FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
postprocess_template.c
Go to the documentation of this file.
1 /*
2  * Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at)
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License as published by
8  * the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20 
21 /**
22  * @file
23  * mmx/mmx2/3dnow postprocess code.
24  */
25 
26 #include "libavutil/x86/asm.h"
27 
28 /* A single TEMPLATE_PP_* should be defined (to 1) when this template is
29  * included. The following macros will define its dependencies to 1 as well
30  * (like MMX2 depending on MMX), and will define to 0 all the others. Every
31  * TEMPLATE_PP_* need to be undef at the end. */
32 
33 #ifdef TEMPLATE_PP_C
34 # define RENAME(a) a ## _C
35 #else
36 # define TEMPLATE_PP_C 0
37 #endif
38 
39 #ifdef TEMPLATE_PP_ALTIVEC
40 # define RENAME(a) a ## _altivec
41 #else
42 # define TEMPLATE_PP_ALTIVEC 0
43 #endif
44 
45 #ifdef TEMPLATE_PP_MMX
46 # define RENAME(a) a ## _MMX
47 #else
48 # define TEMPLATE_PP_MMX 0
49 #endif
50 
51 #ifdef TEMPLATE_PP_MMXEXT
52 # undef TEMPLATE_PP_MMX
53 # define TEMPLATE_PP_MMX 1
54 # define RENAME(a) a ## _MMX2
55 #else
56 # define TEMPLATE_PP_MMXEXT 0
57 #endif
58 
59 #ifdef TEMPLATE_PP_3DNOW
60 # undef TEMPLATE_PP_MMX
61 # define TEMPLATE_PP_MMX 1
62 # define RENAME(a) a ## _3DNow
63 #else
64 # define TEMPLATE_PP_3DNOW 0
65 #endif
66 
67 #ifdef TEMPLATE_PP_SSE2
68 # undef TEMPLATE_PP_MMX
69 # define TEMPLATE_PP_MMX 1
70 # undef TEMPLATE_PP_MMXEXT
71 # define TEMPLATE_PP_MMXEXT 1
72 # define RENAME(a) a ## _SSE2
73 #else
74 # define TEMPLATE_PP_SSE2 0
75 #endif
76 
77 #undef REAL_PAVGB
78 #undef PAVGB
79 #undef PMINUB
80 #undef PMAXUB
81 
82 #if TEMPLATE_PP_MMXEXT
83 #define REAL_PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
84 #elif TEMPLATE_PP_3DNOW
85 #define REAL_PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
86 #endif
87 #define PAVGB(a,b) REAL_PAVGB(a,b)
88 
89 #if TEMPLATE_PP_MMXEXT
90 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
91 #elif TEMPLATE_PP_MMX
92 #define PMINUB(b,a,t) \
93  "movq " #a ", " #t " \n\t"\
94  "psubusb " #b ", " #t " \n\t"\
95  "psubb " #t ", " #a " \n\t"
96 #endif
97 
98 #if TEMPLATE_PP_MMXEXT
99 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
100 #elif TEMPLATE_PP_MMX
101 #define PMAXUB(a,b) \
102  "psubusb " #a ", " #b " \n\t"\
103  "paddb " #a ", " #b " \n\t"
104 #endif
105 
106 //FIXME? |255-0| = 1 (should not be a problem ...)
107 #if TEMPLATE_PP_MMX
108 /**
109  * Check if the middle 8x8 Block in the given 8x16 block is flat
110  */
111 static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){
112  int numEq= 0, dcOk;
113  src+= stride*4; // src points to begin of the 8x8 Block
114  __asm__ volatile(
115  "movq %0, %%mm7 \n\t"
116  "movq %1, %%mm6 \n\t"
117  : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
118  );
119 
120  __asm__ volatile(
121  "lea (%2, %3), %%"REG_a" \n\t"
122 // 0 1 2 3 4 5 6 7 8 9
123 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
124 
125  "movq (%2), %%mm0 \n\t"
126  "movq (%%"REG_a"), %%mm1 \n\t"
127  "movq %%mm0, %%mm3 \n\t"
128  "movq %%mm0, %%mm4 \n\t"
129  PMAXUB(%%mm1, %%mm4)
130  PMINUB(%%mm1, %%mm3, %%mm5)
131  "psubb %%mm1, %%mm0 \n\t" // mm0 = difference
132  "paddb %%mm7, %%mm0 \n\t"
133  "pcmpgtb %%mm6, %%mm0 \n\t"
134 
135  "movq (%%"REG_a",%3), %%mm2 \n\t"
136  PMAXUB(%%mm2, %%mm4)
137  PMINUB(%%mm2, %%mm3, %%mm5)
138  "psubb %%mm2, %%mm1 \n\t"
139  "paddb %%mm7, %%mm1 \n\t"
140  "pcmpgtb %%mm6, %%mm1 \n\t"
141  "paddb %%mm1, %%mm0 \n\t"
142 
143  "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
144  PMAXUB(%%mm1, %%mm4)
145  PMINUB(%%mm1, %%mm3, %%mm5)
146  "psubb %%mm1, %%mm2 \n\t"
147  "paddb %%mm7, %%mm2 \n\t"
148  "pcmpgtb %%mm6, %%mm2 \n\t"
149  "paddb %%mm2, %%mm0 \n\t"
150 
151  "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
152 
153  "movq (%2, %3, 4), %%mm2 \n\t"
154  PMAXUB(%%mm2, %%mm4)
155  PMINUB(%%mm2, %%mm3, %%mm5)
156  "psubb %%mm2, %%mm1 \n\t"
157  "paddb %%mm7, %%mm1 \n\t"
158  "pcmpgtb %%mm6, %%mm1 \n\t"
159  "paddb %%mm1, %%mm0 \n\t"
160 
161  "movq (%%"REG_a"), %%mm1 \n\t"
162  PMAXUB(%%mm1, %%mm4)
163  PMINUB(%%mm1, %%mm3, %%mm5)
164  "psubb %%mm1, %%mm2 \n\t"
165  "paddb %%mm7, %%mm2 \n\t"
166  "pcmpgtb %%mm6, %%mm2 \n\t"
167  "paddb %%mm2, %%mm0 \n\t"
168 
169  "movq (%%"REG_a", %3), %%mm2 \n\t"
170  PMAXUB(%%mm2, %%mm4)
171  PMINUB(%%mm2, %%mm3, %%mm5)
172  "psubb %%mm2, %%mm1 \n\t"
173  "paddb %%mm7, %%mm1 \n\t"
174  "pcmpgtb %%mm6, %%mm1 \n\t"
175  "paddb %%mm1, %%mm0 \n\t"
176 
177  "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
178  PMAXUB(%%mm1, %%mm4)
179  PMINUB(%%mm1, %%mm3, %%mm5)
180  "psubb %%mm1, %%mm2 \n\t"
181  "paddb %%mm7, %%mm2 \n\t"
182  "pcmpgtb %%mm6, %%mm2 \n\t"
183  "paddb %%mm2, %%mm0 \n\t"
184  "psubusb %%mm3, %%mm4 \n\t"
185 
186  " \n\t"
187 #if TEMPLATE_PP_MMXEXT
188  "pxor %%mm7, %%mm7 \n\t"
189  "psadbw %%mm7, %%mm0 \n\t"
190 #else
191  "movq %%mm0, %%mm1 \n\t"
192  "psrlw $8, %%mm0 \n\t"
193  "paddb %%mm1, %%mm0 \n\t"
194  "movq %%mm0, %%mm1 \n\t"
195  "psrlq $16, %%mm0 \n\t"
196  "paddb %%mm1, %%mm0 \n\t"
197  "movq %%mm0, %%mm1 \n\t"
198  "psrlq $32, %%mm0 \n\t"
199  "paddb %%mm1, %%mm0 \n\t"
200 #endif
201  "movq %4, %%mm7 \n\t" // QP,..., QP
202  "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
203  "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0
204  "packssdw %%mm4, %%mm4 \n\t"
205  "movd %%mm0, %0 \n\t"
206  "movd %%mm4, %1 \n\t"
207 
208  : "=r" (numEq), "=r" (dcOk)
209  : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
210  : "%"REG_a
211  );
212 
213  numEq= (-numEq) &0xFF;
214  if(numEq > c->ppMode.flatnessThreshold){
215  if(dcOk) return 0;
216  else return 1;
217  }else{
218  return 2;
219  }
220 }
221 #endif //TEMPLATE_PP_MMX
222 
223 /**
224  * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
225  * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
226  */
227 #if !TEMPLATE_PP_ALTIVEC
228 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
229 {
230 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
231  src+= stride*3;
232  __asm__ volatile( //"movv %0 %1 %2\n\t"
233  "movq %2, %%mm0 \n\t" // QP,..., QP
234  "pxor %%mm4, %%mm4 \n\t"
235 
236  "movq (%0), %%mm6 \n\t"
237  "movq (%0, %1), %%mm5 \n\t"
238  "movq %%mm5, %%mm1 \n\t"
239  "movq %%mm6, %%mm2 \n\t"
240  "psubusb %%mm6, %%mm5 \n\t"
241  "psubusb %%mm1, %%mm2 \n\t"
242  "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
243  "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
244  "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
245 
246  "pand %%mm2, %%mm6 \n\t"
247  "pandn %%mm1, %%mm2 \n\t"
248  "por %%mm2, %%mm6 \n\t"// First Line to Filter
249 
250  "movq (%0, %1, 8), %%mm5 \n\t"
251  "lea (%0, %1, 4), %%"REG_a" \n\t"
252  "lea (%0, %1, 8), %%"REG_c" \n\t"
253  "sub %1, %%"REG_c" \n\t"
254  "add %1, %0 \n\t" // %0 points to line 1 not 0
255  "movq (%0, %1, 8), %%mm7 \n\t"
256  "movq %%mm5, %%mm1 \n\t"
257  "movq %%mm7, %%mm2 \n\t"
258  "psubusb %%mm7, %%mm5 \n\t"
259  "psubusb %%mm1, %%mm2 \n\t"
260  "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
261  "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
262  "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
263 
264  "pand %%mm2, %%mm7 \n\t"
265  "pandn %%mm1, %%mm2 \n\t"
266  "por %%mm2, %%mm7 \n\t" // First Line to Filter
267 
268 
269  // 1 2 3 4 5 6 7 8
270  // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
271  // 6 4 2 2 1 1
272  // 6 4 4 2
273  // 6 8 2
274 
275  "movq (%0, %1), %%mm0 \n\t" // 1
276  "movq %%mm0, %%mm1 \n\t" // 1
277  PAVGB(%%mm6, %%mm0) //1 1 /2
278  PAVGB(%%mm6, %%mm0) //3 1 /4
279 
280  "movq (%0, %1, 4), %%mm2 \n\t" // 1
281  "movq %%mm2, %%mm5 \n\t" // 1
282  PAVGB((%%REGa), %%mm2) // 11 /2
283  PAVGB((%0, %1, 2), %%mm2) // 211 /4
284  "movq %%mm2, %%mm3 \n\t" // 211 /4
285  "movq (%0), %%mm4 \n\t" // 1
286  PAVGB(%%mm4, %%mm3) // 4 211 /8
287  PAVGB(%%mm0, %%mm3) //642211 /16
288  "movq %%mm3, (%0) \n\t" // X
289  // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
290  "movq %%mm1, %%mm0 \n\t" // 1
291  PAVGB(%%mm6, %%mm0) //1 1 /2
292  "movq %%mm4, %%mm3 \n\t" // 1
293  PAVGB((%0,%1,2), %%mm3) // 1 1 /2
294  PAVGB((%%REGa,%1,2), %%mm5) // 11 /2
295  PAVGB((%%REGa), %%mm5) // 211 /4
296  PAVGB(%%mm5, %%mm3) // 2 2211 /8
297  PAVGB(%%mm0, %%mm3) //4242211 /16
298  "movq %%mm3, (%0,%1) \n\t" // X
299  // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
300  PAVGB(%%mm4, %%mm6) //11 /2
301  "movq (%%"REG_c"), %%mm0 \n\t" // 1
302  PAVGB((%%REGa, %1, 2), %%mm0) // 11/2
303  "movq %%mm0, %%mm3 \n\t" // 11/2
304  PAVGB(%%mm1, %%mm0) // 2 11/4
305  PAVGB(%%mm6, %%mm0) //222 11/8
306  PAVGB(%%mm2, %%mm0) //22242211/16
307  "movq (%0, %1, 2), %%mm2 \n\t" // 1
308  "movq %%mm0, (%0, %1, 2) \n\t" // X
