Parent Directory | Revision Log
Revision 1270 - (view) (download)
1 : | edgomez | 200 | /***************************************************************************** |
2 : | edgomez | 201 | * |
3 : | * XVID MPEG-4 VIDEO CODEC | ||
4 : | * - Native API implementation - | ||
5 : | * | ||
6 : | edgomez | 1054 | * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
7 : | * | ||
8 : | edgomez | 851 | * 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 | 201 | * (at your option) any later version. |
12 : | * | ||
13 : | * This program is distributed in the hope that it will be useful, | ||
14 : | edgomez | 851 | * but WITHOUT ANY WARRANTY ; without even the implied warranty of |
15 : | edgomez | 201 | * 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 | 851 | * along with this program ; if not, write to the Free Software |
20 : | edgomez | 201 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
21 : | * | ||
22 : | Isibaar | 1270 | * $Id: xvid.c,v 1.45.2.25 2003-12-17 17:07:38 Isibaar Exp $ |
23 : | edgomez | 410 | * |
24 : | edgomez | 201 | ****************************************************************************/ |
25 : | chenm001 | 274 | |
26 : | edgomez | 851 | #include <stdio.h> |
27 : | #include <stdlib.h> | ||
28 : | #include <string.h> | ||
29 : | #include <time.h> | ||
30 : | |||
31 : | Isibaar | 3 | #include "xvid.h" |
32 : | #include "decoder.h" | ||
33 : | #include "encoder.h" | ||
34 : | #include "bitstream/cbp.h" | ||
35 : | #include "dct/idct.h" | ||
36 : | #include "dct/fdct.h" | ||
37 : | #include "image/colorspace.h" | ||
38 : | #include "image/interpolate8x8.h" | ||
39 : | edgomez | 851 | #include "image/reduced.h" |
40 : | Isibaar | 3 | #include "utils/mem_transfer.h" |
41 : | edgomez | 851 | #include "utils/mbfunctions.h" |
42 : | edgomez | 1174 | #include "quant/quant.h" |
43 : | ia64p | 299 | #include "motion/motion.h" |
44 : | Isibaar | 3 | #include "motion/sad.h" |
45 : | #include "utils/emms.h" | ||
46 : | #include "utils/timer.h" | ||
47 : | Isibaar | 100 | #include "bitstream/mbcoding.h" |
48 : | Isibaar | 1125 | #include "image/qpel.h" |
49 : | Isibaar | 1256 | #include "image/postprocessing.h" |
50 : | Isibaar | 3 | |
51 : | suxen_drol | 1031 | #if defined(_DEBUG) |
52 : | unsigned int xvid_debug = 0; /* xvid debug mask */ | ||
53 : | #endif | ||
54 : | |||
55 : | edgomez | 851 | #if defined(ARCH_IS_IA32) |
56 : | #if defined(_MSC_VER) | ||
57 : | # include <windows.h> | ||
58 : | suxen_drol | 311 | #else |
59 : | edgomez | 851 | # include <signal.h> |
60 : | # include <setjmp.h> | ||
61 : | suxen_drol | 311 | |
62 : | edgomez | 851 | static jmp_buf mark; |
63 : | suxen_drol | 311 | |
64 : | edgomez | 851 | static void |
65 : | sigill_handler(int signal) | ||
66 : | { | ||
67 : | longjmp(mark, 1); | ||
68 : | } | ||
69 : | suxen_drol | 311 | #endif |
70 : | |||
71 : | |||
72 : | /* | ||
73 : | edgomez | 874 | * Calls the funcptr, and returns whether SIGILL (illegal instruction) was |
74 : | * signalled | ||
75 : | * | ||
76 : | * Return values: | ||
77 : | * -1 : could not determine | ||
78 : | * 0 : SIGILL was *not* signalled | ||
79 : | * 1 : SIGILL was signalled | ||
80 : | */ | ||
81 : | suxen_drol | 311 | |
82 : | int | ||
83 : | sigill_check(void (*func)()) | ||
84 : | { | ||
85 : | edgomez | 851 | #if defined(_MSC_VER) |
86 : | suxen_drol | 311 | _try { |
87 : | func(); | ||
88 : | } | ||
89 : | _except(EXCEPTION_EXECUTE_HANDLER) { | ||
90 : | |||
91 : | if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION) | ||
92 : | return 1; | ||
93 : | } | ||
94 : | return 0; | ||
95 : | #else | ||
96 : | void * old_handler; | ||
97 : | int jmpret; | ||
98 : | |||
99 : | |||
100 : | old_handler = signal(SIGILL, sigill_handler); | ||
101 : | if (old_handler == SIG_ERR) | ||
102 : | { | ||
