Parent Directory | Revision Log
Revision 1066 - (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 | 1066 | * $Id: xvid.c,v 1.45.2.8 2003-06-11 14:10:40 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 : | Isibaar | 3 | #include "quant/quant_h263.h" |
43 : | #include "quant/quant_mpeg4.h" | ||
44 : | ia64p | 299 | #include "motion/motion.h" |
45 : | Isibaar | 3 | #include "motion/sad.h" |
46 : | #include "utils/emms.h" | ||
47 : | #include "utils/timer.h" | ||
48 : | Isibaar | 100 | #include "bitstream/mbcoding.h" |
49 : | Isibaar | 3 | |
50 : | suxen_drol | 1031 | #if defined(_DEBUG) |
51 : | unsigned int xvid_debug = 0; /* xvid debug mask */ | ||
52 : | #endif | ||
53 : | |||
54 : | edgomez | 851 | #if defined(ARCH_IS_IA32) |
55 : | suxen_drol | 311 | |
56 : | edgomez | 851 | #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 : | * | ||
154 : | * 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 : | 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 : | suxen_drol | 890 | if (XVID_MAJOR(init->version) != 1) /* v1.x.x */ |
167 : | 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 : | edgomez | 200 | /* Quantization functions */ |
186 : | quant_intra = quant_intra_c; | ||
187 : | Isibaar | 3 | dequant_intra = dequant_intra_c; |
188 : | edgomez | 200 | quant_inter = quant_inter_c; |
189 : | Isibaar | 3 | dequant_inter = dequant_inter_c; |
190 : | |||
191 : | edgomez | 200 | quant4_intra = quant4_intra_c; |
192 : | Isibaar | 3 | dequant4_intra = dequant4_intra_c; |
193 : | edgomez | 200 | quant4_inter = quant4_inter_c; |
194 : | Isibaar | 3 | dequant4_inter = dequant4_inter_c; |
195 : | |||
196 : | edgomez | 200 | /* Block transfer related functions */ |
197 : | Isibaar | 3 | transfer_8to16copy = transfer_8to16copy_c; |
198 : | transfer_16to8copy = transfer_16to8copy_c; | ||
199 : | edgomez | 200 | transfer_8to16sub = transfer_8to16sub_c; |
200 : | edgomez | 851 | transfer_8to16subro = transfer_8to16subro_c; |
201 : | suxen_drol | 118 | transfer_8to16sub2 = transfer_8to16sub2_c; |
202 : | edgomez | 200 | transfer_16to8add = transfer_16to8add_c; |
203 : | transfer8x8_copy = transfer8x8_copy_c; | ||
204 : | Isibaar | 3 | |
205 : | edgomez | 851 | /* Interlacing functions */ |
206 : | MBFieldTest = MBFieldTest_c; | ||
207 : | |||
208 : | edgomez | 200 | /* Image interpolation related functions */ |
209 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_c; | ||
210 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; | ||
211 : | Isibaar | 3 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
212 : | |||
213 : | edgomez | 851 | interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c; |
214 : | interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c; | ||
215 : | interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c; | ||
216 : | |||
217 : | interpolate8x8_lowpass_h = interpolate8x8_lowpass_h_c; | ||
218 : | interpolate8x8_lowpass_v = interpolate8x8_lowpass_v_c; | ||
219 : | interpolate8x8_lowpass_hv = interpolate8x8_lowpass_hv_c; | ||
220 : | |||
221 : | interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_c; | ||
222 : | interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_c; | ||
223 : | |||
224 : | interpolate8x8_avg2 = interpolate8x8_avg2_c; | ||
225 : | interpolate8x8_avg4 = interpolate8x8_avg4_c; | ||
226 : | |||
227 : | /* reduced resoltuion */ | ||
228 : | copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_C; | ||
229 : | add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_C; | ||
230 : | vfilter_31 = xvid_VFilter_31_C; | ||
231 : | hfilter_31 = xvid_HFilter_31_C; | ||
232 : | filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_C; | ||
233 : | filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_C; | ||
234 : | |||
235 : | edgomez | 200 | /* Initialize internal colorspace transformation tables */ |
236 : | Isibaar | 3 | colorspace_init(); |
237 : | |||
238 : | edgomez | 200 | /* All colorspace transformation functions User Format->YV12 */ |
239 : | edgomez | 851 | yv12_to_yv12 = yv12_to_yv12_c; |
240 : | rgb555_to_yv12 = rgb555_to_yv12_c; | ||
241 : | rgb565_to_yv12 = rgb565_to_yv12_c; | ||
242 : | bgr_to_yv12 = bgr_to_yv12_c; | ||
243 : | bgra_to_yv12 = bgra_to_yv12_c; | ||
244 : | abgr_to_yv12 = abgr_to_yv12_c; | ||
245 : | rgba_to_yv12 = rgba_to_yv12_c; | ||
246 : | yuyv_to_yv12 = yuyv_to_yv12_c; | ||
247 : | uyvy_to_yv12 = uyvy_to_yv12_c; | ||
248 : | Isibaar | 3 | |
249 : | edgomez | 851 | rgb555i_to_yv12 = rgb555i_to_yv12_c; |
250 : | rgb565i_to_yv12 = rgb565i_to_yv12_c; | ||
251 : | bgri_to_yv12 = bgri_to_yv12_c; | ||
252 : | bgrai_to_yv12 = bgrai_to_yv12_c; | ||
253 : | abgri_to_yv12 = abgri_to_yv12_c; | ||
254 : | rgbai_to_yv12 = rgbai_to_yv12_c; | ||
255 : | yuyvi_to_yv12 = yuyvi_to_yv12_c; | ||
256 : | uyvyi_to_yv12 = uyvyi_to_yv12_c; | ||
257 : | |||
258 : | |||
259 : | edgomez | 200 | /* All colorspace transformation functions YV12->User format */ |
