2 * Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at)
4 * This file is part of FFmpeg.
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.
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.
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
23 * mmx/mmx2/3dnow postprocess code.
26 #include "libavutil/x86_cpu.h"
34 #define REAL_PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
36 #define REAL_PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
38 #define PAVGB(a,b) REAL_PAVGB(a,b)
41 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
43 #define PMINUB(b,a,t) \
44 "movq " #a ", " #t " \n\t"\
45 "psubusb " #b ", " #t " \n\t"\
46 "psubb " #t ", " #a " \n\t"
50 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
53 "psubusb " #a ", " #b " \n\t"\
54 "paddb " #a ", " #b " \n\t"
57 //FIXME? |255-0| = 1 (should not be a problem ...)
60 * Check if the middle 8x8 Block in the given 8x16 block is flat
62 static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){
64 src+= stride*4; // src points to begin of the 8x8 Block
68 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
72 "lea (%2, %3), %%"REG_a" \n\t"
73 // 0 1 2 3 4 5 6 7 8 9
74 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
76 "movq (%2), %%mm0 \n\t"
77 "movq (%%"REG_a"), %%mm1 \n\t"
78 "movq %%mm0, %%mm3 \n\t"
79 "movq %%mm0, %%mm4 \n\t"
81 PMINUB(%%mm1, %%mm3, %%mm5)
82 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
83 "paddb %%mm7, %%mm0 \n\t"
84 "pcmpgtb %%mm6, %%mm0 \n\t"
86 "movq (%%"REG_a",%3), %%mm2 \n\t"
88 PMINUB(%%mm2, %%mm3, %%mm5)
89 "psubb %%mm2, %%mm1 \n\t"
90 "paddb %%mm7, %%mm1 \n\t"
91 "pcmpgtb %%mm6, %%mm1 \n\t"
92 "paddb %%mm1, %%mm0 \n\t"
94 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
96 PMINUB(%%mm1, %%mm3, %%mm5)
97 "psubb %%mm1, %%mm2 \n\t"
98 "paddb %%mm7, %%mm2 \n\t"
99 "pcmpgtb %%mm6, %%mm2 \n\t"
100 "paddb %%mm2, %%mm0 \n\t"
102 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
104 "movq (%2, %3, 4), %%mm2 \n\t"
106 PMINUB(%%mm2, %%mm3, %%mm5)
107 "psubb %%mm2, %%mm1 \n\t"
108 "paddb %%mm7, %%mm1 \n\t"
109 "pcmpgtb %%mm6, %%mm1 \n\t"
110 "paddb %%mm1, %%mm0 \n\t"
112 "movq (%%"REG_a"), %%mm1 \n\t"
114 PMINUB(%%mm1, %%mm3, %%mm5)
115 "psubb %%mm1, %%mm2 \n\t"
116 "paddb %%mm7, %%mm2 \n\t"
117 "pcmpgtb %%mm6, %%mm2 \n\t"
118 "paddb %%mm2, %%mm0 \n\t"
120 "movq (%%"REG_a", %3), %%mm2 \n\t"
122 PMINUB(%%mm2, %%mm3, %%mm5)
123 "psubb %%mm2, %%mm1 \n\t"
124 "paddb %%mm7, %%mm1 \n\t"
125 "pcmpgtb %%mm6, %%mm1 \n\t"
126 "paddb %%mm1, %%mm0 \n\t"
128 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
130 PMINUB(%%mm1, %%mm3, %%mm5)
131 "psubb %%mm1, %%mm2 \n\t"
132 "paddb %%mm7, %%mm2 \n\t"
133 "pcmpgtb %%mm6, %%mm2 \n\t"
134 "paddb %%mm2, %%mm0 \n\t"
135 "psubusb %%mm3, %%mm4 \n\t"
139 "pxor %%mm7, %%mm7 \n\t"
140 "psadbw %%mm7, %%mm0 \n\t"
142 "movq %%mm0, %%mm1 \n\t"
143 "psrlw $8, %%mm0 \n\t"
144 "paddb %%mm1, %%mm0 \n\t"
145 "movq %%mm0, %%mm1 \n\t"
146 "psrlq $16, %%mm0 \n\t"
147 "paddb %%mm1, %%mm0 \n\t"
148 "movq %%mm0, %%mm1 \n\t"
149 "psrlq $32, %%mm0 \n\t"
150 "paddb %%mm1, %%mm0 \n\t"
152 "movq %4, %%mm7 \n\t" // QP,..., QP
153 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
154 "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0
155 "packssdw %%mm4, %%mm4 \n\t"
156 "movd %%mm0, %0 \n\t"
157 "movd %%mm4, %1 \n\t"
159 : "=r" (numEq), "=r" (dcOk)
160 : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
164 numEq= (-numEq) &0xFF;
165 if(numEq > c->ppMode.flatnessThreshold){
175 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
176 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
179 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
181 #if HAVE_MMX2 || HAVE_AMD3DNOW
183 __asm__ volatile( //"movv %0 %1 %2\n\t"
184 "movq %2, %%mm0 \n\t" // QP,..., QP
185 "pxor %%mm4, %%mm4 \n\t"
187 "movq (%0), %%mm6 \n\t"
188 "movq (%0, %1), %%mm5 \n\t"
189 "movq %%mm5, %%mm1 \n\t"
190 "movq %%mm6, %%mm2 \n\t"
191 "psubusb %%mm6, %%mm5 \n\t"
192 "psubusb %%mm1, %%mm2 \n\t"
193 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
194 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
195 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
197 "pand %%mm2, %%mm6 \n\t"
198 "pandn %%mm1, %%mm2 \n\t"
199 "por %%mm2, %%mm6 \n\t"// First Line to Filter
201 "movq (%0, %1, 8), %%mm5 \n\t"
202 "lea (%0, %1, 4), %%"REG_a" \n\t"
203 "lea (%0, %1, 8), %%"REG_c" \n\t"
204 "sub %1, %%"REG_c" \n\t"
205 "add %1, %0 \n\t" // %0 points to line 1 not 0
206 "movq (%0, %1, 8), %%mm7 \n\t"
207 "movq %%mm5, %%mm1 \n\t"
208 "movq %%mm7, %%mm2 \n\t"
209 "psubusb %%mm7, %%mm5 \n\t"
210 "psubusb %%mm1, %%mm2 \n\t"
211 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
212 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
213 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
215 "pand %%mm2, %%mm7 \n\t"
216 "pandn %%mm1, %%mm2 \n\t"
217 "por %%mm2, %%mm7 \n\t" // First Line to Filter
221 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
226 "movq (%0, %1), %%mm0 \n\t" // 1
227 "movq %%mm0, %%mm1 \n\t" // 1
228 PAVGB(%%mm6, %%mm0) //1 1 /2
229 PAVGB(%%mm6, %%mm0) //3 1 /4
231 "movq (%0, %1, 4), %%mm2 \n\t" // 1
232 "movq %%mm2, %%mm5 \n\t" // 1
233 PAVGB((%%REGa), %%mm2) // 11 /2
234 PAVGB((%0, %1, 2), %%mm2) // 211 /4
235 "movq %%mm2, %%mm3 \n\t" // 211 /4
236 "movq (%0), %%mm4 \n\t" // 1
237 PAVGB(%%mm4, %%mm3) // 4 211 /8
238 PAVGB(%%mm0, %%mm3) //642211 /16
239 "movq %%mm3, (%0) \n\t" // X
240 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
241 "movq %%mm1, %%mm0 \n\t" // 1
242 PAVGB(%%mm6, %%mm0) //1 1 /2
243 "movq %%mm4, %%mm3 \n\t" // 1
244 PAVGB((%0,%1,2), %%mm3) // 1 1 /2
245 PAVGB((%%REGa,%1,2), %%mm5) // 11 /2
246 PAVGB((%%REGa), %%mm5) // 211 /4
247 PAVGB(%%mm5, %%mm3) // 2 2211 /8
248 PAVGB(%%mm0, %%mm3) //4242211 /16
249 "movq %%mm3, (%0,%1) \n\t" // X
250 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
251 PAVGB(%%mm4, %%mm6) //11 /2
252 "movq (%%"REG_c"), %%mm0 \n\t" // 1
253 PAVGB((%%REGa, %1, 2), %%mm0) // 11/2
254 "movq %%mm0, %%mm3 \n\t" // 11/2
255 PAVGB(%%mm1, %%mm0) // 2 11/4
256 PAVGB(%%mm6, %%mm0) //222 11/8
257 PAVGB(%%mm2, %%mm0) //22242211/16
258 "movq (%0, %1, 2), %%mm2 \n\t" // 1
259 "movq %%mm0, (%0, %1, 2) \n\t" // X
260 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
261 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
262 PAVGB((%%REGc), %%mm0) // 11 /2
263 PAVGB(%%mm0, %%mm6) //11 11 /4
264 PAVGB(%%mm1, %%mm4) // 11 /2
265 PAVGB(%%mm2, %%mm1) // 11 /2
266 PAVGB(%%mm1, %%mm6) //1122 11 /8
267 PAVGB(%%mm5, %%mm6) //112242211 /16
268 "movq (%%"REG_a"), %%mm5 \n\t" // 1
269 "movq %%mm6, (%%"REG_a") \n\t" // X
270 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
271 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1
272 PAVGB(%%mm7, %%mm6) // 11 /2
273 PAVGB(%%mm4, %%mm6) // 11 11 /4
274 PAVGB(%%mm3, %%mm6) // 11 2211 /8
275 PAVGB(%%mm5, %%mm2) // 11 /2
276 "movq (%0, %1, 4), %%mm4 \n\t" // 1
277 PAVGB(%%mm4, %%mm2) // 112 /4
278 PAVGB(%%mm2, %%mm6) // 112242211 /16
279 "movq %%mm6, (%0, %1, 4) \n\t" // X
280 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
281 PAVGB(%%mm7, %%mm1) // 11 2 /4
282 PAVGB(%%mm4, %%mm5) // 11 /2
283 PAVGB(%%mm5, %%mm0) // 11 11 /4
284 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1
285 PAVGB(%%mm6, %%mm1) // 11 4 2 /8
286 PAVGB(%%mm0, %%mm1) // 11224222 /16
287 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X
288 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
289 PAVGB((%%REGc), %%mm2) // 112 4 /8
290 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
291 PAVGB(%%mm0, %%mm6) // 1 1 /2
292 PAVGB(%%mm7, %%mm6) // 1 12 /4
293 PAVGB(%%mm2, %%mm6) // 1122424 /4
294 "movq %%mm6, (%%"REG_c") \n\t" // X
295 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
296 PAVGB(%%mm7, %%mm5) // 11 2 /4
297 PAVGB(%%mm7, %%mm5) // 11 6 /8
299 PAVGB(%%mm3, %%mm0) // 112 /4
300 PAVGB(%%mm0, %%mm5) // 112246 /16
301 "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X
305 : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
308 #else //HAVE_MMX2 || HAVE_AMD3DNOW
309 const int l1= stride;
310 const int l2= stride + l1;
311 const int l3= stride + l2;
312 const int l4= stride + l3;
313 const int l5= stride + l4;
314 const int l6= stride + l5;
315 const int l7= stride + l6;
316 const int l8= stride + l7;
317 const int l9= stride + l8;
320 for(x=0; x<BLOCK_SIZE; x++){
321 const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
322 const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
325 sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
326 sums[1] = sums[0] - first + src[l4];
327 sums[2] = sums[1] - first + src[l5];
328 sums[3] = sums[2] - first + src[l6];
329 sums[4] = sums[3] - first + src[l7];
330 sums[5] = sums[4] - src[l1] + src[l8];
331 sums[6] = sums[5] - src[l2] + last;
332 sums[7] = sums[6] - src[l3] + last;
333 sums[8] = sums[7] - src[l4] + last;
334 sums[9] = sums[8] - src[l5] + last;
336 src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
337 src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
338 src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
339 src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
340 src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
341 src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
342 src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
343 src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
347 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
349 #endif //HAVE_ALTIVEC
352 * Experimental Filter 1
353 * will not damage linear gradients
354 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
355 * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
356 * MMX2 version does correct clipping C version does not
358 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
360 #if HAVE_MMX2 || HAVE_AMD3DNOW
364 "pxor %%mm7, %%mm7 \n\t" // 0
365 "lea (%0, %1), %%"REG_a" \n\t"
366 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
367 // 0 1 2 3 4 5 6 7 8 9
368 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
369 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
370 "movq (%0, %1, 4), %%mm1 \n\t" // line 4
371 "movq %%mm1, %%mm2 \n\t" // line 4
372 "psubusb %%mm0, %%mm1 \n\t"
373 "psubusb %%mm2, %%mm0 \n\t"
374 "por %%mm1, %%mm0 \n\t" // |l2 - l3|
375 "movq (%%"REG_c"), %%mm3 \n\t" // line 5
376 "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6
377 "movq %%mm3, %%mm5 \n\t" // line 5
378 "psubusb %%mm4, %%mm3 \n\t"
379 "psubusb %%mm5, %%mm4 \n\t"
380 "por %%mm4, %%mm3 \n\t" // |l5 - l6|
381 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
382 "movq %%mm2, %%mm1 \n\t" // line 4
383 "psubusb %%mm5, %%mm2 \n\t"
384 "movq %%mm2, %%mm4 \n\t"
385 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
386 "psubusb %%mm1, %%mm5 \n\t"
387 "por %%mm5, %%mm4 \n\t" // |l4 - l5|
388 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
389 "movq %%mm4, %%mm3 \n\t" // d
390 "movq %2, %%mm0 \n\t"
391 "paddusb %%mm0, %%mm0 \n\t"
392 "psubusb %%mm0, %%mm4 \n\t"
393 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
394 "psubusb "MANGLE(b01)", %%mm3 \n\t"
395 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
397 PAVGB(%%mm7, %%mm3) // d/2
398 "movq %%mm3, %%mm1 \n\t" // d/2
399 PAVGB(%%mm7, %%mm3) // d/4
400 PAVGB(%%mm1, %%mm3) // 3*d/8
402 "movq (%0, %1, 4), %%mm0 \n\t" // line 4
403 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
404 "psubusb %%mm3, %%mm0 \n\t"
405 "pxor %%mm2, %%mm0 \n\t"
406 "movq %%mm0, (%0, %1, 4) \n\t" // line 4
408 "movq (%%"REG_c"), %%mm0 \n\t" // line 5
409 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
410 "paddusb %%mm3, %%mm0 \n\t"
411 "pxor %%mm2, %%mm0 \n\t"
412 "movq %%mm0, (%%"REG_c") \n\t" // line 5
414 PAVGB(%%mm7, %%mm1) // d/4
416 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
417 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
418 "psubusb %%mm1, %%mm0 \n\t"
419 "pxor %%mm2, %%mm0 \n\t"
420 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3
422 "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6
423 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
424 "paddusb %%mm1, %%mm0 \n\t"
425 "pxor %%mm2, %%mm0 \n\t"
426 "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6
428 PAVGB(%%mm7, %%mm1) // d/8
430 "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2
431 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
432 "psubusb %%mm1, %%mm0 \n\t"
433 "pxor %%mm2, %%mm0 \n\t"
434 "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2
436 "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7
437 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
438 "paddusb %%mm1, %%mm0 \n\t"
439 "pxor %%mm2, %%mm0 \n\t"
440 "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7
443 : "r" (src), "r" ((x86_reg)stride), "m" (co->pQPb)
446 #else //HAVE_MMX2 || HAVE_AMD3DNOW
448 const int l1= stride;
449 const int l2= stride + l1;
450 const int l3= stride + l2;
451 const int l4= stride + l3;
452 const int l5= stride + l4;
453 const int l6= stride + l5;
454 const int l7= stride + l6;
455 // const int l8= stride + l7;
456 // const int l9= stride + l8;
460 for(x=0; x<BLOCK_SIZE; x++){
461 int a= src[l3] - src[l4];
462 int b= src[l4] - src[l5];
463 int c= src[l5] - src[l6];
465 int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
469 int v = d * FFSIGN(-b);