103 : | return -1; | ||
104 : | } | ||
105 : | |||
106 : | jmpret = setjmp(mark); | ||
107 : | if (jmpret == 0) | ||
108 : | { | ||
109 : | func(); | ||
110 : | } | ||
111 : | |||
112 : | signal(SIGILL, old_handler); | ||
113 : | |||
114 : | return jmpret; | ||
115 : | #endif | ||
116 : | } | ||
117 : | #endif | ||
118 : | |||
119 : | edgomez | 851 | |
120 : | /* detect cpu flags */ | ||
121 : | static unsigned int | ||
122 : | detect_cpu_flags() | ||
123 : | { | ||
124 : | /* enable native assembly optimizations by default */ | ||
125 : | unsigned int cpu_flags = XVID_CPU_ASM; | ||
126 : | |||
127 : | #if defined(ARCH_IS_IA32) | ||
128 : | cpu_flags |= check_cpu_features(); | ||
129 : | if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger)) | ||
130 : | cpu_flags &= ~XVID_CPU_SSE; | ||
131 : | |||
132 : | if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger)) | ||
133 : | cpu_flags &= ~XVID_CPU_SSE2; | ||
134 : | #endif | ||
135 : | |||
136 : | #if defined(ARCH_IS_PPC) | ||
137 : | #if defined(ARCH_IS_PPC_ALTIVEC) | ||
138 : | cpu_flags |= XVID_CPU_ALTIVEC; | ||
139 : | #endif | ||
140 : | #endif | ||
141 : | |||
142 : | return cpu_flags; | ||
143 : | } | ||
144 : | |||
145 : | |||
146 : | edgomez | 200 | /***************************************************************************** |
147 : | * XviD Init Entry point | ||
148 : | * | ||
149 : | * Well this function initialize all internal function pointers according | ||
150 : | * to the CPU features forced by the library client or autodetected (depending | ||
151 : | * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all | ||
152 : | * image colorspace transformation tables. | ||
153 : | edgomez | 1161 | * |
154 : | edgomez | 200 | * Returned value : XVID_ERR_OK |
155 : | * + API_VERSION in the input XVID_INIT_PARAM structure | ||
156 : | * + core build " " " " " | ||
157 : | * | ||
158 : | ****************************************************************************/ | ||
159 : | |||
160 : | edgomez | 851 | |
161 : | edgomez | 1161 | static |
162 : | suxen_drol | 890 | int xvid_gbl_init(xvid_gbl_init_t * init) |
163 : | Isibaar | 3 | { |
164 : | suxen_drol | 890 | unsigned int cpu_flags; |
165 : | Isibaar | 3 | |
166 : | edgomez | 1107 | if (XVID_VERSION_MAJOR(init->version) != 1) /* v1.x.x */ |
167 : | suxen_drol | 890 | return XVID_ERR_VERSION; |
168 : | suxen_drol | 234 | |
169 : | suxen_drol | 890 | cpu_flags = (init->cpu_flags & XVID_CPU_FORCE) ? init->cpu_flags : detect_cpu_flags(); |
170 : | suxen_drol | 234 | |
171 : | edgomez | 200 | /* Initialize the function pointers */ |
172 : | Isibaar | 3 | idct_int32_init(); |
173 : | Isibaar | 100 | init_vlc_tables(); |
174 : | |||
175 : | edgomez | 200 | /* Fixed Point Forward/Inverse DCT transformations */ |
176 : | Isibaar | 3 | fdct = fdct_int32; |
177 : | idct = idct_int32; | ||
178 : | |||
179 : | edgomez | 200 | /* Only needed on PPC Altivec archs */ |
180 : | canard | 115 | sadInit = 0; |
181 : | edgomez | 195 | |
182 : | edgomez | 200 | /* Restore FPU context : emms_c is a nop functions */ |
183 : | Isibaar | 3 | emms = emms_c; |
184 : | |||
185 : | Isibaar | 1125 | /* Qpel stuff */ |
186 : | xvid_QP_Funcs = &xvid_QP_Funcs_C; | ||
187 : | xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_C; | ||
188 : | edgomez | 1127 | xvid_Init_QP(); |
189 : | Isibaar | 1125 | |
190 : | edgomez | 200 | /* Quantization functions */ |
191 : | edgomez | 1174 | quant_h263_intra = quant_h263_intra_c; |
192 : | quant_h263_inter = quant_h263_inter_c; | ||
193 : | dequant_h263_intra = dequant_h263_intra_c; | ||
194 : | dequant_h263_inter = dequant_h263_inter_c; | ||
195 : | Isibaar | 3 | |
196 : | edgomez | 1174 | quant_mpeg_intra = quant_mpeg_intra_c; |
197 : | quant_mpeg_inter = quant_mpeg_inter_c; | ||