309  // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
310  "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
311  PAVGB((%%REGc), %%mm0) // 11 /2
312  PAVGB(%%mm0, %%mm6) //11 11 /4
313  PAVGB(%%mm1, %%mm4) // 11 /2
314  PAVGB(%%mm2, %%mm1) // 11 /2
315  PAVGB(%%mm1, %%mm6) //1122 11 /8
316  PAVGB(%%mm5, %%mm6) //112242211 /16
317  "movq (%%"REG_a"), %%mm5 \n\t" // 1
318  "movq %%mm6, (%%"REG_a") \n\t" // X
319  // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
320  "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1
321  PAVGB(%%mm7, %%mm6) // 11 /2
322  PAVGB(%%mm4, %%mm6) // 11 11 /4
323  PAVGB(%%mm3, %%mm6) // 11 2211 /8
324  PAVGB(%%mm5, %%mm2) // 11 /2
325  "movq (%0, %1, 4), %%mm4 \n\t" // 1
326  PAVGB(%%mm4, %%mm2) // 112 /4
327  PAVGB(%%mm2, %%mm6) // 112242211 /16
328  "movq %%mm6, (%0, %1, 4) \n\t" // X
329  // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
330  PAVGB(%%mm7, %%mm1) // 11 2 /4
331  PAVGB(%%mm4, %%mm5) // 11 /2
332  PAVGB(%%mm5, %%mm0) // 11 11 /4
333  "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1
334  PAVGB(%%mm6, %%mm1) // 11 4 2 /8
335  PAVGB(%%mm0, %%mm1) // 11224222 /16
336  "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X
337  // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
338  PAVGB((%%REGc), %%mm2) // 112 4 /8
339  "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
340  PAVGB(%%mm0, %%mm6) // 1 1 /2
341  PAVGB(%%mm7, %%mm6) // 1 12 /4
342  PAVGB(%%mm2, %%mm6) // 1122424 /4
343  "movq %%mm6, (%%"REG_c") \n\t" // X
344  // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
345  PAVGB(%%mm7, %%mm5) // 11 2 /4
346  PAVGB(%%mm7, %%mm5) // 11 6 /8
347 
348  PAVGB(%%mm3, %%mm0) // 112 /4
349  PAVGB(%%mm0, %%mm5) // 112246 /16
350  "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X
351  "sub %1, %0 \n\t"
352 
353  :
354  : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
355  : "%"REG_a, "%"REG_c
356  );
357 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
358  const int l1= stride;
359  const int l2= stride + l1;
360  const int l3= stride + l2;
361  const int l4= stride + l3;
362  const int l5= stride + l4;
363  const int l6= stride + l5;
364  const int l7= stride + l6;
365  const int l8= stride + l7;
366  const int l9= stride + l8;
367  int x;
368  src+= stride*3;
369  for(x=0; x<BLOCK_SIZE; x++){
370  const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
371  const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
372 
373  int sums[10];
374  sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
375  sums[1] = sums[0] - first + src[l4];
376  sums[2] = sums[1] - first + src[l5];
377  sums[3] = sums[2] - first + src[l6];
378  sums[4] = sums[3] - first + src[l7];
379  sums[5] = sums[4] - src[l1] + src[l8];
380  sums[6] = sums[5] - src[l2] + last;
381  sums[7] = sums[6] - src[l3] + last;
382  sums[8] = sums[7] - src[l4] + last;
383  sums[9] = sums[8] - src[l5] + last;
384 
385  src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
386  src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
387  src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
388  src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
389  src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
390  src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
391  src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
392  src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
393 
394  src++;
395  }
396 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
397 }
398 #endif //TEMPLATE_PP_ALTIVEC
399 
400 /**
401  * Experimental Filter 1
402  * will not damage linear gradients
403  * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
404  * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
405  * MMX2 version does correct clipping C version does not
406  */
407 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
408 {
409 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
410  src+= stride*3;
411 
412  __asm__ volatile(
413  "pxor %%mm7, %%mm7 \n\t" // 0
414  "lea (%0, %1), %%"REG_a" \n\t"
415  "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
416 // 0 1 2 3 4 5 6 7 8 9
417 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
418  "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
419  "movq (%0, %1, 4), %%mm1 \n\t" // line 4
420  "movq %%mm1, %%mm2 \n\t" // line 4
421  "psubusb %%mm0, %%mm1 \n\t"
422  "psubusb %%mm2, %%mm0 \n\t"
423  "por %%mm1, %%mm0 \n\t" // |l2 - l3|
424  "movq (%%"REG_c"), %%mm3 \n\t" // line 5
425  "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6
426  "movq %%mm3, %%mm5 \n\t" // line 5
427  "psubusb %%mm4, %%mm3 \n\t"
428  "psubusb %%mm5, %%mm4 \n\t"
429  "por %%mm4, %%mm3 \n\t" // |l5 - l6|
430  PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
431  "movq %%mm2, %%mm1 \n\t" // line 4
432  "psubusb %%mm5, %%mm2 \n\t"
433  "movq %%mm2, %%mm4 \n\t"
434  "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
435  "psubusb %%mm1, %%mm5 \n\t"
436  "por %%mm5, %%mm4 \n\t" // |l4 - l5|
437  "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
438  "movq %%mm4, %%mm3 \n\t" // d
439  "movq %2, %%mm0 \n\t"
440  "paddusb %%mm0, %%mm0 \n\t"
441  "psubusb %%mm0, %%mm4 \n\t"
442  "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
443  "psubusb "MANGLE(b01)", %%mm3 \n\t"
444  "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
445 
446  PAVGB(%%mm7, %%mm3) // d/2
447  "movq %%mm3, %%mm1 \n\t" // d/2
448  PAVGB(%%mm7, %%mm3) // d/4
449  PAVGB(%%mm1, %%mm3) // 3*d/8
450 
451  "movq (%0, %1, 4), %%mm0 \n\t" // line 4
452  "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
453  "psubusb %%mm3, %%mm0 \n\t"
454  "pxor %%mm2, %%mm0 \n\t"
455  "movq %%mm0, (%0, %1, 4) \n\t" // line 4
456 
457  "movq (%%"REG_c"), %%mm0 \n\t" // line 5
458  "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
459  "paddusb %%mm3, %%mm0 \n\t"
460  "pxor %%mm2, %%mm0 \n\t"
461  "movq %%mm0, (%%"REG_c") \n\t" // line 5
462 
463  PAVGB(%%mm7, %%mm1) // d/4
464 
465  "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
466  "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
467  "psubusb %%mm1, %%mm0 \n\t"
468  "pxor %%mm2, %%mm0 \n\t"
469  "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3
470 
471  "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6
472  "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
473  "paddusb %%mm1, %%mm0 \n\t"
474  "pxor %%mm2, %%mm0 \n\t"
475  "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6
476 
477  PAVGB(%%mm7, %%mm1) // d/8
478 
479  "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2
480  "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
481  "psubusb %%mm1, %%mm0 \n\t"
482  "pxor %%mm2, %%mm0 \n\t"
483  "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2
484 
485  "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7
486  "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
487  "paddusb %%mm1, %%mm0 \n\t"
488  "pxor %%mm2, %%mm0 \n\t"
489  "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7
490 
491  :
492  : "r" (src), "r" ((x86_reg)stride), "m" (co->pQPb)
493  : "%"REG_a, "%"REG_c
494  );
495 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
496 
497  const int l1= stride;
498  const int l2= stride + l1;
499  const int l3= stride + l2;
500  const int l4= stride + l3;
501  const int l5= stride + l4;
502  const int l6= stride + l5;
503  const int l7= stride + l6;
504 // const int l8= stride + l7;
505 // const int l9= stride + l8;
506  int x;
507 
508  src+= stride*3;
509  for(x=0; x<BLOCK_SIZE; x++){
510  int a= src[l3] - src[l4];
511  int b= src[l4] - src[l5];
512  int c= src[l5] - src[l6];
513 
514  int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
515  d= FFMAX(d, 0);
516 
517  if(d < co->QP*2){
518  int v = d * FFSIGN(-b);
519 
520  src[l2] +=v>>3;
521  src[l3] +=v>>2;
522  src[l4] +=(3*v)>>3;
523  src[l5] -=(3*v)>>3;
524  src[l6] -=v>>2;
525  src[l7] -=v>>3;
526  }
527  src++;
528  }
529 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
530 }
531 
532 #if !TEMPLATE_PP_ALTIVEC
533 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
534 {
535 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
536 /*
537  uint8_t tmp[16];
538  const int l1= stride;
539  const int l2= stride + l1;
540  const int l3= stride + l2;
541  const int l4= (int)tmp - (int)src - stride*3;
542  const int l5= (int)tmp - (int)src - stride*3 + 8;
543  const int l6= stride*3 + l3;
544  const int l7= stride + l6;
545  const int l8= stride + l7;
546 
547  memcpy(tmp, src+stride*7, 8);
548  memcpy(tmp+8, src+stride*8, 8);
549 */
550  src+= stride*4;
551  __asm__ volatile(
552 
553 #if 0 //slightly more accurate and slightly slower
554  "pxor %%mm7, %%mm7 \n\t" // 0
555  "lea (%0, %1), %%"REG_a" \n\t"
556  "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
557 // 0 1 2 3 4 5 6 7
558 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
