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

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.72 - (view) (download)

1 : edgomez 1.16 /*****************************************************************************
2 : edgomez 1.17 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : suxen_drol 1.49 * Copyright(C) 2001-2004 Peter Ross <pross@xvid.org>
7 : edgomez 1.48 *
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 : chl 1.72 * $Id: xvid.c,v 1.71 2006/11/11 22:06:44 chl Exp $
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 :     #include "utils/mem_transfer.h"
40 : edgomez 1.41 #include "utils/mbfunctions.h"
41 : edgomez 1.48 #include "quant/quant.h"
42 : ia64p 1.30 #include "motion/motion.h"
43 : Skal 1.69 #include "motion/gmc.h"
44 : Isibaar 1.1 #include "motion/sad.h"
45 :     #include "utils/emms.h"
46 :     #include "utils/timer.h"
47 : Isibaar 1.9 #include "bitstream/mbcoding.h"
48 : edgomez 1.48 #include "image/qpel.h"
49 :     #include "image/postprocessing.h"
50 :    
51 :     #if defined(_DEBUG)
52 :     unsigned int xvid_debug = 0; /* xvid debug mask */
53 :     #endif
54 : Isibaar 1.1
55 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
56 : edgomez 1.41 # include <windows.h>
57 : edgomez 1.60 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
58 : edgomez 1.41 # include <signal.h>
59 :     # include <setjmp.h>
60 : suxen_drol 1.31
61 : edgomez 1.41 static jmp_buf mark;
62 : suxen_drol 1.31
63 : edgomez 1.41 static void
64 :     sigill_handler(int signal)
65 :     {
66 :     longjmp(mark, 1);
67 :     }
68 : suxen_drol 1.31 #endif
69 :    
70 :    
71 :     /*
72 : edgomez 1.43 * 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 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
81 : edgomez 1.50 static int
82 : suxen_drol 1.31 sigill_check(void (*func)())
83 :     {
84 :     _try {
85 :     func();
86 : edgomez 1.50 } _except(EXCEPTION_EXECUTE_HANDLER) {
87 : suxen_drol 1.31
88 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
89 : edgomez 1.50 return(1);
90 : suxen_drol 1.31 }
91 : edgomez 1.50 return(0);
92 :     }
93 : edgomez 1.60 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
94 : edgomez 1.50 static int
95 :     sigill_check(void (*func)())
96 :     {
97 :     void *old_handler;
98 : suxen_drol 1.31 int jmpret;
99 :    
100 : edgomez 1.50 /* Set our SIGILL handler */
101 :     old_handler = signal(SIGILL, sigill_handler);
102 : suxen_drol 1.31
103 : edgomez 1.50 /* Check for error */
104 :     if (old_handler == SIG_ERR) {
105 :     return(-1);
106 : suxen_drol 1.31 }
107 :    
108 : edgomez 1.50 /* Save stack context, so if func triggers a SIGILL, we can still roll
109 :     * back to a valid CPU state */
110 :     jmpret = setjmp(mark);
111 :    
112 :     /* If setjmp returned directly, then its returned value is 0, and we still
113 :     * have to test the passed func. Otherwise it means the stack context has
114 :     * been restored by a longjmp() call, which in our case happens only in the
115 :     * signal handler */
116 :     if (jmpret == 0) {
117 : suxen_drol 1.31 func();
118 :     }
119 :    
120 : edgomez 1.50 /* Restore old signal handler */
121 : suxen_drol 1.31 signal(SIGILL, old_handler);
122 :    
123 : edgomez 1.50 return(jmpret);
124 : suxen_drol 1.31 }
125 :     #endif
126 :    
127 : edgomez 1.41
128 :     /* detect cpu flags */
129 :     static unsigned int
130 : suxen_drol 1.65 detect_cpu_flags(void)
131 : edgomez 1.41 {
132 :     /* enable native assembly optimizations by default */
133 :     unsigned int cpu_flags = XVID_CPU_ASM;
134 :    
135 : edgomez 1.60 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
136 : edgomez 1.41 cpu_flags |= check_cpu_features();
137 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
138 :     cpu_flags &= ~XVID_CPU_SSE;
139 :    
140 :     if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger))
141 :     cpu_flags &= ~XVID_CPU_SSE2;
142 :     #endif
143 :    
144 :     #if defined(ARCH_IS_PPC)
145 : edgomez 1.50 if (!sigill_check(altivec_trigger))
146 :     cpu_flags |= XVID_CPU_ALTIVEC;
147 : edgomez 1.41 #endif
148 :    
149 :     return cpu_flags;
150 :     }
151 :    
152 :    
153 : edgomez 1.16 /*****************************************************************************
154 :     * XviD Init Entry point
155 :     *
156 :     * Well this function initialize all internal function pointers according
157 :     * to the CPU features forced by the library client or autodetected (depending
158 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
159 :     * image colorspace transformation tables.