198 : | dequant_mpeg_intra = dequant_mpeg_intra_c; | ||
199 : | dequant_mpeg_inter = dequant_mpeg_inter_c; | ||
200 : | Isibaar | 3 | |
201 : | edgomez | 200 | /* Block transfer related functions */ |
202 : | Isibaar | 3 | transfer_8to16copy = transfer_8to16copy_c; |
203 : | transfer_16to8copy = transfer_16to8copy_c; | ||
204 : | edgomez | 200 | transfer_8to16sub = transfer_8to16sub_c; |
205 : | edgomez | 851 | transfer_8to16subro = transfer_8to16subro_c; |
206 : | suxen_drol | 118 | transfer_8to16sub2 = transfer_8to16sub2_c; |
207 : | edgomez | 200 | transfer_16to8add = transfer_16to8add_c; |
208 : | transfer8x8_copy = transfer8x8_copy_c; | ||
209 : | Isibaar | 3 | |
210 : | edgomez | 851 | /* Interlacing functions */ |
211 : | MBFieldTest = MBFieldTest_c; | ||
212 : | |||
213 : | edgomez | 200 | /* Image interpolation related functions */ |
214 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c; | ||
215 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; | ||
216 : | Isibaar | 3 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
217 : | |||
218 : | edgomez | 851 | interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c; |
219 : | interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c; | ||
220 : | interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c; | ||
221 : | |||
222 : | interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c; | ||
223 : | interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c; | ||
224 : | interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c; | ||
225 : | |||
226 : | interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c; | ||
227 : | interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c; | ||
228 : | |||
229 : | interpolate8x8_avg2 = interpolate8x8_avg2_c; | ||
230 : | interpolate8x8_avg4 = interpolate8x8_avg4_c; | ||
231 : | |||
232 : | edgomez | 1174 | /* reduced resolution */ |
233 : | edgomez | 851 | copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_C; |
234 : | add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_C; | ||
235 : | vfilter_31 = xvid_VFilter_31_C; | ||
236 : | hfilter_31 = xvid_HFilter_31_C; | ||
237 : | filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_C; | ||
238 : | filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_C; | ||
239 : | |||
240 : | edgomez | 200 | /* Initialize internal colorspace transformation tables */ |
241 : | Isibaar | 3 | colorspace_init(); |
242 : | |||
243 : | edgomez | 200 | /* All colorspace transformation functions User Format->YV12 */ |
244 : | edgomez | 851 | yv12_to_yv12 = yv12_to_yv12_c; |
245 : | rgb555_to_yv12 = rgb555_to_yv12_c; | ||
246 : | rgb565_to_yv12 = rgb565_to_yv12_c; | ||
247 : | bgr_to_yv12 = bgr_to_yv12_c; | ||
248 : | bgra_to_yv12 = bgra_to_yv12_c; | ||
249 : | abgr_to_yv12 = abgr_to_yv12_c; | ||
250 : | rgba_to_yv12 = rgba_to_yv12_c; | ||
251 : | yuyv_to_yv12 = yuyv_to_yv12_c; | ||
252 : | uyvy_to_yv12 = uyvy_to_yv12_c; | ||
253 : | Isibaar | 3 | |
254 : | edgomez | 851 | rgb555i_to_yv12 = rgb555i_to_yv12_c; |
255 : | rgb565i_to_yv12 = rgb565i_to_yv12_c; | ||
256 : | bgri_to_yv12 = bgri_to_yv12_c; | ||
257 : | bgrai_to_yv12 = bgrai_to_yv12_c; | ||
258 : | abgri_to_yv12 = abgri_to_yv12_c; | ||
259 : | rgbai_to_yv12 = rgbai_to_yv12_c; | ||
260 : | yuyvi_to_yv12 = yuyvi_to_yv12_c; | ||
261 : | uyvyi_to_yv12 = uyvyi_to_yv12_c; | ||
262 : | |||
263 : | |||
264 : | edgomez | 200 | /* All colorspace transformation functions YV12->User format */ |
265 : | edgomez | 851 | yv12_to_rgb555 = yv12_to_rgb555_c; |
266 : | yv12_to_rgb565 = yv12_to_rgb565_c; | ||
267 : | yv12_to_bgr = yv12_to_bgr_c; | ||
268 : | yv12_to_bgra = yv12_to_bgra_c; | ||
269 : | yv12_to_abgr = yv12_to_abgr_c; | ||
270 : | yv12_to_rgba = yv12_to_rgba_c; | ||
271 : | yv12_to_yuyv = yv12_to_yuyv_c; | ||