559 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
560 
561 
562  "movq (%0, %1, 2), %%mm0 \n\t" // l2
563  "movq (%0), %%mm1 \n\t" // l0
564  "movq %%mm0, %%mm2 \n\t" // l2
565  PAVGB(%%mm7, %%mm0) // ~l2/2
566  PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
567  PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
568 
569  "movq (%%"REG_a"), %%mm1 \n\t" // l1
570  "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3
571  "movq %%mm1, %%mm4 \n\t" // l1
572  PAVGB(%%mm7, %%mm1) // ~l1/2
573  PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
574  PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
575 
576  "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
577  "psubusb %%mm1, %%mm0 \n\t"
578  "psubusb %%mm4, %%mm1 \n\t"
579  "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
580 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
581 
582  "movq (%0, %1, 4), %%mm0 \n\t" // l4
583  "movq %%mm0, %%mm4 \n\t" // l4
584  PAVGB(%%mm7, %%mm0) // ~l4/2
585  PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
586  PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
587 
588  "movq (%%"REG_c"), %%mm2 \n\t" // l5
589  "movq %%mm3, %%mm5 \n\t" // l3
590  PAVGB(%%mm7, %%mm3) // ~l3/2
591  PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
592  PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
593 
594  "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
595  "psubusb %%mm3, %%mm0 \n\t"
596  "psubusb %%mm6, %%mm3 \n\t"
597  "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
598  "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
599 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
600 
601  "movq (%%"REG_c", %1), %%mm6 \n\t" // l6
602  "movq %%mm6, %%mm5 \n\t" // l6
603  PAVGB(%%mm7, %%mm6) // ~l6/2
604  PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
605  PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
606 
607  "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7
608  "movq %%mm2, %%mm4 \n\t" // l5
609  PAVGB(%%mm7, %%mm2) // ~l5/2
610  PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
611  PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
612 
613  "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
614  "psubusb %%mm2, %%mm6 \n\t"
615  "psubusb %%mm4, %%mm2 \n\t"
616  "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
617 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
618 
619 
620  PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
621  "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
622  "paddusb "MANGLE(b01)", %%mm4 \n\t"
623  "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
624  "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
625  "pand %%mm4, %%mm3 \n\t"
626 
627  "movq %%mm3, %%mm1 \n\t"
628 // "psubusb "MANGLE(b01)", %%mm3 \n\t"
629  PAVGB(%%mm7, %%mm3)
630  PAVGB(%%mm7, %%mm3)
631  "paddusb %%mm1, %%mm3 \n\t"
632 // "paddusb "MANGLE(b01)", %%mm3 \n\t"
633 
634  "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3
635  "movq (%0, %1, 4), %%mm5 \n\t" //l4
636  "movq (%0, %1, 4), %%mm4 \n\t" //l4
637  "psubusb %%mm6, %%mm5 \n\t"
638  "psubusb %%mm4, %%mm6 \n\t"
639  "por %%mm6, %%mm5 \n\t" // |l3-l4|
640  "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
641  "pxor %%mm6, %%mm0 \n\t"
642  "pand %%mm0, %%mm3 \n\t"
643  PMINUB(%%mm5, %%mm3, %%mm0)
644 
645  "psubusb "MANGLE(b01)", %%mm3 \n\t"
646  PAVGB(%%mm7, %%mm3)
647 
648  "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
649  "movq (%0, %1, 4), %%mm2 \n\t"
650  "pxor %%mm6, %%mm0 \n\t"
651  "pxor %%mm6, %%mm2 \n\t"
652  "psubb %%mm3, %%mm0 \n\t"
653  "paddb %%mm3, %%mm2 \n\t"
654  "pxor %%mm6, %%mm0 \n\t"
655  "pxor %%mm6, %%mm2 \n\t"
656  "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
657  "movq %%mm2, (%0, %1, 4) \n\t"
658 #endif //0
659 
660  "lea (%0, %1), %%"REG_a" \n\t"
661  "pcmpeqb %%mm6, %%mm6 \n\t" // -1
662 // 0 1 2 3 4 5 6 7
663 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
664 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
665 
666 
667  "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3
668  "movq (%0, %1, 4), %%mm0 \n\t" // l4
669  "pxor %%mm6, %%mm1 \n\t" // -l3-1
670  PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
671 // mm1=-l3-1, mm0=128-q
672 
673  "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5
674  "movq (%%"REG_a", %1), %%mm3 \n\t" // l2
675  "pxor %%mm6, %%mm2 \n\t" // -l5-1
676  "movq %%mm2, %%mm5 \n\t" // -l5-1
677  "movq "MANGLE(b80)", %%mm4 \n\t" // 128
678  "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
679  PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
680  PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
681  PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
682  PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
683 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
684 
685  "movq (%%"REG_a"), %%mm2 \n\t" // l1
686  "pxor %%mm6, %%mm2 \n\t" // -l1-1
687  PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
688  PAVGB((%0), %%mm1) // (l0-l3+256)/2
689  "movq "MANGLE(b80)", %%mm3 \n\t" // 128
690  PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
691  PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
692  PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
693 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
694 
695  PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2
696  "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7
697  "pxor %%mm6, %%mm1 \n\t" // -l7-1
698  PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
699  "movq "MANGLE(b80)", %%mm2 \n\t" // 128
700  PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
701  PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
702  PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
703 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
704 
705  "movq "MANGLE(b00)", %%mm1 \n\t" // 0
706  "movq "MANGLE(b00)", %%mm5 \n\t" // 0
707  "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
708  "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
709  PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
710  PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
711  PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
712 
713 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
714 
715  "movq "MANGLE(b00)", %%mm7 \n\t" // 0
716  "movq %2, %%mm2 \n\t" // QP
717  PAVGB(%%mm6, %%mm2) // 128 + QP/2
718  "psubb %%mm6, %%mm2 \n\t"
719 
720  "movq %%mm4, %%mm1 \n\t"
721  "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
722  "pxor %%mm1, %%mm4 \n\t"
723  "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
724  "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
725  "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
726 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
727 
728  "movq %%mm4, %%mm3 \n\t" // d
729  "psubusb "MANGLE(b01)", %%mm4 \n\t"
730  PAVGB(%%mm7, %%mm4) // d/32
731  PAVGB(%%mm7, %%mm4) // (d + 32)/64
732  "paddb %%mm3, %%mm4 \n\t" // 5d/64
733  "pand %%mm2, %%mm4 \n\t"
734 
735  "movq "MANGLE(b80)", %%mm5 \n\t" // 128
736  "psubb %%mm0, %%mm5 \n\t" // q
737  "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
738  "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
739  "pxor %%mm7, %%mm5 \n\t"
740 
741  PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
742  "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
743 
744  "pand %%mm7, %%mm4 \n\t"
745  "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
746  "movq (%0, %1, 4), %%mm2 \n\t"
747  "pxor %%mm1, %%mm0 \n\t"
748  "pxor %%mm1, %%mm2 \n\t"
749  "paddb %%mm4, %%mm0 \n\t"
750  "psubb %%mm4, %%mm2 \n\t"
751  "pxor %%mm1, %%mm0 \n\t"
752  "pxor %%mm1, %%mm2 \n\t"
753  "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
754  "movq %%mm2, (%0, %1, 4) \n\t"
755 
756  :
757  : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
758  : "%"REG_a, "%"REG_c
759  );
760 
761 /*
762  {
763  int x;
764  src-= stride;
765  for(x=0; x<BLOCK_SIZE; x++){
766  const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
767  if(FFABS(middleEnergy)< 8*QP){
768  const int q=(src[l4] - src[l5])/2;
769  const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
770  const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
771 
772  int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
773  d= FFMAX(d, 0);
774 
775  d= (5*d + 32) >> 6;
776  d*= FFSIGN(-middleEnergy);
777 
778  if(q>0){
779  d= d<0 ? 0 : d;
780  d= d>q ? q : d;
781  }else{
782  d= d>0 ? 0 : d;
783  d= d<q ? q : d;
784  }
785 
786  src[l4]-= d;
787  src[l5]+= d;
788  }
789  src++;
790  }
791  src-=8;
792  for(x=0; x<8; x++){
793  int y;
794  for(y=4; y<6; y++){
795  int d= src[x+y*stride] - tmp[x+(y-4)*8];
796  int ad= FFABS(d);
797  static int max=0;
798  static int sum=0;
799  static int num=0;
800  static int bias=0;