260 : | edgomez | 851 | yv12_to_rgb555 = yv12_to_rgb555_c; |
261 : | yv12_to_rgb565 = yv12_to_rgb565_c; | ||
262 : | yv12_to_bgr = yv12_to_bgr_c; | ||
263 : | yv12_to_bgra = yv12_to_bgra_c; | ||
264 : | yv12_to_abgr = yv12_to_abgr_c; | ||
265 : | yv12_to_rgba = yv12_to_rgba_c; | ||
266 : | yv12_to_yuyv = yv12_to_yuyv_c; | ||
267 : | yv12_to_uyvy = yv12_to_uyvy_c; | ||
268 : | |||
269 : | yv12_to_rgb555i = yv12_to_rgb555i_c; | ||
270 : | yv12_to_rgb565i = yv12_to_rgb565i_c; | ||
271 : | yv12_to_bgri = yv12_to_bgri_c; | ||
272 : | yv12_to_bgrai = yv12_to_bgrai_c; | ||
273 : | yv12_to_abgri = yv12_to_abgri_c; | ||
274 : | yv12_to_rgbai = yv12_to_rgbai_c; | ||
275 : | yv12_to_yuyvi = yv12_to_yuyvi_c; | ||
276 : | yv12_to_uyvyi = yv12_to_uyvyi_c; | ||
277 : | Isibaar | 3 | |
278 : | edgomez | 200 | /* Functions used in motion estimation algorithms */ |
279 : | Isibaar | 3 | calc_cbp = calc_cbp_c; |
280 : | edgomez | 200 | sad16 = sad16_c; |
281 : | suxen_drol | 329 | sad8 = sad8_c; |
282 : | edgomez | 200 | sad16bi = sad16bi_c; |
283 : | suxen_drol | 329 | sad8bi = sad8bi_c; |
284 : | edgomez | 200 | dev16 = dev16_c; |
285 : | edgomez | 851 | sad16v = sad16v_c; |
286 : | suxen_drol | 329 | |
287 : | edgomez | 874 | /* Halfpel8_Refine = Halfpel8_Refine_c; */ |
288 : | Isibaar | 3 | |
289 : | edgomez | 851 | #if defined(ARCH_IS_IA32) |
290 : | edgomez | 200 | |
291 : | edgomez | 1058 | if ((cpu_flags & XVID_CPU_ASM)) { |
292 : | edgomez | 851 | vfilter_31 = xvid_VFilter_31_x86; |
293 : | hfilter_31 = xvid_HFilter_31_x86; | ||
294 : | } | ||
295 : | |||
296 : | if ((cpu_flags & XVID_CPU_MMX) || (cpu_flags & XVID_CPU_MMXEXT) || | ||
297 : | (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) || | ||
298 : | (cpu_flags & XVID_CPU_SSE) || (cpu_flags & XVID_CPU_SSE2)) | ||
299 : | { | ||
300 : | /* Restore FPU context : emms_c is a nop functions */ | ||
301 : | emms = emms_mmx; | ||
302 : | } | ||
303 : | |||
304 : | if ((cpu_flags & XVID_CPU_MMX)) { | ||
305 : | |||
306 : | edgomez | 200 | /* Forward and Inverse Discrete Cosine Transformation functions */ |
307 : | Isibaar | 3 | fdct = fdct_mmx; |
308 : | Isibaar | 1066 | idct = idct_mmx; |
309 : | Isibaar | 3 | |
310 : | edgomez | 200 | /* Quantization related functions */ |
311 : | quant_intra = quant_intra_mmx; | ||
312 : | Isibaar | 3 | dequant_intra = dequant_intra_mmx; |
313 : | edgomez | 200 | quant_inter = quant_inter_mmx; |
314 : | Isibaar | 3 | dequant_inter = dequant_inter_mmx; |
315 : | |||
316 : | edgomez | 200 | quant4_intra = quant4_intra_mmx; |
317 : | Isibaar | 3 | dequant4_intra = dequant4_intra_mmx; |
318 : | edgomez | 200 | quant4_inter = quant4_inter_mmx; |
319 : | Isibaar | 3 | dequant4_inter = dequant4_inter_mmx; |
320 : | |||
321 : | edgomez | 200 | /* Block related functions */ |
322 : | Isibaar | 3 | transfer_8to16copy = transfer_8to16copy_mmx; |
323 : | transfer_16to8copy = transfer_16to8copy_mmx; | ||
324 : | edgomez | 200 | transfer_8to16sub = transfer_8to16sub_mmx; |
325 : | edgomez | 851 | transfer_8to16subro = transfer_8to16subro_mmx; |
326 : | edgomez | 236 | transfer_8to16sub2 = transfer_8to16sub2_mmx; |
327 : | edgomez | 200 | transfer_16to8add = transfer_16to8add_mmx; |
328 : | transfer8x8_copy = transfer8x8_copy_mmx; | ||
329 : | Isibaar | 3 | |
330 : | edgomez | 851 | /* Interlacing Functions */ |
331 : | MBFieldTest = MBFieldTest_mmx; | ||
332 : | edgomez | 236 | |
333 : | edgomez | 200 | /* Image Interpolation related functions */ |
334 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_mmx; | ||
335 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; | ||
336 : | Isibaar | 3 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
337 : | |||
338 : | edgomez | 851 | interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx; |
339 : | interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx; | ||
340 : | |||
341 : | interpolate8x8_avg2 = interpolate8x8_avg2_mmx; | ||
342 : | interpolate8x8_avg4 = interpolate8x8_avg4_mmx; | ||
343 : | |||
344 : | /* reduced resolution */ | ||
345 : | copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_mmx; | ||
346 : | add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_mmx; | ||
347 : | hfilter_31 = xvid_HFilter_31_mmx; | ||
348 : | filter_18x18_to_8x8 = xvid_Filter_18x18_To_8x8_mmx; | ||
349 : | filter_diff_18x18_to_8x8 = xvid_Filter_Diff_18x18_To_8x8_mmx; | ||
350 : | |||
351 : | /* image input xxx_to_yv12 related functions */ | ||
352 : | yv12_to_yv12 = yv12_to_yv12_mmx; | ||
353 : | bgr_to_yv12 = bgr_to_yv12_mmx; | ||
354 : | bgra_to_yv12 = bgra_to_yv12_mmx; | ||
355 : | edgomez | 200 | yuyv_to_yv12 = yuyv_to_yv12_mmx; |
356 : | uyvy_to_yv12 = uyvy_to_yv12_mmx; | ||
357 : | Isibaar | 3 | |
358 : | edgomez | 851 | /* image output yv12_to_xxx related functions */ |
359 : | yv12_to_bgr = yv12_to_bgr_mmx; | ||
360 : | yv12_to_bgra = yv12_to_bgra_mmx; | ||
361 : | edgomez | 200 | yv12_to_yuyv = yv12_to_yuyv_mmx; |
362 : | yv12_to_uyvy = yv12_to_uyvy_mmx; | ||