160 : edgomez 1.48 *
161 : edgomez 1.16 * Returned value : XVID_ERR_OK
162 :     * + API_VERSION in the input XVID_INIT_PARAM structure
163 :     * + core build " " " " "
164 :     *
165 :     ****************************************************************************/
166 :    
167 : edgomez 1.41
168 : edgomez 1.48 static
169 :     int xvid_gbl_init(xvid_gbl_init_t * init)
170 : Isibaar 1.1 {
171 : edgomez 1.48 unsigned int cpu_flags;
172 : suxen_drol 1.21
173 : edgomez 1.48 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
174 :     return XVID_ERR_VERSION;
175 : suxen_drol 1.31
176 : edgomez 1.48 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
177 : Isibaar 1.1
178 : edgomez 1.16 /* Initialize the function pointers */
179 : Isibaar 1.1 idct_int32_init();
180 : Isibaar 1.9 init_vlc_tables();
181 :    
182 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
183 : Isibaar 1.1 fdct = fdct_int32;
184 :     idct = idct_int32;
185 :    
186 : edgomez 1.16 /* Only needed on PPC Altivec archs */
187 : suxen_drol 1.65 sadInit = NULL;
188 : edgomez 1.15
189 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
190 : Isibaar 1.1 emms = emms_c;
191 :    
192 : edgomez 1.48 /* Qpel stuff */
193 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
194 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
195 :     xvid_Init_QP();
196 :    
197 : edgomez 1.16 /* Quantization functions */
198 : edgomez 1.48 quant_h263_intra = quant_h263_intra_c;
199 :     quant_h263_inter = quant_h263_inter_c;
200 :     dequant_h263_intra = dequant_h263_intra_c;
201 :     dequant_h263_inter = dequant_h263_inter_c;
202 :    
203 :     quant_mpeg_intra = quant_mpeg_intra_c;
204 :     quant_mpeg_inter = quant_mpeg_inter_c;
205 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
206 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
207 : Isibaar 1.1
208 : edgomez 1.16 /* Block transfer related functions */
209 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
210 :     transfer_16to8copy = transfer_16to8copy_c;
211 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
212 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_c;
213 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
214 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_c;
215 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
216 :     transfer8x8_copy = transfer8x8_copy_c;
217 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_c;
218 : Isibaar 1.1
219 : edgomez 1.41 /* Interlacing functions */
220 :     MBFieldTest = MBFieldTest_c;
221 :    
222 : edgomez 1.16 /* Image interpolation related functions */
223 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
224 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
225 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
226 :    
227 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_c;
228 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_c;
229 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_c;
230 :    
231 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c;
232 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c;
233 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c;
234 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c;
235 :    
236 : edgomez 1.41 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
237 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
238 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
239 :    
240 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
241 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
242 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
243 :    
244 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
245 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
246 :    
247 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
248 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
249 :    
250 : suxen_drol 1.49 /* postprocessing */
251 :     image_brightness = image_brightness_c;
252 :    
253 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
254 : Isibaar 1.1 colorspace_init();
255 :    
256 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
257 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_c;
258 :     rgb555_to_yv12 = rgb555_to_yv12_c;
259 :     rgb565_to_yv12 = rgb565_to_yv12_c;
260 : chl 1.71 rgb_to_yv12 = rgb_to_yv12_c;