801 
802  if(max<ad) max=ad;
803  sum+= ad>3 ? 1 : 0;
804  if(ad>3){
805  src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
806  }
807  if(y==4) bias+=d;
808  num++;
809  if(num%1000000 == 0){
810  av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias);
811  }
812  }
813  }
814 }
815 */
816 #elif TEMPLATE_PP_MMX
817  DECLARE_ALIGNED(8, uint64_t, tmp)[4]; // make space for 4 8-byte vars
818  src+= stride*4;
819  __asm__ volatile(
820  "pxor %%mm7, %%mm7 \n\t"
821 // 0 1 2 3 4 5 6 7
822 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
823 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
824 
825  "movq (%0), %%mm0 \n\t"
826  "movq %%mm0, %%mm1 \n\t"
827  "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
828  "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
829 
830  "movq (%0, %1), %%mm2 \n\t"
831  "lea (%0, %1, 2), %%"REG_a" \n\t"
832  "movq %%mm2, %%mm3 \n\t"
833  "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
834  "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
835 
836  "movq (%%"REG_a"), %%mm4 \n\t"
837  "movq %%mm4, %%mm5 \n\t"
838  "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
839  "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
840 
841  "paddw %%mm0, %%mm0 \n\t" // 2L0
842  "paddw %%mm1, %%mm1 \n\t" // 2H0
843  "psubw %%mm4, %%mm2 \n\t" // L1 - L2
844  "psubw %%mm5, %%mm3 \n\t" // H1 - H2
845  "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
846  "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
847 
848  "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
849  "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
850  "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
851  "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
852 
853  "movq (%%"REG_a", %1), %%mm2 \n\t"
854  "movq %%mm2, %%mm3 \n\t"
855  "punpcklbw %%mm7, %%mm2 \n\t" // L3
856  "punpckhbw %%mm7, %%mm3 \n\t" // H3
857 
858  "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
859  "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
860  "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
861  "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
862  "movq %%mm0, (%3) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
863  "movq %%mm1, 8(%3) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
864 
865  "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
866  "movq %%mm0, %%mm1 \n\t"
867  "punpcklbw %%mm7, %%mm0 \n\t" // L4
868  "punpckhbw %%mm7, %%mm1 \n\t" // H4
869 
870  "psubw %%mm0, %%mm2 \n\t" // L3 - L4
871  "psubw %%mm1, %%mm3 \n\t" // H3 - H4
872  "movq %%mm2, 16(%3) \n\t" // L3 - L4
873  "movq %%mm3, 24(%3) \n\t" // H3 - H4
874  "paddw %%mm4, %%mm4 \n\t" // 2L2
875  "paddw %%mm5, %%mm5 \n\t" // 2H2
876  "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
877  "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
878 
879  "lea (%%"REG_a", %1), %0 \n\t"
880  "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
881  "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
882  "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
883  "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
884 //50 opcodes so far
885  "movq (%0, %1, 2), %%mm2 \n\t"
886  "movq %%mm2, %%mm3 \n\t"
887  "punpcklbw %%mm7, %%mm2 \n\t" // L5
888  "punpckhbw %%mm7, %%mm3 \n\t" // H5
889  "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
890  "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
891  "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
892  "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
893 
894  "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
895  "punpcklbw %%mm7, %%mm6 \n\t" // L6
896  "psubw %%mm6, %%mm2 \n\t" // L5 - L6
897  "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
898  "punpckhbw %%mm7, %%mm6 \n\t" // H6
899  "psubw %%mm6, %%mm3 \n\t" // H5 - H6
900 
901  "paddw %%mm0, %%mm0 \n\t" // 2L4
902  "paddw %%mm1, %%mm1 \n\t" // 2H4
903  "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
904  "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
905 
906  "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
907  "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
908  "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
909  "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
910 
911  "movq (%0, %1, 4), %%mm2 \n\t"
912  "movq %%mm2, %%mm3 \n\t"
913  "punpcklbw %%mm7, %%mm2 \n\t" // L7
914  "punpckhbw %%mm7, %%mm3 \n\t" // H7
915 
916  "paddw %%mm2, %%mm2 \n\t" // 2L7
917  "paddw %%mm3, %%mm3 \n\t" // 2H7
918  "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
919  "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
920 
921  "movq (%3), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
922  "movq 8(%3), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
923 
924 #if TEMPLATE_PP_MMXEXT
925  "movq %%mm7, %%mm6 \n\t" // 0
926  "psubw %%mm0, %%mm6 \n\t"
927  "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
928  "movq %%mm7, %%mm6 \n\t" // 0
929  "psubw %%mm1, %%mm6 \n\t"
930  "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
931  "movq %%mm7, %%mm6 \n\t" // 0
932  "psubw %%mm2, %%mm6 \n\t"
933  "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
934  "movq %%mm7, %%mm6 \n\t" // 0
935  "psubw %%mm3, %%mm6 \n\t"
936  "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
937 #else
938  "movq %%mm7, %%mm6 \n\t" // 0
939  "pcmpgtw %%mm0, %%mm6 \n\t"
940  "pxor %%mm6, %%mm0 \n\t"
941  "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
942  "movq %%mm7, %%mm6 \n\t" // 0
943  "pcmpgtw %%mm1, %%mm6 \n\t"
944  "pxor %%mm6, %%mm1 \n\t"
945  "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
946  "movq %%mm7, %%mm6 \n\t" // 0
947  "pcmpgtw %%mm2, %%mm6 \n\t"
948  "pxor %%mm6, %%mm2 \n\t"
949  "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
950  "movq %%mm7, %%mm6 \n\t" // 0
951  "pcmpgtw %%mm3, %%mm6 \n\t"
952  "pxor %%mm6, %%mm3 \n\t"
953  "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
954 #endif
955 
956 #if TEMPLATE_PP_MMXEXT
957  "pminsw %%mm2, %%mm0 \n\t"
958  "pminsw %%mm3, %%mm1 \n\t"
959 #else
960  "movq %%mm0, %%mm6 \n\t"
961  "psubusw %%mm2, %%mm6 \n\t"
962  "psubw %%mm6, %%mm0 \n\t"
963  "movq %%mm1, %%mm6 \n\t"
964  "psubusw %%mm3, %%mm6 \n\t"
965  "psubw %%mm6, %%mm1 \n\t"
966 #endif
967 
968  "movd %2, %%mm2 \n\t" // QP
969  "punpcklbw %%mm7, %%mm2 \n\t"
970 
971  "movq %%mm7, %%mm6 \n\t" // 0
972  "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
973  "pxor %%mm6, %%mm4 \n\t"
974  "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
975  "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
976  "pxor %%mm7, %%mm5 \n\t"
977  "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
978 // 100 opcodes
979  "psllw $3, %%mm2 \n\t" // 8QP
980  "movq %%mm2, %%mm3 \n\t" // 8QP
981  "pcmpgtw %%mm4, %%mm2 \n\t"
982  "pcmpgtw %%mm5, %%mm3 \n\t"
983  "pand %%mm2, %%mm4 \n\t"
984  "pand %%mm3, %%mm5 \n\t"
985 
986 
987  "psubusw %%mm0, %%mm4 \n\t" // hd
988  "psubusw %%mm1, %%mm5 \n\t" // ld
989 
990 
991  "movq "MANGLE(w05)", %%mm2 \n\t" // 5
992  "pmullw %%mm2, %%mm4 \n\t"
993  "pmullw %%mm2, %%mm5 \n\t"
994  "movq "MANGLE(w20)", %%mm2 \n\t" // 32
995  "paddw %%mm2, %%mm4 \n\t"
996  "paddw %%mm2, %%mm5 \n\t"
997  "psrlw $6, %%mm4 \n\t"
998  "psrlw $6, %%mm5 \n\t"
999 
1000  "movq 16(%3), %%mm0 \n\t" // L3 - L4
1001  "movq 24(%3), %%mm1 \n\t" // H3 - H4
1002 
1003  "pxor %%mm2, %%mm2 \n\t"
1004  "pxor %%mm3, %%mm3 \n\t"
1005 
1006  "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1007  "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
1008  "pxor %%mm2, %%mm0 \n\t"
1009  "pxor %%mm3, %%mm1 \n\t"
1010  "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
1011  "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
1012  "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
1013  "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
1014 
1015  "pxor %%mm6, %%mm2 \n\t"
1016  "pxor %%mm7, %%mm3 \n\t"
1017  "pand %%mm2, %%mm4 \n\t"
1018  "pand %%mm3, %%mm5 \n\t"
1019 
1020 #if TEMPLATE_PP_MMXEXT
1021  "pminsw %%mm0, %%mm4 \n\t"
1022  "pminsw %%mm1, %%mm5 \n\t"
1023 #else
1024  "movq %%mm4, %%mm2 \n\t"
1025  "psubusw %%mm0, %%mm2 \n\t"
1026  "psubw %%mm2, %%mm4 \n\t"
1027  "movq %%mm5, %%mm2 \n\t"
1028  "psubusw %%mm1, %%mm2 \n\t"
1029  "psubw %%mm2, %%mm5 \n\t"
1030 #endif
1031  "pxor %%mm6, %%mm4 \n\t"
1032  "pxor %%mm7, %%mm5 \n\t"
1033  "psubw %%mm6, %%mm4 \n\t"
1034  "psubw %%mm7, %%mm5 \n\t"
1035  "packsswb %%mm5, %%mm4 \n\t"
1036  "movq (%0), %%mm0 \n\t"
1037  "paddb %%mm4, %%mm0 \n\t"
1038  "movq %%mm0, (%0) \n\t"
1039  "movq (%0, %1), %%mm0 \n\t"
1040  "psubb %%mm4, %%mm0 \n\t"
1041  "movq %%mm0, (%0, %1) \n\t"
1042 
1043  : "+r" (src)
1044  : "r" ((x86_reg)stride), "m" (c->pQPb), "r"(tmp)
1045  : "%"REG_a
1046  );
1047 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1048  const int l1= stride;