272 : | yv12_to_uyvy = yv12_to_uyvy_c; | ||
273 : | edgomez | 1161 | |
274 : | edgomez | 851 | yv12_to_rgb555i = yv12_to_rgb555i_c; |
275 : | yv12_to_rgb565i = yv12_to_rgb565i_c; | ||
276 : | yv12_to_bgri = yv12_to_bgri_c; | ||
277 : | yv12_to_bgrai = yv12_to_bgrai_c; | ||
278 : | yv12_to_abgri = yv12_to_abgri_c; | ||
279 : | yv12_to_rgbai = yv12_to_rgbai_c; | ||
280 : | yv12_to_yuyvi = yv12_to_yuyvi_c; | ||
281 : | yv12_to_uyvyi = yv12_to_uyvyi_c; | ||
282 : | Isibaar | 3 | |
283 : | edgomez | 200 | /* Functions used in motion estimation algorithms */ |
284 : | edgomez | 1205 | calc_cbp = calc_cbp_c; |
285 : | sad16 = sad16_c; | ||
286 : | sad8 = sad8_c; | ||
287 : | sad16bi = sad16bi_c; | ||
288 : | sad8bi = sad8bi_c; | ||
289 : | dev16 = dev16_c; | ||
290 : | sad16v = sad16v_c; | ||
291 : | sse8_16bit = sse8_16bit_c; | ||
292 : | edgomez | 1161 | |
293 : | edgomez | 851 | #if defined(ARCH_IS_IA32) |
294 : | edgomez | 200 | |
295 : | edgomez | 1058 | if ((cpu_flags & XVID_CPU_ASM)) { |
296 : | edgomez | 851 | vfilter_31 = xvid_VFilter_31_x86; |
297 : | hfilter_31 = xvid_HFilter_31_x86; | ||
298 : | } | ||
299 : | |||
300 : | if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || | ||
301 : | (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) || | ||
302 : | (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2)) | ||
303 : | { | ||
304 : | /* Restore FPU context : emms_c is a nop functions */ | ||
305 : | emms = emms_mmx; | ||
306 : | } | ||
307 : | |||
308 : | if ((cpu_flags & XVID_CPU_MMX)) { | ||
309 : | |||
310 : | edgomez | 200 | /* Forward and Inverse Discrete Cosine Transformation functions */ |
311 : | edgomez | 1190 | fdct = fdct_mmx_skal; |
312 : | Isibaar | 1066 | idct = idct_mmx; |
313 : | Isibaar | 3 | |
314 : | Isibaar | 1125 | /* Qpel stuff */ |
315 : | xvid_QP_Funcs = &xvid_QP_Funcs_mmx; | ||
316 : | xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_mmx; | ||
317 : | |||
318 : | edgomez | 200 | /* Quantization related functions */ |
319 : | edgomez | 1174 | quant_h263_intra = quant_h263_intra_mmx; |
320 : | quant_h263_inter = quant_h263_inter_mmx; | ||
321 : | dequant_h263_intra = dequant_h263_intra_mmx; | ||
322 : | dequant_h263_inter = dequant_h263_inter_mmx; | ||
323 : | Isibaar | 3 | |
324 : | edgomez | 1174 | quant_mpeg_intra = quant_mpeg_intra_mmx; |
325 : | quant_mpeg_inter = quant_mpeg_inter_mmx; | ||
326 : | dequant_mpeg_intra = dequant_mpeg_intra_mmx; | ||
327 : | dequant_mpeg_inter = dequant_mpeg_inter_mmx; | ||
328 : | Isibaar | 3 | |
329 : | edgomez | 200 | /* Block related functions */ |
330 : | Isibaar | 3 | transfer_8to16copy = transfer_8to16copy_mmx; |
331 : | transfer_16to8copy = transfer_16to8copy_mmx; | ||
332 : | edgomez | 200 | transfer_8to16sub = transfer_8to16sub_mmx; |
333 : | edgomez | 851 | transfer_8to16subro = transfer_8to16subro_mmx; |
334 : | edgomez | 236 | transfer_8to16sub2 = transfer_8to16sub2_mmx; |
335 : | edgomez | 200 | transfer_16to8add = transfer_16to8add_mmx; |
336 : | transfer8x8_copy = transfer8x8_copy_mmx; | ||
337 : | Isibaar | 3 | |
338 : | edgomez | 851 | /* Interlacing Functions */ |
339 : | MBFieldTest = MBFieldTest_mmx; | ||
340 : | edgomez | 236 | |
341 : | edgomez | 200 | /* Image Interpolation related functions */ |
342 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx; | ||
343 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; | ||
344 : | Isibaar | 3 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
345 : | |||
346 : | edgomez | 851 | interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx; |
347 : | interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx; | ||
348 : | |||
349 : | interpolate8x8_avg2 = interpolate8x8_avg2_mmx; | ||
350 : | interpolate8x8_avg4 = interpolate8x8_avg4_mmx; | ||
351 : | |||
352 : | /* reduced resolution */ | ||
353 : | copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_mmx; | ||
354 : | add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_mmx; | ||
355 : | hfilter_31 = xvid_HFilter_31_mmx; | ||
356 : | filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_mmx; | ||
357 : | filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_mmx; | ||
358 : | |||
359 : | /* image input xxx_to_yv12 related functions */ | ||
360 : | yv12_to_yv12 = yv12_to_yv12_mmx; | ||
361 : | bgr_to_yv12 = bgr_to_yv12_mmx; | ||
362 : | bgra_to_yv12 = bgra_to_yv12_mmx; | ||
363 : | edgomez | 200 | yuyv_to_yv12 = yuyv_to_yv12_mmx; |
364 : | uyvy_to_yv12 = uyvy_to_yv12_mmx; | ||
365 : | Isibaar | 3 | |
366 : | edgomez | 851 | /* image output yv12_to_xxx related functions */ |
367 : | yv12_to_bgr = yv12_to_bgr_mmx; | ||
368 : | yv12_to_bgra = yv12_to_bgra_mmx; | ||
369 : | edgomez | 200 | yv12_to_yuyv = yv12_to_yuyv_mmx; |
370 : | yv12_to_uyvy = yv12_to_uyvy_mmx; | ||
371 : | edgomez | 1161 | |
372 : | edgomez | 851 | yv12_to_yuyvi = yv12_to_yuyvi_mmx; |
373 : | yv12_to_uyvyi = yv12_to_uyvyi_mmx; | ||
374 : | Isibaar | 3 | |
375 : | edgomez | 200 | /* Motion estimation related functions */ |
376 : | Isibaar | 3 | calc_cbp = calc_cbp_mmx; |
377 : | edgomez | 200 | sad16 = sad16_mmx; |
378 : | sad8 = sad8_mmx; | ||
379 : | suxen_drol | 329 | sad16bi = sad16bi_mmx; |
380 : | sad8bi = sad8bi_mmx; | ||
381 : | edgomez | 200 | dev16 = dev16_mmx; |
382 : | edgomez | 851 | sad16v = sad16v_mmx; |
383 : | edgomez | 1205 | sse8_16bit = sse8_16bit_mmx; |
384 : | Isibaar | 3 | } |
385 : | |||
386 : | suxen_drol | 329 | /* these 3dnow functions are faster than mmx, but slower than xmm. */ |
387 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_3DNOW)) { |
388 : | suxen_drol | 329 | |
389 : | edgomez | 851 | emms = emms_3dn; |
390 : | |||
391 : | suxen_drol | 329 | /* ME functions */ |
392 : | sad16bi = sad16bi_3dn; | ||
393 : | sad8bi = sad8bi_3dn; | ||
394 : | edgomez | 851 | |
395 : | yuyv_to_yv12 = yuyv_to_yv12_3dn; | ||
396 : | uyvy_to_yv12 = uyvy_to_yv12_3dn; | ||
397 : | suxen_drol | 329 | } |
398 : | |||
399 : | |||
400 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_MMXEXT)) { |
401 : | edgomez | 200 | |
402 : | edgomez | 1190 | /* DCT */ |
403 : | fdct = fdct_xmm_skal; | ||
404 : | Isibaar | 3 | idct = idct_xmm; |
405 : | edgomez | 200 | |
406 : | /* Interpolation */ | ||
407 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm; | ||
408 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; | ||
409 : | h | 38 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
410 : | edgomez | 200 | |
411 : | edgomez | 851 | /* reduced resolution */ |
412 : | copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm; | ||
413 : | add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm; | ||
414 : | |||
415 : | chenm001 | 274 | /* Quantization */ |
416 : | edgomez | 1174 | quant_mpeg_intra = quant_mpeg_intra_xmm; |
417 : | quant_mpeg_inter = quant_mpeg_inter_xmm; | ||
418 : | edgomez | 851 | |
419 : | edgomez | 1174 | dequant_h263_intra = dequant_h263_intra_xmm; |
420 : | dequant_h263_inter = dequant_h263_inter_xmm; | ||
421 : | chenm001 | 274 | |
422 : | edgomez | 218 | /* Buffer transfer */ |
423 : | transfer_8to16sub2 = transfer_8to16sub2_xmm; | ||
424 : | |||
425 : | edgomez | 200 | /* Colorspace transformation */ |
426 : | edgomez | 851 | yv12_to_yv12 = yv12_to_yv12_xmm; |
427 : | yuyv_to_yv12 = yuyv_to_yv12_xmm; | ||
428 : | uyvy_to_yv12 = uyvy_to_yv12_xmm; | ||
429 : | Isibaar | 3 | |
430 : | edgomez | 200 | /* ME functions */ |
431 : | Isibaar | 3 | sad16 = sad16_xmm; |
432 : | suxen_drol | 329 | sad8 = sad8_xmm; |
433 : | chenm001 | 274 | sad16bi = sad16bi_xmm; |
434 : | suxen_drol | 329 | sad8bi = sad8bi_xmm; |
435 : | Isibaar | 3 | dev16 = dev16_xmm; |