261 : edgomez 1.41 bgr_to_yv12 = bgr_to_yv12_c;
262 :     bgra_to_yv12 = bgra_to_yv12_c;
263 :     abgr_to_yv12 = abgr_to_yv12_c;
264 :     rgba_to_yv12 = rgba_to_yv12_c;
265 : edgomez 1.48 argb_to_yv12 = argb_to_yv12_c;
266 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_c;
267 :     uyvy_to_yv12 = uyvy_to_yv12_c;
268 :    
269 :     rgb555i_to_yv12 = rgb555i_to_yv12_c;
270 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
271 :     bgri_to_yv12 = bgri_to_yv12_c;
272 :     bgrai_to_yv12 = bgrai_to_yv12_c;
273 :     abgri_to_yv12 = abgri_to_yv12_c;
274 :     rgbai_to_yv12 = rgbai_to_yv12_c;
275 : edgomez 1.48 argbi_to_yv12 = argbi_to_yv12_c;
276 : edgomez 1.41 yuyvi_to_yv12 = yuyvi_to_yv12_c;
277 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
278 :    
279 : edgomez 1.16 /* All colorspace transformation functions YV12->User format */
280 : edgomez 1.41 yv12_to_rgb555 = yv12_to_rgb555_c;
281 :     yv12_to_rgb565 = yv12_to_rgb565_c;
282 : chl 1.71 yv12_to_rgb = yv12_to_rgb_c;
283 : edgomez 1.41 yv12_to_bgr = yv12_to_bgr_c;
284 :     yv12_to_bgra = yv12_to_bgra_c;
285 :     yv12_to_abgr = yv12_to_abgr_c;
286 :     yv12_to_rgba = yv12_to_rgba_c;
287 : edgomez 1.48 yv12_to_argb = yv12_to_argb_c;
288 : edgomez 1.41 yv12_to_yuyv = yv12_to_yuyv_c;
289 :     yv12_to_uyvy = yv12_to_uyvy_c;
290 : edgomez 1.48
291 : edgomez 1.41 yv12_to_rgb555i = yv12_to_rgb555i_c;
292 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
293 :     yv12_to_bgri = yv12_to_bgri_c;
294 :     yv12_to_bgrai = yv12_to_bgrai_c;
295 :     yv12_to_abgri = yv12_to_abgri_c;
296 :     yv12_to_rgbai = yv12_to_rgbai_c;
297 : edgomez 1.48 yv12_to_argbi = yv12_to_argbi_c;
298 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_c;
299 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
300 : Isibaar 1.1
301 : edgomez 1.16 /* Functions used in motion estimation algorithms */
302 : edgomez 1.48 calc_cbp = calc_cbp_c;
303 :     sad16 = sad16_c;
304 :     sad8 = sad8_c;
305 :     sad16bi = sad16bi_c;
306 :     sad8bi = sad8bi_c;
307 :     dev16 = dev16_c;
308 :     sad16v = sad16v_c;
309 :     sse8_16bit = sse8_16bit_c;
310 : edgomez 1.52 sse8_8bit = sse8_8bit_c;
311 : Isibaar 1.1
312 : Skal 1.69 init_GMC(cpu_flags);
313 :    
314 : edgomez 1.41 #if defined(ARCH_IS_IA32)
315 :    
316 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
317 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
318 :     (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2))
319 :     {
320 :     /* Restore FPU context : emms_c is a nop functions */
321 :     emms = emms_mmx;
322 :     }
323 :    
324 :     if ((cpu_flags & XVID_CPU_MMX)) {
325 : edgomez 1.16
326 :     /* Forward and Inverse Discrete Cosine Transformation functions */
327 : edgomez 1.48 fdct = fdct_mmx_skal;
328 : Isibaar 1.47 idct = idct_mmx;
329 : Isibaar 1.1
330 : edgomez 1.48 /* Qpel stuff */
331 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
332 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
333 :    
334 : edgomez 1.16 /* Quantization related functions */
335 : edgomez 1.48 quant_h263_intra = quant_h263_intra_mmx;
336 :     quant_h263_inter = quant_h263_inter_mmx;
337 :     dequant_h263_intra = dequant_h263_intra_mmx;
338 :     dequant_h263_inter = dequant_h263_inter_mmx;
339 :    
340 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
341 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
342 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
343 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
344 : Isibaar 1.1
345 : edgomez 1.16 /* Block related functions */
346 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
347 :     transfer_16to8copy = transfer_16to8copy_mmx;
348 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
349 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_mmx;
350 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
351 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
352 :     transfer8x8_copy = transfer8x8_copy_mmx;
353 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_mmx;
354 : edgomez 1.22
355 : edgomez 1.41 /* Interlacing Functions */
356 :     MBFieldTest = MBFieldTest_mmx;
357 : edgomez 1.16
358 :     /* Image Interpolation related functions */
359 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
360 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
