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

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

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1950 - (view) (download)

1 : edgomez 200 /*****************************************************************************
2 : edgomez 201 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - Native API implementation -
5 :     *
6 : suxen_drol 1397 * Copyright(C) 2001-2004 Peter Ross <pross@xvid.org>
7 : Isibaar 1883 * 2002-2010 Michael Militzer <isibaar@xvid.org>
8 : edgomez 1382 *
9 : edgomez 851 * This program is free software ; you can redistribute it and/or modify
10 :     * it under the terms of the GNU General Public License as published by
11 :     * the Free Software Foundation ; either version 2 of the License, or
12 : edgomez 201 * (at your option) any later version.
13 :     *
14 :     * This program is distributed in the hope that it will be useful,
15 : edgomez 851 * but WITHOUT ANY WARRANTY ; without even the implied warranty of
16 : edgomez 201 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 :     * GNU General Public License for more details.
18 :     *
19 :     * You should have received a copy of the GNU General Public License
20 : edgomez 851 * along with this program ; if not, write to the Free Software
21 : edgomez 201 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 :     *
23 : Isibaar 1950 * $Id: xvid.c,v 1.85.2.2 2011-02-14 17:21:00 Isibaar Exp $
24 : edgomez 410 *
25 : edgomez 201 ****************************************************************************/
26 : chenm001 274
27 : edgomez 851 #include <stdio.h>
28 :     #include <stdlib.h>
29 :     #include <string.h>
30 :     #include <time.h>
31 :    
32 : Isibaar 1835 #if !defined(_WIN32)
33 :     #include <unistd.h>
34 :     #endif
35 :    
36 : Isibaar 1873 #if defined(__APPLE__) && defined(__MACH__) && !defined(_SC_NPROCESSORS_CONF)
37 :     #include <sys/types.h>
38 :     #include <sys/sysctl.h>
39 :     #ifdef MAX
40 :     #undef MAX
41 :     #endif
42 :     #ifdef MIN
43 :     #undef MIN
44 :     #endif
45 :     #endif
46 :    
47 : Isibaar 1874 #if defined(__amigaos4__)
48 :     #include <exec/exec.h>
49 :     #include <proto/exec.h>
50 :     #endif
51 :    
52 : Isibaar 3 #include "xvid.h"
53 :     #include "decoder.h"
54 :     #include "encoder.h"
55 :     #include "bitstream/cbp.h"
56 :     #include "dct/idct.h"
57 :     #include "dct/fdct.h"
58 :     #include "image/colorspace.h"
59 :     #include "image/interpolate8x8.h"
60 :     #include "utils/mem_transfer.h"
61 : edgomez 851 #include "utils/mbfunctions.h"
62 : edgomez 1382 #include "quant/quant.h"
63 : ia64p 299 #include "motion/motion.h"
64 : Skal 1709 #include "motion/gmc.h"
65 : Isibaar 3 #include "motion/sad.h"
66 :     #include "utils/emms.h"
67 :     #include "utils/timer.h"
68 : Isibaar 100 #include "bitstream/mbcoding.h"
69 : edgomez 1382 #include "image/qpel.h"
70 :     #include "image/postprocessing.h"
71 : Isibaar 3
72 : edgomez 1382 #if defined(_DEBUG)
73 :     unsigned int xvid_debug = 0; /* xvid debug mask */
74 :     #endif
75 : suxen_drol 311
76 : edgomez 1586 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
77 : edgomez 851 # include <windows.h>
78 : Isibaar 1874 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || (defined(ARCH_IS_PPC) && !defined(__amigaos4__))
79 : edgomez 851 # include <signal.h>
80 :     # include <setjmp.h>
81 : suxen_drol 311
82 : edgomez 851 static jmp_buf mark;
83 : suxen_drol 311
84 : edgomez 851 static void
85 :     sigill_handler(int signal)
86 :     {
87 :     longjmp(mark, 1);
88 :     }
89 : suxen_drol 311 #endif
90 :    
91 :     /*
92 : edgomez 874 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
93 :     * signalled
94 :     *
95 :     * Return values:
96 :     * -1 : could not determine
97 :     * 0 : SIGILL was *not* signalled
98 :     * 1 : SIGILL was signalled
99 :     */
100 : edgomez 1586 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
101 : edgomez 1412 static int
102 : suxen_drol 311 sigill_check(void (*func)())
103 :     {
104 :     _try {
105 :     func();
106 : edgomez 1412 } _except(EXCEPTION_EXECUTE_HANDLER) {
107 : suxen_drol 311
108 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
109 : edgomez 1412 return(1);
110 : suxen_drol 311 }
111 : edgomez 1412 return(0);
112 :     }
113 : Isibaar 1874 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || (defined(ARCH_IS_PPC) && !defined(__amigaos4__))