1049  const int l2= stride + l1;
1050  const int l3= stride + l2;
1051  const int l4= stride + l3;
1052  const int l5= stride + l4;
1053  const int l6= stride + l5;
1054  const int l7= stride + l6;
1055  const int l8= stride + l7;
1056 // const int l9= stride + l8;
1057  int x;
1058  src+= stride*3;
1059  for(x=0; x<BLOCK_SIZE; x++){
1060  const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1061  if(FFABS(middleEnergy) < 8*c->QP){
1062  const int q=(src[l4] - src[l5])/2;
1063  const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1064  const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1065 
1066  int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
1067  d= FFMAX(d, 0);
1068 
1069  d= (5*d + 32) >> 6;
1070  d*= FFSIGN(-middleEnergy);
1071 
1072  if(q>0){
1073  d= d<0 ? 0 : d;
1074  d= d>q ? q : d;
1075  }else{
1076  d= d>0 ? 0 : d;
1077  d= d<q ? q : d;
1078  }
1079 
1080  src[l4]-= d;
1081  src[l5]+= d;
1082  }
1083  src++;
1084  }
1085 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1086 }
1087 #endif //TEMPLATE_PP_ALTIVEC
1088 
1089 #if !TEMPLATE_PP_ALTIVEC
1090 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
1091 {
1092 #if HAVE_7REGS && (TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW)
1093  DECLARE_ALIGNED(8, uint64_t, tmp)[3];
1094  __asm__ volatile(
1095  "pxor %%mm6, %%mm6 \n\t"
1096  "pcmpeqb %%mm7, %%mm7 \n\t"
1097  "movq %2, %%mm0 \n\t"
1098  "punpcklbw %%mm6, %%mm0 \n\t"
1099  "psrlw $1, %%mm0 \n\t"
1100  "psubw %%mm7, %%mm0 \n\t"
1101  "packuswb %%mm0, %%mm0 \n\t"
1102  "movq %%mm0, %3 \n\t"
1103 
1104  "lea (%0, %1), %%"REG_a" \n\t"
1105  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1106 
1107 // 0 1 2 3 4 5 6 7 8 9
1108 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1109 
1110 #undef REAL_FIND_MIN_MAX
1111 #undef FIND_MIN_MAX
1112 #if TEMPLATE_PP_MMXEXT
1113 #define REAL_FIND_MIN_MAX(addr)\
1114  "movq " #addr ", %%mm0 \n\t"\
1115  "pminub %%mm0, %%mm7 \n\t"\
1116  "pmaxub %%mm0, %%mm6 \n\t"
1117 #else
1118 #define REAL_FIND_MIN_MAX(addr)\
1119  "movq " #addr ", %%mm0 \n\t"\
1120  "movq %%mm7, %%mm1 \n\t"\
1121  "psubusb %%mm0, %%mm6 \n\t"\
1122  "paddb %%mm0, %%mm6 \n\t"\
1123  "psubusb %%mm0, %%mm1 \n\t"\
1124  "psubb %%mm1, %%mm7 \n\t"
1125 #endif
1126 #define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr)
1127 
1128 FIND_MIN_MAX((%%REGa))
1129 FIND_MIN_MAX((%%REGa, %1))
1130 FIND_MIN_MAX((%%REGa, %1, 2))
1131 FIND_MIN_MAX((%0, %1, 4))
1132 FIND_MIN_MAX((%%REGd))
1133 FIND_MIN_MAX((%%REGd, %1))
1134 FIND_MIN_MAX((%%REGd, %1, 2))
1135 FIND_MIN_MAX((%0, %1, 8))
1136 
1137  "movq %%mm7, %%mm4 \n\t"
1138  "psrlq $8, %%mm7 \n\t"
1139 #if TEMPLATE_PP_MMXEXT
1140  "pminub %%mm4, %%mm7 \n\t" // min of pixels
1141  "pshufw $0xF9, %%mm7, %%mm4 \n\t"
1142  "pminub %%mm4, %%mm7 \n\t" // min of pixels
1143  "pshufw $0xFE, %%mm7, %%mm4 \n\t"
1144  "pminub %%mm4, %%mm7 \n\t"
1145 #else
1146  "movq %%mm7, %%mm1 \n\t"
1147  "psubusb %%mm4, %%mm1 \n\t"
1148  "psubb %%mm1, %%mm7 \n\t"
1149  "movq %%mm7, %%mm4 \n\t"
1150  "psrlq $16, %%mm7 \n\t"
1151  "movq %%mm7, %%mm1 \n\t"
1152  "psubusb %%mm4, %%mm1 \n\t"
1153  "psubb %%mm1, %%mm7 \n\t"
1154  "movq %%mm7, %%mm4 \n\t"
1155  "psrlq $32, %%mm7 \n\t"
1156  "movq %%mm7, %%mm1 \n\t"
1157  "psubusb %%mm4, %%mm1 \n\t"
1158  "psubb %%mm1, %%mm7 \n\t"
1159 #endif
1160 
1161 
1162  "movq %%mm6, %%mm4 \n\t"
1163  "psrlq $8, %%mm6 \n\t"
1164 #if TEMPLATE_PP_MMXEXT
1165  "pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1166  "pshufw $0xF9, %%mm6, %%mm4 \n\t"
1167  "pmaxub %%mm4, %%mm6 \n\t"
1168  "pshufw $0xFE, %%mm6, %%mm4 \n\t"
1169  "pmaxub %%mm4, %%mm6 \n\t"
1170 #else
1171  "psubusb %%mm4, %%mm6 \n\t"
1172  "paddb %%mm4, %%mm6 \n\t"
1173  "movq %%mm6, %%mm4 \n\t"
1174  "psrlq $16, %%mm6 \n\t"
1175  "psubusb %%mm4, %%mm6 \n\t"
1176  "paddb %%mm4, %%mm6 \n\t"
1177  "movq %%mm6, %%mm4 \n\t"
1178  "psrlq $32, %%mm6 \n\t"
1179  "psubusb %%mm4, %%mm6 \n\t"
1180  "paddb %%mm4, %%mm6 \n\t"
1181 #endif
1182  "movq %%mm6, %%mm0 \n\t" // max
1183  "psubb %%mm7, %%mm6 \n\t" // max - min
1184  "push %4 \n\t"
1185  "movd %%mm6, %k4 \n\t"
1186  "cmpb "MANGLE(deringThreshold)", %b4 \n\t"
1187  "pop %4 \n\t"
1188  " jb 1f \n\t"
1189  PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1190  "punpcklbw %%mm7, %%mm7 \n\t"
1191  "punpcklbw %%mm7, %%mm7 \n\t"
1192  "punpcklbw %%mm7, %%mm7 \n\t"
1193  "movq %%mm7, (%4) \n\t"
1194 
1195  "movq (%0), %%mm0 \n\t" // L10
1196  "movq %%mm0, %%mm1 \n\t" // L10
1197  "movq %%mm0, %%mm2 \n\t" // L10
1198  "psllq $8, %%mm1 \n\t"
1199  "psrlq $8, %%mm2 \n\t"
1200  "movd -4(%0), %%mm3 \n\t"
1201  "movd 8(%0), %%mm4 \n\t"
1202  "psrlq $24, %%mm3 \n\t"
1203  "psllq $56, %%mm4 \n\t"
1204  "por %%mm3, %%mm1 \n\t" // L00
1205  "por %%mm4, %%mm2 \n\t" // L20
1206  "movq %%mm1, %%mm3 \n\t" // L00
1207  PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1208  PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1209  "psubusb %%mm7, %%mm0 \n\t"
1210  "psubusb %%mm7, %%mm2 \n\t"
1211  "psubusb %%mm7, %%mm3 \n\t"
1212  "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1213  "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1214  "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1215  "paddb %%mm2, %%mm0 \n\t"
1216  "paddb %%mm3, %%mm0 \n\t"
1217 
1218  "movq (%%"REG_a"), %%mm2 \n\t" // L11
1219  "movq %%mm2, %%mm3 \n\t" // L11
1220  "movq %%mm2, %%mm4 \n\t" // L11
1221  "psllq $8, %%mm3 \n\t"
1222  "psrlq $8, %%mm4 \n\t"
1223  "movd -4(%%"REG_a"), %%mm5 \n\t"
1224  "movd 8(%%"REG_a"), %%mm6 \n\t"
1225  "psrlq $24, %%mm5 \n\t"
1226  "psllq $56, %%mm6 \n\t"
1227  "por %%mm5, %%mm3 \n\t" // L01
1228  "por %%mm6, %%mm4 \n\t" // L21
1229  "movq %%mm3, %%mm5 \n\t" // L01
1230  PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1231  PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1232  "psubusb %%mm7, %%mm2 \n\t"
1233  "psubusb %%mm7, %%mm4 \n\t"
1234  "psubusb %%mm7, %%mm5 \n\t"
1235  "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1236  "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1237  "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1238  "paddb %%mm4, %%mm2 \n\t"
1239  "paddb %%mm5, %%mm2 \n\t"
1240 // 0, 2, 3, 1
1241 #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1242  "movq " #src ", " #sx " \n\t" /* src[0] */\
1243  "movq " #sx ", " #lx " \n\t" /* src[0] */\
1244  "movq " #sx ", " #t0 " \n\t" /* src[0] */\
1245  "psllq $8, " #lx " \n\t"\
1246  "psrlq $8, " #t0 " \n\t"\
1247  "movd -4" #src ", " #t1 " \n\t"\
1248  "psrlq $24, " #t1 " \n\t"\
1249  "por " #t1 ", " #lx " \n\t" /* src[-1] */\
1250  "movd 8" #src ", " #t1 " \n\t"\
1251  "psllq $56, " #t1 " \n\t"\
1252  "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1253  "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1254  PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1255  PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1256  PAVGB(lx, pplx) \
1257  "movq " #lx ", 8(%4) \n\t"\
1258  "movq (%4), " #lx " \n\t"\
1259  "psubusb " #lx ", " #t1 " \n\t"\
1260  "psubusb " #lx ", " #t0 " \n\t"\
1261  "psubusb " #lx ", " #sx " \n\t"\
1262  "movq "MANGLE(b00)", " #lx " \n\t"\
1263  "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1264  "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1265  "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1266  "paddb " #t1 ", " #t0 " \n\t"\
1267  "paddb " #t0 ", " #sx " \n\t"\
1268 \
1269  PAVGB(plx, pplx) /* filtered */\
1270  "movq " #dst ", " #t0 " \n\t" /* dst */\
1271  "movq " #t0 ", " #t1 " \n\t" /* dst */\
1272  "psubusb %3, " #t0 " \n\t"\
1273  "paddusb %3, " #t1 " \n\t"\
1274  PMAXUB(t0, pplx)\
1275  PMINUB(t1, pplx, t0)\
1276  "paddb " #sx ", " #ppsx " \n\t"\
1277  "paddb " #psx ", " #ppsx " \n\t"\
1278  "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1279  "pand "MANGLE(b08)", " #ppsx " \n\t"\
1280  "pcmpeqb " #lx ", " #ppsx " \n\t"\
1281  "pand " #ppsx ", " #pplx " \n\t"\
1282  "pandn " #dst ", " #ppsx " \n\t"\
1283  "por " #pplx ", " #ppsx " \n\t"\
1284  "movq " #ppsx ", " #dst " \n\t"\
1285  "movq 8(%4), " #lx " \n\t"
1286 
1287 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1288  REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1)
1289 /*
1290 0000000
1291 1111111
1292 
1293 1111110
1294 1111101
1295 1111100
1296 1111011
1297 1111010
1298 1111001
1299 
1300 1111000
1301 1110111
1302 
1303 */
1304 //DERING_CORE(dst ,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
1305 DERING_CORE((%%REGa) ,(%%REGa, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1306 DERING_CORE((%%REGa, %1) ,(%%REGa, %1, 2),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1307 DERING_CORE((%%REGa, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1308 DERING_CORE((%0, %1, 4) ,(%%REGd) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1309 DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1310 DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1311 DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1312 DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1313 
1314  "1: \n\t"
1315  : : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb), "m"(c->pQPb2), "q"(tmp)
1316  : "%"REG_a, "%"REG_d, "%"REG_SP
1317  );
1318 #else // HAVE_7REGS && (TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW)
1319  int y;
1320  int min=255;
1321  int max=0;
1322  int avg;
1323  uint8_t *p;
1324  int s[10];
1325  const int QP2= c->QP/2 + 1;
1326 
1327  src --;
1328  for(y=1; y<9; y++){
1329  int x;
1330  p= src + stride*y;
1331  for(x=1; x<9; x++){
1332  p++;
1333  if(*p > max) max= *p;
1334  if(*p < min) min= *p;
1335  }
1336  }
1337  avg= (min + max + 1)>>1;
1338 
1339  if(max - min <deringThreshold) return;
1340 
1341  for(y=0; y<10; y++){
1342  int t = 0;
1343 
1344  if(src[stride*y + 0] > avg) t+= 1;
1345  if(src[stride*y + 1] > avg) t+= 2;
1346  if(src[stride*y + 2] > avg) t+= 4;
1347  if(src[stride*y + 3] > avg) t+= 8;
1348  if(src[stride*y + 4] > avg) t+= 16;
1349  if(src[stride*y + 5] > avg) t+= 32;
1350  if(src[stride*y + 6] > avg) t+= 64;
1351  if(src[stride*y + 7] > avg) t+= 128;
1352  if(src[stride*y + 8] > avg) t+= 256;
1353  if(src[stride*y + 9] > avg) t+= 512;
1354 
1355  t |= (~t)<<16;
1356  t &= (t<<1) & (t>>1);
1357  s[y] = t;
1358  }
1359 
1360  for(y=1; y<9; y++){
1361  int t = s[y-1] & s[y] & s[y+1];
1362  t|= t>>16;
1363  s[y-1]= t;
1364  }
1365 
1366  for(y=1; y<9; y++){
1367  int x;
1368  int t = s[y-1];
1369 
1370  p= src + stride*y;
1371  for(x=1; x<9; x++){
1372  p++;
1373  if(t & (1<<x)){
1374  int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1375  +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
1376  +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1377  f= (f + 8)>>4;
1378 
1379 #ifdef DEBUG_DERING_THRESHOLD
1380  __asm__ volatile("emms\n\t":);
1381  {
1382  static long long numPixels=0;
1383  if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1384 // if((max-min)<20 || (max-min)*QP<200)
1385 // if((max-min)*QP < 500)
1386 // if(max-min<QP/2)
1387  if(max-min < 20){
1388  static int numSkipped=0;
1389  static int errorSum=0;
1390  static int worstQP=0;
1391  static int worstRange=0;
1392  static int worstDiff=0;
1393  int diff= (f - *p);
1394  int absDiff= FFABS(diff);
1395  int error= diff*diff;
1396 
1397  if(x==1 || x==8 || y==1 || y==8) continue;
1398 
1399  numSkipped++;
1400  if(absDiff > worstDiff){
1401  worstDiff= absDiff;
1402  worstQP= QP;
1403  worstRange= max-min;
1404  }
1405  errorSum+= error;
1406 
1407  if(1024LL*1024LL*1024LL % numSkipped == 0){
1408  av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, "
1409  "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1410  (float)errorSum/numSkipped, numSkipped, worstQP, worstRange,
1411  worstDiff, (float)numSkipped/numPixels);
1412  }
1413  }
1414  }
1415 #endif
1416  if (*p + QP2 < f) *p= *p + QP2;
1417  else if(*p - QP2 > f) *p= *p - QP2;
1418  else *p=f;
1419  }
1420  }
1421  }
1422 #ifdef DEBUG_DERING_THRESHOLD
1423  if(max-min < 20){
1424  for(y=1; y<9; y++){
1425  int x;
1426  int t = 0;
1427  p= src + stride*y;
1428  for(x=1; x<9; x++){
1429  p++;
1430  *p = FFMIN(*p + 20, 255);
1431  }
1432  }
1433 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1434  }
1435 #endif
1436 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1437 }
1438 #endif //TEMPLATE_PP_ALTIVEC
1439 
1440 /**
1441  * Deinterlace the given block by linearly interpolating every second line.
1442  * will be called for every 8x8 block and can read & write from line 4-15
1443  * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1444  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1445  */
1446 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1447 {
1448 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1449  src+= 4*stride;
1450  __asm__ volatile(
1451  "lea (%0, %1), %%"REG_a" \n\t"
1452  "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
1453 // 0 1 2 3 4 5 6 7 8 9
1454 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
1455 
1456  "movq (%0), %%mm0 \n\t"
1457  "movq (%%"REG_a", %1), %%mm1 \n\t"
1458  PAVGB(%%mm1, %%mm0)
1459  "movq %%mm0, (%%"REG_a") \n\t"
1460  "movq (%0, %1, 4), %%mm0 \n\t"
1461  PAVGB(%%mm0, %%mm1)
1462  "movq %%mm1, (%%"REG_a", %1, 2) \n\t"
1463  "movq (%%"REG_c", %1), %%mm1 \n\t"
1464  PAVGB(%%mm1, %%mm0)
1465  "movq %%mm0, (%%"REG_c") \n\t"
1466  "movq (%0, %1, 8), %%mm0 \n\t"
1467  PAVGB(%%mm0, %%mm1)
1468  "movq %%mm1, (%%"REG_c", %1, 2) \n\t"
1469 
1470  : : "r" (src), "r" ((x86_reg)stride)
1471  : "%"REG_a, "%"REG_c
1472  );
1473 #else
1474  int a, b, x;
1475  src+= 4*stride;
1476 
1477  for(x=0; x<2; x++){
1478  a= *(uint32_t*)&src[stride*0];
1479  b= *(uint32_t*)&src[stride*2];
1480  *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1481  a= *(uint32_t*)&src[stride*4];
1482  *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1483  b= *(uint32_t*)&src[stride*6];
1484  *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1485  a= *(uint32_t*)&src[stride*8];
1486  *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1487  src += 4;
1488  }
1489 #endif
1490 }
1491 
1492 /**
1493  * Deinterlace the given block by cubic interpolating every second line.
1494  * will be called for every 8x8 block and can read & write from line 4-15
1495  * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1496  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1497  * this filter will read lines 3-15 and write 7-13
1498  */
1499 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1500 {
1501 #if TEMPLATE_PP_SSE2 || TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1502  src+= stride*3;
1503  __asm__ volatile(
1504  "lea (%0, %1), %%"REG_a" \n\t"
1505  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1506  "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t"
1507  "add %1, %%"REG_c" \n\t"
1508 #if TEMPLATE_PP_SSE2
1509  "pxor %%xmm7, %%xmm7 \n\t"
1510 #define REAL_DEINT_CUBIC(a,b,c,d,e)\
1511  "movq " #a ", %%xmm0 \n\t"\
1512  "movq " #b ", %%xmm1 \n\t"\
1513  "movq " #d ", %%xmm2 \n\t"\
1514  "movq " #e ", %%xmm3 \n\t"\
1515  "pavgb %%xmm2, %%xmm1 \n\t"\
1516  "pavgb %%xmm3, %%xmm0 \n\t"\
1517  "punpcklbw %%xmm7, %%xmm0 \n\t"\
1518  "punpcklbw %%xmm7, %%xmm1 \n\t"\
1519  "psubw %%xmm1, %%xmm0 \n\t"\
1520  "psraw $3, %%xmm0 \n\t"\
1521  "psubw %%xmm0, %%xmm1 \n\t"\
1522  "packuswb %%xmm1, %%xmm1 \n\t"\
1523  "movlps %%xmm1, " #c " \n\t"
1524 #else //TEMPLATE_PP_SSE2
1525  "pxor %%mm7, %%mm7 \n\t"
1526 // 0 1 2 3 4 5 6 7 8 9 10
1527 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1528 
1529 #define REAL_DEINT_CUBIC(a,b,c,d,e)\
1530  "movq " #a ", %%mm0 \n\t"\
1531  "movq " #b ", %%mm1 \n\t"\
1532  "movq " #d ", %%mm2 \n\t"\
1533  "movq " #e ", %%mm3 \n\t"\
1534  PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\
1535  PAVGB(%%mm3, %%mm0) /* (a+e) /2 */\
1536  "movq %%mm0, %%mm2 \n\t"\
1537  "punpcklbw %%mm7, %%mm0 \n\t"\
1538  "punpckhbw %%mm7, %%mm2 \n\t"\
1539  "movq %%mm1, %%mm3 \n\t"\
1540  "punpcklbw %%mm7, %%mm1 \n\t"\
1541  "punpckhbw %%mm7, %%mm3 \n\t"\
1542  "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\
1543  "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\
1544  "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\
1545  "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\
1546  "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\
1547  "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\
1548  "packuswb %%mm3, %%mm1 \n\t"\
1549  "movq %%mm1, " #c " \n\t"
1550 #endif //TEMPLATE_PP_SSE2
1551 #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e)
1552 
1553 DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1))
1554 DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8))
1555 DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc))
1556 DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2))
1557 
1558  : : "r" (src), "r" ((x86_reg)stride)
1559  :
1560 #if TEMPLATE_PP_SSE2
1561  XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm7",)
1562 #endif
1563  "%"REG_a, "%"REG_d, "%"REG_c
1564  );
1565 #undef REAL_DEINT_CUBIC
1566 #else //TEMPLATE_PP_SSE2 || TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1567  int x;
1568  src+= stride*3;
1569  for(x=0; x<8; x++){
1570  src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4);
1571  src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4);
1572  src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4);
1573  src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4);
1574  src++;
1575  }
1576 #endif //TEMPLATE_PP_SSE2 || TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1577 }
1578 
1579 /**
1580  * Deinterlace the given block by filtering every second line with a (-1 4 2 4 -1) filter.
1581  * will be called for every 8x8 block and can read & write from line 4-15
1582  * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1583  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1584  * this filter will read lines 4-13 and write 5-11
1585  */
1586 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
1587 {
1588 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1589  src+= stride*4;
1590  __asm__ volatile(
1591  "lea (%0, %1), %%"REG_a" \n\t"
1592  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1593  "pxor %%mm7, %%mm7 \n\t"
1594  "movq (%2), %%mm0 \n\t"
1595 // 0 1 2 3 4 5 6 7 8 9 10
1596 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1597 
1598 #define REAL_DEINT_FF(a,b,c,d)\
1599  "movq " #a ", %%mm1 \n\t"\
1600  "movq " #b ", %%mm2 \n\t"\
1601  "movq " #c ", %%mm3 \n\t"\
1602  "movq " #d ", %%mm4 \n\t"\
1603  PAVGB(%%mm3, %%mm1) \
1604  PAVGB(%%mm4, %%mm0) \
1605  "movq %%mm0, %%mm3 \n\t"\
1606  "punpcklbw %%mm7, %%mm0 \n\t"\
1607  "punpckhbw %%mm7, %%mm3 \n\t"\
1608  "movq %%mm1, %%mm4 \n\t"\
1609  "punpcklbw %%mm7, %%mm1 \n\t"\
1610  "punpckhbw %%mm7, %%mm4 \n\t"\
1611  "psllw $2, %%mm1 \n\t"\
1612  "psllw $2, %%mm4 \n\t"\
1613  "psubw %%mm0, %%mm1 \n\t"\
1614  "psubw %%mm3, %%mm4 \n\t"\
1615  "movq %%mm2, %%mm5 \n\t"\
1616  "movq %%mm2, %%mm0 \n\t"\
1617  "punpcklbw %%mm7, %%mm2 \n\t"\
1618  "punpckhbw %%mm7, %%mm5 \n\t"\
1619  "paddw %%mm2, %%mm1 \n\t"\
1620  "paddw %%mm5, %%mm4 \n\t"\
1621  "psraw $2, %%mm1 \n\t"\
1622  "psraw $2, %%mm4 \n\t"\
1623  "packuswb %%mm4, %%mm1 \n\t"\
1624  "movq %%mm1, " #b " \n\t"\
1625 
1626 #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d)
1627 
1628 DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2))
1629 DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1630 DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2))
1631 DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1632 
1633  "movq %%mm0, (%2) \n\t"
1634  : : "r" (src), "r" ((x86_reg)stride), "r"(tmp)
1635  : "%"REG_a, "%"REG_d
1636  );
1637 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1638  int x;
1639  src+= stride*4;
1640  for(x=0; x<8; x++){
1641  int t1= tmp[x];
1642  int t2= src[stride*1];
1643 
1644  src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3);
1645  t1= src[stride*4];
1646  src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3);
1647  t2= src[stride*6];
1648  src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3);
1649  t1= src[stride*8];
1650  src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3);
1651  tmp[x]= t1;
1652 
1653  src++;
1654  }
1655 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1656 }
1657 
1658 /**
1659  * Deinterlace the given block by filtering every line with a (-1 2 6 2 -1) filter.
1660  * will be called for every 8x8 block and can read & write from line 4-15
1661  * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1662  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1663  * this filter will read lines 4-13 and write 4-11
1664  */
1665 static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2)
1666 {
1667 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1668  src+= stride*4;
1669  __asm__ volatile(
1670  "lea (%0, %1), %%"REG_a" \n\t"
1671  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1672  "pxor %%mm7, %%mm7 \n\t"
1673  "movq (%2), %%mm0 \n\t"
1674  "movq (%3), %%mm1 \n\t"
1675 // 0 1 2 3 4 5 6 7 8 9 10
1676 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1677 
1678 #define REAL_DEINT_L5(t1,t2,a,b,c)\
1679  "movq " #a ", %%mm2 \n\t"\
1680  "movq " #b ", %%mm3 \n\t"\
1681  "movq " #c ", %%mm4 \n\t"\
1682  PAVGB(t2, %%mm3) \
1683  PAVGB(t1, %%mm4) \
1684  "movq %%mm2, %%mm5 \n\t"\
1685  "movq %%mm2, " #t1 " \n\t"\
1686  "punpcklbw %%mm7, %%mm2 \n\t"\
1687  "punpckhbw %%mm7, %%mm5 \n\t"\
1688  "movq %%mm2, %%mm6 \n\t"\
1689  "paddw %%mm2, %%mm2 \n\t"\
1690  "paddw %%mm6, %%mm2 \n\t"\
1691  "movq %%mm5, %%mm6 \n\t"\
1692  "paddw %%mm5, %%mm5 \n\t"\
1693  "paddw %%mm6, %%mm5 \n\t"\
1694  "movq %%mm3, %%mm6 \n\t"\
1695  "punpcklbw %%mm7, %%mm3 \n\t"\
1696  "punpckhbw %%mm7, %%mm6 \n\t"\
1697  "paddw %%mm3, %%mm3 \n\t"\
1698  "paddw %%mm6, %%mm6 \n\t"\
1699  "paddw %%mm3, %%mm2 \n\t"\
1700  "paddw %%mm6, %%mm5 \n\t"\
1701  "movq %%mm4, %%mm6 \n\t"\
1702  "punpcklbw %%mm7, %%mm4 \n\t"\
1703  "punpckhbw %%mm7, %%mm6 \n\t"\
1704  "psubw %%mm4, %%mm2 \n\t"\
1705  "psubw %%mm6, %%mm5 \n\t"\
1706  "psraw $2, %%mm2 \n\t"\
1707  "psraw $2, %%mm5 \n\t"\
1708  "packuswb %%mm5, %%mm2 \n\t"\
1709  "movq %%mm2, " #a " \n\t"\
1710 
1711 #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c)
1712 
1713 DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) )
1714 DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2))
1715 DEINT_L5(%%mm0, %%mm1, (%%REGa, %1) , (%%REGa, %1, 2), (%0, %1, 4) )
1716 DEINT_L5(%%mm1, %%mm0, (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1717 DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) )
1718 DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2))
1719 DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) )
1720 DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1721 
1722  "movq %%mm0, (%2) \n\t"
1723  "movq %%mm1, (%3) \n\t"
1724  : : "r" (src), "r" ((x86_reg)stride), "r"(tmp), "r"(tmp2)
1725  : "%"REG_a, "%"REG_d
1726  );
1727 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1728  int x;
1729  src+= stride*4;
1730  for(x=0; x<8; x++){
1731  int t1= tmp[x];
1732  int t2= tmp2[x];
1733  int t3= src[0];
1734 
1735  src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
1736  t1= src[stride*1];
1737  src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
1738  t2= src[stride*2];
1739  src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
1740  t3= src[stride*3];
1741  src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
1742  t1= src[stride*4];
1743  src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
1744  t2= src[stride*5];
1745  src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
1746  t3= src[stride*6];
1747  src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
1748  t1= src[stride*7];
1749  src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);
1750 
1751  tmp[x]= t3;
1752  tmp2[x]= t1;
1753 
1754  src++;
1755  }
1756 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1757 }
1758 
1759 /**
1760  * Deinterlace the given block by filtering all lines with a (1 2 1) filter.
1761  * will be called for every 8x8 block and can read & write from line 4-15
1762  * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1763  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1764  * this filter will read lines 4-13 and write 4-11
1765  */
1766 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp)
1767 {
1768 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1769  src+= 4*stride;
1770  __asm__ volatile(
1771  "lea (%0, %1), %%"REG_a" \n\t"
1772  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1773 // 0 1 2 3 4 5 6 7 8 9
1774 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1775 
1776  "movq (%2), %%mm0 \n\t" // L0
1777  "movq (%%"REG_a"), %%mm1 \n\t" // L2
1778  PAVGB(%%mm1, %%mm0) // L0+L2
1779  "movq (%0), %%mm2 \n\t" // L1
1780  PAVGB(%%mm2, %%mm0)
1781  "movq %%mm0, (%0) \n\t"
1782  "movq (%%"REG_a", %1), %%mm0 \n\t" // L3
1783  PAVGB(%%mm0, %%mm2) // L1+L3
1784  PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1785  "movq %%mm2, (%%"REG_a") \n\t"
1786  "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4
1787  PAVGB(%%mm2, %%mm1) // L2+L4
1788  PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1789  "movq %%mm1, (%%"REG_a", %1) \n\t"
1790  "movq (%0, %1, 4), %%mm1 \n\t" // L5
1791  PAVGB(%%mm1, %%mm0) // L3+L5
1792  PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1793  "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
1794  "movq (%%"REG_d"), %%mm0 \n\t" // L6
1795  PAVGB(%%mm0, %%mm2) // L4+L6
1796  PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1797  "movq %%mm2, (%0, %1, 4) \n\t"
1798  "movq (%%"REG_d", %1), %%mm2 \n\t" // L7
1799  PAVGB(%%mm2, %%mm1) // L5+L7
1800  PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1801  "movq %%mm1, (%%"REG_d") \n\t"
1802  "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8
1803  PAVGB(%%mm1, %%mm0) // L6+L8
1804  PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1805  "movq %%mm0, (%%"REG_d", %1) \n\t"
1806  "movq (%0, %1, 8), %%mm0 \n\t" // L9
1807  PAVGB(%%mm0, %%mm2) // L7+L9
1808  PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1809  "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1810  "movq %%mm1, (%2) \n\t"
1811 
1812  : : "r" (src), "r" ((x86_reg)stride), "r" (tmp)
1813  : "%"REG_a, "%"REG_d
1814  );
1815 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1816  int a, b, c, x;
1817  src+= 4*stride;
1818 
1819  for(x=0; x<2; x++){
1820  a= *(uint32_t*)&tmp[stride*0];
1821  b= *(uint32_t*)&src[stride*0];
1822  c= *(uint32_t*)&src[stride*1];
1823  a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1824  *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1825 
1826  a= *(uint32_t*)&src[stride*2];
1827  b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1828  *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1829 
1830  b= *(uint32_t*)&src[stride*3];
1831  c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1832  *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1833 
1834  c= *(uint32_t*)&src[stride*4];
1835  a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1836  *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1837 
1838  a= *(uint32_t*)&src[stride*5];
1839  b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1840  *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1841 
1842  b= *(uint32_t*)&src[stride*6];
1843  c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1844  *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1845 
1846  c= *(uint32_t*)&src[stride*7];
1847  a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1848  *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1849 
1850  a= *(uint32_t*)&src[stride*8];
1851  b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1852  *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1853 
1854  *(uint32_t*)&tmp[stride*0]= c;
1855  src += 4;
1856  tmp += 4;
1857  }
1858 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
1859 }
1860 
1861 /**
1862  * Deinterlace the given block by applying a median filter to every second line.