361 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
362 :    
363 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_mmx;
364 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_mmx;
365 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_mmx;
366 :    
367 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx;
368 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx;
369 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx;
370 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx;
371 :    
372 : edgomez 1.41 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
373 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
374 :    
375 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
376 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
377 : suxen_drol 1.49
378 :     /* postprocessing */
379 :     image_brightness = image_brightness_mmx;
380 : edgomez 1.41
381 :     /* image input xxx_to_yv12 related functions */
382 :     yv12_to_yv12 = yv12_to_yv12_mmx;
383 :     bgr_to_yv12 = bgr_to_yv12_mmx;
384 : chl 1.72 rgb_to_yv12 = rgb_to_yv12_mmx;
385 : edgomez 1.41 bgra_to_yv12 = bgra_to_yv12_mmx;
386 : chl 1.72 rgba_to_yv12 = rgba_to_yv12_mmx;
387 : edgomez 1.16 yuyv_to_yv12 = yuyv_to_yv12_mmx;
388 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
389 : Isibaar 1.1
390 : edgomez 1.41 /* image output yv12_to_xxx related functions */
391 :     yv12_to_bgr = yv12_to_bgr_mmx;
392 :     yv12_to_bgra = yv12_to_bgra_mmx;
393 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
394 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
395 : edgomez 1.48
396 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
397 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
398 : Isibaar 1.1
399 : edgomez 1.16 /* Motion estimation related functions */
400 : edgomez 1.52 calc_cbp = calc_cbp_mmx;
401 :     sad16 = sad16_mmx;
402 :     sad8 = sad8_mmx;
403 :     sad16bi = sad16bi_mmx;
404 :     sad8bi = sad8bi_mmx;
405 :     dev16 = dev16_mmx;
406 :     sad16v = sad16v_mmx;
407 : edgomez 1.48 sse8_16bit = sse8_16bit_mmx;
408 : edgomez 1.52 sse8_8bit = sse8_8bit_mmx;
409 : Isibaar 1.1 }
410 :    
411 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
412 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
413 :    
414 :     emms = emms_3dn;
415 : suxen_drol 1.33
416 :     /* ME functions */
417 :     sad16bi = sad16bi_3dn;
418 :     sad8bi = sad8bi_3dn;
419 : edgomez 1.41
420 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
421 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
422 : suxen_drol 1.33 }
423 :    
424 :    
425 : edgomez 1.41 if ((cpu_flags & XVID_CPU_MMXEXT)) {
426 : edgomez 1.16
427 : edgomez 1.48 /* DCT */
428 :     fdct = fdct_xmm_skal;
429 : Isibaar 1.47 idct = idct_xmm;
430 : edgomez 1.16
431 :     /* Interpolation */
432 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
433 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
434 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
435 : edgomez 1.55
436 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_xmm;
437 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_xmm;
438 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_xmm;
439 :    
440 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm;
441 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm;
442 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm;
443 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm;
444 : Isibaar 1.25
445 : chenm001 1.29 /* Quantization */
446 : edgomez 1.48 quant_mpeg_inter = quant_mpeg_inter_xmm;
447 : edgomez 1.41
448 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_xmm;
449 :     dequant_h263_inter = dequant_h263_inter_xmm;
450 : chenm001 1.29
451 : edgomez 1.19 /* Buffer transfer */
452 :     transfer_8to16sub2 = transfer_8to16sub2_xmm;
453 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_xmm;
454 : edgomez 1.16
455 :     /* Colorspace transformation */
456 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_xmm;
457 :     yuyv_to_yv12 = yuyv_to_yv12_xmm;
458 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
459 : Isibaar 1.1
460 : edgomez 1.16 /* ME functions */
461 : Isibaar 1.1 sad16 = sad16_xmm;
462 : suxen_drol 1.33 sad8 = sad8_xmm;
463 : chenm001 1.29 sad16bi = sad16bi_xmm;
464 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
465 : Isibaar 1.1 dev16 = dev16_xmm;
466 : edgomez 1.41 sad16v = sad16v_xmm;
467 : Isibaar 1.1 }
468 :    
469 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
470 : edgomez 1.16
471 :     /* Interpolation */
472 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
473 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
474 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
475 : suxen_drol 1.64
476 :     interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dn;
477 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dn;
478 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dn;
479 : Isibaar 1.1 }
480 :    
481 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
482 :    
483 :     /* Buffer transfer */
484 :     transfer_8to16copy = transfer_8to16copy_3dne;
485 :     transfer_16to8copy = transfer_16to8copy_3dne;
486 :     transfer_8to16sub = transfer_8to16sub_3dne;
487 :     transfer_8to16subro = transfer_8to16subro_3dne;
488 :     transfer_16to8add = transfer_16to8add_3dne;
489 :     transfer8x8_copy = transfer8x8_copy_3dne;
490 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_3dne;
491 : edgomez 1.41
492 : edgomez 1.51 if ((cpu_flags & XVID_CPU_MMXEXT)) {
493 :     /* Inverse DCT */
494 :     idct = idct_3dne;
495 :    
496 :     /* Buffer transfer */
497 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
498 :    
499 :     /* Interpolation */
500 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
501 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
502 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
503 :    
504 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dne;
505 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dne;
506 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dne;
507 :    
508 : edgomez 1.51 /* Quantization */
509 :     quant_h263_intra = quant_h263_intra_3dne; /* cmov only */
510 :     quant_h263_inter = quant_h263_inter_3dne;
511 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */
512 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
513 :     dequant_h263_intra = dequant_h263_intra_3dne;
514 :     dequant_h263_inter = dequant_h263_inter_3dne;
515 :    
516 :     /* ME functions */
517 :     calc_cbp = calc_cbp_3dne;
518 :    
519 :     sad16 = sad16_3dne;
520 :     sad8 = sad8_3dne;
521 :     sad16bi = sad16bi_3dne;
522 :     sad8bi = sad8bi_3dne;
523 :     dev16 = dev16_3dne;
524 :     }
525 : edgomez 1.41 }
526 :    
527 :     if ((cpu_flags & XVID_CPU_SSE2)) {
528 : edgomez 1.16
529 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
530 :    
531 : edgomez 1.16 /* Quantization */
532 : edgomez 1.48 quant_h263_intra = quant_h263_intra_sse2;
533 :     quant_h263_inter = quant_h263_inter_sse2;
534 :     dequant_h263_intra = dequant_h263_intra_sse2;
535 :     dequant_h263_inter = dequant_h263_inter_sse2;
536 : h 1.13
537 : edgomez 1.48 /* SAD operators */
538 : edgomez 1.16 sad16 = sad16_sse2;
539 :     dev16 = dev16_sse2;
540 : edgomez 1.48
541 : Skal 1.63 /* DCT operators */
542 : edgomez 1.48 fdct = fdct_sse2_skal;
543 : suxen_drol 1.64 /* idct = idct_sse2_skal; */ /* Is now IEEE1180 and Walken compliant. Disabled until fully tested. */
544 : suxen_drol 1.54
545 :     /* postprocessing */
546 :     image_brightness = image_brightness_sse2;
547 : h 1.12 }
548 : edgomez 1.48 #endif /* ARCH_IS_IA32 */
549 : Isibaar 1.18
550 : edgomez 1.41 #if defined(ARCH_IS_IA64)
551 : edgomez 1.43 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
552 : Isibaar 1.18 idct_ia64_init();
553 :     fdct = fdct_ia64;
554 : edgomez 1.43 idct = idct_ia64; /*not yet working, crashes */
555 : Isibaar 1.18 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
556 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
557 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
558 :     sad16 = sad16_ia64;
559 :     sad16bi = sad16bi_ia64;
560 :     sad8 = sad8_ia64;
561 :     dev16 = dev16_ia64;
562 : edgomez 1.43 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
563 : edgomez 1.48 quant_h263_intra = quant_h263_intra_ia64;
564 :     quant_h263_inter = quant_h263_inter_ia64;
565 :     dequant_h263_intra = dequant_h263_intra_ia64;
566 :     dequant_h263_inter = dequant_h263_inter_ia64;
567 : Isibaar 1.18 transfer_8to16copy = transfer_8to16copy_ia64;
568 :     transfer_16to8copy = transfer_16to8copy_ia64;
569 :     transfer_8to16sub = transfer_8to16sub_ia64;