114 : edgomez 1412 static int
115 :     sigill_check(void (*func)())
116 :     {
117 :     void *old_handler;
118 : suxen_drol 311 int jmpret;
119 :    
120 : edgomez 1412 /* Set our SIGILL handler */
121 :     old_handler = signal(SIGILL, sigill_handler);
122 : suxen_drol 311
123 : edgomez 1412 /* Check for error */
124 :     if (old_handler == SIG_ERR) {
125 :     return(-1);
126 : suxen_drol 311 }
127 :    
128 : edgomez 1412 /* Save stack context, so if func triggers a SIGILL, we can still roll
129 :     * back to a valid CPU state */
130 :     jmpret = setjmp(mark);
131 :    
132 :     /* If setjmp returned directly, then its returned value is 0, and we still
133 :     * have to test the passed func. Otherwise it means the stack context has
134 :     * been restored by a longjmp() call, which in our case happens only in the
135 :     * signal handler */
136 :     if (jmpret == 0) {
137 : suxen_drol 311 func();
138 :     }
139 :    
140 : edgomez 1412 /* Restore old signal handler */
141 : suxen_drol 311 signal(SIGILL, old_handler);
142 :    
143 : edgomez 1412 return(jmpret);
144 : suxen_drol 311 }
145 :     #endif
146 :    
147 : edgomez 851
148 :     /* detect cpu flags */
149 :     static unsigned int
150 : suxen_drol 1653 detect_cpu_flags(void)
151 : edgomez 851 {
152 :     /* enable native assembly optimizations by default */
153 :     unsigned int cpu_flags = XVID_CPU_ASM;
154 :    
155 : edgomez 1586 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
156 : edgomez 851 cpu_flags |= check_cpu_features();
157 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
158 :     cpu_flags &= ~XVID_CPU_SSE;
159 :    
160 : Isibaar 1794 if ((cpu_flags & (XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41)) && sigill_check(sse2_os_trigger))
161 :     cpu_flags &= ~(XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41);
162 : edgomez 851 #endif
163 :    
164 :     #if defined(ARCH_IS_PPC)
165 : Isibaar 1874 #if defined(__amigaos4__)
166 :     {
167 :     uint32_t vector_unit = VECTORTYPE_NONE;
168 :     IExec->GetCPUInfoTags(GCIT_VectorUnit, &vector_unit, TAG_END);
169 :     if (vector_unit == VECTORTYPE_ALTIVEC) {
170 :     cpu_flags |= XVID_CPU_ALTIVEC;
171 :     }
172 :     }
173 :     #else
174 : edgomez 1412 if (!sigill_check(altivec_trigger))
175 :     cpu_flags |= XVID_CPU_ALTIVEC;
176 : edgomez 851 #endif
177 : Isibaar 1874 #endif
178 : edgomez 851
179 :     return cpu_flags;
180 :     }
181 :    
182 :    
183 : edgomez 200 /*****************************************************************************
184 : Isibaar 1883 * Xvid Init Entry point
185 : edgomez 200 *
186 :     * Well this function initialize all internal function pointers according
187 :     * to the CPU features forced by the library client or autodetected (depending
188 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
189 :     * image colorspace transformation tables.
190 : edgomez 1382 *
191 : edgomez 200 * Returned value : XVID_ERR_OK
192 :     * + API_VERSION in the input XVID_INIT_PARAM structure
193 :     * + core build " " " " "
194 :     *
195 :     ****************************************************************************/
196 :    
197 : edgomez 851
198 : edgomez 1382 static
199 :     int xvid_gbl_init(xvid_gbl_init_t * init)
200 : Isibaar 3 {
201 : edgomez 1382 unsigned int cpu_flags;
202 : Isibaar 3
203 : edgomez 1382 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
204 :     return XVID_ERR_VERSION;
205 : suxen_drol 234
206 : edgomez 1382 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
207 : suxen_drol 234
208 : edgomez 200 /* Initialize the function pointers */
209 : Isibaar 3 idct_int32_init();
210 : Isibaar 100 init_vlc_tables();
211 :    
212 : edgomez 200 /* Fixed Point Forward/Inverse DCT transformations */
213 : Isibaar 3 fdct = fdct_int32;
214 :     idct = idct_int32;
215 :    
216 : edgomez 200 /* Only needed on PPC Altivec archs */
217 : suxen_drol 1653 sadInit = NULL;
218 : edgomez 195
219 : edgomez 200 /* Restore FPU context : emms_c is a nop functions */
220 : Isibaar 3 emms = emms_c;
221 :    
222 : edgomez 1382 /* Qpel stuff */
223 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
224 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
225 :     xvid_Init_QP();
226 :    
227 : edgomez 200 /* Quantization functions */
228 : edgomez 1382 quant_h263_intra = quant_h263_intra_c;
229 :     quant_h263_inter = quant_h263_inter_c;
230 :     dequant_h263_intra = dequant_h263_intra_c;
