2 Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at)
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program; if not, write to the Free Software
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
24 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
25 #elif defined (HAVE_3DNOW)
26 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
30 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
31 #elif defined (HAVE_MMX)
32 #define PMINUB(b,a,t) \
33 "movq " #a ", " #t " \n\t"\
34 "psubusb " #b ", " #t " \n\t"\
35 "psubb " #t ", " #a " \n\t"
39 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
40 #elif defined (HAVE_MMX)
42 "psubusb " #a ", " #b " \n\t"\
43 "paddb " #a ", " #b " \n\t"
47 //FIXME? |255-0| = 1 (shouldnt be a problem ...)
50 * Check if the middle 8x8 Block in the given 8x16 block is flat
52 static inline int RENAME(isVertDC)(uint8_t src[], int stride, PPContext *c){
54 src+= stride*4; // src points to begin of the 8x8 Block
56 "leal (%1, %2), %%eax \n\t"
57 // 0 1 2 3 4 5 6 7 8 9
58 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
59 "movq %3, %%mm7 \n\t" // mm7 = 0x7F
60 "movq %4, %%mm6 \n\t" // mm6 = 0x7D
61 "movq (%1), %%mm0 \n\t"
62 "movq (%%eax), %%mm1 \n\t"
63 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
64 "paddb %%mm7, %%mm0 \n\t"
65 "pcmpgtb %%mm6, %%mm0 \n\t"
67 "movq (%%eax,%2), %%mm2 \n\t"
68 "psubb %%mm2, %%mm1 \n\t"
69 "paddb %%mm7, %%mm1 \n\t"
70 "pcmpgtb %%mm6, %%mm1 \n\t"
71 "paddb %%mm1, %%mm0 \n\t"
73 "movq (%%eax, %2, 2), %%mm1 \n\t"
74 "psubb %%mm1, %%mm2 \n\t"
75 "paddb %%mm7, %%mm2 \n\t"
76 "pcmpgtb %%mm6, %%mm2 \n\t"
77 "paddb %%mm2, %%mm0 \n\t"
79 "leal (%%eax, %2, 4), %%eax \n\t"
81 "movq (%1, %2, 4), %%mm2 \n\t"
82 "psubb %%mm2, %%mm1 \n\t"
83 "paddb %%mm7, %%mm1 \n\t"
84 "pcmpgtb %%mm6, %%mm1 \n\t"
85 "paddb %%mm1, %%mm0 \n\t"
87 "movq (%%eax), %%mm1 \n\t"
88 "psubb %%mm1, %%mm2 \n\t"
89 "paddb %%mm7, %%mm2 \n\t"
90 "pcmpgtb %%mm6, %%mm2 \n\t"
91 "paddb %%mm2, %%mm0 \n\t"
93 "movq (%%eax, %2), %%mm2 \n\t"
94 "psubb %%mm2, %%mm1 \n\t"
95 "paddb %%mm7, %%mm1 \n\t"
96 "pcmpgtb %%mm6, %%mm1 \n\t"
97 "paddb %%mm1, %%mm0 \n\t"
99 "movq (%%eax, %2, 2), %%mm1 \n\t"
100 "psubb %%mm1, %%mm2 \n\t"
101 "paddb %%mm7, %%mm2 \n\t"
102 "pcmpgtb %%mm6, %%mm2 \n\t"
103 "paddb %%mm2, %%mm0 \n\t"
107 "pxor %%mm7, %%mm7 \n\t"
108 "psadbw %%mm7, %%mm0 \n\t"
110 "movq %%mm0, %%mm1 \n\t"
111 "psrlw $8, %%mm0 \n\t"
112 "paddb %%mm1, %%mm0 \n\t"
113 "movq %%mm0, %%mm1 \n\t"
114 "psrlq $16, %%mm0 \n\t"
115 "paddb %%mm1, %%mm0 \n\t"
116 "movq %%mm0, %%mm1 \n\t"
117 "psrlq $32, %%mm0 \n\t"
118 "paddb %%mm1, %%mm0 \n\t"
120 "movd %%mm0, %0 \n\t"
122 : "r" (src), "r" (stride), "m" (c->mmxDcOffset), "m" (c->mmxDcThreshold)
125 numEq= (-numEq) &0xFF;
126 return numEq > c->ppMode.flatnessThreshold;
130 static inline int RENAME(isVertMinMaxOk)(uint8_t src[], int stride, PPContext *c)
136 "movq (%1, %2), %%mm0 \n\t"
137 "movq (%1, %2, 8), %%mm1 \n\t"
138 "movq %%mm0, %%mm2 \n\t"
139 "psubusb %%mm1, %%mm0 \n\t"
140 "psubusb %%mm2, %%mm1 \n\t"
141 "por %%mm1, %%mm0 \n\t" // ABS Diff
143 "movq %3, %%mm7 \n\t" // QP,..., QP
144 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
145 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0
146 "packssdw %%mm0, %%mm0 \n\t"
147 "movd %%mm0, %0 \n\t"
149 : "r" (src), "r" (stride), "m" (c->pQPb)
156 for(x=0; x<BLOCK_SIZE; x++)
158 if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
166 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
167 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
169 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
171 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
173 asm volatile( //"movv %0 %1 %2\n\t"
174 "movq %2, %%mm0 \n\t" // QP,..., QP
175 "pxor %%mm4, %%mm4 \n\t"
177 "movq (%0), %%mm6 \n\t"
178 "movq (%0, %1), %%mm5 \n\t"
179 "movq %%mm5, %%mm1 \n\t"
180 "movq %%mm6, %%mm2 \n\t"
181 "psubusb %%mm6, %%mm5 \n\t"
182 "psubusb %%mm1, %%mm2 \n\t"
183 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
184 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
185 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
187 "pand %%mm2, %%mm6 \n\t"
188 "pandn %%mm1, %%mm2 \n\t"
189 "por %%mm2, %%mm6 \n\t"// First Line to Filter
191 "movq (%0, %1, 8), %%mm5 \n\t"
192 "leal (%0, %1, 4), %%eax \n\t"
193 "leal (%0, %1, 8), %%ecx \n\t"
194 "subl %1, %%ecx \n\t"
195 "addl %1, %0 \n\t" // %0 points to line 1 not 0
196 "movq (%0, %1, 8), %%mm7 \n\t"
197 "movq %%mm5, %%mm1 \n\t"
198 "movq %%mm7, %%mm2 \n\t"
199 "psubusb %%mm7, %%mm5 \n\t"
200 "psubusb %%mm1, %%mm2 \n\t"
201 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
202 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
203 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
205 "pand %%mm2, %%mm7 \n\t"
206 "pandn %%mm1, %%mm2 \n\t"
207 "por %%mm2, %%mm7 \n\t" // First Line to Filter
211 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
216 "movq (%0, %1), %%mm0 \n\t" // 1
217 "movq %%mm0, %%mm1 \n\t" // 1
218 PAVGB(%%mm6, %%mm0) //1 1 /2
219 PAVGB(%%mm6, %%mm0) //3 1 /4
221 "movq (%0, %1, 4), %%mm2 \n\t" // 1
222 "movq %%mm2, %%mm5 \n\t" // 1
223 PAVGB((%%eax), %%mm2) // 11 /2
224 PAVGB((%0, %1, 2), %%mm2) // 211 /4
225 "movq %%mm2, %%mm3 \n\t" // 211 /4
226 "movq (%0), %%mm4 \n\t" // 1
227 PAVGB(%%mm4, %%mm3) // 4 211 /8
228 PAVGB(%%mm0, %%mm3) //642211 /16
229 "movq %%mm3, (%0) \n\t" // X
230 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
231 "movq %%mm1, %%mm0 \n\t" // 1
232 PAVGB(%%mm6, %%mm0) //1 1 /2
233 "movq %%mm4, %%mm3 \n\t" // 1
234 PAVGB((%0,%1,2), %%mm3) // 1 1 /2
235 PAVGB((%%eax,%1,2), %%mm5) // 11 /2
236 PAVGB((%%eax), %%mm5) // 211 /4
237 PAVGB(%%mm5, %%mm3) // 2 2211 /8
238 PAVGB(%%mm0, %%mm3) //4242211 /16
239 "movq %%mm3, (%0,%1) \n\t" // X
240 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
241 PAVGB(%%mm4, %%mm6) //11 /2
242 "movq (%%ecx), %%mm0 \n\t" // 1
243 PAVGB((%%eax, %1, 2), %%mm0) // 11/2
244 "movq %%mm0, %%mm3 \n\t" // 11/2
245 PAVGB(%%mm1, %%mm0) // 2 11/4
246 PAVGB(%%mm6, %%mm0) //222 11/8
247 PAVGB(%%mm2, %%mm0) //22242211/16
248 "movq (%0, %1, 2), %%mm2 \n\t" // 1
249 "movq %%mm0, (%0, %1, 2) \n\t" // X
250 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
251 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1
252 PAVGB((%%ecx), %%mm0) // 11 /2
253 PAVGB(%%mm0, %%mm6) //11 11 /4
254 PAVGB(%%mm1, %%mm4) // 11 /2
255 PAVGB(%%mm2, %%mm1) // 11 /2
256 PAVGB(%%mm1, %%mm6) //1122 11 /8
257 PAVGB(%%mm5, %%mm6) //112242211 /16
258 "movq (%%eax), %%mm5 \n\t" // 1
259 "movq %%mm6, (%%eax) \n\t" // X
260 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
261 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1
262 PAVGB(%%mm7, %%mm6) // 11 /2
263 PAVGB(%%mm4, %%mm6) // 11 11 /4
264 PAVGB(%%mm3, %%mm6) // 11 2211 /8
265 PAVGB(%%mm5, %%mm2) // 11 /2
266 "movq (%0, %1, 4), %%mm4 \n\t" // 1
267 PAVGB(%%mm4, %%mm2) // 112 /4
268 PAVGB(%%mm2, %%mm6) // 112242211 /16
269 "movq %%mm6, (%0, %1, 4) \n\t" // X
270 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
271 PAVGB(%%mm7, %%mm1) // 11 2 /4
272 PAVGB(%%mm4, %%mm5) // 11 /2
273 PAVGB(%%mm5, %%mm0) // 11 11 /4
274 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1
275 PAVGB(%%mm6, %%mm1) // 11 4 2 /8
276 PAVGB(%%mm0, %%mm1) // 11224222 /16
277 "movq %%mm1, (%%eax, %1, 2) \n\t" // X
278 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
279 PAVGB((%%ecx), %%mm2) // 112 4 /8
280 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1
281 PAVGB(%%mm0, %%mm6) // 1 1 /2
282 PAVGB(%%mm7, %%mm6) // 1 12 /4
283 PAVGB(%%mm2, %%mm6) // 1122424 /4
284 "movq %%mm6, (%%ecx) \n\t" // X
285 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
286 PAVGB(%%mm7, %%mm5) // 11 2 /4
287 PAVGB(%%mm7, %%mm5) // 11 6 /8
289 PAVGB(%%mm3, %%mm0) // 112 /4
290 PAVGB(%%mm0, %%mm5) // 112246 /16
291 "movq %%mm5, (%%eax, %1, 4) \n\t" // X
295 : "r" (src), "r" (stride), "m" (c->pQPb)
299 const int l1= stride;
300 const int l2= stride + l1;
301 const int l3= stride + l2;
302 const int l4= stride + l3;
303 const int l5= stride + l4;
304 const int l6= stride + l5;
305 const int l7= stride + l6;
306 const int l8= stride + l7;
307 const int l9= stride + l8;
310 for(x=0; x<BLOCK_SIZE; x++)
312 const int first= ABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
313 const int last= ABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
316 sums[0] = first + src[l1];
317 sums[1] = src[l1] + src[l2];
318 sums[2] = src[l2] + src[l3];
319 sums[3] = src[l3] + src[l4];
320 sums[4] = src[l4] + src[l5];
321 sums[5] = src[l5] + src[l6];
322 sums[6] = src[l6] + src[l7];
323 sums[7] = src[l7] + src[l8];
324 sums[8] = src[l8] + last;
326 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
327 src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
328 src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
329 src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
330 src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
331 src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
332 src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4;
333 src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
342 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
343 * values are correctly clipped (MMX2)
344 * values are wraparound (C)
345 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient
352 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
354 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
358 "pxor %%mm7, %%mm7 \n\t" // 0
359 "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
360 "leal (%0, %1), %%eax \n\t"
361 "leal (%%eax, %1, 4), %%ecx \n\t"
362 // 0 1 2 3 4 5 6 7 8 9
363 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
364 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
365 "movq %%mm0, %%mm1 \n\t" // QP,..., QP
366 "paddusb "MANGLE(b02)", %%mm0 \n\t"
367 "psrlw $2, %%mm0 \n\t"
368 "pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4
369 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
370 "movq (%0, %1, 4), %%mm2 \n\t" // line 4
371 "movq (%%ecx), %%mm3 \n\t" // line 5
372 "movq %%mm2, %%mm4 \n\t" // line 4
373 "pcmpeqb %%mm5, %%mm5 \n\t" // -1
374 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1
376 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2
377 "psubusb %%mm3, %%mm4 \n\t"
378 "psubusb %%mm2, %%mm3 \n\t"
379 "por %%mm3, %%mm4 \n\t" // |l4 - l5|
380 "psubusb %%mm0, %%mm4 \n\t"
381 "pcmpeqb %%mm7, %%mm4 \n\t"