570 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
571 :     transfer_16to8add = transfer_16to8add_ia64;
572 :     transfer8x8_copy = transfer8x8_copy_ia64;
573 :     }
574 : edgomez 1.48 #endif
575 : edgomez 1.16
576 : edgomez 1.41 #if defined(ARCH_IS_PPC)
577 : edgomez 1.50 if ((cpu_flags & XVID_CPU_ALTIVEC)) {
578 :     /* sad operators */
579 : edgomez 1.56 sad16 = sad16_altivec_c;
580 :     sad16bi = sad16bi_altivec_c;
581 :     sad8 = sad8_altivec_c;
582 :     dev16 = dev16_altivec_c;
583 : edgomez 1.50
584 :     sse8_16bit = sse8_16bit_altivec_c;
585 :    
586 :     /* mem transfer */
587 :     transfer_8to16copy = transfer_8to16copy_altivec_c;
588 :     transfer_16to8copy = transfer_16to8copy_altivec_c;
589 :     transfer_8to16sub = transfer_8to16sub_altivec_c;
590 :     transfer_8to16subro = transfer_8to16subro_altivec_c;
591 :     transfer_8to16sub2 = transfer_8to16sub2_altivec_c;
592 :     transfer_16to8add = transfer_16to8add_altivec_c;
593 :     transfer8x8_copy = transfer8x8_copy_altivec_c;
594 :    
595 :     /* Inverse DCT */
596 :     idct = idct_altivec_c;
597 :    
598 :     /* Interpolation */
599 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c;
600 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c;
601 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c;
602 : edgomez 1.58
603 : edgomez 1.50 interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c;
604 :     interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c;
605 : edgomez 1.56
606 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c;
607 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c;
608 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c;
609 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c;
610 : edgomez 1.50
611 :     /* Colorspace conversion */
612 :     bgra_to_yv12 = bgra_to_yv12_altivec_c;
613 :     abgr_to_yv12 = abgr_to_yv12_altivec_c;
614 :     rgba_to_yv12 = rgba_to_yv12_altivec_c;
615 :     argb_to_yv12 = argb_to_yv12_altivec_c;
616 :    
617 :     yuyv_to_yv12 = yuyv_to_yv12_altivec_c;
618 :     uyvy_to_yv12 = uyvy_to_yv12_altivec_c;
619 :    
620 :     yv12_to_yuyv = yv12_to_yuyv_altivec_c;
621 :     yv12_to_uyvy = yv12_to_uyvy_altivec_c;
622 :    
623 :     /* Quantization */
624 :     quant_h263_intra = quant_h263_intra_altivec_c;
625 :     quant_h263_inter = quant_h263_inter_altivec_c;
626 :     dequant_h263_intra = dequant_h263_intra_altivec_c;
627 :     dequant_h263_inter = dequant_h263_inter_altivec_c;
628 : edgomez 1.62
629 :     dequant_mpeg_intra = dequant_mpeg_intra_altivec_c;
630 :     dequant_mpeg_inter = dequant_mpeg_inter_altivec_c;
631 : edgomez 1.56
632 :     /* Qpel stuff */
633 :     xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C;
634 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C;
635 : edgomez 1.50 }
636 : canard 1.6 #endif
637 : edgomez 1.41
638 : edgomez 1.60 #if defined(ARCH_IS_X86_64)
639 :     /* For now, only XVID_CPU_ASM is looked for, so user can still
640 :     * disable asm usage the usual way. When Intel EMT64 cpus will
641 :     * be out, maybe we'll have to check more precisely what cpu
642 :     * features there really are. */
643 :     if (cpu_flags & XVID_CPU_ASM) {
644 :     /* SIMD state flusher */
645 : Isibaar 1.67 emms = emms_mmx;
646 : edgomez 1.60
647 :     /* DCT operators */
648 :     fdct = fdct_skal_x86_64;
649 :     idct = idct_x86_64;
650 :    
651 :     /* SAD operators */
652 :     sad16 = sad16_x86_64;
653 :     sad8 = sad8_x86_64;
654 :     sad16bi = sad16bi_x86_64;
655 :     sad8bi = sad8bi_x86_64;
656 :     dev16 = dev16_x86_64;
657 :     sad16v = sad16v_x86_64;
658 :     sse8_16bit = sse8_16bit_x86_64;
659 :     sse8_8bit = sse8_8bit_x86_64;
660 :    
661 :     /* Interpolation operators */
662 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_x86_64;
663 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_x86_64;
664 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_x86_64;
665 :    
666 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_x86_64;
667 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_x86_64;
668 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_x86_64;
669 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_x86_64;
670 :    
671 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_x86_64;
672 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_x86_64;
673 :    
674 :     interpolate8x8_avg2 = interpolate8x8_avg2_x86_64;
675 :     interpolate8x8_avg4 = interpolate8x8_avg4_x86_64;
676 :    
677 :     /* Quantization related functions */
678 :     quant_h263_intra = quant_h263_intra_x86_64;
679 :     quant_h263_inter = quant_h263_inter_x86_64;
680 :     dequant_h263_intra = dequant_h263_intra_x86_64;
681 :     dequant_h263_inter = dequant_h263_inter_x86_64;
682 : syskin 1.70 /*quant_mpeg_intra = quant_mpeg_intra_x86_64; fix me! */
683 : edgomez 1.60 quant_mpeg_inter = quant_mpeg_inter_x86_64;
684 :     dequant_mpeg_intra = dequant_mpeg_intra_x86_64;
685 :     dequant_mpeg_inter = dequant_mpeg_inter_x86_64;
686 :    
687 :     /* Block related functions */
688 :     transfer_8to16copy = transfer_8to16copy_x86_64;
689 :     transfer_16to8copy = transfer_16to8copy_x86_64;
690 :     transfer_8to16sub = transfer_8to16sub_x86_64;
691 :     transfer_8to16subro = transfer_8to16subro_x86_64;
692 :     transfer_8to16sub2 = transfer_8to16sub2_x86_64;
693 :     transfer_8to16sub2ro= transfer_8to16sub2ro_x86_64;
694 :     transfer_16to8add = transfer_16to8add_x86_64;
695 :     transfer8x8_copy = transfer8x8_copy_x86_64;
696 :    
697 :     /* Qpel stuff */
698 :     xvid_QP_Funcs = &xvid_QP_Funcs_x86_64;
699 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_x86_64;
700 :    
701 :     /* Interlacing Functions */
702 :     MBFieldTest = MBFieldTest_x86_64;
703 :     }
704 :     #endif
705 :    
706 : edgomez 1.48 #if defined(_DEBUG)
707 :     xvid_debug = init->debug;
708 :     #endif
709 :    
710 : edgomez 1.50 return(0);
711 : edgomez 1.41 }
712 :    
713 :    
714 :     static int
715 : edgomez 1.48 xvid_gbl_info(xvid_gbl_info_t * info)
716 : edgomez 1.41 {
717 : edgomez 1.48 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
718 :     return XVID_ERR_VERSION;
719 : edgomez 1.41
720 : edgomez 1.48 info->actual_version = XVID_VERSION;
721 : suxen_drol 1.66 info->build = "xvid-1.2.0-dev";
722 : edgomez 1.48 info->cpu_flags = detect_cpu_flags();
723 : suxen_drol 1.68 info->num_threads = 0;
724 : edgomez 1.41
725 : suxen_drol 1.68 #if defined(WIN32)
726 :     {
727 :     DWORD dwProcessAffinityMask, dwSystemAffinityMask;
728 :     if (GetProcessAffinityMask(GetCurrentProcess(), &dwProcessAffinityMask, &dwSystemAffinityMask)) {
729 :     int i;
730 :     for(i=0; i<32; i++) {
731 :     if ((dwProcessAffinityMask & (1<<i)))
732 :     info->num_threads++;
733 :     }
734 :     }
735 :     }
736 : edgomez 1.48 #endif
737 : edgomez 1.41
738 :     return 0;
739 :     }
740 :    
741 :    
742 : edgomez 1.48 static int
743 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
744 : edgomez 1.41 {
745 : edgomez 1.48 int width;
746 :     int height;
747 :     int width2;
748 :     int height2;
749 :     IMAGE img;
750 : edgomez 1.41
751 : edgomez 1.48 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
752 :     return XVID_ERR_VERSION;
753 : edgomez 1.41
754 : edgomez 1.48 #if 0
755 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
756 :     #endif
757 :     width = convert->width;
758 :     height = convert->height;
759 :     width2 = convert->width/2;
760 :     height2 = convert->height/2;
761 : edgomez 1.41
762 : edgomez 1.48 switch (convert->input.csp & ~XVID_CSP_VFLIP)
763 : edgomez 1.41 {
764 : edgomez 1.48 case XVID_CSP_YV12 :
765 :     img.y = convert->input.plane[0];
766 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
767 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
768 :     image_output(&img, width, height, width,
769 :     (uint8_t**)convert->output.plane, convert->output.stride,
770 :     convert->output.csp, convert->interlacing);
771 :     break;
772 : edgomez 1.41
773 : edgomez 1.48 default :
774 :     return XVID_ERR_FORMAT;
775 : edgomez 1.41 }
776 :    
777 :    
778 : edgomez 1.48 emms();
779 : edgomez 1.41 return 0;
780 :     }
781 :    
782 : edgomez 1.48 /*****************************************************************************
783 :     * XviD Global Entry point
784 :     *
785 :     * Well this function initialize all internal function pointers according
786 :     * to the CPU features forced by the library client or autodetected (depending
787 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
788 :     * image colorspace transformation tables.
