Mercurial > libavcodec.hg
annotate dsputil.c @ 296:c1a8a1b4a24b libavcodec
sizeof(s->block) isnt 64*6*2 anymore bugfix
mpeg12 decoding optimization
author | michaelni |
---|---|
date | Thu, 28 Mar 2002 13:41:04 +0000 |
parents | 944632089814 |
children | 8cf5507e6ca5 |
rev | line source |
---|---|
0 | 1 /* |
2 * DSP utils | |
3 * Copyright (c) 2000, 2001 Gerard Lantau. | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
256 | 18 * |
19 * gmc & q-pel support by Michael Niedermayer <michaelni@gmx.at> | |
0 | 20 */ |
21 #include <stdlib.h> | |
22 #include <stdio.h> | |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
23 #include <math.h> |
0 | 24 #include "avcodec.h" |
25 #include "dsputil.h" | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
26 #include "simple_idct.h" |
0 | 27 |
19
82d4c9be9873
MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents:
6
diff
changeset
|
28 void (*ff_idct)(DCTELEM *block); |
0 | 29 void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size); |
30 void (*put_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size); | |
31 void (*add_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size); | |
255 | 32 void (*gmc1)(UINT8 *dst, UINT8 *src, int srcStride, int h, int x16, int y16, int rounder); |
296 | 33 void (*clear_blocks)(DCTELEM *blocks); |
0 | 34 |
35 op_pixels_abs_func pix_abs16x16; | |
36 op_pixels_abs_func pix_abs16x16_x2; | |
37 op_pixels_abs_func pix_abs16x16_y2; | |
38 op_pixels_abs_func pix_abs16x16_xy2; | |
39 | |
294 | 40 op_pixels_abs_func pix_abs8x8; |
41 op_pixels_abs_func pix_abs8x8_x2; | |
42 op_pixels_abs_func pix_abs8x8_y2; | |
43 op_pixels_abs_func pix_abs8x8_xy2; | |
44 | |
50 | 45 UINT8 cropTbl[256 + 2 * MAX_NEG_CROP]; |
0 | 46 UINT32 squareTbl[512]; |
47 | |
34 | 48 extern UINT16 default_intra_matrix[64]; |
49 extern UINT16 default_non_intra_matrix[64]; | |
50 | |
51 UINT8 zigzag_direct[64] = { | |
52 0, 1, 8, 16, 9, 2, 3, 10, | |
53 17, 24, 32, 25, 18, 11, 4, 5, | |
54 12, 19, 26, 33, 40, 48, 41, 34, | |
55 27, 20, 13, 6, 7, 14, 21, 28, | |
56 35, 42, 49, 56, 57, 50, 43, 36, | |
57 29, 22, 15, 23, 30, 37, 44, 51, | |
58 58, 59, 52, 45, 38, 31, 39, 46, | |
59 53, 60, 61, 54, 47, 55, 62, 63 | |
60 }; | |
61 | |
220 | 62 /* not permutated inverse zigzag_direct + 1 for MMX quantizer */ |
63 UINT16 __align8 inv_zigzag_direct16[64]; | |
64 | |
65 /* not permutated zigzag_direct for MMX quantizer */ | |
66 UINT8 zigzag_direct_noperm[64]; | |
67 | |
34 | 68 UINT8 ff_alternate_horizontal_scan[64] = { |
69 0, 1, 2, 3, 8, 9, 16, 17, | |
70 10, 11, 4, 5, 6, 7, 15, 14, | |
71 13, 12, 19, 18, 24, 25, 32, 33, | |
72 26, 27, 20, 21, 22, 23, 28, 29, | |
73 30, 31, 34, 35, 40, 41, 48, 49, | |
74 42, 43, 36, 37, 38, 39, 44, 45, | |
75 46, 47, 50, 51, 56, 57, 58, 59, | |
76 52, 53, 54, 55, 60, 61, 62, 63, | |
77 }; | |
78 | |
79 UINT8 ff_alternate_vertical_scan[64] = { | |
80 0, 8, 16, 24, 1, 9, 2, 10, | |
81 17, 25, 32, 40, 48, 56, 57, 49, | |
82 41, 33, 26, 18, 3, 11, 4, 12, | |
83 19, 27, 34, 42, 50, 58, 35, 43, | |
84 51, 59, 20, 28, 5, 13, 6, 14, | |
85 21, 29, 36, 44, 52, 60, 37, 45, | |
86 53, 61, 22, 30, 7, 15, 23, 31, | |
87 38, 46, 54, 62, 39, 47, 55, 63, | |
88 }; | |
89 | |
209 | 90 /* Input permutation for the simple_idct_mmx */ |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
91 static UINT8 simple_mmx_permutation[64]={ |
209 | 92 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D, |
93 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D, | |
94 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D, | |
95 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F, | |
96 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F, | |
97 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D, | |
98 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F, | |
99 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F, | |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
100 }; |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
101 |
220 | 102 /* a*inverse[b]>>32 == a/b for all 0<=a<=65536 && 2<=b<=255 */ |
103 UINT32 inverse[256]={ | |
104 0, 4294967295U,2147483648U,1431655766, 1073741824, 858993460, 715827883, 613566757, | |
105 536870912, 477218589, 429496730, 390451573, 357913942, 330382100, 306783379, 286331154, | |
106 268435456, 252645136, 238609295, 226050911, 214748365, 204522253, 195225787, 186737709, | |
107 178956971, 171798692, 165191050, 159072863, 153391690, 148102321, 143165577, 138547333, | |
108 134217728, 130150525, 126322568, 122713352, 119304648, 116080198, 113025456, 110127367, | |
109 107374183, 104755300, 102261127, 99882961, 97612894, 95443718, 93368855, 91382283, | |
110 89478486, 87652394, 85899346, 84215046, 82595525, 81037119, 79536432, 78090315, | |
111 76695845, 75350304, 74051161, 72796056, 71582789, 70409300, 69273667, 68174085, | |
112 67108864, 66076420, 65075263, 64103990, 63161284, 62245903, 61356676, 60492498, | |
113 59652324, 58835169, 58040099, 57266231, 56512728, 55778797, 55063684, 54366675, | |
114 53687092, 53024288, 52377650, 51746594, 51130564, 50529028, 49941481, 49367441, | |
115 48806447, 48258060, 47721859, 47197443, 46684428, 46182445, 45691142, 45210183, | |
116 44739243, 44278014, 43826197, 43383509, 42949673, 42524429, 42107523, 41698712, | |
117 41297763, 40904451, 40518560, 40139882, 39768216, 39403370, 39045158, 38693400, | |
118 38347923, 38008561, 37675152, 37347542, 37025581, 36709123, 36398028, 36092163, | |
119 35791395, 35495598, 35204650, 34918434, 34636834, 34359739, 34087043, 33818641, | |
120 33554432, 33294321, 33038210, 32786010, 32537632, 32292988, 32051995, 31814573, | |
121 31580642, 31350127, 31122952, 30899046, 30678338, 30460761, 30246249, 30034737, | |
122 29826162, 29620465, 29417585, 29217465, 29020050, 28825284, 28633116, 28443493, | |
123 28256364, 28071682, 27889399, 27709467, 27531842, 27356480, 27183338, 27012373, | |
124 26843546, 26676816, 26512144, 26349493, 26188825, 26030105, 25873297, 25718368, | |
125 25565282, 25414008, 25264514, 25116768, 24970741, 24826401, 24683721, 24542671, | |
126 24403224, 24265352, 24129030, 23994231, 23860930, 23729102, 23598722, 23469767, | |
127 23342214, 23216040, 23091223, 22967740, 22845571, 22724695, 22605092, 22486740, | |
128 22369622, 22253717, 22139007, 22025474, 21913099, 21801865, 21691755, 21582751, | |
129 21474837, 21367997, 21262215, 21157475, 21053762, 20951060, 20849356, 20748635, | |
130 20648882, 20550083, 20452226, 20355296, 20259280, 20164166, 20069941, 19976593, | |
131 19884108, 19792477, 19701685, 19611723, 19522579, 19434242, 19346700, 19259944, | |
132 19173962, 19088744, 19004281, 18920561, 18837576, 18755316, 18673771, 18592933, | |
133 18512791, 18433337, 18354562, 18276457, 18199014, 18122225, 18046082, 17970575, | |
134 17895698, 17821442, 17747799, 17674763, 17602325, 17530479, 17459217, 17388532, | |
135 17318417, 17248865, 17179870, 17111424, 17043522, 16976156, 16909321, 16843010, | |
136 }; | |
137 | |
200 | 138 /* used to skip zeros at the end */ |
139 UINT8 zigzag_end[64]; | |
140 | |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
141 UINT8 permutation[64]; |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
142 //UINT8 invPermutation[64]; |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
143 |
200 | 144 static void build_zigzag_end() |
145 { | |
146 int lastIndex; | |
147 int lastIndexAfterPerm=0; | |
148 for(lastIndex=0; lastIndex<64; lastIndex++) | |
149 { | |
150 if(zigzag_direct[lastIndex] > lastIndexAfterPerm) | |
151 lastIndexAfterPerm= zigzag_direct[lastIndex]; | |
152 zigzag_end[lastIndex]= lastIndexAfterPerm + 1; | |
153 } | |
154 } | |
155 | |
0 | 156 void get_pixels_c(DCTELEM *block, const UINT8 *pixels, int line_size) |
157 { | |
158 DCTELEM *p; | |
159 const UINT8 *pix; | |
160 int i; | |
161 | |
162 /* read the pixels */ | |
163 p = block; | |
164 pix = pixels; | |
165 for(i=0;i<8;i++) { | |
166 p[0] = pix[0]; | |
167 p[1] = pix[1]; | |
168 p[2] = pix[2]; | |
169 p[3] = pix[3]; | |
170 p[4] = pix[4]; | |
171 p[5] = pix[5]; | |
172 p[6] = pix[6]; | |
173 p[7] = pix[7]; | |
174 pix += line_size; | |
175 p += 8; | |
176 } | |
177 } | |
178 | |
179 void put_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size) | |
180 { | |
181 const DCTELEM *p; | |
182 UINT8 *pix; | |
183 int i; | |
184 UINT8 *cm = cropTbl + MAX_NEG_CROP; | |
185 | |
186 /* read the pixels */ | |
187 p = block; | |
188 pix = pixels; | |
189 for(i=0;i<8;i++) { | |
190 pix[0] = cm[p[0]]; | |
191 pix[1] = cm[p[1]]; | |
192 pix[2] = cm[p[2]]; | |
193 pix[3] = cm[p[3]]; | |
194 pix[4] = cm[p[4]]; | |
195 pix[5] = cm[p[5]]; | |
196 pix[6] = cm[p[6]]; | |
197 pix[7] = cm[p[7]]; | |
198 pix += line_size; | |
199 p += 8; | |
200 } | |
201 } | |
202 | |
203 void add_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size) | |
204 { | |
205 const DCTELEM *p; | |
206 UINT8 *pix; | |
207 int i; | |
208 UINT8 *cm = cropTbl + MAX_NEG_CROP; | |
209 | |
210 /* read the pixels */ | |
211 p = block; | |
212 pix = pixels; | |
213 for(i=0;i<8;i++) { | |
214 pix[0] = cm[pix[0] + p[0]]; | |
215 pix[1] = cm[pix[1] + p[1]]; | |
216 pix[2] = cm[pix[2] + p[2]]; | |
217 pix[3] = cm[pix[3] + p[3]]; | |
218 pix[4] = cm[pix[4] + p[4]]; | |
219 pix[5] = cm[pix[5] + p[5]]; | |
220 pix[6] = cm[pix[6] + p[6]]; | |
221 pix[7] = cm[pix[7] + p[7]]; | |
222 pix += line_size; | |
223 p += 8; | |
224 } | |
225 } | |
226 | |
227 #define PIXOP(BTYPE, OPNAME, OP, INCR) \ | |
228 \ | |
229 static void OPNAME ## _pixels(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
230 { \ | |
231 BTYPE *p; \ | |
232 const UINT8 *pix; \ | |
233 \ | |
234 p = block; \ | |
235 pix = pixels; \ | |
236 do { \ | |
237 OP(p[0], pix[0]); \ | |
238 OP(p[1], pix[1]); \ | |
239 OP(p[2], pix[2]); \ | |
240 OP(p[3], pix[3]); \ | |
241 OP(p[4], pix[4]); \ | |
242 OP(p[5], pix[5]); \ | |
243 OP(p[6], pix[6]); \ | |
244 OP(p[7], pix[7]); \ | |
245 pix += line_size; \ | |
246 p += INCR; \ | |
247 } while (--h);; \ | |
248 } \ | |
249 \ | |
250 static void OPNAME ## _pixels_x2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
251 { \ | |
252 BTYPE *p; \ | |
253 const UINT8 *pix; \ | |
254 \ | |
255 p = block; \ | |
256 pix = pixels; \ | |
257 do { \ | |