231 :     dequant_h263_inter = dequant_h263_inter_c;
232 : Isibaar 3
233 : edgomez 1382 quant_mpeg_intra = quant_mpeg_intra_c;
234 :     quant_mpeg_inter = quant_mpeg_inter_c;
235 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
236 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
237 : Isibaar 3
238 : edgomez 200 /* Block transfer related functions */
239 : Isibaar 3 transfer_8to16copy = transfer_8to16copy_c;
240 :     transfer_16to8copy = transfer_16to8copy_c;
241 : edgomez 200 transfer_8to16sub = transfer_8to16sub_c;
242 : edgomez 851 transfer_8to16subro = transfer_8to16subro_c;
243 : suxen_drol 118 transfer_8to16sub2 = transfer_8to16sub2_c;
244 : syskin 1583 transfer_8to16sub2ro = transfer_8to16sub2ro_c;
245 : edgomez 200 transfer_16to8add = transfer_16to8add_c;
246 :     transfer8x8_copy = transfer8x8_copy_c;
247 : suxen_drol 1632 transfer8x4_copy = transfer8x4_copy_c;
248 : Isibaar 3
249 : edgomez 851 /* Interlacing functions */
250 :     MBFieldTest = MBFieldTest_c;
251 :    
252 : edgomez 200 /* Image interpolation related functions */
253 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
254 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
255 : Isibaar 3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
256 :    
257 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_c;
258 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_c;
259 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_c;
260 :    
261 : edgomez 1530 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c;
262 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c;
263 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c;
264 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c;
265 :    
266 : edgomez 851 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
267 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
268 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
269 :    
270 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
271 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
272 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
273 :    
274 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
275 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
276 :    
277 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
278 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
279 :    
280 : suxen_drol 1397 /* postprocessing */
281 :     image_brightness = image_brightness_c;
282 :    
283 : edgomez 200 /* Initialize internal colorspace transformation tables */
284 : Isibaar 3 colorspace_init();
285 :    
286 : edgomez 200 /* All colorspace transformation functions User Format->YV12 */
287 : edgomez 851 yv12_to_yv12 = yv12_to_yv12_c;
288 :     rgb555_to_yv12 = rgb555_to_yv12_c;
289 :     rgb565_to_yv12 = rgb565_to_yv12_c;
290 : chl 1762 rgb_to_yv12 = rgb_to_yv12_c;
291 : edgomez 851 bgr_to_yv12 = bgr_to_yv12_c;
292 :     bgra_to_yv12 = bgra_to_yv12_c;
293 :     abgr_to_yv12 = abgr_to_yv12_c;
294 :     rgba_to_yv12 = rgba_to_yv12_c;
295 : edgomez 1382 argb_to_yv12 = argb_to_yv12_c;
296 : edgomez 851 yuyv_to_yv12 = yuyv_to_yv12_c;
297 :     uyvy_to_yv12 = uyvy_to_yv12_c;
298 : Isibaar 3
299 : edgomez 851 rgb555i_to_yv12 = rgb555i_to_yv12_c;
300 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
301 :     bgri_to_yv12 = bgri_to_yv12_c;
302 :     bgrai_to_yv12 = bgrai_to_yv12_c;
303 :     abgri_to_yv12 = abgri_to_yv12_c;
304 :     rgbai_to_yv12 = rgbai_to_yv12_c;
305 : edgomez 1382 argbi_to_yv12 = argbi_to_yv12_c;
306 : edgomez 851 yuyvi_to_yv12 = yuyvi_to_yv12_c;
307 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
308 :    
309 : edgomez 200 /* All colorspace transformation functions YV12->User format */
310 : edgomez 851 yv12_to_rgb555 = yv12_to_rgb555_c;
311 :     yv12_to_rgb565 = yv12_to_rgb565_c;
312 : chl 1762 yv12_to_rgb = yv12_to_rgb_c;
313 : edgomez 851 yv12_to_bgr = yv12_to_bgr_c;
314 :     yv12_to_bgra = yv12_to_bgra_c;
315 :     yv12_to_abgr = yv12_to_abgr_c;
316 :     yv12_to_rgba = yv12_to_rgba_c;
317 : edgomez 1382 yv12_to_argb = yv12_to_argb_c;
318 : edgomez 851 yv12_to_yuyv = yv12_to_yuyv_c;
319 :     yv12_to_uyvy = yv12_to_uyvy_c;
320 : edgomez 1382
321 : edgomez 851 yv12_to_rgb555i = yv12_to_rgb555i_c;
322 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
323 :     yv12_to_bgri = yv12_to_bgri_c;
324 :     yv12_to_bgrai = yv12_to_bgrai_c;
325 :     yv12_to_abgri = yv12_to_abgri_c;
326 :     yv12_to_rgbai = yv12_to_rgbai_c;
327 : edgomez 1382 yv12_to_argbi = yv12_to_argbi_c;
328 : edgomez 851 yv12_to_yuyvi = yv12_to_yuyvi_c;
329 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
330 : Isibaar 3
331 : edgomez 200 /* Functions used in motion estimation algorithms */
332 : Isibaar 1909 calc_cbp = calc_cbp_c;
333 :     sad16 = sad16_c;
334 :     sad8 = sad8_c;
335 :     sad16bi = sad16bi_c;
336 :     sad8bi = sad8bi_c;
337 :     dev16 = dev16_c;
338 :     sad16v = sad16v_c;
339 :     sse8_16bit = sse8_16bit_c;
340 :     sse8_8bit = sse8_8bit_c;
341 : Isibaar 3
342 : Isibaar 1909 sseh8_16bit = sseh8_16bit_c;
343 :     coeff8_energy = coeff8_energy_c;
344 :     blocksum8 = blocksum8_c;
345 :    
346 : Skal 1709 init_GMC(cpu_flags);
347 :    
348 : Isibaar 1795 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
349 : edgomez 200
350 : edgomez 851 if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
351 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
352 : Isibaar 1794 (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2) ||
353 :     (cpu_flags & XVID_CPU_SSE3) || (cpu_flags & XVID_CPU_SSE41))
354 : edgomez 851 {
355 :     /* Restore FPU context : emms_c is a nop functions */
356 :     emms = emms_mmx;
357 :     }
358 :    
359 :     if ((cpu_flags & XVID_CPU_MMX)) {
360 :    
361 : edgomez 200 /* Forward and Inverse Discrete Cosine Transformation functions */
362 : edgomez 1382 fdct = fdct_mmx_skal;
363 : Isibaar 1066 idct = idct_mmx;
364 : Isibaar 3
365 : edgomez 1382 /* Qpel stuff */
366 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
367 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
368 :    
369 : edgomez 200 /* Quantization related functions */
370 : edgomez 1382 quant_h263_intra = quant_h263_intra_mmx;
371 : Isibaar 1795 quant_h263_inter = quant_h263_inter_mmx;
372 : edgomez 1382 dequant_h263_intra = dequant_h263_intra_mmx;
373 :     dequant_h263_inter = dequant_h263_inter_mmx;
374 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
375 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
376 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
377 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
378 : Isibaar 3
379 : Isibaar 1795
380 : edgomez 200 /* Block related functions */
381 : Isibaar 3 transfer_8to16copy = transfer_8to16copy_mmx;
382 :     transfer_16to8copy = transfer_16to8copy_mmx;
383 : edgomez 200 transfer_8to16sub = transfer_8to16sub_mmx;
384 : edgomez 851 transfer_8to16subro = transfer_8to16subro_mmx;
385 : edgomez 236 transfer_8to16sub2 = transfer_8to16sub2_mmx;
386 : edgomez 200 transfer_16to8add = transfer_16to8add_mmx;
387 :     transfer8x8_copy = transfer8x8_copy_mmx;
388 : suxen_drol 1632 transfer8x4_copy = transfer8x4_copy_mmx;
389 : Isibaar 3
390 : edgomez 851 /* Interlacing Functions */
391 :     MBFieldTest = MBFieldTest_mmx;
392 : edgomez 236
393 : edgomez 200 /* Image Interpolation related functions */
394 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
395 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
396 : Isibaar 3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
397 :    
398 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_mmx;
399 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_mmx;
400 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_mmx;
401 :    
402 : edgomez 1530 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx;
403 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx;
404 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx;
405 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx;
406 :    
407 : edgomez 851 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
408 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
409 :    
410 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
411 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
412 :    
413 : suxen_drol 1397 /* postprocessing */
414 :     image_brightness = image_brightness_mmx;
415 :    
416 : edgomez 851 /* image input xxx_to_yv12 related functions */
417 : Isibaar 1795
418 : edgomez 851 yv12_to_yv12 = yv12_to_yv12_mmx;
419 : Isibaar 1795
420 : edgomez 851 bgr_to_yv12 = bgr_to_yv12_mmx;
421 : chl 1763 rgb_to_yv12 = rgb_to_yv12_mmx;
422 : edgomez 851 bgra_to_yv12 = bgra_to_yv12_mmx;
423 : chl 1763 rgba_to_yv12 = rgba_to_yv12_mmx;
424 : edgomez 200 yuyv_to_yv12 = yuyv_to_yv12_mmx;
425 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
426 : Isibaar 3
427 : edgomez 851 /* image output yv12_to_xxx related functions */
428 :     yv12_to_bgr = yv12_to_bgr_mmx;
429 :     yv12_to_bgra = yv12_to_bgra_mmx;
430 : edgomez 200 yv12_to_yuyv = yv12_to_yuyv_mmx;
431 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
432 : edgomez 1382
433 : edgomez 851 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
434 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
435 : Isibaar 3
436 : edgomez 200 /* Motion estimation related functions */
437 : edgomez 1424 calc_cbp = calc_cbp_mmx;
438 :     sad16 = sad16_mmx;
439 :     sad8 = sad8_mmx;
440 :     sad16bi = sad16bi_mmx;
441 :     sad8bi = sad8bi_mmx;
442 :     dev16 = dev16_mmx;
443 :     sad16v = sad16v_mmx;
444 : edgomez 1382 sse8_16bit = sse8_16bit_mmx;
445 : edgomez 1424 sse8_8bit = sse8_8bit_mmx;
446 : Isibaar 3 }
447 :    
448 : suxen_drol 329 /* these 3dnow functions are faster than mmx, but slower than xmm. */
449 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOW)) {
450 : suxen_drol 329
451 : edgomez 851 emms = emms_3dn;
452 :    
453 : suxen_drol 329 /* ME functions */
454 :     sad16bi = sad16bi_3dn;
455 :     sad8bi = sad8bi_3dn;
456 : edgomez 851
457 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
458 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
459 : Isibaar 1795
460 : suxen_drol 329 }
461 :    
462 :    
463 : edgomez 851 if ((cpu_flags & XVID_CPU_MMXEXT)) {
464 : edgomez 200
465 : edgomez 1382 /* DCT */
466 :     fdct = fdct_xmm_skal;
467 : Isibaar 1066 idct = idct_xmm;
468 : edgomez 200
469 :     /* Interpolation */
470 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
471 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
472 : h 38 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
473 : edgomez 1530
474 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_xmm;
475 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_xmm;
476 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_xmm;
477 :    
478 : edgomez 1530 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm;
479 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm;
480 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm;
481 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm;
482 : edgomez 200
483 : Isibaar 1795 /* Quantization */
484 : edgomez 1382 quant_mpeg_inter = quant_mpeg_inter_xmm;
485 : edgomez 851
486 : edgomez 1382 dequant_h263_intra = dequant_h263_intra_xmm;
487 :     dequant_h263_inter = dequant_h263_inter_xmm;
488 : chenm001 274
489 : Isibaar 1795 /* Buffer transfer */
490 : edgomez 218 transfer_8to16sub2 = transfer_8to16sub2_xmm;
491 : syskin 1583 transfer_8to16sub2ro = transfer_8to16sub2ro_xmm;
492 : edgomez 218
493 : edgomez 200 /* Colorspace transformation */
494 : Isibaar 1795 /* yv12_to_yv12 = yv12_to_yv12_xmm; */ /* appears to be slow on many machines */
495 : edgomez 851 yuyv_to_yv12 = yuyv_to_yv12_xmm;
496 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
497 : Isibaar 3
498 : edgomez 200 /* ME functions */
499 : Isibaar 3 sad16 = sad16_xmm;
500 : suxen_drol 329 sad8 = sad8_xmm;
501 : chenm001 274 sad16bi = sad16bi_xmm;
502 : suxen_drol 329 sad8bi = sad8bi_xmm;
503 : Isibaar 3 dev16 = dev16_xmm;
504 : edgomez 851 sad16v = sad16v_xmm;
505 : Isibaar 3 }
506 :    
507 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOW)) {
508 : edgomez 200
509 :     /* Interpolation */
510 : Isibaar 3 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
511 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
512 : h 40 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
513 : suxen_drol 1632
514 :     interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dn;
515 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dn;
516 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dn;
517 : Isibaar 3 }
518 :    
519 : edgomez 851 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
520 : edgomez 200
521 : edgomez 851 /* Buffer transfer */
522 :     transfer_8to16copy = transfer_8to16copy_3dne;
523 :     transfer_16to8copy = transfer_16to8copy_3dne;
524 :     transfer_8to16sub = transfer_8to16sub_3dne;
525 :     transfer_8to16subro = transfer_8to16subro_3dne;
526 :     transfer_16to8add = transfer_16to8add_3dne;
527 :     transfer8x8_copy = transfer8x8_copy_3dne;
528 : suxen_drol 1632 transfer8x4_copy = transfer8x4_copy_3dne;
529 : edgomez 851
530 : edgomez 1423 if ((cpu_flags & XVID_CPU_MMXEXT)) {
531 :     /* Inverse DCT */
532 :     idct = idct_3dne;
533 : edgomez 851
534 : edgomez 1423 /* Buffer transfer */
535 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
536 : edgomez 851
537 : edgomez 1423 /* Interpolation */
538 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
539 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
540 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
541 :    
542 : suxen_drol 1632 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dne;
543 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dne;
544 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dne;
545 :    
546 : Isibaar 1795 /* Quantization */
547 : edgomez 1423 quant_h263_intra = quant_h263_intra_3dne; /* cmov only */
548 :     quant_h263_inter = quant_h263_inter_3dne;
549 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */
550 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
551 :     dequant_h263_intra = dequant_h263_intra_3dne;
552 :     dequant_h263_inter = dequant_h263_inter_3dne;
553 :    
554 : Isibaar 1795 /* ME functions */
555 : edgomez 1423 sad16 = sad16_3dne;
556 :     sad8 = sad8_3dne;
557 :     sad16bi = sad16bi_3dne;
558 :     sad8bi = sad8bi_3dne;
559 :     dev16 = dev16_3dne;
560 :     }
561 : edgomez 851 }
562 : Isibaar 1795
563 : edgomez 851 if ((cpu_flags & XVID_CPU_SSE2)) {
564 :    
565 : chenm001 274 calc_cbp = calc_cbp_sse2;
566 :    
567 : edgomez 200 /* Quantization */
568 : edgomez 1382 quant_h263_intra = quant_h263_intra_sse2;
569 :     quant_h263_inter = quant_h263_inter_sse2;
570 :     dequant_h263_intra = dequant_h263_intra_sse2;
571 :     dequant_h263_inter = dequant_h263_inter_sse2;
572 : h 135
573 : edgomez 1382 /* SAD operators */
574 : Isibaar 1909 sad16 = sad16_sse2;
575 :     dev16 = dev16_sse2;
576 : edgomez 1382
577 : Isibaar 1909 /* PSNR-HVS-M distortion metric */
578 :     sseh8_16bit = sseh8_16bit_sse2;
579 :     coeff8_energy = coeff8_energy_sse2;
580 :     blocksum8 = blocksum8_sse2;
581 :    
582 : Skal 1618 /* DCT operators */
583 : edgomez 1382 fdct = fdct_sse2_skal;
584 : Isibaar 1805 idct = idct_sse2_skal; /* Is now IEEE1180 and Walken compliant. */
585 : suxen_drol 1493
586 :     /* postprocessing */
587 :     image_brightness = image_brightness_sse2;
588 : Isibaar 1795
589 : h 126 }
590 : Isibaar 1764
591 :     if ((cpu_flags & XVID_CPU_SSE3)) {
592 :    
593 :     /* SAD operators */
594 :     sad16 = sad16_sse3;
595 :     dev16 = dev16_sse3;
596 :     }
597 : Isibaar 1794
598 : edgomez 1382 #endif /* ARCH_IS_IA32 */
599 : edgomez 200
600 : edgomez 851 #if defined(ARCH_IS_IA64)
601 : edgomez 874 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
602 : Isibaar 209 idct_ia64_init();
603 :     fdct = fdct_ia64;
604 : edgomez 874 idct = idct_ia64; /*not yet working, crashes */
605 : Isibaar 209 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
606 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
607 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
608 :     sad16 = sad16_ia64;
609 :     sad16bi = sad16bi_ia64;
610 :     sad8 = sad8_ia64;
611 :     dev16 = dev16_ia64;
612 : edgomez 874 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
613 : edgomez 1382 quant_h263_intra = quant_h263_intra_ia64;
614 :     quant_h263_inter = quant_h263_inter_ia64;
615 :     dequant_h263_intra = dequant_h263_intra_ia64;
616 :     dequant_h263_inter = dequant_h263_inter_ia64;
617 : Isibaar 209 transfer_8to16copy = transfer_8to16copy_ia64;
618 :     transfer_16to8copy = transfer_16to8copy_ia64;
619 :     transfer_8to16sub = transfer_8to16sub_ia64;
620 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
621 :     transfer_16to8add = transfer_16to8add_ia64;
622 :     transfer8x8_copy = transfer8x8_copy_ia64;
623 :     }
624 : edgomez 1382 #endif
625 : Isibaar 209
626 : edgomez 851 #if defined(ARCH_IS_PPC)
627 : edgomez 1412 if ((cpu_flags & XVID_CPU_ALTIVEC)) {
628 :     /* sad operators */
629 : edgomez 1557 sad16 = sad16_altivec_c;
630 :     sad16bi = sad16bi_altivec_c;
631 :     sad8 = sad8_altivec_c;
632 :     dev16 = dev16_altivec_c;
633 : edgomez 1412
634 :     sse8_16bit = sse8_16bit_altivec_c;
635 :    
636 :     /* mem transfer */
637 :     transfer_8to16copy = transfer_8to16copy_altivec_c;
638 :     transfer_16to8copy = transfer_16to8copy_altivec_c;
639 :     transfer_8to16sub = transfer_8to16sub_altivec_c;
640 :     transfer_8to16subro = transfer_8to16subro_altivec_c;
641 :     transfer_8to16sub2 = transfer_8to16sub2_altivec_c;
642 :     transfer_16to8add = transfer_16to8add_altivec_c;
643 :     transfer8x8_copy = transfer8x8_copy_altivec_c;
644 :    
645 :     /* Inverse DCT */
646 :     idct = idct_altivec_c;
647 :    
648 :     /* Interpolation */
649 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c;
650 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c;
651 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c;
652 : edgomez 1570
653 : edgomez 1412 interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c;
654 :     interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c;
655 : edgomez 1557
656 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c;
657 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c;
658 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c;
659 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c;
660 : edgomez 1412
661 :     /* Colorspace conversion */
662 :     bgra_to_yv12 = bgra_to_yv12_altivec_c;
663 :     abgr_to_yv12 = abgr_to_yv12_altivec_c;
664 :     rgba_to_yv12 = rgba_to_yv12_altivec_c;
665 :     argb_to_yv12 = argb_to_yv12_altivec_c;
666 :    
667 :     yuyv_to_yv12 = yuyv_to_yv12_altivec_c;
668 :     uyvy_to_yv12 = uyvy_to_yv12_altivec_c;
669 :    
670 :     yv12_to_yuyv = yv12_to_yuyv_altivec_c;
671 :     yv12_to_uyvy = yv12_to_uyvy_altivec_c;
672 :    
673 :     /* Quantization */
674 :     quant_h263_intra = quant_h263_intra_altivec_c;
675 :     quant_h263_inter = quant_h263_inter_altivec_c;
676 :     dequant_h263_intra = dequant_h263_intra_altivec_c;
677 :     dequant_h263_inter = dequant_h263_inter_altivec_c;
678 : edgomez 1609
679 :     dequant_mpeg_intra = dequant_mpeg_intra_altivec_c;
680 :     dequant_mpeg_inter = dequant_mpeg_inter_altivec_c;
681 : edgomez 1557
682 :     /* Qpel stuff */
683 :     xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C;
684 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C;
685 : edgomez 1412 }
686 : canard 52 #endif
687 : edgomez 851
688 : edgomez 1382 #if defined(_DEBUG)
689 :     xvid_debug = init->debug;
690 :     #endif
691 :    
692 : edgomez 1412 return(0);
693 : edgomez 851 }
694 :    
695 :    
696 :     static int
697 : edgomez 1382 xvid_gbl_info(xvid_gbl_info_t * info)
698 : edgomez 851 {
699 : edgomez 1382 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
700 :     return XVID_ERR_VERSION;
701 : edgomez 851
702 : edgomez 1382 info->actual_version = XVID_VERSION;
703 : Isibaar 1950 info->build = "xvid-1.3.0";
704 : edgomez 1382 info->cpu_flags = detect_cpu_flags();
705 : Isibaar 1873 info->num_threads = 0; /* single-thread */
706 : edgomez 851
707 : Isibaar 1795 #if defined(_WIN32)
708 : Isibaar 1827
709 : suxen_drol 1688 {
710 : Isibaar 1827 SYSTEM_INFO siSysInfo;
711 :     GetSystemInfo(&siSysInfo);
712 :     info->num_threads = siSysInfo.dwNumberOfProcessors; /* number of _logical_ cores */
713 : suxen_drol 1688 }
714 : Isibaar 1827
715 : Isibaar 1873 #elif defined(_SC_NPROCESSORS_CONF) /* should be available on Apple too actually */
716 : Isibaar 1795
717 :     info->num_threads = sysconf(_SC_NPROCESSORS_CONF);
718 :    
719 : Isibaar 1873 #elif defined(__APPLE__) && defined(__MACH__)
720 :    
721 :     {
722 :     size_t len;
723 :     int mib[2], ncpu;
724 :    
725 :     mib[0] = CTL_HW;
726 :     mib[1] = HW_NCPU;
727 :     len = sizeof(ncpu);
728 :     if (sysctl(mib, 2, &ncpu, &len, NULL, 0) == 0)
729 :     info -> num_threads = ncpu;
730 :     else
731 :     info -> num_threads = 1;
732 :     }
733 :    
734 : Isibaar 1874 #elif defined(__amigaos4__)
735 :    
736 :     {
737 :     uint32_t num_threads = 1;
738 :     IExec->GetCPUInfoTags(GCIT_NumberOfCPUs, &num_threads, TAG_END);
739 :     info->num_threads = num_threads;
740 :     }
741 :    
742 : edgomez 1382 #endif
743 : edgomez 851
744 :     return 0;
745 :     }
746 :    
747 :    
748 : edgomez 1382 static int
749 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
750 : edgomez 851 {
751 : edgomez 1382 int width;
752 :     int height;
753 :     int width2;
754 :     int height2;
755 :     IMAGE img;
756 : edgomez 851
757 : edgomez 1382 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
758 :     return XVID_ERR_VERSION;
759 : edgomez 851
760 : edgomez 1382 #if 0
761 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
762 :     #endif
763 :     width = convert->width;
764 :     height = convert->height;
765 :     width2 = convert->width/2;
766 :     height2 = convert->height/2;
767 : edgomez 851
768 : edgomez 1382 switch (convert->input.csp & ~XVID_CSP_VFLIP)
769 : edgomez 851 {
770 : edgomez 1382 case XVID_CSP_YV12 :
771 :     img.y = convert->input.plane[0];
772 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
773 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
774 :     image_output(&img, width, height, width,
775 :     (uint8_t**)convert->output.plane, convert->output.stride,
776 :     convert->output.csp, convert->interlacing);
777 :     break;
778 : Isibaar 1897
779 :     case XVID_CSP_INTERNAL :
780 :     img.y = (uint8_t*)convert->input.plane[0];
781 :     img.u = (uint8_t*)convert->input.plane[1];
782 :     img.v = (uint8_t*)convert->input.plane[2];
783 :     image_output(&img, width, height, convert->input.stride[0],
784 :     (uint8_t**)convert->output.plane, convert->output.stride,
785 :     convert->output.csp, convert->interlacing);
786 :     break;
787 : edgomez 851
788 : edgomez 1382 default :
789 :     return XVID_ERR_FORMAT;
790 : edgomez 851 }
791 :    
792 :    
793 : edgomez 1382 emms();
794 : edgomez 851 return 0;
795 :     }
796 :    
797 : edgomez 1382 /*****************************************************************************
798 : Isibaar 1883 * Xvid Global Entry point
799 : edgomez 1382 *
800 :     * Well this function initialize all internal function pointers according
801 :     * to the CPU features forced by the library client or autodetected (depending
802 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
803 :     * image colorspace transformation tables.
804 :     *
805 :     ****************************************************************************/
806 : edgomez 851
807 :    
808 :     int
809 : edgomez 1382 xvid_global(void *handle,
810 : edgomez 851 int opt,
811 :     void *param1,
812 :     void *param2)
813 :     {
814 :     switch(opt)
815 :     {
816 : edgomez 1382 case XVID_GBL_INIT :
817 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
818 : edgomez 851
819 : edgomez 1382 case XVID_GBL_INFO :
820 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
821 : edgomez 851
822 : edgomez 1382 case XVID_GBL_CONVERT :
823 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
824 :    
825 : edgomez 851 default :
826 :     return XVID_ERR_FAIL;
827 :     }
828 :     }
829 :    
830 : edgomez 200 /*****************************************************************************
831 : Isibaar 1883 * Xvid Native decoder entry point
832 : edgomez 200 *
833 :     * This function is just a wrapper to all the option cases.
834 :     *
835 :     * Returned values : XVID_ERR_FAIL when opt is invalid
836 :     * else returns the wrapped function result
837 :     *
838 :     ****************************************************************************/
839 :    
840 : edgomez 195 int
841 :     xvid_decore(void *handle,
842 :     int opt,
843 :     void *param1,
844 :     void *param2)
845 : Isibaar 3 {
846 : edgomez 195 switch (opt) {
847 :     case XVID_DEC_CREATE:
848 : edgomez 1382 return decoder_create((xvid_dec_create_t *) param1);
849 : Isibaar 3
850 : edgomez 195 case XVID_DEC_DESTROY:
851 :     return decoder_destroy((DECODER *) handle);
852 :    
853 : edgomez 1382 case XVID_DEC_DECODE:
854 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
855 :    
856 : Isibaar 3 default:
857 : edgomez 195 return XVID_ERR_FAIL;
858 :     }
859 : Isibaar 3 }
860 :    
861 :    
862 : edgomez 200 /*****************************************************************************
863 : Isibaar 1883 * Xvid Native encoder entry point
864 : edgomez 200 *
865 :     * This function is just a wrapper to all the option cases.
866 :     *
867 :     * Returned values : XVID_ERR_FAIL when opt is invalid
868 :     * else returns the wrapped function result
869 :     *
870 :     ****************************************************************************/
871 :    
872 : edgomez 195 int
873 :     xvid_encore(void *handle,
874 :     int opt,
875 :     void *param1,
876 :     void *param2)
877 : Isibaar 3 {
878 : edgomez 195 switch (opt) {
879 :     case XVID_ENC_ENCODE:
880 : edgomez 851
881 : edgomez 1382 return enc_encode((Encoder *) handle,
882 :     (xvid_enc_frame_t *) param1,
883 :     (xvid_enc_stats_t *) param2);
884 : Isibaar 3
885 : edgomez 195 case XVID_ENC_CREATE:
886 : edgomez 1382 return enc_create((xvid_enc_create_t *) param1);
887 : Isibaar 3
888 : edgomez 195 case XVID_ENC_DESTROY:
889 : edgomez 1382 return enc_destroy((Encoder *) handle);
890 : edgomez 195
891 : Isibaar 3 default:
892 : edgomez 195 return XVID_ERR_FAIL;
893 :     }
894 : Isibaar 3 }

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