436 : | edgomez | 851 | sad16v = sad16v_xmm; |
437 : | Isibaar | 3 | } |
438 : | |||
439 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_3DNOW)) { |
440 : | edgomez | 200 | |
441 : | /* Interpolation */ | ||
442 : | Isibaar | 3 | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn; |
443 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn; | ||
444 : | h | 40 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn; |
445 : | Isibaar | 3 | } |
446 : | |||
447 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
448 : | edgomez | 200 | |
449 : | edgomez | 851 | /* Inverse DCT */ |
450 : | idct = idct_3dne; | ||
451 : | |||
452 : | /* Buffer transfer */ | ||
453 : | transfer_8to16copy = transfer_8to16copy_3dne; | ||
454 : | transfer_16to8copy = transfer_16to8copy_3dne; | ||
455 : | transfer_8to16sub = transfer_8to16sub_3dne; | ||
456 : | transfer_8to16subro = transfer_8to16subro_3dne; | ||
457 : | transfer_8to16sub2 = transfer_8to16sub2_3dne; | ||
458 : | transfer_16to8add = transfer_16to8add_3dne; | ||
459 : | transfer8x8_copy = transfer8x8_copy_3dne; | ||
460 : | |||
461 : | /* Quantization */ | ||
462 : | edgomez | 1174 | quant_h263_intra = quant_h263_intra_3dne; |
463 : | quant_h263_inter = quant_h263_inter_3dne; | ||
464 : | dequant_mpeg_intra = dequant_mpeg_intra_3dne; | ||
465 : | dequant_mpeg_inter = dequant_mpeg_inter_3dne; | ||
466 : | dequant_h263_intra = dequant_h263_intra_3dne; | ||
467 : | dequant_h263_inter = dequant_h263_inter_3dne; | ||
468 : | edgomez | 851 | |
469 : | /* ME functions */ | ||
470 : | calc_cbp = calc_cbp_3dne; | ||
471 : | sad16 = sad16_3dne; | ||
472 : | sad8 = sad8_3dne; | ||
473 : | sad16bi = sad16bi_3dne; | ||
474 : | sad8bi = sad8bi_3dne; | ||
475 : | dev16 = dev16_3dne; | ||
476 : | |||
477 : | /* Interpolation */ | ||
478 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne; | ||
479 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; | ||
480 : | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; | ||
481 : | } | ||
482 : | |||
483 : | if ((cpu_flags & XVID_CPU_SSE2)) { | ||
484 : | |||
485 : | chenm001 | 274 | calc_cbp = calc_cbp_sse2; |
486 : | |||
487 : | edgomez | 200 | /* Quantization */ |
488 : | edgomez | 1174 | quant_h263_intra = quant_h263_intra_sse2; |
489 : | quant_h263_inter = quant_h263_inter_sse2; | ||
490 : | dequant_h263_intra = dequant_h263_intra_sse2; | ||
491 : | dequant_h263_inter = dequant_h263_inter_sse2; | ||
492 : | h | 135 | |
493 : | edgomez | 1195 | /* SAD operators */ |
494 : | edgomez | 200 | sad16 = sad16_sse2; |
495 : | dev16 = dev16_sse2; | ||
496 : | edgomez | 1195 | |
497 : | edgomez | 1261 | /* DCT operators |
498 : | * no iDCT because it's not "Walken matching" */ | ||
499 : | edgomez | 1195 | fdct = fdct_sse2_skal; |
500 : | h | 126 | } |
501 : | edgomez | 1261 | #endif /* ARCH_IS_IA32 */ |
502 : | edgomez | 200 | |
503 : | edgomez | 851 | #if defined(ARCH_IS_IA64) |
504 : | edgomez | 874 | if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */ |
505 : | Isibaar | 209 | idct_ia64_init(); |
506 : | fdct = fdct_ia64; | ||
507 : | edgomez | 874 | idct = idct_ia64; /*not yet working, crashes */ |
508 : | Isibaar | 209 | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64; |
509 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64; | ||
510 : | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64; | ||
511 : | sad16 = sad16_ia64; | ||
512 : | sad16bi = sad16bi_ia64; | ||
513 : | sad8 = sad8_ia64; | ||
514 : | dev16 = dev16_ia64; | ||
515 : | edgomez | 874 | /* Halfpel8_Refine = Halfpel8_Refine_ia64; */ |
516 : | edgomez | 1174 | quant_h263_intra = quant_h263_intra_ia64; |
517 : | quant_h263_inter = quant_h263_inter_ia64; | ||
518 : | dequant_h263_intra = dequant_h263_intra_ia64; | ||
519 : | dequant_h263_inter = dequant_h263_inter_ia64; | ||
520 : | Isibaar | 209 | transfer_8to16copy = transfer_8to16copy_ia64; |
521 : | transfer_16to8copy = transfer_16to8copy_ia64; | ||
522 : | transfer_8to16sub = transfer_8to16sub_ia64; | ||
523 : | transfer_8to16sub2 = transfer_8to16sub2_ia64; | ||
524 : | transfer_16to8add = transfer_16to8add_ia64; | ||
525 : | transfer8x8_copy = transfer8x8_copy_ia64; | ||
526 : | } | ||
527 : | edgomez | 1161 | #endif |
528 : | Isibaar | 209 | |
529 : | edgomez | 851 | #if defined(ARCH_IS_PPC) |
530 : | if ((cpu_flags & XVID_CPU_ASM)) | ||
531 : | { | ||
532 : | calc_cbp = calc_cbp_ppc; | ||
533 : | } | ||
534 : | |||
535 : | if ((cpu_flags & XVID_CPU_ALTIVEC)) | ||
536 : | { | ||
537 : | calc_cbp = calc_cbp_altivec; | ||
538 : | fdct = fdct_altivec; | ||
539 : | idct = idct_altivec; | ||
540 : | sadInit = sadInit_altivec; | ||
541 : | sad16 = sad16_altivec; | ||
542 : | sad8 = sad8_altivec; | ||
543 : | dev16 = dev16_altivec; | ||
544 : | } | ||
545 : | canard | 52 | #endif |
546 : | edgomez | 851 | |
547 : | suxen_drol | 1031 | #if defined(_DEBUG) |
548 : | xvid_debug = init->debug; | ||
549 : | #endif | ||
550 : | |||
551 : | return 0; | ||
552 : | edgomez | 851 | } |
553 : | |||
554 : | |||
555 : | suxen_drol | 890 | static int |
556 : | xvid_gbl_info(xvid_gbl_info_t * info) | ||
557 : | { | ||
558 : | edgomez | 1107 | if (XVID_VERSION_MAJOR(info->version) != 1) /* v1.x.x */ |
559 : | suxen_drol | 890 | return XVID_ERR_VERSION; |
560 : | edgomez | 851 | |
561 : | suxen_drol | 890 | info->actual_version = XVID_VERSION; |
562 : | edgomez | 1226 | info->build = "xvid-1.0.0"; |
563 : | suxen_drol | 890 | info->cpu_flags = detect_cpu_flags(); |
564 : | |||
565 : | #if defined(_SMP) && defined(WIN32) | ||
566 : | info->num_threads = pthread_num_processors_np();; | ||
567 : | #else | ||
568 : | info->num_threads = 0; | ||
569 : | #endif | ||
570 : | |||
571 : | return 0; | ||
572 : | } | ||
573 : | |||
574 : | |||
575 : | edgomez | 851 | static int |
576 : | suxen_drol | 890 | xvid_gbl_convert(xvid_gbl_convert_t* convert) |
577 : | edgomez | 851 | { |
578 : | suxen_drol | 890 | int width; |
579 : | int height; | ||
580 : | int width2; | ||
581 : | int height2; | ||
582 : | edgomez | 851 | IMAGE img; |
583 : | |||
584 : | edgomez | 1107 | if (XVID_VERSION_MAJOR(convert->version) != 1) /* v1.x.x */ |
585 : | suxen_drol | 890 | return XVID_ERR_VERSION; |
586 : | |||
587 : | edgomez | 1053 | #if 0 |
588 : | const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP); | ||
589 : | #endif | ||
590 : | suxen_drol | 890 | width = convert->width; |
591 : | height = convert->height; | ||
592 : | width2 = convert->width/2; | ||
593 : | height2 = convert->height/2; | ||
594 : | |||
595 : | switch (convert->input.csp & ~XVID_CSP_VFLIP) | ||
596 : | edgomez | 851 | { |
597 : | case XVID_CSP_YV12 : | ||
598 : | suxen_drol | 890 | img.y = convert->input.plane[0]; |
599 : | edgomez | 1161 | img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height; |
600 : | suxen_drol | 890 | img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2; |
601 : | edgomez | 851 | image_output(&img, width, height, width, |
602 : | suxen_drol | 890 | (uint8_t**)convert->output.plane, convert->output.stride, |
603 : | convert->output.csp, convert->interlacing); | ||
604 : | edgomez | 851 | break; |
605 : | |||
606 : | default : | ||
607 : | return XVID_ERR_FORMAT; | ||
608 : | } | ||
609 : | |||
610 : | |||
611 : | emms(); | ||
612 : | suxen_drol | 890 | return 0; |
613 : | edgomez | 851 | } |
614 : | |||
615 : | suxen_drol | 890 | /***************************************************************************** |
616 : | * XviD Global Entry point | ||
617 : | * | ||
618 : | * Well this function initialize all internal function pointers according | ||
619 : | * to the CPU features forced by the library client or autodetected (depending | ||
620 : | * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all | ||
621 : | * image colorspace transformation tables. | ||
622 : | edgomez | 1161 | * |
623 : | suxen_drol | 890 | ****************************************************************************/ |
624 : | |||
625 : | |||
626 : | edgomez | 851 | int |
627 : | suxen_drol | 890 | xvid_global(void *handle, |
628 : | edgomez | 851 | int opt, |
629 : | void *param1, | ||
630 : | void *param2) | ||
631 : | { | ||
632 : | switch(opt) | ||
633 : | { | ||
634 : | suxen_drol | 890 | case XVID_GBL_INIT : |
635 : | return xvid_gbl_init((xvid_gbl_init_t*)param1); | ||
636 : | edgomez | 851 | |
637 : | suxen_drol | 890 | case XVID_GBL_INFO : |
638 : | return xvid_gbl_info((xvid_gbl_info_t*)param1); | ||
639 : | edgomez | 851 | |
640 : | suxen_drol | 890 | case XVID_GBL_CONVERT : |
641 : | return xvid_gbl_convert((xvid_gbl_convert_t*)param1); | ||
642 : | |||
643 : | edgomez | 851 | default : |
644 : | return XVID_ERR_FAIL; | ||
645 : | } | ||
646 : | } | ||
647 : | |||
648 : | edgomez | 200 | /***************************************************************************** |
649 : | * XviD Native decoder entry point | ||
650 : | * | ||
651 : | * This function is just a wrapper to all the option cases. | ||
652 : | * | ||
653 : | * Returned values : XVID_ERR_FAIL when opt is invalid | ||
654 : | * else returns the wrapped function result | ||
655 : | * | ||
656 : | ****************************************************************************/ | ||
657 : | |||
658 : | edgomez | 195 | int |
659 : | xvid_decore(void *handle, | ||
660 : | int opt, | ||
661 : | void *param1, | ||
662 : | void *param2) | ||
663 : | Isibaar | 3 | { |
664 : | edgomez | 195 | switch (opt) { |
665 : | case XVID_DEC_CREATE: | ||
666 : | suxen_drol | 890 | return decoder_create((xvid_dec_create_t *) param1); |
667 : | Isibaar | 3 | |
668 : | edgomez | 195 | case XVID_DEC_DESTROY: |
669 : | return decoder_destroy((DECODER *) handle); | ||
670 : | |||
671 : | suxen_drol | 890 | case XVID_DEC_DECODE: |
672 : | return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2); | ||
673 : | |||
674 : | Isibaar | 3 | default: |
675 : | edgomez | 195 | return XVID_ERR_FAIL; |
676 : | } | ||
677 : | Isibaar | 3 | } |
678 : | |||
679 : | |||
680 : | edgomez | 200 | /***************************************************************************** |
681 : | * XviD Native encoder entry point | ||
682 : | * | ||
683 : | * This function is just a wrapper to all the option cases. | ||
684 : | * | ||
685 : | * Returned values : XVID_ERR_FAIL when opt is invalid | ||
686 : | * else returns the wrapped function result | ||
687 : | * | ||
688 : | ****************************************************************************/ | ||
689 : | |||
690 : | edgomez | 195 | int |
691 : | xvid_encore(void *handle, | ||
692 : | int opt, | ||
693 : | void *param1, | ||
694 : | void *param2) | ||
695 : | Isibaar | 3 | { |
696 : | edgomez | 195 | switch (opt) { |
697 : | case XVID_ENC_ENCODE: | ||
698 : | edgomez | 851 | |
699 : | suxen_drol | 890 | return enc_encode((Encoder *) handle, |
700 : | (xvid_enc_frame_t *) param1, | ||
701 : | (xvid_enc_stats_t *) param2); | ||
702 : | Isibaar | 3 | |
703 : | edgomez | 195 | case XVID_ENC_CREATE: |
704 : | suxen_drol | 948 | return enc_create((xvid_enc_create_t *) param1); |
705 : | Isibaar | 3 | |
706 : | edgomez | 195 | case XVID_ENC_DESTROY: |
707 : | suxen_drol | 890 | return enc_destroy((Encoder *) handle); |
708 : | edgomez | 195 | |
709 : | Isibaar | 3 | default: |
710 : | edgomez | 195 | return XVID_ERR_FAIL; |
711 : | } | ||
712 : | Isibaar | 3 | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |