simplify ALIGN_MASK definition
[libav.git] / libpostproc / postprocess_template.c
1 /*
2 * Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at)
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 /**
22 * @file postprocess_template.c
23 * mmx/mmx2/3dnow postprocess code.
24 */
25
26 #include "x86_cpu.h"
27
28 #define ALIGN_MASK "$-8"
29
30 #undef PAVGB
31 #undef PMINUB
32 #undef PMAXUB
33
34 #ifdef HAVE_MMX2
35 #define REAL_PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
36 #elif defined (HAVE_3DNOW)
37 #define REAL_PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
38 #endif
39 #define PAVGB(a,b) REAL_PAVGB(a,b)
40
41 #ifdef HAVE_MMX2
42 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
43 #elif defined (HAVE_MMX)
44 #define PMINUB(b,a,t) \
45 "movq " #a ", " #t " \n\t"\
46 "psubusb " #b ", " #t " \n\t"\
47 "psubb " #t ", " #a " \n\t"
48 #endif
49
50 #ifdef HAVE_MMX2
51 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
52 #elif defined (HAVE_MMX)
53 #define PMAXUB(a,b) \
54 "psubusb " #a ", " #b " \n\t"\
55 "paddb " #a ", " #b " \n\t"
56 #endif
57
58 //FIXME? |255-0| = 1 (should not be a problem ...)
59 #ifdef HAVE_MMX
60 /**
61 * Check if the middle 8x8 Block in the given 8x16 block is flat
62 */
63 static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){
64 int numEq= 0, dcOk;
65 src+= stride*4; // src points to begin of the 8x8 Block
66 asm volatile(
67 "movq %0, %%mm7 \n\t"
68 "movq %1, %%mm6 \n\t"
69 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
70 );
71
72 asm volatile(
73 "lea (%2, %3), %%"REG_a" \n\t"
74 // 0 1 2 3 4 5 6 7 8 9
75 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
76
77 "movq (%2), %%mm0 \n\t"
78 "movq (%%"REG_a"), %%mm1 \n\t"
79 "movq %%mm0, %%mm3 \n\t"
80 "movq %%mm0, %%mm4 \n\t"
81 PMAXUB(%%mm1, %%mm4)
82 PMINUB(%%mm1, %%mm3, %%mm5)
83 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
84 "paddb %%mm7, %%mm0 \n\t"
85 "pcmpgtb %%mm6, %%mm0 \n\t"
86
87 "movq (%%"REG_a",%3), %%mm2 \n\t"
88 PMAXUB(%%mm2, %%mm4)
89 PMINUB(%%mm2, %%mm3, %%mm5)
90 "psubb %%mm2, %%mm1 \n\t"
91 "paddb %%mm7, %%mm1 \n\t"
92 "pcmpgtb %%mm6, %%mm1 \n\t"
93 "paddb %%mm1, %%mm0 \n\t"
94
95 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
96 PMAXUB(%%mm1, %%mm4)
97 PMINUB(%%mm1, %%mm3, %%mm5)
98 "psubb %%mm1, %%mm2 \n\t"
99 "paddb %%mm7, %%mm2 \n\t"
100 "pcmpgtb %%mm6, %%mm2 \n\t"
101 "paddb %%mm2, %%mm0 \n\t"
102
103 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
104
105 "movq (%2, %3, 4), %%mm2 \n\t"
106 PMAXUB(%%mm2, %%mm4)
107 PMINUB(%%mm2, %%mm3, %%mm5)
108 "psubb %%mm2, %%mm1 \n\t"
109 "paddb %%mm7, %%mm1 \n\t"
110 "pcmpgtb %%mm6, %%mm1 \n\t"
111 "paddb %%mm1, %%mm0 \n\t"
112
113 "movq (%%"REG_a"), %%mm1 \n\t"
114 PMAXUB(%%mm1, %%mm4)
115 PMINUB(%%mm1, %%mm3, %%mm5)
116 "psubb %%mm1, %%mm2 \n\t"
117 "paddb %%mm7, %%mm2 \n\t"
118 "pcmpgtb %%mm6, %%mm2 \n\t"
119 "paddb %%mm2, %%mm0 \n\t"
120
121 "movq (%%"REG_a", %3), %%mm2 \n\t"
122 PMAXUB(%%mm2, %%mm4)
123 PMINUB(%%mm2, %%mm3, %%mm5)
124 "psubb %%mm2, %%mm1 \n\t"
125 "paddb %%mm7, %%mm1 \n\t"
126 "pcmpgtb %%mm6, %%mm1 \n\t"
127 "paddb %%mm1, %%mm0 \n\t"
128
129 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
130 PMAXUB(%%mm1, %%mm4)
131 PMINUB(%%mm1, %%mm3, %%mm5)
132 "psubb %%mm1, %%mm2 \n\t"
133 "paddb %%mm7, %%mm2 \n\t"
134 "pcmpgtb %%mm6, %%mm2 \n\t"
135 "paddb %%mm2, %%mm0 \n\t"
136 "psubusb %%mm3, %%mm4 \n\t"
137
138 " \n\t"
139 #ifdef HAVE_MMX2
140 "pxor %%mm7, %%mm7 \n\t"
141 "psadbw %%mm7, %%mm0 \n\t"
142 #else
143 "movq %%mm0, %%mm1 \n\t"
144 "psrlw $8, %%mm0 \n\t"
145 "paddb %%mm1, %%mm0 \n\t"
146 "movq %%mm0, %%mm1 \n\t"
147 "psrlq $16, %%mm0 \n\t"
148 "paddb %%mm1, %%mm0 \n\t"
149 "movq %%mm0, %%mm1 \n\t"
150 "psrlq $32, %%mm0 \n\t"
151 "paddb %%mm1, %%mm0 \n\t"
152 #endif
153 "movq %4, %%mm7 \n\t" // QP,..., QP
154 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
155 "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0
156 "packssdw %%mm4, %%mm4 \n\t"
157 "movd %%mm0, %0 \n\t"
158 "movd %%mm4, %1 \n\t"
159
160 : "=r" (numEq), "=r" (dcOk)
161 : "r" (src), "r" ((long)stride), "m" (c->pQPb)
162 : "%"REG_a
163 );
164
165 numEq= (-numEq) &0xFF;
166 if(numEq > c->ppMode.flatnessThreshold){
167 if(dcOk) return 0;
168 else return 1;
169 }else{
170 return 2;
171 }
172 }
173 #endif //HAVE_MMX
174
175 /**
176 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
177 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
178 */
179 #ifndef HAVE_ALTIVEC
180 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
181 {
182 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
183 src+= stride*3;
184 asm volatile( //"movv %0 %1 %2\n\t"
185 "movq %2, %%mm0 \n\t" // QP,..., QP
186 "pxor %%mm4, %%mm4 \n\t"
187
188 "movq (%0), %%mm6 \n\t"
189 "movq (%0, %1), %%mm5 \n\t"
190 "movq %%mm5, %%mm1 \n\t"
191 "movq %%mm6, %%mm2 \n\t"
192 "psubusb %%mm6, %%mm5 \n\t"
193 "psubusb %%mm1, %%mm2 \n\t"
194 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
195 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
196 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
197
198 "pand %%mm2, %%mm6 \n\t"
199 "pandn %%mm1, %%mm2 \n\t"
200 "por %%mm2, %%mm6 \n\t"// First Line to Filter
201
202 "movq (%0, %1, 8), %%mm5 \n\t"
203 "lea (%0, %1, 4), %%"REG_a" \n\t"
204 "lea (%0, %1, 8), %%"REG_c" \n\t"
205 "sub %1, %%"REG_c" \n\t"
206 "add %1, %0 \n\t" // %0 points to line 1 not 0
207 "movq (%0, %1, 8), %%mm7 \n\t"
208 "movq %%mm5, %%mm1 \n\t"
209 "movq %%mm7, %%mm2 \n\t"
210 "psubusb %%mm7, %%mm5 \n\t"
211 "psubusb %%mm1, %%mm2 \n\t"
212 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
213 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
214 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
215
216 "pand %%mm2, %%mm7 \n\t"
217 "pandn %%mm1, %%mm2 \n\t"
218 "por %%mm2, %%mm7 \n\t" // First Line to Filter
219
220
221 // 1 2 3 4 5 6 7 8
222 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
223 // 6 4 2 2 1 1
224 // 6 4 4 2
225 // 6 8 2
226
227 "movq (%0, %1), %%mm0 \n\t" // 1
228 "movq %%mm0, %%mm1 \n\t" // 1
229 PAVGB(%%mm6, %%mm0) //1 1 /2
230 PAVGB(%%mm6, %%mm0) //3 1 /4
231
232 "movq (%0, %1, 4), %%mm2 \n\t" // 1
233 "movq %%mm2, %%mm5 \n\t" // 1
234 PAVGB((%%REGa), %%mm2) // 11 /2
235 PAVGB((%0, %1, 2), %%mm2) // 211 /4
236 "movq %%mm2, %%mm3 \n\t" // 211 /4
237 "movq (%0), %%mm4 \n\t" // 1
238 PAVGB(%%mm4, %%mm3) // 4 211 /8
239 PAVGB(%%mm0, %%mm3) //642211 /16
240 "movq %%mm3, (%0) \n\t" // X
241 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
242 "movq %%mm1, %%mm0 \n\t" // 1
243 PAVGB(%%mm6, %%mm0) //1 1 /2
244 "movq %%mm4, %%mm3 \n\t" // 1
245 PAVGB((%0,%1,2), %%mm3) // 1 1 /2
246 PAVGB((%%REGa,%1,2), %%mm5) // 11 /2
247 PAVGB((%%REGa), %%mm5) // 211 /4
248 PAVGB(%%mm5, %%mm3) // 2 2211 /8
249 PAVGB(%%mm0, %%mm3) //4242211 /16
250 "movq %%mm3, (%0,%1) \n\t" // X
251 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
252 PAVGB(%%mm4, %%mm6) //11 /2
253 "movq (%%"REG_c"), %%mm0 \n\t" // 1
254 PAVGB((%%REGa, %1, 2), %%mm0) // 11/2
255 "movq %%mm0, %%mm3 \n\t" // 11/2
256 PAVGB(%%mm1, %%mm0) // 2 11/4
257 PAVGB(%%mm6, %%mm0) //222 11/8
258 PAVGB(%%mm2, %%mm0) //22242211/16
259 "movq (%0, %1, 2), %%mm2 \n\t" // 1
260 "movq %%mm0, (%0, %1, 2) \n\t" // X
261 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
262 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
263 PAVGB((%%REGc), %%mm0) // 11 /2
264 PAVGB(%%mm0, %%mm6) //11 11 /4
265 PAVGB(%%mm1, %%mm4) // 11 /2
266 PAVGB(%%mm2, %%mm1) // 11 /2
267 PAVGB(%%mm1, %%mm6) //1122 11 /8
268 PAVGB(%%mm5, %%mm6) //112242211 /16
269 "movq (%%"REG_a"), %%mm5 \n\t" // 1
270 "movq %%mm6, (%%"REG_a") \n\t" // X
271 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
272 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1
273 PAVGB(%%mm7, %%mm6) // 11 /2
274 PAVGB(%%mm4, %%mm6) // 11 11 /4
275 PAVGB(%%mm3, %%mm6) // 11 2211 /8
276 PAVGB(%%mm5, %%mm2) // 11 /2
277 "movq (%0, %1, 4), %%mm4 \n\t" // 1
278 PAVGB(%%mm4, %%mm2) // 112 /4
279 PAVGB(%%mm2, %%mm6) // 112242211 /16
280 "movq %%mm6, (%0, %1, 4) \n\t" // X
281 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
282 PAVGB(%%mm7, %%mm1) // 11 2 /4
283 PAVGB(%%mm4, %%mm5) // 11 /2
284 PAVGB(%%mm5, %%mm0) // 11 11 /4
285 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1
286 PAVGB(%%mm6, %%mm1) // 11 4 2 /8
287 PAVGB(%%mm0, %%mm1) // 11224222 /16
288 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X
289 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
290 PAVGB((%%REGc), %%mm2) // 112 4 /8
291 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
292 PAVGB(%%mm0, %%mm6) // 1 1 /2
293 PAVGB(%%mm7, %%mm6) // 1 12 /4
294 PAVGB(%%mm2, %%mm6) // 1122424 /4
295 "movq %%mm6, (%%"REG_c") \n\t" // X
296 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
297 PAVGB(%%mm7, %%mm5) // 11 2 /4
298 PAVGB(%%mm7, %%mm5) // 11 6 /8
299
300 PAVGB(%%mm3, %%mm0) // 112 /4
301 PAVGB(%%mm0, %%mm5) // 112246 /16
302 "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X
303 "sub %1, %0 \n\t"
304
305 :
306 : "r" (src), "r" ((long)stride), "m" (c->pQPb)
307 : "%"REG_a, "%"REG_c
308 );
309 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
310 const int l1= stride;
311 const int l2= stride + l1;
312 const int l3= stride + l2;
313 const int l4= stride + l3;
314 const int l5= stride + l4;
315 const int l6= stride + l5;
316 const int l7= stride + l6;
317 const int l8= stride + l7;
318 const int l9= stride + l8;
319 int x;
320 src+= stride*3;
321 for(x=0; x<BLOCK_SIZE; x++)
322 {
323 const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
324 const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
325
326 int sums[10];
327 sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
328 sums[1] = sums[0] - first + src[l4];
329 sums[2] = sums[1] - first + src[l5];
330 sums[3] = sums[2] - first + src[l6];
331 sums[4] = sums[3] - first + src[l7];
332 sums[5] = sums[4] - src[l1] + src[l8];
333 sums[6] = sums[5] - src[l2] + last;
334 sums[7] = sums[6] - src[l3] + last;
335 sums[8] = sums[7] - src[l4] + last;
336 sums[9] = sums[8] - src[l5] + last;
337
338 src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
339 src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
340 src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
341 src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
342 src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
343 src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
344 src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
345 src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
346
347 src++;
348 }
349 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
350 }
351 #endif //HAVE_ALTIVEC
352
353 #if 0
354 /**
355 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
356 * values are correctly clipped (MMX2)
357 * values are wraparound (C)
358 * Conclusion: It is fast, but introduces ugly horizontal patterns
359 * if there is a continuous gradient.
360 0 8 16 24
361 x = 8
362 x/2 = 4
363 x/8 = 1
364 1 12 12 23
365 */
366 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
367 {
368 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
369 src+= stride*3;
370 // FIXME rounding
371 asm volatile(
372 "pxor %%mm7, %%mm7 \n\t" // 0
373 "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
374 "leal (%0, %1), %%"REG_a" \n\t"
375 "leal (%%"REG_a", %1, 4), %%"REG_c" \n\t"
376 // 0 1 2 3 4 5 6 7 8 9
377 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
378 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
379 "movq %%mm0, %%mm1 \n\t" // QP,..., QP
380 "paddusb "MANGLE(b02)", %%mm0 \n\t"
381 "psrlw $2, %%mm0 \n\t"
382 "pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4
383 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
384 "movq (%0, %1, 4), %%mm2 \n\t" // line 4
385 "movq (%%"REG_c"), %%mm3 \n\t" // line 5
386 "movq %%mm2, %%mm4 \n\t" // line 4
387 "pcmpeqb %%mm5, %%mm5 \n\t" // -1
388 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1
389 PAVGB(%%mm3, %%mm5)
390 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2
391 "psubusb %%mm3, %%mm4 \n\t"
392 "psubusb %%mm2, %%mm3 \n\t"
393 "por %%mm3, %%mm4 \n\t" // |l4 - l5|
394 "psubusb %%mm0, %%mm4 \n\t"
395 "pcmpeqb %%mm7, %%mm4 \n\t"
396 "pand %%mm4, %%mm5 \n\t" // d/2
397
398 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80
399 "paddb %%mm5, %%mm2 \n\t"
400 // "psubb %%mm6, %%mm2 \n\t"
401 "movq %%mm2, (%0,%1, 4) \n\t"
402
403 "movq (%%"REG_c"), %%mm2 \n\t"
404 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80
405 "psubb %%mm5, %%mm2 \n\t"
406 // "psubb %%mm6, %%mm2 \n\t"
407 "movq %%mm2, (%%"REG_c") \n\t"
408
409 "paddb %%mm6, %%mm5 \n\t"
410 "psrlw $2, %%mm5 \n\t"
411 "pand "MANGLE(b3F)", %%mm5 \n\t"
412 "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
413
414 "movq (%%"REG_a", %1, 2), %%mm2 \n\t"
415 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
416 "paddsb %%mm5, %%mm2 \n\t"
417 "psubb %%mm6, %%mm2 \n\t"
418 "movq %%mm2, (%%"REG_a", %1, 2) \n\t"
419
420 "movq (%%"REG_c", %1), %%mm2 \n\t"
421 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80
422 "psubsb %%mm5, %%mm2 \n\t"
423 "psubb %%mm6, %%mm2 \n\t"
424 "movq %%mm2, (%%"REG_c", %1) \n\t"
425
426 :
427 : "r" (src), "r" ((long)stride)
428 : "%"REG_a, "%"REG_c
429 );
430 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
431 const int l1= stride;
432 const int l2= stride + l1;
433 const int l3= stride + l2;
434 const int l4= stride + l3;
435 const int l5= stride + l4;
436 const int l6= stride + l5;
437 // const int l7= stride + l6;
438 // const int l8= stride + l7;
439 // const int l9= stride + l8;
440 int x;
441 const int QP15= QP + (QP>>2);
442 src+= stride*3;
443 for(x=0; x<BLOCK_SIZE; x++)
444 {
445 const int v = (src[x+l5] - src[x+l4]);
446 if(FFABS(v) < QP15)
447 {
448 src[x+l3] +=v>>3;
449 src[x+l4] +=v>>1;
450 src[x+l5] -=v>>1;
451 src[x+l6] -=v>>3;
452
453 }
454 }
455
456 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
457 }
458 #endif //0
459
460 /**
461 * Experimental Filter 1
462 * will not damage linear gradients
463 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
464 * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
465 * MMX2 version does correct clipping C version does not
466 */
467 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
468 {
469 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
470 src+= stride*3;
471
472 asm volatile(
473 "pxor %%mm7, %%mm7 \n\t" // 0
474 "lea (%0, %1), %%"REG_a" \n\t"
475 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
476 // 0 1 2 3 4 5 6 7 8 9
477 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
478 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
479 "movq (%0, %1, 4), %%mm1 \n\t" // line 4
480 "movq %%mm1, %%mm2 \n\t" // line 4
481 "psubusb %%mm0, %%mm1 \n\t"
482 "psubusb %%mm2, %%mm0 \n\t"
483 "por %%mm1, %%mm0 \n\t" // |l2 - l3|
484 "movq (%%"REG_c"), %%mm3 \n\t" // line 5
485 "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6
486 "movq %%mm3, %%mm5 \n\t" // line 5
487 "psubusb %%mm4, %%mm3 \n\t"
488 "psubusb %%mm5, %%mm4 \n\t"
489 "por %%mm4, %%mm3 \n\t" // |l5 - l6|
490 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
491 "movq %%mm2, %%mm1 \n\t" // line 4
492 "psubusb %%mm5, %%mm2 \n\t"
493 "movq %%mm2, %%mm4 \n\t"
494 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
495 "psubusb %%mm1, %%mm5 \n\t"
496 "por %%mm5, %%mm4 \n\t" // |l4 - l5|
497 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
498 "movq %%mm4, %%mm3 \n\t" // d
499 "movq %2, %%mm0 \n\t"
500 "paddusb %%mm0, %%mm0 \n\t"
501 "psubusb %%mm0, %%mm4 \n\t"
502 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
503 "psubusb "MANGLE(b01)", %%mm3 \n\t"
504 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
505
506 PAVGB(%%mm7, %%mm3) // d/2
507 "movq %%mm3, %%mm1 \n\t" // d/2
508 PAVGB(%%mm7, %%mm3) // d/4
509 PAVGB(%%mm1, %%mm3) // 3*d/8
510
511 "movq (%0, %1, 4), %%mm0 \n\t" // line 4
512 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
513 "psubusb %%mm3, %%mm0 \n\t"
514 "pxor %%mm2, %%mm0 \n\t"
515 "movq %%mm0, (%0, %1, 4) \n\t" // line 4
516
517 "movq (%%"REG_c"), %%mm0 \n\t" // line 5
518 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
519 "paddusb %%mm3, %%mm0 \n\t"
520 "pxor %%mm2, %%mm0 \n\t"
521 "movq %%mm0, (%%"REG_c") \n\t" // line 5
522
523 PAVGB(%%mm7, %%mm1) // d/4
524
525 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
526 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
527 "psubusb %%mm1, %%mm0 \n\t"
528 "pxor %%mm2, %%mm0 \n\t"
529 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3
530
531 "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6
532 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
533 "paddusb %%mm1, %%mm0 \n\t"
534 "pxor %%mm2, %%mm0 \n\t"
535 "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6
536
537 PAVGB(%%mm7, %%mm1) // d/8
538
539 "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2
540 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
541 "psubusb %%mm1, %%mm0 \n\t"
542 "pxor %%mm2, %%mm0 \n\t"
543 "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2
544
545 "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7
546 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
547 "paddusb %%mm1, %%mm0 \n\t"
548 "pxor %%mm2, %%mm0 \n\t"
549 "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7
550
551 :
552 : "r" (src), "r" ((long)stride), "m" (co->pQPb)
553 : "%"REG_a, "%"REG_c
554 );
555 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
556
557 const int l1= stride;
558 const int l2= stride + l1;
559 const int l3= stride + l2;
560 const int l4= stride + l3;
561 const int l5= stride + l4;
562 const int l6= stride + l5;
563 const int l7= stride + l6;
564 // const int l8= stride + l7;
565 // const int l9= stride + l8;
566 int x;
567
568 src+= stride*3;
569 for(x=0; x<BLOCK_SIZE; x++)
570 {
571 int a= src[l3] - src[l4];
572 int b= src[l4] - src[l5];
573 int c= src[l5] - src[l6];
574
575 int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
576 d= FFMAX(d, 0);
577
578 if(d < co->QP*2)
579 {
580 int v = d * FFSIGN(-b);
581
582 src[l2] +=v>>3;
583 src[l3] +=v>>2;
584 src[l4] +=(3*v)>>3;
585 src[l5] -=(3*v)>>3;
586 src[l6] -=v>>2;
587 src[l7] -=v>>3;
588
589 }
590 src++;
591 }
592 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
593 }
594
595 #ifndef HAVE_ALTIVEC
596 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
597 {
598 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
599 /*
600 uint8_t tmp[16];
601 const int l1= stride;
602 const int l2= stride + l1;
603 const int l3= stride + l2;
604 const int l4= (int)tmp - (int)src - stride*3;
605 const int l5= (int)tmp - (int)src - stride*3 + 8;
606 const int l6= stride*3 + l3;
607 const int l7= stride + l6;
608 const int l8= stride + l7;
609
610 memcpy(tmp, src+stride*7, 8);
611 memcpy(tmp+8, src+stride*8, 8);
612 */
613 src+= stride*4;
614 asm volatile(
615
616 #if 0 //sligtly more accurate and slightly slower
617 "pxor %%mm7, %%mm7 \n\t" // 0
618 "lea (%0, %1), %%"REG_a" \n\t"
619 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
620 // 0 1 2 3 4 5 6 7
621 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
622 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
623
624
625 "movq (%0, %1, 2), %%mm0 \n\t" // l2
626 "movq (%0), %%mm1 \n\t" // l0
627 "movq %%mm0, %%mm2 \n\t" // l2
628 PAVGB(%%mm7, %%mm0) // ~l2/2
629 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
630 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
631
632 "movq (%%"REG_a"), %%mm1 \n\t" // l1
633 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3
634 "movq %%mm1, %%mm4 \n\t" // l1
635 PAVGB(%%mm7, %%mm1) // ~l1/2
636 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
637 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
638
639 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
640 "psubusb %%mm1, %%mm0 \n\t"
641 "psubusb %%mm4, %%mm1 \n\t"
642 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
643 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
644
645 "movq (%0, %1, 4), %%mm0 \n\t" // l4
646 "movq %%mm0, %%mm4 \n\t" // l4
647 PAVGB(%%mm7, %%mm0) // ~l4/2
648 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
649 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
650
651 "movq (%%"REG_c"), %%mm2 \n\t" // l5
652 "movq %%mm3, %%mm5 \n\t" // l3
653 PAVGB(%%mm7, %%mm3) // ~l3/2
654 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
655 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
656
657 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
658 "psubusb %%mm3, %%mm0 \n\t"
659 "psubusb %%mm6, %%mm3 \n\t"
660 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
661 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
662 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
663
664 "movq (%%"REG_c", %1), %%mm6 \n\t" // l6
665 "movq %%mm6, %%mm5 \n\t" // l6
666 PAVGB(%%mm7, %%mm6) // ~l6/2
667 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
668 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
669
670 "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7
671 "movq %%mm2, %%mm4 \n\t" // l5
672 PAVGB(%%mm7, %%mm2) // ~l5/2
673 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
674 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
675
676 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
677 "psubusb %%mm2, %%mm6 \n\t"
678 "psubusb %%mm4, %%mm2 \n\t"
679 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
680 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
681
682
683 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
684 "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
685 "paddusb "MANGLE(b01)", %%mm4 \n\t"
686 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
687 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
688 "pand %%mm4, %%mm3 \n\t"
689
690 "movq %%mm3, %%mm1 \n\t"
691 // "psubusb "MANGLE(b01)", %%mm3 \n\t"
692 PAVGB(%%mm7, %%mm3)
693 PAVGB(%%mm7, %%mm3)
694 "paddusb %%mm1, %%mm3 \n\t"
695 // "paddusb "MANGLE(b01)", %%mm3 \n\t"
696
697 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3
698 "movq (%0, %1, 4), %%mm5 \n\t" //l4
699 "movq (%0, %1, 4), %%mm4 \n\t" //l4
700 "psubusb %%mm6, %%mm5 \n\t"
701 "psubusb %%mm4, %%mm6 \n\t"
702 "por %%mm6, %%mm5 \n\t" // |l3-l4|
703 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
704 "pxor %%mm6, %%mm0 \n\t"
705 "pand %%mm0, %%mm3 \n\t"
706 PMINUB(%%mm5, %%mm3, %%mm0)
707
708 "psubusb "MANGLE(b01)", %%mm3 \n\t"
709 PAVGB(%%mm7, %%mm3)
710
711 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
712 "movq (%0, %1, 4), %%mm2 \n\t"
713 "pxor %%mm6, %%mm0 \n\t"
714 "pxor %%mm6, %%mm2 \n\t"
715 "psubb %%mm3, %%mm0 \n\t"
716 "paddb %%mm3, %%mm2 \n\t"
717 "pxor %%mm6, %%mm0 \n\t"
718 "pxor %%mm6, %%mm2 \n\t"
719 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
720 "movq %%mm2, (%0, %1, 4) \n\t"
721 #endif //0
722
723 "lea (%0, %1), %%"REG_a" \n\t"
724 "pcmpeqb %%mm6, %%mm6 \n\t" // -1
725 // 0 1 2 3 4 5 6 7
726 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
727 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
728
729
730 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3
731 "movq (%0, %1, 4), %%mm0 \n\t" // l4
732 "pxor %%mm6, %%mm1 \n\t" // -l3-1
733 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
734 // mm1=-l3-1, mm0=128-q
735
736 "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5
737 "movq (%%"REG_a", %1), %%mm3 \n\t" // l2
738 "pxor %%mm6, %%mm2 \n\t" // -l5-1
739 "movq %%mm2, %%mm5 \n\t" // -l5-1
740 "movq "MANGLE(b80)", %%mm4 \n\t" // 128
741 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
742 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
743 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
744 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
745 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
746 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
747
748 "movq (%%"REG_a"), %%mm2 \n\t" // l1
749 "pxor %%mm6, %%mm2 \n\t" // -l1-1
750 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
751 PAVGB((%0), %%mm1) // (l0-l3+256)/2
752 "movq "MANGLE(b80)", %%mm3 \n\t" // 128
753 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
754 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
755 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
756 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
757
758 PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2
759 "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7
760 "pxor %%mm6, %%mm1 \n\t" // -l7-1
761 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
762 "movq "MANGLE(b80)", %%mm2 \n\t" // 128
763 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
764 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
765 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
766 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
767
768 "movq "MANGLE(b00)", %%mm1 \n\t" // 0
769 "movq "MANGLE(b00)", %%mm5 \n\t" // 0
770 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
771 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
772 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
773 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
774 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
775
776 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
777
778 "movq "MANGLE(b00)", %%mm7 \n\t" // 0
779 "movq %2, %%mm2 \n\t" // QP
780 PAVGB(%%mm6, %%mm2) // 128 + QP/2
781 "psubb %%mm6, %%mm2 \n\t"
782
783 "movq %%mm4, %%mm1 \n\t"
784 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
785 "pxor %%mm1, %%mm4 \n\t"
786 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
787 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
788 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
789 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
790
791 "movq %%mm4, %%mm3 \n\t" // d
792 "psubusb "MANGLE(b01)", %%mm4 \n\t"
793 PAVGB(%%mm7, %%mm4) // d/32
794 PAVGB(%%mm7, %%mm4) // (d + 32)/64
795 "paddb %%mm3, %%mm4 \n\t" // 5d/64
796 "pand %%mm2, %%mm4 \n\t"
797
798 "movq "MANGLE(b80)", %%mm5 \n\t" // 128
799 "psubb %%mm0, %%mm5 \n\t" // q
800 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
801 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
802 "pxor %%mm7, %%mm5 \n\t"
803
804 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
805 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
806
807 "pand %%mm7, %%mm4 \n\t"
808 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
809 "movq (%0, %1, 4), %%mm2 \n\t"
810 "pxor %%mm1, %%mm0 \n\t"
811 "pxor %%mm1, %%mm2 \n\t"
812 "paddb %%mm4, %%mm0 \n\t"
813 "psubb %%mm4, %%mm2 \n\t"
814 "pxor %%mm1, %%mm0 \n\t"
815 "pxor %%mm1, %%mm2 \n\t"
816 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
817 "movq %%mm2, (%0, %1, 4) \n\t"
818
819 :
820 : "r" (src), "r" ((long)stride), "m" (c->pQPb)
821 : "%"REG_a, "%"REG_c
822 );
823
824 /*
825 {
826 int x;
827 src-= stride;
828 for(x=0; x<BLOCK_SIZE; x++)
829 {
830 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
831 if(FFABS(middleEnergy)< 8*QP)
832 {
833 const int q=(src[l4] - src[l5])/2;
834 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
835 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
836
837 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
838 d= FFMAX(d, 0);
839
840 d= (5*d + 32) >> 6;
841 d*= FFSIGN(-middleEnergy);
842
843 if(q>0)
844 {
845 d= d<0 ? 0 : d;
846 d= d>q ? q : d;
847 }
848 else
849 {
850 d= d>0 ? 0 : d;
851 d= d<q ? q : d;
852 }
853
854 src[l4]-= d;
855 src[l5]+= d;
856 }
857 src++;
858 }
859 src-=8;
860 for(x=0; x<8; x++)
861 {
862 int y;
863 for(y=4; y<6; y++)
864 {
865 int d= src[x+y*stride] - tmp[x+(y-4)*8];
866 int ad= FFABS(d);
867 static int max=0;
868 static int sum=0;
869 static int num=0;
870 static int bias=0;
871
872 if(max<ad) max=ad;
873 sum+= ad>3 ? 1 : 0;
874 if(ad>3)
875 {
876 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
877 }
878 if(y==4) bias+=d;
879 num++;
880 if(num%1000000 == 0)
881 {
882 av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias);
883 }
884 }
885 }
886 }
887 */
888 #elif defined (HAVE_MMX)
889 src+= stride*4;
890 asm volatile(
891 "pxor %%mm7, %%mm7 \n\t"
892 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
893 "and "ALIGN_MASK", %%"REG_c" \n\t" // align
894 // 0 1 2 3 4 5 6 7
895 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
896 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
897
898 "movq (%0), %%mm0 \n\t"
899 "movq %%mm0, %%mm1 \n\t"
900 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
901 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
902
903 "movq (%0, %1), %%mm2 \n\t"
904 "lea (%0, %1, 2), %%"REG_a" \n\t"
905 "movq %%mm2, %%mm3 \n\t"
906 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
907 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
908
909 "movq (%%"REG_a"), %%mm4 \n\t"
910 "movq %%mm4, %%mm5 \n\t"
911 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
912 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
913
914 "paddw %%mm0, %%mm0 \n\t" // 2L0
915 "paddw %%mm1, %%mm1 \n\t" // 2H0
916 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
917 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
918 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
919 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
920
921 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
922 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
923 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
924 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
925
926 "movq (%%"REG_a", %1), %%mm2 \n\t"
927 "movq %%mm2, %%mm3 \n\t"
928 "punpcklbw %%mm7, %%mm2 \n\t" // L3
929 "punpckhbw %%mm7, %%mm3 \n\t" // H3
930
931 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
932 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
933 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
934 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
935 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
936 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
937
938 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
939 "movq %%mm0, %%mm1 \n\t"
940 "punpcklbw %%mm7, %%mm0 \n\t" // L4
941 "punpckhbw %%mm7, %%mm1 \n\t" // H4
942
943 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
944 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
945 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
946 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
947 "paddw %%mm4, %%mm4 \n\t" // 2L2
948 "paddw %%mm5, %%mm5 \n\t" // 2H2
949 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
950 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
951
952 "lea (%%"REG_a", %1), %0 \n\t"
953 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
954 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
955 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
956 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
957 //50 opcodes so far
958 "movq (%0, %1, 2), %%mm2 \n\t"
959 "movq %%mm2, %%mm3 \n\t"
960 "punpcklbw %%mm7, %%mm2 \n\t" // L5
961 "punpckhbw %%mm7, %%mm3 \n\t" // H5
962 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
963 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
964 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
965 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
966
967 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
968 "punpcklbw %%mm7, %%mm6 \n\t" // L6
969 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
970 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
971 "punpckhbw %%mm7, %%mm6 \n\t" // H6
972 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
973
974 "paddw %%mm0, %%mm0 \n\t" // 2L4
975 "paddw %%mm1, %%mm1 \n\t" // 2H4
976 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
977 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
978
979 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
980 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
981 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
982 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
983
984 "movq (%0, %1, 4), %%mm2 \n\t"
985 "movq %%mm2, %%mm3 \n\t"
986 "punpcklbw %%mm7, %%mm2 \n\t" // L7
987 "punpckhbw %%mm7, %%mm3 \n\t" // H7
988
989 "paddw %%mm2, %%mm2 \n\t" // 2L7
990 "paddw %%mm3, %%mm3 \n\t" // 2H7
991 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
992 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
993
994 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
995 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
996
997 #ifdef HAVE_MMX2
998 "movq %%mm7, %%mm6 \n\t" // 0
999 "psubw %%mm0, %%mm6 \n\t"
1000 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1001 "movq %%mm7, %%mm6 \n\t" // 0
1002 "psubw %%mm1, %%mm6 \n\t"
1003 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1004 "movq %%mm7, %%mm6 \n\t" // 0
1005 "psubw %%mm2, %%mm6 \n\t"
1006 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1007 "movq %%mm7, %%mm6 \n\t" // 0
1008 "psubw %%mm3, %%mm6 \n\t"
1009 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1010 #else
1011 "movq %%mm7, %%mm6 \n\t" // 0
1012 "pcmpgtw %%mm0, %%mm6 \n\t"
1013 "pxor %%mm6, %%mm0 \n\t"
1014 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1015 "movq %%mm7, %%mm6 \n\t" // 0
1016 "pcmpgtw %%mm1, %%mm6 \n\t"
1017 "pxor %%mm6, %%mm1 \n\t"
1018 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1019 "movq %%mm7, %%mm6 \n\t" // 0
1020 "pcmpgtw %%mm2, %%mm6 \n\t"
1021 "pxor %%mm6, %%mm2 \n\t"
1022 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1023 "movq %%mm7, %%mm6 \n\t" // 0
1024 "pcmpgtw %%mm3, %%mm6 \n\t"
1025 "pxor %%mm6, %%mm3 \n\t"
1026 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1027 #endif
1028
1029 #ifdef HAVE_MMX2
1030 "pminsw %%mm2, %%mm0 \n\t"
1031 "pminsw %%mm3, %%mm1 \n\t"
1032 #else
1033 "movq %%mm0, %%mm6 \n\t"
1034 "psubusw %%mm2, %%mm6 \n\t"
1035 "psubw %%mm6, %%mm0 \n\t"
1036 "movq %%mm1, %%mm6 \n\t"
1037 "psubusw %%mm3, %%mm6 \n\t"
1038 "psubw %%mm6, %%mm1 \n\t"
1039 #endif
1040
1041 "movd %2, %%mm2 \n\t" // QP
1042 "punpcklbw %%mm7, %%mm2 \n\t"
1043
1044 "movq %%mm7, %%mm6 \n\t" // 0
1045 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1046 "pxor %%mm6, %%mm4 \n\t"
1047 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1048 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1049 "pxor %%mm7, %%mm5 \n\t"
1050 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1051 // 100 opcodes
1052 "psllw $3, %%mm2 \n\t" // 8QP
1053 "movq %%mm2, %%mm3 \n\t" // 8QP
1054 "pcmpgtw %%mm4, %%mm2 \n\t"
1055 "pcmpgtw %%mm5, %%mm3 \n\t"
1056 "pand %%mm2, %%mm4 \n\t"
1057 "pand %%mm3, %%mm5 \n\t"
1058
1059
1060 "psubusw %%mm0, %%mm4 \n\t" // hd
1061 "psubusw %%mm1, %%mm5 \n\t" // ld
1062
1063
1064 "movq "MANGLE(w05)", %%mm2 \n\t" // 5
1065 "pmullw %%mm2, %%mm4 \n\t"
1066 "pmullw %%mm2, %%mm5 \n\t"
1067 "movq "MANGLE(w20)", %%mm2 \n\t" // 32
1068 "paddw %%mm2, %%mm4 \n\t"
1069 "paddw %%mm2, %%mm5 \n\t"
1070 "psrlw $6, %%mm4 \n\t"
1071 "psrlw $6, %%mm5 \n\t"
1072
1073 "movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4
1074 "movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4
1075
1076 "pxor %%mm2, %%mm2 \n\t"
1077 "pxor %%mm3, %%mm3 \n\t"
1078
1079 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1080 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
1081 "pxor %%mm2, %%mm0 \n\t"
1082 "pxor %%mm3, %%mm1 \n\t"
1083 "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
1084 "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
1085 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
1086 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
1087
1088 "pxor %%mm6, %%mm2 \n\t"
1089 "pxor %%mm7, %%mm3 \n\t"
1090 "pand %%mm2, %%mm4 \n\t"
1091 "pand %%mm3, %%mm5 \n\t"
1092
1093 #ifdef HAVE_MMX2
1094 "pminsw %%mm0, %%mm4 \n\t"
1095 "pminsw %%mm1, %%mm5 \n\t"
1096 #else
1097 "movq %%mm4, %%mm2 \n\t"
1098 "psubusw %%mm0, %%mm2 \n\t"
1099 "psubw %%mm2, %%mm4 \n\t"
1100 "movq %%mm5, %%mm2 \n\t"
1101 "psubusw %%mm1, %%mm2 \n\t"
1102 "psubw %%mm2, %%mm5 \n\t"
1103 #endif
1104 "pxor %%mm6, %%mm4 \n\t"
1105 "pxor %%mm7, %%mm5 \n\t"
1106 "psubw %%mm6, %%mm4 \n\t"
1107 "psubw %%mm7, %%mm5 \n\t"
1108 "packsswb %%mm5, %%mm4 \n\t"
1109 "movq (%0), %%mm0 \n\t"
1110 "paddb %%mm4, %%mm0 \n\t"
1111 "movq %%mm0, (%0) \n\t"
1112 "movq (%0, %1), %%mm0 \n\t"
1113 "psubb %%mm4, %%mm0 \n\t"
1114 "movq %%mm0, (%0, %1) \n\t"
1115
1116 : "+r" (src)
1117 : "r" ((long)stride), "m" (c->pQPb)
1118 : "%"REG_a, "%"REG_c
1119 );
1120 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1121 const int l1= stride;
1122 const int l2= stride + l1;
1123 const int l3= stride + l2;
1124 const int l4= stride + l3;
1125 const int l5= stride + l4;
1126 const int l6= stride + l5;
1127 const int l7= stride + l6;
1128 const int l8= stride + l7;
1129 // const int l9= stride + l8;
1130 int x;
1131 src+= stride*3;
1132 for(x=0; x<BLOCK_SIZE; x++)
1133 {
1134 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1135 if(FFABS(middleEnergy) < 8*c->QP)
1136 {
1137 const int q=(src[l4] - src[l5])/2;
1138 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1139 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1140
1141 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
1142 d= FFMAX(d, 0);
1143
1144 d= (5*d + 32) >> 6;
1145 d*= FFSIGN(-middleEnergy);
1146
1147 if(q>0)
1148 {
1149 d= d<0 ? 0 : d;
1150 d= d>q ? q : d;
1151 }
1152 else
1153 {
1154 d= d>0 ? 0 : d;
1155 d= d<q ? q : d;
1156 }
1157
1158 src[l4]-= d;
1159 src[l5]+= d;
1160 }
1161 src++;
1162 }
1163 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1164 }
1165 #endif //HAVE_ALTIVEC
1166
1167 #ifndef HAVE_ALTIVEC
1168 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
1169 {
1170 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1171 asm volatile(
1172 "pxor %%mm6, %%mm6 \n\t"
1173 "pcmpeqb %%mm7, %%mm7 \n\t"
1174 "movq %2, %%mm0 \n\t"
1175 "punpcklbw %%mm6, %%mm0 \n\t"
1176 "psrlw $1, %%mm0 \n\t"
1177 "psubw %%mm7, %%mm0 \n\t"
1178 "packuswb %%mm0, %%mm0 \n\t"
1179 "movq %%mm0, %3 \n\t"
1180
1181 "lea (%0, %1), %%"REG_a" \n\t"
1182 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1183
1184 // 0 1 2 3 4 5 6 7 8 9
1185 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1186
1187 #undef FIND_MIN_MAX
1188 #ifdef HAVE_MMX2
1189 #define REAL_FIND_MIN_MAX(addr)\
1190 "movq " #addr ", %%mm0 \n\t"\
1191 "pminub %%mm0, %%mm7 \n\t"\
1192 "pmaxub %%mm0, %%mm6 \n\t"
1193 #else
1194 #define REAL_FIND_MIN_MAX(addr)\
1195 "movq " #addr ", %%mm0 \n\t"\
1196 "movq %%mm7, %%mm1 \n\t"\
1197 "psubusb %%mm0, %%mm6 \n\t"\
1198 "paddb %%mm0, %%mm6 \n\t"\
1199 "psubusb %%mm0, %%mm1 \n\t"\
1200 "psubb %%mm1, %%mm7 \n\t"
1201 #endif
1202 #define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr)
1203
1204 FIND_MIN_MAX((%%REGa))
1205 FIND_MIN_MAX((%%REGa, %1))
1206 FIND_MIN_MAX((%%REGa, %1, 2))
1207 FIND_MIN_MAX((%0, %1, 4))
1208 FIND_MIN_MAX((%%REGd))
1209 FIND_MIN_MAX((%%REGd, %1))
1210 FIND_MIN_MAX((%%REGd, %1, 2))
1211 FIND_MIN_MAX((%0, %1, 8))
1212
1213 "movq %%mm7, %%mm4 \n\t"
1214 "psrlq $8, %%mm7 \n\t"
1215 #ifdef HAVE_MMX2
1216 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1217 "pshufw $0xF9, %%mm7, %%mm4 \n\t"
1218 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1219 "pshufw $0xFE, %%mm7, %%mm4 \n\t"
1220 "pminub %%mm4, %%mm7 \n\t"
1221 #else
1222 "movq %%mm7, %%mm1 \n\t"
1223 "psubusb %%mm4, %%mm1 \n\t"
1224 "psubb %%mm1, %%mm7 \n\t"
1225 "movq %%mm7, %%mm4 \n\t"
1226 "psrlq $16, %%mm7 \n\t"
1227 "movq %%mm7, %%mm1 \n\t"
1228 "psubusb %%mm4, %%mm1 \n\t"
1229 "psubb %%mm1, %%mm7 \n\t"
1230 "movq %%mm7, %%mm4 \n\t"
1231 "psrlq $32, %%mm7 \n\t"
1232 "movq %%mm7, %%mm1 \n\t"
1233 "psubusb %%mm4, %%mm1 \n\t"
1234 "psubb %%mm1, %%mm7 \n\t"
1235 #endif
1236
1237
1238 "movq %%mm6, %%mm4 \n\t"
1239 "psrlq $8, %%mm6 \n\t"
1240 #ifdef HAVE_MMX2
1241 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1242 "pshufw $0xF9, %%mm6, %%mm4 \n\t"
1243 "pmaxub %%mm4, %%mm6 \n\t"
1244 "pshufw $0xFE, %%mm6, %%mm4 \n\t"
1245 "pmaxub %%mm4, %%mm6 \n\t"
1246 #else
1247 "psubusb %%mm4, %%mm6 \n\t"
1248 "paddb %%mm4, %%mm6 \n\t"
1249 "movq %%mm6, %%mm4 \n\t"
1250 "psrlq $16, %%mm6 \n\t"
1251 "psubusb %%mm4, %%mm6 \n\t"
1252 "paddb %%mm4, %%mm6 \n\t"
1253 "movq %%mm6, %%mm4 \n\t"
1254 "psrlq $32, %%mm6 \n\t"
1255 "psubusb %%mm4, %%mm6 \n\t"
1256 "paddb %%mm4, %%mm6 \n\t"
1257 #endif
1258 "movq %%mm6, %%mm0 \n\t" // max
1259 "psubb %%mm7, %%mm6 \n\t" // max - min
1260 "movd %%mm6, %%ecx \n\t"
1261 "cmpb "MANGLE(deringThreshold)", %%cl \n\t"
1262 " jb 1f \n\t"
1263 "lea -24(%%"REG_SP"), %%"REG_c" \n\t"
1264 "and "ALIGN_MASK", %%"REG_c" \n\t"
1265 PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1266 "punpcklbw %%mm7, %%mm7 \n\t"
1267 "punpcklbw %%mm7, %%mm7 \n\t"
1268 "punpcklbw %%mm7, %%mm7 \n\t"
1269 "movq %%mm7, (%%"REG_c") \n\t"
1270
1271 "movq (%0), %%mm0 \n\t" // L10
1272 "movq %%mm0, %%mm1 \n\t" // L10
1273 "movq %%mm0, %%mm2 \n\t" // L10
1274 "psllq $8, %%mm1 \n\t"
1275 "psrlq $8, %%mm2 \n\t"
1276 "movd -4(%0), %%mm3 \n\t"
1277 "movd 8(%0), %%mm4 \n\t"
1278 "psrlq $24, %%mm3 \n\t"
1279 "psllq $56, %%mm4 \n\t"
1280 "por %%mm3, %%mm1 \n\t" // L00
1281 "por %%mm4, %%mm2 \n\t" // L20
1282 "movq %%mm1, %%mm3 \n\t" // L00
1283 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1284 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1285 "psubusb %%mm7, %%mm0 \n\t"
1286 "psubusb %%mm7, %%mm2 \n\t"
1287 "psubusb %%mm7, %%mm3 \n\t"
1288 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1289 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1290 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1291 "paddb %%mm2, %%mm0 \n\t"
1292 "paddb %%mm3, %%mm0 \n\t"
1293
1294 "movq (%%"REG_a"), %%mm2 \n\t" // L11
1295 "movq %%mm2, %%mm3 \n\t" // L11
1296 "movq %%mm2, %%mm4 \n\t" // L11
1297 "psllq $8, %%mm3 \n\t"
1298 "psrlq $8, %%mm4 \n\t"
1299 "movd -4(%%"REG_a"), %%mm5 \n\t"
1300 "movd 8(%%"REG_a"), %%mm6 \n\t"
1301 "psrlq $24, %%mm5 \n\t"
1302 "psllq $56, %%mm6 \n\t"
1303 "por %%mm5, %%mm3 \n\t" // L01
1304 "por %%mm6, %%mm4 \n\t" // L21
1305 "movq %%mm3, %%mm5 \n\t" // L01
1306 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1307 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1308 "psubusb %%mm7, %%mm2 \n\t"
1309 "psubusb %%mm7, %%mm4 \n\t"
1310 "psubusb %%mm7, %%mm5 \n\t"
1311 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1312 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1313 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1314 "paddb %%mm4, %%mm2 \n\t"
1315 "paddb %%mm5, %%mm2 \n\t"
1316 // 0, 2, 3, 1
1317 #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1318 "movq " #src ", " #sx " \n\t" /* src[0] */\
1319 "movq " #sx ", " #lx " \n\t" /* src[0] */\
1320 "movq " #sx ", " #t0 " \n\t" /* src[0] */\
1321 "psllq $8, " #lx " \n\t"\
1322 "psrlq $8, " #t0 " \n\t"\
1323 "movd -4" #src ", " #t1 " \n\t"\
1324 "psrlq $24, " #t1 " \n\t"\
1325 "por " #t1 ", " #lx " \n\t" /* src[-1] */\
1326 "movd 8" #src ", " #t1 " \n\t"\
1327 "psllq $56, " #t1 " \n\t"\
1328 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1329 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1330 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1331 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1332 PAVGB(lx, pplx) \
1333 "movq " #lx ", 8(%%"REG_c") \n\t"\
1334 "movq (%%"REG_c"), " #lx " \n\t"\
1335 "psubusb " #lx ", " #t1 " \n\t"\
1336 "psubusb " #lx ", " #t0 " \n\t"\
1337 "psubusb " #lx ", " #sx " \n\t"\
1338 "movq "MANGLE(b00)", " #lx " \n\t"\
1339 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1340 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1341 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1342 "paddb " #t1 ", " #t0 " \n\t"\
1343 "paddb " #t0 ", " #sx " \n\t"\
1344 \
1345 PAVGB(plx, pplx) /* filtered */\
1346 "movq " #dst ", " #t0 " \n\t" /* dst */\
1347 "movq " #t0 ", " #t1 " \n\t" /* dst */\
1348 "psubusb %3, " #t0 " \n\t"\
1349 "paddusb %3, " #t1 " \n\t"\
1350 PMAXUB(t0, pplx)\
1351 PMINUB(t1, pplx, t0)\
1352 "paddb " #sx ", " #ppsx " \n\t"\
1353 "paddb " #psx ", " #ppsx " \n\t"\
1354 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1355 "pand "MANGLE(b08)", " #ppsx " \n\t"\
1356 "pcmpeqb " #lx ", " #ppsx " \n\t"\
1357 "pand " #ppsx ", " #pplx " \n\t"\
1358 "pandn " #dst ", " #ppsx " \n\t"\
1359 "por " #pplx ", " #ppsx " \n\t"\
1360 "movq " #ppsx ", " #dst " \n\t"\
1361 "movq 8(%%"REG_c"), " #lx " \n\t"
1362
1363 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1364 REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1)
1365 /*
1366 0000000
1367 1111111
1368
1369 1111110
1370 1111101
1371 1111100
1372 1111011
1373 1111010
1374 1111001
1375
1376 1111000
1377 1110111
1378
1379 */
1380 //DERING_CORE(dst ,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
1381 DERING_CORE((%%REGa) ,(%%REGa, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1382 DERING_CORE((%%REGa, %1) ,(%%REGa, %1, 2),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1383 DERING_CORE((%%REGa, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1384 DERING_CORE((%0, %1, 4) ,(%%REGd) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1385 DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1386 DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1387 DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1388 DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1389
1390 "1: \n\t"
1391 : : "r" (src), "r" ((long)stride), "m" (c->pQPb), "m"(c->pQPb2)
1392 : "%"REG_a, "%"REG_d, "%"REG_c
1393 );
1394 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1395 int y;
1396 int min=255;
1397 int max=0;
1398 int avg;
1399 uint8_t *p;
1400 int s[10];
1401 const int QP2= c->QP/2 + 1;
1402
1403 for(y=1; y<9; y++)
1404 {
1405 int x;
1406 p= src + stride*y;
1407 for(x=1; x<9; x++)
1408 {
1409 p++;
1410 if(*p > max) max= *p;
1411 if(*p < min) min= *p;
1412 }
1413 }
1414 avg= (min + max + 1)>>1;
1415
1416 if(max - min <deringThreshold) return;
1417
1418 for(y=0; y<10; y++)
1419 {
1420 int t = 0;
1421
1422 if(src[stride*y + 0] > avg) t+= 1;
1423 if(src[stride*y + 1] > avg) t+= 2;
1424 if(src[stride*y + 2] > avg) t+= 4;
1425 if(src[stride*y + 3] > avg) t+= 8;
1426 if(src[stride*y + 4] > avg) t+= 16;
1427 if(src[stride*y + 5] > avg) t+= 32;
1428 if(src[stride*y + 6] > avg) t+= 64;
1429 if(src[stride*y + 7] > avg) t+= 128;
1430 if(src[stride*y + 8] > avg) t+= 256;
1431 if(src[stride*y + 9] > avg) t+= 512;
1432
1433 t |= (~t)<<16;
1434 t &= (t<<1) & (t>>1);
1435 s[y] = t;
1436 }
1437
1438 for(y=1; y<9; y++)
1439 {
1440 int t = s[y-1] & s[y] & s[y+1];
1441 t|= t>>16;
1442 s[y-1]= t;
1443 }
1444
1445 for(y=1; y<9; y++)
1446 {
1447 int x;
1448 int t = s[y-1];
1449
1450 p= src + stride*y;
1451 for(x=1; x<9; x++)
1452 {
1453 p++;
1454 if(t & (1<<x))
1455 {
1456 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1457 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
1458 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1459 f= (f + 8)>>4;
1460
1461 #ifdef DEBUG_DERING_THRESHOLD
1462 asm volatile("emms\n\t":);
1463 {
1464 static long long numPixels=0;
1465 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1466 // if((max-min)<20 || (max-min)*QP<200)
1467 // if((max-min)*QP < 500)
1468 // if(max-min<QP/2)
1469 if(max-min < 20)
1470 {
1471 static int numSkiped=0;
1472 static int errorSum=0;
1473 static int worstQP=0;
1474 static int worstRange=0;
1475 static int worstDiff=0;
1476 int diff= (f - *p);
1477 int absDiff= FFABS(diff);
1478 int error= diff*diff;
1479
1480 if(x==1 || x==8 || y==1 || y==8) continue;
1481
1482 numSkiped++;
1483 if(absDiff > worstDiff)
1484 {
1485 worstDiff= absDiff;
1486 worstQP= QP;
1487 worstRange= max-min;
1488 }
1489 errorSum+= error;
1490
1491 if(1024LL*1024LL*1024LL % numSkiped == 0)
1492 {
1493 av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, "
1494 "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1495 (float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
1496 worstDiff, (float)numSkiped/numPixels);
1497 }
1498 }
1499 }
1500 #endif
1501 if (*p + QP2 < f) *p= *p + QP2;
1502 else if(*p - QP2 > f) *p= *p - QP2;
1503 else *p=f;
1504 }
1505 }
1506 }
1507 #ifdef DEBUG_DERING_THRESHOLD
1508 if(max-min < 20)
1509 {
1510 for(y=1; y<9; y++)
1511 {
1512 int x;
1513 int t = 0;
1514 p= src + stride*y;
1515 for(x=1; x<9; x++)
1516 {
1517 p++;
1518 *p = FFMIN(*p + 20, 255);
1519 }
1520 }
1521 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1522 }
1523 #endif
1524 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1525 }
1526 #endif //HAVE_ALTIVEC
1527
1528 /**
1529 * Deinterlaces the given block by linearly interpolating every second line.
1530 * will be called for every 8x8 block and can read & write from line 4-15
1531 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1532 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1533 */
1534 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1535 {
1536 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1537 src+= 4*stride;
1538 asm volatile(
1539 "lea (%0, %1), %%"REG_a" \n\t"
1540 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
1541 // 0 1 2 3 4 5 6 7 8 9
1542 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
1543
1544 "movq (%0), %%mm0 \n\t"
1545 "movq (%%"REG_a", %1), %%mm1 \n\t"
1546 PAVGB(%%mm1, %%mm0)
1547 "movq %%mm0, (%%"REG_a") \n\t"
1548 "movq (%0, %1, 4), %%mm0 \n\t"
1549 PAVGB(%%mm0, %%mm1)
1550 "movq %%mm1, (%%"REG_a", %1, 2) \n\t"
1551 "movq (%%"REG_c", %1), %%mm1 \n\t"
1552 PAVGB(%%mm1, %%mm0)
1553 "movq %%mm0, (%%"REG_c") \n\t"
1554 "movq (%0, %1, 8), %%mm0 \n\t"
1555 PAVGB(%%mm0, %%mm1)
1556 "movq %%mm1, (%%"REG_c", %1, 2) \n\t"
1557
1558 : : "r" (src), "r" ((long)stride)
1559 : "%"REG_a, "%"REG_c
1560 );
1561 #else
1562 int a, b, x;
1563 src+= 4*stride;
1564
1565 for(x=0; x<2; x++){
1566 a= *(uint32_t*)&src[stride*0];
1567 b= *(uint32_t*)&src[stride*2];
1568 *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1569 a= *(uint32_t*)&src[stride*4];
1570 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1571 b= *(uint32_t*)&src[stride*6];
1572 *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1573 a= *(uint32_t*)&src[stride*8];
1574 *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1575 src += 4;
1576 }
1577 #endif
1578 }
1579
1580 /**
1581 * Deinterlaces the given block by cubic interpolating every second line.
1582 * will be called for every 8x8 block and can read & write from line 4-15
1583 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1584 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1585 * this filter will read lines 3-15 and write 7-13
1586 */
1587 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1588 {
1589 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1590 src+= stride*3;
1591 asm volatile(
1592 "lea (%0, %1), %%"REG_a" \n\t"
1593 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1594 "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t"
1595 "add %1, %%"REG_c" \n\t"
1596 "pxor %%mm7, %%mm7 \n\t"
1597 // 0 1 2 3 4 5 6 7 8 9 10
1598 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1599
1600 #define REAL_DEINT_CUBIC(a,b,c,d,e)\
1601 "movq " #a ", %%mm0 \n\t"\
1602 "movq " #b ", %%mm1 \n\t"\
1603 "movq " #d ", %%mm2 \n\t"\
1604 "movq " #e ", %%mm3 \n\t"\
1605 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\
1606 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\
1607 "movq %%mm0, %%mm2 \n\t"\
1608 "punpcklbw %%mm7, %%mm0 \n\t"\
1609 "punpckhbw %%mm7, %%mm2 \n\t"\
1610 "movq %%mm1, %%mm3 \n\t"\
1611 "punpcklbw %%mm7, %%mm1 \n\t"\
1612 "punpckhbw %%mm7, %%mm3 \n\t"\
1613 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\
1614 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\
1615 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\
1616 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\
1617 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\
1618 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\
1619 "packuswb %%mm3, %%mm1 \n\t"\
1620 "movq %%mm1, " #c " \n\t"
1621 #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e)
1622
1623 DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1))
1624 DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8))
1625 DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc))
1626 DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2))
1627
1628 : : "r" (src), "r" ((long)stride)
1629 : "%"REG_a, "%"REG_d, "%"REG_c
1630 );
1631 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1632 int x;
1633 src+= stride*3;
1634 for(x=0; x<8; x++)
1635 {
1636 src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4);
1637 src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4);
1638 src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4);
1639 src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4);
1640 src++;
1641 }
1642 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1643 }
1644
1645 /**
1646 * Deinterlaces the given block by filtering every second line with a (-1 4 2 4 -1) filter.