382 "pand %%mm4, %%mm5 \n\t" // d/2
384 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80
385 "paddb %%mm5, %%mm2 \n\t"
386 // "psubb %%mm6, %%mm2 \n\t"
387 "movq %%mm2, (%0,%1, 4) \n\t"
389 "movq (%%ecx), %%mm2 \n\t"
390 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80
391 "psubb %%mm5, %%mm2 \n\t"
392 // "psubb %%mm6, %%mm2 \n\t"
393 "movq %%mm2, (%%ecx) \n\t"
395 "paddb %%mm6, %%mm5 \n\t"
396 "psrlw $2, %%mm5 \n\t"
397 "pand "MANGLE(b3F)", %%mm5 \n\t"
398 "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
400 "movq (%%eax, %1, 2), %%mm2 \n\t"
401 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
402 "paddsb %%mm5, %%mm2 \n\t"
403 "psubb %%mm6, %%mm2 \n\t"
404 "movq %%mm2, (%%eax, %1, 2) \n\t"
406 "movq (%%ecx, %1), %%mm2 \n\t"
407 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80
408 "psubsb %%mm5, %%mm2 \n\t"
409 "psubb %%mm6, %%mm2 \n\t"
410 "movq %%mm2, (%%ecx, %1) \n\t"
413 : "r" (src), "r" (stride)
417 const int l1= stride;
418 const int l2= stride + l1;
419 const int l3= stride + l2;
420 const int l4= stride + l3;
421 const int l5= stride + l4;
422 const int l6= stride + l5;
423 // const int l7= stride + l6;
424 // const int l8= stride + l7;
425 // const int l9= stride + l8;
427 const int QP15= QP + (QP>>2);
429 for(x=0; x<BLOCK_SIZE; x++)
431 const int v = (src[x+l5] - src[x+l4]);
447 * Experimental Filter 1
448 * will not damage linear gradients
449 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
450 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
451 * MMX2 version does correct clipping C version doesnt
453 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
455 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
459 "pxor %%mm7, %%mm7 \n\t" // 0
460 "leal (%0, %1), %%eax \n\t"
461 "leal (%%eax, %1, 4), %%ecx \n\t"
462 // 0 1 2 3 4 5 6 7 8 9
463 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
464 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3
465 "movq (%0, %1, 4), %%mm1 \n\t" // line 4
466 "movq %%mm1, %%mm2 \n\t" // line 4
467 "psubusb %%mm0, %%mm1 \n\t"
468 "psubusb %%mm2, %%mm0 \n\t"
469 "por %%mm1, %%mm0 \n\t" // |l2 - l3|
470 "movq (%%ecx), %%mm3 \n\t" // line 5
471 "movq (%%ecx, %1), %%mm4 \n\t" // line 6
472 "movq %%mm3, %%mm5 \n\t" // line 5
473 "psubusb %%mm4, %%mm3 \n\t"
474 "psubusb %%mm5, %%mm4 \n\t"
475 "por %%mm4, %%mm3 \n\t" // |l5 - l6|
476 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
477 "movq %%mm2, %%mm1 \n\t" // line 4
478 "psubusb %%mm5, %%mm2 \n\t"
479 "movq %%mm2, %%mm4 \n\t"
480 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
481 "psubusb %%mm1, %%mm5 \n\t"
482 "por %%mm5, %%mm4 \n\t" // |l4 - l5|
483 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
484 "movq %%mm4, %%mm3 \n\t" // d
485 "movq %2, %%mm0 \n\t"
486 "paddusb %%mm0, %%mm0 \n\t"
487 "psubusb %%mm0, %%mm4 \n\t"
488 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
489 "psubusb "MANGLE(b01)", %%mm3 \n\t"
490 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
492 PAVGB(%%mm7, %%mm3) // d/2
493 "movq %%mm3, %%mm1 \n\t" // d/2
494 PAVGB(%%mm7, %%mm3) // d/4
495 PAVGB(%%mm1, %%mm3) // 3*d/8
497 "movq (%0, %1, 4), %%mm0 \n\t" // line 4
498 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
499 "psubusb %%mm3, %%mm0 \n\t"
500 "pxor %%mm2, %%mm0 \n\t"
501 "movq %%mm0, (%0, %1, 4) \n\t" // line 4
503 "movq (%%ecx), %%mm0 \n\t" // line 5
504 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
505 "paddusb %%mm3, %%mm0 \n\t"
506 "pxor %%mm2, %%mm0 \n\t"
507 "movq %%mm0, (%%ecx) \n\t" // line 5
509 PAVGB(%%mm7, %%mm1) // d/4
511 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3
512 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
513 "psubusb %%mm1, %%mm0 \n\t"
514 "pxor %%mm2, %%mm0 \n\t"
515 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3
517 "movq (%%ecx, %1), %%mm0 \n\t" // line 6
518 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
519 "paddusb %%mm1, %%mm0 \n\t"
520 "pxor %%mm2, %%mm0 \n\t"
521 "movq %%mm0, (%%ecx, %1) \n\t" // line 6
523 PAVGB(%%mm7, %%mm1) // d/8
525 "movq (%%eax, %1), %%mm0 \n\t" // line 2
526 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
527 "psubusb %%mm1, %%mm0 \n\t"
528 "pxor %%mm2, %%mm0 \n\t"
529 "movq %%mm0, (%%eax, %1) \n\t" // line 2
531 "movq (%%ecx, %1, 2), %%mm0 \n\t" // line 7
532 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
533 "paddusb %%mm1, %%mm0 \n\t"
534 "pxor %%mm2, %%mm0 \n\t"
535 "movq %%mm0, (%%ecx, %1, 2) \n\t" // line 7
538 : "r" (src), "r" (stride), "m" (co->pQPb)
543 const int l1= stride;
544 const int l2= stride + l1;
545 const int l3= stride + l2;
546 const int l4= stride + l3;
547 const int l5= stride + l4;
548 const int l6= stride + l5;
549 const int l7= stride + l6;
550 // const int l8= stride + l7;
551 // const int l9= stride + l8;
555 for(x=0; x<BLOCK_SIZE; x++)
557 int a= src[l3] - src[l4];
558 int b= src[l4] - src[l5];
559 int c= src[l5] - src[l6];
561 int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
566 int v = d * SIGN(-b);
581 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
583 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
586 const int l1= stride;
587 const int l2= stride + l1;
588 const int l3= stride + l2;
589 const int l4= (int)tmp - (int)src - stride*3;
590 const int l5= (int)tmp - (int)src - stride*3 + 8;
591 const int l6= stride*3 + l3;
592 const int l7= stride + l6;
593 const int l8= stride + l7;
595 memcpy(tmp, src+stride*7, 8);
596 memcpy(tmp+8, src+stride*8, 8);
601 #if 0 //sligtly more accurate and slightly slower
602 "pxor %%mm7, %%mm7 \n\t" // 0
603 "leal (%0, %1), %%eax \n\t"
604 "leal (%%eax, %1, 4), %%ecx \n\t"
606 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
607 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
610 "movq (%0, %1, 2), %%mm0 \n\t" // l2
611 "movq (%0), %%mm1 \n\t" // l0
612 "movq %%mm0, %%mm2 \n\t" // l2
613 PAVGB(%%mm7, %%mm0) // ~l2/2
614 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
615 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
617 "movq (%%eax), %%mm1 \n\t" // l1
618 "movq (%%eax, %1, 2), %%mm3 \n\t" // l3
619 "movq %%mm1, %%mm4 \n\t" // l1
620 PAVGB(%%mm7, %%mm1) // ~l1/2
621 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
622 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
624 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
625 "psubusb %%mm1, %%mm0 \n\t"
626 "psubusb %%mm4, %%mm1 \n\t"
627 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
628 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
630 "movq (%0, %1, 4), %%mm0 \n\t" // l4
631 "movq %%mm0, %%mm4 \n\t" // l4
632 PAVGB(%%mm7, %%mm0) // ~l4/2
633 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
634 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
636 "movq (%%ecx), %%mm2 \n\t" // l5
637 "movq %%mm3, %%mm5 \n\t" // l3
638 PAVGB(%%mm7, %%mm3) // ~l3/2
639 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
640 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
642 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
643 "psubusb %%mm3, %%mm0 \n\t"
644 "psubusb %%mm6, %%mm3 \n\t"
645 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
646 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
647 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
649 "movq (%%ecx, %1), %%mm6 \n\t" // l6
650 "movq %%mm6, %%mm5 \n\t" // l6
651 PAVGB(%%mm7, %%mm6) // ~l6/2
652 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
653 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
655 "movq (%%ecx, %1, 2), %%mm5 \n\t" // l7
656 "movq %%mm2, %%mm4 \n\t" // l5
657 PAVGB(%%mm7, %%mm2) // ~l5/2
658 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
659 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
661 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
662 "psubusb %%mm2, %%mm6 \n\t"
663 "psubusb %%mm4, %%mm2 \n\t"
664 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
665 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
668 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
669 "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
670 "paddusb "MANGLE(b01)", %%mm4 \n\t"
671 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
672 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
673 "pand %%mm4, %%mm3 \n\t"
675 "movq %%mm3, %%mm1 \n\t"
676 // "psubusb "MANGLE(b01)", %%mm3 \n\t"
679 "paddusb %%mm1, %%mm3 \n\t"
680 // "paddusb "MANGLE(b01)", %%mm3 \n\t"
682 "movq (%%eax, %1, 2), %%mm6 \n\t" //l3
683 "movq (%0, %1, 4), %%mm5 \n\t" //l4
684 "movq (%0, %1, 4), %%mm4 \n\t" //l4
685 "psubusb %%mm6, %%mm5 \n\t"
686 "psubusb %%mm4, %%mm6 \n\t"
687 "por %%mm6, %%mm5 \n\t" // |l3-l4|
688 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
689 "pxor %%mm6, %%mm0 \n\t"
690 "pand %%mm0, %%mm3 \n\t"
691 PMINUB(%%mm5, %%mm3, %%mm0)
693 "psubusb "MANGLE(b01)", %%mm3 \n\t"
696 "movq (%%eax, %1, 2), %%mm0 \n\t"
697 "movq (%0, %1, 4), %%mm2 \n\t"
698 "pxor %%mm6, %%mm0 \n\t"
699 "pxor %%mm6, %%mm2 \n\t"
700 "psubb %%mm3, %%mm0 \n\t"
701 "paddb %%mm3, %%mm2 \n\t"
702 "pxor %%mm6, %%mm0 \n\t"
703 "pxor %%mm6, %%mm2 \n\t"
704 "movq %%mm0, (%%eax, %1, 2) \n\t"
705 "movq %%mm2, (%0, %1, 4) \n\t"
708 "leal (%0, %1), %%eax \n\t"
709 "pcmpeqb %%mm6, %%mm6 \n\t" // -1
711 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
712 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
715 "movq (%%eax, %1, 2), %%mm1 \n\t" // l3
716 "movq (%0, %1, 4), %%mm0 \n\t" // l4
717 "pxor %%mm6, %%mm1 \n\t" // -l3-1
718 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
719 // mm1=-l3-1, mm0=128-q
721 "movq (%%eax, %1, 4), %%mm2 \n\t" // l5
722 "movq (%%eax, %1), %%mm3 \n\t" // l2
723 "pxor %%mm6, %%mm2 \n\t" // -l5-1
724 "movq %%mm2, %%mm5 \n\t" // -l5-1
725 "movq "MANGLE(b80)", %%mm4 \n\t" // 128
726 "leal (%%eax, %1, 4), %%ecx \n\t"
727 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
728 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
729 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
730 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
731 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
733 "movq (%%eax), %%mm2 \n\t" // l1
734 "pxor %%mm6, %%mm2 \n\t" // -l1-1
735 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
736 PAVGB((%0), %%mm1) // (l0-l3+256)/2
737 "movq "MANGLE(b80)", %%mm3 \n\t" // 128
738 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
739 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
740 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