480 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
484 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
486 #if HAVE_MMX2 || HAVE_AMD3DNOW
489 const int l1= stride;
490 const int l2= stride + l1;
491 const int l3= stride + l2;
492 const int l4= (int)tmp - (int)src - stride*3;
493 const int l5= (int)tmp - (int)src - stride*3 + 8;
494 const int l6= stride*3 + l3;
495 const int l7= stride + l6;
496 const int l8= stride + l7;
498 memcpy(tmp, src+stride*7, 8);
499 memcpy(tmp+8, src+stride*8, 8);
504 #if 0 //slightly more accurate and slightly slower
505 "pxor %%mm7, %%mm7 \n\t" // 0
506 "lea (%0, %1), %%"REG_a" \n\t"
507 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
509 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
510 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
513 "movq (%0, %1, 2), %%mm0 \n\t" // l2
514 "movq (%0), %%mm1 \n\t" // l0
515 "movq %%mm0, %%mm2 \n\t" // l2
516 PAVGB(%%mm7, %%mm0) // ~l2/2
517 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
518 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
520 "movq (%%"REG_a"), %%mm1 \n\t" // l1
521 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3
522 "movq %%mm1, %%mm4 \n\t" // l1
523 PAVGB(%%mm7, %%mm1) // ~l1/2
524 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
525 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
527 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
528 "psubusb %%mm1, %%mm0 \n\t"
529 "psubusb %%mm4, %%mm1 \n\t"
530 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
531 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
533 "movq (%0, %1, 4), %%mm0 \n\t" // l4
534 "movq %%mm0, %%mm4 \n\t" // l4
535 PAVGB(%%mm7, %%mm0) // ~l4/2
536 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
537 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
539 "movq (%%"REG_c"), %%mm2 \n\t" // l5
540 "movq %%mm3, %%mm5 \n\t" // l3
541 PAVGB(%%mm7, %%mm3) // ~l3/2
542 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
543 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
545 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
546 "psubusb %%mm3, %%mm0 \n\t"
547 "psubusb %%mm6, %%mm3 \n\t"
548 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
549 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
550 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
552 "movq (%%"REG_c", %1), %%mm6 \n\t" // l6
553 "movq %%mm6, %%mm5 \n\t" // l6
554 PAVGB(%%mm7, %%mm6) // ~l6/2
555 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
556 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
558 "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7
559 "movq %%mm2, %%mm4 \n\t" // l5
560 PAVGB(%%mm7, %%mm2) // ~l5/2
561 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
562 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
564 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
565 "psubusb %%mm2, %%mm6 \n\t"
566 "psubusb %%mm4, %%mm2 \n\t"
567 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
568 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
571 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
572 "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
573 "paddusb "MANGLE(b01)", %%mm4 \n\t"
574 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
575 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
576 "pand %%mm4, %%mm3 \n\t"
578 "movq %%mm3, %%mm1 \n\t"
579 // "psubusb "MANGLE(b01)", %%mm3 \n\t"
582 "paddusb %%mm1, %%mm3 \n\t"
583 // "paddusb "MANGLE(b01)", %%mm3 \n\t"
585 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3
586 "movq (%0, %1, 4), %%mm5 \n\t" //l4
587 "movq (%0, %1, 4), %%mm4 \n\t" //l4
588 "psubusb %%mm6, %%mm5 \n\t"
589 "psubusb %%mm4, %%mm6 \n\t"
590 "por %%mm6, %%mm5 \n\t" // |l3-l4|
591 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
592 "pxor %%mm6, %%mm0 \n\t"
593 "pand %%mm0, %%mm3 \n\t"
594 PMINUB(%%mm5, %%mm3, %%mm0)
596 "psubusb "MANGLE(b01)", %%mm3 \n\t"
599 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
600 "movq (%0, %1, 4), %%mm2 \n\t"
601 "pxor %%mm6, %%mm0 \n\t"
602 "pxor %%mm6, %%mm2 \n\t"
603 "psubb %%mm3, %%mm0 \n\t"
604 "paddb %%mm3, %%mm2 \n\t"
605 "pxor %%mm6, %%mm0 \n\t"
606 "pxor %%mm6, %%mm2 \n\t"
607 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
608 "movq %%mm2, (%0, %1, 4) \n\t"
611 "lea (%0, %1), %%"REG_a" \n\t"
612 "pcmpeqb %%mm6, %%mm6 \n\t" // -1
614 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
615 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
618 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3
619 "movq (%0, %1, 4), %%mm0 \n\t" // l4
620 "pxor %%mm6, %%mm1 \n\t" // -l3-1
621 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
622 // mm1=-l3-1, mm0=128-q
624 "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5
625 "movq (%%"REG_a", %1), %%mm3 \n\t" // l2
626 "pxor %%mm6, %%mm2 \n\t" // -l5-1
627 "movq %%mm2, %%mm5 \n\t" // -l5-1
628 "movq "MANGLE(b80)", %%mm4 \n\t" // 128
629 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
630 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
631 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
632 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
633 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
634 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
636 "movq (%%"REG_a"), %%mm2 \n\t" // l1
637 "pxor %%mm6, %%mm2 \n\t" // -l1-1
638 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
639 PAVGB((%0), %%mm1) // (l0-l3+256)/2
640 "movq "MANGLE(b80)", %%mm3 \n\t" // 128
641 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
642 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
643 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
644 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
646 PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2
647 "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7
648 "pxor %%mm6, %%mm1 \n\t" // -l7-1
649 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
650 "movq "MANGLE(b80)", %%mm2 \n\t" // 128
651 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
652 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
653 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
654 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
656 "movq "MANGLE(b00)", %%mm1 \n\t" // 0
657 "movq "MANGLE(b00)", %%mm5 \n\t" // 0
658 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
659 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
660 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
661 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
662 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
664 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
666 "movq "MANGLE(b00)", %%mm7 \n\t" // 0
667 "movq %2, %%mm2 \n\t" // QP
668 PAVGB(%%mm6, %%mm2) // 128 + QP/2
669 "psubb %%mm6, %%mm2 \n\t"
671 "movq %%mm4, %%mm1 \n\t"
672 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
673 "pxor %%mm1, %%mm4 \n\t"
674 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
675 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
676 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
677 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
679 "movq %%mm4, %%mm3 \n\t" // d
680 "psubusb "MANGLE(b01)", %%mm4 \n\t"
681 PAVGB(%%mm7, %%mm4) // d/32
682 PAVGB(%%mm7, %%mm4) // (d + 32)/64
683 "paddb %%mm3, %%mm4 \n\t" // 5d/64
684 "pand %%mm2, %%mm4 \n\t"
686 "movq "MANGLE(b80)", %%mm5 \n\t" // 128
687 "psubb %%mm0, %%mm5 \n\t" // q
688 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
689 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
690 "pxor %%mm7, %%mm5 \n\t"
692 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
693 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
695 "pand %%mm7, %%mm4 \n\t"
696 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
697 "movq (%0, %1, 4), %%mm2 \n\t"
698 "pxor %%mm1, %%mm0 \n\t"
699 "pxor %%mm1, %%mm2 \n\t"
700 "paddb %%mm4, %%mm0 \n\t"
701 "psubb %%mm4, %%mm2 \n\t"
702 "pxor %%mm1, %%mm0 \n\t"
703 "pxor %%mm1, %%mm2 \n\t"
704 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
705 "movq %%mm2, (%0, %1, 4) \n\t"
708 : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
716 for(x=0; x<BLOCK_SIZE; x++){
717 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
718 if(FFABS(middleEnergy)< 8*QP){
719 const int q=(src[l4] - src[l5])/2;
720 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
721 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
723 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
727 d*= FFSIGN(-middleEnergy);
746 int d= src[x+y*stride] - tmp[x+(y-4)*8];
756 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
760 if(num%1000000 == 0){
761 av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias);
768 DECLARE_ALIGNED(8, uint64_t, tmp)[4]; // make space for 4 8-byte vars
771 "pxor %%mm7, %%mm7 \n\t"
773 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
774 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
776 "movq (%0), %%mm0 \n\t"
777 "movq %%mm0, %%mm1 \n\t"
778 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
779 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
781 "movq (%0, %1), %%mm2 \n\t"
782 "lea (%0, %1, 2), %%"REG_a" \n\t"
783 "movq %%mm2, %%mm3 \n\t"
784 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
785 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
787 "movq (%%"REG_a"), %%mm4 \n\t"
788 "movq %%mm4, %%mm5 \n\t"
789 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
790 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
792 "paddw %%mm0, %%mm0 \n\t" // 2L0
793 "paddw %%mm1, %%mm1 \n\t" // 2H0
794 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
795 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
796 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
797 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
799 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
800 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
801 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
802 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
804 "movq (%%"REG_a", %1), %%mm2 \n\t"
805 "movq %%mm2, %%mm3 \n\t"
806 "punpcklbw %%mm7, %%mm2 \n\t" // L3
807 "punpckhbw %%mm7, %%mm3 \n\t" // H3
809 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
810 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
811 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
812 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
813 "movq %%mm0, (%3) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
814 "movq %%mm1, 8(%3) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
816 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
817 "movq %%mm0, %%mm1 \n\t"
818 "punpcklbw %%mm7, %%mm0 \n\t" // L4
819 "punpckhbw %%mm7, %%mm1 \n\t" // H4
821 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
822 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
823 "movq %%mm2, 16(%3) \n\t" // L3 - L4
824 "movq %%mm3, 24(%3) \n\t" // H3 - H4
825 "paddw %%mm4, %%mm4 \n\t" // 2L2
826 "paddw %%mm5, %%mm5 \n\t" // 2H2
827 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
828 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
830 "lea (%%"REG_a", %1), %0 \n\t"
831 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
832 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
833 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
834 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
836 "movq (%0, %1, 2), %%mm2 \n\t"
837 "movq %%mm2, %%mm3 \n\t"
838 "punpcklbw %%mm7, %%mm2 \n\t" // L5
839 "punpckhbw %%mm7, %%mm3 \n\t" // H5
840 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
841 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
842 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
843 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
845 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
846 "punpcklbw %%mm7, %%mm6 \n\t" // L6
847 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
848 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
849 "punpckhbw %%mm7, %%mm6 \n\t" // H6
850 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
852 "paddw %%mm0, %%mm0 \n\t" // 2L4
853 "paddw %%mm1, %%mm1 \n\t" // 2H4
854 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
855 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
857 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
858 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
859 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
860 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
862 "movq (%0, %1, 4), %%mm2 \n\t"
863 "movq %%mm2, %%mm3 \n\t"
864 "punpcklbw %%mm7, %%mm2 \n\t" // L7
865 "punpckhbw %%mm7, %%mm3 \n\t" // H7
867 "paddw %%mm2, %%mm2 \n\t" // 2L7
868 "paddw %%mm3, %%mm3 \n\t" // 2H7
869 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
870 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
872 "movq (%3), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
873 "movq 8(%3), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
876 "movq %%mm7, %%mm6 \n\t" // 0
877 "psubw %%mm0, %%mm6 \n\t"
878 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
879 "movq %%mm7, %%mm6 \n\t" // 0
880 "psubw %%mm1, %%mm6 \n\t"
881 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
882 "movq %%mm7, %%mm6 \n\t" // 0
883 "psubw %%mm2, %%mm6 \n\t"
884 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
885 "movq %%mm7, %%mm6 \n\t" // 0
886 "psubw %%mm3, %%mm6 \n\t"
887 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
889 "movq %%mm7, %%mm6 \n\t" // 0
890 "pcmpgtw %%mm0, %%mm6 \n\t"
891 "pxor %%mm6, %%mm0 \n\t"
892 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
893 "movq %%mm7, %%mm6 \n\t" // 0
894 "pcmpgtw %%mm1, %%mm6 \n\t"
895 "pxor %%mm6, %%mm1 \n\t"
896 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
897 "movq %%mm7, %%mm6 \n\t" // 0
898 "pcmpgtw %%mm2, %%mm6 \n\t"
899 "pxor %%mm6, %%mm2 \n\t"
900 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
901 "movq %%mm7, %%mm6 \n\t" // 0
902 "pcmpgtw %%mm3, %%mm6 \n\t"
903 "pxor %%mm6, %%mm3 \n\t"
904 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
908 "pminsw %%mm2, %%mm0 \n\t"
909 "pminsw %%mm3, %%mm1 \n\t"
911 "movq %%mm0, %%mm6 \n\t"
912 "psubusw %%mm2, %%mm6 \n\t"
913 "psubw %%mm6, %%mm0 \n\t"
914 "movq %%mm1, %%mm6 \n\t"
915 "psubusw %%mm3, %%mm6 \n\t"
916 "psubw %%mm6, %%mm1 \n\t"
919 "movd %2, %%mm2 \n\t" // QP
920 "punpcklbw %%mm7, %%mm2 \n\t"
922 "movq %%mm7, %%mm6 \n\t" // 0
923 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
924 "pxor %%mm6, %%mm4 \n\t"
925 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
926 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
927 "pxor %%mm7, %%mm5 \n\t"
928 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
930 "psllw $3, %%mm2 \n\t" // 8QP
931 "movq %%mm2, %%mm3 \n\t" // 8QP
932 "pcmpgtw %%mm4, %%mm2 \n\t"
933 "pcmpgtw %%mm5, %%mm3 \n\t"
934 "pand %%mm2, %%mm4 \n\t"
935 "pand %%mm3, %%mm5 \n\t"
938 "psubusw %%mm0, %%mm4 \n\t" // hd
939 "psubusw %%mm1, %%mm5 \n\t" // ld
942 "movq "MANGLE(w05)", %%mm2 \n\t" // 5
943 "pmullw %%mm2, %%mm4 \n\t"
944 "pmullw %%mm2, %%mm5 \n\t"
945 "movq "MANGLE(w20)", %%mm2 \n\t" // 32
946 "paddw %%mm2, %%mm4 \n\t"
947 "paddw %%mm2, %%mm5 \n\t"
948 "psrlw $6, %%mm4 \n\t"
949 "psrlw $6, %%mm5 \n\t"
951 "movq 16(%3), %%mm0 \n\t" // L3 - L4
952 "movq 24(%3), %%mm1 \n\t" // H3 - H4
954 "pxor %%mm2, %%mm2 \n\t"
955 "pxor %%mm3, %%mm3 \n\t"
957 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
958 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
959 "pxor %%mm2, %%mm0 \n\t"
960 "pxor %%mm3, %%mm1 \n\t"
961 "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
962 "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
963 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
964 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
966 "pxor %%mm6, %%mm2 \n\t"
967 "pxor %%mm7, %%mm3 \n\t"
968 "pand %%mm2, %%mm4 \n\t"
969 "pand %%mm3, %%mm5 \n\t"
972 "pminsw %%mm0, %%mm4 \n\t"
973 "pminsw %%mm1, %%mm5 \n\t"
975 "movq %%mm4, %%mm2 \n\t"
976 "psubusw %%mm0, %%mm2 \n\t"
977 "psubw %%mm2, %%mm4 \n\t"
978 "movq %%mm5, %%mm2 \n\t"
979 "psubusw %%mm1, %%mm2 \n\t"
980 "psubw %%mm2, %%mm5 \n\t"
982 "pxor %%mm6, %%mm4 \n\t"
983 "pxor %%mm7, %%mm5 \n\t"
984 "psubw %%mm6, %%mm4 \n\t"
985 "psubw %%mm7, %%mm5 \n\t"
986 "packsswb %%mm5, %%mm4 \n\t"
987 "movq (%0), %%mm0 \n\t"
988 "paddb %%mm4, %%mm0 \n\t"
989 "movq %%mm0, (%0) \n\t"
990 "movq (%0, %1), %%mm0 \n\t"
991 "psubb %%mm4, %%mm0 \n\t"
992 "movq %%mm0, (%0, %1) \n\t"
995 : "r" ((x86_reg)stride), "m" (c->pQPb), "r"(tmp)
998 #else //HAVE_MMX2 || HAVE_AMD3DNOW
999 const int l1= stride;
1000 const int l2= stride + l1;
1001 const int l3= stride + l2;
1002 const int l4= stride + l3;
1003 const int l5= stride + l4;
1004 const int l6= stride + l5;
1005 const int l7= stride + l6;
1006 const int l8= stride + l7;
1007 // const int l9= stride + l8;
1010 for(x=0; x<BLOCK_SIZE; x++){
1011 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1012 if(FFABS(middleEnergy) < 8*c->QP){
1013 const int q=(src[l4] - src[l5])/2;
1014 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1015 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1017 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
1021 d*= FFSIGN(-middleEnergy);
1036 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
1038 #endif //HAVE_ALTIVEC
1041 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
1043 #if HAVE_MMX2 || HAVE_AMD3DNOW
1044 DECLARE_ALIGNED(8, uint64_t, tmp)[3];
1046 "pxor %%mm6, %%mm6 \n\t"
1047 "pcmpeqb %%mm7, %%mm7 \n\t"
1048 "movq %2, %%mm0 \n\t"
1049 "punpcklbw %%mm6, %%mm0 \n\t"
1050 "psrlw $1, %%mm0 \n\t"
1051 "psubw %%mm7, %%mm0 \n\t"
1052 "packuswb %%mm0, %%mm0 \n\t"
1053 "movq %%mm0, %3 \n\t"
1055 "lea (%0, %1), %%"REG_a" \n\t"
1056 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1058 // 0 1 2 3 4 5 6 7 8 9
1059 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1061 #undef REAL_FIND_MIN_MAX
1064 #define REAL_FIND_MIN_MAX(addr)\
1065 "movq " #addr ", %%mm0 \n\t"\
1066 "pminub %%mm0, %%mm7 \n\t"\
1067 "pmaxub %%mm0, %%mm6 \n\t"
1069 #define REAL_FIND_MIN_MAX(addr)\
1070 "movq " #addr ", %%mm0 \n\t"\
1071 "movq %%mm7, %%mm1 \n\t"\
1072 "psubusb %%mm0, %%mm6 \n\t"\
1073 "paddb %%mm0, %%mm6 \n\t"\
1074 "psubusb %%mm0, %%mm1 \n\t"\
1075 "psubb %%mm1, %%mm7 \n\t"
1077 #define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr)
1079 FIND_MIN_MAX((%%REGa))
1080 FIND_MIN_MAX((%%REGa, %1))
1081 FIND_MIN_MAX((%%REGa, %1, 2))
1082 FIND_MIN_MAX((%0, %1, 4))
1083 FIND_MIN_MAX((%%REGd))
1084 FIND_MIN_MAX((%%REGd, %1))
1085 FIND_MIN_MAX((%%REGd, %1, 2))
1086 FIND_MIN_MAX((%0, %1, 8))
1088 "movq %%mm7, %%mm4 \n\t"
1089 "psrlq $8, %%mm7 \n\t"
1091 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1092 "pshufw $0xF9, %%mm7, %%mm4 \n\t"
1093 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1094 "pshufw $0xFE, %%mm7, %%mm4 \n\t"
1095 "pminub %%mm4, %%mm7 \n\t"
1097 "movq %%mm7, %%mm1 \n\t"
1098 "psubusb %%mm4, %%mm1 \n\t"
1099 "psubb %%mm1, %%mm7 \n\t"
1100 "movq %%mm7, %%mm4 \n\t"
1101 "psrlq $16, %%mm7 \n\t"
1102 "movq %%mm7, %%mm1 \n\t"
1103 "psubusb %%mm4, %%mm1 \n\t"
1104 "psubb %%mm1, %%mm7 \n\t"
1105 "movq %%mm7, %%mm4 \n\t"
1106 "psrlq $32, %%mm7 \n\t"
1107 "movq %%mm7, %%mm1 \n\t"
1108 "psubusb %%mm4, %%mm1 \n\t"
1109 "psubb %%mm1, %%mm7 \n\t"
1113 "movq %%mm6, %%mm4 \n\t"
1114 "psrlq $8, %%mm6 \n\t"
1116 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1117 "pshufw $0xF9, %%mm6, %%mm4 \n\t"
1118 "pmaxub %%mm4, %%mm6 \n\t"
1119 "pshufw $0xFE, %%mm6, %%mm4 \n\t"
1120 "pmaxub %%mm4, %%mm6 \n\t"
1122 "psubusb %%mm4, %%mm6 \n\t"
1123 "paddb %%mm4, %%mm6 \n\t"
1124 "movq %%mm6, %%mm4 \n\t"
1125 "psrlq $16, %%mm6 \n\t"
1126 "psubusb %%mm4, %%mm6 \n\t"
1127 "paddb %%mm4, %%mm6 \n\t"
1128 "movq %%mm6, %%mm4 \n\t"
1129 "psrlq $32, %%mm6 \n\t"
1130 "psubusb %%mm4, %%mm6 \n\t"
1131 "paddb %%mm4, %%mm6 \n\t"
1133 "movq %%mm6, %%mm0 \n\t" // max
1134 "psubb %%mm7, %%mm6 \n\t" // max - min
1136 "movd %%mm6, %k4 \n\t"
1137 "cmpb "MANGLE(deringThreshold)", %b4 \n\t"
1140 PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1141 "punpcklbw %%mm7, %%mm7 \n\t"
1142 "punpcklbw %%mm7, %%mm7 \n\t"
1143 "punpcklbw %%mm7, %%mm7 \n\t"
1144 "movq %%mm7, (%4) \n\t"
1146 "movq (%0), %%mm0 \n\t" // L10
1147 "movq %%mm0, %%mm1 \n\t" // L10
1148 "movq %%mm0, %%mm2 \n\t" // L10
1149 "psllq $8, %%mm1 \n\t"
1150 "psrlq $8, %%mm2 \n\t"
1151 "movd -4(%0), %%mm3 \n\t"
1152 "movd 8(%0), %%mm4 \n\t"
1153 "psrlq $24, %%mm3 \n\t"
1154 "psllq $56, %%mm4 \n\t"
1155 "por %%mm3, %%mm1 \n\t" // L00
1156 "por %%mm4, %%mm2 \n\t" // L20
1157 "movq %%mm1, %%mm3 \n\t" // L00
1158 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1159 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1160 "psubusb %%mm7, %%mm0 \n\t"
1161 "psubusb %%mm7, %%mm2 \n\t"
1162 "psubusb %%mm7, %%mm3 \n\t"
1163 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1164 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1165 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1166 "paddb %%mm2, %%mm0 \n\t"
1167 "paddb %%mm3, %%mm0 \n\t"
1169 "movq (%%"REG_a"), %%mm2 \n\t" // L11
1170 "movq %%mm2, %%mm3 \n\t" // L11
1171 "movq %%mm2, %%mm4 \n\t" // L11
1172 "psllq $8, %%mm3 \n\t"
1173 "psrlq $8, %%mm4 \n\t"
1174 "movd -4(%%"REG_a"), %%mm5 \n\t"
1175 "movd 8(%%"REG_a"), %%mm6 \n\t"
1176 "psrlq $24, %%mm5 \n\t"
1177 "psllq $56, %%mm6 \n\t"
1178 "por %%mm5, %%mm3 \n\t" // L01
1179 "por %%mm6, %%mm4 \n\t" // L21
1180 "movq %%mm3, %%mm5 \n\t" // L01
1181 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1182 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1183 "psubusb %%mm7, %%mm2 \n\t"
1184 "psubusb %%mm7, %%mm4 \n\t"
1185 "psubusb %%mm7, %%mm5 \n\t"
1186 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1187 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1188 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1189 "paddb %%mm4, %%mm2 \n\t"
1190 "paddb %%mm5, %%mm2 \n\t"
1192 #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1193 "movq " #src ", " #sx " \n\t" /* src[0] */\
1194 "movq " #sx ", " #lx " \n\t" /* src[0] */\
1195 "movq " #sx ", " #t0 " \n\t" /* src[0] */\
1196 "psllq $8, " #lx " \n\t"\
1197 "psrlq $8, " #t0 " \n\t"\
1198 "movd -4" #src ", " #t1 " \n\t"\
1199 "psrlq $24, " #t1 " \n\t"\
1200 "por " #t1 ", " #lx " \n\t" /* src[-1] */\
1201 "movd 8" #src ", " #t1 " \n\t"\
1202 "psllq $56, " #t1 " \n\t"\
1203 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1204 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1205 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1206 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1208 "movq " #lx ", 8(%4) \n\t"\
1209 "movq (%4), " #lx " \n\t"\
1210 "psubusb " #lx ", " #t1 " \n\t"\
1211 "psubusb " #lx ", " #t0 " \n\t"\
1212 "psubusb " #lx ", " #sx " \n\t"\
1213 "movq "MANGLE(b00)", " #lx " \n\t"\
1214 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1215 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1216 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1217 "paddb " #t1 ", " #t0 " \n\t"\
1218 "paddb " #t0 ", " #sx " \n\t"\
1220 PAVGB(plx, pplx) /* filtered */\
1221 "movq " #dst ", " #t0 " \n\t" /* dst */\
1222 "movq " #t0 ", " #t1 " \n\t" /* dst */\
1223 "psubusb %3, " #t0 " \n\t"\
1224 "paddusb %3, " #t1 " \n\t"\
1226 PMINUB(t1, pplx, t0)\
1227 "paddb " #sx ", " #ppsx " \n\t"\
1228 "paddb " #psx ", " #ppsx " \n\t"\
1229 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1230 "pand "MANGLE(b08)", " #ppsx " \n\t"\
1231 "pcmpeqb " #lx ", " #ppsx " \n\t"\
1232 "pand " #ppsx ", " #pplx " \n\t"\
1233 "pandn " #dst ", " #ppsx " \n\t"\
1234 "por " #pplx ", " #ppsx " \n\t"\
1235 "movq " #ppsx ", " #dst " \n\t"\
1236 "movq 8(%4), " #lx " \n\t"
1238 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1239 REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1)
1255 //DERING_CORE(dst ,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
1256 DERING_CORE((%%REGa) ,(%%REGa, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1257 DERING_CORE((%%REGa, %1) ,(%%REGa, %1, 2),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1258 DERING_CORE((%%REGa, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1259 DERING_CORE((%0, %1, 4) ,(%%REGd) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1260 DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1261 DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1262 DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1263 DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1266 : : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb), "m"(c->pQPb2), "q"(tmp)
1267 : "%"REG_a, "%"REG_d
1269 #else //HAVE_MMX2 || HAVE_AMD3DNOW
1276 const int QP2= c->QP/2 + 1;
1283 if(*p > max) max= *p;
1284 if(*p < min) min= *p;
1287 avg= (min + max + 1)>>1;
1289 if(max - min <deringThreshold) return;
1291 for(y=0; y<10; y++){
1294 if(src[stride*y + 0] > avg) t+= 1;
1295 if(src[stride*y + 1] > avg) t+= 2;
1296 if(src[stride*y + 2] > avg) t+= 4;
1297 if(src[stride*y + 3] > avg) t+= 8;
1298 if(src[stride*y + 4] > avg) t+= 16;
1299 if(src[stride*y + 5] > avg) t+= 32;
1300 if(src[stride*y + 6] > avg) t+= 64;
1301 if(src[stride*y + 7] > avg) t+= 128;
1302 if(src[stride*y + 8] > avg) t+= 256;
1303 if(src[stride*y + 9] > avg) t+= 512;
1306 t &= (t<<1) & (t>>1);
1311 int t = s[y-1] & s[y] & s[y+1];
1324 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1325 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
1326 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1329 #ifdef DEBUG_DERING_THRESHOLD
1330 __asm__ volatile("emms\n\t":);
1332 static long long numPixels=0;
1333 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1334 // if((max-min)<20 || (max-min)*QP<200)
1335 // if((max-min)*QP < 500)
1338 static int numSkipped=0;
1339 static int errorSum=0;
1340 static int worstQP=0;
1341 static int worstRange=0;
1342 static int worstDiff=0;
1344 int absDiff= FFABS(diff);
1345 int error= diff*diff;
1347 if(x==1 || x==8 || y==1 || y==8) continue;
1350 if(absDiff > worstDiff){
1353 worstRange= max-min;
1357 if(1024LL*1024LL*1024LL % numSkipped == 0){
1358 av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, "
1359 "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1360 (float)errorSum/numSkipped, numSkipped, worstQP, worstRange,
1361 worstDiff, (float)numSkipped/numPixels);
1366 if (*p + QP2 < f) *p= *p + QP2;
1367 else if(*p - QP2 > f) *p= *p - QP2;
1372 #ifdef DEBUG_DERING_THRESHOLD
1380 *p = FFMIN(*p + 20, 255);
1383 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1386 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
1388 #endif //HAVE_ALTIVEC
1391 * Deinterlace the given block by linearly interpolating every second line.
1392 * will be called for every 8x8 block and can read & write from line 4-15
1393 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1394 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1396 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1398 #if HAVE_MMX2 || HAVE_AMD3DNOW
1401 "lea (%0, %1), %%"REG_a" \n\t"
1402 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
1403 // 0 1 2 3 4 5 6 7 8 9
1404 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
1406 "movq (%0), %%mm0 \n\t"
1407 "movq (%%"REG_a", %1), %%mm1 \n\t"
1409 "movq %%mm0, (%%"REG_a") \n\t"
1410 "movq (%0, %1, 4), %%mm0 \n\t"
1412 "movq %%mm1, (%%"REG_a", %1, 2) \n\t"
1413 "movq (%%"REG_c", %1), %%mm1 \n\t"
1415 "movq %%mm0, (%%"REG_c") \n\t"
1416 "movq (%0, %1, 8), %%mm0 \n\t"
1418 "movq %%mm1, (%%"REG_c", %1, 2) \n\t"
1420 : : "r" (src), "r" ((x86_reg)stride)
1421 : "%"REG_a, "%"REG_c
1428 a= *(uint32_t*)&src[stride*0];
1429 b= *(uint32_t*)&src[stride*2];
1430 *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1431 a= *(uint32_t*)&src[stride*4];
1432 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1433 b= *(uint32_t*)&src[stride*6];
1434 *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1435 a= *(uint32_t*)&src[stride*8];
1436 *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1443 * Deinterlace the given block by cubic interpolating every second line.
1444 * will be called for every 8x8 block and can read & write from line 4-15
1445 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1446 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1447 * this filter will read lines 3-15 and write 7-13
1449 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1451 #if HAVE_MMX2 || HAVE_AMD3DNOW
1454 "lea (%0, %1), %%"REG_a" \n\t"
1455 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1456 "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t"
1457 "add %1, %%"REG_c" \n\t"
1458 "pxor %%mm7, %%mm7 \n\t"
1459 // 0 1 2 3 4 5 6 7 8 9 10
1460 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1462 #define REAL_DEINT_CUBIC(a,b,c,d,e)\
1463 "movq " #a ", %%mm0 \n\t"\
1464 "movq " #b ", %%mm1 \n\t"\
1465 "movq " #d ", %%mm2 \n\t"\
1466 "movq " #e ", %%mm3 \n\t"\
1467 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\
1468 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\
1469 "movq %%mm0, %%mm2 \n\t"\
1470 "punpcklbw %%mm7, %%mm0 \n\t"\
1471 "punpckhbw %%mm7, %%mm2 \n\t"\
1472 "movq %%mm1, %%mm3 \n\t"\
1473 "punpcklbw %%mm7, %%mm1 \n\t"\
1474 "punpckhbw %%mm7, %%mm3 \n\t"\
1475 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\
1476 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\
1477 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\
1478 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\
1479 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\
1480 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\
1481 "packuswb %%mm3, %%mm1 \n\t"\
1482 "movq %%mm1, " #c " \n\t"
1483 #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e)
1485 DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1))
1486 DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8))
1487 DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc))
1488 DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2))
1490 : : "r" (src), "r" ((x86_reg)stride)
1491 : "%"REG_a, "%"REG_d, "%"REG_c
1493 #else //HAVE_MMX2 || HAVE_AMD3DNOW
1497 src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4);
1498 src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4);
1499 src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4);
1500 src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4);
1503 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
1507 * Deinterlace the given block by filtering every second line with a (-1 4 2 4 -1) filter.
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 already, but can be read, too.
1510 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1511 * this filter will read lines 4-13 and write 5-11
1513 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
1515 #if HAVE_MMX2 || HAVE_AMD3DNOW
1518 "lea (%0, %1), %%"REG_a" \n\t"
1519 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1520 "pxor %%mm7, %%mm7 \n\t"
1521 "movq (%2), %%mm0 \n\t"
1522 // 0 1 2 3 4 5 6 7 8 9 10
1523 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1525 #define REAL_DEINT_FF(a,b,c,d)\
1526 "movq " #a ", %%mm1 \n\t"\
1527 "movq " #b ", %%mm2 \n\t"\
1528 "movq " #c ", %%mm3 \n\t"\
1529 "movq " #d ", %%mm4 \n\t"\
1530 PAVGB(%%mm3, %%mm1) \
1531 PAVGB(%%mm4, %%mm0) \
1532 "movq %%mm0, %%mm3 \n\t"\
1533 "punpcklbw %%mm7, %%mm0 \n\t"\
1534 "punpckhbw %%mm7, %%mm3 \n\t"\
1535 "movq %%mm1, %%mm4 \n\t"\
1536 "punpcklbw %%mm7, %%mm1 \n\t"\
1537 "punpckhbw %%mm7, %%mm4 \n\t"\
1538 "psllw $2, %%mm1 \n\t"\
1539 "psllw $2, %%mm4 \n\t"\
1540 "psubw %%mm0, %%mm1 \n\t"\
1541 "psubw %%mm3, %%mm4 \n\t"\
1542 "movq %%mm2, %%mm5 \n\t"\
1543 "movq %%mm2, %%mm0 \n\t"\
1544 "punpcklbw %%mm7, %%mm2 \n\t"\
1545 "punpckhbw %%mm7, %%mm5 \n\t"\
1546 "paddw %%mm2, %%mm1 \n\t"\
1547 "paddw %%mm5, %%mm4 \n\t"\
1548 "psraw $2, %%mm1 \n\t"\
1549 "psraw $2, %%mm4 \n\t"\
1550 "packuswb %%mm4, %%mm1 \n\t"\
1551 "movq %%mm1, " #b " \n\t"\
1553 #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d)
1555 DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2))
1556 DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1557 DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2))
1558 DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1560 "movq %%mm0, (%2) \n\t"
1561 : : "r" (src), "r" ((x86_reg)stride), "r"(tmp)
1562 : "%"REG_a, "%"REG_d
1564 #else //HAVE_MMX2 || HAVE_AMD3DNOW
1569 int t2= src[stride*1];
1571 src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3);
1573 src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3);
1575 src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3);
1577 src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3);
1582 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
1586 * Deinterlace the given block by filtering every line with a (-1 2 6 2 -1) filter.
1587 * will be called for every 8x8 block and can read & write from line 4-15
1588 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1589 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1590 * this filter will read lines 4-13 and write 4-11
1592 static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2)
1594 #if HAVE_MMX2 || HAVE_AMD3DNOW
1597 "lea (%0, %1), %%"REG_a" \n\t"
1598 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1599 "pxor %%mm7, %%mm7 \n\t"
1600 "movq (%2), %%mm0 \n\t"
1601 "movq (%3), %%mm1 \n\t"
1602 // 0 1 2 3 4 5 6 7 8 9 10
1603 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1605 #define REAL_DEINT_L5(t1,t2,a,b,c)\
1606 "movq " #a ", %%mm2 \n\t"\
1607 "movq " #b ", %%mm3 \n\t"\
1608 "movq " #c ", %%mm4 \n\t"\
1611 "movq %%mm2, %%mm5 \n\t"\
1612 "movq %%mm2, " #t1 " \n\t"\
1613 "punpcklbw %%mm7, %%mm2 \n\t"\
1614 "punpckhbw %%mm7, %%mm5 \n\t"\
1615 "movq %%mm2, %%mm6 \n\t"\
1616 "paddw %%mm2, %%mm2 \n\t"\
1617 "paddw %%mm6, %%mm2 \n\t"\
1618 "movq %%mm5, %%mm6 \n\t"\
1619 "paddw %%mm5, %%mm5 \n\t"\
1620 "paddw %%mm6, %%mm5 \n\t"\
1621 "movq %%mm3, %%mm6 \n\t"\
1622 "punpcklbw %%mm7, %%mm3 \n\t"\
1623 "punpckhbw %%mm7, %%mm6 \n\t"\
1624 "paddw %%mm3, %%mm3 \n\t"\
1625 "paddw %%mm6, %%mm6 \n\t"\
1626 "paddw %%mm3, %%mm2 \n\t"\
1627 "paddw %%mm6, %%mm5 \n\t"\
1628 "movq %%mm4, %%mm6 \n\t"\
1629 "punpcklbw %%mm7, %%mm4 \n\t"\
1630 "punpckhbw %%mm7, %%mm6 \n\t"\
1631 "psubw %%mm4, %%mm2 \n\t"\
1632 "psubw %%mm6, %%mm5 \n\t"\
1633 "psraw $2, %%mm2 \n\t"\
1634 "psraw $2, %%mm5 \n\t"\
1635 "packuswb %%mm5, %%mm2 \n\t"\
1636 "movq %%mm2, " #a " \n\t"\
1638 #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c)
1640 DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) )
1641 DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2))
1642 DEINT_L5(%%mm0, %%mm1, (%%REGa, %1) , (%%REGa, %1, 2), (%0, %1, 4) )
1643 DEINT_L5(%%mm1, %%mm0, (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1644 DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) )
1645 DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2))
1646 DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) )
1647 DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1649 "movq %%mm0, (%2) \n\t"
1650 "movq %%mm1, (%3) \n\t"
1651 : : "r" (src), "r" ((x86_reg)stride), "r"(tmp), "r"(tmp2)
1652 : "%"REG_a, "%"REG_d
1654 #else //HAVE_MMX2 || HAVE_AMD3DNOW
1662 src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
1664 src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
1666 src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
1668 src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
1670 src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
1672 src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
1674 src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
1676 src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);
1683 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
1687 * Deinterlace the given block by filtering all lines with a (1 2 1) filter.
1688 * will be called for every 8x8 block and can read & write from line 4-15
1689 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1690 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1691 * this filter will read lines 4-13 and write 4-11
1693 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp)
1695 #if HAVE_MMX2 || HAVE_AMD3DNOW
1698 "lea (%0, %1), %%"REG_a" \n\t"
1699 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1700 // 0 1 2 3 4 5 6 7 8 9
1701 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1703 "movq (%2), %%mm0 \n\t" // L0
1704 "movq (%%"REG_a"), %%mm1 \n\t" // L2
1705 PAVGB(%%mm1, %%mm0) // L0+L2
1706 "movq (%0), %%mm2 \n\t" // L1
1708 "movq %%mm0, (%0) \n\t"
1709 "movq (%%"REG_a", %1), %%mm0 \n\t" // L3
1710 PAVGB(%%mm0, %%mm2) // L1+L3
1711 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1712 "movq %%mm2, (%%"REG_a") \n\t"
1713 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4
1714 PAVGB(%%mm2, %%mm1) // L2+L4
1715 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1716 "movq %%mm1, (%%"REG_a", %1) \n\t"
1717 "movq (%0, %1, 4), %%mm1 \n\t" // L5
1718 PAVGB(%%mm1, %%mm0) // L3+L5
1719 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1720 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
1721 "movq (%%"REG_d"), %%mm0 \n\t" // L6
1722 PAVGB(%%mm0, %%mm2) // L4+L6
1723 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1724 "movq %%mm2, (%0, %1, 4) \n\t"
1725 "movq (%%"REG_d", %1), %%mm2 \n\t" // L7
1726 PAVGB(%%mm2, %%mm1) // L5+L7
1727 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1728 "movq %%mm1, (%%"REG_d") \n\t"
1729 "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8
1730 PAVGB(%%mm1, %%mm0) // L6+L8
1731 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1732 "movq %%mm0, (%%"REG_d", %1) \n\t"
1733 "movq (%0, %1, 8), %%mm0 \n\t" // L9
1734 PAVGB(%%mm0, %%mm2) // L7+L9
1735 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1736 "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1737 "movq %%mm1, (%2) \n\t"
1739 : : "r" (src), "r" ((x86_reg)stride), "r" (tmp)
1740 : "%"REG_a, "%"REG_d
1742 #else //HAVE_MMX2 || HAVE_AMD3DNOW
1747 a= *(uint32_t*)&tmp[stride*0];
1748 b= *(uint32_t*)&src[stride*0];
1749 c= *(uint32_t*)&src[stride*1];
1750 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1751 *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1753 a= *(uint32_t*)&src[stride*2];
1754 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1755 *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1757 b= *(uint32_t*)&src[stride*3];
1758 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1759 *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1761 c= *(uint32_t*)&src[stride*4];
1762 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1763 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1765 a= *(uint32_t*)&src[stride*5];
1766 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1767 *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1769 b= *(uint32_t*)&src[stride*6];
1770 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1771 *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1773 c= *(uint32_t*)&src[stride*7];
1774 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1775 *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1777 a= *(uint32_t*)&src[stride*8];
1778 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1779 *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1781 *(uint32_t*)&tmp[stride*0]= c;
1785 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
1789 * Deinterlace the given block by applying a median filter to every second line.
1790 * will be called for every 8x8 block and can read & write from line 4-15,
1791 * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1792 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1794 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1800 "lea (%0, %1), %%"REG_a" \n\t"
1801 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1802 // 0 1 2 3 4 5 6 7 8 9
1803 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1805 "movq (%0), %%mm0 \n\t" //
1806 "movq (%%"REG_a", %1), %%mm2 \n\t" //
1807 "movq (%%"REG_a"), %%mm1 \n\t" //
1808 "movq %%mm0, %%mm3 \n\t"
1809 "pmaxub %%mm1, %%mm0 \n\t" //
1810 "pminub %%mm3, %%mm1 \n\t" //
1811 "pmaxub %%mm2, %%mm1 \n\t" //
1812 "pminub %%mm1, %%mm0 \n\t"
1813 "movq %%mm0, (%%"REG_a") \n\t"
1815 "movq (%0, %1, 4), %%mm0 \n\t" //
1816 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" //
1817 "movq %%mm2, %%mm3 \n\t"
1818 "pmaxub %%mm1, %%mm2 \n\t" //
1819 "pminub %%mm3, %%mm1 \n\t" //
1820 "pmaxub %%mm0, %%mm1 \n\t" //
1821 "pminub %%mm1, %%mm2 \n\t"
1822 "movq %%mm2, (%%"REG_a", %1, 2) \n\t"
1824 "movq (%%"REG_d"), %%mm2 \n\t" //
1825 "movq (%%"REG_d", %1), %%mm1 \n\t" //
1826 "movq %%mm2, %%mm3 \n\t"
1827 "pmaxub %%mm0, %%mm2 \n\t" //
1828 "pminub %%mm3, %%mm0 \n\t" //
1829 "pmaxub %%mm1, %%mm0 \n\t" //
1830 "pminub %%mm0, %%mm2 \n\t"
1831 "movq %%mm2, (%%"REG_d") \n\t"
1833 "movq (%%"REG_d", %1, 2), %%mm2 \n\t" //
1834 "movq (%0, %1, 8), %%mm0 \n\t" //
1835 "movq %%mm2, %%mm3 \n\t"
1836 "pmaxub %%mm0, %%mm2 \n\t" //
1837 "pminub %%mm3, %%mm0 \n\t" //
1838 "pmaxub %%mm1, %%mm0 \n\t" //
1839 "pminub %%mm0, %%mm2 \n\t"
1840 "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1843 : : "r" (src), "r" ((x86_reg)stride)
1844 : "%"REG_a, "%"REG_d
1847 #else // MMX without MMX2
1849 "lea (%0, %1), %%"REG_a" \n\t"
1850 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1851 // 0 1 2 3 4 5 6 7 8 9
1852 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1853 "pxor %%mm7, %%mm7 \n\t"
1855 #define REAL_MEDIAN(a,b,c)\
1856 "movq " #a ", %%mm0 \n\t"\
1857 "movq " #b ", %%mm2 \n\t"\
1858 "movq " #c ", %%mm1 \n\t"\
1859 "movq %%mm0, %%mm3 \n\t"\
1860 "movq %%mm1, %%mm4 \n\t"\
1861 "movq %%mm2, %%mm5 \n\t"\
1862 "psubusb %%mm1, %%mm3 \n\t"\
1863 "psubusb %%mm2, %%mm4 \n\t"\
1864 "psubusb %%mm0, %%mm5 \n\t"\
1865 "pcmpeqb %%mm7, %%mm3 \n\t"\
1866 "pcmpeqb %%mm7, %%mm4 \n\t"\
1867 "pcmpeqb %%mm7, %%mm5 \n\t"\
1868 "movq %%mm3, %%mm6 \n\t"\
1869 "pxor %%mm4, %%mm3 \n\t"\
1870 "pxor %%mm5, %%mm4 \n\t"\
1871 "pxor %%mm6, %%mm5 \n\t"\
1872 "por %%mm3, %%mm1 \n\t"\
1873 "por %%mm4, %%mm2 \n\t"\
1874 "por %%mm5, %%mm0 \n\t"\
1875 "pand %%mm2, %%mm0 \n\t"\
1876 "pand %%mm1, %%mm0 \n\t"\
1877 "movq %%mm0, " #b " \n\t"
1878 #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c)
1880 MEDIAN((%0) , (%%REGa) , (%%REGa, %1))
1881 MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4))
1882 MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1))
1883 MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8))
1885 : : "r" (src), "r" ((x86_reg)stride)
1886 : "%"REG_a, "%"REG_d
1892 // FIXME - there should be a way to do a few columns in parallel like w/mmx
1894 uint8_t *colsrc = src;
1895 for (y=0; y<4; y++){
1896 int a, b, c, d, e, f;
1898 b = colsrc[stride ];
1899 c = colsrc[stride*2];
1903 colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f));
1913 * Transpose and shift the given 8x8 Block into dst1 and dst2.
1915 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
1918 "lea (%0, %1), %%"REG_a" \n\t"
1919 // 0 1 2 3 4 5 6 7 8 9
1920 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1921 "movq (%0), %%mm0 \n\t" // 12345678
1922 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
1923 "movq %%mm0, %%mm2 \n\t" // 12345678
1924 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
1925 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
1927 "movq (%%"REG_a", %1), %%mm1 \n\t"
1928 "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
1929 "movq %%mm1, %%mm4 \n\t"
1930 "punpcklbw %%mm3, %%mm1 \n\t"
1931 "punpckhbw %%mm3, %%mm4 \n\t"
1933 "movq %%mm0, %%mm3 \n\t"
1934 "punpcklwd %%mm1, %%mm0 \n\t"
1935 "punpckhwd %%mm1, %%mm3 \n\t"
1936 "movq %%mm2, %%mm1 \n\t"
1937 "punpcklwd %%mm4, %%mm2 \n\t"
1938 "punpckhwd %%mm4, %%mm1 \n\t"
1940 "movd %%mm0, 128(%2) \n\t"
1941 "psrlq $32, %%mm0 \n\t"
1942 "movd %%mm0, 144(%2) \n\t"
1943 "movd %%mm3, 160(%2) \n\t"
1944 "psrlq $32, %%mm3 \n\t"
1945 "movd %%mm3, 176(%2) \n\t"
1946 "movd %%mm3, 48(%3) \n\t"
1947 "movd %%mm2, 192(%2) \n\t"
1948 "movd %%mm2, 64(%3) \n\t"
1949 "psrlq $32, %%mm2 \n\t"
1950 "movd %%mm2, 80(%3) \n\t"
1951 "movd %%mm1, 96(%3) \n\t"
1952 "psrlq $32, %%mm1 \n\t"
1953 "movd %%mm1, 112(%3) \n\t"
1955 "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t"
1957 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
1958 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
1959 "movq %%mm0, %%mm2 \n\t" // 12345678
1960 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
1961 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
1963 "movq (%%"REG_a", %1), %%mm1 \n\t"
1964 "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
1965 "movq %%mm1, %%mm4 \n\t"
1966 "punpcklbw %%mm3, %%mm1 \n\t"
1967 "punpckhbw %%mm3, %%mm4 \n\t"
1969 "movq %%mm0, %%mm3 \n\t"
1970 "punpcklwd %%mm1, %%mm0 \n\t"
1971 "punpckhwd %%mm1, %%mm3 \n\t"
1972 "movq %%mm2, %%mm1 \n\t"
1973 "punpcklwd %%mm4, %%mm2 \n\t"
1974 "punpckhwd %%mm4, %%mm1 \n\t"
1976 "movd %%mm0, 132(%2) \n\t"
1977 "psrlq $32, %%mm0 \n\t"
1978 "movd %%mm0, 148(%2) \n\t"
1979 "movd %%mm3, 164(%2) \n\t"
1980 "psrlq $32, %%mm3 \n\t"
1981 "movd %%mm3, 180(%2) \n\t"
1982 "movd %%mm3, 52(%3) \n\t"
1983 "movd %%mm2, 196(%2) \n\t"
1984 "movd %%mm2, 68(%3) \n\t"
1985 "psrlq $32, %%mm2 \n\t"
1986 "movd %%mm2, 84(%3) \n\t"
1987 "movd %%mm1, 100(%3) \n\t"
1988 "psrlq $32, %%mm1 \n\t"
1989 "movd %%mm1, 116(%3) \n\t"
1992 :: "r" (src), "r" ((x86_reg)srcStride), "r" (dst1), "r" (dst2)
1998 * Transpose the given 8x8 block.
2000 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
2003 "lea (%0, %1), %%"REG_a" \n\t"
2004 "lea (%%"REG_a",%1,4), %%"REG_d" \n\t"
2005 // 0 1 2 3 4 5 6 7 8 9
2006 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2007 "movq (%2), %%mm0 \n\t" // 12345678
2008 "movq 16(%2), %%mm1 \n\t" // abcdefgh
2009 "movq %%mm0, %%mm2 \n\t" // 12345678
2010 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2011 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2013 "movq 32(%2), %%mm1 \n\t"
2014 "movq 48(%2), %%mm3 \n\t"
2015 "movq %%mm1, %%mm4 \n\t"
2016 "punpcklbw %%mm3, %%mm1 \n\t"
2017 "punpckhbw %%mm3, %%mm4 \n\t"
2019 "movq %%mm0, %%mm3 \n\t"
2020 "punpcklwd %%mm1, %%mm0 \n\t"
2021 "punpckhwd %%mm1, %%mm3 \n\t"
2022 "movq %%mm2, %%mm1 \n\t"
2023 "punpcklwd %%mm4, %%mm2 \n\t"
2024 "punpckhwd %%mm4, %%mm1 \n\t"
2026 "movd %%mm0, (%0) \n\t"
2027 "psrlq $32, %%mm0 \n\t"
2028 "movd %%mm0, (%%"REG_a") \n\t"
2029 "movd %%mm3, (%%"REG_a", %1) \n\t"
2030 "psrlq $32, %%mm3 \n\t"
2031 "movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2032 "movd %%mm2, (%0, %1, 4) \n\t"
2033 "psrlq $32, %%mm2 \n\t"
2034 "movd %%mm2, (%%"REG_d") \n\t"
2035 "movd %%mm1, (%%"REG_d", %1) \n\t"
2036 "psrlq $32, %%mm1 \n\t"
2037 "movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2040 "movq 64(%2), %%mm0 \n\t" // 12345678
2041 "movq 80(%2), %%mm1 \n\t" // abcdefgh
2042 "movq %%mm0, %%mm2 \n\t" // 12345678
2043 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2044 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2046 "movq 96(%2), %%mm1 \n\t"
2047 "movq 112(%2), %%mm3 \n\t"
2048 "movq %%mm1, %%mm4 \n\t"
2049 "punpcklbw %%mm3, %%mm1 \n\t"
2050 "punpckhbw %%mm3, %%mm4 \n\t"
2052 "movq %%mm0, %%mm3 \n\t"
2053 "punpcklwd %%mm1, %%mm0 \n\t"
2054 "punpckhwd %%mm1, %%mm3 \n\t"
2055 "movq %%mm2, %%mm1 \n\t"
2056 "punpcklwd %%mm4, %%mm2 \n\t"
2057 "punpckhwd %%mm4, %%mm1 \n\t"
2059 "movd %%mm0, 4(%0) \n\t"
2060 "psrlq $32, %%mm0 \n\t"
2061 "movd %%mm0, 4(%%"REG_a") \n\t"
2062 "movd %%mm3, 4(%%"REG_a", %1) \n\t"
2063 "psrlq $32, %%mm3 \n\t"
2064 "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t"
2065 "movd %%mm2, 4(%0, %1, 4) \n\t"
2066 "psrlq $32, %%mm2 \n\t"
2067 "movd %%mm2, 4(%%"REG_d") \n\t"
2068 "movd %%mm1, 4(%%"REG_d", %1) \n\t"
2069 "psrlq $32, %%mm1 \n\t"
2070 "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t"
2072 :: "r" (dst), "r" ((x86_reg)dstStride), "r" (src)
2073 : "%"REG_a, "%"REG_d
2077 //static long test=0;
2080 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2081 uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise)
2083 // to save a register (FIXME do this outside of the loops)
2084 tempBlurredPast[127]= maxNoise[0];
2085 tempBlurredPast[128]= maxNoise[1];
2086 tempBlurredPast[129]= maxNoise[2];
2088 #define FAST_L2_DIFF
2089 //#define L1_DIFF //u should change the thresholds too if u try that one
2090 #if HAVE_MMX2 || HAVE_AMD3DNOW
2092 "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride
2093 "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride
2094 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2095 // 0 1 2 3 4 5 6 7 8 9
2096 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
2098 #ifdef L1_DIFF //needs mmx2
2099 "movq (%0), %%mm0 \n\t" // L0
2100 "psadbw (%1), %%mm0 \n\t" // |L0-R0|
2101 "movq (%0, %2), %%mm1 \n\t" // L1
2102 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2103 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2104 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2105 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2106 "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3|
2108 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2109 "paddw %%mm1, %%mm0 \n\t"
2110 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2111 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2112 "paddw %%mm2, %%mm0 \n\t"
2113 "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5|
2114 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2115 "paddw %%mm3, %%mm0 \n\t"
2116 "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6|
2117 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2118 "paddw %%mm4, %%mm0 \n\t"
2119 "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7|
2120 "paddw %%mm5, %%mm6 \n\t"
2121 "paddw %%mm7, %%mm6 \n\t"
2122 "paddw %%mm6, %%mm0 \n\t"
2124 #if defined (FAST_L2_DIFF)
2125 "pcmpeqb %%mm7, %%mm7 \n\t"
2126 "movq "MANGLE(b80)", %%mm6 \n\t"
2127 "pxor %%mm0, %%mm0 \n\t"
2128 #define REAL_L2_DIFF_CORE(a, b)\
2129 "movq " #a ", %%mm5 \n\t"\
2130 "movq " #b ", %%mm2 \n\t"\
2131 "pxor %%mm7, %%mm2 \n\t"\
2132 PAVGB(%%mm2, %%mm5)\
2133 "paddb %%mm6, %%mm5 \n\t"\
2134 "movq %%mm5, %%mm2 \n\t"\
2135 "psllw $8, %%mm5 \n\t"\
2136 "pmaddwd %%mm5, %%mm5 \n\t"\
2137 "pmaddwd %%mm2, %%mm2 \n\t"\
2138 "paddd %%mm2, %%mm5 \n\t"\
2139 "psrld $14, %%mm5 \n\t"\
2140 "paddd %%mm5, %%mm0 \n\t"
2142 #else //defined (FAST_L2_DIFF)
2143 "pxor %%mm7, %%mm7 \n\t"
2144 "pxor %%mm0, %%mm0 \n\t"
2145 #define REAL_L2_DIFF_CORE(a, b)\
2146 "movq " #a ", %%mm5 \n\t"\
2147 "movq " #b ", %%mm2 \n\t"\
2148 "movq %%mm5, %%mm1 \n\t"\
2149 "movq %%mm2, %%mm3 \n\t"\
2150 "punpcklbw %%mm7, %%mm5 \n\t"\
2151 "punpckhbw %%mm7, %%mm1 \n\t"\
2152 "punpcklbw %%mm7, %%mm2 \n\t"\
2153 "punpckhbw %%mm7, %%mm3 \n\t"\
2154 "psubw %%mm2, %%mm5 \n\t"\
2155 "psubw %%mm3, %%mm1 \n\t"\
2156 "pmaddwd %%mm5, %%mm5 \n\t"\
2157 "pmaddwd %%mm1, %%mm1 \n\t"\
2158 "paddd %%mm1, %%mm5 \n\t"\
2159 "paddd %%mm5, %%mm0 \n\t"
2161 #endif //defined (FAST_L2_DIFF)
2163 #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b)
2165 L2_DIFF_CORE((%0) , (%1))
2166 L2_DIFF_CORE((%0, %2) , (%1, %2))
2167 L2_DIFF_CORE((%0, %2, 2) , (%1, %2, 2))
2168 L2_DIFF_CORE((%0, %%REGa) , (%1, %%REGa))
2169 L2_DIFF_CORE((%0, %2, 4) , (%1, %2, 4))
2170 L2_DIFF_CORE((%0, %%REGd) , (%1, %%REGd))
2171 L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2))
2172 L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc))
2176 "movq %%mm0, %%mm4 \n\t"
2177 "psrlq $32, %%mm0 \n\t"
2178 "paddd %%mm0, %%mm4 \n\t"
2179 "movd %%mm4, %%ecx \n\t"
2180 "shll $2, %%ecx \n\t"
2181 "mov %3, %%"REG_d" \n\t"
2182 "addl -4(%%"REG_d"), %%ecx \n\t"
2183 "addl 4(%%"REG_d"), %%ecx \n\t"
2184 "addl -1024(%%"REG_d"), %%ecx \n\t"
2185 "addl $4, %%ecx \n\t"
2186 "addl 1024(%%"REG_d"), %%ecx \n\t"
2187 "shrl $3, %%ecx \n\t"
2188 "movl %%ecx, (%%"REG_d") \n\t"
2190 // "mov %3, %%"REG_c" \n\t"
2191 // "mov %%"REG_c", test \n\t"
2193 "cmpl 512(%%"REG_d"), %%ecx \n\t"
2195 "cmpl 516(%%"REG_d"), %%ecx \n\t"
2198 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2199 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2200 "movq (%0), %%mm0 \n\t" // L0
2201 "movq (%0, %2), %%mm1 \n\t" // L1
2202 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2203 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2204 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2205 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2206 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2207 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2208 "movq %%mm0, (%1) \n\t" // L0
2209 "movq %%mm1, (%1, %2) \n\t" // L1
2210 "movq %%mm2, (%1, %2, 2) \n\t" // L2
2211 "movq %%mm3, (%1, %%"REG_a") \n\t" // L3
2212 "movq %%mm4, (%1, %2, 4) \n\t" // L4
2213 "movq %%mm5, (%1, %%"REG_d") \n\t" // L5
2214 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6
2215 "movq %%mm7, (%1, %%"REG_c") \n\t" // L7
2219 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2220 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2221 "movq (%0), %%mm0 \n\t" // L0
2222 PAVGB((%1), %%mm0) // L0
2223 "movq (%0, %2), %%mm1 \n\t" // L1
2224 PAVGB((%1, %2), %%mm1) // L1
2225 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2226 PAVGB((%1, %2, 2), %%mm2) // L2
2227 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2228 PAVGB((%1, %%REGa), %%mm3) // L3
2229 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2230 PAVGB((%1, %2, 4), %%mm4) // L4
2231 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2232 PAVGB((%1, %%REGd), %%mm5) // L5
2233 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2234 PAVGB((%1, %%REGa, 2), %%mm6) // L6
2235 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2236 PAVGB((%1, %%REGc), %%mm7) // L7
2237 "movq %%mm0, (%1) \n\t" // R0
2238 "movq %%mm1, (%1, %2) \n\t" // R1
2239 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2240 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2241 "movq %%mm4, (%1, %2, 4) \n\t" // R4
2242 "movq %%mm5, (%1, %%"REG_d") \n\t" // R5
2243 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6
2244 "movq %%mm7, (%1, %%"REG_c") \n\t" // R7
2245 "movq %%mm0, (%0) \n\t" // L0
2246 "movq %%mm1, (%0, %2) \n\t" // L1
2247 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2248 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2249 "movq %%mm4, (%0, %2, 4) \n\t" // L4
2250 "movq %%mm5, (%0, %%"REG_d") \n\t" // L5
2251 "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6
2252 "movq %%mm7, (%0, %%"REG_c") \n\t" // L7
2256 "cmpl 508(%%"REG_d"), %%ecx \n\t"
2259 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2260 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2261 "movq (%0), %%mm0 \n\t" // L0
2262 "movq (%0, %2), %%mm1 \n\t" // L1
2263 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2264 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2265 "movq (%1), %%mm4 \n\t" // R0
2266 "movq (%1, %2), %%mm5 \n\t" // R1
2267 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2268 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2277 "movq %%mm0, (%1) \n\t" // R0
2278 "movq %%mm1, (%1, %2) \n\t" // R1
2279 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2280 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2281 "movq %%mm0, (%0) \n\t" // L0
2282 "movq %%mm1, (%0, %2) \n\t" // L1
2283 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2284 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2286 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2287 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2288 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2289 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2290 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2291 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2292 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2293 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2302 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2303 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2304 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2305 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2306 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2307 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2308 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2309 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2313 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2314 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2315 "movq (%0), %%mm0 \n\t" // L0
2316 "movq (%0, %2), %%mm1 \n\t" // L1
2317 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2318 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2319 "movq (%1), %%mm4 \n\t" // R0
2320 "movq (%1, %2), %%mm5 \n\t" // R1
2321 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2322 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2335 "movq %%mm0, (%1) \n\t" // R0
2336 "movq %%mm1, (%1, %2) \n\t" // R1
2337 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2338 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2339 "movq %%mm0, (%0) \n\t" // L0
2340 "movq %%mm1, (%0, %2) \n\t" // L1
2341 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2342 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2344 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2345 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2346 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2347 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2348 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2349 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2350 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2351 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2364 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2365 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2366 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2367 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2368 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2369 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2370 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2371 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2375 :: "r" (src), "r" (tempBlurred), "r"((x86_reg)stride), "m" (tempBlurredPast)
2376 : "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2378 #else //HAVE_MMX2 || HAVE_AMD3DNOW
2388 int ref= tempBlurred[ x + y*stride ];
2389 int cur= src[ x + y*stride ];
2391 // if(x==0 || x==7) d1+= d1>>1;
2392 // if(y==0 || y==7) d1+= d1>>1;
2401 +(*(tempBlurredPast-256))
2402 +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1))
2403 +(*(tempBlurredPast+256))
2406 // ((*tempBlurredPast)*3 + d + 2)>>2;
2411 64 32 16 8 4 2 1 (1)
2412 64 48 36 27 20 15 11 (33) (approx)
2413 64 56 49 43 37 33 29 (200) (approx)
2415 if(d > maxNoise[1]){
2416 if(d < maxNoise[2]){
2420 int ref= tempBlurred[ x + y*stride ];
2421 int cur= src[ x + y*stride ];
2422 tempBlurred[ x + y*stride ]=
2423 src[ x + y*stride ]=
2431 tempBlurred[ x + y*stride ]= src[ x + y*stride ];
2436 if(d < maxNoise[0]){
2440 int ref= tempBlurred[ x + y*stride ];
2441 int cur= src[ x + y*stride ];
2442 tempBlurred[ x + y*stride ]=
2443 src[ x + y*stride ]=
2444 (ref*7 + cur + 4)>>3;
2451 int ref= tempBlurred[ x + y*stride ];
2452 int cur= src[ x + y*stride ];
2453 tempBlurred[ x + y*stride ]=
2454 src[ x + y*stride ]=
2455 (ref*3 + cur + 2)>>2;
2461 #endif //HAVE_MMX2 || HAVE_AMD3DNOW
2463 #endif //HAVE_ALTIVEC
2467 * accurate deblock filter
2469 static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){
2470 int64_t dc_mask, eq_mask, both_masks;
2471 int64_t sums[10*8*2];
2472 src+= step*3; // src points to begin of the 8x8 Block
2475 "movq %0, %%mm7 \n\t"
2476 "movq %1, %%mm6 \n\t"
2477 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
2481 "lea (%2, %3), %%"REG_a" \n\t"
2482 // 0 1 2 3 4 5 6 7 8 9
2483 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
2485 "movq (%2), %%mm0 \n\t"
2486 "movq (%%"REG_a"), %%mm1 \n\t"
2487 "movq %%mm1, %%mm3 \n\t"
2488 "movq %%mm1, %%mm4 \n\t"
2489 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
2490 "paddb %%mm7, %%mm0 \n\t"
2491 "pcmpgtb %%mm6, %%mm0 \n\t"
2493 "movq (%%"REG_a",%3), %%mm2 \n\t"
2494 PMAXUB(%%mm2, %%mm4)
2495 PMINUB(%%mm2, %%mm3, %%mm5)
2496 "psubb %%mm2, %%mm1 \n\t"
2497 "paddb %%mm7, %%mm1 \n\t"
2498 "pcmpgtb %%mm6, %%mm1 \n\t"
2499 "paddb %%mm1, %%mm0 \n\t"
2501 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2502 PMAXUB(%%mm1, %%mm4)
2503 PMINUB(%%mm1, %%mm3, %%mm5)
2504 "psubb %%mm1, %%mm2 \n\t"
2505 "paddb %%mm7, %%mm2 \n\t"
2506 "pcmpgtb %%mm6, %%mm2 \n\t"
2507 "paddb %%mm2, %%mm0 \n\t"
2509 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
2511 "movq (%2, %3, 4), %%mm2 \n\t"
2512 PMAXUB(%%mm2, %%mm4)
2513 PMINUB(%%mm2, %%mm3, %%mm5)
2514 "psubb %%mm2, %%mm1 \n\t"
2515 "paddb %%mm7, %%mm1 \n\t"
2516 "pcmpgtb %%mm6, %%mm1 \n\t"
2517 "paddb %%mm1, %%mm0 \n\t"
2519 "movq (%%"REG_a"), %%mm1 \n\t"
2520 PMAXUB(%%mm1, %%mm4)
2521 PMINUB(%%mm1, %%mm3, %%mm5)
2522 "psubb %%mm1, %%mm2 \n\t"
2523 "paddb %%mm7, %%mm2 \n\t"
2524 "pcmpgtb %%mm6, %%mm2 \n\t"
2525 "paddb %%mm2, %%mm0 \n\t"
2527 "movq (%%"REG_a", %3), %%mm2 \n\t"
2528 PMAXUB(%%mm2, %%mm4)
2529 PMINUB(%%mm2, %%mm3, %%mm5)
2530 "psubb %%mm2, %%mm1 \n\t"
2531 "paddb %%mm7, %%mm1 \n\t"
2532 "pcmpgtb %%mm6, %%mm1 \n\t"
2533 "paddb %%mm1, %%mm0 \n\t"
2535 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2536 PMAXUB(%%mm1, %%mm4)
2537 PMINUB(%%mm1, %%mm3, %%mm5)
2538 "psubb %%mm1, %%mm2 \n\t"
2539 "paddb %%mm7, %%mm2 \n\t"
2540 "pcmpgtb %%mm6, %%mm2 \n\t"
2541 "paddb %%mm2, %%mm0 \n\t"
2543 "movq (%2, %3, 8), %%mm2 \n\t"
2544 PMAXUB(%%mm2, %%mm4)
2545 PMINUB(%%mm2, %%mm3, %%mm5)
2546 "psubb %%mm2, %%mm1 \n\t"
2547 "paddb %%mm7, %%mm1 \n\t"
2548 "pcmpgtb %%mm6, %%mm1 \n\t"
2549 "paddb %%mm1, %%mm0 \n\t"
2551 "movq (%%"REG_a", %3, 4), %%mm1 \n\t"
2552 "psubb %%mm1, %%mm2 \n\t"
2553 "paddb %%mm7, %%mm2 \n\t"
2554 "pcmpgtb %%mm6, %%mm2 \n\t"
2555 "paddb %%mm2, %%mm0 \n\t"
2556 "psubusb %%mm3, %%mm4 \n\t"
2558 "pxor %%mm6, %%mm6 \n\t"
2559 "movq %4, %%mm7 \n\t" // QP,..., QP
2560 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
2561 "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0
2562 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2563 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2564 "movq %%mm7, %1 \n\t"
2566 "movq %5, %%mm7 \n\t"
2567 "punpcklbw %%mm7, %%mm7 \n\t"
2568 "punpcklbw %%mm7, %%mm7 \n\t"
2569 "punpcklbw %%mm7, %%mm7 \n\t"
2570 "psubb %%mm0, %%mm6 \n\t"
2571 "pcmpgtb %%mm7, %%mm6 \n\t"
2572 "movq %%mm6, %0 \n\t"
2574 : "=m" (eq_mask), "=m" (dc_mask)
2575 : "r" (src), "r" ((x86_reg)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
2579 both_masks = dc_mask & eq_mask;
2582 x86_reg offset= -8*step;
2583 int64_t *temp_sums= sums;
2586 "movq %2, %%mm0 \n\t" // QP,..., QP
2587 "pxor %%mm4, %%mm4 \n\t"
2589 "movq (%0), %%mm6 \n\t"
2590 "movq (%0, %1), %%mm5 \n\t"
2591 "movq %%mm5, %%mm1 \n\t"
2592 "movq %%mm6, %%mm2 \n\t"
2593 "psubusb %%mm6, %%mm5 \n\t"
2594 "psubusb %%mm1, %%mm2 \n\t"
2595 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2596 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2597 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2599 "pxor %%mm6, %%mm1 \n\t"
2600 "pand %%mm0, %%mm1 \n\t"
2601 "pxor %%mm1, %%mm6 \n\t"
2604 "movq (%0, %1, 8), %%mm5 \n\t"
2605 "add %1, %0 \n\t" // %0 points to line 1 not 0
2606 "movq (%0, %1, 8), %%mm7 \n\t"
2607 "movq %%mm5, %%mm1 \n\t"
2608 "movq %%mm7, %%mm2 \n\t"
2609 "psubusb %%mm7, %%mm5 \n\t"
2610 "psubusb %%mm1, %%mm2 \n\t"
2611 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2612 "movq %2, %%mm0 \n\t" // QP,..., QP
2613 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2614 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2616 "pxor %%mm7, %%mm1 \n\t"
2617 "pand %%mm0, %%mm1 \n\t"
2618 "pxor %%mm1, %%mm7 \n\t"
2620 "movq %%mm6, %%mm5 \n\t"
2621 "punpckhbw %%mm4, %%mm6 \n\t"
2622 "punpcklbw %%mm4, %%mm5 \n\t"
2623 // 4:0 5/6:First 7:Last
2625 "movq %%mm5, %%mm0 \n\t"
2626 "movq %%mm6, %%mm1 \n\t"
2627 "psllw $2, %%mm0 \n\t"
2628 "psllw $2, %%mm1 \n\t"
2629 "paddw "MANGLE(w04)", %%mm0 \n\t"
2630 "paddw "MANGLE(w04)", %%mm1 \n\t"
2633 "movq (%0), %%mm2 \n\t"\
2634 "movq (%0), %%mm3 \n\t"\
2636 "punpcklbw %%mm4, %%mm2 \n\t"\
2637 "punpckhbw %%mm4, %%mm3 \n\t"\
2638 "paddw %%mm2, %%mm0 \n\t"\
2639 "paddw %%mm3, %%mm1 \n\t"
2642 "movq (%0), %%mm2 \n\t"\
2643 "movq (%0), %%mm3 \n\t"\
2645 "punpcklbw %%mm4, %%mm2 \n\t"\
2646 "punpckhbw %%mm4, %%mm3 \n\t"\
2647 "psubw %%mm2, %%mm0 \n\t"\
2648 "psubw %%mm3, %%mm1 \n\t"
2654 "movq %%mm0, (%3) \n\t"
2655 "movq %%mm1, 8(%3) \n\t"
2658 "psubw %%mm5, %%mm0 \n\t"
2659 "psubw %%mm6, %%mm1 \n\t"
2660 "movq %%mm0, 16(%3) \n\t"
2661 "movq %%mm1, 24(%3) \n\t"
2664 "psubw %%mm5, %%mm0 \n\t"
2665 "psubw %%mm6, %%mm1 \n\t"
2666 "movq %%mm0, 32(%3) \n\t"
2667 "movq %%mm1, 40(%3) \n\t"
2670 "psubw %%mm5, %%mm0 \n\t"
2671 "psubw %%mm6, %%mm1 \n\t"
2672 "movq %%mm0, 48(%3) \n\t"
2673 "movq %%mm1, 56(%3) \n\t"
2676 "psubw %%mm5, %%mm0 \n\t"
2677 "psubw %%mm6, %%mm1 \n\t"
2678 "movq %%mm0, 64(%3) \n\t"
2679 "movq %%mm1, 72(%3) \n\t"
2681 "movq %%mm7, %%mm6 \n\t"
2682 "punpckhbw %%mm4, %%mm7 \n\t"
2683 "punpcklbw %%mm4, %%mm6 \n\t"
2689 "movq %%mm0, 80(%3) \n\t"
2690 "movq %%mm1, 88(%3) \n\t"
2693 "paddw %%mm6, %%mm0 \n\t"
2694 "paddw %%mm7, %%mm1 \n\t"
2695 "movq %%mm0, 96(%3) \n\t"
2696 "movq %%mm1, 104(%3) \n\t"
2699 "paddw %%mm6, %%mm0 \n\t"
2700 "paddw %%mm7, %%mm1 \n\t"
2701 "movq %%mm0, 112(%3) \n\t"
2702 "movq %%mm1, 120(%3) \n\t"
2705 "paddw %%mm6, %%mm0 \n\t"
2706 "paddw %%mm7, %%mm1 \n\t"
2707 "movq %%mm0, 128(%3) \n\t"
2708 "movq %%mm1, 136(%3) \n\t"
2711 "paddw %%mm6, %%mm0 \n\t"
2712 "paddw %%mm7, %%mm1 \n\t"
2713 "movq %%mm0, 144(%3) \n\t"
2714 "movq %%mm1, 152(%3) \n\t"
2716 "mov %4, %0 \n\t" //FIXME
2719 : "r" ((x86_reg)step), "m" (c->pQPb), "r"(sums), "g"(src)
2722 src+= step; // src points to begin of the 8x8 Block
2725 "movq %4, %%mm6 \n\t"
2726 "pcmpeqb %%mm5, %%mm5 \n\t"
2727 "pxor %%mm6, %%mm5 \n\t"
2728 "pxor %%mm7, %%mm7 \n\t"
2731 "movq (%1), %%mm0 \n\t"
2732 "movq 8(%1), %%mm1 \n\t"
2733 "paddw 32(%1), %%mm0 \n\t"
2734 "paddw 40(%1), %%mm1 \n\t"
2735 "movq (%0, %3), %%mm2 \n\t"
2736 "movq %%mm2, %%mm3 \n\t"
2737 "movq %%mm2, %%mm4 \n\t"
2738 "punpcklbw %%mm7, %%mm2 \n\t"
2739 "punpckhbw %%mm7, %%mm3 \n\t"
2740 "paddw %%mm2, %%mm0 \n\t"
2741 "paddw %%mm3, %%mm1 \n\t"
2742 "paddw %%mm2, %%mm0 \n\t"
2743 "paddw %%mm3, %%mm1 \n\t"
2744 "psrlw $4, %%mm0 \n\t"
2745 "psrlw $4, %%mm1 \n\t"
2746 "packuswb %%mm1, %%mm0 \n\t"
2747 "pand %%mm6, %%mm0 \n\t"
2748 "pand %%mm5, %%mm4 \n\t"
2749 "por %%mm4, %%mm0 \n\t"
2750 "movq %%mm0, (%0, %3) \n\t"
2755 : "+r"(offset), "+r"(temp_sums)
2756 : "r" ((x86_reg)step), "r"(src - offset), "m"(both_masks)
2759 src+= step; // src points to begin of the 8x8 Block
2761 if(eq_mask != -1LL){
2762 uint8_t *temp_src= src;
2763 DECLARE_ALIGNED(8, uint64_t, tmp)[4]; // make space for 4 8-byte vars
2765 "pxor %%mm7, %%mm7 \n\t"
2766 // 0 1 2 3 4 5 6 7 8 9
2767 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1
2769 "movq (%0), %%mm0 \n\t"
2770 "movq %%mm0, %%mm1 \n\t"
2771 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
2772 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
2774 "movq (%0, %1), %%mm2 \n\t"
2775 "lea (%0, %1, 2), %%"REG_a" \n\t"
2776 "movq %%mm2, %%mm3 \n\t"
2777 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
2778 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
2780 "movq (%%"REG_a"), %%mm4 \n\t"
2781 "movq %%mm4, %%mm5 \n\t"
2782 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
2783 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
2785 "paddw %%mm0, %%mm0 \n\t" // 2L0
2786 "paddw %%mm1, %%mm1 \n\t" // 2H0
2787 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
2788 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
2789 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
2790 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
2792 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
2793 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
2794 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
2795 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
2797 "movq (%%"REG_a", %1), %%mm2 \n\t"
2798 "movq %%mm2, %%mm3 \n\t"
2799 "punpcklbw %%mm7, %%mm2 \n\t" // L3
2800 "punpckhbw %%mm7, %%mm3 \n\t" // H3
2802 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
2803 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
2804 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2805 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2806 "movq %%mm0, (%4) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2807 "movq %%mm1, 8(%4) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2809 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
2810 "movq %%mm0, %%mm1 \n\t"
2811 "punpcklbw %%mm7, %%mm0 \n\t" // L4
2812 "punpckhbw %%mm7, %%mm1 \n\t" // H4
2814 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
2815 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
2816 "movq %%mm2, 16(%4) \n\t" // L3 - L4
2817 "movq %%mm3, 24(%4) \n\t" // H3 - H4
2818 "paddw %%mm4, %%mm4 \n\t" // 2L2
2819 "paddw %%mm5, %%mm5 \n\t" // 2H2
2820 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
2821 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
2823 "lea (%%"REG_a", %1), %0 \n\t"
2824 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
2825 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
2826 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
2827 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
2829 "movq (%0, %1, 2), %%mm2 \n\t"
2830 "movq %%mm2, %%mm3 \n\t"
2831 "punpcklbw %%mm7, %%mm2 \n\t" // L5
2832 "punpckhbw %%mm7, %%mm3 \n\t" // H5
2833 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
2834 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
2835 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
2836 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
2838 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
2839 "punpcklbw %%mm7, %%mm6 \n\t" // L6
2840 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
2841 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
2842 "punpckhbw %%mm7, %%mm6 \n\t" // H6
2843 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
2845 "paddw %%mm0, %%mm0 \n\t" // 2L4
2846 "paddw %%mm1, %%mm1 \n\t" // 2H4
2847 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
2848 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
2850 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
2851 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
2852 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
2853 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
2855 "movq (%0, %1, 4), %%mm2 \n\t"
2856 "movq %%mm2, %%mm3 \n\t"
2857 "punpcklbw %%mm7, %%mm2 \n\t" // L7
2858 "punpckhbw %%mm7, %%mm3 \n\t" // H7
2860 "paddw %%mm2, %%mm2 \n\t" // 2L7
2861 "paddw %%mm3, %%mm3 \n\t" // 2H7
2862 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
2863 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
2865 "movq (%4), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2866 "movq 8(%4), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2869 "movq %%mm7, %%mm6 \n\t" // 0
2870 "psubw %%mm0, %%mm6 \n\t"
2871 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
2872 "movq %%mm7, %%mm6 \n\t" // 0
2873 "psubw %%mm1, %%mm6 \n\t"
2874 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
2875 "movq %%mm7, %%mm6 \n\t" // 0
2876 "psubw %%mm2, %%mm6 \n\t"
2877 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
2878 "movq %%mm7, %%mm6 \n\t" // 0
2879 "psubw %%mm3, %%mm6 \n\t"
2880 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
2882 "movq %%mm7, %%mm6 \n\t" // 0
2883 "pcmpgtw %%mm0, %%mm6 \n\t"
2884 "pxor %%mm6, %%mm0 \n\t"
2885 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
2886 "movq %%mm7, %%mm6 \n\t" // 0
2887 "pcmpgtw %%mm1, %%mm6 \n\t"
2888 "pxor %%mm6, %%mm1 \n\t"
2889 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
2890 "movq %%mm7, %%mm6 \n\t" // 0
2891 "pcmpgtw %%mm2, %%mm6 \n\t"
2892 "pxor %%mm6, %%mm2 \n\t"
2893 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
2894 "movq %%mm7, %%mm6 \n\t" // 0
2895 "pcmpgtw %%mm3, %%mm6 \n\t"
2896 "pxor %%mm6, %%mm3 \n\t"
2897 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
2901 "pminsw %%mm2, %%mm0 \n\t"
2902 "pminsw %%mm3, %%mm1 \n\t"
2904 "movq %%mm0, %%mm6 \n\t"
2905 "psubusw %%mm2, %%mm6 \n\t"
2906 "psubw %%mm6, %%mm0 \n\t"
2907 "movq %%mm1, %%mm6 \n\t"
2908 "psubusw %%mm3, %%mm6 \n\t"
2909 "psubw %%mm6, %%mm1 \n\t"
2912 "movd %2, %%mm2 \n\t" // QP
2913 "punpcklbw %%mm7, %%mm2 \n\t"
2915 "movq %%mm7, %%mm6 \n\t" // 0
2916 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
2917 "pxor %%mm6, %%mm4 \n\t"
2918 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
2919 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
2920 "pxor %%mm7, %%mm5 \n\t"
2921 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
2923 "psllw $3, %%mm2 \n\t" // 8QP
2924 "movq %%mm2, %%mm3 \n\t" // 8QP
2925 "pcmpgtw %%mm4, %%mm2 \n\t"
2926 "pcmpgtw %%mm5, %%mm3 \n\t"
2927 "pand %%mm2, %%mm4 \n\t"
2928 "pand %%mm3, %%mm5 \n\t"
2931 "psubusw %%mm0, %%mm4 \n\t" // hd
2932 "psubusw %%mm1, %%mm5 \n\t" // ld
2935 "movq "MANGLE(w05)", %%mm2 \n\t" // 5
2936 "pmullw %%mm2, %%mm4 \n\t"
2937 "pmullw %%mm2, %%mm5 \n\t"
2938 "movq "MANGLE(w20)", %%mm2 \n\t" // 32
2939 "paddw %%mm2, %%mm4 \n\t"
2940 "paddw %%mm2, %%mm5 \n\t"
2941 "psrlw $6, %%mm4 \n\t"
2942 "psrlw $6, %%mm5 \n\t"
2944 "movq 16(%4), %%mm0 \n\t" // L3 - L4
2945 "movq 24(%4), %%mm1 \n\t" // H3 - H4
2947 "pxor %%mm2, %%mm2 \n\t"
2948 "pxor %%mm3, %%mm3 \n\t"
2950 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
2951 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
2952 "pxor %%mm2, %%mm0 \n\t"
2953 "pxor %%mm3, %%mm1 \n\t"
2954 "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
2955 "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
2956 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
2957 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
2959 "pxor %%mm6, %%mm2 \n\t"
2960 "pxor %%mm7, %%mm3 \n\t"
2961 "pand %%mm2, %%mm4 \n\t"
2962 "pand %%mm3, %%mm5 \n\t"
2965 "pminsw %%mm0, %%mm4 \n\t"
2966 "pminsw %%mm1, %%mm5 \n\t"
2968 "movq %%mm4, %%mm2 \n\t"
2969 "psubusw %%mm0, %%mm2 \n\t"
2970 "psubw %%mm2, %%mm4 \n\t"
2971 "movq %%mm5, %%mm2 \n\t"
2972 "psubusw %%mm1, %%mm2 \n\t"
2973 "psubw %%mm2, %%mm5 \n\t"
2975 "pxor %%mm6, %%mm4 \n\t"
2976 "pxor %%mm7, %%mm5 \n\t"
2977 "psubw %%mm6, %%mm4 \n\t"
2978 "psubw %%mm7, %%mm5 \n\t"
2979 "packsswb %%mm5, %%mm4 \n\t"
2980 "movq %3, %%mm1 \n\t"
2981 "pandn %%mm4, %%mm1 \n\t"
2982 "movq (%0), %%mm0 \n\t"
2983 "paddb %%mm1, %%mm0 \n\t"
2984 "movq %%mm0, (%0) \n\t"
2985 "movq (%0, %1), %%mm0 \n\t"
2986 "psubb %%mm1, %%mm0 \n\t"
2987 "movq %%mm0, (%0, %1) \n\t"
2990 : "r" ((x86_reg)step), "m" (c->pQPb), "m"(eq_mask), "r"(tmp)
2995 STOP_TIMER("step16")
3003 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
3004 const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c);
3007 * Copy a block from src to dst and fixes the blacklevel.
3008 * levelFix == 0 -> do not touch the brighness & contrast
3010 #undef REAL_SCALED_CPY
3013 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, const uint8_t src[], int srcStride,
3014 int levelFix, int64_t *packedOffsetAndScale)
3022 "movq (%%"REG_a"), %%mm2 \n\t" // packedYOffset
3023 "movq 8(%%"REG_a"), %%mm3 \n\t" // packedYScale
3024 "lea (%2,%4), %%"REG_a" \n\t"
3025 "lea (%3,%5), %%"REG_d" \n\t"
3026 "pxor %%mm4, %%mm4 \n\t"
3028 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \
3029 "movq " #src1 ", %%mm0 \n\t"\
3030 "movq " #src1 ", %%mm5 \n\t"\
3031 "movq " #src2 ", %%mm1 \n\t"\
3032 "movq " #src2 ", %%mm6 \n\t"\
3033 "punpcklbw %%mm0, %%mm0 \n\t"\
3034 "punpckhbw %%mm5, %%mm5 \n\t"\
3035 "punpcklbw %%mm1, %%mm1 \n\t"\
3036 "punpckhbw %%mm6, %%mm6 \n\t"\
3037 "pmulhuw %%mm3, %%mm0 \n\t"\
3038 "pmulhuw %%mm3, %%mm5 \n\t"\
3039 "pmulhuw %%mm3, %%mm1 \n\t"\
3040 "pmulhuw %%mm3, %%mm6 \n\t"\
3041 "psubw %%mm2, %%mm0 \n\t"\
3042 "psubw %%mm2, %%mm5 \n\t"\
3043 "psubw %%mm2, %%mm1 \n\t"\
3044 "psubw %%mm2, %%mm6 \n\t"\
3045 "packuswb %%mm5, %%mm0 \n\t"\
3046 "packuswb %%mm6, %%mm1 \n\t"\
3047 "movq %%mm0, " #dst1 " \n\t"\
3048 "movq %%mm1, " #dst2 " \n\t"\
3051 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \
3052 "movq " #src1 ", %%mm0 \n\t"\
3053 "movq " #src1 ", %%mm5 \n\t"\
3054 "punpcklbw %%mm4, %%mm0 \n\t"\
3055 "punpckhbw %%mm4, %%mm5 \n\t"\
3056 "psubw %%mm2, %%mm0 \n\t"\
3057 "psubw %%mm2, %%mm5 \n\t"\
3058 "movq " #src2 ", %%mm1 \n\t"\
3059 "psllw $6, %%mm0 \n\t"\
3060 "psllw $6, %%mm5 \n\t"\
3061 "pmulhw %%mm3, %%mm0 \n\t"\
3062 "movq " #src2 ", %%mm6 \n\t"\
3063 "pmulhw %%mm3, %%mm5 \n\t"\
3064 "punpcklbw %%mm4, %%mm1 \n\t"\
3065 "punpckhbw %%mm4, %%mm6 \n\t"\
3066 "psubw %%mm2, %%mm1 \n\t"\
3067 "psubw %%mm2, %%mm6 \n\t"\
3068 "psllw $6, %%mm1 \n\t"\
3069 "psllw $6, %%mm6 \n\t"\
3070 "pmulhw %%mm3, %%mm1 \n\t"\
3071 "pmulhw %%mm3, %%mm6 \n\t"\
3072 "packuswb %%mm5, %%mm0 \n\t"\
3073 "packuswb %%mm6, %%mm1 \n\t"\
3074 "movq %%mm0, " #dst1 " \n\t"\
3075 "movq %%mm1, " #dst2 " \n\t"\
3078 #define SCALED_CPY(src1, src2, dst1, dst2)\
3079 REAL_SCALED_CPY(src1, src2, dst1, dst2)
3081 SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5))
3082 SCALED_CPY((%2, %4, 2), (%%REGa, %4, 2), (%3, %5, 2), (%%REGd, %5, 2))
3083 SCALED_CPY((%2, %4, 4), (%%REGa, %4, 4), (%3, %5, 4), (%%REGd, %5, 4))
3084 "lea (%%"REG_a",%4,4), %%"REG_a" \n\t"
3085 "lea (%%"REG_d",%5,4), %%"REG_d" \n\t"
3086 SCALED_CPY((%%REGa, %4), (%%REGa, %4, 2), (%%REGd, %5), (%%REGd, %5, 2))
3089 : "=&a" (packedOffsetAndScale)
3090 : "0" (packedOffsetAndScale),
3093 "r" ((x86_reg)srcStride),
3094 "r" ((x86_reg)dstStride)
3099 memcpy( &(dst[dstStride*i]),
3100 &(src[srcStride*i]), BLOCK_SIZE);
3105 "lea (%0,%2), %%"REG_a" \n\t"
3106 "lea (%1,%3), %%"REG_d" \n\t"
3108 #define REAL_SIMPLE_CPY(src1, src2, dst1, dst2) \
3109 "movq " #src1 ", %%mm0 \n\t"\
3110 "movq " #src2 ", %%mm1 \n\t"\
3111 "movq %%mm0, " #dst1 " \n\t"\
3112 "movq %%mm1, " #dst2 " \n\t"\
3114 #define SIMPLE_CPY(src1, src2, dst1, dst2)\
3115 REAL_SIMPLE_CPY(src1, src2, dst1, dst2)
3117 SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3))
3118 SIMPLE_CPY((%0, %2, 2), (%%REGa, %2, 2), (%1, %3, 2), (%%REGd, %3, 2))
3119 SIMPLE_CPY((%0, %2, 4), (%%REGa, %2, 4), (%1, %3, 4), (%%REGd, %3, 4))
3120 "lea (%%"REG_a",%2,4), %%"REG_a" \n\t"
3121 "lea (%%"REG_d",%3,4), %%"REG_d" \n\t"
3122 SIMPLE_CPY((%%REGa, %2), (%%REGa, %2, 2), (%%REGd, %3), (%%REGd, %3, 2))
3126 "r" ((x86_reg)srcStride),
3127 "r" ((x86_reg)dstStride)
3128 : "%"REG_a, "%"REG_d
3132 memcpy( &(dst[dstStride*i]),
3133 &(src[srcStride*i]), BLOCK_SIZE);
3139 * Duplicate the given 8 src pixels ? times upward
3141 static inline void RENAME(duplicate)(uint8_t src[], int stride)
3145 "movq (%0), %%mm0 \n\t"
3147 "movq %%mm0, (%0) \n\t"
3148 "movq %%mm0, (%0, %1) \n\t"
3149 "movq %%mm0, (%0, %1, 2) \n\t"
3151 : "r" ((x86_reg)-stride)
3164 * Filter array of bytes (Y or U or V values)
3166 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
3167 const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)
3169 DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access
3171 #ifdef COMPILE_TIME_MODE
3172 const int mode= COMPILE_TIME_MODE;
3174 const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;
3176 int black=0, white=255; // blackest black and whitest white in the picture
3177 int QPCorrecture= 256*256;
3184 const int qpHShift= isColor ? 4-c.hChromaSubSample : 4;
3185 const int qpVShift= isColor ? 4-c.vChromaSubSample : 4;
3188 uint64_t * const yHistogram= c.yHistogram;
3189 uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride;
3190 uint8_t * const tempDst= dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride;
3191 //const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
3194 for(i=0; i<57; i++){
3195 int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;
3196 int threshold= offset*2 + 1;
3197 c.mmxDcOffset[i]= 0x7F - offset;
3198 c.mmxDcThreshold[i]= 0x7F - threshold;
3199 c.mmxDcOffset[i]*= 0x0101010101010101LL;
3200 c.mmxDcThreshold[i]*= 0x0101010101010101LL;
3204 if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
3205 else if( (mode & LINEAR_BLEND_DEINT_FILTER)
3206 || (mode & FFMPEG_DEINT_FILTER)
3207 || (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14;
3208 else if( (mode & V_DEBLOCK)
3209 || (mode & LINEAR_IPOL_DEINT_FILTER)
3210 || (mode & MEDIAN_DEINT_FILTER)
3211 || (mode & V_A_DEBLOCK)) copyAhead=13;
3212 else if(mode & V_X1_FILTER) copyAhead=11;
3213 // else if(mode & V_RK1_FILTER) copyAhead=10;
3214 else if(mode & DERING) copyAhead=9;
3222 uint64_t maxClipped;
3227 // first frame is fscked so we ignore it
3228 if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256;
3230 for(i=0; i<256; i++){
3231 sum+= yHistogram[i];
3234 /* We always get a completely black picture first. */
3235 maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold);
3238 for(black=255; black>0; black--){
3239 if(clipped < maxClipped) break;
3240 clipped-= yHistogram[black];
3244 for(white=0; white<256; white++){
3245 if(clipped < maxClipped) break;
3246 clipped-= yHistogram[white];
3249 scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);
3252 c.packedYScale= (uint16_t)(scale*256.0 + 0.5);
3253 c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;
3255 c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);
3256 c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;
3259 c.packedYOffset|= c.packedYOffset<<32;
3260 c.packedYOffset|= c.packedYOffset<<16;
3262 c.packedYScale|= c.packedYScale<<32;
3263 c.packedYScale|= c.packedYScale<<16;
3265 if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5);
3266 else QPCorrecture= 256*256;
3268 c.packedYScale= 0x0100010001000100LL;
3270 QPCorrecture= 256*256;
3273 /* copy & deinterlace first row of blocks */
3276 const uint8_t *srcBlock= &(src[y*srcStride]);
3277 uint8_t *dstBlock= tempDst + dstStride;
3279 // From this point on it is guaranteed that we can read and write 16 lines downward
3280 // finish 1 block before the next otherwise we might have a problem
3281 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3282 for(x=0; x<width; x+=BLOCK_SIZE){
3286 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3287 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3288 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3289 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3293 "mov %4, %%"REG_a" \n\t"
3294 "shr $2, %%"REG_a" \n\t"
3295 "and $6, %%"REG_a" \n\t"
3296 "add %5, %%"REG_a" \n\t"
3297 "mov %%"REG_a", %%"REG_d" \n\t"
3298 "imul %1, %%"REG_a" \n\t"
3299 "imul %3, %%"REG_d" \n\t"
3300 "prefetchnta 32(%%"REG_a", %0) \n\t"
3301 "prefetcht0 32(%%"REG_d", %2) \n\t"
3302 "add %1, %%"REG_a" \n\t"
3303 "add %3, %%"REG_d" \n\t"
3304 "prefetchnta 32(%%"REG_a", %0) \n\t"
3305 "prefetcht0 32(%%"REG_d", %2) \n\t"
3306 :: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride),
3307 "g" ((x86_reg)x), "g" ((x86_reg)copyAhead)
3308 : "%"REG_a, "%"REG_d
3312 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3313 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3314 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3315 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3316 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3320 RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
3321 srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3323 RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
3325 if(mode & LINEAR_IPOL_DEINT_FILTER)
3326 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3327 else if(mode & LINEAR_BLEND_DEINT_FILTER)
3328 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3329 else if(mode & MEDIAN_DEINT_FILTER)
3330 RENAME(deInterlaceMedian)(dstBlock, dstStride);
3331 else if(mode & CUBIC_IPOL_DEINT_FILTER)
3332 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3333 else if(mode & FFMPEG_DEINT_FILTER)
3334 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3335 else if(mode & LOWPASS5_DEINT_FILTER)
3336 RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3337 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3338 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3343 if(width==FFABS(dstStride))
3344 linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride);
3347 for(i=0; i<copyAhead; i++){
3348 memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width);
3353 for(y=0; y<height; y+=BLOCK_SIZE){
3354 //1% speedup if these are here instead of the inner loop
3355 const uint8_t *srcBlock= &(src[y*srcStride]);
3356 uint8_t *dstBlock= &(dst[y*dstStride]);
3358 uint8_t *tempBlock1= c.tempBlocks;
3359 uint8_t *tempBlock2= c.tempBlocks + 8;
3361 const int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
3362 int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)];
3364 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
3365 if not than use a temporary buffer */
3368 /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
3369 blockcopy to dst later */
3370 linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
3371 FFMAX(height-y-copyAhead, 0), srcStride);
3373 /* duplicate last line of src to fill the void up to line (copyAhead+7) */
3374 for(i=FFMAX(height-y, 8); i<copyAhead+8; i++)
3375 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), FFABS(srcStride));
3377 /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
3378 linecpy(tempDst, dstBlock - dstStride, FFMIN(height-y+1, copyAhead+1), dstStride);
3380 /* duplicate last line of dst to fill the void up to line (copyAhead) */
3381 for(i=height-y+1; i<=copyAhead; i++)
3382 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), FFABS(dstStride));
3384 dstBlock= tempDst + dstStride;
3388 // From this point on it is guaranteed that we can read and write 16 lines downward
3389 // finish 1 block before the next otherwise we might have a problem
3390 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3391 for(x=0; x<width; x+=BLOCK_SIZE){
3392 const int stride= dstStride;
3397 QP= QPptr[x>>qpHShift];
3398 c.nonBQP= nonBQPptr[x>>qpHShift];
3401 QP= (QP* QPCorrecture + 256*128)>>16;
3402 c.nonBQP= nonBQPptr[x>>4];
3403 c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
3404 yHistogram[ srcBlock[srcStride*12 + 4] ]++;
3409 "movd %1, %%mm7 \n\t"
3410 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
3411 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
3412 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
3413 "movq %%mm7, %0 \n\t"
3422 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3423 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3424 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3425 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3429 "mov %4, %%"REG_a" \n\t"
3430 "shr $2, %%"REG_a" \n\t"
3431 "and $6, %%"REG_a" \n\t"
3432 "add %5, %%"REG_a" \n\t"
3433 "mov %%"REG_a", %%"REG_d" \n\t"
3434 "imul %1, %%"REG_a" \n\t"
3435 "imul %3, %%"REG_d" \n\t"
3436 "prefetchnta 32(%%"REG_a", %0) \n\t"
3437 "prefetcht0 32(%%"REG_d", %2) \n\t"
3438 "add %1, %%"REG_a" \n\t"
3439 "add %3, %%"REG_d" \n\t"
3440 "prefetchnta 32(%%"REG_a", %0) \n\t"
3441 "prefetcht0 32(%%"REG_d", %2) \n\t"
3442 :: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride),
3443 "g" ((x86_reg)x), "g" ((x86_reg)copyAhead)
3444 : "%"REG_a, "%"REG_d
3448 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3449 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3450 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3451 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3452 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3456 RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
3457 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3459 if(mode & LINEAR_IPOL_DEINT_FILTER)
3460 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3461 else if(mode & LINEAR_BLEND_DEINT_FILTER)
3462 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3463 else if(mode & MEDIAN_DEINT_FILTER)
3464 RENAME(deInterlaceMedian)(dstBlock, dstStride);
3465 else if(mode & CUBIC_IPOL_DEINT_FILTER)
3466 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3467 else if(mode & FFMPEG_DEINT_FILTER)
3468 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3469 else if(mode & LOWPASS5_DEINT_FILTER)
3470 RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3471 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3472 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3475 /* only deblock if we have 2 blocks */
3477 if(mode & V_X1_FILTER)
3478 RENAME(vertX1Filter)(dstBlock, stride, &c);
3479 else if(mode & V_DEBLOCK){
3480 const int t= RENAME(vertClassify)(dstBlock, stride, &c);
3483 RENAME(doVertLowPass)(dstBlock, stride, &c);
3485 RENAME(doVertDefFilter)(dstBlock, stride, &c);
3486 }else if(mode & V_A_DEBLOCK){
3487 RENAME(do_a_deblock)(dstBlock, stride, 1, &c);
3492 RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
3494 /* check if we have a previous block to deblock it with dstBlock */
3497 if(mode & H_X1_FILTER)
3498 RENAME(vertX1Filter)(tempBlock1, 16, &c);
3499 else if(mode & H_DEBLOCK){
3501 const int t= RENAME(vertClassify)(tempBlock1, 16, &c);
3502 //STOP_TIMER("dc & minmax")
3504 RENAME(doVertLowPass)(tempBlock1, 16, &c);
3506 RENAME(doVertDefFilter)(tempBlock1, 16, &c);
3507 }else if(mode & H_A_DEBLOCK){
3508 RENAME(do_a_deblock)(tempBlock1, 16, 1, &c);
3511 RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
3514 if(mode & H_X1_FILTER)
3515 horizX1Filter(dstBlock-4, stride, QP);
3516 else if(mode & H_DEBLOCK){
3518 DECLARE_ALIGNED(16, unsigned char, tempBlock)[272];
3520 transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride);
3522 t = vertClassify_altivec(tempBlock-48, 16, &c);
3524 doVertLowPass_altivec(tempBlock-48, 16, &c);
3525 transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3528 doVertDefFilter_altivec(tempBlock-48, 16, &c);
3529 transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3532 const int t= RENAME(horizClassify)(dstBlock-4, stride, &c);
3535 RENAME(doHorizLowPass)(dstBlock-4, stride, &c);
3537 RENAME(doHorizDefFilter)(dstBlock-4, stride, &c);
3539 }else if(mode & H_A_DEBLOCK){
3540 RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c);
3544 //FIXME filter first line
3545 if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);
3548 if(mode & TEMP_NOISE_FILTER)
3550 RENAME(tempNoiseReducer)(dstBlock-8, stride,
3551 c.tempBlurred[isColor] + y*dstStride + x,
3552 c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3),
3553 c.ppMode.maxTmpNoise);
3561 tmpXchg= tempBlock1;
3562 tempBlock1= tempBlock2;
3563 tempBlock2 = tmpXchg;
3568 if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);
3571 if((mode & TEMP_NOISE_FILTER)){
3572 RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
3573 c.tempBlurred[isColor] + y*dstStride + x,
3574 c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3),
3575 c.ppMode.maxTmpNoise);
3578 /* did we use a tmp buffer for the last lines*/
3580 uint8_t *dstBlock= &(dst[y*dstStride]);
3581 if(width==FFABS(dstStride))
3582 linecpy(dstBlock, tempDst + dstStride, height-y, dstStride);
3585 for(i=0; i<height-y; i++){
3586 memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width);
3591 for(x=0; x<width; x+=32){
3593 i+= dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
3594 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
3595 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
3596 + dstBlock[x +13*dstStride]
3597 + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
3601 __asm__ volatile("femms");
3603 __asm__ volatile("emms");
3606 #ifdef DEBUG_BRIGHTNESS
3610 for(i=0; i<256; i++)
3611 if(yHistogram[i] > max) max=yHistogram[i];
3613 for(i=1; i<256; i++){
3615 int start=yHistogram[i-1]/(max/256+1);
3616 int end=yHistogram[i]/(max/256+1);
3617 int inc= end > start ? 1 : -1;
3618 for(x=start; x!=end+inc; x+=inc)
3619 dst[ i*dstStride + x]+=128;
3622 for(i=0; i<100; i+=2){
3623 dst[ (white)*dstStride + i]+=128;
3624 dst[ (black)*dstStride + i]+=128;
3629 *c2= c; //copy local context back