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

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.63 - (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 : Skal 1.63 * $Id: xvid.c,v 1.62 2005/04/04 23:49:37 edgomez 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 : Isibaar 1.1 #include "motion/sad.h"
44 :     #include "utils/emms.h"
45 :     #include "utils/timer.h"
46 : Isibaar 1.9 #include "bitstream/mbcoding.h"
47 : edgomez 1.48 #include "image/qpel.h"
48 :     #include "image/postprocessing.h"
49 :    
50 :     #if defined(_DEBUG)
51 :     unsigned int xvid_debug = 0; /* xvid debug mask */
52 :     #endif
53 : Isibaar 1.1
54 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
55 : edgomez 1.41 # include <windows.h>
56 : edgomez 1.60 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
57 : edgomez 1.41 # include <signal.h>
58 :     # include <setjmp.h>
59 : suxen_drol 1.31
60 : edgomez 1.41 static jmp_buf mark;
61 : suxen_drol 1.31
62 : edgomez 1.41 static void
63 :     sigill_handler(int signal)
64 :     {
65 :     longjmp(mark, 1);
66 :     }
67 : suxen_drol 1.31 #endif
68 :    
69 :    
70 :     /*
71 : edgomez 1.43 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
72 :     * signalled
73 :     *
74 :     * Return values:
75 :     * -1 : could not determine
76 :     * 0 : SIGILL was *not* signalled
77 :     * 1 : SIGILL was signalled
78 :     */
79 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
80 : edgomez 1.50 static int
81 : suxen_drol 1.31 sigill_check(void (*func)())
82 :     {
83 :     _try {
84 :     func();
85 : edgomez 1.50 } _except(EXCEPTION_EXECUTE_HANDLER) {
86 : suxen_drol 1.31
87 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
88 : edgomez 1.50 return(1);
89 : suxen_drol 1.31 }
90 : edgomez 1.50 return(0);
91 :     }
92 : edgomez 1.60 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
93 : edgomez 1.50 static int
94 :     sigill_check(void (*func)())
95 :     {
96 :     void *old_handler;
97 : suxen_drol 1.31 int jmpret;
98 :    
99 : edgomez 1.50 /* Set our SIGILL handler */
100 :     old_handler = signal(SIGILL, sigill_handler);
101 : suxen_drol 1.31
102 : edgomez 1.50 /* Check for error */
103 :     if (old_handler == SIG_ERR) {
104 :     return(-1);
105 : suxen_drol 1.31 }
106 :    
107 : edgomez 1.50 /* Save stack context, so if func triggers a SIGILL, we can still roll
108 :     * back to a valid CPU state */
109 :     jmpret = setjmp(mark);
110 :    
111 :     /* If setjmp returned directly, then its returned value is 0, and we still
112 :     * have to test the passed func. Otherwise it means the stack context has
113 :     * been restored by a longjmp() call, which in our case happens only in the
114 :     * signal handler */
115 :     if (jmpret == 0) {
116 : suxen_drol 1.31 func();
117 :     }
118 :    
119 : edgomez 1.50 /* Restore old signal handler */
120 : suxen_drol 1.31 signal(SIGILL, old_handler);
121 :    
122 : edgomez 1.50 return(jmpret);
123 : suxen_drol 1.31 }
124 :     #endif
125 :    
126 : edgomez 1.41
127 :     /* detect cpu flags */
128 :     static unsigned int
129 :     detect_cpu_flags()
130 :     {
131 :     /* enable native assembly optimizations by default */
132 :     unsigned int cpu_flags = XVID_CPU_ASM;
133 :    
134 : edgomez 1.60 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
135 : edgomez 1.41 cpu_flags |= check_cpu_features();
136 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
137 :     cpu_flags &= ~XVID_CPU_SSE;
138 :    
139 :     if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger))
140 :     cpu_flags &= ~XVID_CPU_SSE2;
141 :     #endif
142 :    
143 :     #if defined(ARCH_IS_PPC)
144 : edgomez 1.50 if (!sigill_check(altivec_trigger))
145 :     cpu_flags |= XVID_CPU_ALTIVEC;
146 : edgomez 1.41 #endif
147 :    
148 :     return cpu_flags;
149 :     }
150 :    
151 :    
152 : edgomez 1.16 /*****************************************************************************
153 :     * XviD Init Entry point
154 :     *
155 :     * Well this function initialize all internal function pointers according
156 :     * to the CPU features forced by the library client or autodetected (depending
157 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
158 :     * image colorspace transformation tables.
159 : edgomez 1.48 *
160 : edgomez 1.16 * Returned value : XVID_ERR_OK
161 :     * + API_VERSION in the input XVID_INIT_PARAM structure
162 :     * + core build " " " " "
163 :     *
164 :     ****************************************************************************/
165 :    
166 : edgomez 1.41
167 : edgomez 1.48 static
168 :     int xvid_gbl_init(xvid_gbl_init_t * init)
169 : Isibaar 1.1 {
170 : edgomez 1.48 unsigned int cpu_flags;
171 : suxen_drol 1.21
172 : edgomez 1.48 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
173 :     return XVID_ERR_VERSION;
174 : suxen_drol 1.31
175 : edgomez 1.48 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
176 : Isibaar 1.1
177 : edgomez 1.16 /* Initialize the function pointers */
178 : Isibaar 1.1 idct_int32_init();
179 : Isibaar 1.9 init_vlc_tables();
180 :    
181 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
182 : Isibaar 1.1 fdct = fdct_int32;
183 :     idct = idct_int32;
184 :    
185 : edgomez 1.16 /* Only needed on PPC Altivec archs */
186 : canard 1.10 sadInit = 0;
187 : edgomez 1.15
188 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
189 : Isibaar 1.1 emms = emms_c;
190 :    
191 : edgomez 1.48 /* Qpel stuff */
192 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
193 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
194 :     xvid_Init_QP();
195 :    
196 : edgomez 1.16 /* Quantization functions */
197 : edgomez 1.48 quant_h263_intra = quant_h263_intra_c;
198 :     quant_h263_inter = quant_h263_inter_c;
199 :     dequant_h263_intra = dequant_h263_intra_c;
200 :     dequant_h263_inter = dequant_h263_inter_c;
201 :    
202 :     quant_mpeg_intra = quant_mpeg_intra_c;
203 :     quant_mpeg_inter = quant_mpeg_inter_c;
204 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
205 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
206 : Isibaar 1.1
207 : edgomez 1.16 /* Block transfer related functions */
208 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
209 :     transfer_16to8copy = transfer_16to8copy_c;
210 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
211 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_c;
212 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
213 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_c;
214 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
215 :     transfer8x8_copy = transfer8x8_copy_c;
216 : Isibaar 1.1
217 : edgomez 1.41 /* Interlacing functions */
218 :     MBFieldTest = MBFieldTest_c;
219 :    
220 : edgomez 1.16 /* Image interpolation related functions */
221 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
222 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
223 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
224 :    
225 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c;
226 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c;
227 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c;
228 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c;
229 :    
230 : edgomez 1.41 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
231 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
232 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
233 :    
234 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
235 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
236 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
237 :    
238 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
239 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
240 :    
241 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
242 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
243 :    
244 : suxen_drol 1.49 /* postprocessing */
245 :     image_brightness = image_brightness_c;
246 :    
247 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
248 : Isibaar 1.1 colorspace_init();
249 :    
250 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
251 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_c;
252 :     rgb555_to_yv12 = rgb555_to_yv12_c;
253 :     rgb565_to_yv12 = rgb565_to_yv12_c;
254 :     bgr_to_yv12 = bgr_to_yv12_c;
255 :     bgra_to_yv12 = bgra_to_yv12_c;
256 :     abgr_to_yv12 = abgr_to_yv12_c;
257 :     rgba_to_yv12 = rgba_to_yv12_c;
258 : edgomez 1.48 argb_to_yv12 = argb_to_yv12_c;
259 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_c;
260 :     uyvy_to_yv12 = uyvy_to_yv12_c;
261 :    
262 :     rgb555i_to_yv12 = rgb555i_to_yv12_c;
263 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
264 :     bgri_to_yv12 = bgri_to_yv12_c;
265 :     bgrai_to_yv12 = bgrai_to_yv12_c;
266 :     abgri_to_yv12 = abgri_to_yv12_c;
267 :     rgbai_to_yv12 = rgbai_to_yv12_c;
268 : edgomez 1.48 argbi_to_yv12 = argbi_to_yv12_c;
269 : edgomez 1.41 yuyvi_to_yv12 = yuyvi_to_yv12_c;
270 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
271 :    
272 : edgomez 1.16 /* All colorspace transformation functions YV12->User format */
273 : edgomez 1.41 yv12_to_rgb555 = yv12_to_rgb555_c;
274 :     yv12_to_rgb565 = yv12_to_rgb565_c;
275 :     yv12_to_bgr = yv12_to_bgr_c;
276 :     yv12_to_bgra = yv12_to_bgra_c;
277 :     yv12_to_abgr = yv12_to_abgr_c;
278 :     yv12_to_rgba = yv12_to_rgba_c;
279 : edgomez 1.48 yv12_to_argb = yv12_to_argb_c;
280 : edgomez 1.41 yv12_to_yuyv = yv12_to_yuyv_c;
281 :     yv12_to_uyvy = yv12_to_uyvy_c;
282 : edgomez 1.48
283 : edgomez 1.41 yv12_to_rgb555i = yv12_to_rgb555i_c;
284 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
285 :     yv12_to_bgri = yv12_to_bgri_c;
286 :     yv12_to_bgrai = yv12_to_bgrai_c;
287 :     yv12_to_abgri = yv12_to_abgri_c;
288 :     yv12_to_rgbai = yv12_to_rgbai_c;
289 : edgomez 1.48 yv12_to_argbi = yv12_to_argbi_c;
290 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_c;
291 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
292 : Isibaar 1.1
293 : edgomez 1.16 /* Functions used in motion estimation algorithms */
294 : edgomez 1.48 calc_cbp = calc_cbp_c;
295 :     sad16 = sad16_c;
296 :     sad8 = sad8_c;
297 :     sad16bi = sad16bi_c;
298 :     sad8bi = sad8bi_c;
299 :     dev16 = dev16_c;
300 :     sad16v = sad16v_c;
301 :     sse8_16bit = sse8_16bit_c;
302 : edgomez 1.52 sse8_8bit = sse8_8bit_c;
303 : Isibaar 1.1
304 : edgomez 1.41 #if defined(ARCH_IS_IA32)
305 :    
306 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
307 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
308 :     (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2))
309 :     {
310 :     /* Restore FPU context : emms_c is a nop functions */
311 :     emms = emms_mmx;
312 :     }
313 :    
314 :     if ((cpu_flags & XVID_CPU_MMX)) {
315 : edgomez 1.16
316 :     /* Forward and Inverse Discrete Cosine Transformation functions */
317 : edgomez 1.48 fdct = fdct_mmx_skal;
318 : Isibaar 1.47 idct = idct_mmx;
319 : Isibaar 1.1
320 : edgomez 1.48 /* Qpel stuff */
321 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
322 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
323 :    
324 : edgomez 1.16 /* Quantization related functions */
325 : edgomez 1.48 quant_h263_intra = quant_h263_intra_mmx;
326 :     quant_h263_inter = quant_h263_inter_mmx;
327 :     dequant_h263_intra = dequant_h263_intra_mmx;
328 :     dequant_h263_inter = dequant_h263_inter_mmx;
329 :    
330 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
331 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
332 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
333 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
334 : Isibaar 1.1
335 : edgomez 1.16 /* Block related functions */
336 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
337 :     transfer_16to8copy = transfer_16to8copy_mmx;
338 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
339 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_mmx;
340 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
341 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
342 :     transfer8x8_copy = transfer8x8_copy_mmx;
343 : edgomez 1.22
344 : edgomez 1.41 /* Interlacing Functions */
345 :     MBFieldTest = MBFieldTest_mmx;
346 : edgomez 1.16
347 :     /* Image Interpolation related functions */
348 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
349 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
350 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
351 :    
352 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx;
353 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx;
354 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx;
355 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx;
356 :    
357 : edgomez 1.41 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
358 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
359 :    
360 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
361 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
362 : suxen_drol 1.49
363 :     /* postprocessing */
364 :     image_brightness = image_brightness_mmx;
365 : edgomez 1.41
366 :     /* image input xxx_to_yv12 related functions */
367 :     yv12_to_yv12 = yv12_to_yv12_mmx;
368 :     bgr_to_yv12 = bgr_to_yv12_mmx;
369 :     bgra_to_yv12 = bgra_to_yv12_mmx;
370 : edgomez 1.16 yuyv_to_yv12 = yuyv_to_yv12_mmx;
371 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
372 : Isibaar 1.1
373 : edgomez 1.41 /* image output yv12_to_xxx related functions */
374 :     yv12_to_bgr = yv12_to_bgr_mmx;
375 :     yv12_to_bgra = yv12_to_bgra_mmx;
376 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
377 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
378 : edgomez 1.48
379 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
380 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
381 : Isibaar 1.1
382 : edgomez 1.16 /* Motion estimation related functions */
383 : edgomez 1.52 calc_cbp = calc_cbp_mmx;
384 :     sad16 = sad16_mmx;
385 :     sad8 = sad8_mmx;
386 :     sad16bi = sad16bi_mmx;
387 :     sad8bi = sad8bi_mmx;
388 :     dev16 = dev16_mmx;
389 :     sad16v = sad16v_mmx;
390 : edgomez 1.48 sse8_16bit = sse8_16bit_mmx;
391 : edgomez 1.52 sse8_8bit = sse8_8bit_mmx;
392 : Isibaar 1.1 }
393 :    
394 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
395 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
396 :    
397 :     emms = emms_3dn;
398 : suxen_drol 1.33
399 :     /* ME functions */
400 :     sad16bi = sad16bi_3dn;
401 :     sad8bi = sad8bi_3dn;
402 : edgomez 1.41
403 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
404 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
405 : suxen_drol 1.33 }
406 :    
407 :    
408 : edgomez 1.41 if ((cpu_flags & XVID_CPU_MMXEXT)) {
409 : edgomez 1.16
410 : edgomez 1.48 /* DCT */
411 :     fdct = fdct_xmm_skal;
412 : Isibaar 1.47 idct = idct_xmm;
413 : edgomez 1.16
414 :     /* Interpolation */
415 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
416 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
417 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
418 : edgomez 1.55
419 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm;
420 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm;
421 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm;
422 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm;
423 : Isibaar 1.25
424 : chenm001 1.29 /* Quantization */
425 : edgomez 1.48 quant_mpeg_intra = quant_mpeg_intra_xmm;
426 :     quant_mpeg_inter = quant_mpeg_inter_xmm;
427 : edgomez 1.41
428 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_xmm;
429 :     dequant_h263_inter = dequant_h263_inter_xmm;
430 : chenm001 1.29
431 : edgomez 1.19 /* Buffer transfer */
432 :     transfer_8to16sub2 = transfer_8to16sub2_xmm;
433 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_xmm;
434 : edgomez 1.16
435 :     /* Colorspace transformation */
436 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_xmm;
437 :     yuyv_to_yv12 = yuyv_to_yv12_xmm;
438 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
439 : Isibaar 1.1
440 : edgomez 1.16 /* ME functions */
441 : Isibaar 1.1 sad16 = sad16_xmm;
442 : suxen_drol 1.33 sad8 = sad8_xmm;
443 : chenm001 1.29 sad16bi = sad16bi_xmm;
444 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
445 : Isibaar 1.1 dev16 = dev16_xmm;
446 : edgomez 1.41 sad16v = sad16v_xmm;
447 : Isibaar 1.1 }
448 :    
449 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
450 : edgomez 1.16
451 :     /* Interpolation */
452 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
453 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
454 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
455 : Isibaar 1.1 }
456 :    
457 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
458 :    
459 :     /* Buffer transfer */
460 :     transfer_8to16copy = transfer_8to16copy_3dne;
461 :     transfer_16to8copy = transfer_16to8copy_3dne;
462 :     transfer_8to16sub = transfer_8to16sub_3dne;
463 :     transfer_8to16subro = transfer_8to16subro_3dne;
464 :     transfer_16to8add = transfer_16to8add_3dne;
465 :     transfer8x8_copy = transfer8x8_copy_3dne;
466 :    
467 : edgomez 1.51 if ((cpu_flags & XVID_CPU_MMXEXT)) {
468 :     /* Inverse DCT */
469 :     idct = idct_3dne;
470 :    
471 :     /* Buffer transfer */
472 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
473 :    
474 :     /* Interpolation */
475 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
476 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
477 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
478 :    
479 :     /* Quantization */
480 :     quant_h263_intra = quant_h263_intra_3dne; /* cmov only */
481 :     quant_h263_inter = quant_h263_inter_3dne;
482 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */
483 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
484 :     dequant_h263_intra = dequant_h263_intra_3dne;
485 :     dequant_h263_inter = dequant_h263_inter_3dne;
486 :    
487 :     /* ME functions */
488 :     calc_cbp = calc_cbp_3dne;
489 :    
490 :     sad16 = sad16_3dne;
491 :     sad8 = sad8_3dne;
492 :     sad16bi = sad16bi_3dne;
493 :     sad8bi = sad8bi_3dne;
494 :     dev16 = dev16_3dne;
495 :     }
496 : edgomez 1.41 }
497 :    
498 :     if ((cpu_flags & XVID_CPU_SSE2)) {
499 : edgomez 1.16
500 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
501 :    
502 : edgomez 1.16 /* Quantization */
503 : edgomez 1.48 quant_h263_intra = quant_h263_intra_sse2;
504 :     quant_h263_inter = quant_h263_inter_sse2;
505 :     dequant_h263_intra = dequant_h263_intra_sse2;
506 :     dequant_h263_inter = dequant_h263_inter_sse2;
507 : h 1.13
508 : edgomez 1.48 /* SAD operators */
509 : edgomez 1.16 sad16 = sad16_sse2;
510 :     dev16 = dev16_sse2;
511 : edgomez 1.48
512 : Skal 1.63 /* DCT operators */
513 : edgomez 1.48 fdct = fdct_sse2_skal;
514 : Skal 1.63 /* idct = idct_sse2_skal; */ /* Is now IEEE1180 and Walken compliant. Disabled until fully tested. */
515 : suxen_drol 1.54
516 :     /* postprocessing */
517 :     image_brightness = image_brightness_sse2;
518 : h 1.12 }
519 : edgomez 1.48 #endif /* ARCH_IS_IA32 */
520 : Isibaar 1.18
521 : edgomez 1.41 #if defined(ARCH_IS_IA64)
522 : edgomez 1.43 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
523 : Isibaar 1.18 idct_ia64_init();
524 :     fdct = fdct_ia64;
525 : edgomez 1.43 idct = idct_ia64; /*not yet working, crashes */
526 : Isibaar 1.18 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
527 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
528 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
529 :     sad16 = sad16_ia64;
530 :     sad16bi = sad16bi_ia64;
531 :     sad8 = sad8_ia64;
532 :     dev16 = dev16_ia64;
533 : edgomez 1.43 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
534 : edgomez 1.48 quant_h263_intra = quant_h263_intra_ia64;
535 :     quant_h263_inter = quant_h263_inter_ia64;
536 :     dequant_h263_intra = dequant_h263_intra_ia64;
537 :     dequant_h263_inter = dequant_h263_inter_ia64;
538 : Isibaar 1.18 transfer_8to16copy = transfer_8to16copy_ia64;
539 :     transfer_16to8copy = transfer_16to8copy_ia64;
540 :     transfer_8to16sub = transfer_8to16sub_ia64;
541 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
542 :     transfer_16to8add = transfer_16to8add_ia64;
543 :     transfer8x8_copy = transfer8x8_copy_ia64;
544 :     }
545 : edgomez 1.48 #endif
546 : edgomez 1.16
547 : edgomez 1.41 #if defined(ARCH_IS_PPC)
548 : edgomez 1.50 if ((cpu_flags & XVID_CPU_ALTIVEC)) {
549 :     /* sad operators */
550 : edgomez 1.56 sad16 = sad16_altivec_c;
551 :     sad16bi = sad16bi_altivec_c;
552 :     sad8 = sad8_altivec_c;
553 :     dev16 = dev16_altivec_c;
554 : edgomez 1.50
555 :     sse8_16bit = sse8_16bit_altivec_c;
556 :    
557 :     /* mem transfer */
558 :     transfer_8to16copy = transfer_8to16copy_altivec_c;
559 :     transfer_16to8copy = transfer_16to8copy_altivec_c;
560 :     transfer_8to16sub = transfer_8to16sub_altivec_c;
561 :     transfer_8to16subro = transfer_8to16subro_altivec_c;
562 :     transfer_8to16sub2 = transfer_8to16sub2_altivec_c;
563 :     transfer_16to8add = transfer_16to8add_altivec_c;
564 :     transfer8x8_copy = transfer8x8_copy_altivec_c;
565 :    
566 :     /* Inverse DCT */
567 :     idct = idct_altivec_c;
568 :    
569 :     /* Interpolation */
570 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c;
571 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c;
572 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c;
573 : edgomez 1.58
574 : edgomez 1.50 interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c;
575 :     interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c;
576 : edgomez 1.56
577 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c;
578 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c;
579 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c;
580 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c;
581 : edgomez 1.50
582 :     /* Colorspace conversion */
583 :     bgra_to_yv12 = bgra_to_yv12_altivec_c;
584 :     abgr_to_yv12 = abgr_to_yv12_altivec_c;
585 :     rgba_to_yv12 = rgba_to_yv12_altivec_c;
586 :     argb_to_yv12 = argb_to_yv12_altivec_c;
587 :    
588 :     yuyv_to_yv12 = yuyv_to_yv12_altivec_c;
589 :     uyvy_to_yv12 = uyvy_to_yv12_altivec_c;
590 :    
591 :     yv12_to_yuyv = yv12_to_yuyv_altivec_c;
592 :     yv12_to_uyvy = yv12_to_uyvy_altivec_c;
593 :    
594 :     /* Quantization */
595 :     quant_h263_intra = quant_h263_intra_altivec_c;
596 :     quant_h263_inter = quant_h263_inter_altivec_c;
597 :     dequant_h263_intra = dequant_h263_intra_altivec_c;
598 :     dequant_h263_inter = dequant_h263_inter_altivec_c;
599 : edgomez 1.62
600 :     dequant_mpeg_intra = dequant_mpeg_intra_altivec_c;
601 :     dequant_mpeg_inter = dequant_mpeg_inter_altivec_c;
602 : edgomez 1.56
603 :     /* Qpel stuff */
604 :     xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C;
605 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C;
606 : edgomez 1.50 }
607 : canard 1.6 #endif
608 : edgomez 1.41
609 : edgomez 1.60 #if defined(ARCH_IS_X86_64)
610 :     /* For now, only XVID_CPU_ASM is looked for, so user can still
611 :     * disable asm usage the usual way. When Intel EMT64 cpus will
612 :     * be out, maybe we'll have to check more precisely what cpu
613 :     * features there really are. */
614 :     if (cpu_flags & XVID_CPU_ASM) {
615 :     /* SIMD state flusher */
616 :     emms = emms_3dn;
617 :    
618 :     /* DCT operators */
619 :     fdct = fdct_skal_x86_64;
620 :     idct = idct_x86_64;
621 :    
622 :     /* SAD operators */
623 :     sad16 = sad16_x86_64;
624 :     sad8 = sad8_x86_64;
625 :     sad16bi = sad16bi_x86_64;
626 :     sad8bi = sad8bi_x86_64;
627 :     dev16 = dev16_x86_64;
628 :     sad16v = sad16v_x86_64;
629 :     sse8_16bit = sse8_16bit_x86_64;
630 :     sse8_8bit = sse8_8bit_x86_64;
631 :    
632 :     /* Interpolation operators */
633 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_x86_64;
634 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_x86_64;
635 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_x86_64;
636 :    
637 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_x86_64;
638 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_x86_64;
639 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_x86_64;
640 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_x86_64;
641 :    
642 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_x86_64;
643 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_x86_64;
644 :    
645 :     interpolate8x8_avg2 = interpolate8x8_avg2_x86_64;
646 :     interpolate8x8_avg4 = interpolate8x8_avg4_x86_64;
647 :    
648 :     /* Quantization related functions */
649 :     quant_h263_intra = quant_h263_intra_x86_64;
650 :     quant_h263_inter = quant_h263_inter_x86_64;
651 :     dequant_h263_intra = dequant_h263_intra_x86_64;
652 :     dequant_h263_inter = dequant_h263_inter_x86_64;
653 :     quant_mpeg_intra = quant_mpeg_intra_x86_64;
654 :     quant_mpeg_inter = quant_mpeg_inter_x86_64;
655 :     dequant_mpeg_intra = dequant_mpeg_intra_x86_64;
656 :     dequant_mpeg_inter = dequant_mpeg_inter_x86_64;
657 :    
658 :     /* Block related functions */
659 :     transfer_8to16copy = transfer_8to16copy_x86_64;
660 :     transfer_16to8copy = transfer_16to8copy_x86_64;
661 :     transfer_8to16sub = transfer_8to16sub_x86_64;
662 :     transfer_8to16subro = transfer_8to16subro_x86_64;
663 :     transfer_8to16sub2 = transfer_8to16sub2_x86_64;
664 :     transfer_8to16sub2ro= transfer_8to16sub2ro_x86_64;
665 :     transfer_16to8add = transfer_16to8add_x86_64;
666 :     transfer8x8_copy = transfer8x8_copy_x86_64;
667 :    
668 :     /* Qpel stuff */
669 :     xvid_QP_Funcs = &xvid_QP_Funcs_x86_64;
670 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_x86_64;
671 :    
672 :     /* Interlacing Functions */
673 :     MBFieldTest = MBFieldTest_x86_64;
674 :     }
675 :     #endif
676 :    
677 : edgomez 1.48 #if defined(_DEBUG)
678 :     xvid_debug = init->debug;
679 :     #endif
680 :    
681 : edgomez 1.50 return(0);
682 : edgomez 1.41 }
683 :    
684 :    
685 :     static int
686 : edgomez 1.48 xvid_gbl_info(xvid_gbl_info_t * info)
687 : edgomez 1.41 {
688 : edgomez 1.48 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
689 :     return XVID_ERR_VERSION;
690 : edgomez 1.41
691 : edgomez 1.48 info->actual_version = XVID_VERSION;
692 : edgomez 1.62 info->build = "xvid-1.1.0-beta2";
693 : edgomez 1.48 info->cpu_flags = detect_cpu_flags();
694 : edgomez 1.41
695 : edgomez 1.48 #if defined(_SMP) && defined(WIN32)
696 :     info->num_threads = pthread_num_processors_np();;
697 :     #else
698 :     info->num_threads = 0;
699 :     #endif
700 : edgomez 1.41
701 :     return 0;
702 :     }
703 :    
704 :    
705 : edgomez 1.48 static int
706 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
707 : edgomez 1.41 {
708 : edgomez 1.48 int width;
709 :     int height;
710 :     int width2;
711 :     int height2;
712 :     IMAGE img;
713 : edgomez 1.41
714 : edgomez 1.48 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
715 :     return XVID_ERR_VERSION;
716 : edgomez 1.41
717 : edgomez 1.48 #if 0
718 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
719 :     #endif
720 :     width = convert->width;
721 :     height = convert->height;
722 :     width2 = convert->width/2;
723 :     height2 = convert->height/2;
724 : edgomez 1.41
725 : edgomez 1.48 switch (convert->input.csp & ~XVID_CSP_VFLIP)
726 : edgomez 1.41 {
727 : edgomez 1.48 case XVID_CSP_YV12 :
728 :     img.y = convert->input.plane[0];
729 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
730 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
731 :     image_output(&img, width, height, width,
732 :     (uint8_t**)convert->output.plane, convert->output.stride,
733 :     convert->output.csp, convert->interlacing);
734 :     break;
735 : edgomez 1.41
736 : edgomez 1.48 default :
737 :     return XVID_ERR_FORMAT;
738 : edgomez 1.41 }
739 :    
740 :    
741 : edgomez 1.48 emms();
742 : edgomez 1.41 return 0;
743 :     }
744 :    
745 : edgomez 1.48 /*****************************************************************************
746 :     * XviD Global Entry point
747 :     *
748 :     * Well this function initialize all internal function pointers according
749 :     * to the CPU features forced by the library client or autodetected (depending
750 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
751 :     * image colorspace transformation tables.
752 :     *
753 :     ****************************************************************************/
754 : Isibaar 1.1
755 : edgomez 1.41
756 :     int
757 : edgomez 1.48 xvid_global(void *handle,
758 : edgomez 1.41 int opt,
759 :     void *param1,
760 :     void *param2)
761 :     {
762 :     switch(opt)
763 :     {
764 : edgomez 1.48 case XVID_GBL_INIT :
765 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
766 : edgomez 1.41
767 : edgomez 1.48 case XVID_GBL_INFO :
768 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
769 :    
770 :     case XVID_GBL_CONVERT :
771 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
772 : edgomez 1.41
773 :     default :
774 :     return XVID_ERR_FAIL;
775 :     }
776 :     }
777 :    
778 : edgomez 1.16 /*****************************************************************************
779 :     * XviD Native decoder entry point
780 :     *
781 :     * This function is just a wrapper to all the option cases.
782 :     *
783 :     * Returned values : XVID_ERR_FAIL when opt is invalid
784 :     * else returns the wrapped function result
785 :     *
786 :     ****************************************************************************/
787 :    
788 : edgomez 1.15 int
789 :     xvid_decore(void *handle,
790 :     int opt,
791 :     void *param1,
792 :     void *param2)
793 : Isibaar 1.1 {
794 : edgomez 1.15 switch (opt) {
795 :     case XVID_DEC_CREATE:
796 : edgomez 1.48 return decoder_create((xvid_dec_create_t *) param1);
797 : edgomez 1.15
798 :     case XVID_DEC_DESTROY:
799 :     return decoder_destroy((DECODER *) handle);
800 : Isibaar 1.1
801 : edgomez 1.48 case XVID_DEC_DECODE:
802 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
803 :    
804 : Isibaar 1.1 default:
805 : edgomez 1.15 return XVID_ERR_FAIL;
806 :     }
807 : Isibaar 1.1 }
808 :    
809 : edgomez 1.16
810 :     /*****************************************************************************
811 :     * XviD Native encoder entry point
812 :     *
813 :     * This function is just a wrapper to all the option cases.
814 :     *
815 :     * Returned values : XVID_ERR_FAIL when opt is invalid
816 :     * else returns the wrapped function result
817 :     *
818 :     ****************************************************************************/
819 : Isibaar 1.1
820 : edgomez 1.15 int
821 :     xvid_encore(void *handle,
822 :     int opt,
823 :     void *param1,
824 :     void *param2)
825 : Isibaar 1.1 {
826 : edgomez 1.15 switch (opt) {
827 :     case XVID_ENC_ENCODE:
828 : edgomez 1.41
829 : edgomez 1.48 return enc_encode((Encoder *) handle,
830 :     (xvid_enc_frame_t *) param1,
831 :     (xvid_enc_stats_t *) param2);
832 : edgomez 1.15
833 :     case XVID_ENC_CREATE:
834 : edgomez 1.48 return enc_create((xvid_enc_create_t *) param1);
835 : edgomez 1.15
836 :     case XVID_ENC_DESTROY:
837 : edgomez 1.48 return enc_destroy((Encoder *) handle);
838 : Isibaar 1.1
839 :     default:
840 : edgomez 1.15 return XVID_ERR_FAIL;
841 :     }
842 : Isibaar 1.1 }

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