741 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
743 PAVGB((%%ecx, %1), %%mm5) // (l6-l5+256)/2
744 "movq (%%ecx, %1, 2), %%mm1 \n\t" // l7
745 "pxor %%mm6, %%mm1 \n\t" // -l7-1
746 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
747 "movq "MANGLE(b80)", %%mm2 \n\t" // 128
748 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
749 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
750 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
751 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
753 "movq "MANGLE(b00)", %%mm1 \n\t" // 0
754 "movq "MANGLE(b00)", %%mm5 \n\t" // 0
755 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
756 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
757 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
758 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
759 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
761 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
763 "movq "MANGLE(b00)", %%mm7 \n\t" // 0
764 "movq %2, %%mm2 \n\t" // QP
765 PAVGB(%%mm6, %%mm2) // 128 + QP/2
766 "psubb %%mm6, %%mm2 \n\t"
768 "movq %%mm4, %%mm1 \n\t"
769 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
770 "pxor %%mm1, %%mm4 \n\t"
771 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
772 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
773 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
774 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
776 "movq %%mm4, %%mm3 \n\t" // d
777 "psubusb "MANGLE(b01)", %%mm4 \n\t"
778 PAVGB(%%mm7, %%mm4) // d/32
779 PAVGB(%%mm7, %%mm4) // (d + 32)/64
780 "paddb %%mm3, %%mm4 \n\t" // 5d/64
781 "pand %%mm2, %%mm4 \n\t"
783 "movq "MANGLE(b80)", %%mm5 \n\t" // 128
784 "psubb %%mm0, %%mm5 \n\t" // q
785 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
786 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
787 "pxor %%mm7, %%mm5 \n\t"
789 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
790 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
792 "pand %%mm7, %%mm4 \n\t"
793 "movq (%%eax, %1, 2), %%mm0 \n\t"
794 "movq (%0, %1, 4), %%mm2 \n\t"
795 "pxor %%mm1, %%mm0 \n\t"
796 "pxor %%mm1, %%mm2 \n\t"
797 "paddb %%mm4, %%mm0 \n\t"
798 "psubb %%mm4, %%mm2 \n\t"
799 "pxor %%mm1, %%mm0 \n\t"
800 "pxor %%mm1, %%mm2 \n\t"
801 "movq %%mm0, (%%eax, %1, 2) \n\t"
802 "movq %%mm2, (%0, %1, 4) \n\t"
805 : "r" (src), "r" (stride), "m" (c->pQPb)
813 for(x=0; x<BLOCK_SIZE; x++)
815 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
816 if(ABS(middleEnergy)< 8*QP)
818 const int q=(src[l4] - src[l5])/2;
819 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
820 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
822 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
826 d*= SIGN(-middleEnergy);
850 int d= src[x+y*stride] - tmp[x+(y-4)*8];
861 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
867 printf(" %d %d %d %d\n", num, sum, max, bias);
873 #elif defined (HAVE_MMX)
877 "pxor %%mm7, %%mm7 \n\t"
878 "leal (%0, %1), %%eax \n\t"
879 "leal (%%eax, %1, 4), %%edx \n\t"
880 "leal -40(%%esp), %%ecx \n\t" // make space for 4 8-byte vars
881 "andl $0xFFFFFFF8, %%ecx \n\t" // align
883 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
884 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
886 "movq (%0), %%mm0 \n\t"
887 "movq %%mm0, %%mm1 \n\t"
888 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
889 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
891 "movq (%%eax), %%mm2 \n\t"
892 "movq %%mm2, %%mm3 \n\t"
893 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
894 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
896 "movq (%%eax, %1), %%mm4 \n\t"
897 "movq %%mm4, %%mm5 \n\t"
898 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
899 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
901 "paddw %%mm0, %%mm0 \n\t" // 2L0
902 "paddw %%mm1, %%mm1 \n\t" // 2H0
903 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
904 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
905 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
906 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
908 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
909 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
910 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
911 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
913 "movq (%%eax, %1, 2), %%mm2 \n\t"
914 "movq %%mm2, %%mm3 \n\t"
915 "punpcklbw %%mm7, %%mm2 \n\t" // L3
916 "punpckhbw %%mm7, %%mm3 \n\t" // H3
918 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
919 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
920 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
921 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
922 "movq %%mm0, (%%ecx) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
923 "movq %%mm1, 8(%%ecx) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
925 "movq (%0, %1, 4), %%mm0 \n\t"
926 "movq %%mm0, %%mm1 \n\t"
927 "punpcklbw %%mm7, %%mm0 \n\t" // L4
928 "punpckhbw %%mm7, %%mm1 \n\t" // H4
930 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
931 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
932 "movq %%mm2, 16(%%ecx) \n\t" // L3 - L4
933 "movq %%mm3, 24(%%ecx) \n\t" // H3 - H4
934 "paddw %%mm4, %%mm4 \n\t" // 2L2
935 "paddw %%mm5, %%mm5 \n\t" // 2H2
936 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
937 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
939 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
940 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
941 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
942 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
944 "movq (%%edx), %%mm2 \n\t"
945 "movq %%mm2, %%mm3 \n\t"
946 "punpcklbw %%mm7, %%mm2 \n\t" // L5
947 "punpckhbw %%mm7, %%mm3 \n\t" // H5
948 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
949 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
950 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
951 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
953 "movq (%%edx, %1), %%mm6 \n\t"
954 "punpcklbw %%mm7, %%mm6 \n\t" // L6
955 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
956 "movq (%%edx, %1), %%mm6 \n\t"
957 "punpckhbw %%mm7, %%mm6 \n\t" // H6
958 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
960 "paddw %%mm0, %%mm0 \n\t" // 2L4
961 "paddw %%mm1, %%mm1 \n\t" // 2H4
962 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
963 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
965 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
966 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
967 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
968 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
970 "movq (%%edx, %1, 2), %%mm2 \n\t"
971 "movq %%mm2, %%mm3 \n\t"
972 "punpcklbw %%mm7, %%mm2 \n\t" // L7
973 "punpckhbw %%mm7, %%mm3 \n\t" // H7
975 "paddw %%mm2, %%mm2 \n\t" // 2L7
976 "paddw %%mm3, %%mm3 \n\t" // 2H7
977 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
978 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
980 "movq (%%ecx), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
981 "movq 8(%%ecx), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
984 "movq %%mm7, %%mm6 \n\t" // 0
985 "psubw %%mm0, %%mm6 \n\t"
986 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
987 "movq %%mm7, %%mm6 \n\t" // 0
988 "psubw %%mm1, %%mm6 \n\t"
989 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
990 "movq %%mm7, %%mm6 \n\t" // 0
991 "psubw %%mm2, %%mm6 \n\t"
992 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
993 "movq %%mm7, %%mm6 \n\t" // 0
994 "psubw %%mm3, %%mm6 \n\t"
995 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
997 "movq %%mm7, %%mm6 \n\t" // 0
998 "pcmpgtw %%mm0, %%mm6 \n\t"
999 "pxor %%mm6, %%mm0 \n\t"
1000 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1001 "movq %%mm7, %%mm6 \n\t" // 0
1002 "pcmpgtw %%mm1, %%mm6 \n\t"
1003 "pxor %%mm6, %%mm1 \n\t"
1004 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1005 "movq %%mm7, %%mm6 \n\t" // 0
1006 "pcmpgtw %%mm2, %%mm6 \n\t"
1007 "pxor %%mm6, %%mm2 \n\t"
1008 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1009 "movq %%mm7, %%mm6 \n\t" // 0
1010 "pcmpgtw %%mm3, %%mm6 \n\t"
1011 "pxor %%mm6, %%mm3 \n\t"
1012 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1016 "pminsw %%mm2, %%mm0 \n\t"
1017 "pminsw %%mm3, %%mm1 \n\t"
1019 "movq %%mm0, %%mm6 \n\t"
1020 "psubusw %%mm2, %%mm6 \n\t"
1021 "psubw %%mm6, %%mm0 \n\t"
1022 "movq %%mm1, %%mm6 \n\t"
1023 "psubusw %%mm3, %%mm6 \n\t"
1024 "psubw %%mm6, %%mm1 \n\t"
1027 "movq %%mm7, %%mm6 \n\t" // 0
1028 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1029 "pxor %%mm6, %%mm4 \n\t"
1030 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1031 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1032 "pxor %%mm7, %%mm5 \n\t"
1033 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1035 "movd %2, %%mm2 \n\t" // QP
1036 "psllw $3, %%mm2 \n\t" // 8QP
1037 "movq %%mm2, %%mm3 \n\t" // 8QP
1038 "pcmpgtw %%mm4, %%mm2 \n\t"
1039 "pcmpgtw %%mm5, %%mm3 \n\t"
1040 "pand %%mm2, %%mm4 \n\t"
1041 "pand %%mm3, %%mm5 \n\t"
1044 "psubusw %%mm0, %%mm4 \n\t" // hd
1045 "psubusw %%mm1, %%mm5 \n\t" // ld
1048 "movq "MANGLE(w05)", %%mm2 \n\t" // 5
1049 "pmullw %%mm2, %%mm4 \n\t"
1050 "pmullw %%mm2, %%mm5 \n\t"
1051 "movq "MANGLE(w20)", %%mm2 \n\t" // 32
1052 "paddw %%mm2, %%mm4 \n\t"
1053 "paddw %%mm2, %%mm5 \n\t"
1054 "psrlw $6, %%mm4 \n\t"
1055 "psrlw $6, %%mm5 \n\t"
1057 "movq 16(%%ecx), %%mm0 \n\t" // L3 - L4
1058 "movq 24(%%ecx), %%mm1 \n\t" // H3 - H4
1060 "pxor %%mm2, %%mm2 \n\t"
1061 "pxor %%mm3, %%mm3 \n\t"
1063 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1064 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
1065 "pxor %%mm2, %%mm0 \n\t"
1066 "pxor %%mm3, %%mm1 \n\t"
1067 "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
1068 "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
1069 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
1070 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
1072 "pxor %%mm6, %%mm2 \n\t"
1073 "pxor %%mm7, %%mm3 \n\t"
1074 "pand %%mm2, %%mm4 \n\t"
1075 "pand %%mm3, %%mm5 \n\t"
1078 "pminsw %%mm0, %%mm4 \n\t"
1079 "pminsw %%mm1, %%mm5 \n\t"
1081 "movq %%mm4, %%mm2 \n\t"
1082 "psubusw %%mm0, %%mm2 \n\t"
1083 "psubw %%mm2, %%mm4 \n\t"
1084 "movq %%mm5, %%mm2 \n\t"
1085 "psubusw %%mm1, %%mm2 \n\t"
1086 "psubw %%mm2, %%mm5 \n\t"
1088 "pxor %%mm6, %%mm4 \n\t"