789 :     *
790 :     ****************************************************************************/
791 : Isibaar 1.1
792 : edgomez 1.41
793 :     int
794 : edgomez 1.48 xvid_global(void *handle,
795 : edgomez 1.41 int opt,
796 :     void *param1,
797 :     void *param2)
798 :     {
799 :     switch(opt)
800 :     {
801 : edgomez 1.48 case XVID_GBL_INIT :
802 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
803 : edgomez 1.41
804 : edgomez 1.48 case XVID_GBL_INFO :
805 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
806 :    
807 :     case XVID_GBL_CONVERT :
808 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
809 : edgomez 1.41
810 :     default :
811 :     return XVID_ERR_FAIL;
812 :     }
813 :     }
814 :    
815 : edgomez 1.16 /*****************************************************************************
816 :     * XviD Native decoder entry point
817 :     *
818 :     * This function is just a wrapper to all the option cases.
819 :     *
820 :     * Returned values : XVID_ERR_FAIL when opt is invalid
821 :     * else returns the wrapped function result
822 :     *
823 :     ****************************************************************************/
824 :    
825 : edgomez 1.15 int
826 :     xvid_decore(void *handle,
827 :     int opt,
828 :     void *param1,
829 :     void *param2)
830 : Isibaar 1.1 {
831 : edgomez 1.15 switch (opt) {
832 :     case XVID_DEC_CREATE:
833 : edgomez 1.48 return decoder_create((xvid_dec_create_t *) param1);
834 : edgomez 1.15
835 :     case XVID_DEC_DESTROY:
836 :     return decoder_destroy((DECODER *) handle);
837 : Isibaar 1.1
838 : edgomez 1.48 case XVID_DEC_DECODE:
839 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
840 :    
841 : Isibaar 1.1 default:
842 : edgomez 1.15 return XVID_ERR_FAIL;
843 :     }
844 : Isibaar 1.1 }
845 :    
846 : edgomez 1.16
847 :     /*****************************************************************************
848 :     * XviD Native encoder entry point
849 :     *
850 :     * This function is just a wrapper to all the option cases.
851 :     *
852 :     * Returned values : XVID_ERR_FAIL when opt is invalid
853 :     * else returns the wrapped function result
854 :     *
855 :     ****************************************************************************/
856 : Isibaar 1.1
857 : edgomez 1.15 int
858 :     xvid_encore(void *handle,
859 :     int opt,
860 :     void *param1,
861 :     void *param2)
862 : Isibaar 1.1 {
863 : edgomez 1.15 switch (opt) {
864 :     case XVID_ENC_ENCODE:
865 : edgomez 1.41
866 : edgomez 1.48 return enc_encode((Encoder *) handle,
867 :     (xvid_enc_frame_t *) param1,
868 :     (xvid_enc_stats_t *) param2);
869 : edgomez 1.15
870 :     case XVID_ENC_CREATE:
871 : edgomez 1.48 return enc_create((xvid_enc_create_t *) param1);
872 : edgomez 1.15
873 :     case XVID_ENC_DESTROY:
874 : edgomez 1.48 return enc_destroy((Encoder *) handle);
875 : Isibaar 1.1
876 :     default:
877 : edgomez 1.15 return XVID_ERR_FAIL;
878 :     }
879 : Isibaar 1.1 }

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