363 : | edgomez | 851 | |
364 : | yv12_to_yuyvi = yv12_to_yuyvi_mmx; | ||
365 : | yv12_to_uyvyi = yv12_to_uyvyi_mmx; | ||
366 : | Isibaar | 3 | |
367 : | edgomez | 200 | /* Motion estimation related functions */ |
368 : | Isibaar | 3 | calc_cbp = calc_cbp_mmx; |
369 : | edgomez | 200 | sad16 = sad16_mmx; |
370 : | sad8 = sad8_mmx; | ||
371 : | suxen_drol | 329 | sad16bi = sad16bi_mmx; |
372 : | sad8bi = sad8bi_mmx; | ||
373 : | edgomez | 200 | dev16 = dev16_mmx; |
374 : | edgomez | 851 | sad16v = sad16v_mmx; |
375 : | Isibaar | 3 | } |
376 : | |||
377 : | suxen_drol | 329 | /* these 3dnow functions are faster than mmx, but slower than xmm. */ |
378 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_3DNOW)) { |
379 : | suxen_drol | 329 | |
380 : | edgomez | 851 | emms = emms_3dn; |
381 : | |||
382 : | suxen_drol | 329 | /* ME functions */ |
383 : | sad16bi = sad16bi_3dn; | ||
384 : | sad8bi = sad8bi_3dn; | ||
385 : | edgomez | 851 | |
386 : | yuyv_to_yv12 = yuyv_to_yv12_3dn; | ||
387 : | uyvy_to_yv12 = uyvy_to_yv12_3dn; | ||
388 : | suxen_drol | 329 | } |
389 : | |||
390 : | |||
391 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_MMXEXT)) { |
392 : | edgomez | 200 | |
393 : | /* Inverse DCT */ | ||
394 : | Isibaar | 3 | idct = idct_xmm; |
395 : | edgomez | 200 | |
396 : | /* Interpolation */ | ||
397 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_xmm; | ||
398 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; | ||
399 : | h | 38 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
400 : | edgomez | 200 | |
401 : | edgomez | 851 | /* reduced resolution */ |
402 : | copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm; | ||
403 : | add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm; | ||
404 : | |||
405 : | chenm001 | 274 | /* Quantization */ |
406 : | edgomez | 851 | quant4_intra = quant4_intra_xmm; |
407 : | quant4_inter = quant4_inter_xmm; | ||
408 : | |||
409 : | chenm001 | 274 | dequant_intra = dequant_intra_xmm; |
410 : | dequant_inter = dequant_inter_xmm; | ||
411 : | |||
412 : | edgomez | 218 | /* Buffer transfer */ |
413 : | transfer_8to16sub2 = transfer_8to16sub2_xmm; | ||
414 : | |||
415 : | edgomez | 200 | /* Colorspace transformation */ |
416 : | edgomez | 851 | yv12_to_yv12 = yv12_to_yv12_xmm; |
417 : | yuyv_to_yv12 = yuyv_to_yv12_xmm; | ||
418 : | uyvy_to_yv12 = uyvy_to_yv12_xmm; | ||
419 : | Isibaar | 3 | |
420 : | edgomez | 200 | /* ME functions */ |
421 : | Isibaar | 3 | sad16 = sad16_xmm; |
422 : | suxen_drol | 329 | sad8 = sad8_xmm; |
423 : | chenm001 | 274 | sad16bi = sad16bi_xmm; |
424 : | suxen_drol | 329 | sad8bi = sad8bi_xmm; |
425 : | Isibaar | 3 | dev16 = dev16_xmm; |
426 : | edgomez | 851 | sad16v = sad16v_xmm; |
427 : | Isibaar | 3 | } |
428 : | |||
429 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_3DNOW)) { |
430 : | edgomez | 200 | |
431 : | /* Interpolation */ | ||
432 : | Isibaar | 3 | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dn; |
433 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dn; | ||
434 : | h | 40 | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dn; |
435 : | Isibaar | 3 | } |
436 : | |||
437 : | edgomez | 851 | if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
438 : | edgomez | 200 | |
439 : | edgomez | 851 | /* Inverse DCT */ |
440 : | idct = idct_3dne; | ||
441 : | |||
442 : | /* Buffer transfer */ | ||
443 : | transfer_8to16copy = transfer_8to16copy_3dne; | ||
444 : | transfer_16to8copy = transfer_16to8copy_3dne; | ||
445 : | transfer_8to16sub = transfer_8to16sub_3dne; | ||
446 : | transfer_8to16subro = transfer_8to16subro_3dne; | ||
447 : | transfer_8to16sub2 = transfer_8to16sub2_3dne; | ||
448 : | transfer_16to8add = transfer_16to8add_3dne; | ||
449 : | transfer8x8_copy = transfer8x8_copy_3dne; | ||
450 : | |||
451 : | /* Quantization */ | ||
452 : | dequant4_intra = dequant4_intra_3dne; | ||
453 : | dequant4_inter = dequant4_inter_3dne; | ||
454 : | quant_intra = quant_intra_3dne; | ||
455 : | quant_inter = quant_inter_3dne; | ||
456 : | dequant_intra = dequant_intra_3dne; | ||
457 : | dequant_inter = dequant_inter_3dne; | ||
458 : | |||
459 : | /* ME functions */ | ||
460 : | calc_cbp = calc_cbp_3dne; | ||
461 : | sad16 = sad16_3dne; | ||
462 : | sad8 = sad8_3dne; | ||
463 : | sad16bi = sad16bi_3dne; | ||
464 : | sad8bi = sad8bi_3dne; | ||
465 : | dev16 = dev16_3dne; | ||
466 : | |||
467 : | /* Interpolation */ | ||
468 : | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne; | ||
469 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; | ||
470 : | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; | ||
471 : | } | ||
472 : | |||
473 : | |||
474 : | if ((cpu_flags & XVID_CPU_SSE2)) { | ||
475 : | |||
476 : | chenm001 | 274 | calc_cbp = calc_cbp_sse2; |
477 : | |||
478 : | edgomez | 200 | /* Quantization */ |
479 : | quant_intra = quant_intra_sse2; | ||
480 : | Isibaar | 154 | dequant_intra = dequant_intra_sse2; |