1089 "pxor %%mm7, %%mm5 \n\t"
1090 "psubw %%mm6, %%mm4 \n\t"
1091 "psubw %%mm7, %%mm5 \n\t"
1092 "packsswb %%mm5, %%mm4 \n\t"
1093 "movq (%%eax, %1, 2), %%mm0 \n\t"
1094 "paddb %%mm4, %%mm0 \n\t"
1095 "movq %%mm0, (%%eax, %1, 2) \n\t"
1096 "movq (%0, %1, 4), %%mm0 \n\t"
1097 "psubb %%mm4, %%mm0 \n\t"
1098 "movq %%mm0, (%0, %1, 4) \n\t"
1101 : "r" (src), "r" (stride), "m" (c->pQPb)
1102 : "%eax", "%edx", "%ecx"
1105 const int l1= stride;
1106 const int l2= stride + l1;
1107 const int l3= stride + l2;
1108 const int l4= stride + l3;
1109 const int l5= stride + l4;
1110 const int l6= stride + l5;
1111 const int l7= stride + l6;
1112 const int l8= stride + l7;
1113 // const int l9= stride + l8;
1116 for(x=0; x<BLOCK_SIZE; x++)
1118 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1119 if(ABS(middleEnergy) < 8*c->QP)
1121 const int q=(src[l4] - src[l5])/2;
1122 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1123 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1125 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
1129 d*= SIGN(-middleEnergy);
1150 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
1152 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1154 "pxor %%mm6, %%mm6 \n\t"
1155 "pcmpeqb %%mm7, %%mm7 \n\t"
1156 "movq %2, %%mm0 \n\t"
1157 "punpcklbw %%mm6, %%mm0 \n\t"
1158 "psrlw $1, %%mm0 \n\t"
1159 "psubw %%mm7, %%mm0 \n\t"
1160 "packuswb %%mm0, %%mm0 \n\t"
1161 "movq %%mm0, %3 \n\t"
1163 "leal (%0, %1), %%eax \n\t"
1164 "leal (%%eax, %1, 4), %%edx \n\t"
1166 // 0 1 2 3 4 5 6 7 8 9
1167 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1171 #define FIND_MIN_MAX(addr)\
1172 "movq " #addr ", %%mm0 \n\t"\
1173 "pminub %%mm0, %%mm7 \n\t"\
1174 "pmaxub %%mm0, %%mm6 \n\t"
1176 #define FIND_MIN_MAX(addr)\
1177 "movq " #addr ", %%mm0 \n\t"\
1178 "movq %%mm7, %%mm1 \n\t"\
1179 "psubusb %%mm0, %%mm6 \n\t"\
1180 "paddb %%mm0, %%mm6 \n\t"\
1181 "psubusb %%mm0, %%mm1 \n\t"\
1182 "psubb %%mm1, %%mm7 \n\t"
1185 FIND_MIN_MAX((%%eax))
1186 FIND_MIN_MAX((%%eax, %1))
1187 FIND_MIN_MAX((%%eax, %1, 2))
1188 FIND_MIN_MAX((%0, %1, 4))
1189 FIND_MIN_MAX((%%edx))
1190 FIND_MIN_MAX((%%edx, %1))
1191 FIND_MIN_MAX((%%edx, %1, 2))
1192 FIND_MIN_MAX((%0, %1, 8))
1194 "movq %%mm7, %%mm4 \n\t"
1195 "psrlq $8, %%mm7 \n\t"
1197 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1198 "pshufw $0xF9, %%mm7, %%mm4 \n\t"
1199 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1200 "pshufw $0xFE, %%mm7, %%mm4 \n\t"
1201 "pminub %%mm4, %%mm7 \n\t"
1203 "movq %%mm7, %%mm1 \n\t"
1204 "psubusb %%mm4, %%mm1 \n\t"
1205 "psubb %%mm1, %%mm7 \n\t"
1206 "movq %%mm7, %%mm4 \n\t"
1207 "psrlq $16, %%mm7 \n\t"
1208 "movq %%mm7, %%mm1 \n\t"
1209 "psubusb %%mm4, %%mm1 \n\t"
1210 "psubb %%mm1, %%mm7 \n\t"
1211 "movq %%mm7, %%mm4 \n\t"
1212 "psrlq $32, %%mm7 \n\t"
1213 "movq %%mm7, %%mm1 \n\t"
1214 "psubusb %%mm4, %%mm1 \n\t"
1215 "psubb %%mm1, %%mm7 \n\t"
1219 "movq %%mm6, %%mm4 \n\t"
1220 "psrlq $8, %%mm6 \n\t"
1222 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1223 "pshufw $0xF9, %%mm6, %%mm4 \n\t"
1224 "pmaxub %%mm4, %%mm6 \n\t"
1225 "pshufw $0xFE, %%mm6, %%mm4 \n\t"
1226 "pmaxub %%mm4, %%mm6 \n\t"
1228 "psubusb %%mm4, %%mm6 \n\t"
1229 "paddb %%mm4, %%mm6 \n\t"
1230 "movq %%mm6, %%mm4 \n\t"
1231 "psrlq $16, %%mm6 \n\t"
1232 "psubusb %%mm4, %%mm6 \n\t"
1233 "paddb %%mm4, %%mm6 \n\t"
1234 "movq %%mm6, %%mm4 \n\t"
1235 "psrlq $32, %%mm6 \n\t"
1236 "psubusb %%mm4, %%mm6 \n\t"
1237 "paddb %%mm4, %%mm6 \n\t"
1239 "movq %%mm6, %%mm0 \n\t" // max
1240 "psubb %%mm7, %%mm6 \n\t" // max - min
1241 "movd %%mm6, %%ecx \n\t"
1242 "cmpb "MANGLE(deringThreshold)", %%cl \n\t"
1244 "leal -24(%%esp), %%ecx \n\t"
1245 "andl $0xFFFFFFF8, %%ecx \n\t"
1246 PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1247 "punpcklbw %%mm7, %%mm7 \n\t"
1248 "punpcklbw %%mm7, %%mm7 \n\t"
1249 "punpcklbw %%mm7, %%mm7 \n\t"
1250 "movq %%mm7, (%%ecx) \n\t"
1252 "movq (%0), %%mm0 \n\t" // L10
1253 "movq %%mm0, %%mm1 \n\t" // L10
1254 "movq %%mm0, %%mm2 \n\t" // L10
1255 "psllq $8, %%mm1 \n\t"
1256 "psrlq $8, %%mm2 \n\t"
1257 "movd -4(%0), %%mm3 \n\t"
1258 "movd 8(%0), %%mm4 \n\t"
1259 "psrlq $24, %%mm3 \n\t"
1260 "psllq $56, %%mm4 \n\t"
1261 "por %%mm3, %%mm1 \n\t" // L00
1262 "por %%mm4, %%mm2 \n\t" // L20
1263 "movq %%mm1, %%mm3 \n\t" // L00
1264 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1265 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1266 "psubusb %%mm7, %%mm0 \n\t"
1267 "psubusb %%mm7, %%mm2 \n\t"
1268 "psubusb %%mm7, %%mm3 \n\t"
1269 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1270 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1271 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1272 "paddb %%mm2, %%mm0 \n\t"
1273 "paddb %%mm3, %%mm0 \n\t"
1275 "movq (%%eax), %%mm2 \n\t" // L11
1276 "movq %%mm2, %%mm3 \n\t" // L11
1277 "movq %%mm2, %%mm4 \n\t" // L11
1278 "psllq $8, %%mm3 \n\t"
1279 "psrlq $8, %%mm4 \n\t"
1280 "movd -4(%%eax), %%mm5 \n\t"
1281 "movd 8(%%eax), %%mm6 \n\t"
1282 "psrlq $24, %%mm5 \n\t"
1283 "psllq $56, %%mm6 \n\t"
1284 "por %%mm5, %%mm3 \n\t" // L01
1285 "por %%mm6, %%mm4 \n\t" // L21
1286 "movq %%mm3, %%mm5 \n\t" // L01
1287 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1288 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1289 "psubusb %%mm7, %%mm2 \n\t"
1290 "psubusb %%mm7, %%mm4 \n\t"
1291 "psubusb %%mm7, %%mm5 \n\t"
1292 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1293 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1294 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1295 "paddb %%mm4, %%mm2 \n\t"
1296 "paddb %%mm5, %%mm2 \n\t"
1298 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1299 "movq " #src ", " #sx " \n\t" /* src[0] */\
1300 "movq " #sx ", " #lx " \n\t" /* src[0] */\
1301 "movq " #sx ", " #t0 " \n\t" /* src[0] */\
1302 "psllq $8, " #lx " \n\t"\
1303 "psrlq $8, " #t0 " \n\t"\
1304 "movd -4" #src ", " #t1 " \n\t"\
1305 "psrlq $24, " #t1 " \n\t"\
1306 "por " #t1 ", " #lx " \n\t" /* src[-1] */\
1307 "movd 8" #src ", " #t1 " \n\t"\
1308 "psllq $56, " #t1 " \n\t"\
1309 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1310 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1311 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1312 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1314 "movq " #lx ", 8(%%ecx) \n\t"\
1315 "movq (%%ecx), " #lx " \n\t"\
1316 "psubusb " #lx ", " #t1 " \n\t"\
1317 "psubusb " #lx ", " #t0 " \n\t"\
1318 "psubusb " #lx ", " #sx " \n\t"\
1319 "movq "MANGLE(b00)", " #lx " \n\t"\
1320 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1321 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1322 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1323 "paddb " #t1 ", " #t0 " \n\t"\
1324 "paddb " #t0 ", " #sx " \n\t"\
1326 PAVGB(plx, pplx) /* filtered */\
1327 "movq " #dst ", " #t0 " \n\t" /* dst */\
1328 "movq " #t0 ", " #t1 " \n\t" /* dst */\
1329 "psubusb %3, " #t0 " \n\t"\
1330 "paddusb %3, " #t1 " \n\t"\
1332 PMINUB(t1, pplx, t0)\
1333 "paddb " #sx ", " #ppsx " \n\t"\
1334 "paddb " #psx ", " #ppsx " \n\t"\
1335 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1336 "pand "MANGLE(b08)", " #ppsx " \n\t"\
1337 "pcmpeqb " #lx ", " #ppsx " \n\t"\
1338 "pand " #ppsx ", " #pplx " \n\t"\
1339 "pandn " #dst ", " #ppsx " \n\t"\
1340 "por " #pplx ", " #ppsx " \n\t"\
1341 "movq " #ppsx ", " #dst " \n\t"\
1342 "movq 8(%%ecx), " #lx " \n\t"
1359 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
1360 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1361 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1362 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1363 DERING_CORE((%0, %1, 4),(%%edx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1364 DERING_CORE((%%edx),(%%edx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1365 DERING_CORE((%%edx, %1), (%%edx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1366 DERING_CORE((%%edx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1367 DERING_CORE((%0, %1, 8),(%%edx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1370 : : "r" (src), "r" (stride), "m" (c->pQPb), "m"(c->pQPb2)
1371 : "%eax", "%edx", "%ecx"
1380 const int QP2= c->QP/2 + 1;
1389 if(*p > max) max= *p;
1390 if(*p < min) min= *p;
1393 avg= (min + max + 1)>>1;
1395 if(max - min <deringThreshold) return;
1401 if(src[stride*y + 0] > avg) t+= 1;
1402 if(src[stride*y + 1] > avg) t+= 2;
1403 if(src[stride*y + 2] > avg) t+= 4;
1404 if(src[stride*y + 3] > avg) t+= 8;
1405 if(src[stride*y + 4] > avg) t+= 16;
1406 if(src[stride*y + 5] > avg) t+= 32;
1407 if(src[stride*y + 6] > avg) t+= 64;
1408 if(src[stride*y + 7] > avg) t+= 128;
1409 if(src[stride*y + 8] > avg) t+= 256;
1410 if(src[stride*y + 9] > avg) t+= 512;
1413 t &= (t<<1) & (t>>1);
1419 int t = s[y-1] & s[y] & s[y+1];
1435 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1436 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
1437 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1440 #ifdef DEBUG_DERING_THRESHOLD
1441 asm volatile("emms\n\t":);
1443 static long long numPixels=0;
1444 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1445 // if((max-min)<20 || (max-min)*QP<200)
1446 // if((max-min)*QP < 500)
1450 static int numSkiped=0;
1451 static int errorSum=0;
1452 static int worstQP=0;
1453 static int worstRange=0;
1454 static int worstDiff=0;
1456 int absDiff= ABS(diff);
1457 int error= diff*diff;
1459 if(x==1 || x==8 || y==1 || y==8) continue;
1462 if(absDiff > worstDiff)
1466 worstRange= max-min;
1470 if(1024LL*1024LL*1024LL % numSkiped == 0)
1472 printf( "sum:%1.3f, skip:%d, wQP:%d, "
1473 "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1474 (float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
1475 worstDiff, (float)numSkiped/numPixels);
1480 if (*p + QP2 < f) *p= *p + QP2;
1481 else if(*p - QP2 > f) *p= *p - QP2;
1486 #ifdef DEBUG_DERING_THRESHOLD
1497 *p = MIN(*p + 20, 255);
1500 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1507 * Deinterlaces the given block
1508 * will be called for every 8x8 block and can read & write from line 4-15
1509 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1510 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1512 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1514 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1517 "leal (%0, %1), %%eax \n\t"
1518 "leal (%%eax, %1, 4), %%ecx \n\t"
1519 // 0 1 2 3 4 5 6 7 8 9
1520 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
1522 "movq (%0), %%mm0 \n\t"
1523 "movq (%%eax, %1), %%mm1 \n\t"
1525 "movq %%mm0, (%%eax) \n\t"
1526 "movq (%0, %1, 4), %%mm0 \n\t"
1528 "movq %%mm1, (%%eax, %1, 2) \n\t"
1529 "movq (%%ecx, %1), %%mm1 \n\t"
1531 "movq %%mm0, (%%ecx) \n\t"
1532 "movq (%0, %1, 8), %%mm0 \n\t"
1534 "movq %%mm1, (%%ecx, %1, 2) \n\t"
1536 : : "r" (src), "r" (stride)
1544 src[stride] = (src[0] + src[stride*2])>>1;
1545 src[stride*3] = (src[stride*2] + src[stride*4])>>1;
1546 src[stride*5] = (src[stride*4] + src[stride*6])>>1;
1547 src[stride*7] = (src[stride*6] + src[stride*8])>>1;
1554 * Deinterlaces the given block
1555 * will be called for every 8x8 block and can read & write from line 4-15
1556 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1557 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1558 * this filter will read lines 3-15 and write 7-13
1559 * no cliping in C version
1561 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1563 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1566 "leal (%0, %1), %%eax \n\t"
1567 "leal (%%eax, %1, 4), %%edx \n\t"
1568 "leal (%%edx, %1, 4), %%ecx \n\t"
1569 "addl %1, %%ecx \n\t"
1570 "pxor %%mm7, %%mm7 \n\t"
1571 // 0 1 2 3 4 5 6 7 8 9 10
1572 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1574 #define DEINT_CUBIC(a,b,c,d,e)\
1575 "movq " #a ", %%mm0 \n\t"\
1576 "movq " #b ", %%mm1 \n\t"\
1577 "movq " #d ", %%mm2 \n\t"\
1578 "movq " #e ", %%mm3 \n\t"\
1579 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\
1580 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\
1581 "movq %%mm0, %%mm2 \n\t"\
1582 "punpcklbw %%mm7, %%mm0 \n\t"\
1583 "punpckhbw %%mm7, %%mm2 \n\t"\
1584 "movq %%mm1, %%mm3 \n\t"\
1585 "punpcklbw %%mm7, %%mm1 \n\t"\
1586 "punpckhbw %%mm7, %%mm3 \n\t"\
1587 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\
1588 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\
1589 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\
1590 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\
1591 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\
1592 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\
1593 "packuswb %%mm3, %%mm1 \n\t"\
1594 "movq %%mm1, " #c " \n\t"
1596 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx, %1))
1597 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%edx), (%%edx, %1), (%0, %1, 8))
1598 DEINT_CUBIC((%0, %1, 4), (%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%ecx))
1599 DEINT_CUBIC((%%edx, %1), (%0, %1, 8), (%%edx, %1, 4), (%%ecx), (%%ecx, %1, 2))
1601 : : "r" (src), "r" (stride)
1602 : "%eax", "%edx", "ecx"
1609 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4;
1610 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4;
1611 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4;
1612 src[stride*9] = (-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4;
1619 * Deinterlaces the given block
1620 * will be called for every 8x8 block and can read & write from line 4-15
1621 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1622 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1623 * this filter will read lines 4-13 and write 5-11
1624 * no cliping in C version
1626 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
1628 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1631 "leal (%0, %1), %%eax \n\t"
1632 "leal (%%eax, %1, 4), %%edx \n\t"
1633 "pxor %%mm7, %%mm7 \n\t"
1634 "movq (%2), %%mm0 \n\t"
1635 // 0 1 2 3 4 5 6 7 8 9 10
1636 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1638 #define DEINT_FF(a,b,c,d)\
1639 "movq " #a ", %%mm1 \n\t"\
1640 "movq " #b ", %%mm2 \n\t"\
1641 "movq " #c ", %%mm3 \n\t"\
1642 "movq " #d ", %%mm4 \n\t"\
1643 PAVGB(%%mm3, %%mm1) \
1644 PAVGB(%%mm4, %%mm0) \
1645 "movq %%mm0, %%mm3 \n\t"\
1646 "punpcklbw %%mm7, %%mm0 \n\t"\
1647 "punpckhbw %%mm7, %%mm3 \n\t"\
1648 "movq %%mm1, %%mm4 \n\t"\
1649 "punpcklbw %%mm7, %%mm1 \n\t"\
1650 "punpckhbw %%mm7, %%mm4 \n\t"\
1651 "psllw $2, %%mm1 \n\t"\
1652 "psllw $2, %%mm4 \n\t"\
1653 "psubw %%mm0, %%mm1 \n\t"\
1654 "psubw %%mm3, %%mm4 \n\t"\
1655 "movq %%mm2, %%mm5 \n\t"\
1656 "movq %%mm2, %%mm0 \n\t"\
1657 "punpcklbw %%mm7, %%mm2 \n\t"\
1658 "punpckhbw %%mm7, %%mm5 \n\t"\
1659 "paddw %%mm2, %%mm1 \n\t"\
1660 "paddw %%mm5, %%mm4 \n\t"\
1661 "psraw $2, %%mm1 \n\t"\
1662 "psraw $2, %%mm4 \n\t"\
1663 "packuswb %%mm4, %%mm1 \n\t"\
1664 "movq %%mm1, " #b " \n\t"\
1666 DEINT_FF((%0) , (%%eax) , (%%eax, %1), (%%eax, %1, 2))
1667 DEINT_FF((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx) )
1668 DEINT_FF((%0, %1, 4), (%%edx) , (%%edx, %1), (%%edx, %1, 2))
1669 DEINT_FF((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%edx, %1, 4))
1671 "movq %%mm0, (%2) \n\t"
1672 : : "r" (src), "r" (stride), "r"(tmp)
1681 int t2= src[stride*1];
1683 src[stride*1]= (-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3;
1685 src[stride*3]= (-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3;
1687 src[stride*5]= (-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3;
1689 src[stride*7]= (-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3;
1698 * Deinterlaces the given block
1699 * will be called for every 8x8 block and can read & write from line 4-15
1700 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1701 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1702 * will shift the image up by 1 line (FIXME if this is a problem)
1703 * this filter will read lines 4-13 and write 4-11
1705 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride)
1707 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1710 "leal (%0, %1), %%eax \n\t"
1711 "leal (%%eax, %1, 4), %%edx \n\t"
1712 // 0 1 2 3 4 5 6 7 8 9
1713 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1715 "movq (%0), %%mm0 \n\t" // L0
1716 "movq (%%eax, %1), %%mm1 \n\t" // L2
1717 PAVGB(%%mm1, %%mm0) // L0+L2
1718 "movq (%%eax), %%mm2 \n\t" // L1
1720 "movq %%mm0, (%0) \n\t"
1721 "movq (%%eax, %1, 2), %%mm0 \n\t" // L3
1722 PAVGB(%%mm0, %%mm2) // L1+L3
1723 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1724 "movq %%mm2, (%%eax) \n\t"
1725 "movq (%0, %1, 4), %%mm2 \n\t" // L4
1726 PAVGB(%%mm2, %%mm1) // L2+L4
1727 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1728 "movq %%mm1, (%%eax, %1) \n\t"
1729 "movq (%%edx), %%mm1 \n\t" // L5
1730 PAVGB(%%mm1, %%mm0) // L3+L5
1731 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1732 "movq %%mm0, (%%eax, %1, 2) \n\t"
1733 "movq (%%edx, %1), %%mm0 \n\t" // L6
1734 PAVGB(%%mm0, %%mm2) // L4+L6
1735 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1736 "movq %%mm2, (%0, %1, 4) \n\t"
1737 "movq (%%edx, %1, 2), %%mm2 \n\t" // L7
1738 PAVGB(%%mm2, %%mm1) // L5+L7
1739 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1740 "movq %%mm1, (%%edx) \n\t"
1741 "movq (%0, %1, 8), %%mm1 \n\t" // L8
1742 PAVGB(%%mm1, %%mm0) // L6+L8
1743 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1744 "movq %%mm0, (%%edx, %1) \n\t"
1745 "movq (%%edx, %1, 4), %%mm0 \n\t" // L9
1746 PAVGB(%%mm0, %%mm2) // L7+L9
1747 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1748 "movq %%mm2, (%%edx, %1, 2) \n\t"
1751 : : "r" (src), "r" (stride)
1759 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2;
1760 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2;
1761 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
1762 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
1763 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
1764 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
1765 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
1766 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
1773 * Deinterlaces the given block
1774 * will be called for every 8x8 block and can read & write from line 4-15,
1775 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1776 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1778 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1784 "leal (%0, %1), %%eax \n\t"
1785 "leal (%%eax, %1, 4), %%edx \n\t"
1786 // 0 1 2 3 4 5 6 7 8 9
1787 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1789 "movq (%0), %%mm0 \n\t" //
1790 "movq (%%eax, %1), %%mm2 \n\t" //
1791 "movq (%%eax), %%mm1 \n\t" //
1792 "movq %%mm0, %%mm3 \n\t"
1793 "pmaxub %%mm1, %%mm0 \n\t" //
1794 "pminub %%mm3, %%mm1 \n\t" //
1795 "pmaxub %%mm2, %%mm1 \n\t" //
1796 "pminub %%mm1, %%mm0 \n\t"
1797 "movq %%mm0, (%%eax) \n\t"
1799 "movq (%0, %1, 4), %%mm0 \n\t" //
1800 "movq (%%eax, %1, 2), %%mm1 \n\t" //
1801 "movq %%mm2, %%mm3 \n\t"
1802 "pmaxub %%mm1, %%mm2 \n\t" //
1803 "pminub %%mm3, %%mm1 \n\t" //
1804 "pmaxub %%mm0, %%mm1 \n\t" //
1805 "pminub %%mm1, %%mm2 \n\t"
1806 "movq %%mm2, (%%eax, %1, 2) \n\t"
1808 "movq (%%edx), %%mm2 \n\t" //
1809 "movq (%%edx, %1), %%mm1 \n\t" //
1810 "movq %%mm2, %%mm3 \n\t"
1811 "pmaxub %%mm0, %%mm2 \n\t" //
1812 "pminub %%mm3, %%mm0 \n\t" //
1813 "pmaxub %%mm1, %%mm0 \n\t" //
1814 "pminub %%mm0, %%mm2 \n\t"
1815 "movq %%mm2, (%%edx) \n\t"
1817 "movq (%%edx, %1, 2), %%mm2 \n\t" //
1818 "movq (%0, %1, 8), %%mm0 \n\t" //
1819 "movq %%mm2, %%mm3 \n\t"
1820 "pmaxub %%mm0, %%mm2 \n\t" //
1821 "pminub %%mm3, %%mm0 \n\t" //
1822 "pmaxub %%mm1, %%mm0 \n\t" //
1823 "pminub %%mm0, %%mm2 \n\t"
1824 "movq %%mm2, (%%edx, %1, 2) \n\t"
1827 : : "r" (src), "r" (stride)
1831 #else // MMX without MMX2
1833 "leal (%0, %1), %%eax \n\t"
1834 "leal (%%eax, %1, 4), %%edx \n\t"
1835 // 0 1 2 3 4 5 6 7 8 9
1836 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1837 "pxor %%mm7, %%mm7 \n\t"
1839 #define MEDIAN(a,b,c)\
1840 "movq " #a ", %%mm0 \n\t"\
1841 "movq " #b ", %%mm2 \n\t"\
1842 "movq " #c ", %%mm1 \n\t"\
1843 "movq %%mm0, %%mm3 \n\t"\
1844 "movq %%mm1, %%mm4 \n\t"\
1845 "movq %%mm2, %%mm5 \n\t"\
1846 "psubusb %%mm1, %%mm3 \n\t"\
1847 "psubusb %%mm2, %%mm4 \n\t"\
1848 "psubusb %%mm0, %%mm5 \n\t"\
1849 "pcmpeqb %%mm7, %%mm3 \n\t"\
1850 "pcmpeqb %%mm7, %%mm4 \n\t"\
1851 "pcmpeqb %%mm7, %%mm5 \n\t"\
1852 "movq %%mm3, %%mm6 \n\t"\
1853 "pxor %%mm4, %%mm3 \n\t"\
1854 "pxor %%mm5, %%mm4 \n\t"\
1855 "pxor %%mm6, %%mm5 \n\t"\
1856 "por %%mm3, %%mm1 \n\t"\
1857 "por %%mm4, %%mm2 \n\t"\
1858 "por %%mm5, %%mm0 \n\t"\
1859 "pand %%mm2, %%mm0 \n\t"\
1860 "pand %%mm1, %%mm0 \n\t"\
1861 "movq %%mm0, " #b " \n\t"
1863 MEDIAN((%0), (%%eax), (%%eax, %1))
1864 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
1865 MEDIAN((%0, %1, 4), (%%edx), (%%edx, %1))
1866 MEDIAN((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8))
1868 : : "r" (src), "r" (stride)
1878 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2;
1879 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2;
1880 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
1881 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
1882 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
1883 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
1884 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
1885 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
1893 * transposes and shift the given 8x8 Block into dst1 and dst2
1895 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
1898 "leal (%0, %1), %%eax \n\t"
1899 // 0 1 2 3 4 5 6 7 8 9
1900 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1901 "movq (%0), %%mm0 \n\t" // 12345678
1902 "movq (%%eax), %%mm1 \n\t" // abcdefgh
1903 "movq %%mm0, %%mm2 \n\t" // 12345678
1904 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
1905 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
1907 "movq (%%eax, %1), %%mm1 \n\t"
1908 "movq (%%eax, %1, 2), %%mm3 \n\t"
1909 "movq %%mm1, %%mm4 \n\t"
1910 "punpcklbw %%mm3, %%mm1 \n\t"
1911 "punpckhbw %%mm3, %%mm4 \n\t"
1913 "movq %%mm0, %%mm3 \n\t"
1914 "punpcklwd %%mm1, %%mm0 \n\t"
1915 "punpckhwd %%mm1, %%mm3 \n\t"
1916 "movq %%mm2, %%mm1 \n\t"
1917 "punpcklwd %%mm4, %%mm2 \n\t"
1918 "punpckhwd %%mm4, %%mm1 \n\t"
1920 "movd %%mm0, 128(%2) \n\t"
1921 "psrlq $32, %%mm0 \n\t"
1922 "movd %%mm0, 144(%2) \n\t"
1923 "movd %%mm3, 160(%2) \n\t"
1924 "psrlq $32, %%mm3 \n\t"
1925 "movd %%mm3, 176(%2) \n\t"
1926 "movd %%mm3, 48(%3) \n\t"
1927 "movd %%mm2, 192(%2) \n\t"
1928 "movd %%mm2, 64(%3) \n\t"
1929 "psrlq $32, %%mm2 \n\t"
1930 "movd %%mm2, 80(%3) \n\t"
1931 "movd %%mm1, 96(%3) \n\t"
1932 "psrlq $32, %%mm1 \n\t"
1933 "movd %%mm1, 112(%3) \n\t"
1935 "leal (%%eax, %1, 4), %%eax \n\t"
1937 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
1938 "movq (%%eax), %%mm1 \n\t" // abcdefgh
1939 "movq %%mm0, %%mm2 \n\t" // 12345678
1940 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
1941 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
1943 "movq (%%eax, %1), %%mm1 \n\t"
1944 "movq (%%eax, %1, 2), %%mm3 \n\t"
1945 "movq %%mm1, %%mm4 \n\t"
1946 "punpcklbw %%mm3, %%mm1 \n\t"
1947 "punpckhbw %%mm3, %%mm4 \n\t"
1949 "movq %%mm0, %%mm3 \n\t"
1950 "punpcklwd %%mm1, %%mm0 \n\t"
1951 "punpckhwd %%mm1, %%mm3 \n\t"
1952 "movq %%mm2, %%mm1 \n\t"
1953 "punpcklwd %%mm4, %%mm2 \n\t"
1954 "punpckhwd %%mm4, %%mm1 \n\t"
1956 "movd %%mm0, 132(%2) \n\t"
1957 "psrlq $32, %%mm0 \n\t"
1958 "movd %%mm0, 148(%2) \n\t"
1959 "movd %%mm3, 164(%2) \n\t"
1960 "psrlq $32, %%mm3 \n\t"
1961 "movd %%mm3, 180(%2) \n\t"
1962 "movd %%mm3, 52(%3) \n\t"
1963 "movd %%mm2, 196(%2) \n\t"
1964 "movd %%mm2, 68(%3) \n\t"
1965 "psrlq $32, %%mm2 \n\t"
1966 "movd %%mm2, 84(%3) \n\t"
1967 "movd %%mm1, 100(%3) \n\t"
1968 "psrlq $32, %%mm1 \n\t"
1969 "movd %%mm1, 116(%3) \n\t"
1972 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)
1978 * transposes the given 8x8 block
1980 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
1983 "leal (%0, %1), %%eax \n\t"
1984 "leal (%%eax, %1, 4), %%edx \n\t"
1985 // 0 1 2 3 4 5 6 7 8 9
1986 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1987 "movq (%2), %%mm0 \n\t" // 12345678
1988 "movq 16(%2), %%mm1 \n\t" // abcdefgh
1989 "movq %%mm0, %%mm2 \n\t" // 12345678
1990 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
1991 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
1993 "movq 32(%2), %%mm1 \n\t"
1994 "movq 48(%2), %%mm3 \n\t"
1995 "movq %%mm1, %%mm4 \n\t"
1996 "punpcklbw %%mm3, %%mm1 \n\t"
1997 "punpckhbw %%mm3, %%mm4 \n\t"
1999 "movq %%mm0, %%mm3 \n\t"
2000 "punpcklwd %%mm1, %%mm0 \n\t"
2001 "punpckhwd %%mm1, %%mm3 \n\t"
2002 "movq %%mm2, %%mm1 \n\t"
2003 "punpcklwd %%mm4, %%mm2 \n\t"
2004 "punpckhwd %%mm4, %%mm1 \n\t"
2006 "movd %%mm0, (%0) \n\t"
2007 "psrlq $32, %%mm0 \n\t"
2008 "movd %%mm0, (%%eax) \n\t"
2009 "movd %%mm3, (%%eax, %1) \n\t"
2010 "psrlq $32, %%mm3 \n\t"
2011 "movd %%mm3, (%%eax, %1, 2) \n\t"
2012 "movd %%mm2, (%0, %1, 4) \n\t"
2013 "psrlq $32, %%mm2 \n\t"
2014 "movd %%mm2, (%%edx) \n\t"
2015 "movd %%mm1, (%%edx, %1) \n\t"
2016 "psrlq $32, %%mm1 \n\t"
2017 "movd %%mm1, (%%edx, %1, 2) \n\t"
2020 "movq 64(%2), %%mm0 \n\t" // 12345678
2021 "movq 80(%2), %%mm1 \n\t" // abcdefgh
2022 "movq %%mm0, %%mm2 \n\t" // 12345678
2023 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2024 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2026 "movq 96(%2), %%mm1 \n\t"
2027 "movq 112(%2), %%mm3 \n\t"
2028 "movq %%mm1, %%mm4 \n\t"
2029 "punpcklbw %%mm3, %%mm1 \n\t"
2030 "punpckhbw %%mm3, %%mm4 \n\t"
2032 "movq %%mm0, %%mm3 \n\t"
2033 "punpcklwd %%mm1, %%mm0 \n\t"
2034 "punpckhwd %%mm1, %%mm3 \n\t"
2035 "movq %%mm2, %%mm1 \n\t"
2036 "punpcklwd %%mm4, %%mm2 \n\t"
2037 "punpckhwd %%mm4, %%mm1 \n\t"
2039 "movd %%mm0, 4(%0) \n\t"
2040 "psrlq $32, %%mm0 \n\t"
2041 "movd %%mm0, 4(%%eax) \n\t"
2042 "movd %%mm3, 4(%%eax, %1) \n\t"
2043 "psrlq $32, %%mm3 \n\t"
2044 "movd %%mm3, 4(%%eax, %1, 2) \n\t"
2045 "movd %%mm2, 4(%0, %1, 4) \n\t"
2046 "psrlq $32, %%mm2 \n\t"
2047 "movd %%mm2, 4(%%edx) \n\t"
2048 "movd %%mm1, 4(%%edx, %1) \n\t"
2049 "psrlq $32, %%mm1 \n\t"
2050 "movd %%mm1, 4(%%edx, %1, 2) \n\t"
2052 :: "r" (dst), "r" (dstStride), "r" (src)
2057 //static int test=0;
2059 static void inline RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2060 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2062 // to save a register (FIXME do this outside of the loops)
2063 tempBluredPast[127]= maxNoise[0];
2064 tempBluredPast[128]= maxNoise[1];
2065 tempBluredPast[129]= maxNoise[2];
2067 #define FAST_L2_DIFF
2068 //#define L1_DIFF //u should change the thresholds too if u try that one
2069 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2071 "leal (%2, %2, 2), %%eax \n\t" // 3*stride
2072 "leal (%2, %2, 4), %%edx \n\t" // 5*stride
2073 "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
2074 // 0 1 2 3 4 5 6 7 8 9
2075 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
2077 #ifdef L1_DIFF //needs mmx2
2078 "movq (%0), %%mm0 \n\t" // L0
2079 "psadbw (%1), %%mm0 \n\t" // |L0-R0|
2080 "movq (%0, %2), %%mm1 \n\t" // L1
2081 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2082 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2083 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2084 "movq (%0, %%eax), %%mm3 \n\t" // L3
2085 "psadbw (%1, %%eax), %%mm3 \n\t" // |L3-R3|
2087 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2088 "paddw %%mm1, %%mm0 \n\t"
2089 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2090 "movq (%0, %%edx), %%mm5 \n\t" // L5
2091 "paddw %%mm2, %%mm0 \n\t"
2092 "psadbw (%1, %%edx), %%mm5 \n\t" // |L5-R5|
2093 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
2094 "paddw %%mm3, %%mm0 \n\t"
2095 "psadbw (%1, %%eax, 2), %%mm6 \n\t" // |L6-R6|
2096 "movq (%0, %%ecx), %%mm7 \n\t" // L7
2097 "paddw %%mm4, %%mm0 \n\t"
2098 "psadbw (%1, %%ecx), %%mm7 \n\t" // |L7-R7|
2099 "paddw %%mm5, %%mm6 \n\t"
2100 "paddw %%mm7, %%mm6 \n\t"
2101 "paddw %%mm6, %%mm0 \n\t"
2102 #elif defined (FAST_L2_DIFF)
2103 "pcmpeqb %%mm7, %%mm7 \n\t"
2104 "movq "MANGLE(b80)", %%mm6 \n\t"
2105 "pxor %%mm0, %%mm0 \n\t"
2106 #define L2_DIFF_CORE(a, b)\
2107 "movq " #a ", %%mm5 \n\t"\
2108 "movq " #b ", %%mm2 \n\t"\
2109 "pxor %%mm7, %%mm2 \n\t"\
2110 PAVGB(%%mm2, %%mm5)\
2111 "paddb %%mm6, %%mm5 \n\t"\
2112 "movq %%mm5, %%mm2 \n\t"\
2113 "psllw $8, %%mm5 \n\t"\
2114 "pmaddwd %%mm5, %%mm5 \n\t"\
2115 "pmaddwd %%mm2, %%mm2 \n\t"\
2116 "paddd %%mm2, %%mm5 \n\t"\
2117 "psrld $14, %%mm5 \n\t"\
2118 "paddd %%mm5, %%mm0 \n\t"
2120 L2_DIFF_CORE((%0), (%1))
2121 L2_DIFF_CORE((%0, %2), (%1, %2))
2122 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2123 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2124 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2125 L2_DIFF_CORE((%0, %%edx), (%1, %%edx))
2126 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2127 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2130 "pxor %%mm7, %%mm7 \n\t"
2131 "pxor %%mm0, %%mm0 \n\t"
2132 #define L2_DIFF_CORE(a, b)\
2133 "movq " #a ", %%mm5 \n\t"\
2134 "movq " #b ", %%mm2 \n\t"\
2135 "movq %%mm5, %%mm1 \n\t"\
2136 "movq %%mm2, %%mm3 \n\t"\
2137 "punpcklbw %%mm7, %%mm5 \n\t"\
2138 "punpckhbw %%mm7, %%mm1 \n\t"\
2139 "punpcklbw %%mm7, %%mm2 \n\t"\
2140 "punpckhbw %%mm7, %%mm3 \n\t"\
2141 "psubw %%mm2, %%mm5 \n\t"\
2142 "psubw %%mm3, %%mm1 \n\t"\
2143 "pmaddwd %%mm5, %%mm5 \n\t"\
2144 "pmaddwd %%mm1, %%mm1 \n\t"\
2145 "paddd %%mm1, %%mm5 \n\t"\
2146 "paddd %%mm5, %%mm0 \n\t"
2148 L2_DIFF_CORE((%0), (%1))
2149 L2_DIFF_CORE((%0, %2), (%1, %2))
2150 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2151 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2152 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2153 L2_DIFF_CORE((%0, %%edx), (%1, %%edx))
2154 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2155 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2159 "movq %%mm0, %%mm4 \n\t"
2160 "psrlq $32, %%mm0 \n\t"
2161 "paddd %%mm0, %%mm4 \n\t"
2162 "movd %%mm4, %%ecx \n\t"
2163 "shll $2, %%ecx \n\t"
2164 "movl %3, %%edx \n\t"
2165 "addl -4(%%edx), %%ecx \n\t"
2166 "addl 4(%%edx), %%ecx \n\t"
2167 "addl -1024(%%edx), %%ecx \n\t"
2168 "addl $4, %%ecx \n\t"
2169 "addl 1024(%%edx), %%ecx \n\t"
2170 "shrl $3, %%ecx \n\t"
2171 "movl %%ecx, (%%edx) \n\t"
2173 // "movl %3, %%ecx \n\t"
2174 // "movl %%ecx, test \n\t"
2176 "cmpl 512(%%edx), %%ecx \n\t"
2178 "cmpl 516(%%edx), %%ecx \n\t"
2181 "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
2182 "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
2183 "movq (%0), %%mm0 \n\t" // L0
2184 "movq (%0, %2), %%mm1 \n\t" // L1
2185 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2186 "movq (%0, %%eax), %%mm3 \n\t" // L3
2187 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2188 "movq (%0, %%edx), %%mm5 \n\t" // L5
2189 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
2190 "movq (%0, %%ecx), %%mm7 \n\t" // L7
2191 "movq %%mm0, (%1) \n\t" // L0
2192 "movq %%mm1, (%1, %2) \n\t" // L1
2193 "movq %%mm2, (%1, %2, 2) \n\t" // L2
2194 "movq %%mm3, (%1, %%eax) \n\t" // L3
2195 "movq %%mm4, (%1, %2, 4) \n\t" // L4
2196 "movq %%mm5, (%1, %%edx) \n\t" // L5
2197 "movq %%mm6, (%1, %%eax, 2) \n\t" // L6
2198 "movq %%mm7, (%1, %%ecx) \n\t" // L7
2202 "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
2203 "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
2204 "movq (%0), %%mm0 \n\t" // L0
2205 PAVGB((%1), %%mm0) // L0
2206 "movq (%0, %2), %%mm1 \n\t" // L1
2207 PAVGB((%1, %2), %%mm1) // L1
2208 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2209 PAVGB((%1, %2, 2), %%mm2) // L2
2210 "movq (%0, %%eax), %%mm3 \n\t" // L3
2211 PAVGB((%1, %%eax), %%mm3) // L3
2212 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2213 PAVGB((%1, %2, 4), %%mm4) // L4
2214 "movq (%0, %%edx), %%mm5 \n\t" // L5
2215 PAVGB((%1, %%edx), %%mm5) // L5
2216 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
2217 PAVGB((%1, %%eax, 2), %%mm6) // L6
2218 "movq (%0, %%ecx), %%mm7 \n\t" // L7
2219 PAVGB((%1, %%ecx), %%mm7) // L7
2220 "movq %%mm0, (%1) \n\t" // R0
2221 "movq %%mm1, (%1, %2) \n\t" // R1
2222 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2223 "movq %%mm3, (%1, %%eax) \n\t" // R3
2224 "movq %%mm4, (%1, %2, 4) \n\t" // R4
2225 "movq %%mm5, (%1, %%edx) \n\t" // R5
2226 "movq %%mm6, (%1, %%eax, 2) \n\t" // R6
2227 "movq %%mm7, (%1, %%ecx) \n\t" // R7
2228 "movq %%mm0, (%0) \n\t" // L0
2229 "movq %%mm1, (%0, %2) \n\t" // L1
2230 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2231 "movq %%mm3, (%0, %%eax) \n\t" // L3
2232 "movq %%mm4, (%0, %2, 4) \n\t" // L4
2233 "movq %%mm5, (%0, %%edx) \n\t" // L5
2234 "movq %%mm6, (%0, %%eax, 2) \n\t" // L6
2235 "movq %%mm7, (%0, %%ecx) \n\t" // L7
2239 "cmpl 508(%%edx), %%ecx \n\t"
2242 "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
2243 "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
2244 "movq (%0), %%mm0 \n\t" // L0
2245 "movq (%0, %2), %%mm1 \n\t" // L1
2246 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2247 "movq (%0, %%eax), %%mm3 \n\t" // L3
2248 "movq (%1), %%mm4 \n\t" // R0
2249 "movq (%1, %2), %%mm5 \n\t" // R1
2250 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2251 "movq (%1, %%eax), %%mm7 \n\t" // R3
2260 "movq %%mm0, (%1) \n\t" // R0
2261 "movq %%mm1, (%1, %2) \n\t" // R1
2262 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2263 "movq %%mm3, (%1, %%eax) \n\t" // R3
2264 "movq %%mm0, (%0) \n\t" // L0
2265 "movq %%mm1, (%0, %2) \n\t" // L1
2266 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2267 "movq %%mm3, (%0, %%eax) \n\t" // L3
2269 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2270 "movq (%0, %%edx), %%mm1 \n\t" // L5
2271 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6
2272 "movq (%0, %%ecx), %%mm3 \n\t" // L7
2273 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2274 "movq (%1, %%edx), %%mm5 \n\t" // R5
2275 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6
2276 "movq (%1, %%ecx), %%mm7 \n\t" // R7
2285 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2286 "movq %%mm1, (%1, %%edx) \n\t" // R5
2287 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6
2288 "movq %%mm3, (%1, %%ecx) \n\t" // R7
2289 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2290 "movq %%mm1, (%0, %%edx) \n\t" // L5
2291 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6
2292 "movq %%mm3, (%0, %%ecx) \n\t" // L7
2296 "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
2297 "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
2298 "movq (%0), %%mm0 \n\t" // L0
2299 "movq (%0, %2), %%mm1 \n\t" // L1
2300 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2301 "movq (%0, %%eax), %%mm3 \n\t" // L3
2302 "movq (%1), %%mm4 \n\t" // R0
2303 "movq (%1, %2), %%mm5 \n\t" // R1
2304 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2305 "movq (%1, %%eax), %%mm7 \n\t" // R3
2318 "movq %%mm0, (%1) \n\t" // R0
2319 "movq %%mm1, (%1, %2) \n\t" // R1
2320 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2321 "movq %%mm3, (%1, %%eax) \n\t" // R3
2322 "movq %%mm0, (%0) \n\t" // L0
2323 "movq %%mm1, (%0, %2) \n\t" // L1
2324 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2325 "movq %%mm3, (%0, %%eax) \n\t" // L3
2327 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2328 "movq (%0, %%edx), %%mm1 \n\t" // L5
2329 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6
2330 "movq (%0, %%ecx), %%mm3 \n\t" // L7
2331 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2332 "movq (%1, %%edx), %%mm5 \n\t" // R5
2333 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6
2334 "movq (%1, %%ecx), %%mm7 \n\t" // R7
2347 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2348 "movq %%mm1, (%1, %%edx) \n\t" // R5
2349 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6
2350 "movq %%mm3, (%1, %%ecx) \n\t" // R7
2351 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2352 "movq %%mm1, (%0, %%edx) \n\t" // L5
2353 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6
2354 "movq %%mm3, (%0, %%ecx) \n\t" // L7
2358 :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast)
2359 : "%eax", "%edx", "%ecx", "memory"
2361 //printf("%d\n", test);
2374 int ref= tempBlured[ x + y*stride ];
2375 int cur= src[ x + y*stride ];
2377 // if(x==0 || x==7) d1+= d1>>1;
2378 // if(y==0 || y==7) d1+= d1>>1;
2387 +(*(tempBluredPast-256))
2388 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
2389 +(*(tempBluredPast+256))
2392 // ((*tempBluredPast)*3 + d + 2)>>2;
2394 //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]);
2398 64 32 16 8 4 2 1 (1)
2399 64 48 36 27 20 15 11 (33) (approx)
2400 64 56 49 43 37 33 29 (200) (approx)
2411 int ref= tempBlured[ x + y*stride ];
2412 int cur= src[ x + y*stride ];
2413 tempBlured[ x + y*stride ]=
2414 src[ x + y*stride ]=
2426 tempBlured[ x + y*stride ]= src[ x + y*stride ];
2440 int ref= tempBlured[ x + y*stride ];
2441 int cur= src[ x + y*stride ];
2442 tempBlured[ x + y*stride ]=
2443 src[ x + y*stride ]=
2444 (ref*7 + cur + 4)>>3;
2455 int ref= tempBlured[ x + y*stride ];
2456 int cur= src[ x + y*stride ];
2457 tempBlured[ x + y*stride ]=
2458 src[ x + y*stride ]=
2459 (ref*3 + cur + 2)>>2;
2468 static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2469 QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c);
2472 * Copies a block from src to dst and fixes the blacklevel
2473 * levelFix == 0 -> dont touch the brighness & contrast
2477 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
2478 int levelFix, int64_t *packedOffsetAndScale)
2487 "movq (%%eax), %%mm2 \n\t" // packedYOffset
2488 "movq 8(%%eax), %%mm3 \n\t" // packedYScale
2489 "leal (%2,%4), %%eax \n\t"
2490 "leal (%3,%5), %%edx \n\t"
2491 "pxor %%mm4, %%mm4 \n\t"
2493 #define SCALED_CPY(src1, src2, dst1, dst2) \
2494 "movq " #src1 ", %%mm0 \n\t"\
2495 "movq " #src1 ", %%mm5 \n\t"\
2496 "movq " #src2 ", %%mm1 \n\t"\
2497 "movq " #src2 ", %%mm6 \n\t"\
2498 "punpcklbw %%mm0, %%mm0 \n\t"\
2499 "punpckhbw %%mm5, %%mm5 \n\t"\
2500 "punpcklbw %%mm1, %%mm1 \n\t"\
2501 "punpckhbw %%mm6, %%mm6 \n\t"\
2502 "pmulhuw %%mm3, %%mm0 \n\t"\
2503 "pmulhuw %%mm3, %%mm5 \n\t"\
2504 "pmulhuw %%mm3, %%mm1 \n\t"\
2505 "pmulhuw %%mm3, %%mm6 \n\t"\
2506 "psubw %%mm2, %%mm0 \n\t"\
2507 "psubw %%mm2, %%mm5 \n\t"\
2508 "psubw %%mm2, %%mm1 \n\t"\
2509 "psubw %%mm2, %%mm6 \n\t"\
2510 "packuswb %%mm5, %%mm0 \n\t"\
2511 "packuswb %%mm6, %%mm1 \n\t"\
2512 "movq %%mm0, " #dst1 " \n\t"\
2513 "movq %%mm1, " #dst2 " \n\t"\
2516 #define SCALED_CPY(src1, src2, dst1, dst2) \
2517 "movq " #src1 ", %%mm0 \n\t"\
2518 "movq " #src1 ", %%mm5 \n\t"\
2519 "punpcklbw %%mm4, %%mm0 \n\t"\
2520 "punpckhbw %%mm4, %%mm5 \n\t"\
2521 "psubw %%mm2, %%mm0 \n\t"\
2522 "psubw %%mm2, %%mm5 \n\t"\
2523 "movq " #src2 ", %%mm1 \n\t"\
2524 "psllw $6, %%mm0 \n\t"\
2525 "psllw $6, %%mm5 \n\t"\
2526 "pmulhw %%mm3, %%mm0 \n\t"\
2527 "movq " #src2 ", %%mm6 \n\t"\
2528 "pmulhw %%mm3, %%mm5 \n\t"\
2529 "punpcklbw %%mm4, %%mm1 \n\t"\
2530 "punpckhbw %%mm4, %%mm6 \n\t"\
2531 "psubw %%mm2, %%mm1 \n\t"\
2532 "psubw %%mm2, %%mm6 \n\t"\
2533 "psllw $6, %%mm1 \n\t"\
2534 "psllw $6, %%mm6 \n\t"\
2535 "pmulhw %%mm3, %%mm1 \n\t"\
2536 "pmulhw %%mm3, %%mm6 \n\t"\
2537 "packuswb %%mm5, %%mm0 \n\t"\
2538 "packuswb %%mm6, %%mm1 \n\t"\
2539 "movq %%mm0, " #dst1 " \n\t"\
2540 "movq %%mm1, " #dst2 " \n\t"\
2544 SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5))
2545 SCALED_CPY((%2, %4, 2), (%%eax, %4, 2), (%3, %5, 2), (%%edx, %5, 2))
2546 SCALED_CPY((%2, %4, 4), (%%eax, %4, 4), (%3, %5, 4), (%%edx, %5, 4))
2547 "leal (%%eax,%4,4), %%eax \n\t"
2548 "leal (%%edx,%5,4), %%edx \n\t"
2549 SCALED_CPY((%%eax, %4), (%%eax, %4, 2), (%%edx, %5), (%%edx, %5, 2))
2552 : "=&a" (packedOffsetAndScale)
2553 : "0" (packedOffsetAndScale),
2562 memcpy( &(dst[dstStride*i]),
2563 &(src[srcStride*i]), BLOCK_SIZE);
2570 "leal (%0,%2), %%eax \n\t"
2571 "leal (%1,%3), %%edx \n\t"
2573 #define SIMPLE_CPY(src1, src2, dst1, dst2) \
2574 "movq " #src1 ", %%mm0 \n\t"\
2575 "movq " #src2 ", %%mm1 \n\t"\
2576 "movq %%mm0, " #dst1 " \n\t"\
2577 "movq %%mm1, " #dst2 " \n\t"\
2579 SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3))
2580 SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%edx, %3, 2))
2581 SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%edx, %3, 4))
2582 "leal (%%eax,%2,4), %%eax \n\t"
2583 "leal (%%edx,%3,4), %%edx \n\t"
2584 SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%edx, %3), (%%edx, %3, 2))
2594 memcpy( &(dst[dstStride*i]),
2595 &(src[srcStride*i]), BLOCK_SIZE);
2601 * Duplicates the given 8 src pixels ? times upward
2603 static inline void RENAME(duplicate)(uint8_t src[], int stride)
2607 "movq (%0), %%mm0 \n\t"
2609 "movq %%mm0, (%0) \n\t"
2610 "movq %%mm0, (%0, %1) \n\t"
2611 "movq %%mm0, (%0, %1, 2) \n\t"
2627 * Filters array of bytes (Y or U or V values)
2629 static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2630 QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)
2632 PPContext __attribute__((aligned(8))) c= *c2; //copy to stack for faster access
2634 #ifdef COMPILE_TIME_MODE
2635 const int mode= COMPILE_TIME_MODE;
2637 const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;
2639 int black=0, white=255; // blackest black and whitest white in the picture
2640 int QPCorrecture= 256*256;
2645 uint64_t * const yHistogram= c.yHistogram;
2646 uint8_t * const tempSrc= c.tempSrc;
2647 uint8_t * const tempDst= c.tempDst;
2649 c.dcOffset= c.ppMode.maxDcDiff;
2650 c.dcThreshold= c.ppMode.maxDcDiff*2 + 1;
2653 c.mmxDcOffset= 0x7F - c.dcOffset;
2654 c.mmxDcThreshold= 0x7F - c.dcThreshold;
2656 c.mmxDcOffset*= 0x0101010101010101LL;
2657 c.mmxDcThreshold*= 0x0101010101010101LL;
2660 if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
2661 else if( (mode & LINEAR_BLEND_DEINT_FILTER)
2662 || (mode & FFMPEG_DEINT_FILTER)) copyAhead=14;
2663 else if( (mode & V_DEBLOCK)
2664 || (mode & LINEAR_IPOL_DEINT_FILTER)
2665 || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
2666 else if(mode & V_X1_FILTER) copyAhead=11;
2667 // else if(mode & V_RK1_FILTER) copyAhead=10;
2668 else if(mode & DERING) copyAhead=9;
2677 uint64_t maxClipped;
2682 // first frame is fscked so we ignore it
2683 if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256;
2685 for(i=0; i<256; i++)
2687 sum+= yHistogram[i];
2688 // printf("%d ", yHistogram[i]);
2692 /* we allways get a completly black picture first */
2693 maxClipped= (uint64_t)(sum * maxClippedThreshold);
2696 for(black=255; black>0; black--)
2698 if(clipped < maxClipped) break;
2699 clipped-= yHistogram[black];
2703 for(white=0; white<256; white++)
2705 if(clipped < maxClipped) break;
2706 clipped-= yHistogram[white];
2709 scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);
2712 c.packedYScale= (uint16_t)(scale*256.0 + 0.5);
2713 c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;
2715 c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);
2716 c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;
2719 c.packedYOffset|= c.packedYOffset<<32;
2720 c.packedYOffset|= c.packedYOffset<<16;
2722 c.packedYScale|= c.packedYScale<<32;
2723 c.packedYScale|= c.packedYScale<<16;
2725 if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5);
2726 else QPCorrecture= 256*256;
2730 c.packedYScale= 0x0100010001000100LL;
2732 QPCorrecture= 256*256;
2735 /* copy & deinterlace first row of blocks */
2738 uint8_t *srcBlock= &(src[y*srcStride]);
2739 uint8_t *dstBlock= tempDst + dstStride;
2741 // From this point on it is guranteed that we can read and write 16 lines downward
2742 // finish 1 block before the next otherwise we´ll might have a problem
2743 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
2744 for(x=0; x<width; x+=BLOCK_SIZE)
2749 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
2750 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
2751 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
2752 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
2756 "movl %4, %%eax \n\t"
2757 "shrl $2, %%eax \n\t"
2758 "andl $6, %%eax \n\t"
2759 "addl %5, %%eax \n\t"
2760 "movl %%eax, %%edx \n\t"
2761 "imul %1, %%eax \n\t"
2762 "imul %3, %%edx \n\t"
2763 "prefetchnta 32(%%eax, %0) \n\t"
2764 "prefetcht0 32(%%edx, %2) \n\t"
2765 "addl %1, %%eax \n\t"
2766 "addl %3, %%edx \n\t"
2767 "prefetchnta 32(%%eax, %0) \n\t"
2768 "prefetcht0 32(%%edx, %2) \n\t"
2769 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
2770 "m" (x), "m" (copyAhead)
2774 #elif defined(HAVE_3DNOW)
2775 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
2776 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
2777 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
2778 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
2779 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
2783 RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
2784 srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
2786 RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
2788 if(mode & LINEAR_IPOL_DEINT_FILTER)
2789 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
2790 else if(mode & LINEAR_BLEND_DEINT_FILTER)
2791 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
2792 else if(mode & MEDIAN_DEINT_FILTER)
2793 RENAME(deInterlaceMedian)(dstBlock, dstStride);
2794 else if(mode & CUBIC_IPOL_DEINT_FILTER)
2795 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
2796 else if(mode & FFMPEG_DEINT_FILTER)
2797 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
2798 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
2799 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
2804 memcpy(dst, tempDst + 9*dstStride, copyAhead*dstStride );
2808 for(y=0; y<height; y+=BLOCK_SIZE)
2810 //1% speedup if these are here instead of the inner loop
2811 uint8_t *srcBlock= &(src[y*srcStride]);
2812 uint8_t *dstBlock= &(dst[y*dstStride]);
2814 uint8_t *tempBlock1= c.tempBlocks;
2815 uint8_t *tempBlock2= c.tempBlocks + 8;
2818 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride];
2819 int QPDelta= isColor ? (-1) : 1<<31;
2823 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
2824 if not than use a temporary buffer */
2828 /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
2829 blockcopy to dst later */
2830 memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
2831 srcStride*MAX(height-y-copyAhead, 0) );
2833 /* duplicate last line of src to fill the void upto line (copyAhead+7) */
2834 for(i=MAX(height-y, 8); i<copyAhead+8; i++)
2835 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);
2837 /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
2838 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );
2840 /* duplicate last line of dst to fill the void upto line (copyAhead) */
2841 for(i=height-y+1; i<=copyAhead; i++)
2842 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);
2844 dstBlock= tempDst + dstStride;
2849 // From this point on it is guranteed that we can read and write 16 lines downward
2850 // finish 1 block before the next otherwise we´ll might have a problem
2851 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
2852 for(x=0; x<width; x+=BLOCK_SIZE)
2854 const int stride= dstStride;
2862 "sbbl %%eax, %%eax \n\t"
2863 "shll $2, %%eax \n\t"
2864 "subl %%eax, %0 \n\t"
2865 : "+r" (QPptr), "+m" (QPFrac)
2871 QPs[(y>>3)*QPStride + (x>>3)]:
2872 QPs[(y>>4)*QPStride + (x>>4)];
2876 QP= (QP* QPCorrecture + 256*128)>>16;
2877 yHistogram[ srcBlock[srcStride*12 + 4] ]++;
2879 //printf("%d ", QP);
2883 "movd %1, %%mm7 \n\t"
2884 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
2885 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
2886 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
2887 "movq %%mm7, %0 \n\t"
2896 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
2897 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
2898 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
2899 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
2903 "movl %4, %%eax \n\t"
2904 "shrl $2, %%eax \n\t"
2905 "andl $6, %%eax \n\t"
2906 "addl %5, %%eax \n\t"
2907 "movl %%eax, %%edx \n\t"
2908 "imul %1, %%eax \n\t"
2909 "imul %3, %%edx \n\t"
2910 "prefetchnta 32(%%eax, %0) \n\t"
2911 "prefetcht0 32(%%edx, %2) \n\t"
2912 "addl %1, %%eax \n\t"
2913 "addl %3, %%edx \n\t"
2914 "prefetchnta 32(%%eax, %0) \n\t"
2915 "prefetcht0 32(%%edx, %2) \n\t"
2916 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
2917 "m" (x), "m" (copyAhead)
2921 #elif defined(HAVE_3DNOW)
2922 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
2923 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
2924 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
2925 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
2926 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
2930 RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
2931 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
2933 if(mode & LINEAR_IPOL_DEINT_FILTER)
2934 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
2935 else if(mode & LINEAR_BLEND_DEINT_FILTER)
2936 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
2937 else if(mode & MEDIAN_DEINT_FILTER)
2938 RENAME(deInterlaceMedian)(dstBlock, dstStride);
2939 else if(mode & CUBIC_IPOL_DEINT_FILTER)
2940 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
2941 else if(mode & FFMPEG_DEINT_FILTER)
2942 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
2943 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
2944 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
2947 /* only deblock if we have 2 blocks */
2950 if(mode & V_X1_FILTER)
2951 RENAME(vertX1Filter)(dstBlock, stride, &c);
2952 else if(mode & V_DEBLOCK)
2954 if( RENAME(isVertDC)(dstBlock, stride, &c))
2956 if(RENAME(isVertMinMaxOk)(dstBlock, stride, &c))
2957 RENAME(doVertLowPass)(dstBlock, stride, &c);
2960 RENAME(doVertDefFilter)(dstBlock, stride, &c);
2965 RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
2967 /* check if we have a previous block to deblock it with dstBlock */
2971 if(mode & H_X1_FILTER)
2972 RENAME(vertX1Filter)(tempBlock1, 16, &c);
2973 else if(mode & H_DEBLOCK)
2975 if( RENAME(isVertDC)(tempBlock1, 16, &c))
2977 if(RENAME(isVertMinMaxOk)(tempBlock1, 16, &c))
2978 RENAME(doVertLowPass)(tempBlock1, 16, &c);
2981 RENAME(doVertDefFilter)(tempBlock1, 16, &c);
2984 RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
2987 if(mode & H_X1_FILTER)
2988 horizX1Filter(dstBlock-4, stride, QP);
2989 else if(mode & H_DEBLOCK)
2991 if( isHorizDC(dstBlock-4, stride, &c))
2993 if(isHorizMinMaxOk(dstBlock-4, stride, QP))
2994 doHorizLowPass(dstBlock-4, stride, QP);
2997 doHorizDefFilter(dstBlock-4, stride, QP);
3002 //FIXME filter first line
3003 if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);
3006 if(mode & TEMP_NOISE_FILTER)
3008 RENAME(tempNoiseReducer)(dstBlock-8, stride,
3009 c.tempBlured[isColor] + y*dstStride + x,
3010 c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
3011 c.ppMode.maxTmpNoise);
3019 tmpXchg= tempBlock1;
3020 tempBlock1= tempBlock2;
3021 tempBlock2 = tmpXchg;
3027 if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);
3030 if((mode & TEMP_NOISE_FILTER))
3032 RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
3033 c.tempBlured[isColor] + y*dstStride + x,
3034 c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
3035 c.ppMode.maxTmpNoise);
3038 /* did we use a tmp buffer for the last lines*/
3041 uint8_t *dstBlock= &(dst[y*dstStride]);
3042 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
3045 for(x=0; x<width; x+=32)
3048 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
3049 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
3050 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
3051 // + dstBlock[x +13*dstStride]
3052 // + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
3056 asm volatile("femms");
3057 #elif defined (HAVE_MMX)
3058 asm volatile("emms");
3061 #ifdef DEBUG_BRIGHTNESS
3066 for(i=0; i<256; i++)
3067 if(yHistogram[i] > max) max=yHistogram[i];
3069 for(i=1; i<256; i++)
3072 int start=yHistogram[i-1]/(max/256+1);
3073 int end=yHistogram[i]/(max/256+1);
3074 int inc= end > start ? 1 : -1;
3075 for(x=start; x!=end+inc; x+=inc)
3076 dst[ i*dstStride + x]+=128;
3079 for(i=0; i<100; i+=2)
3081 dst[ (white)*dstStride + i]+=128;
3082 dst[ (black)*dstStride + i]+=128;
3088 *c2= c; //copy local context back