgomez 200 /* All colorspace transformation functions YV12->User format */
284 : edgomez 851 yv12_to_rgb555 = yv12_to_rgb555_c;
285 :     yv12_to_rgb565 = yv12_to_rgb565_c;
286 : chl 1762 yv12_to_rgb = yv12_to_rgb_c;
287 : edgomez 851 yv12_to_bgr = yv12_to_bgr_c;
288 :     yv12_to_bgra = yv12_to_bgra_c;
289 :     yv12_to_abgr = yv12_to_abgr_c;
290 :     yv12_to_rgba = yv12_to_rgba_c;
291 : edgomez 1382 yv12_to_argb = yv12_to_argb_c;
292 : edgomez 851 yv12_to_yuyv = yv12_to_yuyv_c;
293 :     yv12_to_uyvy = yv12_to_uyvy_c;
294 : edgomez 1382
295 : edgomez 851 yv12_to_rgb555i = yv12_to_rgb555i_c;
296 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
297 :     yv12_to_bgri = yv12_to_bgri_c;
298 :     yv12_to_bgrai = yv12_to_bgrai_c;
299 :     yv12_to_abgri = yv12_to_abgri_c;
300 :     yv12_to_rgbai = yv12_to_rgbai_c;
301 : edgomez 1382 yv12_to_argbi = yv12_to_argbi_c;
302 : edgomez 851 yv12_to_yuyvi = yv12_to_yuyvi_c;
303 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
304 : Isibaar 3
305 : edgomez 200 /* Functions used in motion estimation algorithms */
306 : edgomez 1382 calc_cbp = calc_cbp_c;
307 :     sad16 = sad16_c;
308 :     sad8 = sad8_c;
309 :     sad16bi = sad16bi_c;
310 :     sad8bi = sad8bi_c;
311 :     dev16 = dev16_c;
312 :     sad16v = sad16v_c;
313 :     sse8_16bit = sse8_16bit_c;
314 : edgomez 1424 sse8_8bit = sse8_8bit_c;
315 : Isibaar 3
316 : Skal 1709 init_GMC(cpu_flags);
317 :    
318 : Isibaar 1795 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
319 : edgomez 200
320 : edgomez 851 if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
321 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
322 : Isibaar 1794 (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2) ||
323 :     (cpu_flags & XVID_CPU_SSE3) || (cpu_flags & XVID_CPU_SSE41))
324 : edgomez 851 {
325 :     /* Restore FPU context : emms_c is a nop functions */
326 :     emms = emms_mmx;
327 :     }
328 :    
329 :     if ((cpu_flags & XVID_CPU_MMX)) {
330 :    
331 : edgomez 200 /* Forward and Inverse Discrete Cosine Transformation functions */
332 : edgomez 1382 fdct = fdct_mmx_skal;
333 : Isibaar 1066 idct = idct_mmx;
334 : Isibaar 3
335 : edgomez 1382 /* Qpel stuff */
336 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
337 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
338 :    
339 : edgomez 200 /* Quantization related functions */
340 : edgomez 1382 quant_h263_intra = quant_h263_intra_mmx;
341 : Isibaar 1795 quant_h263_inter = quant_h263_inter_mmx;
342 : edgomez 1382 dequant_h263_intra = dequant_h263_intra_mmx;
343 :     dequant_h263_inter = dequant_h263_inter_mmx;
344 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
345 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
346 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
347 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
348 : Isibaar 3
349 : Isibaar 1795
350 : edgomez 200 /* Block related functions */
351 : Isibaar 3 transfer_8to16copy = transfer_8to16copy_mmx;
352 :     transfer_16to8copy = transfer_16to8copy_mmx;
353 : edgomez 200 transfer_8to16sub = transfer_8to16sub_mmx;
354 : edgomez 851 transfer_8to16subro = transfer_8to16subro_mmx;
355 : edgomez 236 transfer_8to16sub2 = transfer_8to16sub2_mmx;
356 : edgomez 200 transfer_16to8add = transfer_16to8add_mmx;
357 :     transfer8x8_copy = transfer8x8_copy_mmx;
358 : suxen_drol 1632 transfer8x4_copy = transfer8x4_copy_mmx;
359 : Isibaar 3
360 : edgomez 851 /* Interlacing Functions */
361 :     MBFieldTest = MBFieldTest_mmx;
362 : edgomez 236
363 : edgomez 200 /* Image Interpolation related functions */
364 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
365 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
366 : Isibaar 3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
367 :    
368 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_mmx;
369 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_mmx;
370 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_mmx;
371 :    
372 : edgomez 1530 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx;
373 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx;
374 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx;
375 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx;
376 :    
377 : edgomez 851 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
378 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
379 :    
380 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
381 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
382 :    
383 : suxen_drol 1397 /* postprocessing */
384 :     image_brightness = image_brightness_mmx;
385 :    
386 : edgomez 851 /* image input xxx_to_yv12 related functions */
387 : Isibaar 1795
388 : edgomez 851 yv12_to_yv12 = yv12_to_yv12_mmx;
389 : Isibaar 1795
390 : edgomez 851 bgr_to_yv12 = bgr_to_yv12_mmx;
391 : chl 1763 rgb_to_yv12 = rgb_to_yv12_mmx;
392 : edgomez 851 bgra_to_yv12 = bgra_to_yv12_mmx;
393 : chl 1763 rgba_to_yv12 = rgba_to_yv12_mmx;
394 : edgomez 200 yuyv_to_yv12 = yuyv_to_yv12_mmx;
395 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
396 : Isibaar 3
397 : edgomez 851 /* image output yv12_to_xxx related functions */
398 :     yv12_to_bgr = yv12_to_bgr_mmx;
399 :     yv12_to_bgra = yv12_to_bgra_mmx;
400 : edgomez 200 yv12_to_yuyv = yv12_to_yuyv_mmx;
401 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
402 : edgomez 1382
403 : edgomez 851 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
404 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
405 : Isibaar 3
406 : edgomez 200 /* Motion estimation related functions */
407 : edgomez 1424 calc_cbp = calc_cbp_mmx;
408 :     sad16 = sad16_mmx;
409 :     sad8 = sad8_mmx;
410 :     sad16bi = sad16bi_mmx;
411 :     sad8bi = sad8bi_mmx;
412 :     dev16 = dev16_mmx;
413 :     sad16v = sad16v_mmx;
414 : edgomez 1382 sse8_16bit = sse8_16bit_mmx;
415 : edgomez 1424 sse8_8bit = sse8_8bit_mmx;
416 : Isibaar 3 }
417 :    
418 : suxen_drol 329 /* these 3dnow functions are faster than mmx, but slower than xmm. */
419 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOW)) {
420 : suxen_drol 329
421 : edgomez 851 emms = emms_3dn;
422 :    
423 : suxen_drol 329 /* ME functions */
424 :     sad16bi = sad16bi_3dn;
425 :     sad8bi = sad8bi_3dn;
426 : edgomez 851
427 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
428 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
429 : Isibaar 1795
430 : suxen_drol 329 }
431 :    
432 :    
433 : edgomez 851 if ((cpu_flags & XVID_CPU_MMXEXT)) {
434 : edgomez 200
435 : edgomez 1382 /* DCT */
436 :     fdct = fdct_xmm_skal;
437 : Isibaar 1066 idct = idct_xmm;
438 : edgomez 200
439 :     /* Interpolation */
440 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
441 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
442 : h 38 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
443 : edgomez 1530
444 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_xmm;
445 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_xmm;
446 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_xmm;
447 :    
448 : edgomez 1530 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm;
449 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm;
450 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm;
451 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm;
452 : edgomez 200
453 : Isibaar 1795 /* Quantization */
454 : edgomez 1382 quant_mpeg_inter = quant_mpeg_inter_xmm;
455 : edgomez 851
456 : edgomez 1382 dequant_h263_intra = dequant_h263_intra_xmm;
457 :     dequant_h263_inter = dequant_h263_inter_xmm;
458 : chenm001 274
459 : Isibaar 1795 /* Buffer transfer */
460 : edgomez 218 transfer_8to16sub2 = transfer_8to16sub2_xmm;
461 : syskin 1583 transfer_8to16sub2ro = transfer_8to16sub2ro_xmm;
462 : edgomez 218
463 : edgomez 200 /* Colorspace transformation */
464 : Isibaar 1795 /* yv12_to_yv12 = yv12_to_yv12_xmm; */ /* appears to be slow on many machines */
465 : edgomez 851 yuyv_to_yv12 = yuyv_to_yv12_xmm;
466 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
467 : Isibaar 3
468 : edgomez 200 /* ME functions */
469 : Isibaar 3 sad16 = sad16_xmm;
470 : suxen_drol 329 sad8 = sad8_xmm;
471 : chenm001 274 sad16bi = sad16bi_xmm;
472 : suxen_drol 329 sad8bi = sad8bi_xmm;
473 : Isibaar 3 dev16 = dev16_xmm;
474 : edgomez 851 sad16v = sad16v_xmm;
475 : Isibaar 3 }
476 :    
477 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOW)) {
478 : edgomez 200
479 :     /* Interpolation */
480 : Isibaar 3 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
481 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
482 : h 40 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
483 : suxen_drol 1632
484 :     interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dn;
485 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dn;
486 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dn;
487 : Isibaar 3 }
488 :    
489 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
490 : edgomez 200
491 : edgomez 851 /* Buffer transfer */
492 :     transfer_8to16copy = transfer_8to16copy_3dne;
493 :     transfer_16to8copy = transfer_16to8copy_3dne;
494 :     transfer_8to16sub = transfer_8to16sub_3dne;
495 :     transfer_8to16subro = transfer_8to16subro_3dne;
496 :     transfer_16to8add = transfer_16to8add_3dne;
497 :     transfer8x8_copy = transfer8x8_copy_3dne;
498 : suxen_drol 1632 transfer8x4_copy = transfer8x4_copy_3dne;
499 : edgomez 851
500 : edgomez 1423 if ((cpu_flags & XVID_CPU_MMXEXT)) {
501 :     /* Inverse DCT */
502 :     idct = idct_3dne;
503 : edgomez 851
504 : edgomez 1423 /* Buffer transfer */
505 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
506 : edgomez 851
507 : edgomez 1423 /* Interpolation */
508 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
509 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
510 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
511 :    
512 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dne;
513 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dne;
514 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dne;
515 :    
516 : Isibaar 1795 /* Quantization */
517 : edgomez 1423 quant_h263_intra = quant_h263_intra_3dne; /* cmov only */
518 :     quant_h263_inter = quant_h263_inter_3dne;
519 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */
520 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
521 :     dequant_h263_intra = dequant_h263_intra_3dne;
522 :     dequant_h263_inter = dequant_h263_inter_3dne;
523 :    
524 : Isibaar 1795 /* ME functions */
525 : edgomez 1423 sad16 = sad16_3dne;
526 :     sad8 = sad8_3dne;
527 :     sad16bi = sad16bi_3dne;
528 :     sad8bi = sad8bi_3dne;
529 :     dev16 = dev16_3dne;
530 :     }
531 : edgomez 851 }
532 : Isibaar 1795
533 : edgomez 851 if ((cpu_flags & XVID_CPU_SSE2)) {
534 :    
535 : chenm001 274 calc_cbp = calc_cbp_sse2;
536 :    
537 : edgomez 200 /* Quantization */
538 : edgomez 1382 quant_h263_intra = quant_h263_intra_sse2;
539 :     quant_h263_inter = quant_h263_inter_sse2;
540 :     dequant_h263_intra = dequant_h263_intra_sse2;
541 :     dequant_h263_inter = dequant_h263_inter_sse2;
542 : h 135