1647 * will be called for every 8x8 block and can read & write from line 4-15
1648 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1649 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1650 * this filter will read lines 4-13 and write 5-11
1651 */
1652 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
1653 {
1654 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1655 src+= stride*4;
1656 asm volatile(
1657 "lea (%0, %1), %%"REG_a" \n\t"
1658 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1659 "pxor %%mm7, %%mm7 \n\t"
1660 "movq (%2), %%mm0 \n\t"
1661 // 0 1 2 3 4 5 6 7 8 9 10
1662 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1663
1664 #define REAL_DEINT_FF(a,b,c,d)\
1665 "movq " #a ", %%mm1 \n\t"\
1666 "movq " #b ", %%mm2 \n\t"\
1667 "movq " #c ", %%mm3 \n\t"\
1668 "movq " #d ", %%mm4 \n\t"\
1669 PAVGB(%%mm3, %%mm1) \
1670 PAVGB(%%mm4, %%mm0) \
1671 "movq %%mm0, %%mm3 \n\t"\
1672 "punpcklbw %%mm7, %%mm0 \n\t"\
1673 "punpckhbw %%mm7, %%mm3 \n\t"\
1674 "movq %%mm1, %%mm4 \n\t"\
1675 "punpcklbw %%mm7, %%mm1 \n\t"\
1676 "punpckhbw %%mm7, %%mm4 \n\t"\
1677 "psllw $2, %%mm1 \n\t"\
1678 "psllw $2, %%mm4 \n\t"\
1679 "psubw %%mm0, %%mm1 \n\t"\
1680 "psubw %%mm3, %%mm4 \n\t"\
1681 "movq %%mm2, %%mm5 \n\t"\
1682 "movq %%mm2, %%mm0 \n\t"\
1683 "punpcklbw %%mm7, %%mm2 \n\t"\
1684 "punpckhbw %%mm7, %%mm5 \n\t"\
1685 "paddw %%mm2, %%mm1 \n\t"\
1686 "paddw %%mm5, %%mm4 \n\t"\
1687 "psraw $2, %%mm1 \n\t"\
1688 "psraw $2, %%mm4 \n\t"\
1689 "packuswb %%mm4, %%mm1 \n\t"\
1690 "movq %%mm1, " #b " \n\t"\
1691
1692 #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d)
1693
1694 DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2))
1695 DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1696 DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2))
1697 DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1698
1699 "movq %%mm0, (%2) \n\t"
1700 : : "r" (src), "r" ((long)stride), "r"(tmp)
1701 : "%"REG_a, "%"REG_d
1702 );
1703 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1704 int x;
1705 src+= stride*4;
1706 for(x=0; x<8; x++)
1707 {
1708 int t1= tmp[x];
1709 int t2= src[stride*1];
1710
1711 src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3);
1712 t1= src[stride*4];
1713 src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3);
1714 t2= src[stride*6];
1715 src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3);
1716 t1= src[stride*8];
1717 src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3);
1718 tmp[x]= t1;
1719
1720 src++;
1721 }
1722 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1723 }
1724
1725 /**
1726 * Deinterlaces the given block by filtering every line with a (-1 2 6 2 -1) filter.
1727 * will be called for every 8x8 block and can read & write from line 4-15
1728 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1729 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1730 * this filter will read lines 4-13 and write 4-11
1731 */
1732 static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2)
1733 {
1734 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1735 src+= stride*4;
1736 asm volatile(
1737 "lea (%0, %1), %%"REG_a" \n\t"
1738 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1739 "pxor %%mm7, %%mm7 \n\t"
1740 "movq (%2), %%mm0 \n\t"
1741 "movq (%3), %%mm1 \n\t"
1742 // 0 1 2 3 4 5 6 7 8 9 10
1743 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1744
1745 #define REAL_DEINT_L5(t1,t2,a,b,c)\
1746 "movq " #a ", %%mm2 \n\t"\
1747 "movq " #b ", %%mm3 \n\t"\
1748 "movq " #c ", %%mm4 \n\t"\
1749 PAVGB(t2, %%mm3) \
1750 PAVGB(t1, %%mm4) \
1751 "movq %%mm2, %%mm5 \n\t"\
1752 "movq %%mm2, " #t1 " \n\t"\
1753 "punpcklbw %%mm7, %%mm2 \n\t"\
1754 "punpckhbw %%mm7, %%mm5 \n\t"\
1755 "movq %%mm2, %%mm6 \n\t"\
1756 "paddw %%mm2, %%mm2 \n\t"\
1757 "paddw %%mm6, %%mm2 \n\t"\
1758 "movq %%mm5, %%mm6 \n\t"\
1759 "paddw %%mm5, %%mm5 \n\t"\
1760 "paddw %%mm6, %%mm5 \n\t"\
1761 "movq %%mm3, %%mm6 \n\t"\
1762 "punpcklbw %%mm7, %%mm3 \n\t"\
1763 "punpckhbw %%mm7, %%mm6 \n\t"\
1764 "paddw %%mm3, %%mm3 \n\t"\
1765 "paddw %%mm6, %%mm6 \n\t"\
1766 "paddw %%mm3, %%mm2 \n\t"\
1767 "paddw %%mm6, %%mm5 \n\t"\
1768 "movq %%mm4, %%mm6 \n\t"\
1769 "punpcklbw %%mm7, %%mm4 \n\t"\
1770 "punpckhbw %%mm7, %%mm6 \n\t"\
1771 "psubw %%mm4, %%mm2 \n\t"\
1772 "psubw %%mm6, %%mm5 \n\t"\
1773 "psraw $2, %%mm2 \n\t"\
1774 "psraw $2, %%mm5 \n\t"\
1775 "packuswb %%mm5, %%mm2 \n\t"\
1776 "movq %%mm2, " #a " \n\t"\
1777
1778 #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c)
1779
1780 DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) )
1781 DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2))
1782 DEINT_L5(%%mm0, %%mm1, (%%REGa, %1) , (%%REGa, %1, 2), (%0, %1, 4) )
1783 DEINT_L5(%%mm1, %%mm0, (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1784 DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) )
1785 DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2))
1786 DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) )
1787 DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1788
1789 "movq %%mm0, (%2) \n\t"
1790 "movq %%mm1, (%3) \n\t"
1791 : : "r" (src), "r" ((long)stride), "r"(tmp), "r"(tmp2)
1792 : "%"REG_a, "%"REG_d
1793 );
1794 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1795 int x;
1796 src+= stride*4;
1797 for(x=0; x<8; x++)
1798 {
1799 int t1= tmp[x];
1800 int t2= tmp2[x];
1801 int t3= src[0];
1802
1803 src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
1804 t1= src[stride*1];
1805 src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
1806 t2= src[stride*2];
1807 src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
1808 t3= src[stride*3];
1809 src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
1810 t1= src[stride*4];
1811 src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
1812 t2= src[stride*5];
1813 src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
1814 t3= src[stride*6];
1815 src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
1816 t1= src[stride*7];
1817 src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);
1818
1819 tmp[x]= t3;
1820 tmp2[x]= t1;
1821
1822 src++;
1823 }
1824 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1825 }
1826
1827 /**
1828 * Deinterlaces the given block by filtering all lines with a (1 2 1) filter.
1829 * will be called for every 8x8 block and can read & write from line 4-15
1830 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1831 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1832 * this filter will read lines 4-13 and write 4-11
1833 */
1834 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp)
1835 {
1836 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1837 src+= 4*stride;
1838 asm volatile(
1839 "lea (%0, %1), %%"REG_a" \n\t"
1840 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1841 // 0 1 2 3 4 5 6 7 8 9
1842 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1843
1844 "movq (%2), %%mm0 \n\t" // L0
1845 "movq (%%"REG_a"), %%mm1 \n\t" // L2
1846 PAVGB(%%mm1, %%mm0) // L0+L2
1847 "movq (%0), %%mm2 \n\t" // L1
1848 PAVGB(%%mm2, %%mm0)
1849 "movq %%mm0, (%0) \n\t"
1850 "movq (%%"REG_a", %1), %%mm0 \n\t" // L3
1851 PAVGB(%%mm0, %%mm2) // L1+L3
1852 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1853 "movq %%mm2, (%%"REG_a") \n\t"
1854 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4
1855 PAVGB(%%mm2, %%mm1) // L2+L4
1856 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1857 "movq %%mm1, (%%"REG_a", %1) \n\t"
1858 "movq (%0, %1, 4), %%mm1 \n\t" // L5
1859 PAVGB(%%mm1, %%mm0) // L3+L5
1860 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1861 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
1862 "movq (%%"REG_d"), %%mm0 \n\t" // L6
1863 PAVGB(%%mm0, %%mm2) // L4+L6
1864 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1865 "movq %%mm2, (%0, %1, 4) \n\t"
1866 "movq (%%"REG_d", %1), %%mm2 \n\t" // L7
1867 PAVGB(%%mm2, %%mm1) // L5+L7
1868 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1869 "movq %%mm1, (%%"REG_d") \n\t"
1870 "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8
1871 PAVGB(%%mm1, %%mm0) // L6+L8
1872 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1873 "movq %%mm0, (%%"REG_d", %1) \n\t"
1874 "movq (%0, %1, 8), %%mm0 \n\t" // L9
1875 PAVGB(%%mm0, %%mm2) // L7+L9
1876 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1877 "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1878 "movq %%mm1, (%2) \n\t"
1879
1880 : : "r" (src), "r" ((long)stride), "r" (tmp)
1881 : "%"REG_a, "%"REG_d
1882 );
1883 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1884 int a, b, c, x;
1885 src+= 4*stride;
1886
1887 for(x=0; x<2; x++){
1888 a= *(uint32_t*)&tmp[stride*0];
1889 b= *(uint32_t*)&src[stride*0];
1890 c= *(uint32_t*)&src[stride*1];
1891 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1892 *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1893
1894 a= *(uint32_t*)&src[stride*2];
1895 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1896 *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1897
1898 b= *(uint32_t*)&src[stride*3];
1899 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1900 *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1901
1902 c= *(uint32_t*)&src[stride*4];
1903 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1904 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1905
1906 a= *(uint32_t*)&src[stride*5];
1907 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1908 *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1909
1910 b= *(uint32_t*)&src[stride*6];
1911 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1912 *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1913
1914 c= *(uint32_t*)&src[stride*7];
1915 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1916 *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1917
1918 a= *(uint32_t*)&src[stride*8];
1919 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1920 *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1921
1922 *(uint32_t*)&tmp[stride*0]= c;
1923 src += 4;
1924 tmp += 4;
1925 }
1926 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1927 }
1928
1929 /**
1930 * Deinterlaces the given block by applying a median filter to every second line.
1931 * will be called for every 8x8 block and can read & write from line 4-15,
1932 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1933 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1934 */
1935 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1936 {
1937 #ifdef HAVE_MMX
1938 src+= 4*stride;
1939 #ifdef HAVE_MMX2
1940 asm volatile(
1941 "lea (%0, %1), %%"REG_a" \n\t"
1942 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1943 // 0 1 2 3 4 5 6 7 8 9
1944 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1945
1946 "movq (%0), %%mm0 \n\t" //
1947 "movq (%%"REG_a", %1), %%mm2 \n\t" //
1948 "movq (%%"REG_a"), %%mm1 \n\t" //
1949 "movq %%mm0, %%mm3 \n\t"
1950 "pmaxub %%mm1, %%mm0 \n\t" //
1951 "pminub %%mm3, %%mm1 \n\t" //
1952 "pmaxub %%mm2, %%mm1 \n\t" //
1953 "pminub %%mm1, %%mm0 \n\t"
1954 "movq %%mm0, (%%"REG_a") \n\t"
1955
1956 "movq (%0, %1, 4), %%mm0 \n\t" //
1957 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" //
1958 "movq %%mm2, %%mm3 \n\t"
1959 "pmaxub %%mm1, %%mm2 \n\t" //
1960 "pminub %%mm3, %%mm1 \n\t" //
1961 "pmaxub %%mm0, %%mm1 \n\t" //
1962 "pminub %%mm1, %%mm2 \n\t"
1963 "movq %%mm2, (%%"REG_a", %1, 2) \n\t"
1964
1965 "movq (%%"REG_d"), %%mm2 \n\t" //
1966 "movq (%%"REG_d", %1), %%mm1 \n\t" //
1967 "movq %%mm2, %%mm3 \n\t"
1968 "pmaxub %%mm0, %%mm2 \n\t" //
1969 "pminub %%mm3, %%mm0 \n\t" //
1970 "pmaxub %%mm1, %%mm0 \n\t" //
1971 "pminub %%mm0, %%mm2 \n\t"
1972 "movq %%mm2, (%%"REG_d") \n\t"
1973
1974 "movq (%%"REG_d", %1, 2), %%mm2 \n\t" //
1975 "movq (%0, %1, 8), %%mm0 \n\t" //
1976 "movq %%mm2, %%mm3 \n\t"
1977 "pmaxub %%mm0, %%mm2 \n\t" //
1978 "pminub %%mm3, %%mm0 \n\t" //
1979 "pmaxub %%mm1, %%mm0 \n\t" //
1980 "pminub %%mm0, %%mm2 \n\t"
1981 "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1982
1983
1984 : : "r" (src), "r" ((long)stride)
1985 : "%"REG_a, "%"REG_d
1986 );
1987
1988 #else // MMX without MMX2
1989 asm volatile(
1990 "lea (%0, %1), %%"REG_a" \n\t"
1991 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1992 // 0 1 2 3 4 5 6 7 8 9
1993 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1994 "pxor %%mm7, %%mm7 \n\t"
1995
1996 #define REAL_MEDIAN(a,b,c)\
1997 "movq " #a ", %%mm0 \n\t"\
1998 "movq " #b ", %%mm2 \n\t"\
1999 "movq " #c ", %%mm1 \n\t"\
2000 "movq %%mm0, %%mm3 \n\t"\
2001 "movq %%mm1, %%mm4 \n\t"\
2002 "movq %%mm2, %%mm5 \n\t"\
2003 "psubusb %%mm1, %%mm3 \n\t"\
2004 "psubusb %%mm2, %%mm4 \n\t"\
2005 "psubusb %%mm0, %%mm5 \n\t"\
2006 "pcmpeqb %%mm7, %%mm3 \n\t"\
2007 "pcmpeqb %%mm7, %%mm4 \n\t"\
2008 "pcmpeqb %%mm7, %%mm5 \n\t"\
2009 "movq %%mm3, %%mm6 \n\t"\
2010 "pxor %%mm4, %%mm3 \n\t"\
2011 "pxor %%mm5, %%mm4 \n\t"\
2012 "pxor %%mm6, %%mm5 \n\t"\
2013 "por %%mm3, %%mm1 \n\t"\
2014 "por %%mm4, %%mm2 \n\t"\
2015 "por %%mm5, %%mm0 \n\t"\
2016 "pand %%mm2, %%mm0 \n\t"\
2017 "pand %%mm1, %%mm0 \n\t"\
2018 "movq %%mm0, " #b " \n\t"
2019 #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c)
2020
2021 MEDIAN((%0) , (%%REGa) , (%%REGa, %1))
2022 MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4))
2023 MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1))
2024 MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8))
2025
2026 : : "r" (src), "r" ((long)stride)
2027 : "%"REG_a, "%"REG_d
2028 );
2029 #endif //HAVE_MMX2
2030 #else //HAVE_MMX
2031 int x, y;
2032 src+= 4*stride;
2033 // FIXME - there should be a way to do a few columns in parallel like w/mmx
2034 for(x=0; x<8; x++)
2035 {
2036 uint8_t *colsrc = src;
2037 for (y=0; y<4; y++)
2038 {
2039 int a, b, c, d, e, f;
2040 a = colsrc[0 ];
2041 b = colsrc[stride ];
2042 c = colsrc[stride*2];
2043 d = (a-b)>>31;
2044 e = (b-c)>>31;
2045 f = (c-a)>>31;
2046 colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f));
2047 colsrc += stride*2;
2048 }
2049 src++;
2050 }
2051 #endif //HAVE_MMX
2052 }
2053
2054 #ifdef HAVE_MMX
2055 /**
2056 * transposes and shift the given 8x8 Block into dst1 and dst2
2057 */
2058 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
2059 {
2060 asm(
2061 "lea (%0, %1), %%"REG_a" \n\t"
2062 // 0 1 2 3 4 5 6 7 8 9
2063 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2064 "movq (%0), %%mm0 \n\t" // 12345678
2065 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
2066 "movq %%mm0, %%mm2 \n\t" // 12345678
2067 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2068 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2069
2070 "movq (%%"REG_a", %1), %%mm1 \n\t"
2071 "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2072 "movq %%mm1, %%mm4 \n\t"
2073 "punpcklbw %%mm3, %%mm1 \n\t"
2074 "punpckhbw %%mm3, %%mm4 \n\t"
2075
2076 "movq %%mm0, %%mm3 \n\t"
2077 "punpcklwd %%mm1, %%mm0 \n\t"
2078 "punpckhwd %%mm1, %%mm3 \n\t"
2079 "movq %%mm2, %%mm1 \n\t"
2080 "punpcklwd %%mm4, %%mm2 \n\t"
2081 "punpckhwd %%mm4, %%mm1 \n\t"
2082
2083 "movd %%mm0, 128(%2) \n\t"
2084 "psrlq $32, %%mm0 \n\t"
2085 "movd %%mm0, 144(%2) \n\t"
2086 "movd %%mm3, 160(%2) \n\t"
2087 "psrlq $32, %%mm3 \n\t"
2088 "movd %%mm3, 176(%2) \n\t"
2089 "movd %%mm3, 48(%3) \n\t"
2090 "movd %%mm2, 192(%2) \n\t"
2091 "movd %%mm2, 64(%3) \n\t"
2092 "psrlq $32, %%mm2 \n\t"
2093 "movd %%mm2, 80(%3) \n\t"
2094 "movd %%mm1, 96(%3) \n\t"
2095 "psrlq $32, %%mm1 \n\t"
2096 "movd %%mm1, 112(%3) \n\t"
2097
2098 "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t"
2099
2100 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
2101 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
2102 "movq %%mm0, %%mm2 \n\t" // 12345678
2103 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2104 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2105
2106 "movq (%%"REG_a", %1), %%mm1 \n\t"
2107 "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2108 "movq %%mm1, %%mm4 \n\t"
2109 "punpcklbw %%mm3, %%mm1 \n\t"
2110 "punpckhbw %%mm3, %%mm4 \n\t"
2111
2112 "movq %%mm0, %%mm3 \n\t"
2113 "punpcklwd %%mm1, %%mm0 \n\t"
2114 "punpckhwd %%mm1, %%mm3 \n\t"
2115 "movq %%mm2, %%mm1 \n\t"
2116 "punpcklwd %%mm4, %%mm2 \n\t"
2117 "punpckhwd %%mm4, %%mm1 \n\t"
2118
2119 "movd %%mm0, 132(%2) \n\t"
2120 "psrlq $32, %%mm0 \n\t"
2121 "movd %%mm0, 148(%2) \n\t"
2122 "movd %%mm3, 164(%2) \n\t"
2123 "psrlq $32, %%mm3 \n\t"
2124 "movd %%mm3, 180(%2) \n\t"
2125 "movd %%mm3, 52(%3) \n\t"
2126 "movd %%mm2, 196(%2) \n\t"
2127 "movd %%mm2, 68(%3) \n\t"
2128 "psrlq $32, %%mm2 \n\t"
2129 "movd %%mm2, 84(%3) \n\t"
2130 "movd %%mm1, 100(%3) \n\t"
2131 "psrlq $32, %%mm1 \n\t"
2132 "movd %%mm1, 116(%3) \n\t"
2133
2134
2135 :: "r" (src), "r" ((long)srcStride), "r" (dst1), "r" (dst2)
2136 : "%"REG_a
2137 );
2138 }
2139
2140 /**
2141 * transposes the given 8x8 block
2142 */
2143 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
2144 {
2145 asm(
2146 "lea (%0, %1), %%"REG_a" \n\t"
2147 "lea (%%"REG_a",%1,4), %%"REG_d" \n\t"
2148 // 0 1 2 3 4 5 6 7 8 9
2149 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2150 "movq (%2), %%mm0 \n\t" // 12345678
2151 "movq 16(%2), %%mm1 \n\t" // abcdefgh
2152 "movq %%mm0, %%mm2 \n\t" // 12345678
2153 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2154 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2155
2156 "movq 32(%2), %%mm1 \n\t"
2157 "movq 48(%2), %%mm3 \n\t"
2158 "movq %%mm1, %%mm4 \n\t"
2159 "punpcklbw %%mm3, %%mm1 \n\t"
2160 "punpckhbw %%mm3, %%mm4 \n\t"
2161
2162 "movq %%mm0, %%mm3 \n\t"
2163 "punpcklwd %%mm1, %%mm0 \n\t"
2164 "punpckhwd %%mm1, %%mm3 \n\t"
2165 "movq %%mm2, %%mm1 \n\t"
2166 "punpcklwd %%mm4, %%mm2 \n\t"
2167 "punpckhwd %%mm4, %%mm1 \n\t"
2168
2169 "movd %%mm0, (%0) \n\t"
2170 "psrlq $32, %%mm0 \n\t"
2171 "movd %%mm0, (%%"REG_a") \n\t"
2172 "movd %%mm3, (%%"REG_a", %1) \n\t"
2173 "psrlq $32, %%mm3 \n\t"
2174 "movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2175 "movd %%mm2, (%0, %1, 4) \n\t"
2176 "psrlq $32, %%mm2 \n\t"
2177 "movd %%mm2, (%%"REG_d") \n\t"
2178 "movd %%mm1, (%%"REG_d", %1) \n\t"
2179 "psrlq $32, %%mm1 \n\t"
2180 "movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2181
2182
2183 "movq 64(%2), %%mm0 \n\t" // 12345678
2184 "movq 80(%2), %%mm1 \n\t" // abcdefgh
2185 "movq %%mm0, %%mm2 \n\t" // 12345678
2186 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2187 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2188
2189 "movq 96(%2), %%mm1 \n\t"
2190 "movq 112(%2), %%mm3 \n\t"
2191 "movq %%mm1, %%mm4 \n\t"
2192 "punpcklbw %%mm3, %%mm1 \n\t"
2193 "punpckhbw %%mm3, %%mm4 \n\t"
2194
2195 "movq %%mm0, %%mm3 \n\t"
2196 "punpcklwd %%mm1, %%mm0 \n\t"
2197 "punpckhwd %%mm1, %%mm3 \n\t"
2198 "movq %%mm2, %%mm1 \n\t"
2199 "punpcklwd %%mm4, %%mm2 \n\t"
2200 "punpckhwd %%mm4, %%mm1 \n\t"
2201
2202 "movd %%mm0, 4(%0) \n\t"
2203 "psrlq $32, %%mm0 \n\t"
2204 "movd %%mm0, 4(%%"REG_a") \n\t"
2205 "movd %%mm3, 4(%%"REG_a", %1) \n\t"
2206 "psrlq $32, %%mm3 \n\t"
2207 "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t"
2208 "movd %%mm2, 4(%0, %1, 4) \n\t"
2209 "psrlq $32, %%mm2 \n\t"
2210 "movd %%mm2, 4(%%"REG_d") \n\t"
2211 "movd %%mm1, 4(%%"REG_d", %1) \n\t"
2212 "psrlq $32, %%mm1 \n\t"
2213 "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t"
2214
2215 :: "r" (dst), "r" ((long)dstStride), "r" (src)
2216 : "%"REG_a, "%"REG_d
2217 );
2218 }
2219 #endif //HAVE_MMX
2220 //static long test=0;
2221
2222 #ifndef HAVE_ALTIVEC
2223 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2224 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2225 {
2226 // to save a register (FIXME do this outside of the loops)
2227 tempBluredPast[127]= maxNoise[0];
2228 tempBluredPast[128]= maxNoise[1];
2229 tempBluredPast[129]= maxNoise[2];
2230
2231 #define FAST_L2_DIFF
2232 //#define L1_DIFF //u should change the thresholds too if u try that one
2233 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2234 asm volatile(
2235 "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride
2236 "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride
2237 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2238 // 0 1 2 3 4 5 6 7 8 9
2239 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
2240 //FIXME reorder?
2241 #ifdef L1_DIFF //needs mmx2
2242 "movq (%0), %%mm0 \n\t" // L0
2243 "psadbw (%1), %%mm0 \n\t" // |L0-R0|
2244 "movq (%0, %2), %%mm1 \n\t" // L1
2245 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2246 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2247 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2248 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2249 "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3|
2250
2251 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2252 "paddw %%mm1, %%mm0 \n\t"
2253 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2254 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2255 "paddw %%mm2, %%mm0 \n\t"
2256 "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5|
2257 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2258 "paddw %%mm3, %%mm0 \n\t"
2259 "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6|
2260 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2261 "paddw %%mm4, %%mm0 \n\t"
2262 "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7|
2263 "paddw %%mm5, %%mm6 \n\t"
2264 "paddw %%mm7, %%mm6 \n\t"
2265 "paddw %%mm6, %%mm0 \n\t"
2266 #else //L1_DIFF
2267 #if defined (FAST_L2_DIFF)
2268 "pcmpeqb %%mm7, %%mm7 \n\t"
2269 "movq "MANGLE(b80)", %%mm6 \n\t"
2270 "pxor %%mm0, %%mm0 \n\t"
2271 #define REAL_L2_DIFF_CORE(a, b)\
2272 "movq " #a ", %%mm5 \n\t"\
2273 "movq " #b ", %%mm2 \n\t"\
2274 "pxor %%mm7, %%mm2 \n\t"\
2275 PAVGB(%%mm2, %%mm5)\
2276 "paddb %%mm6, %%mm5 \n\t"\
2277 "movq %%mm5, %%mm2 \n\t"\
2278 "psllw $8, %%mm5 \n\t"\
2279 "pmaddwd %%mm5, %%mm5 \n\t"\
2280 "pmaddwd %%mm2, %%mm2 \n\t"\
2281 "paddd %%mm2, %%mm5 \n\t"\
2282 "psrld $14, %%mm5 \n\t"\
2283 "paddd %%mm5, %%mm0 \n\t"
2284
2285 #else //defined (FAST_L2_DIFF)
2286 "pxor %%mm7, %%mm7 \n\t"
2287 "pxor %%mm0, %%mm0 \n\t"
2288 #define REAL_L2_DIFF_CORE(a, b)\
2289 "movq " #a ", %%mm5 \n\t"\
2290 "movq " #b ", %%mm2 \n\t"\
2291 "movq %%mm5, %%mm1 \n\t"\
2292 "movq %%mm2, %%mm3 \n\t"\
2293 "punpcklbw %%mm7, %%mm5 \n\t"\
2294 "punpckhbw %%mm7, %%mm1 \n\t"\
2295 "punpcklbw %%mm7, %%mm2 \n\t"\
2296 "punpckhbw %%mm7, %%mm3 \n\t"\
2297 "psubw %%mm2, %%mm5 \n\t"\
2298 "psubw %%mm3, %%mm1 \n\t"\
2299 "pmaddwd %%mm5, %%mm5 \n\t"\
2300 "pmaddwd %%mm1, %%mm1 \n\t"\
2301 "paddd %%mm1, %%mm5 \n\t"\
2302 "paddd %%mm5, %%mm0 \n\t"
2303
2304 #endif //defined (FAST_L2_DIFF)
2305
2306 #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b)
2307
2308 L2_DIFF_CORE((%0) , (%1))
2309 L2_DIFF_CORE((%0, %2) , (%1, %2))
2310 L2_DIFF_CORE((%0, %2, 2) , (%1, %2, 2))
2311 L2_DIFF_CORE((%0, %%REGa) , (%1, %%REGa))
2312 L2_DIFF_CORE((%0, %2, 4) , (%1, %2, 4))
2313 L2_DIFF_CORE((%0, %%REGd) , (%1, %%REGd))
2314 L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2))
2315 L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc))
2316
2317 #endif //L1_DIFF
2318
2319 "movq %%mm0, %%mm4 \n\t"
2320 "psrlq $32, %%mm0 \n\t"
2321 "paddd %%mm0, %%mm4 \n\t"
2322 "movd %%mm4, %%ecx \n\t"
2323 "shll $2, %%ecx \n\t"
2324 "mov %3, %%"REG_d" \n\t"
2325 "addl -4(%%"REG_d"), %%ecx \n\t"
2326 "addl 4(%%"REG_d"), %%ecx \n\t"
2327 "addl -1024(%%"REG_d"), %%ecx \n\t"
2328 "addl $4, %%ecx \n\t"
2329 "addl 1024(%%"REG_d"), %%ecx \n\t"
2330 "shrl $3, %%ecx \n\t"
2331 "movl %%ecx, (%%"REG_d") \n\t"
2332
2333 // "mov %3, %%"REG_c" \n\t"
2334 // "mov %%"REG_c", test \n\t"
2335 // "jmp 4f \n\t"
2336 "cmpl 512(%%"REG_d"), %%ecx \n\t"
2337 " jb 2f \n\t"
2338 "cmpl 516(%%"REG_d"), %%ecx \n\t"
2339 " jb 1f \n\t"
2340
2341 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2342 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2343 "movq (%0), %%mm0 \n\t" // L0
2344 "movq (%0, %2), %%mm1 \n\t" // L1
2345 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2346 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2347 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2348 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2349 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2350 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2351 "movq %%mm0, (%1) \n\t" // L0
2352 "movq %%mm1, (%1, %2) \n\t" // L1
2353 "movq %%mm2, (%1, %2, 2) \n\t" // L2
2354 "movq %%mm3, (%1, %%"REG_a") \n\t" // L3
2355 "movq %%mm4, (%1, %2, 4) \n\t" // L4
2356 "movq %%mm5, (%1, %%"REG_d") \n\t" // L5
2357 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6
2358 "movq %%mm7, (%1, %%"REG_c") \n\t" // L7
2359 "jmp 4f \n\t"
2360
2361 "1: \n\t"
2362 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2363 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2364 "movq (%0), %%mm0 \n\t" // L0
2365 PAVGB((%1), %%mm0) // L0
2366 "movq (%0, %2), %%mm1 \n\t" // L1
2367 PAVGB((%1, %2), %%mm1) // L1
2368 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2369 PAVGB((%1, %2, 2), %%mm2) // L2
2370 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2371 PAVGB((%1, %%REGa), %%mm3) // L3
2372 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2373 PAVGB((%1, %2, 4), %%mm4) // L4
2374 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2375 PAVGB((%1, %%REGd), %%mm5) // L5
2376 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2377 PAVGB((%1, %%REGa, 2), %%mm6) // L6
2378 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2379 PAVGB((%1, %%REGc), %%mm7) // L7
2380 "movq %%mm0, (%1) \n\t" // R0
2381 "movq %%mm1, (%1, %2) \n\t" // R1
2382 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2383 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2384 "movq %%mm4, (%1, %2, 4) \n\t" // R4
2385 "movq %%mm5, (%1, %%"REG_d") \n\t" // R5
2386 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6
2387 "movq %%mm7, (%1, %%"REG_c") \n\t" // R7
2388 "movq %%mm0, (%0) \n\t" // L0
2389 "movq %%mm1, (%0, %2) \n\t" // L1
2390 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2391 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2392 "movq %%mm4, (%0, %2, 4) \n\t" // L4
2393 "movq %%mm5, (%0, %%"REG_d") \n\t" // L5
2394 "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6
2395 "movq %%mm7, (%0, %%"REG_c") \n\t" // L7
2396 "jmp 4f \n\t"
2397
2398 "2: \n\t"
2399 "cmpl 508(%%"REG_d"), %%ecx \n\t"
2400 " jb 3f \n\t"
2401
2402 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2403 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2404 "movq (%0), %%mm0 \n\t" // L0
2405 "movq (%0, %2), %%mm1 \n\t" // L1
2406 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2407 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2408 "movq (%1), %%mm4 \n\t" // R0
2409 "movq (%1, %2), %%mm5 \n\t" // R1
2410 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2411 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2412 PAVGB(%%mm4, %%mm0)
2413 PAVGB(%%mm5, %%mm1)
2414 PAVGB(%%mm6, %%mm2)
2415 PAVGB(%%mm7, %%mm3)
2416 PAVGB(%%mm4, %%mm0)
2417 PAVGB(%%mm5, %%mm1)
2418 PAVGB(%%mm6, %%mm2)
2419 PAVGB(%%mm7, %%mm3)
2420 "movq %%mm0, (%1) \n\t" // R0
2421 "movq %%mm1, (%1, %2) \n\t" // R1
2422 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2423 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2424 "movq %%mm0, (%0) \n\t" // L0
2425 "movq %%mm1, (%0, %2) \n\t" // L1
2426 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2427 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2428
2429 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2430 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2431 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2432 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2433 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2434 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2435 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2436 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2437 PAVGB(%%mm4, %%mm0)
2438 PAVGB(%%mm5, %%mm1)
2439 PAVGB(%%mm6, %%mm2)
2440 PAVGB(%%mm7, %%mm3)
2441 PAVGB(%%mm4, %%mm0)
2442 PAVGB(%%mm5, %%mm1)
2443 PAVGB(%%mm6, %%mm2)
2444 PAVGB(%%mm7, %%mm3)
2445 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2446 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2447 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2448 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2449 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2450 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2451 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2452 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2453 "jmp 4f \n\t"
2454
2455 "3: \n\t"
2456 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2457 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2458 "movq (%0), %%mm0 \n\t" // L0
2459 "movq (%0, %2), %%mm1 \n\t" // L1
2460 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2461 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2462 "movq (%1), %%mm4 \n\t" // R0
2463 "movq (%1, %2), %%mm5 \n\t" // R1
2464 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2465 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2466 PAVGB(%%mm4, %%mm0)
2467 PAVGB(%%mm5, %%mm1)
2468 PAVGB(%%mm6, %%mm2)
2469 PAVGB(%%mm7, %%mm3)
2470 PAVGB(%%mm4, %%mm0)
2471 PAVGB(%%mm5, %%mm1)
2472 PAVGB(%%mm6, %%mm2)
2473 PAVGB(%%mm7, %%mm3)
2474 PAVGB(%%mm4, %%mm0)
2475 PAVGB(%%mm5, %%mm1)
2476 PAVGB(%%mm6, %%mm2)
2477 PAVGB(%%mm7, %%mm3)
2478 "movq %%mm0, (%1) \n\t" // R0
2479 "movq %%mm1, (%1, %2) \n\t" // R1
2480 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2481 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2482 "movq %%mm0, (%0) \n\t" // L0
2483 "movq %%mm1, (%0, %2) \n\t" // L1
2484 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2485 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2486
2487 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2488 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2489 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2490 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2491 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2492 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2493 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2494 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2495 PAVGB(%%mm4, %%mm0)
2496 PAVGB(%%mm5, %%mm1)
2497 PAVGB(%%mm6, %%mm2)
2498 PAVGB(%%mm7, %%mm3)
2499 PAVGB(%%mm4, %%mm0)
2500 PAVGB(%%mm5, %%mm1)
2501 PAVGB(%%mm6, %%mm2)
2502 PAVGB(%%mm7, %%mm3)
2503 PAVGB(%%mm4, %%mm0)
2504 PAVGB(%%mm5, %%mm1)
2505 PAVGB(%%mm6, %%mm2)
2506 PAVGB(%%mm7, %%mm3)
2507 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2508 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2509 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2510 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2511 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2512 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2513 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2514 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2515
2516 "4: \n\t"
2517
2518 :: "r" (src), "r" (tempBlured), "r"((long)stride), "m" (tempBluredPast)
2519 : "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2520 );
2521 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2522 {
2523 int y;
2524 int d=0;
2525 // int sysd=0;
2526 int i;
2527
2528 for(y=0; y<8; y++)
2529 {
2530 int x;
2531 for(x=0; x<8; x++)
2532 {
2533 int ref= tempBlured[ x + y*stride ];
2534 int cur= src[ x + y*stride ];
2535 int d1=ref - cur;
2536 // if(x==0 || x==7) d1+= d1>>1;
2537 // if(y==0 || y==7) d1+= d1>>1;
2538 // d+= FFABS(d1);
2539 d+= d1*d1;
2540 // sysd+= d1;
2541 }
2542 }
2543 i=d;
2544 d= (
2545 4*d
2546 +(*(tempBluredPast-256))
2547 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
2548 +(*(tempBluredPast+256))
2549 +4)>>3;
2550 *tempBluredPast=i;
2551 // ((*tempBluredPast)*3 + d + 2)>>2;
2552
2553 /*
2554 Switch between
2555 1 0 0 0 0 0 0 (0)
2556 64 32 16 8 4 2 1 (1)
2557 64 48 36 27 20 15 11 (33) (approx)
2558 64 56 49 43 37 33 29 (200) (approx)
2559 */
2560 if(d > maxNoise[1])
2561 {
2562 if(d < maxNoise[2])
2563 {
2564 for(y=0; y<8; y++)
2565 {
2566 int x;
2567 for(x=0; x<8; x++)
2568 {
2569 int ref= tempBlured[ x + y*stride ];
2570 int cur= src[ x + y*stride ];
2571 tempBlured[ x + y*stride ]=
2572 src[ x + y*stride ]=
2573 (ref + cur + 1)>>1;
2574 }
2575 }
2576 }
2577 else
2578 {
2579 for(y=0; y<8; y++)
2580 {
2581 int x;
2582 for(x=0; x<8; x++)
2583 {
2584 tempBlured[ x + y*stride ]= src[ x + y*stride ];
2585 }
2586 }
2587 }
2588 }
2589 else
2590 {
2591 if(d < maxNoise[0])
2592 {
2593 for(y=0; y<8; y++)
2594 {
2595 int x;
2596 for(x=0; x<8; x++)
2597 {
2598 int ref= tempBlured[ x + y*stride ];
2599 int cur= src[ x + y*stride ];
2600 tempBlured[ x + y*stride ]=
2601 src[ x + y*stride ]=
2602 (ref*7 + cur + 4)>>3;
2603 }
2604 }
2605 }
2606 else
2607 {
2608 for(y=0; y<8; y++)
2609 {
2610 int x;
2611 for(x=0; x<8; x++)
2612 {
2613 int ref= tempBlured[ x + y*stride ];
2614 int cur= src[ x + y*stride ];
2615 tempBlured[ x + y*stride ]=
2616 src[ x + y*stride ]=
2617 (ref*3 + cur + 2)>>2;
2618 }
2619 }
2620 }
2621 }
2622 }
2623 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2624 }
2625 #endif //HAVE_ALTIVEC
2626
2627 #ifdef HAVE_MMX
2628 /**
2629 * accurate deblock filter
2630 */
2631 static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){
2632 int64_t dc_mask, eq_mask, both_masks;
2633 int64_t sums[10*8*2];
2634 src+= step*3; // src points to begin of the 8x8 Block
2635 //START_TIMER
2636 asm volatile(
2637 "movq %0, %%mm7 \n\t"
2638 "movq %1, %%mm6 \n\t"
2639 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
2640 );
2641
2642 asm volatile(
2643 "lea (%2, %3), %%"REG_a" \n\t"
2644 // 0 1 2 3 4 5 6 7 8 9
2645 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
2646
2647 "movq (%2), %%mm0 \n\t"
2648 "movq (%%"REG_a"), %%mm1 \n\t"
2649 "movq %%mm1, %%mm3 \n\t"
2650 "movq %%mm1, %%mm4 \n\t"
2651 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
2652 "paddb %%mm7, %%mm0 \n\t"
2653 "pcmpgtb %%mm6, %%mm0 \n\t"
2654
2655 "movq (%%"REG_a",%3), %%mm2 \n\t"
2656 PMAXUB(%%mm2, %%mm4)
2657 PMINUB(%%mm2, %%mm3, %%mm5)
2658 "psubb %%mm2, %%mm1 \n\t"
2659 "paddb %%mm7, %%mm1 \n\t"
2660 "pcmpgtb %%mm6, %%mm1 \n\t"
2661 "paddb %%mm1, %%mm0 \n\t"
2662
2663 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2664 PMAXUB(%%mm1, %%mm4)
2665 PMINUB(%%mm1, %%mm3, %%mm5)
2666 "psubb %%mm1, %%mm2 \n\t"
2667 "paddb %%mm7, %%mm2 \n\t"
2668 "pcmpgtb %%mm6, %%mm2 \n\t"
2669 "paddb %%mm2, %%mm0 \n\t"
2670
2671 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
2672
2673 "movq (%2, %3, 4), %%mm2 \n\t"
2674 PMAXUB(%%mm2, %%mm4)
2675 PMINUB(%%mm2, %%mm3, %%mm5)
2676 "psubb %%mm2, %%mm1 \n\t"
2677 "paddb %%mm7, %%mm1 \n\t"
2678 "pcmpgtb %%mm6, %%mm1 \n\t"
2679 "paddb %%mm1, %%mm0 \n\t"
2680
2681 "movq (%%"REG_a"), %%mm1 \n\t"
2682 PMAXUB(%%mm1, %%mm4)
2683 PMINUB(%%mm1, %%mm3, %%mm5)
2684 "psubb %%mm1, %%mm2 \n\t"
2685 "paddb %%mm7, %%mm2 \n\t"
2686 "pcmpgtb %%mm6, %%mm2 \n\t"
2687 "paddb %%mm2, %%mm0 \n\t"
2688
2689 "movq (%%"REG_a", %3), %%mm2 \n\t"
2690 PMAXUB(%%mm2, %%mm4)
2691 PMINUB(%%mm2, %%mm3, %%mm5)
2692 "psubb %%mm2, %%mm1 \n\t"
2693 "paddb %%mm7, %%mm1 \n\t"
2694 "pcmpgtb %%mm6, %%mm1 \n\t"
2695 "paddb %%mm1, %%mm0 \n\t"
2696
2697 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2698 PMAXUB(%%mm1, %%mm4)
2699 PMINUB(%%mm1, %%mm3, %%mm5)
2700 "psubb %%mm1, %%mm2 \n\t"
2701 "paddb %%mm7, %%mm2 \n\t"
2702 "pcmpgtb %%mm6, %%mm2 \n\t"
2703 "paddb %%mm2, %%mm0 \n\t"
2704
2705 "movq (%2, %3, 8), %%mm2 \n\t"
2706 PMAXUB(%%mm2, %%mm4)
2707 PMINUB(%%mm2, %%mm3, %%mm5)
2708 "psubb %%mm2, %%mm1 \n\t"
2709 "paddb %%mm7, %%mm1 \n\t"
2710 "pcmpgtb %%mm6, %%mm1 \n\t"
2711 "paddb %%mm1, %%mm0 \n\t"
2712
2713 "movq (%%"REG_a", %3, 4), %%mm1 \n\t"
2714 "psubb %%mm1, %%mm2 \n\t"
2715 "paddb %%mm7, %%mm2 \n\t"
2716 "pcmpgtb %%mm6, %%mm2 \n\t"
2717 "paddb %%mm2, %%mm0 \n\t"
2718 "psubusb %%mm3, %%mm4 \n\t"
2719
2720 "pxor %%mm6, %%mm6 \n\t"
2721 "movq %4, %%mm7 \n\t" // QP,..., QP
2722 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
2723 "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0
2724 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2725 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2726 "movq %%mm7, %1 \n\t"
2727
2728 "movq %5, %%mm7 \n\t"
2729 "punpcklbw %%mm7, %%mm7 \n\t"
2730 "punpcklbw %%mm7, %%mm7 \n\t"
2731 "punpcklbw %%mm7, %%mm7 \n\t"
2732 "psubb %%mm0, %%mm6 \n\t"
2733 "pcmpgtb %%mm7, %%mm6 \n\t"
2734 "movq %%mm6, %0 \n\t"
2735
2736 : "=m" (eq_mask), "=m" (dc_mask)
2737 : "r" (src), "r" ((long)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
2738 : "%"REG_a
2739 );
2740
2741 both_masks = dc_mask & eq_mask;
2742
2743 if(both_masks){
2744 long offset= -8*step;
2745 int64_t *temp_sums= sums;
2746
2747 asm volatile(
2748 "movq %2, %%mm0 \n\t" // QP,..., QP
2749 "pxor %%mm4, %%mm4 \n\t"
2750
2751 "movq (%0), %%mm6 \n\t"
2752 "movq (%0, %1), %%mm5 \n\t"
2753 "movq %%mm5, %%mm1 \n\t"
2754 "movq %%mm6, %%mm2 \n\t"
2755 "psubusb %%mm6, %%mm5 \n\t"
2756 "psubusb %%mm1, %%mm2 \n\t"
2757 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2758 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2759 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2760
2761 "pxor %%mm6, %%mm1 \n\t"
2762 "pand %%mm0, %%mm1 \n\t"
2763 "pxor %%mm1, %%mm6 \n\t"
2764 // 0:QP 6:First
2765
2766 "movq (%0, %1, 8), %%mm5 \n\t"
2767 "add %1, %0 \n\t" // %0 points to line 1 not 0
2768 "movq (%0, %1, 8), %%mm7 \n\t"
2769 "movq %%mm5, %%mm1 \n\t"
2770 "movq %%mm7, %%mm2 \n\t"
2771 "psubusb %%mm7, %%mm5 \n\t"
2772 "psubusb %%mm1, %%mm2 \n\t"
2773 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2774 "movq %2, %%mm0 \n\t" // QP,..., QP
2775 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2776 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2777
2778 "pxor %%mm7, %%mm1 \n\t"
2779 "pand %%mm0, %%mm1 \n\t"
2780 "pxor %%mm1, %%mm7 \n\t"
2781
2782 "movq %%mm6, %%mm5 \n\t"
2783 "punpckhbw %%mm4, %%mm6 \n\t"
2784 "punpcklbw %%mm4, %%mm5 \n\t"
2785 // 4:0 5/6:First 7:Last
2786
2787 "movq %%mm5, %%mm0 \n\t"
2788 "movq %%mm6, %%mm1 \n\t"
2789 "psllw $2, %%mm0 \n\t"
2790 "psllw $2, %%mm1 \n\t"
2791 "paddw "MANGLE(w04)", %%mm0 \n\t"
2792 "paddw "MANGLE(w04)", %%mm1 \n\t"
2793
2794 #define NEXT\
2795 "movq (%0), %%mm2 \n\t"\
2796 "movq (%0), %%mm3 \n\t"\
2797 "add %1, %0 \n\t"\
2798 "punpcklbw %%mm4, %%mm2 \n\t"\
2799 "punpckhbw %%mm4, %%mm3 \n\t"\
2800 "paddw %%mm2, %%mm0 \n\t"\
2801 "paddw %%mm3, %%mm1 \n\t"
2802
2803 #define PREV\
2804 "movq (%0), %%mm2 \n\t"\
2805 "movq (%0), %%mm3 \n\t"\
2806 "add %1, %0 \n\t"\
2807 "punpcklbw %%mm4, %%mm2 \n\t"\
2808 "punpckhbw %%mm4, %%mm3 \n\t"\
2809 "psubw %%mm2, %%mm0 \n\t"\
2810 "psubw %%mm3, %%mm1 \n\t"
2811
2812
2813 NEXT //0
2814 NEXT //1
2815 NEXT //2
2816 "movq %%mm0, (%3) \n\t"
2817 "movq %%mm1, 8(%3) \n\t"
2818
2819 NEXT //3
2820 "psubw %%mm5, %%mm0 \n\t"
2821 "psubw %%mm6, %%mm1 \n\t"
2822 "movq %%mm0, 16(%3) \n\t"
2823 "movq %%mm1, 24(%3) \n\t"
2824
2825 NEXT //4
2826 "psubw %%mm5, %%mm0 \n\t"
2827 "psubw %%mm6, %%mm1 \n\t"
2828 "movq %%mm0, 32(%3) \n\t"
2829 "movq %%mm1, 40(%3) \n\t"
2830
2831 NEXT //5
2832 "psubw %%mm5, %%mm0 \n\t"
2833 "psubw %%mm6, %%mm1 \n\t"
2834 "movq %%mm0, 48(%3) \n\t"
2835 "movq %%mm1, 56(%3) \n\t"
2836
2837 NEXT //6
2838 "psubw %%mm5, %%mm0 \n\t"
2839 "psubw %%mm6, %%mm1 \n\t"
2840 "movq %%mm0, 64(%3) \n\t"
2841 "movq %%mm1, 72(%3) \n\t"
2842
2843 "movq %%mm7, %%mm6 \n\t"
2844 "punpckhbw %%mm4, %%mm7 \n\t"
2845 "punpcklbw %%mm4, %%mm6 \n\t"
2846
2847 NEXT //7
2848 "mov %4, %0 \n\t"
2849 "add %1, %0 \n\t"
2850 PREV //0
2851 "movq %%mm0, 80(%3) \n\t"
2852 "movq %%mm1, 88(%3) \n\t"
2853
2854 PREV //1
2855 "paddw %%mm6, %%mm0 \n\t"
2856 "paddw %%mm7, %%mm1 \n\t"
2857 "movq %%mm0, 96(%3) \n\t"
2858 "movq %%mm1, 104(%3) \n\t"
2859
2860 PREV //2
2861 "paddw %%mm6, %%mm0 \n\t"
2862 "paddw %%mm7, %%mm1 \n\t"
2863 "movq %%mm0, 112(%3) \n\t"
2864 "movq %%mm1, 120(%3) \n\t"
2865
2866 PREV //3
2867 "paddw %%mm6, %%mm0 \n\t"
2868 "paddw %%mm7, %%mm1 \n\t"
2869 "movq %%mm0, 128(%3) \n\t"
2870 "movq %%mm1, 136(%3) \n\t"
2871
2872 PREV //4
2873 "paddw %%mm6, %%mm0 \n\t"
2874 "paddw %%mm7, %%mm1 \n\t"
2875 "movq %%mm0, 144(%3) \n\t"
2876 "movq %%mm1, 152(%3) \n\t"
2877
2878 "mov %4, %0 \n\t" //FIXME
2879
2880 : "+&r"(src)
2881 : "r" ((long)step), "m" (c->pQPb), "r"(sums), "g"(src)
2882 );
2883
2884 src+= step; // src points to begin of the 8x8 Block
2885
2886 asm volatile(
2887 "movq %4, %%mm6 \n\t"
2888 "pcmpeqb %%mm5, %%mm5 \n\t"
2889 "pxor %%mm6, %%mm5 \n\t"
2890 "pxor %%mm7, %%mm7 \n\t"
2891
2892 "1: \n\t"
2893 "movq (%1), %%mm0 \n\t"
2894 "movq 8(%1), %%mm1 \n\t"
2895 "paddw 32(%1), %%mm0 \n\t"
2896 "paddw 40(%1), %%mm1 \n\t"
2897 "movq (%0, %3), %%mm2 \n\t"
2898 "movq %%mm2, %%mm3 \n\t"
2899 "movq %%mm2, %%mm4 \n\t"
2900 "punpcklbw %%mm7, %%mm2 \n\t"
2901 "punpckhbw %%mm7, %%mm3 \n\t"
2902 "paddw %%mm2, %%mm0 \n\t"
2903 "paddw %%mm3, %%mm1 \n\t"
2904 "paddw %%mm2, %%mm0 \n\t"
2905 "paddw %%mm3, %%mm1 \n\t"
2906 "psrlw $4, %%mm0 \n\t"
2907 "psrlw $4, %%mm1 \n\t"
2908 "packuswb %%mm1, %%mm0 \n\t"
2909 "pand %%mm6, %%mm0 \n\t"
2910 "pand %%mm5, %%mm4 \n\t"
2911 "por %%mm4, %%mm0 \n\t"
2912 "movq %%mm0, (%0, %3) \n\t"
2913 "add $16, %1 \n\t"
2914 "add %2, %0 \n\t"
2915 " js 1b \n\t"
2916
2917 : "+r"(offset), "+r"(temp_sums)
2918 : "r" ((long)step), "r"(src - offset), "m"(both_masks)
2919 );
2920 }else
2921 src+= step; // src points to begin of the 8x8 Block
2922
2923 if(eq_mask != -1LL){
2924 uint8_t *temp_src= src;
2925 asm volatile(
2926 "pxor %%mm7, %%mm7 \n\t"
2927 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
2928 "and "ALIGN_MASK", %%"REG_c" \n\t" // align
2929 // 0 1 2 3 4 5 6 7 8 9
2930 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1
2931
2932 "movq (%0), %%mm0 \n\t"
2933 "movq %%mm0, %%mm1 \n\t"
2934 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
2935 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
2936
2937 "movq (%0, %1), %%mm2 \n\t"
2938 "lea (%0, %1, 2), %%"REG_a" \n\t"
2939 "movq %%mm2, %%mm3 \n\t"
2940 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
2941 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
2942
2943 "movq (%%"REG_a"), %%mm4 \n\t"
2944 "movq %%mm4, %%mm5 \n\t"
2945 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
2946 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
2947
2948 "paddw %%mm0, %%mm0 \n\t" // 2L0
2949 "paddw %%mm1, %%mm1 \n\t" // 2H0
2950 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
2951 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
2952 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
2953 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
2954
2955 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
2956 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
2957 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
2958 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
2959
2960 "movq (%%"REG_a", %1), %%mm2 \n\t"
2961 "movq %%mm2, %%mm3 \n\t"
2962 "punpcklbw %%mm7, %%mm2 \n\t" // L3
2963 "punpckhbw %%mm7, %%mm3 \n\t" // H3
2964
2965 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
2966 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
2967 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2968 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2969 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2970 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2971
2972 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
2973 "movq %%mm0, %%mm1 \n\t"
2974 "punpcklbw %%mm7, %%mm0 \n\t" // L4
2975 "punpckhbw %%mm7, %%mm1 \n\t" // H4
2976
2977 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
2978 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
2979 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
2980 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
2981 "paddw %%mm4, %%mm4 \n\t" // 2L2
2982 "paddw %%mm5, %%mm5 \n\t" // 2H2
2983 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
2984 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
2985
2986 "lea (%%"REG_a", %1), %0 \n\t"
2987 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
2988 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
2989 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
2990 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
2991 //50 opcodes so far
2992 "movq (%0, %1, 2), %%mm2 \n\t"
2993 "movq %%mm2, %%mm3 \n\t"
2994 "punpcklbw %%mm7, %%mm2 \n\t" // L5
2995 "punpckhbw %%mm7, %%mm3 \n\t" // H5
2996 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
2997 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
2998 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
2999 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
3000
3001 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
3002 "punpcklbw %%mm7, %%mm6 \n\t" // L6
3003 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
3004 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
3005 "punpckhbw %%mm7, %%mm6 \n\t" // H6
3006 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
3007
3008 "paddw %%mm0, %%mm0 \n\t" // 2L4
3009 "paddw %%mm1, %%mm1 \n\t" // 2H4
3010 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
3011 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
3012
3013 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
3014 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
3015 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
3016 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
3017
3018 "movq (%0, %1, 4), %%mm2 \n\t"
3019 "movq %%mm2, %%mm3 \n\t"
3020 "punpcklbw %%mm7, %%mm2 \n\t" // L7
3021 "punpckhbw %%mm7, %%mm3 \n\t" // H7
3022
3023 "paddw %%mm2, %%mm2 \n\t" // 2L7
3024 "paddw %%mm3, %%mm3 \n\t" // 2H7
3025 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
3026 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
3027
3028 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
3029 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
3030
3031 #ifdef HAVE_MMX2
3032 "movq %%mm7, %%mm6 \n\t" // 0
3033 "psubw %%mm0, %%mm6 \n\t"
3034 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
3035 "movq %%mm7, %%mm6 \n\t" // 0
3036 "psubw %%mm1, %%mm6 \n\t"
3037 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
3038 "movq %%mm7, %%mm6 \n\t" // 0
3039 "psubw %%mm2, %%mm6 \n\t"
3040 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
3041 "movq %%mm7, %%mm6 \n\t" // 0
3042 "psubw %%mm3, %%mm6 \n\t"
3043 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3044 #else
3045 "movq %%mm7, %%mm6 \n\t" // 0
3046 "pcmpgtw %%mm0, %%mm6 \n\t"
3047 "pxor %%mm6, %%mm0 \n\t"
3048 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
3049 "movq %%mm7, %%mm6 \n\t" // 0
3050 "pcmpgtw %%mm1, %%mm6 \n\t"
3051 "pxor %%mm6, %%mm1 \n\t"
3052 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
3053 "movq %%mm7, %%mm6 \n\t" // 0
3054 "pcmpgtw %%mm2, %%mm6 \n\t"
3055 "pxor %%mm6, %%mm2 \n\t"
3056 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
3057 "movq %%mm7, %%mm6 \n\t" // 0
3058 "pcmpgtw %%mm3, %%mm6 \n\t"
3059 "pxor %%mm6, %%mm3 \n\t"
3060 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3061 #endif
3062
3063 #ifdef HAVE_MMX2
3064 "pminsw %%mm2, %%mm0 \n\t"
3065 "pminsw %%mm3, %%mm1 \n\t"
3066 #else
3067 "movq %%mm0, %%mm6 \n\t"
3068 "psubusw %%mm2, %%mm6 \n\t"
3069 "psubw %%mm6, %%mm0 \n\t"
3070 "movq %%mm1, %%mm6 \n\t"
3071 "psubusw %%mm3, %%mm6 \n\t"
3072 "psubw %%mm6, %%mm1 \n\t"
3073 #endif
3074
3075 "movd %2, %%mm2 \n\t" // QP
3076 "punpcklbw %%mm7, %%mm2 \n\t"
3077
3078 "movq %%mm7, %%mm6 \n\t" // 0
3079 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
3080 "pxor %%mm6, %%mm4 \n\t"
3081 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
3082 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
3083 "pxor %%mm7, %%mm5 \n\t"
3084 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
3085 // 100 opcodes
3086 "psllw $3, %%mm2 \n\t" // 8QP
3087 "movq %%mm2, %%mm3 \n\t" // 8QP
3088 "pcmpgtw %%mm4, %%mm2 \n\t"
3089 "pcmpgtw %%mm5, %%mm3 \n\t"