258 OP(p[0], avg2(pix[0], pix[1])); \ | |
259 OP(p[1], avg2(pix[1], pix[2])); \ | |
260 OP(p[2], avg2(pix[2], pix[3])); \ | |
261 OP(p[3], avg2(pix[3], pix[4])); \ | |
262 OP(p[4], avg2(pix[4], pix[5])); \ | |
263 OP(p[5], avg2(pix[5], pix[6])); \ | |
264 OP(p[6], avg2(pix[6], pix[7])); \ | |
265 OP(p[7], avg2(pix[7], pix[8])); \ | |
266 pix += line_size; \ | |
267 p += INCR; \ | |
268 } while (--h); \ | |
269 } \ | |
270 \ | |
271 static void OPNAME ## _pixels_y2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
272 { \ | |
273 BTYPE *p; \ | |
274 const UINT8 *pix; \ | |
275 const UINT8 *pix1; \ | |
276 \ | |
277 p = block; \ | |
278 pix = pixels; \ | |
279 pix1 = pixels + line_size; \ | |
280 do { \ | |
281 OP(p[0], avg2(pix[0], pix1[0])); \ | |
282 OP(p[1], avg2(pix[1], pix1[1])); \ | |
283 OP(p[2], avg2(pix[2], pix1[2])); \ | |
284 OP(p[3], avg2(pix[3], pix1[3])); \ | |
285 OP(p[4], avg2(pix[4], pix1[4])); \ | |
286 OP(p[5], avg2(pix[5], pix1[5])); \ | |
287 OP(p[6], avg2(pix[6], pix1[6])); \ | |
288 OP(p[7], avg2(pix[7], pix1[7])); \ | |
289 pix += line_size; \ | |
290 pix1 += line_size; \ | |
291 p += INCR; \ | |
292 } while(--h); \ | |
293 } \ | |
294 \ | |
295 static void OPNAME ## _pixels_xy2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
296 { \ | |
297 BTYPE *p; \ | |
298 const UINT8 *pix; \ | |
299 const UINT8 *pix1; \ | |
300 \ | |
301 p = block; \ | |
302 pix = pixels; \ | |
303 pix1 = pixels + line_size; \ | |
304 do { \ | |
305 OP(p[0], avg4(pix[0], pix[1], pix1[0], pix1[1])); \ | |
306 OP(p[1], avg4(pix[1], pix[2], pix1[1], pix1[2])); \ | |
307 OP(p[2], avg4(pix[2], pix[3], pix1[2], pix1[3])); \ | |
308 OP(p[3], avg4(pix[3], pix[4], pix1[3], pix1[4])); \ | |
309 OP(p[4], avg4(pix[4], pix[5], pix1[4], pix1[5])); \ | |
310 OP(p[5], avg4(pix[5], pix[6], pix1[5], pix1[6])); \ | |
311 OP(p[6], avg4(pix[6], pix[7], pix1[6], pix1[7])); \ | |
312 OP(p[7], avg4(pix[7], pix[8], pix1[7], pix1[8])); \ | |
313 pix += line_size; \ | |
314 pix1 += line_size; \ | |
315 p += INCR; \ | |
316 } while(--h); \ | |
317 } \ | |
318 \ | |
319 void (*OPNAME ## _pixels_tab[4])(BTYPE *block, const UINT8 *pixels, int line_size, int h) = { \ | |
320 OPNAME ## _pixels, \ | |
321 OPNAME ## _pixels_x2, \ | |
322 OPNAME ## _pixels_y2, \ | |
323 OPNAME ## _pixels_xy2, \ | |
324 }; | |
325 | |
326 | |
327 /* rounding primitives */ | |
328 #define avg2(a,b) ((a+b+1)>>1) | |
329 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2) | |
330 | |
331 #define op_put(a, b) a = b | |
332 #define op_avg(a, b) a = avg2(a, b) | |
333 #define op_sub(a, b) a -= b | |
334 | |
335 PIXOP(UINT8, put, op_put, line_size) | |
336 PIXOP(UINT8, avg, op_avg, line_size) | |
337 | |
338 PIXOP(DCTELEM, sub, op_sub, 8) | |
339 | |
340 /* not rounding primitives */ | |
341 #undef avg2 | |
342 #undef avg4 | |
343 #define avg2(a,b) ((a+b)>>1) | |
344 #define avg4(a,b,c,d) ((a+b+c+d+1)>>2) | |
345 | |
346 PIXOP(UINT8, put_no_rnd, op_put, line_size) | |
347 PIXOP(UINT8, avg_no_rnd, op_avg, line_size) | |
348 | |
349 /* motion estimation */ | |
350 | |
351 #undef avg2 | |
352 #undef avg4 | |
353 #define avg2(a,b) ((a+b+1)>>1) | |
354 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2) | |
355 | |
255 | 356 static void gmc1_c(UINT8 *dst, UINT8 *src, int srcStride, int h, int x16, int y16, int rounder) |
357 { | |
358 const int A=(16-x16)*(16-y16); | |
359 const int B=( x16)*(16-y16); | |
360 const int C=(16-x16)*( y16); | |
361 const int D=( x16)*( y16); | |
362 int i; | |
363 rounder= 128 - rounder; | |
364 | |
365 for(i=0; i<h; i++) | |
366 { | |
367 dst[0]= (A*src[0] + B*src[1] + C*src[srcStride+0] + D*src[srcStride+1] + rounder)>>8; | |
368 dst[1]= (A*src[1] + B*src[2] + C*src[srcStride+1] + D*src[srcStride+2] + rounder)>>8; | |
369 dst[2]= (A*src[2] + B*src[3] + C*src[srcStride+2] + D*src[srcStride+3] + rounder)>>8; | |
370 dst[3]= (A*src[3] + B*src[4] + C*src[srcStride+3] + D*src[srcStride+4] + rounder)>>8; | |
371 dst[4]= (A*src[4] + B*src[5] + C*src[srcStride+4] + D*src[srcStride+5] + rounder)>>8; | |
372 dst[5]= (A*src[5] + B*src[6] + C*src[srcStride+5] + D*src[srcStride+6] + rounder)>>8; | |
373 dst[6]= (A*src[6] + B*src[7] + C*src[srcStride+6] + D*src[srcStride+7] + rounder)>>8; | |
374 dst[7]= (A*src[7] + B*src[8] + C*src[srcStride+7] + D*src[srcStride+8] + rounder)>>8; | |
375 dst+= srcStride; | |
376 src+= srcStride; | |
377 } | |
378 } | |
379 | |
380 static void qpel_h_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int h, int r) | |
381 { | |
382 UINT8 *cm = cropTbl + MAX_NEG_CROP; | |
383 int i; | |
384 for(i=0; i<h; i++) | |
385 { | |
294 | 386 dst[0]= cm[(((src[0]+src[1])*20 - (src[0]+src[2])*6 + (src[1]+src[3])*3 - (src[2]+src[4]) + r)>>5)]; |
387 dst[1]= cm[(((src[1]+src[2])*20 - (src[0]+src[3])*6 + (src[0]+src[4])*3 - (src[1]+src[5]) + r)>>5)]; | |
388 dst[2]= cm[(((src[2]+src[3])*20 - (src[1]+src[4])*6 + (src[0]+src[5])*3 - (src[0]+src[6]) + r)>>5)]; | |
389 dst[3]= cm[(((src[3]+src[4])*20 - (src[2]+src[5])*6 + (src[1]+src[6])*3 - (src[0]+src[7]) + r)>>5)]; | |
390 dst[4]= cm[(((src[4]+src[5])*20 - (src[3]+src[6])*6 + (src[2]+src[7])*3 - (src[1]+src[8]) + r)>>5)]; | |
391 dst[5]= cm[(((src[5]+src[6])*20 - (src[4]+src[7])*6 + (src[3]+src[8])*3 - (src[2]+src[8]) + r)>>5)]; | |
392 dst[6]= cm[(((src[6]+src[7])*20 - (src[5]+src[8])*6 + (src[4]+src[8])*3 - (src[3]+src[7]) + r)>>5)]; | |
393 dst[7]= cm[(((src[7]+src[8])*20 - (src[6]+src[8])*6 + (src[5]+src[7])*3 - (src[4]+src[6]) + r)>>5)]; | |
255 | 394 dst+=dstStride; |
395 src+=srcStride; | |
396 } | |
397 } | |
398 | |
399 static void qpel_v_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int w, int r) | |
400 { | |
401 UINT8 *cm = cropTbl + MAX_NEG_CROP; | |
402 int i; | |
403 for(i=0; i<w; i++) | |
404 { | |
405 const int src0= src[0*srcStride]; | |
406 const int src1= src[1*srcStride]; | |
407 const int src2= src[2*srcStride]; | |
408 const int src3= src[3*srcStride]; | |
409 const int src4= src[4*srcStride]; | |
410 const int src5= src[5*srcStride]; | |
411 const int src6= src[6*srcStride]; | |
412 const int src7= src[7*srcStride]; | |
413 const int src8= src[8*srcStride]; | |
294 | 414 dst[0*dstStride]= cm[(((src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4) + r)>>5)]; |
415 dst[1*dstStride]= cm[(((src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5) + r)>>5)]; | |
416 dst[2*dstStride]= cm[(((src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6) + r)>>5)]; | |
417 dst[3*dstStride]= cm[(((src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7) + r)>>5)]; | |
418 dst[4*dstStride]= cm[(((src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8) + r)>>5)]; | |
419 dst[5*dstStride]= cm[(((src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8) + r)>>5)]; | |
420 dst[6*dstStride]= cm[(((src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7) + r)>>5)]; | |
421 dst[7*dstStride]= cm[(((src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6) + r)>>5)]; | |
255 | 422 dst++; |
423 src++; | |
424 } | |
425 } | |
426 | |
427 static inline void put_block(UINT8 *dst, UINT8 *src, int dstStride, int srcStride) | |
428 { | |
429 int i; | |
430 for(i=0; i<8; i++) | |
431 { | |
432 dst[0]= src[0]; | |
433 dst[1]= src[1]; | |
434 dst[2]= src[2]; | |
435 dst[3]= src[3]; | |
436 dst[4]= src[4]; | |
437 dst[5]= src[5]; | |
438 dst[6]= src[6]; | |
439 dst[7]= src[7]; | |
440 dst+=dstStride; | |
441 src+=srcStride; | |
442 } | |
443 } | |
444 | |
445 static inline void avg2_block(UINT8 *dst, UINT8 *src1, UINT8 *src2, int dstStride, int srcStride, int r) | |
446 { | |
447 int i; | |
448 for(i=0; i<8; i++) | |
449 { | |
450 dst[0]= (src1[0] + src2[0] + r)>>1; | |
451 dst[1]= (src1[1] + src2[1] + r)>>1; | |
452 dst[2]= (src1[2] + src2[2] + r)>>1; | |
453 dst[3]= (src1[3] + src2[3] + r)>>1; | |
454 dst[4]= (src1[4] + src2[4] + r)>>1; | |
455 dst[5]= (src1[5] + src2[5] + r)>>1; | |
456 dst[6]= (src1[6] + src2[6] + r)>>1; | |
457 dst[7]= (src1[7] + src2[7] + r)>>1; | |
458 dst+=dstStride; | |
459 src1+=srcStride; | |
460 src2+=8; | |
461 } | |
462 } | |
463 | |
464 static inline void avg4_block(UINT8 *dst, UINT8 *src1, UINT8 *src2, UINT8 *src3, UINT8 *src4, int dstStride, int srcStride, int r) | |
465 { | |
466 int i; | |
467 for(i=0; i<8; i++) | |
468 { | |
469 dst[0]= (src1[0] + src2[0] + src3[0] + src4[0] + r)>>2; | |
470 dst[1]= (src1[1] + src2[1] + src3[1] + src4[1] + r)>>2; | |
471 dst[2]= (src1[2] + src2[2] + src3[2] + src4[2] + r)>>2; | |
472 dst[3]= (src1[3] + src2[3] + src3[3] + src4[3] + r)>>2; | |
473 dst[4]= (src1[4] + src2[4] + src3[4] + src4[4] + r)>>2; | |
474 dst[5]= (src1[5] + src2[5] + src3[5] + src4[5] + r)>>2; | |
475 dst[6]= (src1[6] + src2[6] + src3[6] + src4[6] + r)>>2; | |
476 dst[7]= (src1[7] + src2[7] + src3[7] + src4[7] + r)>>2; | |
477 dst+=dstStride; | |
478 src1+=srcStride; | |
479 src2+=8; | |
256 | 480 src3+=8; |
255 | 481 src4+=8; |
482 } | |
483 } | |
484 | |
485 #define QPEL_MC(r, name) \ | |
486 static void qpel_mc00_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
487 {\ | |
488 put_block(dst, src, dstStride, srcStride);\ | |
489 }\ | |
490 \ | |
491 static void qpel_mc10_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
492 {\ | |
493 UINT8 half[64];\ | |
294 | 494 qpel_h_lowpass(half, src, 8, srcStride, 8, 16-r);\ |
255 | 495 avg2_block(dst, src, half, dstStride, srcStride, 1-r);\ |
496 }\ | |
497 \ | |
498 static void qpel_mc20_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
499 {\ | |
294 | 500 qpel_h_lowpass(dst, src, dstStride, srcStride, 8, 16-r);\ |
255 | 501 }\ |
502 \ | |
503 static void qpel_mc30_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
504 {\ | |
505 UINT8 half[64];\ | |
294 | 506 qpel_h_lowpass(half, src, 8, srcStride, 8, 16-r);\ |
255 | 507 avg2_block(dst, src+1, half, dstStride, srcStride, 1-r);\ |
508 }\ | |
509 \ | |
510 static void qpel_mc01_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
511 {\ | |
512 UINT8 half[64];\ | |
294 | 513 qpel_v_lowpass(half, src, 8, srcStride, 8, 16-r);\ |
255 | 514 avg2_block(dst, src, half, dstStride, srcStride, 1-r);\ |
515 }\ | |
516 \ | |
517 static void qpel_mc02_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
518 {\ | |
294 | 519 qpel_v_lowpass(dst, src, dstStride, srcStride, 8, 16-r);\ |
255 | 520 }\ |
521 \ | |
522 static void qpel_mc03_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
523 {\ | |
524 UINT8 half[64];\ | |
294 | 525 qpel_v_lowpass(half, src, 8, srcStride, 8, 16-r);\ |
255 | 526 avg2_block(dst, src+srcStride, half, dstStride, srcStride, 1-r);\ |
527 }\ | |
528 static void qpel_mc11_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
529 {\ | |
530 UINT8 halfH[72];\ | |
256 | 531 UINT8 halfV[64];\ |
255 | 532 UINT8 halfHV[64];\ |
294 | 533 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
534 qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\ | |
535 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
255 | 536 avg4_block(dst, src, halfH, halfV, halfHV, dstStride, srcStride, 2-r);\ |
537 }\ | |
538 static void qpel_mc31_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
539 {\ | |
540 UINT8 halfH[72];\ | |
256 | 541 UINT8 halfV[64];\ |
255 | 542 UINT8 halfHV[64];\ |
294 | 543 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
544 qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\ | |
545 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
255 | 546 avg4_block(dst, src+1, halfH, halfV, halfHV, dstStride, srcStride, 2-r);\ |
547 }\ | |
548 static void qpel_mc13_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
549 {\ | |
550 UINT8 halfH[72];\ | |
256 | 551 UINT8 halfV[64];\ |
255 | 552 UINT8 halfHV[64];\ |
294 | 553 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
554 qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\ | |
555 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
256 | 556 avg4_block(dst, src+srcStride, halfH+8, halfV, halfHV, dstStride, srcStride, 2-r);\ |
255 | 557 }\ |
558 static void qpel_mc33_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
559 {\ | |
560 UINT8 halfH[72];\ | |
256 | 561 UINT8 halfV[64];\ |
255 | 562 UINT8 halfHV[64];\ |
294 | 563 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
564 qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\ | |
565 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
256 | 566 avg4_block(dst, src+srcStride+1, halfH+8, halfV, halfHV, dstStride, srcStride, 2-r);\ |
255 | 567 }\ |
568 static void qpel_mc21_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
569 {\ | |
570 UINT8 halfH[72];\ | |
571 UINT8 halfHV[64];\ | |
294 | 572 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
573 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
255 | 574 avg2_block(dst, halfH, halfHV, dstStride, 8, 1-r);\ |
575 }\ | |
576 static void qpel_mc23_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
577 {\ | |
578 UINT8 halfH[72];\ | |
579 UINT8 halfHV[64];\ | |
294 | 580 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
581 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
255 | 582 avg2_block(dst, halfH+8, halfHV, dstStride, 8, 1-r);\ |
583 }\ | |
584 static void qpel_mc12_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
585 {\ | |
586 UINT8 halfH[72];\ | |
256 | 587 UINT8 halfV[64];\ |
255 | 588 UINT8 halfHV[64];\ |
294 | 589 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
590 qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\ | |
591 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
256 | 592 avg2_block(dst, halfV, halfHV, dstStride, 8, 1-r);\ |
255 | 593 }\ |
594 static void qpel_mc32_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
595 {\ | |
596 UINT8 halfH[72];\ | |
256 | 597 UINT8 halfV[64];\ |
255 | 598 UINT8 halfHV[64];\ |
294 | 599 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
600 qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\ | |
601 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\ | |
256 | 602 avg2_block(dst, halfV, halfHV, dstStride, 8, 1-r);\ |
255 | 603 }\ |
604 static void qpel_mc22_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\ | |
605 {\ | |
606 UINT8 halfH[72];\ | |
294 | 607 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\ |
608 qpel_v_lowpass(dst, halfH, dstStride, 8, 8, 16-r);\ | |
255 | 609 }\ |
610 qpel_mc_func qpel_mc ## name ## _tab[16]={ \ | |
611 qpel_mc00_c ## name, \ | |
612 qpel_mc10_c ## name, \ | |
613 qpel_mc20_c ## name, \ | |
614 qpel_mc30_c ## name, \ | |
615 qpel_mc01_c ## name, \ | |
616 qpel_mc11_c ## name, \ | |
617 qpel_mc21_c ## name, \ | |
618 qpel_mc31_c ## name, \ | |
619 qpel_mc02_c ## name, \ | |
620 qpel_mc12_c ## name, \ | |
621 qpel_mc22_c ## name, \ | |
622 qpel_mc32_c ## name, \ | |
623 qpel_mc03_c ## name, \ | |
624 qpel_mc13_c ## name, \ | |
625 qpel_mc23_c ## name, \ | |
626 qpel_mc33_c ## name, \ | |
627 }; | |
628 | |
629 QPEL_MC(0, _rnd) | |
630 QPEL_MC(1, _no_rnd) | |
631 | |
294 | 632 int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size) |
0 | 633 { |
634 int s, i; | |
635 | |
636 s = 0; | |
294 | 637 for(i=0;i<16;i++) { |
0 | 638 s += abs(pix1[0] - pix2[0]); |
639 s += abs(pix1[1] - pix2[1]); | |
640 s += abs(pix1[2] - pix2[2]); | |
641 s += abs(pix1[3] - pix2[3]); | |
642 s += abs(pix1[4] - pix2[4]); | |
643 s += abs(pix1[5] - pix2[5]); | |
644 s += abs(pix1[6] - pix2[6]); | |
645 s += abs(pix1[7] - pix2[7]); | |
646 s += abs(pix1[8] - pix2[8]); | |
647 s += abs(pix1[9] - pix2[9]); | |
648 s += abs(pix1[10] - pix2[10]); | |
649 s += abs(pix1[11] - pix2[11]); | |
650 s += abs(pix1[12] - pix2[12]); | |
651 s += abs(pix1[13] - pix2[13]); | |
652 s += abs(pix1[14] - pix2[14]); | |
653 s += abs(pix1[15] - pix2[15]); | |
654 pix1 += line_size; | |
655 pix2 += line_size; | |
656 } | |
657 return s; | |
658 } | |
659 | |
294 | 660 int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size) |
0 | 661 { |
662 int s, i; | |
663 | |
664 s = 0; | |
294 | 665 for(i=0;i<16;i++) { |
0 | 666 s += abs(pix1[0] - avg2(pix2[0], pix2[1])); |
667 s += abs(pix1[1] - avg2(pix2[1], pix2[2])); | |
668 s += abs(pix1[2] - avg2(pix2[2], pix2[3])); | |
669 s += abs(pix1[3] - avg2(pix2[3], pix2[4])); | |
670 s += abs(pix1[4] - avg2(pix2[4], pix2[5])); | |
671 s += abs(pix1[5] - avg2(pix2[5], pix2[6])); | |
672 s += abs(pix1[6] - avg2(pix2[6], pix2[7])); | |
673 s += abs(pix1[7] - avg2(pix2[7], pix2[8])); | |
674 s += abs(pix1[8] - avg2(pix2[8], pix2[9])); | |
675 s += abs(pix1[9] - avg2(pix2[9], pix2[10])); | |
676 s += abs(pix1[10] - avg2(pix2[10], pix2[11])); | |
677 s += abs(pix1[11] - avg2(pix2[11], pix2[12])); | |
678 s += abs(pix1[12] - avg2(pix2[12], pix2[13])); | |
679 s += abs(pix1[13] - avg2(pix2[13], pix2[14])); | |
680 s += abs(pix1[14] - avg2(pix2[14], pix2[15])); | |
681 s += abs(pix1[15] - avg2(pix2[15], pix2[16])); | |
682 pix1 += line_size; | |
683 pix2 += line_size; | |
684 } | |
685 return s; | |
686 } | |
687 | |
294 | 688 int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size) |
0 | 689 { |
690 int s, i; | |
691 UINT8 *pix3 = pix2 + line_size; | |
692 | |
693 s = 0; | |
294 | 694 for(i=0;i<16;i++) { |
0 | 695 s += abs(pix1[0] - avg2(pix2[0], pix3[0])); |
696 s += abs(pix1[1] - avg2(pix2[1], pix3[1])); | |
697 s += abs(pix1[2] - avg2(pix2[2], pix3[2])); | |
698 s += abs(pix1[3] - avg2(pix2[3], pix3[3])); | |
699 s += abs(pix1[4] - avg2(pix2[4], pix3[4])); | |
700 s += abs(pix1[5] - avg2(pix2[5], pix3[5])); | |
701 s += abs(pix1[6] - avg2(pix2[6], pix3[6])); | |
702 s += abs(pix1[7] - avg2(pix2[7], pix3[7])); | |
703 s += abs(pix1[8] - avg2(pix2[8], pix3[8])); | |
704 s += abs(pix1[9] - avg2(pix2[9], pix3[9])); | |
705 s += abs(pix1[10] - avg2(pix2[10], pix3[10])); | |
706 s += abs(pix1[11] - avg2(pix2[11], pix3[11])); | |
707 s += abs(pix1[12] - avg2(pix2[12], pix3[12])); | |
708 s += abs(pix1[13] - avg2(pix2[13], pix3[13])); | |
709 s += abs(pix1[14] - avg2(pix2[14], pix3[14])); | |
710 s += abs(pix1[15] - avg2(pix2[15], pix3[15])); | |
711 pix1 += line_size; | |
712 pix2 += line_size; | |
713 pix3 += line_size; | |
714 } | |
715 return s; | |
716 } | |
717 | |
294 | 718 int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size) |
0 | 719 { |
720 int s, i; | |
721 UINT8 *pix3 = pix2 + line_size; | |
722 | |
723 s = 0; | |
294 | 724 for(i=0;i<16;i++) { |
0 | 725 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1])); |
726 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2])); | |
727 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3])); | |
728 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4])); | |
729 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5])); | |
730 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6])); | |
731 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7])); | |
732 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8])); | |
733 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9])); | |
734 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10])); | |
735 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11])); | |
736 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12])); | |
737 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13])); | |
738 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14])); | |
739 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15])); | |
740 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16])); | |
741 pix1 += line_size; | |
742 pix2 += line_size; | |
743 pix3 += line_size; | |
744 } | |
745 return s; | |
746 } | |
747 | |
294 | 748 int pix_abs8x8_c(UINT8 *pix1, UINT8 *pix2, int line_size) |
749 { | |
750 int s, i; | |
751 | |
752 s = 0; | |
753 for(i=0;i<8;i++) { | |
754 s += abs(pix1[0] - pix2[0]); | |
755 s += abs(pix1[1] - pix2[1]); | |
756 s += abs(pix1[2] - pix2[2]); | |
757 s += abs(pix1[3] - pix2[3]); | |
758 s += abs(pix1[4] - pix2[4]); | |
759 s += abs(pix1[5] - pix2[5]); | |
760 s += abs(pix1[6] - pix2[6]); | |
761 s += abs(pix1[7] - pix2[7]); | |
762 pix1 += line_size; | |
763 pix2 += line_size; | |
764 } | |
765 return s; | |
766 } | |
767 | |
768 int pix_abs8x8_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size) | |
769 { | |
770 int s, i; | |
771 | |
772 s = 0; | |
773 for(i=0;i<8;i++) { | |
774 s += abs(pix1[0] - avg2(pix2[0], pix2[1])); | |
775 s += abs(pix1[1] - avg2(pix2[1], pix2[2])); | |
776 s += abs(pix1[2] - avg2(pix2[2], pix2[3])); | |
777 s += abs(pix1[3] - avg2(pix2[3], pix2[4])); | |
778 s += abs(pix1[4] - avg2(pix2[4], pix2[5])); | |
779 s += abs(pix1[5] - avg2(pix2[5], pix2[6])); | |
780 s += abs(pix1[6] - avg2(pix2[6], pix2[7])); | |
781 s += abs(pix1[7] - avg2(pix2[7], pix2[8])); | |
782 pix1 += line_size; | |
783 pix2 += line_size; | |
784 } | |
785 return s; | |
786 } | |
787 | |
788 int pix_abs8x8_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size) | |
789 { | |
790 int s, i; | |
791 UINT8 *pix3 = pix2 + line_size; | |
792 | |
793 s = 0; | |
794 for(i=0;i<8;i++) { | |
795 s += abs(pix1[0] - avg2(pix2[0], pix3[0])); | |
796 s += abs(pix1[1] - avg2(pix2[1], pix3[1])); | |
797 s += abs(pix1[2] - avg2(pix2[2], pix3[2])); | |
798 s += abs(pix1[3] - avg2(pix2[3], pix3[3])); | |
799 s += abs(pix1[4] - avg2(pix2[4], pix3[4])); | |
800 s += abs(pix1[5] - avg2(pix2[5], pix3[5])); | |
801 s += abs(pix1[6] - avg2(pix2[6], pix3[6])); | |
802 s += abs(pix1[7] - avg2(pix2[7], pix3[7])); | |
803 pix1 += line_size; | |
804 pix2 += line_size; | |
805 pix3 += line_size; | |
806 } | |
807 return s; | |
808 } | |
809 | |
810 int pix_abs8x8_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size) | |
811 { | |
812 int s, i; | |
813 UINT8 *pix3 = pix2 + line_size; | |
814 | |
815 s = 0; | |
816 for(i=0;i<8;i++) { | |
817 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1])); | |
818 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2])); | |
819 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3])); | |
820 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4])); | |
821 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5])); | |
822 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6])); | |
823 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7])); | |
824 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8])); | |
825 pix1 += line_size; | |
826 pix2 += line_size; | |
827 pix3 += line_size; | |
828 } | |
829 return s; | |
830 } | |
831 | |
34 | 832 /* permute block according so that it corresponds to the MMX idct |
833 order */ | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
834 #ifdef SIMPLE_IDCT |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
835 /* general permutation, but perhaps slightly slower */ |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
836 void block_permute(INT16 *block) |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
837 { |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
838 int i; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
839 INT16 temp[64]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
840 |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
841 for(i=0; i<64; i++) temp[ block_permute_op(i) ] = block[i]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
842 |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
843 for(i=0; i<64; i++) block[i] = temp[i]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
844 } |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
845 #else |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
846 |
34 | 847 void block_permute(INT16 *block) |
848 { | |
849 int tmp1, tmp2, tmp3, tmp4, tmp5, tmp6; | |
850 int i; | |
851 | |
852 for(i=0;i<8;i++) { | |
853 tmp1 = block[1]; | |
854 tmp2 = block[2]; | |
855 tmp3 = block[3]; | |
856 tmp4 = block[4]; | |
857 tmp5 = block[5]; | |
858 tmp6 = block[6]; | |
859 block[1] = tmp2; | |
860 block[2] = tmp4; | |
861 block[3] = tmp6; | |
862 block[4] = tmp1; | |
863 block[5] = tmp3; | |
864 block[6] = tmp5; | |
865 block += 8; | |
866 } | |
867 } | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
868 #endif |
34 | 869 |
296 | 870 void clear_blocks_c(DCTELEM *blocks) |
871 { | |
872 memset(blocks, 0, sizeof(DCTELEM)*6*64); | |
873 } | |
874 | |
0 | 875 void dsputil_init(void) |
876 { | |
34 | 877 int i, j; |
88 | 878 int use_permuted_idct; |
0 | 879 |
880 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i; | |
881 for(i=0;i<MAX_NEG_CROP;i++) { | |
882 cropTbl[i] = 0; | |
883 cropTbl[i + MAX_NEG_CROP + 256] = 255; | |
884 } | |
885 | |
886 for(i=0;i<512;i++) { | |
887 squareTbl[i] = (i - 256) * (i - 256); | |
888 } | |
889 | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
890 #ifdef SIMPLE_IDCT |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
891 ff_idct = simple_idct; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
892 #else |
19
82d4c9be9873
MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents:
6
diff
changeset
|
893 ff_idct = j_rev_dct; |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
894 #endif |
0 | 895 get_pixels = get_pixels_c; |
896 put_pixels_clamped = put_pixels_clamped_c; | |
897 add_pixels_clamped = add_pixels_clamped_c; | |
255 | 898 gmc1= gmc1_c; |
296 | 899 clear_blocks= clear_blocks_c; |
0 | 900 |
294 | 901 pix_abs16x16 = pix_abs16x16_c; |
902 pix_abs16x16_x2 = pix_abs16x16_x2_c; | |
903 pix_abs16x16_y2 = pix_abs16x16_y2_c; | |
0 | 904 pix_abs16x16_xy2 = pix_abs16x16_xy2_c; |
294 | 905 pix_abs8x8 = pix_abs8x8_c; |
906 pix_abs8x8_x2 = pix_abs8x8_x2_c; | |
907 pix_abs8x8_y2 = pix_abs8x8_y2_c; | |
908 pix_abs8x8_xy2 = pix_abs8x8_xy2_c; | |
0 | 909 av_fdct = jpeg_fdct_ifast; |
910 | |
88 | 911 use_permuted_idct = 1; |
34 | 912 |
2 | 913 #ifdef HAVE_MMX |
0 | 914 dsputil_init_mmx(); |
915 #endif | |
62 | 916 #ifdef ARCH_ARMV4L |
917 dsputil_init_armv4l(); | |
918 #endif | |
88 | 919 #ifdef HAVE_MLIB |
920 dsputil_init_mlib(); | |
921 use_permuted_idct = 0; | |
922 #endif | |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
209
diff
changeset
|
923 #ifdef ARCH_ALPHA |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
209
diff
changeset
|
924 dsputil_init_alpha(); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
209
diff
changeset
|
925 use_permuted_idct = 0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
209
diff
changeset
|
926 #endif |
88 | 927 |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
928 #ifdef SIMPLE_IDCT |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
929 if(ff_idct == simple_idct) use_permuted_idct=0; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
930 #endif |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
931 |
190
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
932 if(use_permuted_idct) |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
933 #ifdef SIMPLE_IDCT |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
934 for(i=0; i<64; i++) permutation[i]= simple_mmx_permutation[i]; |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
935 #else |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
936 for(i=0; i<64; i++) permutation[i]= (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2); |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
937 #endif |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
938 else |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
939 for(i=0; i<64; i++) permutation[i]=i; |
9e0e56869d05
fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
174
diff
changeset
|
940 |
220 | 941 for(i=0; i<64; i++) inv_zigzag_direct16[zigzag_direct[i]]= i+1; |
942 for(i=0; i<64; i++) zigzag_direct_noperm[i]= zigzag_direct[i]; | |
943 | |
88 | 944 if (use_permuted_idct) { |
945 /* permute for IDCT */ | |
946 for(i=0;i<64;i++) { | |
947 j = zigzag_direct[i]; | |
948 zigzag_direct[i] = block_permute_op(j); | |
949 j = ff_alternate_horizontal_scan[i]; | |
950 ff_alternate_horizontal_scan[i] = block_permute_op(j); | |
951 j = ff_alternate_vertical_scan[i]; | |
952 ff_alternate_vertical_scan[i] = block_permute_op(j); | |
953 } | |
954 block_permute(default_intra_matrix); | |
955 block_permute(default_non_intra_matrix); | |
956 } | |
200 | 957 |
958 build_zigzag_end(); | |
0 | 959 } |
252
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
960 |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
961 void get_psnr(UINT8 *orig_image[3], UINT8 *coded_image[3], |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
962 int orig_linesize[3], int coded_linesize, |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
963 AVCodecContext *avctx) |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
964 { |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
965 int quad, diff, x, y; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
966 UINT8 *orig, *coded; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
967 UINT32 *sq = squareTbl + 256; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
968 |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
969 quad = 0; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
970 diff = 0; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
971 |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
972 /* Luminance */ |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
973 orig = orig_image[0]; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
974 coded = coded_image[0]; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
975 |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
976 for (y=0;y<avctx->height;y++) { |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
977 for (x=0;x<avctx->width;x++) { |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
978 diff = *(orig + x) - *(coded + x); |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
979 quad += sq[diff]; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
980 } |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
981 orig += orig_linesize[0]; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
982 coded += coded_linesize; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
983 } |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
984 |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
985 avctx->psnr_y = (float) quad / (float) (avctx->width * avctx->height); |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
986 |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
987 if (avctx->psnr_y) { |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
988 avctx->psnr_y = (float) (255 * 255) / avctx->psnr_y; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
989 avctx->psnr_y = 10 * (float) log10 (avctx->psnr_y); |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
990 } else |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
991 avctx->psnr_y = 99.99; |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
992 } |
ddb1a0e94cf4
- Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents:
220
diff
changeset
|
993 |