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

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.81 - (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 : Isibaar 1.81 * $Id: xvid.c,v 1.80 2008/12/01 15:06:48 Isibaar 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 1.80 #if !defined(_WIN32)
32 :     #include <unistd.h>
33 :     #endif
34 :    
35 : Isibaar 1.81 #if defined(__APPLE__) && defined(__MACH__) && !defined(_SC_NPROCESSORS_CONF)
36 :     #include <sys/types.h>
37 :     #include <sys/sysctl.h>
38 :     #ifdef MAX
39 :     #undef MAX
40 :     #endif
41 :     #ifdef MIN
42 :     #undef MIN
43 :     #endif
44 :     #endif
45 :    
46 : Isibaar 1.1 #include "xvid.h"
47 :     #include "decoder.h"
48 :     #include "encoder.h"
49 :     #include "bitstream/cbp.h"
50 :     #include "dct/idct.h"
51 :     #include "dct/fdct.h"
52 :     #include "image/colorspace.h"
53 :     #include "image/interpolate8x8.h"
54 :     #include "utils/mem_transfer.h"
55 : edgomez 1.41 #include "utils/mbfunctions.h"
56 : edgomez 1.48 #include "quant/quant.h"
57 : ia64p 1.30 #include "motion/motion.h"
58 : Skal 1.69 #include "motion/gmc.h"
59 : Isibaar 1.1 #include "motion/sad.h"
60 :     #include "utils/emms.h"
61 :     #include "utils/timer.h"
62 : Isibaar 1.9 #include "bitstream/mbcoding.h"
63 : edgomez 1.48 #include "image/qpel.h"
64 :     #include "image/postprocessing.h"
65 :    
66 :     #if defined(_DEBUG)
67 :     unsigned int xvid_debug = 0; /* xvid debug mask */
68 :     #endif
69 : Isibaar 1.1
70 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
71 : edgomez 1.41 # include <windows.h>
72 : edgomez 1.60 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
73 : edgomez 1.41 # include <signal.h>
74 :     # include <setjmp.h>
75 : suxen_drol 1.31
76 : edgomez 1.41 static jmp_buf mark;
77 : suxen_drol 1.31
78 : edgomez 1.41 static void
79 :     sigill_handler(int signal)
80 :     {
81 :     longjmp(mark, 1);
82 :     }
83 : suxen_drol 1.31 #endif
84 :    
85 :    
86 :     /*
87 : edgomez 1.43 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
88 :     * signalled
89 :     *
90 :     * Return values:
91 :     * -1 : could not determine
92 :     * 0 : SIGILL was *not* signalled
93 :     * 1 : SIGILL was signalled
94 :     */
95 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
96 : edgomez 1.50 static int
97 : suxen_drol 1.31 sigill_check(void (*func)())
98 :     {
99 :     _try {
100 :     func();
101 : edgomez 1.50 } _except(EXCEPTION_EXECUTE_HANDLER) {
102 : suxen_drol 1.31
103 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
104 : edgomez 1.50 return(1);
105 : suxen_drol 1.31 }
106 : edgomez 1.50 return(0);
107 :     }
108 : edgomez 1.60 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || defined(ARCH_IS_PPC)
109 : edgomez 1.50 static int
110 :     sigill_check(void (*func)())
111 :     {
112 :     void *old_handler;
113 : suxen_drol 1.31 int jmpret;
114 :    
115 : edgomez 1.50 /* Set our SIGILL handler */
116 :     old_handler = signal(SIGILL, sigill_handler);
117 : suxen_drol 1.31
118 : edgomez 1.50 /* Check for error */
119 :     if (old_handler == SIG_ERR) {
120 :     return(-1);
121 : suxen_drol 1.31 }
122 :    
123 : edgomez 1.50 /* Save stack context, so if func triggers a SIGILL, we can still roll
124 :     * back to a valid CPU state */
125 :     jmpret = setjmp(mark);
126 :    
127 :     /* If setjmp returned directly, then its returned value is 0, and we still
128 :     * have to test the passed func. Otherwise it means the stack context has
129 :     * been restored by a longjmp() call, which in our case happens only in the
130 :     * signal handler */
131 :     if (jmpret == 0) {
132 : suxen_drol 1.31 func();
133 :     }
134 :    
135 : edgomez 1.50 /* Restore old signal handler */
136 : suxen_drol 1.31 signal(SIGILL, old_handler);
137 :    
138 : edgomez 1.50 return(jmpret);
139 : suxen_drol 1.31 }
140 :     #endif
141 :    
142 : edgomez 1.41
143 :     /* detect cpu flags */
144 :     static unsigned int
145 : suxen_drol 1.65 detect_cpu_flags(void)
146 : edgomez 1.41 {
147 :     /* enable native assembly optimizations by default */
148 :     unsigned int cpu_flags = XVID_CPU_ASM;
149 :    