481 : | edgomez | 200 | quant_inter = quant_inter_sse2; |
482 : | Isibaar | 154 | dequant_inter = dequant_inter_sse2; |
483 : | h | 135 | |
484 : | edgomez | 851 | #if defined(EXPERIMENTAL_SSE2_CODE) |
485 : | /* ME; slower than xmm */ | ||
486 : | edgomez | 200 | sad16 = sad16_sse2; |
487 : | dev16 = dev16_sse2; | ||
488 : | edgomez | 851 | #endif |
489 : | edgomez | 200 | /* Forward and Inverse DCT */ |
490 : | edgomez | 1058 | #if 0 /* Both function are known to be unprecise, better keep them deactivated */ |
491 : | edgomez | 200 | idct = idct_sse2; |
492 : | Isibaar | 154 | fdct = fdct_sse2; |
493 : | edgomez | 1058 | #endif |
494 : | h | 126 | } |
495 : | Isibaar | 3 | #endif |
496 : | edgomez | 200 | |
497 : | edgomez | 851 | #if defined(ARCH_IS_IA64) |
498 : | edgomez | 874 | if ((cpu_flags & XVID_CPU_ASM)) { /* use assembler routines? */ |
499 : | Isibaar | 209 | idct_ia64_init(); |
500 : | fdct = fdct_ia64; | ||
501 : | edgomez | 874 | idct = idct_ia64; /*not yet working, crashes */ |
502 : | Isibaar | 209 | interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_ia64; |
503 : | interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_ia64; | ||
504 : | interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_ia64; | ||
505 : | sad16 = sad16_ia64; | ||
506 : | sad16bi = sad16bi_ia64; | ||
507 : | sad8 = sad8_ia64; | ||
508 : | dev16 = dev16_ia64; | ||
509 : | edgomez | 874 | /* Halfpel8_Refine = Halfpel8_Refine_ia64; */ |
510 : | Isibaar | 209 | quant_intra = quant_intra_ia64; |
511 : | dequant_intra = dequant_intra_ia64; | ||
512 : | quant_inter = quant_inter_ia64; | ||
513 : | dequant_inter = dequant_inter_ia64; | ||
514 : | transfer_8to16copy = transfer_8to16copy_ia64; | ||
515 : | transfer_16to8copy = transfer_16to8copy_ia64; | ||
516 : | transfer_8to16sub = transfer_8to16sub_ia64; | ||
517 : | transfer_8to16sub2 = transfer_8to16sub2_ia64; | ||
518 : | transfer_16to8add = transfer_16to8add_ia64; | ||
519 : | transfer8x8_copy = transfer8x8_copy_ia64; | ||
520 : | edgomez | 874 | DPRINTF(DPRINTF_DEBUG, "Using IA-64 assembler routines."); |
521 : | Isibaar | 209 | } |
522 : | #endif | ||
523 : | |||
524 : | edgomez | 851 | #if defined(ARCH_IS_PPC) |
525 : | if ((cpu_flags & XVID_CPU_ASM)) | ||
526 : | { | ||
527 : | calc_cbp = calc_cbp_ppc; | ||
528 : | } | ||
529 : | |||
530 : | if ((cpu_flags & XVID_CPU_ALTIVEC)) | ||
531 : | { | ||
532 : | calc_cbp = calc_cbp_altivec; | ||
533 : | fdct = fdct_altivec; | ||
534 : | idct = idct_altivec; | ||
535 : | sadInit = sadInit_altivec; | ||
536 : | sad16 = sad16_altivec; | ||
537 : | sad8 = sad8_altivec; | ||
538 : | dev16 = dev16_altivec; | ||
539 : | } | ||
540 : | canard | 52 | #endif |
541 : | edgomez | 851 | |
542 : | suxen_drol | 1031 | #if defined(_DEBUG) |
543 : | xvid_debug = init->debug; | ||
544 : | #endif | ||
545 : | |||
546 : | return 0; | ||
547 : | edgomez | 851 | } |
548 : | |||
549 : | |||
550 : | suxen_drol | 890 | static int |
551 : | xvid_gbl_info(xvid_gbl_info_t * info) | ||
552 : | { | ||
553 : | if (XVID_MAJOR(info->version) != 1) /* v1.x.x */ | ||
554 : | return XVID_ERR_VERSION; | ||
555 : | edgomez | 851 | |
556 : | suxen_drol | 890 | info->actual_version = XVID_VERSION; |
557 : | info->build = "dev-api-4"; | ||
558 : | info->cpu_flags = detect_cpu_flags(); | ||
559 : | |||
560 : | #if defined(_SMP) && defined(WIN32) | ||
561 : | info->num_threads = pthread_num_processors_np();; | ||
562 : | #else | ||
563 : | info->num_threads = 0; | ||
564 : | #endif | ||
565 : | |||
566 : | return 0; | ||
567 : | } | ||
568 : | |||
569 : | |||
570 : | edgomez | 851 | static int |
571 : | suxen_drol | 890 | xvid_gbl_convert(xvid_gbl_convert_t* convert) |
572 : | edgomez | 851 | { |
573 : | suxen_drol | 890 | int width; |
574 : | int height; | ||
575 : | int width2; | ||
576 : | int height2; | ||
577 : | edgomez | 851 | IMAGE img; |
578 : | |||
579 : | suxen_drol | 890 | if (XVID_MAJOR(convert->version) != 1) /* v1.x.x */ |
580 : | return XVID_ERR_VERSION; | ||
581 : | |||
582 : | edgomez | 1053 | #if 0 |
583 : | const int flip1 = (convert->input.colorspace & XVID_CSP_VFLIP) ^ (convert->output.colorspace & XVID_CSP_VFLIP); | ||
584 : | #endif | ||
585 : | suxen_drol | 890 | width = convert->width; |
586 : | height = convert->height; | ||
587 : | width2 = convert->width/2; | ||
588 : | height2 = convert->height/2; | ||
589 : | |||
590 : | switch (convert->input.csp & ~XVID_CSP_VFLIP) | ||
591 : | edgomez | 851 | { |
592 : | case XVID_CSP_YV12 : | ||
593 : | suxen_drol | 890 | img.y = convert->input.plane[0]; |
594 : | img.v = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height; | ||
595 : | img.u = (uint8_t*)convert->input.plane[0] + convert->input.stride[0]*height + (convert->input.stride[0]/2)*height2; | ||
596 : | edgomez | 851 | image_output(&img, width, height, width, |
597 : | suxen_drol | 890 | (uint8_t**)convert->output.plane, convert->output.stride, |
598 : | convert->output.csp, convert->interlacing); | ||
599 : | edgomez | 851 | break; |
600 : | |||
601 : | default : | ||
602 : | return XVID_ERR_FORMAT; | ||
603 : | } | ||
604 : | |||
605 : | |||
606 : | emms(); | ||
607 : | suxen_drol | 890 | return 0; |
608 : | edgomez | 851 | } |
609 : | |||
610 : | |||
611 : | |||
612 : | void fill8(uint8_t * block, int size, int value) | ||
613 : | { | ||
614 : | int i; | ||
615 : | for (i = 0; i < size; i++) | ||
616 : | block[i] = value; | ||
617 : | } | ||
618 : | |||
619 : | void fill16(int16_t * block, int size, int value) | ||
620 : | { | ||
621 : | int i; | ||
622 : | for (i = 0; i < size; i++) | ||
623 : | block[i] = value; | ||
624 : | } | ||
625 : | |||
626 : | #define RANDOM(min,max) min + (rand() % (max-min)) | ||
627 : | |||
628 : | void random8(uint8_t * block, int size, int min, int max) | ||
629 : | { | ||
630 : | int i; | ||
631 : | for (i = 0; i < size; i++) | ||
632 : | block[i] = RANDOM(min,max); | ||
633 : | } | ||
634 : | |||
635 : | void random16(int16_t * block, int size, int min, int max) | ||
636 : | { | ||
637 : | int i; | ||
638 : | for (i = 0; i < size; i++) | ||
639 : | block[i] = RANDOM(min,max); | ||
640 : | } | ||
641 : | |||
642 : | int compare16(const int16_t * blockA, const int16_t * blockB, int size) | ||
643 : | { | ||
644 : | int i; | ||
645 : | for (i = 0; i < size; i++) | ||
646 : | if (blockA[i] != blockB[i]) | ||
647 : | return 1; | ||
648 : | |||
649 : | return 0; | ||
650 : | } | ||
651 : | |||
652 : | int diff16(const int16_t * blockA, const int16_t * blockB, int size) | ||
653 : | { | ||
654 : | int i, diff = 0; | ||
655 : | for (i = 0; i < size; i++) | ||
656 : | edgomez | 982 | diff += abs(blockA[i]-blockB[i]); |
657 : | edgomez | 851 | return diff; |
658 : | } | ||
659 : | |||
660 : | |||
661 : | #define XVID_TEST_RANDOM 0x00000001 /* random input data */ | ||
662 : | #define XVID_TEST_VERBOSE 0x00000002 /* verbose error output */ | ||
663 : | |||
664 : | |||
665 : | #define TEST_FORWARD 0x00000001 /* intra */ | ||
666 : | #define TEST_FDCT (TEST_FORWARD) | ||
667 : | #define TEST_IDCT (0) | ||
668 : | |||
669 : | suxen_drol | 860 | static int test_transform(void * funcA, void * funcB, const char * nameB, |
670 : | edgomez | 851 | int test, int flags) |
671 : | { | ||
672 : | int i; | ||
673 : | int64_t timeSTART; | ||
674 : | int64_t timeA = 0; | ||
675 : | int64_t timeB = 0; | ||
676 : | DECLARE_ALIGNED_MATRIX(arrayA, 1, 64, int16_t, CACHE_LINE); | ||
677 : | DECLARE_ALIGNED_MATRIX(arrayB, 1, 64, int16_t, CACHE_LINE); | ||
678 : | int min, max; | ||
679 : | int count = 0; | ||
680 : | |||
681 : | int tmp; | ||
682 : | int min_error = 0x10000*64; | ||
683 : | int max_error = 0; | ||
684 : | |||
685 : | |||
686 : | if ((test & TEST_FORWARD)) /* forward */ | ||
687 : | { | ||
688 : | min = -256; | ||
689 : | max = 255; | ||
690 : | }else{ /* inverse */ | ||
691 : | min = -2048; | ||
692 : | max = 2047; | ||
693 : | } | ||
694 : | |||
695 : | for (i = 0; i < 64*64; i++) | ||
696 : | { | ||
697 : | if ((flags & XVID_TEST_RANDOM)) | ||
698 : | { | ||
699 : | random16(arrayA, 64, min, max); | ||
700 : | }else{ | ||
701 : | fill16(arrayA, 64, i); | ||
702 : | } | ||
703 : | memcpy(arrayB, arrayA, 64*sizeof(int16_t)); | ||
704 : | |||
705 : | if ((test & TEST_FORWARD)) | ||
706 : | { | ||
707 : | timeSTART = read_counter(); | ||
708 : | ((fdctFunc*)funcA)(arrayA); | ||
709 : | timeA += read_counter() - timeSTART; | ||
710 : | |||
711 : | timeSTART = read_counter(); | ||
712 : | ((fdctFunc*)funcB)(arrayB); | ||
713 : | timeB += read_counter() - timeSTART; | ||
714 : | } | ||
715 : | else | ||
716 : | { | ||
717 : | timeSTART = read_counter(); | ||
718 : | ((idctFunc*)funcA)(arrayA); | ||
719 : | timeA += read_counter() - timeSTART; | ||
720 : | |||
721 : | timeSTART = read_counter(); | ||
722 : | ((idctFunc*)funcB)(arrayB); | ||
723 : | timeB += read_counter() - timeSTART; | ||
724 : | } | ||
725 : | |||
726 : | tmp = diff16(arrayA, arrayB, 64) / 64; | ||
727 : | if (tmp > max_error) | ||
728 : | max_error = tmp; | ||
729 : | if (tmp < min_error) | ||
730 : | min_error = tmp; | ||
731 : | |||
732 : | count++; | ||
733 : | } | ||
734 : | |||
735 : | /* print the "average difference" of best/worst transforms */ | ||
736 : | printf("%s:\t%i\t(min_error:%i, max_error:%i)\n", nameB, (int)(timeB / count), min_error, max_error); | ||
737 : | |||
738 : | return 0; | ||
739 : | } | ||
740 : | |||
741 : | |||
742 : | #define TEST_QUANT 0x00000001 /* forward quantization */ | ||
743 : | #define TEST_INTRA 0x00000002 /* intra */ | ||
744 : | #define TEST_QUANT_INTRA (TEST_QUANT|TEST_INTRA) | ||
745 : | #define TEST_QUANT_INTER (TEST_QUANT) | ||
746 : | #define TEST_DEQUANT_INTRA (TEST_INTRA) | ||
747 : | #define TEST_DEQUANT_INTER (0) | ||
748 : | |||
749 : | suxen_drol | 860 | static int test_quant(void * funcA, void * funcB, const char * nameB, |
750 : | edgomez | 851 | int test, int flags) |
751 : | { | ||
752 : | int q,i; | ||
753 : | int64_t timeSTART; | ||
754 : | int64_t timeA = 0; | ||
755 : | int64_t timeB = 0; | ||
756 : | edgomez | 874 | int retA = 0, retB = 0; |
757 : | edgomez | 851 | DECLARE_ALIGNED_MATRIX(arrayX, 1, 64, int16_t, CACHE_LINE); |
758 : | DECLARE_ALIGNED_MATRIX(arrayA, 1, 64, int16_t, CACHE_LINE); | ||
759 : | DECLARE_ALIGNED_MATRIX(arrayB, 1, 64, int16_t, CACHE_LINE); | ||
760 : | int min, max; | ||
761 : | int count = 0; | ||
762 : | int errors = 0; | ||
763 : | |||
764 : | if ((test & TEST_QUANT)) /* quant */ | ||
765 : | { | ||
766 : | min = -2048; | ||
767 : | max = 2047; | ||
768 : | }else{ /* dequant */ | ||
769 : | min = -256; | ||
770 : | max = 255; | ||
771 : | } | ||
772 : | |||
773 : | for (q = 1; q <= 31; q++) /* quantizer */ | ||
774 : | { | ||
775 : | for (i = min; i < max; i++) /* input coeff */ | ||
776 : | { | ||
777 : | if ((flags & XVID_TEST_RANDOM)) | ||
778 : | { | ||
779 : | random16(arrayX, 64, min, max); | ||
780 : | }else{ | ||
781 : | fill16(arrayX, 64, i); | ||
782 : | } | ||
783 : | |||
784 : | if ((test & TEST_INTRA)) /* intra */ | ||
785 : | { | ||
786 : | timeSTART = read_counter(); | ||
787 : | ((quanth263_intraFunc*)funcA)(arrayA, arrayX, q, q); | ||
788 : | timeA += read_counter() - timeSTART; | ||
789 : | |||
790 : | timeSTART = read_counter(); | ||
791 : | ((quanth263_intraFunc*)funcB)(arrayB, arrayX, q, q); | ||
792 : | timeB += read_counter() - timeSTART; | ||
793 : | } | ||
794 : | else /* inter */ | ||
795 : | { | ||
796 : | timeSTART = read_counter(); | ||
797 : | retA = ((quanth263_interFunc*)funcA)(arrayA, arrayX, q); | ||
798 : | timeA += read_counter() - timeSTART; | ||
799 : | |||
800 : | timeSTART = read_counter(); | ||
801 : | retB = ((quanth263_interFunc*)funcB)(arrayB, arrayX, q); | ||
802 : | timeB += read_counter() - timeSTART; | ||
803 : | } | ||
804 : | |||
805 : | /* compare return value from quant_inter, and compare (de)quantiz'd arrays */ | ||
806 : | if ( ((test&TEST_QUANT) && !(test&TEST_INTRA) && retA != retB ) || | ||
807 : | compare16(arrayA, arrayB, 64)) | ||
808 : | { | ||
809 : | errors++; | ||
810 : | if ((flags & XVID_TEST_VERBOSE)) | ||
811 : | printf("%s error: q=%i, i=%i\n", nameB, q, i); | ||
812 : | } | ||
813 : | |||
814 : | count++; | ||
815 : | } | ||
816 : | } | ||
817 : | |||
818 : | printf("%s:\t%i", nameB, (int)(timeB / count)); | ||
819 : | if (errors>0) | ||
820 : | printf("\t(%i errors out of %i)", errors, count); | ||
821 : | printf("\n"); | ||
822 : | |||
823 : | return 0; | ||
824 : | } | ||
825 : | |||
826 : | |||
827 : | |||
828 : | int xvid_init_test(int flags) | ||
829 : | { | ||
830 : | edgomez | 882 | #if defined(ARCH_IS_IA32) |
831 : | int cpu_flags; | ||
832 : | #endif | ||
833 : | edgomez | 851 | |
834 : | edgomez | 879 | printf("XviD tests\n\n"); |
835 : | edgomez | 851 | |
836 : | edgomez | 879 | #if defined(ARCH_IS_IA32) |
837 : | edgomez | 851 | cpu_flags = detect_cpu_flags(); |
838 : | edgomez | 879 | #endif |
839 : | |||
840 : | edgomez | 851 | idct_int32_init(); |
841 : | edgomez | 879 | emms(); |
842 : | edgomez | 851 | |
843 : | edgomez | 882 | srand(time(0)); |
844 : | |||
845 : | edgomez | 879 | /* fDCT test */ |
846 : | edgomez | 851 | printf("--- fdct ---\n"); |
847 : | test_transform(fdct_int32, fdct_int32, "c", TEST_FDCT, flags); | ||
848 : | edgomez | 879 | |
849 : | #if defined(ARCH_IS_IA32) | ||
850 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
851 : | test_transform(fdct_int32, fdct_mmx, "mmx", TEST_FDCT, flags); | ||
852 : | if (cpu_flags & XVID_CPU_SSE2) | ||
853 : | test_transform(fdct_int32, fdct_sse2, "sse2", TEST_FDCT, flags); | ||
854 : | edgomez | 879 | #endif |
855 : | edgomez | 851 | |
856 : | edgomez | 879 | /* iDCT test */ |
857 : | edgomez | 851 | printf("\n--- idct ---\n"); |
858 : | test_transform(idct_int32, idct_int32, "c", TEST_IDCT, flags); | ||
859 : | edgomez | 879 | |
860 : | #if defined(ARCH_IS_IA32) | ||
861 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
862 : | test_transform(idct_int32, idct_mmx, "mmx", TEST_IDCT, flags); | ||
863 : | if (cpu_flags & XVID_CPU_MMXEXT) | ||
864 : | test_transform(idct_int32, idct_xmm, "xmm", TEST_IDCT, flags); | ||
865 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
866 : | test_transform(idct_int32, idct_3dne, "3dne", TEST_IDCT, flags); | ||
867 : | if (cpu_flags & XVID_CPU_SSE2) | ||
868 : | test_transform(idct_int32, idct_sse2, "sse2", TEST_IDCT, flags); | ||
869 : | edgomez | 879 | #endif |
870 : | edgomez | 851 | |
871 : | edgomez | 879 | /* Intra quantization test */ |
872 : | edgomez | 851 | printf("\n--- quant intra ---\n"); |
873 : | test_quant(quant_intra_c, quant_intra_c, "c", TEST_QUANT_INTRA, flags); | ||
874 : | edgomez | 879 | |
875 : | #if defined(ARCH_IS_IA32) | ||
876 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
877 : | test_quant(quant_intra_c, quant_intra_mmx, "mmx", TEST_QUANT_INTRA, flags); | ||
878 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
879 : | test_quant(quant_intra_c, quant_intra_3dne, "3dne", TEST_QUANT_INTRA, flags); | ||
880 : | if (cpu_flags & XVID_CPU_SSE2) | ||
881 : | test_quant(quant_intra_c, quant_intra_sse2, "sse2", TEST_QUANT_INTRA, flags); | ||
882 : | edgomez | 879 | #endif |
883 : | edgomez | 851 | |
884 : | edgomez | 879 | /* Inter quantization test */ |
885 : | edgomez | 851 | printf("\n--- quant inter ---\n"); |
886 : | test_quant(quant_inter_c, quant_inter_c, "c", TEST_QUANT_INTER, flags); | ||
887 : | edgomez | 879 | |
888 : | #if defined(ARCH_IS_IA32) | ||
889 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
890 : | test_quant(quant_inter_c, quant_inter_mmx, "mmx", TEST_QUANT_INTER, flags); | ||
891 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
892 : | test_quant(quant_inter_c, quant_inter_3dne, "3dne", TEST_QUANT_INTER, flags); | ||
893 : | if (cpu_flags & XVID_CPU_SSE2) | ||
894 : | test_quant(quant_inter_c, quant_inter_sse2, "sse2", TEST_QUANT_INTER, flags); | ||
895 : | edgomez | 879 | #endif |
896 : | edgomez | 851 | |
897 : | edgomez | 879 | /* Intra dequantization test */ |
898 : | edgomez | 851 | printf("\n--- dequant intra ---\n"); |
899 : | test_quant(dequant_intra_c, dequant_intra_c, "c", TEST_DEQUANT_INTRA, flags); | ||
900 : | edgomez | 879 | |
901 : | #if defined(ARCH_IS_IA32) | ||
902 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
903 : | test_quant(dequant_intra_c, dequant_intra_mmx, "mmx", TEST_DEQUANT_INTRA, flags); | ||
904 : | if (cpu_flags & XVID_CPU_MMXEXT) | ||
905 : | test_quant(dequant_intra_c, dequant_intra_xmm, "xmm", TEST_DEQUANT_INTRA, flags); | ||
906 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
907 : | test_quant(dequant_intra_c, dequant_intra_3dne, "3dne", TEST_DEQUANT_INTRA, flags); | ||
908 : | if (cpu_flags & XVID_CPU_SSE2) | ||
909 : | test_quant(dequant_intra_c, dequant_intra_sse2, "sse2", TEST_DEQUANT_INTRA, flags); | ||
910 : | edgomez | 879 | #endif |
911 : | edgomez | 851 | |
912 : | edgomez | 879 | /* Inter dequantization test */ |
913 : | edgomez | 851 | printf("\n--- dequant inter ---\n"); |
914 : | test_quant(dequant_inter_c, dequant_inter_c, "c", TEST_DEQUANT_INTER, flags); | ||
915 : | edgomez | 879 | |
916 : | #if defined(ARCH_IS_IA32) | ||
917 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
918 : | test_quant(dequant_inter_c, dequant_inter_mmx, "mmx", TEST_DEQUANT_INTER, flags); | ||
919 : | if (cpu_flags & XVID_CPU_MMXEXT) | ||
920 : | test_quant(dequant_inter_c, dequant_inter_xmm, "xmm", TEST_DEQUANT_INTER, flags); | ||
921 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
922 : | test_quant(dequant_inter_c, dequant_inter_3dne, "3dne", TEST_DEQUANT_INTER, flags); | ||
923 : | if (cpu_flags & XVID_CPU_SSE2) | ||
924 : | test_quant(dequant_inter_c, dequant_inter_sse2, "sse2", TEST_DEQUANT_INTER, flags); | ||
925 : | edgomez | 879 | #endif |
926 : | edgomez | 851 | |
927 : | edgomez | 879 | /* Intra quantization test */ |
928 : | printf("\n--- quant4 intra ---\n"); | ||
929 : | edgomez | 851 | test_quant(quant4_intra_c, quant4_intra_c, "c", TEST_QUANT_INTRA, flags); |
930 : | edgomez | 879 | |
931 : | #if defined(ARCH_IS_IA32) | ||
932 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
933 : | test_quant(quant4_intra_c, quant4_intra_mmx, "mmx", TEST_QUANT_INTRA, flags); | ||
934 : | if (cpu_flags & XVID_CPU_MMXEXT) | ||
935 : | test_quant(quant4_intra_c, quant4_intra_xmm, "xmm", TEST_QUANT_INTRA, flags); | ||
936 : | edgomez | 879 | #endif |
937 : | edgomez | 851 | |
938 : | edgomez | 879 | /* Inter quantization test */ |
939 : | printf("\n--- quant4 inter ---\n"); | ||
940 : | edgomez | 851 | test_quant(quant4_inter_c, quant4_inter_c, "c", TEST_QUANT_INTER, flags); |
941 : | edgomez | 879 | |
942 : | #if defined(ARCH_IS_IA32) | ||
943 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
944 : | test_quant(quant4_inter_c, quant4_inter_mmx, "mmx", TEST_QUANT_INTER, flags); | ||
945 : | if (cpu_flags & XVID_CPU_MMXEXT) | ||
946 : | test_quant(quant4_inter_c, quant4_inter_xmm, "xmm", TEST_QUANT_INTER, flags); | ||
947 : | edgomez | 879 | #endif |
948 : | edgomez | 851 | |
949 : | edgomez | 879 | /* Intra dequantization test */ |
950 : | printf("\n--- dequant4 intra ---\n"); | ||
951 : | edgomez | 851 | test_quant(dequant4_intra_c, dequant4_intra_c, "c", TEST_DEQUANT_INTRA, flags); |
952 : | edgomez | 879 | |
953 : | #if defined(ARCH_IS_IA32) | ||
954 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
955 : | test_quant(dequant4_intra_c, dequant4_intra_mmx, "mmx", TEST_DEQUANT_INTRA, flags); | ||
956 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
957 : | test_quant(dequant4_intra_c, dequant4_intra_3dne, "3dne", TEST_DEQUANT_INTRA, flags); | ||
958 : | edgomez | 879 | #endif |
959 : | edgomez | 851 | |
960 : | edgomez | 879 | /* Inter dequantization test */ |
961 : | printf("\n--- dequant4 inter ---\n"); | ||
962 : | edgomez | 851 | test_quant(dequant4_inter_c, dequant4_inter_c, "c", TEST_DEQUANT_INTER, flags); |
963 : | edgomez | 879 | |
964 : | #if defined(ARCH_IS_IA32) | ||
965 : | edgomez | 851 | if (cpu_flags & XVID_CPU_MMX) |
966 : | test_quant(dequant4_inter_c, dequant4_inter_mmx, "mmx", TEST_DEQUANT_INTER, flags); | ||
967 : | if (cpu_flags & XVID_CPU_3DNOWEXT) | ||
968 : | test_quant(dequant4_inter_c, dequant4_inter_3dne, "3dne", TEST_DEQUANT_INTER, flags); | ||
969 : | edgomez | 879 | #endif |
970 : | edgomez | 851 | |
971 : | edgomez | 879 | emms(); |
972 : | edgomez | 851 | |
973 : | suxen_drol | 890 | return 0; |
974 : | Isibaar | 3 | } |
975 : | |||
976 : | edgomez | 851 | |
977 : | suxen_drol | 890 | /***************************************************************************** |
978 : | * XviD Global Entry point | ||
979 : | * | ||
980 : | * Well this function initialize all internal function pointers according | ||
981 : | * to the CPU features forced by the library client or autodetected (depending | ||
982 : | * on the XVID_CPU_FORCE flag). It also initializes vlc coding tables and all | ||
983 : | * image colorspace transformation tables. | ||
984 : | * | ||
985 : | ****************************************************************************/ | ||
986 : | |||
987 : | |||
988 : | edgomez | 851 | int |
989 : | suxen_drol | 890 | xvid_global(void *handle, |
990 : | edgomez | 851 | int opt, |
991 : | void *param1, | ||
992 : | void *param2) | ||
993 : | { | ||
994 : | switch(opt) | ||
995 : | { | ||
996 : | suxen_drol | 890 | case XVID_GBL_INIT : |
997 : | return xvid_gbl_init((xvid_gbl_init_t*)param1); | ||
998 : | edgomez | 851 | |
999 : | suxen_drol | 890 | case XVID_GBL_INFO : |
1000 : | return xvid_gbl_info((xvid_gbl_info_t*)param1); | ||
1001 : | edgomez | 851 | |
1002 : | suxen_drol | 890 | case XVID_GBL_CONVERT : |
1003 : | return xvid_gbl_convert((xvid_gbl_convert_t*)param1); | ||
1004 : | |||
1005 : | case XVID_GBL_TEST : | ||
1006 : | edgomez | 879 | { |
1007 : | ptr_t flags = (ptr_t)param1; | ||
1008 : | return xvid_init_test((int)flags); | ||
1009 : | } | ||
1010 : | edgomez | 851 | default : |
1011 : | return XVID_ERR_FAIL; | ||
1012 : | } | ||
1013 : | } | ||
1014 : | |||
1015 : | edgomez | 200 | /***************************************************************************** |
1016 : | * XviD Native decoder entry point | ||
1017 : | * | ||
1018 : | * This function is just a wrapper to all the option cases. | ||
1019 : | * | ||
1020 : | * Returned values : XVID_ERR_FAIL when opt is invalid | ||
1021 : | * else returns the wrapped function result | ||
1022 : | * | ||
1023 : | ****************************************************************************/ | ||
1024 : | |||
1025 : | edgomez | 195 | int |
1026 : | xvid_decore(void *handle, | ||
1027 : | int opt, | ||
1028 : | void *param1, | ||
1029 : | void *param2) | ||
1030 : | Isibaar | 3 | { |
1031 : | edgomez | 195 | switch (opt) { |
1032 : | case XVID_DEC_CREATE: | ||
1033 : | suxen_drol | 890 | return decoder_create((xvid_dec_create_t *) param1); |
1034 : | Isibaar | 3 | |
1035 : | edgomez | 195 | case XVID_DEC_DESTROY: |
1036 : | return decoder_destroy((DECODER *) handle); | ||
1037 : | |||
1038 : | suxen_drol | 890 | case XVID_DEC_DECODE: |
1039 : | return decoder_decode((DECODER *) handle, (xvid_dec_frame_t *) param1, (xvid_dec_stats_t*) param2); | ||
1040 : | |||
1041 : | Isibaar | 3 | default: |
1042 : | edgomez | 195 | return XVID_ERR_FAIL; |
1043 : | } | ||
1044 : | Isibaar | 3 | } |
1045 : | |||
1046 : | |||
1047 : | edgomez | 200 | /***************************************************************************** |
1048 : | * XviD Native encoder entry point | ||
1049 : | * | ||
1050 : | * This function is just a wrapper to all the option cases. | ||
1051 : | * | ||
1052 : | * Returned values : XVID_ERR_FAIL when opt is invalid | ||
1053 : | * else returns the wrapped function result | ||
1054 : | * | ||
1055 : | ****************************************************************************/ | ||
1056 : | |||
1057 : | edgomez | 195 | int |
1058 : | xvid_encore(void *handle, | ||
1059 : | int opt, | ||
1060 : | void *param1, | ||
1061 : | void *param2) | ||
1062 : | Isibaar | 3 | { |
1063 : | edgomez | 195 | switch (opt) { |
1064 : | case XVID_ENC_ENCODE: | ||
1065 : | edgomez | 851 | |
1066 : | suxen_drol | 890 | return enc_encode((Encoder *) handle, |
1067 : | (xvid_enc_frame_t *) param1, | ||
1068 : | (xvid_enc_stats_t *) param2); | ||
1069 : | Isibaar | 3 | |
1070 : | edgomez | 195 | case XVID_ENC_CREATE: |
1071 : | suxen_drol | 948 | return enc_create((xvid_enc_create_t *) param1); |
1072 : | Isibaar | 3 | |
1073 : | edgomez | 195 | case XVID_ENC_DESTROY: |
1074 : | suxen_drol | 890 | return enc_destroy((Encoder *) handle); |
1075 : | edgomez | 195 | |
1076 : | Isibaar | 3 | default: |
1077 : | edgomez | 195 | return XVID_ERR_FAIL; |
1078 : | } | ||
1079 : | Isibaar | 3 | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |