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

Annotation of /xvidcore/src/xvid.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.82 - (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 1.82 * $Id: xvid.c,v 1.81 2009/06/02 13:06:49 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 1.82 #if defined(__amigaos4__)
47 :     #include <exec/exec.h>
48 :     #include <proto/exec.h>
49 :     #endif
50 :    
51 : Isibaar 1.1 #include "xvid.h"
52 :     #include "decoder.h"
53 :     #include "encoder.h"
54 :     #include "bitstream/cbp.h"
55 :     #include "dct/idct.h"
56 :     #include "dct/fdct.h"
57 :     #include "image/colorspace.h"
58 :     #include "image/interpolate8x8.h"
59 :     #include "utils/mem_transfer.h"
60 : edgomez 1.41 #include "utils/mbfunctions.h"
61 : edgomez 1.48 #include "quant/quant.h"
62 : ia64p 1.30 #include "motion/motion.h"
63 : Skal 1.69 #include "motion/gmc.h"
64 : Isibaar 1.1 #include "motion/sad.h"
65 :     #include "utils/emms.h"
66 :     #include "utils/timer.h"
67 : Isibaar 1.9 #include "bitstream/mbcoding.h"
68 : edgomez 1.48 #include "image/qpel.h"
69 :     #include "image/postprocessing.h"
70 :    
71 :     #if defined(_DEBUG)
72 :     unsigned int xvid_debug = 0; /* xvid debug mask */
73 :     #endif
74 : Isibaar 1.1
75 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
76 : edgomez 1.41 # include <windows.h>
77 : Isibaar 1.82 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || (defined(ARCH_IS_PPC) && !defined(__amigaos4__))
78 : edgomez 1.41 # include <signal.h>
79 :     # include <setjmp.h>
80 : suxen_drol 1.31
81 : edgomez 1.41 static jmp_buf mark;
82 : suxen_drol 1.31
83 : edgomez 1.41 static void
84 :     sigill_handler(int signal)
85 :     {
86 :     longjmp(mark, 1);
87 :     }
88 : suxen_drol 1.31 #endif
89 :    
90 :     /*
91 : edgomez 1.43 * Calls the funcptr, and returns whether SIGILL (illegal instruction) was
92 :     * signalled
93 :     *
94 :     * Return values:
95 :     * -1 : could not determine
96 :     * 0 : SIGILL was *not* signalled
97 :     * 1 : SIGILL was signalled
98 :     */
99 : edgomez 1.60 #if (defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)) && defined(_MSC_VER)
100 : edgomez 1.50 static int
101 : suxen_drol 1.31 sigill_check(void (*func)())
102 :     {
103 :     _try {
104 :     func();
105 : edgomez 1.50 } _except(EXCEPTION_EXECUTE_HANDLER) {
106 : suxen_drol 1.31
107 :     if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
108 : edgomez 1.50 return(1);
109 : suxen_drol 1.31 }
110 : edgomez 1.50 return(0);
111 :     }
112 : Isibaar 1.82 #elif defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) || (defined(ARCH_IS_PPC) && !defined(__amigaos4__))
113 : edgomez 1.50 static int
114 :     sigill_check(void (*func)())
115 :     {
116 :     void *old_handler;
117 : suxen_drol 1.31 int jmpret;
118 :    
119 : edgomez 1.50 /* Set our SIGILL handler */
120 :     old_handler = signal(SIGILL, sigill_handler);
121 : suxen_drol 1.31
122 : edgomez 1.50 /* Check for error */
123 :     if (old_handler == SIG_ERR) {
124 :     return(-1);
125 : suxen_drol 1.31 }
126 :    
127 : edgomez 1.50 /* Save stack context, so if func triggers a SIGILL, we can still roll
128 :     * back to a valid CPU state */
129 :     jmpret = setjmp(mark);
130 :    
131 :     /* If setjmp returned directly, then its returned value is 0, and we still
132 :     * have to test the passed func. Otherwise it means the stack context has
133 :     * been restored by a longjmp() call, which in our case happens only in the
134 :     * signal handler */
135 :     if (jmpret == 0) {
136 : suxen_drol 1.31 func();
137 :     }
138 :    
139 : edgomez 1.50 /* Restore old signal handler */
140 : suxen_drol 1.31 signal(SIGILL, old_handler);
141 :    
142 : edgomez 1.50 return(jmpret);
143 : suxen_drol 1.31 }
144 :     #endif
145 :    
146 : edgomez 1.41
147 :     /* detect cpu flags */
148 :     static unsigned int
149 : suxen_drol 1.65 detect_cpu_flags(void)
150 : edgomez 1.41 {
151 :     /* enable native assembly optimizations by default */
152 :     unsigned int cpu_flags = XVID_CPU_ASM;
153 :    
154 : edgomez 1.60 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
155 : edgomez 1.41 cpu_flags |= check_cpu_features();
156 :     if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
157 :     cpu_flags &= ~XVID_CPU_SSE;
158 :    
159 : Isibaar 1.74 if ((cpu_flags & (XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41)) && sigill_check(sse2_os_trigger))
160 :     cpu_flags &= ~(XVID_CPU_SSE2|XVID_CPU_SSE3|XVID_CPU_SSE41);
161 : edgomez 1.41 #endif
162 :    
163 :     #if defined(ARCH_IS_PPC)
164 : Isibaar 1.82 #if defined(__amigaos4__)
165 :     {
166 :     uint32_t vector_unit = VECTORTYPE_NONE;
167 :     IExec->GetCPUInfoTags(GCIT_VectorUnit, &vector_unit, TAG_END);
168 :     if (vector_unit == VECTORTYPE_ALTIVEC) {
169 :     cpu_flags |= XVID_CPU_ALTIVEC;
170 :     }
171 :     }
172 :     #else
173 : edgomez 1.50 if (!sigill_check(altivec_trigger))
174 :     cpu_flags |= XVID_CPU_ALTIVEC;
175 : edgomez 1.41 #endif
176 : Isibaar 1.82 #endif
177 : edgomez 1.41
178 :     return cpu_flags;
179 :     }
180 :    
181 :    
182 : edgomez 1.16 /*****************************************************************************
183 :     * XviD Init Entry point
184 :     *
185 :     * Well this function initialize all internal function pointers according
186 :     * to the CPU features forced by the library client or autodetected (depending
187 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
188 :     * image colorspace transformation tables.
189 : edgomez 1.48 *
190 : edgomez 1.16 * Returned value : XVID_ERR_OK
191 :     * + API_VERSION in the input XVID_INIT_PARAM structure
192 :     * + core build " " " " "
193 :     *
194 :     ****************************************************************************/
195 :    
196 : edgomez 1.41
197 : edgomez 1.48 static
198 :     int xvid_gbl_init(xvid_gbl_init_t * init)
199 : Isibaar 1.1 {
200 : edgomez 1.48 unsigned int cpu_flags;
201 : suxen_drol 1.21
202 : edgomez 1.48 if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */
203 :     return XVID_ERR_VERSION;
204 : suxen_drol 1.31
205 : edgomez 1.48 cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags();
206 : Isibaar 1.1
207 : edgomez 1.16 /* Initialize the function pointers */
208 : Isibaar 1.1 idct_int32_init();
209 : Isibaar 1.9 init_vlc_tables();
210 :    
211 : edgomez 1.16 /* Fixed Point Forward/Inverse DCT transformations */
212 : Isibaar 1.1 fdct = fdct_int32;
213 :     idct = idct_int32;
214 :    
215 : edgomez 1.16 /* Only needed on PPC Altivec archs */
216 : suxen_drol 1.65 sadInit = NULL;
217 : edgomez 1.15
218 : edgomez 1.16 /* Restore FPU context : emms_c is a nop functions */
219 : Isibaar 1.1 emms = emms_c;
220 :    
221 : edgomez 1.48 /* Qpel stuff */
222 :     xvid_QP_Funcs = &xvid_QP_Funcs_C;
223 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C;
224 :     xvid_Init_QP();
225 :    
226 : edgomez 1.16 /* Quantization functions */
227 : edgomez 1.48 quant_h263_intra = quant_h263_intra_c;
228 :     quant_h263_inter = quant_h263_inter_c;
229 :     dequant_h263_intra = dequant_h263_intra_c;
230 :     dequant_h263_inter = dequant_h263_inter_c;
231 :    
232 :     quant_mpeg_intra = quant_mpeg_intra_c;
233 :     quant_mpeg_inter = quant_mpeg_inter_c;
234 :     dequant_mpeg_intra = dequant_mpeg_intra_c;
235 :     dequant_mpeg_inter = dequant_mpeg_inter_c;
236 : Isibaar 1.1
237 : edgomez 1.16 /* Block transfer related functions */
238 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_c;
239 :     transfer_16to8copy = transfer_16to8copy_c;
240 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_c;
241 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_c;
242 : suxen_drol 1.11 transfer_8to16sub2 = transfer_8to16sub2_c;
243 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_c;
244 : edgomez 1.16 transfer_16to8add = transfer_16to8add_c;
245 :     transfer8x8_copy = transfer8x8_copy_c;
246 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_c;
247 : Isibaar 1.1
248 : edgomez 1.41 /* Interlacing functions */
249 :     MBFieldTest = MBFieldTest_c;
250 :    
251 : edgomez 1.16 /* Image interpolation related functions */
252 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c;
253 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c;
254 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c;
255 :    
256 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_c;
257 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_c;
258 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_c;
259 :    
260 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c;
261 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c;
262 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c;
263 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c;
264 :    
265 : edgomez 1.41 interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c;
266 :     interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c;
267 :     interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c;
268 :    
269 :     interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c;
270 :     interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c;
271 :     interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c;
272 :    
273 :     interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c;
274 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c;
275 :    
276 :     interpolate8x8_avg2 = interpolate8x8_avg2_c;
277 :     interpolate8x8_avg4 = interpolate8x8_avg4_c;
278 :    
279 : suxen_drol 1.49 /* postprocessing */
280 :     image_brightness = image_brightness_c;
281 :    
282 : edgomez 1.16 /* Initialize internal colorspace transformation tables */
283 : Isibaar 1.1 colorspace_init();
284 :    
285 : edgomez 1.16 /* All colorspace transformation functions User Format->YV12 */
286 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_c;
287 :     rgb555_to_yv12 = rgb555_to_yv12_c;
288 :     rgb565_to_yv12 = rgb565_to_yv12_c;
289 : chl 1.71 rgb_to_yv12 = rgb_to_yv12_c;
290 : edgomez 1.41 bgr_to_yv12 = bgr_to_yv12_c;
291 :     bgra_to_yv12 = bgra_to_yv12_c;
292 :     abgr_to_yv12 = abgr_to_yv12_c;
293 :     rgba_to_yv12 = rgba_to_yv12_c;
294 : edgomez 1.48 argb_to_yv12 = argb_to_yv12_c;
295 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_c;
296 :     uyvy_to_yv12 = uyvy_to_yv12_c;
297 :    
298 :     rgb555i_to_yv12 = rgb555i_to_yv12_c;
299 :     rgb565i_to_yv12 = rgb565i_to_yv12_c;
300 :     bgri_to_yv12 = bgri_to_yv12_c;
301 :     bgrai_to_yv12 = bgrai_to_yv12_c;
302 :     abgri_to_yv12 = abgri_to_yv12_c;
303 :     rgbai_to_yv12 = rgbai_to_yv12_c;
304 : edgomez 1.48 argbi_to_yv12 = argbi_to_yv12_c;
305 : edgomez 1.41 yuyvi_to_yv12 = yuyvi_to_yv12_c;
306 :     uyvyi_to_yv12 = uyvyi_to_yv12_c;
307 :    
308 : edgomez 1.16 /* All colorspace transformation functions YV12->User format */
309 : edgomez 1.41 yv12_to_rgb555 = yv12_to_rgb555_c;
310 :     yv12_to_rgb565 = yv12_to_rgb565_c;
311 : chl 1.71 yv12_to_rgb = yv12_to_rgb_c;
312 : edgomez 1.41 yv12_to_bgr = yv12_to_bgr_c;
313 :     yv12_to_bgra = yv12_to_bgra_c;
314 :     yv12_to_abgr = yv12_to_abgr_c;
315 :     yv12_to_rgba = yv12_to_rgba_c;
316 : edgomez 1.48 yv12_to_argb = yv12_to_argb_c;
317 : edgomez 1.41 yv12_to_yuyv = yv12_to_yuyv_c;
318 :     yv12_to_uyvy = yv12_to_uyvy_c;
319 : edgomez 1.48
320 : edgomez 1.41 yv12_to_rgb555i = yv12_to_rgb555i_c;
321 :     yv12_to_rgb565i = yv12_to_rgb565i_c;
322 :     yv12_to_bgri = yv12_to_bgri_c;
323 :     yv12_to_bgrai = yv12_to_bgrai_c;
324 :     yv12_to_abgri = yv12_to_abgri_c;
325 :     yv12_to_rgbai = yv12_to_rgbai_c;
326 : edgomez 1.48 yv12_to_argbi = yv12_to_argbi_c;
327 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_c;
328 :     yv12_to_uyvyi = yv12_to_uyvyi_c;
329 : Isibaar 1.1
330 : edgomez 1.16 /* Functions used in motion estimation algorithms */
331 : edgomez 1.48 calc_cbp = calc_cbp_c;
332 :     sad16 = sad16_c;
333 :     sad8 = sad8_c;
334 :     sad16bi = sad16bi_c;
335 :     sad8bi = sad8bi_c;
336 :     dev16 = dev16_c;
337 :     sad16v = sad16v_c;
338 :     sse8_16bit = sse8_16bit_c;
339 : edgomez 1.52 sse8_8bit = sse8_8bit_c;
340 : Isibaar 1.1
341 : Skal 1.69 init_GMC(cpu_flags);
342 :    
343 : Isibaar 1.75 #if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
344 : edgomez 1.41
345 :     if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) ||
346 :     (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
347 : Isibaar 1.74 (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2) ||
348 :     (cpu_flags & XVID_CPU_SSE3) || (cpu_flags & XVID_CPU_SSE41))
349 : edgomez 1.41 {
350 :     /* Restore FPU context : emms_c is a nop functions */
351 :     emms = emms_mmx;
352 :     }
353 :    
354 :     if ((cpu_flags & XVID_CPU_MMX)) {
355 : edgomez 1.16
356 :     /* Forward and Inverse Discrete Cosine Transformation functions */
357 : edgomez 1.48 fdct = fdct_mmx_skal;
358 : Isibaar 1.47 idct = idct_mmx;
359 : Isibaar 1.1
360 : edgomez 1.48 /* Qpel stuff */
361 :     xvid_QP_Funcs = &xvid_QP_Funcs_mmx;
362 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx;
363 :    
364 : edgomez 1.16 /* Quantization related functions */
365 : edgomez 1.48 quant_h263_intra = quant_h263_intra_mmx;
366 : Isibaar 1.75 quant_h263_inter = quant_h263_inter_mmx;
367 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_mmx;
368 :     dequant_h263_inter = dequant_h263_inter_mmx;
369 :     quant_mpeg_intra = quant_mpeg_intra_mmx;
370 :     quant_mpeg_inter = quant_mpeg_inter_mmx;
371 :     dequant_mpeg_intra = dequant_mpeg_intra_mmx;
372 :     dequant_mpeg_inter = dequant_mpeg_inter_mmx;
373 : Isibaar 1.1
374 : Isibaar 1.75
375 : edgomez 1.16 /* Block related functions */
376 : Isibaar 1.1 transfer_8to16copy = transfer_8to16copy_mmx;
377 :     transfer_16to8copy = transfer_16to8copy_mmx;
378 : edgomez 1.16 transfer_8to16sub = transfer_8to16sub_mmx;
379 : edgomez 1.41 transfer_8to16subro = transfer_8to16subro_mmx;
380 : edgomez 1.22 transfer_8to16sub2 = transfer_8to16sub2_mmx;
381 : edgomez 1.16 transfer_16to8add = transfer_16to8add_mmx;
382 :     transfer8x8_copy = transfer8x8_copy_mmx;
383 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_mmx;
384 : edgomez 1.22
385 : edgomez 1.41 /* Interlacing Functions */
386 :     MBFieldTest = MBFieldTest_mmx;
387 : edgomez 1.16
388 :     /* Image Interpolation related functions */
389 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx;
390 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx;
391 : Isibaar 1.1 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx;
392 :    
393 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_mmx;
394 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_mmx;
395 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_mmx;
396 :    
397 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx;
398 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx;
399 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx;
400 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx;
401 :    
402 : edgomez 1.41 interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx;
403 :     interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx;
404 :    
405 :     interpolate8x8_avg2 = interpolate8x8_avg2_mmx;
406 :     interpolate8x8_avg4 = interpolate8x8_avg4_mmx;
407 : suxen_drol 1.49
408 :     /* postprocessing */
409 :     image_brightness = image_brightness_mmx;
410 : edgomez 1.41
411 :     /* image input xxx_to_yv12 related functions */
412 : Isibaar 1.75
413 : edgomez 1.41 yv12_to_yv12 = yv12_to_yv12_mmx;
414 : Isibaar 1.75
415 : edgomez 1.41 bgr_to_yv12 = bgr_to_yv12_mmx;
416 : chl 1.72 rgb_to_yv12 = rgb_to_yv12_mmx;
417 : edgomez 1.41 bgra_to_yv12 = bgra_to_yv12_mmx;
418 : chl 1.72 rgba_to_yv12 = rgba_to_yv12_mmx;
419 : edgomez 1.16 yuyv_to_yv12 = yuyv_to_yv12_mmx;
420 :     uyvy_to_yv12 = uyvy_to_yv12_mmx;
421 : Isibaar 1.1
422 : edgomez 1.41 /* image output yv12_to_xxx related functions */
423 :     yv12_to_bgr = yv12_to_bgr_mmx;
424 :     yv12_to_bgra = yv12_to_bgra_mmx;
425 : edgomez 1.16 yv12_to_yuyv = yv12_to_yuyv_mmx;
426 :     yv12_to_uyvy = yv12_to_uyvy_mmx;
427 : edgomez 1.48
428 : edgomez 1.41 yv12_to_yuyvi = yv12_to_yuyvi_mmx;
429 :     yv12_to_uyvyi = yv12_to_uyvyi_mmx;
430 : Isibaar 1.1
431 : edgomez 1.16 /* Motion estimation related functions */
432 : edgomez 1.52 calc_cbp = calc_cbp_mmx;
433 :     sad16 = sad16_mmx;
434 :     sad8 = sad8_mmx;
435 :     sad16bi = sad16bi_mmx;
436 :     sad8bi = sad8bi_mmx;
437 :     dev16 = dev16_mmx;
438 :     sad16v = sad16v_mmx;
439 : edgomez 1.48 sse8_16bit = sse8_16bit_mmx;
440 : edgomez 1.52 sse8_8bit = sse8_8bit_mmx;
441 : Isibaar 1.1 }
442 :    
443 : suxen_drol 1.33 /* these 3dnow functions are faster than mmx, but slower than xmm. */
444 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
445 :    
446 :     emms = emms_3dn;
447 : suxen_drol 1.33
448 :     /* ME functions */
449 :     sad16bi = sad16bi_3dn;
450 :     sad8bi = sad8bi_3dn;
451 : edgomez 1.41
452 :     yuyv_to_yv12 = yuyv_to_yv12_3dn;
453 :     uyvy_to_yv12 = uyvy_to_yv12_3dn;
454 : Isibaar 1.75
455 : suxen_drol 1.33 }
456 :    
457 :    
458 : edgomez 1.41 if ((cpu_flags & XVID_CPU_MMXEXT)) {
459 : edgomez 1.16
460 : edgomez 1.48 /* DCT */
461 :     fdct = fdct_xmm_skal;
462 : Isibaar 1.47 idct = idct_xmm;
463 : edgomez 1.16
464 :     /* Interpolation */
465 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm;
466 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm;
467 : h 1.3 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm;
468 : edgomez 1.55
469 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_xmm;
470 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_xmm;
471 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_xmm;
472 :    
473 : edgomez 1.55 interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm;
474 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm;
475 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm;
476 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm;
477 : Isibaar 1.25
478 : Isibaar 1.75 /* Quantization */
479 : edgomez 1.48 quant_mpeg_inter = quant_mpeg_inter_xmm;
480 : edgomez 1.41
481 : edgomez 1.48 dequant_h263_intra = dequant_h263_intra_xmm;
482 :     dequant_h263_inter = dequant_h263_inter_xmm;
483 : chenm001 1.29
484 : Isibaar 1.75 /* Buffer transfer */
485 : edgomez 1.19 transfer_8to16sub2 = transfer_8to16sub2_xmm;
486 : syskin 1.59 transfer_8to16sub2ro = transfer_8to16sub2ro_xmm;
487 : edgomez 1.16
488 :     /* Colorspace transformation */
489 : Isibaar 1.75 /* yv12_to_yv12 = yv12_to_yv12_xmm; */ /* appears to be slow on many machines */
490 : edgomez 1.41 yuyv_to_yv12 = yuyv_to_yv12_xmm;
491 :     uyvy_to_yv12 = uyvy_to_yv12_xmm;
492 : Isibaar 1.1
493 : edgomez 1.16 /* ME functions */
494 : Isibaar 1.1 sad16 = sad16_xmm;
495 : suxen_drol 1.33 sad8 = sad8_xmm;
496 : chenm001 1.29 sad16bi = sad16bi_xmm;
497 : suxen_drol 1.33 sad8bi = sad8bi_xmm;
498 : Isibaar 1.1 dev16 = dev16_xmm;
499 : edgomez 1.41 sad16v = sad16v_xmm;
500 : Isibaar 1.1 }
501 :    
502 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOW)) {
503 : edgomez 1.16
504 :     /* Interpolation */
505 : Isibaar 1.1 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn;
506 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn;
507 : h 1.4 interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn;
508 : suxen_drol 1.64
509 :     interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dn;
510 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dn;
511 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dn;
512 : Isibaar 1.1 }
513 :    
514 : edgomez 1.41 if ((cpu_flags & XVID_CPU_3DNOWEXT)) {
515 :    
516 :     /* Buffer transfer */
517 :     transfer_8to16copy = transfer_8to16copy_3dne;
518 :     transfer_16to8copy = transfer_16to8copy_3dne;
519 :     transfer_8to16sub = transfer_8to16sub_3dne;
520 :     transfer_8to16subro = transfer_8to16subro_3dne;
521 :     transfer_16to8add = transfer_16to8add_3dne;
522 :     transfer8x8_copy = transfer8x8_copy_3dne;
523 : suxen_drol 1.64 transfer8x4_copy = transfer8x4_copy_3dne;
524 : edgomez 1.41
525 : edgomez 1.51 if ((cpu_flags & XVID_CPU_MMXEXT)) {
526 :     /* Inverse DCT */
527 :     idct = idct_3dne;
528 :    
529 :     /* Buffer transfer */
530 :     transfer_8to16sub2 = transfer_8to16sub2_3dne;
531 :    
532 :     /* Interpolation */
533 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne;
534 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne;
535 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne;
536 :    
537 : suxen_drol 1.64 interpolate8x4_halfpel_h = interpolate8x4_halfpel_h_3dne;
538 :     interpolate8x4_halfpel_v = interpolate8x4_halfpel_v_3dne;
539 :     interpolate8x4_halfpel_hv = interpolate8x4_halfpel_hv_3dne;
540 :    
541 : Isibaar 1.75 /* Quantization */
542 : edgomez 1.51 quant_h263_intra = quant_h263_intra_3dne; /* cmov only */
543 :     quant_h263_inter = quant_h263_inter_3dne;
544 :     dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */
545 :     dequant_mpeg_inter = dequant_mpeg_inter_3dne;
546 :     dequant_h263_intra = dequant_h263_intra_3dne;
547 :     dequant_h263_inter = dequant_h263_inter_3dne;
548 :    
549 : Isibaar 1.75 /* ME functions */
550 : edgomez 1.51 sad16 = sad16_3dne;
551 :     sad8 = sad8_3dne;
552 :     sad16bi = sad16bi_3dne;
553 :     sad8bi = sad8bi_3dne;
554 :     dev16 = dev16_3dne;
555 :     }
556 : edgomez 1.41 }
557 : Isibaar 1.75
558 : edgomez 1.41 if ((cpu_flags & XVID_CPU_SSE2)) {
559 : edgomez 1.16
560 : chenm001 1.29 calc_cbp = calc_cbp_sse2;
561 :    
562 : edgomez 1.16 /* Quantization */
563 : edgomez 1.48 quant_h263_intra = quant_h263_intra_sse2;
564 :     quant_h263_inter = quant_h263_inter_sse2;
565 :     dequant_h263_intra = dequant_h263_intra_sse2;
566 :     dequant_h263_inter = dequant_h263_inter_sse2;
567 : h 1.13
568 : edgomez 1.48 /* SAD operators */
569 : edgomez 1.16 sad16 = sad16_sse2;
570 :     dev16 = dev16_sse2;
571 : edgomez 1.48
572 : Skal 1.63 /* DCT operators */
573 : edgomez 1.48 fdct = fdct_sse2_skal;
574 : Isibaar 1.76 idct = idct_sse2_skal; /* Is now IEEE1180 and Walken compliant. */
575 : suxen_drol 1.54
576 :     /* postprocessing */
577 :     image_brightness = image_brightness_sse2;
578 : Isibaar 1.75
579 : h 1.12 }
580 : Isibaar 1.73
581 :     if ((cpu_flags & XVID_CPU_SSE3)) {
582 :    
583 :     /* SAD operators */
584 :     sad16 = sad16_sse3;
585 :     dev16 = dev16_sse3;
586 :     }
587 : Isibaar 1.74
588 : edgomez 1.48 #endif /* ARCH_IS_IA32 */
589 : Isibaar 1.18
590 : edgomez 1.41 #if defined(ARCH_IS_IA64)
591 : edgomez 1.43 if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */
592 : Isibaar 1.18 idct_ia64_init();
593 :     fdct = fdct_ia64;
594 : edgomez 1.43 idct = idct_ia64; /*not yet working, crashes */
595 : Isibaar 1.18 interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64;
596 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64;
597 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64;
598 :     sad16 = sad16_ia64;
599 :     sad16bi = sad16bi_ia64;
600 :     sad8 = sad8_ia64;
601 :     dev16 = dev16_ia64;
602 : edgomez 1.43 /* Halfpel8_Refine = Halfpel8_Refine_ia64; */
603 : edgomez 1.48 quant_h263_intra = quant_h263_intra_ia64;
604 :     quant_h263_inter = quant_h263_inter_ia64;
605 :     dequant_h263_intra = dequant_h263_intra_ia64;
606 :     dequant_h263_inter = dequant_h263_inter_ia64;
607 : Isibaar 1.18 transfer_8to16copy = transfer_8to16copy_ia64;
608 :     transfer_16to8copy = transfer_16to8copy_ia64;
609 :     transfer_8to16sub = transfer_8to16sub_ia64;
610 :     transfer_8to16sub2 = transfer_8to16sub2_ia64;
611 :     transfer_16to8add = transfer_16to8add_ia64;
612 :     transfer8x8_copy = transfer8x8_copy_ia64;
613 :     }
614 : edgomez 1.48 #endif
615 : edgomez 1.16
616 : edgomez 1.41 #if defined(ARCH_IS_PPC)
617 : edgomez 1.50 if ((cpu_flags & XVID_CPU_ALTIVEC)) {
618 :     /* sad operators */
619 : edgomez 1.56 sad16 = sad16_altivec_c;
620 :     sad16bi = sad16bi_altivec_c;
621 :     sad8 = sad8_altivec_c;
622 :     dev16 = dev16_altivec_c;
623 : edgomez 1.50
624 :     sse8_16bit = sse8_16bit_altivec_c;
625 :    
626 :     /* mem transfer */
627 :     transfer_8to16copy = transfer_8to16copy_altivec_c;
628 :     transfer_16to8copy = transfer_16to8copy_altivec_c;
629 :     transfer_8to16sub = transfer_8to16sub_altivec_c;
630 :     transfer_8to16subro = transfer_8to16subro_altivec_c;
631 :     transfer_8to16sub2 = transfer_8to16sub2_altivec_c;
632 :     transfer_16to8add = transfer_16to8add_altivec_c;
633 :     transfer8x8_copy = transfer8x8_copy_altivec_c;
634 :    
635 :     /* Inverse DCT */
636 :     idct = idct_altivec_c;
637 :    
638 :     /* Interpolation */
639 :     interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c;
640 :     interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c;
641 :     interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c;
642 : edgomez 1.58
643 : edgomez 1.50 interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c;
644 :     interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c;
645 : edgomez 1.56
646 :     interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c;
647 :     interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c;
648 :     interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c;
649 :     interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c;
650 : edgomez 1.50
651 :     /* Colorspace conversion */
652 :     bgra_to_yv12 = bgra_to_yv12_altivec_c;
653 :     abgr_to_yv12 = abgr_to_yv12_altivec_c;
654 :     rgba_to_yv12 = rgba_to_yv12_altivec_c;
655 :     argb_to_yv12 = argb_to_yv12_altivec_c;
656 :    
657 :     yuyv_to_yv12 = yuyv_to_yv12_altivec_c;
658 :     uyvy_to_yv12 = uyvy_to_yv12_altivec_c;
659 :    
660 :     yv12_to_yuyv = yv12_to_yuyv_altivec_c;
661 :     yv12_to_uyvy = yv12_to_uyvy_altivec_c;
662 :    
663 :     /* Quantization */
664 :     quant_h263_intra = quant_h263_intra_altivec_c;
665 :     quant_h263_inter = quant_h263_inter_altivec_c;
666 :     dequant_h263_intra = dequant_h263_intra_altivec_c;
667 :     dequant_h263_inter = dequant_h263_inter_altivec_c;
668 : edgomez 1.62
669 :     dequant_mpeg_intra = dequant_mpeg_intra_altivec_c;
670 :     dequant_mpeg_inter = dequant_mpeg_inter_altivec_c;
671 : edgomez 1.56
672 :     /* Qpel stuff */
673 :     xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C;
674 :     xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C;
675 : edgomez 1.50 }
676 : canard 1.6 #endif
677 : edgomez 1.41
678 : edgomez 1.48 #if defined(_DEBUG)
679 :     xvid_debug = init->debug;
680 :     #endif
681 :    
682 : edgomez 1.50 return(0);
683 : edgomez 1.41 }
684 :    
685 :    
686 :     static int
687 : edgomez 1.48 xvid_gbl_info(xvid_gbl_info_t * info)
688 : edgomez 1.41 {
689 : edgomez 1.48 if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */
690 :     return XVID_ERR_VERSION;
691 : edgomez 1.41
692 : edgomez 1.48 info->actual_version = XVID_VERSION;
693 : Isibaar 1.78 info->build = "xvid-1.3.0-dev";
694 : edgomez 1.48 info->cpu_flags = detect_cpu_flags();
695 : Isibaar 1.81 info->num_threads = 0; /* single-thread */
696 : edgomez 1.41
697 : Isibaar 1.75 #if defined(_WIN32)
698 : Isibaar 1.79
699 : suxen_drol 1.68 {
700 : Isibaar 1.79 SYSTEM_INFO siSysInfo;
701 :     GetSystemInfo(&siSysInfo);
702 :     info->num_threads = siSysInfo.dwNumberOfProcessors; /* number of _logical_ cores */
703 : suxen_drol 1.68 }
704 : Isibaar 1.79
705 : Isibaar 1.81 #elif defined(_SC_NPROCESSORS_CONF) /* should be available on Apple too actually */
706 : Isibaar 1.75
707 :     info->num_threads = sysconf(_SC_NPROCESSORS_CONF);
708 :    
709 : Isibaar 1.81 #elif defined(__APPLE__) && defined(__MACH__)
710 :    
711 :     {
712 :     size_t len;
713 :     int mib[2], ncpu;
714 :    
715 :     mib[0] = CTL_HW;
716 :     mib[1] = HW_NCPU;
717 :     len = sizeof(ncpu);
718 :     if (sysctl(mib, 2, &ncpu, &len, NULL, 0) == 0)
719 :     info -> num_threads = ncpu;
720 :     else
721 :     info -> num_threads = 1;
722 :     }
723 :    
724 : Isibaar 1.82 #elif defined(__amigaos4__)
725 :    
726 :     {
727 :     uint32_t num_threads = 1;
728 :     IExec->GetCPUInfoTags(GCIT_NumberOfCPUs, &num_threads, TAG_END);
729 :     info->num_threads = num_threads;
730 :     }
731 :    
732 : edgomez 1.48 #endif
733 : edgomez 1.41
734 :     return 0;
735 :     }
736 :    
737 :    
738 : edgomez 1.48 static int
739 :     xvid_gbl_convert(xvid_gbl_convert_t* convert)
740 : edgomez 1.41 {
741 : edgomez 1.48 int width;
742 :     int height;
743 :     int width2;
744 :     int height2;
745 :     IMAGE img;
746 : edgomez 1.41
747 : edgomez 1.48 if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */
748 :     return XVID_ERR_VERSION;
749 : edgomez 1.41
750 : edgomez 1.48 #if 0
751 :     const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP);
752 :     #endif
753 :     width = convert->width;
754 :     height = convert->height;
755 :     width2 = convert->width/2;
756 :     height2 = convert->height/2;
757 : edgomez 1.41
758 : edgomez 1.48 switch (convert->input.csp & ~XVID_CSP_VFLIP)
759 : edgomez 1.41 {
760 : edgomez 1.48 case XVID_CSP_YV12 :
761 :     img.y = convert->input.plane[0];
762 :     img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height;
763 :     img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2;
764 :     image_output(&img, width, height, width,
765 :     (uint8_t**)convert->output.plane, convert->output.stride,
766 :     convert->output.csp, convert->interlacing);
767 :     break;
768 : edgomez 1.41
769 : edgomez 1.48 default :
770 :     return XVID_ERR_FORMAT;
771 : edgomez 1.41 }
772 :    
773 :    
774 : edgomez 1.48 emms();
775 : edgomez 1.41 return 0;
776 :     }
777 :    
778 : edgomez 1.48 /*****************************************************************************
779 :     * XviD Global Entry point
780 :     *
781 :     * Well this function initialize all internal function pointers according
782 :     * to the CPU features forced by the library client or autodetected (depending
783 :     * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all
784 :     * image colorspace transformation tables.
785 :     *
786 :     ****************************************************************************/
787 : Isibaar 1.1
788 : edgomez 1.41
789 :     int
790 : edgomez 1.48 xvid_global(void *handle,
791 : edgomez 1.41 int opt,
792 :     void *param1,
793 :     void *param2)
794 :     {
795 :     switch(opt)
796 :     {
797 : edgomez 1.48 case XVID_GBL_INIT :
798 :     return xvid_gbl_init((xvid_gbl_init_t*)param1);
799 : edgomez 1.41
800 : edgomez 1.48 case XVID_GBL_INFO :
801 :     return xvid_gbl_info((xvid_gbl_info_t*)param1);
802 :    
803 :     case XVID_GBL_CONVERT :
804 :     return xvid_gbl_convert((xvid_gbl_convert_t*)param1);
805 : edgomez 1.41
806 :     default :
807 :     return XVID_ERR_FAIL;
808 :     }
809 :     }
810 :    
811 : edgomez 1.16 /*****************************************************************************
812 :     * XviD Native decoder entry point
813 :     *
814 :     * This function is just a wrapper to all the option cases.
815 :     *
816 :     * Returned values : XVID_ERR_FAIL when opt is invalid
817 :     * else returns the wrapped function result
818 :     *
819 :     ****************************************************************************/
820 :    
821 : edgomez 1.15 int
822 :     xvid_decore(void *handle,
823 :     int opt,
824 :     void *param1,
825 :     void *param2)
826 : Isibaar 1.1 {
827 : edgomez 1.15 switch (opt) {
828 :     case XVID_DEC_CREATE:
829 : edgomez 1.48 return decoder_create((xvid_dec_create_t *) param1);
830 : edgomez 1.15
831 :     case XVID_DEC_DESTROY:
832 :     return decoder_destroy((DECODER *) handle);
833 : Isibaar 1.1
834 : edgomez 1.48 case XVID_DEC_DECODE:
835 :     return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2);
836 :    
837 : Isibaar 1.1 default:
838 : edgomez 1.15 return XVID_ERR_FAIL;
839 :     }
840 : Isibaar 1.1 }
841 :    
842 : edgomez 1.16
843 :     /*****************************************************************************
844 :     * XviD Native encoder entry point
845 :     *
846 :     * This function is just a wrapper to all the option cases.
847 :     *
848 :     * Returned values : XVID_ERR_FAIL when opt is invalid
849 :     * else returns the wrapped function result
850 :     *
851 :     ****************************************************************************/
852 : Isibaar 1.1
853 : edgomez 1.15 int
854 :     xvid_encore(void *handle,
855 :     int opt,
856 :     void *param1,
857 :     void *param2)
858 : Isibaar 1.1 {
859 : edgomez 1.15 switch (opt) {
860 :     case XVID_ENC_ENCODE:
861 : edgomez 1.41
862 : edgomez 1.48 return enc_encode((Encoder *) handle,
863 :     (xvid_enc_frame_t *) param1,
864 :     (xvid_enc_stats_t *) param2);
865 : edgomez 1.15
866 :     case XVID_ENC_CREATE:
867 : edgomez 1.48 return enc_create((xvid_enc_create_t *) param1);
868 : edgomez 1.15
869 :     case XVID_ENC_DESTROY:
870 : edgomez 1.48 return enc_destroy((Encoder *) handle);
871 : Isibaar 1.1
872 :     default:
873 : edgomez 1.15 return XVID_ERR_FAIL;
874 :     }
875 : Isibaar 1.1 }

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