150 : edgomez 1.60 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
151 : edgomez 1.41 cpu_flags |= check_cpu_features();
152 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
153 :     cpu_flags &= ~XVID_CPU_SSE;
154 :    
155 : Isibaar 1.74 if ((cpu_flags & (XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41)) && sigill_check(sse2_os_trigger))
156 :     cpu_flags &= ~(XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41);
157 : edgomez 1.41 #endif
158 :    
159 :     #if defined(ARCH_IS_PPC)
160 : edgomez 1.50 if (!sigill_check(altivec_trigger))
161 :     cpu_flags |= XVID_CPU_ALTIVEC;
162 : edgomez 1.41 #endif
163 :    
164 :     return cpu_flags;
165 :     }
166 :    
167 :    
168 : edgomez 1.16 /*****************************************************************************
169 :     * XviD Init Entry point
170 :     *
171 :     * Well this function initialize all internal function pointers according
172 :     * to the CPU features forced by the library client or autodetected (depending
173 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
174 :     * image colorspace transformation tables.
175 : edgomez 1.48 *
176 : edgomez 1.16 * Returned value : XVID_ERR_OK
177 :     * + API_VERSION in the input XVID_INIT_PARAM structure
178 :     * + core build " " " " "
179 :     *
180 :     ****************************************************************************/
181 :    
182 : edgomez 1.41
183 : edgomez 1.48 static
184 :     int xvid_gbl_init(xvid_gbl_init_t * init)
185 : Isibaar 1.1 {
186 : edgomez 1.48 unsigned int cpu_flags;
187 : suxen_drol 1.21
188 : edgomez 1.48 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
189 :     return XVID_ERR_VERSION;
190 : suxen_drol 1.31
191 : edgomez 1.48 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
192 : Isibaar 1.1
193 : edgomez 1.16 /* Initialize the function pointers */
194 : Isibaar 1.1 idct_int32_init();
195 : Isibaar 1.9 init_vlc_tables();
196 :    
197 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
198 : Isibaar 1.1 fdct = fdct_int32;
199 :     idct = idct_int32;
200 :    
201 : edgomez 1.16 /* Only needed on PPC Altivec archs */
202 : suxen_drol 1.65 sadInit = NULL;
203 : edgomez 1.15
204 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
205 : Isibaar 1.1 emms = emms_c;
206 :    
207 : edgomez 1.48 /* Qpel stuff */
208 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
209 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
210 :     xvid_Init_QP();
211 :    
212 : edgomez 1.16 /* Quantization functions */
213 : edgomez 1.48 quant_h263_intra = quant_h263_intra_c;
214 :     quant_h263_inter = quant_h263_inter_c;
215 :     dequant_h263_intra = dequant_h263_intra_c;
216 :     dequant_h263_inter = dequant_h263_inter_c;
217 :    
218 :     quant_mpeg_intra = quant_mpeg_intra_c;
219 :     quant_mpeg_inter = quant_mpeg_inter_c;
220 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
221 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
222 : Isibaar 1.1
223 : edgomez 1.16 /* Block transfer related functions */
224 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
225 :     transfer_16to8copy = transfer_16to8copy_c;
226 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
227 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_c;
228 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
229 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_c;
230 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
231 :     transfer8x8_copy = transfer8x8_copy_c;
232 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_c;
233 : Isibaar 1.1
234 : edgomez 1.41 /* Interlacing functions */
235 :     MBFieldTest = MBFieldTest_c;
236 :    
237 : edgomez 1.16 /* Image interpolation related functions */
238 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
239 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
240 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
241 :    
242 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_c;
243 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_c;
244 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_c;
245 :    
246 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c;
247 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c;
248 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c;
249 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c;
250 :    
251 : edgomez 1.41 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
252 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
253 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
254 :    
255 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
256 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
257 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
258 :    
259 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
260 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
261 :    
262 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
263 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
264 :    
265 : suxen_drol 1.49 /* postprocessing */
266 :     image_brightness = image_brightness_c;
267 :    
268 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
269 : Isibaar 1.1 colorspace_init();
270 :    
271 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
272 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_c;
273 :     rgb555_to_yv12 = rgb555_to_yv12_c;
274 :     rgb565_to_yv12 = rgb565_to_yv12_c;
275 : chl 1.71 rgb_to_yv12 = rgb_to_yv12_c;
276 : edgomez 1.41 bgr_to_yv12 = bgr_to_yv12_c;
277 :     bgra_to_yv12 = bgra_to_yv12_c;
278 :     abgr_to_yv12 = abgr_to_yv12_c;
279 :     rgba_to_yv12 = rgba_to_yv12_c;
280 : edgomez 1.48 argb_to_yv12 = argb_to_yv12_c;
281 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_c;
282 :     uyvy_to_yv12 = uyvy_to_yv12_c;
283 :    
284 :     rgb555i_to_yv12 = rgb555i_to_yv12_c;
285 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
286 :     bgri_to_yv12 = bgri_to_yv12_c;
287 :     bgrai_to_yv12 = bgrai_to_yv12_c;
288 :     abgri_to_yv12 = abgri_to_yv12_c;
289 :     rgbai_to_yv12 = rgbai_to_yv12_c;
290 : edgomez 1.48 argbi_to_yv12 = argbi_to_yv12_c;
291 : edgomez 1.41 yuyvi_to_yv12 = yuyvi_to_yv12_c;
292 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
293 :    
294 : edgomez 1.16 /* All colorspace transformation functions YV12->User format */
295 : edgomez 1.41 yv12_to_rgb555 = yv12_to_rgb555_c;
296 :     yv12_to_rgb565 = yv12_to_rgb565_c;
297 : chl 1.71 yv12_to_rgb = yv12_to_rgb_c;
298 : edgomez 1.41 yv12_to_bgr = yv12_to_bgr_c;
299 :     yv12_to_bgra = yv12_to_bgra_c;
300 :     yv12_to_abgr = yv12_to_abgr_c;
301 :     yv12_to_rgba = yv12_to_rgba_c;
302 : edgomez 1.48 yv12_to_argb = yv12_to_argb_c;
303 : edgomez 1.41 yv12_to_yuyv = yv12_to_yuyv_c;
304 :     yv12_to_uyvy = yv12_to_uyvy_c;
305 : edgomez 1.48
306 : edgomez 1.41 yv12_to_rgb555i = yv12_to_rgb555i_c;
307 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
308 :     yv12_to_bgri = yv12_to_bgri_c;
309 :     yv12_to_bgrai = yv12_to_bgrai_c;
310 :     yv12_to_abgri = yv12_to_abgri_c;
311 :     yv12_to_rgbai = yv12_to_rgbai_c;
312 : edgomez 1.48 yv12_to_argbi = yv12_to_argbi_c;
313 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_c;
314 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
315 : Isibaar 1.1
316 : edgomez 1.16 /* Functions used in motion estimation algorithms */
317 : edgomez 1.48 calc_cbp = calc_cbp_c;
318 :     sad16 = sad16_c;
319 :     sad8 = sad8_c;
320 :     sad16bi = sad16bi_c;
321 :     sad8bi = sad8bi_c;
322 :     dev16 = dev16_c;
323 :     sad16v = sad16v_c;
324 :     sse8_16bit = sse8_16bit_c;
325 : edgomez 1.52 sse8_8bit = sse8_8bit_c;
326 : Isibaar 1.1
327 : Skal 1.69 init_GMC(cpu_flags);
328 :    
329 : Isibaar 1.75 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
330 : edgomez 1.41
331 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
332 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
333 : Isibaar 1.74 (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2) ||
334 :     (cpu_flags & XVID_CPU_SSE3) || (cpu_flags & XVID_CPU_SSE41))
335 : edgomez 1.41 {
336 :     /* Restore FPU context : emms_c is a nop functions */
337 :     emms = emms_mmx;
338 :     }
339 :    
340 :     if ((cpu_flags & XVID_CPU_MMX)) {
341 : edgomez 1.16
342 :     /* Forward and Inverse Discrete Cosine Transformation functions */
343 : edgomez 1.48 fdct = fdct_mmx_skal;
344 : Isibaar 1.47 idct = idct_mmx;
345 : Isibaar 1.1
346 : edgomez 1.48 /* Qpel stuff */
347 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
348 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
349 :    
350 : edgomez 1.16 /* Quantization related functions */
351 : edgomez 1.48 quant_h263_intra = quant_h263_intra_mmx;
352 : Isibaar 1.75 quant_h263_inter = quant_h263_inter_mmx;
353 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_mmx;
354 :     dequant_h263_inter = dequant_h263_inter_mmx;
355 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
356 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
357 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
358 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
359 : Isibaar 1.1
360 : Isibaar 1.75
361 : edgomez 1.16 /* Block related functions */
362 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
363 :     transfer_16to8copy = transfer_16to8copy_mmx;
364 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
365 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_mmx;
366 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
367 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
368 :     transfer8x8_copy = transfer8x8_copy_mmx;
369 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_mmx;
370 : edgomez 1.22
371 : edgomez 1.41 /* Interlacing Functions */
372 :     MBFieldTest = MBFieldTest_mmx;
373 : edgomez 1.16
374 :     /* Image Interpolation related functions */
375 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
376 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
377 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
378 :    
379 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_mmx;
380 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_mmx;
381 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_mmx;
382 :    
383 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx;
384 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx;
385 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx;
386 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx;
387 :    
388 : edgomez 1.41 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
389 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
390 :    
391 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
392 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
393 : suxen_drol 1.49
394 :     /* postprocessing */
395 :     image_brightness = image_brightness_mmx;
396 : edgomez 1.41
397 :     /* image input xxx_to_yv12 related functions */
398 : Isibaar 1.75
399 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_mmx;
400 : Isibaar 1.75
401 : edgomez 1.41 bgr_to_yv12 = bgr_to_yv12_mmx;
402 : chl 1.72 rgb_to_yv12 = rgb_to_yv12_mmx;
403 : edgomez 1.41 bgra_to_yv12 = bgra_to_yv12_mmx;
404 : chl 1.72 rgba_to_yv12 = rgba_to_yv12_mmx;
405 : edgomez 1.16 yuyv_to_yv12 = yuyv_to_yv12_mmx;
406 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
407 : Isibaar 1.1
408 : edgomez 1.41 /* image output yv12_to_xxx related functions */
409 :     yv12_to_bgr = yv12_to_bgr_mmx;
410 :     yv12_to_bgra = yv12_to_bgra_mmx;
411 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
412 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
413 : edgomez 1.48
414 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
415 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
416 : Isibaar 1.1
417 : edgomez 1.16 /* Motion estimation related functions */
418 : edgomez 1.52 calc_cbp = calc_cbp_mmx;
419 :     sad16 = sad16_mmx;
420 :     sad8 = sad8_mmx;
421 :     sad16bi = sad16bi_mmx;
422 :     sad8bi = sad8bi_mmx;
423 :     dev16 = dev16_mmx;
424 :     sad16v = sad16v_mmx;
425 : edgomez 1.48 sse8_16bit = sse8_16bit_mmx;
426 : edgomez 1.52 sse8_8bit = sse8_8bit_mmx;
427 : Isibaar 1.1 }
428 :    
429 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
430 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
431 :    
432 :     emms = emms_3dn;
433 : suxen_drol 1.33
434 :     /* ME functions */
435 :     sad16bi = sad16bi_3dn;
436 :     sad8bi = sad8bi_3dn;
437 : edgomez 1.41
438 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
439 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
440 : Isibaar 1.75
441 : suxen_drol 1.33 }
442 :    
443 :    
444 : edgomez 1.41 if ((cpu_flags & XVID_CPU_MMXEXT)) {
445 : edgomez 1.16
446 : edgomez 1.48 /* DCT */
447 :     fdct = fdct_xmm_skal;
448 : Isibaar 1.47 idct = idct_xmm;
449 : edgomez 1.16
450 :     /* Interpolation */
451 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
452 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
453 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
454 : edgomez 1.55
455 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_xmm;
456 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_xmm;
457 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_xmm;
458 :    
459 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm;
460 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm;
461 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm;
462 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm;
463 : Isibaar 1.25
464 : Isibaar 1.75 /* Quantization */
465 : edgomez 1.48 quant_mpeg_inter = quant_mpeg_inter_xmm;
466 : edgomez 1.41
467 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_xmm;
468 :     dequant_h263_inter = dequant_h263_inter_xmm;
469 : chenm001 1.29
470 : Isibaar 1.75 /* Buffer transfer */
471 : edgomez 1.19 transfer_8to16sub2 = transfer_8to16sub2_xmm;
472 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_xmm;
473 : edgomez 1.16
474 :     /* Colorspace transformation */
475 : Isibaar 1.75 /* yv12_to_yv12 = yv12_to_yv12_xmm; */ /* appears to be slow on many machines */
476 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_xmm;
477 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
478 : Isibaar 1.1
479 : edgomez 1.16 /* ME functions */
480 : Isibaar 1.1 sad16 = sad16_xmm;
481 : suxen_drol 1.33 sad8 = sad8_xmm;
482 : chenm001 1.29 sad16bi = sad16bi_xmm;
483 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
484 : Isibaar 1.1 dev16 = dev16_xmm;
485 : edgomez 1.41 sad16v = sad16v_xmm;
486 : Isibaar 1.1 }
487 :    
488 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
489 : edgomez 1.16
490 :     /* Interpolation */
491 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
492 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
493 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
494 : suxen_drol 1.64
495 :     interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dn;
496 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dn;
497 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dn;
498 : Isibaar 1.1 }
499 :    
500 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
501 :    
502 :     /* Buffer transfer */
503 :     transfer_8to16copy = transfer_8to16copy_3dne;
504 :     transfer_16to8copy = transfer_16to8copy_3dne;
505 :     transfer_8to16sub = transfer_8to16sub_3dne;
506 :     transfer_8to16subro = transfer_8to16subro_3dne;
507 :     transfer_16to8add = transfer_16to8add_3dne;
508 :     transfer8x8_copy = transfer8x8_copy_3dne;
509 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_3dne;
510 : edgomez 1.41
511 : edgomez 1.51 if ((cpu_flags & XVID_CPU_MMXEXT)) {
512 :     /* Inverse DCT */
513 :     idct = idct_3dne;
514 :    
515 :     /* Buffer transfer */
516 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
517 :    
518 :     /* Interpolation */
519 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
520 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
521 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
522 :    
523 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dne;
524 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dne;
525 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dne;
526 :    
527 : Isibaar 1.75 /* Quantization */
528 : edgomez 1.51 quant_h263_intra = quant_h263_intra_3dne; /* cmov only */
529 :     quant_h263_inter = quant_h263_inter_3dne;
530 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */
531 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
532 :     dequant_h263_intra = dequant_h263_intra_3dne;
533 :     dequant_h263_inter = dequant_h263_inter_3dne;
534 :    
535 : Isibaar 1.75 /* ME functions */
536 : edgomez 1.51 sad16 = sad16_3dne;
537 :     sad8 = sad8_3dne;
538 :     sad16bi = sad16bi_3dne;
539 :     sad8bi = sad8bi_3dne;
540 :     dev16 = dev16_3dne;
541 :     }
542 : edgomez 1.41 }
543 : Isibaar 1.75
544 : edgomez 1.41 if ((cpu_flags & XVID_CPU_SSE2)) {
545 : edgomez 1.16
546 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
547 :    
548 : edgomez 1.16 /* Quantization */
549 : edgomez 1.48 quant_h263_intra = quant_h263_intra_sse2;
550 :     quant_h263_inter = quant_h263_inter_sse2;
551 :     dequant_h263_intra = dequant_h263_intra_sse2;
552 :     dequant_h263_inter = dequant_h263_inter_sse2;
553 : h 1.13
554 : edgomez 1.48 /* SAD operators */
555 : edgomez 1.16 sad16 = sad16_sse2;
556 :     dev16 = dev16_sse2;
557 : edgomez 1.48
558 : Skal 1.63 /* DCT operators */
559 : edgomez 1.48 fdct = fdct_sse2_skal;
560 : Isibaar 1.76 idct = idct_sse2_skal; /* Is now IEEE1180 and Walken compliant. */
561 : suxen_drol 1.54
562 :     /* postprocessing */
563 :     image_brightness = image_brightness_sse2;
564 : Isibaar 1.75
565 : h 1.12 }
566 : Isibaar 1.73
567 :     if ((cpu_flags & XVID_CPU_SSE3)) {
568 :    
569 :     /* SAD operators */
570 :     sad16 = sad16_sse3;
571 :     dev16 = dev16_sse3;
572 :     }
573 : Isibaar 1.74
574 : edgomez 1.48 #endif /* ARCH_IS_IA32 */
575 : Isibaar 1.18
576 : edgomez 1.41 #if defined(ARCH_IS_IA64)
577 : edgomez 1.43 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
578 : Isibaar 1.18 idct_ia64_init();
579 :     fdct = fdct_ia64;
580 : edgomez 1.43 idct = idct_ia64; /*not yet working, crashes */
581 : Isibaar 1.18 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
582 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
583 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
584 :     sad16 = sad16_ia64;
585 :     sad16bi = sad16bi_ia64;
586 :     sad8 = sad8_ia64;
587 :     dev16 = dev16_ia64;
588 : edgomez 1.43 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
589 : edgomez 1.48 quant_h263_intra = quant_h263_intra_ia64;
590 :     quant_h263_inter = quant_h263_inter_ia64;
591 :     dequant_h263_intra = dequant_h263_intra_ia64;
592 :     dequant_h263_inter = dequant_h263_inter_ia64;
593 : Isibaar 1.18 transfer_8to16copy = transfer_8to16copy_ia64;
594 :     transfer_16to8copy = transfer_16to8copy_ia64;
595 :     transfer_8to16sub = transfer_8to16sub_ia64;
596 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
597 :     transfer_16to8add = transfer_16to8add_ia64;
598 :     transfer8x8_copy = transfer8x8_copy_ia64;
599 :     }
600 : edgomez 1.48 #endif
601 : edgomez 1.16
602 : edgomez 1.41 #if defined(ARCH_IS_PPC)
603 : edgomez 1.50 if ((cpu_flags & XVID_CPU_ALTIVEC)) {
604 :     /* sad operators */
605 : edgomez 1.56 sad16 = sad16_altivec_c;
606 :     sad16bi = sad16bi_altivec_c;
607 :     sad8 = sad8_altivec_c;
608 :     dev16 = dev16_altivec_c;
609 : edgomez 1.50
610 :     sse8_16bit = sse8_16bit_altivec_c;
611 :    
612 :     /* mem transfer */
613 :     transfer_8to16copy = transfer_8to16copy_altivec_c;
614 :     transfer_16to8copy = transfer_16to8copy_altivec_c;
615 :     transfer_8to16sub = transfer_8to16sub_altivec_c;
616 :     transfer_8to16subro = transfer_8to16subro_altivec_c;
617 :     transfer_8to16sub2 = transfer_8to16sub2_altivec_c;
618 :     transfer_16to8add = transfer_16to8add_altivec_c;
619 :     transfer8x8_copy = transfer8x8_copy_altivec_c;
620 :    
621 :     /* Inverse DCT */
622 :     idct = idct_altivec_c;
623 :    
624 :     /* Interpolation */
625 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c;
626 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c;
627 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c;
628 : edgomez 1.58
629 : edgomez 1.50 interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c;
630 :     interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c;
631 : edgomez 1.56
632 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c;
633 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c;
634 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c;
635 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c;
636 : edgomez 1.50
637 :     /* Colorspace conversion */
638 :     bgra_to_yv12 = bgra_to_yv12_altivec_c;
639 :     abgr_to_yv12 = abgr_to_yv12_altivec_c;
640 :     rgba_to_yv12 = rgba_to_yv12_altivec_c;
641 :     argb_to_yv12 = argb_to_yv12_altivec_c;
642 :    
643 :     yuyv_to_yv12 = yuyv_to_yv12_altivec_c;
644 :     uyvy_to_yv12 = uyvy_to_yv12_altivec_c;
645 :    
646 :     yv12_to_yuyv = yv12_to_yuyv_altivec_c;
647 :     yv12_to_uyvy = yv12_to_uyvy_altivec_c;
648 :    
649 :     /* Quantization */
650 :     quant_h263_intra = quant_h263_intra_altivec_c;
651 :     quant_h263_inter = quant_h263_inter_altivec_c;
652 :     dequant_h263_intra = dequant_h263_intra_altivec_c;
653 :     dequant_h263_inter = dequant_h263_inter_altivec_c;
654 : edgomez 1.62
655 :     dequant_mpeg_intra = dequant_mpeg_intra_altivec_c;
656 :     dequant_mpeg_inter = dequant_mpeg_inter_altivec_c;
657 : edgomez 1.56
658 :     /* Qpel stuff */
659 :     xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C;
660 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C;
661 : edgomez 1.50 }
662 : canard 1.6 #endif
663 : edgomez 1.41
664 : edgomez 1.48 #if defined(_DEBUG)
665 :     xvid_debug = init->debug;
666 :     #endif
667 :    
668 : edgomez 1.50 return(0);
669 : edgomez 1.41 }
670 :    
671 :    
672 :     static int
673 : edgomez 1.48 xvid_gbl_info(xvid_gbl_info_t * info)
674 : edgomez 1.41 {
675 : edgomez 1.48 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
676 :     return XVID_ERR_VERSION;
677 : edgomez 1.41
678 : edgomez 1.48 info->actual_version = XVID_VERSION;
679 : Isibaar 1.78 info->build = "xvid-1.3.0-dev";
680 : edgomez 1.48 info->cpu_flags = detect_cpu_flags();
681 : Isibaar 1.81 info->num_threads = 0; /* single-thread */
682 : edgomez 1.41
683 : Isibaar 1.75 #if defined(_WIN32)
684 : Isibaar 1.79
685 : suxen_drol 1.68 {
686 : Isibaar 1.79 SYSTEM_INFO siSysInfo;
687 :     GetSystemInfo(&siSysInfo);
688 :     info->num_threads = siSysInfo.dwNumberOfProcessors; /* number of _logical_ cores */
689 : suxen_drol 1.68 }
690 : Isibaar 1.79
691 : Isibaar 1.81 #elif defined(_SC_NPROCESSORS_CONF) /* should be available on Apple too actually */
692 : Isibaar 1.75
693 :     info->num_threads = sysconf(_SC_NPROCESSORS_CONF);
694 :    
695 : Isibaar 1.81 #elif defined(__APPLE__) && defined(__MACH__)
696 :    
697 :     {
698 :     size_t len;
699 :     int mib[2], ncpu;
700 :    
701 :     mib[0] = CTL_HW;
702 :     mib[1] = HW_NCPU;
703 :     len = sizeof(ncpu);
704 :     if (sysctl(mib, 2, &ncpu, &len, NULL, 0) == 0)
705 :     info -> num_threads = ncpu;
706 :     else
707 :     info -> num_threads = 1;
708 :     }
709 :    
710 : edgomez 1.48 #endif
711 : edgomez 1.41
712 :     return 0;
713 :     }
714 :    
715 :    
716 : edgomez 1.48 static int
717 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
718 : edgomez 1.41 {
719 : edgomez 1.48 int width;
720 :     int height;
721 :     int width2;
722 :     int height2;
723 :     IMAGE img;
724 : edgomez 1.41
725 : edgomez 1.48 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
726 :     return XVID_ERR_VERSION;
727 : edgomez 1.41
728 : edgomez 1.48 #if 0
729 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
730 :     #endif
731 :     width = convert->width;
732 :     height = convert->height;
733 :     width2 = convert->width/2;
734 :     height2 = convert->height/2;
735 : edgomez 1.41
736 : edgomez 1.48 switch (convert->input.csp & ~XVID_CSP_VFLIP)
737 : edgomez 1.41 {
738 : edgomez 1.48 case XVID_CSP_YV12 :
739 :     img.y = convert->input.plane[0];
740 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
741 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
742 :     image_output(&img, width, height, width,
743 :     (uint8_t**)convert->output.plane, convert->output.stride,
744 :     convert->output.csp, convert->interlacing);
745 :     break;
746 : edgomez 1.41
747 : edgomez 1.48 default :
748 :     return XVID_ERR_FORMAT;
749 : edgomez 1.41 }
750 :    
751 :    
752 : edgomez 1.48 emms();
753 : edgomez 1.41 return 0;
754 :     }
755 :    
756 : edgomez 1.48 /*****************************************************************************
757 :     * XviD Global Entry point
758 :     *
759 :     * Well this function initialize all internal function pointers according
760 :     * to the CPU features forced by the library client or autodetected (depending
761 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
762 :     * image colorspace transformation tables.
763 :     *
764 :     ****************************************************************************/
765 : Isibaar 1.1
766 : edgomez 1.41
767 :     int
768 : edgomez 1.48 xvid_global(void *handle,
769 : edgomez 1.41 int opt,
770 :     void *param1,
771 :     void *param2)
772 :     {
773 :     switch(opt)
774 :     {
775 : edgomez 1.48 case XVID_GBL_INIT :
776 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
777 : edgomez 1.41
778 : edgomez 1.48 case XVID_GBL_INFO :
779 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
780 :    
781 :     case XVID_GBL_CONVERT :
782 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
783 : edgomez 1.41
784 :     default :
785 :     return XVID_ERR_FAIL;
786 :     }
787 :     }
788 :    
789 : edgomez 1.16 /*****************************************************************************
790 :     * XviD Native decoder entry point
791 :     *
792 :     * This function is just a wrapper to all the option cases.
793 :     *
794 :     * Returned values : XVID_ERR_FAIL when opt is invalid
795 :     * else returns the wrapped function result
796 :     *
797 :     ****************************************************************************/
798 :    
799 : edgomez 1.15 int
800 :     xvid_decore(void *handle,
801 :     int opt,
802 :     void *param1,
803 :     void *param2)
804 : Isibaar 1.1 {
805 : edgomez 1.15 switch (opt) {
806 :     case XVID_DEC_CREATE:
807 : edgomez 1.48 return decoder_create((xvid_dec_create_t *) param1);
808 : edgomez 1.15
809 :     case XVID_DEC_DESTROY:
810 :     return decoder_destroy((DECODER *) handle);
811 : Isibaar 1.1
812 : edgomez 1.48 case XVID_DEC_DECODE:
813 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
814 :    
815 : Isibaar 1.1 default:
816 : edgomez 1.15 return XVID_ERR_FAIL;
817 :     }
818 : Isibaar 1.1 }
819 :    
820 : edgomez 1.16
821 :     /*****************************************************************************
822 :     * XviD Native encoder entry point
823 :     *
824 :     * This function is just a wrapper to all the option cases.
825 :     *
826 :     * Returned values : XVID_ERR_FAIL when opt is invalid
827 :     * else returns the wrapped function result
828 :     *
829 :     ****************************************************************************/
830 : Isibaar 1.1
831 : edgomez 1.15 int
832 :     xvid_encore(void *handle,
833 :     int opt,
834 :     void *param1,
835 :     void *param2)
836 : Isibaar 1.1 {
837 : edgomez 1.15 switch (opt) {
838 :     case XVID_ENC_ENCODE:
839 : edgomez 1.41
840 : edgomez 1.48 return enc_encode((Encoder *) handle,
841 :     (xvid_enc_frame_t *) param1,
842 :     (xvid_enc_stats_t *) param2);
843 : edgomez 1.15
844 :     case XVID_ENC_CREATE:
845 : edgomez 1.48 return enc_create((xvid_enc_create_t *) param1);
846 : edgomez 1.15
847 :     case XVID_ENC_DESTROY:
848 : edgomez 1.48 return enc_destroy((Encoder *) handle);
849 : Isibaar 1.1
850 :     default:
851 : edgomez 1.15 return XVID_ERR_FAIL;
852 :     }
853 : Isibaar 1.1 }

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