543 : edgomez 1382 /* SAD operators */
544 : edgomez 200 sad16 = sad16_sse2;
545 :     dev16 = dev16_sse2;
546 : edgomez 1382
547 : Skal 1618 /* DCT operators */
548 : edgomez 1382 fdct = fdct_sse2_skal;
549 : Isibaar 1805 idct = idct_sse2_skal; /* Is now IEEE1180 and Walken compliant. */
550 : suxen_drol 1493
551 :     /* postprocessing */
552 :     image_brightness = image_brightness_sse2;
553 : Isibaar 1795
554 : h 126 }
555 : Isibaar 1764
556 :     if ((cpu_flags & XVID_CPU_SSE3)) {
557 :    
558 :     /* SAD operators */
559 :     sad16 = sad16_sse3;
560 :     dev16 = dev16_sse3;
561 :     }
562 : Isibaar 1794
563 : edgomez 1382 #endif /* ARCH_IS_IA32 */
564 : edgomez 200
565 : edgomez 851 #if defined(ARCH_IS_IA64)
566 : edgomez 874 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
567 : Isibaar 209 idct_ia64_init();
568 :     fdct = fdct_ia64;
569 : edgomez 874 idct = idct_ia64; /*not yet working, crashes */
570 : Isibaar 209 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
571 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
572 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
573 :     sad16 = sad16_ia64;
574 :     sad16bi = sad16bi_ia64;
575 :     sad8 = sad8_ia64;
576 :     dev16 = dev16_ia64;
577 : edgomez 874 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
578 : edgomez 1382 quant_h263_intra = quant_h263_intra_ia64;
579 :     quant_h263_inter = quant_h263_inter_ia64;
580 :     dequant_h263_intra = dequant_h263_intra_ia64;
581 :     dequant_h263_inter = dequant_h263_inter_ia64;
582 : Isibaar 209 transfer_8to16copy = transfer_8to16copy_ia64;
583 :     transfer_16to8copy = transfer_16to8copy_ia64;
584 :     transfer_8to16sub = transfer_8to16sub_ia64;
585 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
586 :     transfer_16to8add = transfer_16to8add_ia64;
587 :     transfer8x8_copy = transfer8x8_copy_ia64;
588 :     }
589 : edgomez 1382 #endif
590 : Isibaar 209
591 : edgomez 851 #if defined(ARCH_IS_PPC)
592 : edgomez 1412 if ((cpu_flags & XVID_CPU_ALTIVEC)) {
593 :     /* sad operators */
594 : edgomez 1557 sad16 = sad16_altivec_c;
595 :     sad16bi = sad16bi_altivec_c;
596 :     sad8 = sad8_altivec_c;
597 :     dev16 = dev16_altivec_c;
598 : edgomez 1412
599 :     sse8_16bit = sse8_16bit_altivec_c;
600 :    
601 :     /* mem transfer */
602 :     transfer_8to16copy = transfer_8to16copy_altivec_c;
603 :     transfer_16to8copy = transfer_16to8copy_altivec_c;
604 :     transfer_8to16sub = transfer_8to16sub_altivec_c;
605 :     transfer_8to16subro = transfer_8to16subro_altivec_c;
606 :     transfer_8to16sub2 = transfer_8to16sub2_altivec_c;
607 :     transfer_16to8add = transfer_16to8add_altivec_c;
608 :     transfer8x8_copy = transfer8x8_copy_altivec_c;
609 :    
610 :     /* Inverse DCT */
611 :     idct = idct_altivec_c;
612 :    
613 :     /* Interpolation */
614 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c;
615 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c;
616 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c;
617 : edgomez 1570
618 : edgomez 1412 interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c;
619 :     interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c;
620 : edgomez 1557
621 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c;
622 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c;
623 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c;
624 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c;
625 : edgomez 1412
626 :     /* Colorspace conversion */
627 :     bgra_to_yv12 = bgra_to_yv12_altivec_c;
628 :     abgr_to_yv12 = abgr_to_yv12_altivec_c;
629 :     rgba_to_yv12 = rgba_to_yv12_altivec_c;
630 :     argb_to_yv12 = argb_to_yv12_altivec_c;
631 :    
632 :     yuyv_to_yv12 = yuyv_to_yv12_altivec_c;
633 :     uyvy_to_yv12 = uyvy_to_yv12_altivec_c;
634 :    
635 :     yv12_to_yuyv = yv12_to_yuyv_altivec_c;
636 :     yv12_to_uyvy = yv12_to_uyvy_altivec_c;
637 :    
638 :     /* Quantization */
639 :     quant_h263_intra = quant_h263_intra_altivec_c;
640 :     quant_h263_inter = quant_h263_inter_altivec_c;
641 :     dequant_h263_intra = dequant_h263_intra_altivec_c;
642 :     dequant_h263_inter = dequant_h263_inter_altivec_c;
643 : edgomez 1609
644 :     dequant_mpeg_intra = dequant_mpeg_intra_altivec_c;
645 :     dequant_mpeg_inter = dequant_mpeg_inter_altivec_c;
646 : edgomez 1557
647 :     /* Qpel stuff */
648 :     xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C;
649 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C;
650 : edgomez 1412 }
651 : canard 52 #endif
652 : edgomez 851
653 : edgomez 1382 #if defined(_DEBUG)
654 :     xvid_debug = init->debug;
655 :     #endif
656 :    
657 : edgomez 1412 return(0);
658 : edgomez 851 }
659 :    
660 :    
661 :     static int
662 : edgomez 1382 xvid_gbl_info(xvid_gbl_info_t * info)
663 : edgomez 851 {
664 : edgomez 1382 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
665 :     return XVID_ERR_VERSION;
666 : edgomez 851
667 : edgomez 1382 info->actual_version = XVID_VERSION;
668 : Isibaar 1821 info->build = "xvid-1.3.0-dev";
669 : edgomez 1382 info->cpu_flags = detect_cpu_flags();
670 : Isibaar 1827 info->num_threads = 0;
671 : edgomez 851
672 : Isibaar 1795 #if defined(_WIN32)
673 : Isibaar 1827
674 : suxen_drol 1688 {
675 : Isibaar 1827 SYSTEM_INFO siSysInfo;
676 :     GetSystemInfo(&siSysInfo);
677 :     info->num_threads = siSysInfo.dwNumberOfProcessors; /* number of _logical_ cores */
678 : suxen_drol 1688 }
679 : Isibaar 1827
680 : Isibaar 1795 #else
681 :    
682 :     info->num_threads = sysconf(_SC_NPROCESSORS_CONF);
683 :    
684 : edgomez 1382 #endif
685 : edgomez 851
686 :     return 0;
687 :     }
688 :    
689 :    
690 : edgomez 1382 static int
691 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
692 : edgomez 851 {
693 : edgomez 1382 int width;
694 :     int height;
695 :     int width2;
696 :     int height2;
697 :     IMAGE img;
698 : edgomez 851
699 : edgomez 1382 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
700 :     return XVID_ERR_VERSION;
701 : edgomez 851
702 : edgomez 1382 #if 0
703 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
704 :     #endif
705 :     width = convert->width;
706 :     height = convert->height;
707 :     width2 = convert->width/2;
708 :     height2 = convert->height/2;
709 : edgomez 851
710 : edgomez 1382 switch (convert->input.csp & ~XVID_CSP_VFLIP)
711 : edgomez 851 {
712 : edgomez 1382 case XVID_CSP_YV12 :
713 :     img.y = convert->input.plane[0];
714 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
715 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
716 :     image_output(&img, width, height, width,
717 :     (uint8_t**)convert->output.plane, convert->output.stride,
718 :     convert->output.csp, convert->interlacing);
719 :     break;
720 : edgomez 851
721 : edgomez 1382 default :
722 :     return XVID_ERR_FORMAT;
723 : edgomez 851 }
724 :    
725 :    
726 : edgomez 1382 emms();
727 : edgomez 851 return 0;
728 :     }
729 :    
730 : edgomez 1382 /*****************************************************************************
731 :     * XviD Global Entry point
732 :     *
733 :     * Well this function initialize all internal function pointers according
734 :     * to the CPU features forced by the library client or autodetected (depending
735 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
736 :     * image colorspace transformation tables.
737 :     *
738 :     ****************************************************************************/
739 : edgomez 851
740 :    
741 :     int
742 : edgomez 1382 xvid_global(void *handle,
743 : edgomez 851 int opt,
744 :     void *param1,
745 :     void *param2)
746 :     {
747 :     switch(opt)
748 :     {
749 : edgomez 1382 case XVID_GBL_INIT :
750 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
751 : edgomez 851
752 : edgomez 1382 case XVID_GBL_INFO :
753 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
754 : edgomez 851
755 : edgomez 1382 case XVID_GBL_CONVERT :
756 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
757 :    
758 : edgomez 851 default :
759 :     return XVID_ERR_FAIL;
760 :     }
761 :     }
762 :    
763 : edgomez 200 /*****************************************************************************
764 :     * XviD Native decoder entry point
765 :     *
766 :     * This function is just a wrapper to all the option cases.
767 :     *
768 :     * Returned values : XVID_ERR_FAIL when opt is invalid
769 :     * else returns the wrapped function result
770 :     *
771 :     ****************************************************************************/
772 :    
773 : edgomez 195 int
774 :     xvid_decore(void *handle,
775 :     int opt,
776 :     void *param1,
777 :     void *param2)
778 : Isibaar 3 {
779 : edgomez 195 switch (opt) {
780 :     case XVID_DEC_CREATE:
781 : edgomez 1382 return decoder_create((xvid_dec_create_t *) param1);
782 : Isibaar 3
783 : edgomez 195 case XVID_DEC_DESTROY:
784 :     return decoder_destroy((DECODER *) handle);
785 :    
786 : edgomez 1382 case XVID_DEC_DECODE:
787 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
788 :    
789 : Isibaar 3 default:
790 : edgomez 195 return XVID_ERR_FAIL;
791 :     }
792 : Isibaar 3 }
793 :    
794 :    
795 : edgomez 200 /*****************************************************************************
796 :     * XviD Native encoder entry point
797 :     *
798 :     * This function is just a wrapper to all the option cases.
799 :     *
800 :     * Returned values : XVID_ERR_FAIL when opt is invalid
801 :     * else returns the wrapped function result
802 :     *
803 :     ****************************************************************************/
804 :    
805 : edgomez 195 int
806 :     xvid_encore(void *handle,
807 :     int opt,
808 :     void *param1,
809 :     void *param2)
810 : Isibaar 3 {
811 : edgomez 195 switch (opt) {
812 :     case XVID_ENC_ENCODE:
813 : edgomez 851
814 : edgomez 1382 return enc_encode((Encoder *) handle,
815 :     (xvid_enc_frame_t *) param1,
816 :     (xvid_enc_stats_t *) param2);
817 : Isibaar 3
818 : edgomez 195 case XVID_ENC_CREATE:
819 : edgomez 1382 return enc_create((xvid_enc_create_t *) param1);
820 : Isibaar 3
821 : edgomez 195 case XVID_ENC_DESTROY:
822 : edgomez 1382 return enc_destroy((Encoder *) handle);
823 : edgomez 195
824 : Isibaar 3 default:
825 : edgomez 195 return XVID_ERR_FAIL;
826 :     }
827 : Isibaar 3 }

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