1863  * will be called for every 8x8 block and can read & write from line 4-15,
1864  * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1865  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1866  */
1867 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1868 {
1869 #if TEMPLATE_PP_MMX
1870  src+= 4*stride;
1871 #if TEMPLATE_PP_MMXEXT
1872  __asm__ volatile(
1873  "lea (%0, %1), %%"REG_a" \n\t"
1874  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1875 // 0 1 2 3 4 5 6 7 8 9
1876 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1877 
1878  "movq (%0), %%mm0 \n\t"
1879  "movq (%%"REG_a", %1), %%mm2 \n\t"
1880  "movq (%%"REG_a"), %%mm1 \n\t"
1881  "movq %%mm0, %%mm3 \n\t"
1882  "pmaxub %%mm1, %%mm0 \n\t"
1883  "pminub %%mm3, %%mm1 \n\t"
1884  "pmaxub %%mm2, %%mm1 \n\t"
1885  "pminub %%mm1, %%mm0 \n\t"
1886  "movq %%mm0, (%%"REG_a") \n\t"
1887 
1888  "movq (%0, %1, 4), %%mm0 \n\t"
1889  "movq (%%"REG_a", %1, 2), %%mm1 \n\t"
1890  "movq %%mm2, %%mm3 \n\t"
1891  "pmaxub %%mm1, %%mm2 \n\t"
1892  "pminub %%mm3, %%mm1 \n\t"
1893  "pmaxub %%mm0, %%mm1 \n\t"
1894  "pminub %%mm1, %%mm2 \n\t"
1895  "movq %%mm2, (%%"REG_a", %1, 2) \n\t"
1896 
1897  "movq (%%"REG_d"), %%mm2 \n\t"
1898  "movq (%%"REG_d", %1), %%mm1 \n\t"
1899  "movq %%mm2, %%mm3 \n\t"
1900  "pmaxub %%mm0, %%mm2 \n\t"
1901  "pminub %%mm3, %%mm0 \n\t"
1902  "pmaxub %%mm1, %%mm0 \n\t"
1903  "pminub %%mm0, %%mm2 \n\t"
1904  "movq %%mm2, (%%"REG_d") \n\t"
1905 
1906  "movq (%%"REG_d", %1, 2), %%mm2 \n\t"
1907  "movq (%0, %1, 8), %%mm0 \n\t"
1908  "movq %%mm2, %%mm3 \n\t"
1909  "pmaxub %%mm0, %%mm2 \n\t"
1910  "pminub %%mm3, %%mm0 \n\t"
1911  "pmaxub %%mm1, %%mm0 \n\t"
1912  "pminub %%mm0, %%mm2 \n\t"
1913  "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1914 
1915 
1916  : : "r" (src), "r" ((x86_reg)stride)
1917  : "%"REG_a, "%"REG_d
1918  );
1919 
1920 #else // MMX without MMX2
1921  __asm__ volatile(
1922  "lea (%0, %1), %%"REG_a" \n\t"
1923  "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1924 // 0 1 2 3 4 5 6 7 8 9
1925 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1926  "pxor %%mm7, %%mm7 \n\t"
1927 
1928 #define REAL_MEDIAN(a,b,c)\
1929  "movq " #a ", %%mm0 \n\t"\
1930  "movq " #b ", %%mm2 \n\t"\
1931  "movq " #c ", %%mm1 \n\t"\
1932  "movq %%mm0, %%mm3 \n\t"\
1933  "movq %%mm1, %%mm4 \n\t"\
1934  "movq %%mm2, %%mm5 \n\t"\
1935  "psubusb %%mm1, %%mm3 \n\t"\
1936  "psubusb %%mm2, %%mm4 \n\t"\
1937  "psubusb %%mm0, %%mm5 \n\t"\
1938  "pcmpeqb %%mm7, %%mm3 \n\t"\
1939  "pcmpeqb %%mm7, %%mm4 \n\t"\
1940  "pcmpeqb %%mm7, %%mm5 \n\t"\
1941  "movq %%mm3, %%mm6 \n\t"\
1942  "pxor %%mm4, %%mm3 \n\t"\
1943  "pxor %%mm5, %%mm4 \n\t"\
1944  "pxor %%mm6, %%mm5 \n\t"\
1945  "por %%mm3, %%mm1 \n\t"\
1946  "por %%mm4, %%mm2 \n\t"\
1947  "por %%mm5, %%mm0 \n\t"\
1948  "pand %%mm2, %%mm0 \n\t"\
1949  "pand %%mm1, %%mm0 \n\t"\
1950  "movq %%mm0, " #b " \n\t"
1951 #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c)
1952 
1953 MEDIAN((%0) , (%%REGa) , (%%REGa, %1))
1954 MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4))
1955 MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1))
1956 MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8))
1957 
1958  : : "r" (src), "r" ((x86_reg)stride)
1959  : "%"REG_a, "%"REG_d
1960  );
1961 #endif //TEMPLATE_PP_MMXEXT
1962 #else //TEMPLATE_PP_MMX
1963  int x, y;
1964  src+= 4*stride;
1965  // FIXME - there should be a way to do a few columns in parallel like w/mmx
1966  for(x=0; x<8; x++){
1967  uint8_t *colsrc = src;
1968  for (y=0; y<4; y++){
1969  int a, b, c, d, e, f;
1970  a = colsrc[0 ];
1971  b = colsrc[stride ];
1972  c = colsrc[stride*2];
1973  d = (a-b)>>31;
1974  e = (b-c)>>31;
1975  f = (c-a)>>31;
1976  colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f));
1977  colsrc += stride*2;
1978  }
1979  src++;
1980  }
1981 #endif //TEMPLATE_PP_MMX
1982 }
1983 
1984 #if TEMPLATE_PP_MMX
1985 /**
1986  * Transpose and shift the given 8x8 Block into dst1 and dst2.
1987  */
1988 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
1989 {
1990  __asm__(
1991  "lea (%0, %1), %%"REG_a" \n\t"
1992 // 0 1 2 3 4 5 6 7 8 9
1993 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1994  "movq (%0), %%mm0 \n\t" // 12345678
1995  "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
1996  "movq %%mm0, %%mm2 \n\t" // 12345678
1997  "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
1998  "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
1999 
2000  "movq (%%"REG_a", %1), %%mm1 \n\t"
2001  "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2002  "movq %%mm1, %%mm4 \n\t"
2003  "punpcklbw %%mm3, %%mm1 \n\t"
2004  "punpckhbw %%mm3, %%mm4 \n\t"
2005 
2006  "movq %%mm0, %%mm3 \n\t"
2007  "punpcklwd %%mm1, %%mm0 \n\t"
2008  "punpckhwd %%mm1, %%mm3 \n\t"
2009  "movq %%mm2, %%mm1 \n\t"
2010  "punpcklwd %%mm4, %%mm2 \n\t"
2011  "punpckhwd %%mm4, %%mm1 \n\t"
2012 
2013  "movd %%mm0, 128(%2) \n\t"
2014  "psrlq $32, %%mm0 \n\t"
2015  "movd %%mm0, 144(%2) \n\t"
2016  "movd %%mm3, 160(%2) \n\t"
2017  "psrlq $32, %%mm3 \n\t"
2018  "movd %%mm3, 176(%2) \n\t"
2019  "movd %%mm3, 48(%3) \n\t"
2020  "movd %%mm2, 192(%2) \n\t"
2021  "movd %%mm2, 64(%3) \n\t"
2022  "psrlq $32, %%mm2 \n\t"
2023  "movd %%mm2, 80(%3) \n\t"
2024  "movd %%mm1, 96(%3) \n\t"
2025  "psrlq $32, %%mm1 \n\t"
2026  "movd %%mm1, 112(%3) \n\t"
2027 
2028  "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t"
2029 
2030  "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
2031  "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
2032  "movq %%mm0, %%mm2 \n\t" // 12345678
2033  "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2034  "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2035 
2036  "movq (%%"REG_a", %1), %%mm1 \n\t"
2037  "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2038  "movq %%mm1, %%mm4 \n\t"
2039  "punpcklbw %%mm3, %%mm1 \n\t"
2040  "punpckhbw %%mm3, %%mm4 \n\t"
2041 
2042  "movq %%mm0, %%mm3 \n\t"
2043  "punpcklwd %%mm1, %%mm0 \n\t"
2044  "punpckhwd %%mm1, %%mm3 \n\t"
2045  "movq %%mm2, %%mm1 \n\t"
2046  "punpcklwd %%mm4, %%mm2 \n\t"
2047  "punpckhwd %%mm4, %%mm1 \n\t"
2048 
2049  "movd %%mm0, 132(%2) \n\t"
2050  "psrlq $32, %%mm0 \n\t"
2051  "movd %%mm0, 148(%2) \n\t"
2052  "movd %%mm3, 164(%2) \n\t"
2053  "psrlq $32, %%mm3 \n\t"
2054  "movd %%mm3, 180(%2) \n\t"
2055  "movd %%mm3, 52(%3) \n\t"
2056  "movd %%mm2, 196(%2) \n\t"
2057  "movd %%mm2, 68(%3) \n\t"
2058  "psrlq $32, %%mm2 \n\t"
2059  "movd %%mm2, 84(%3) \n\t"
2060  "movd %%mm1, 100(%3) \n\t"
2061  "psrlq $32, %%mm1 \n\t"
2062  "movd %%mm1, 116(%3) \n\t"
2063 
2064 
2065  :: "r" (src), "r" ((x86_reg)srcStride), "r" (dst1), "r" (dst2)
2066  : "%"REG_a
2067  );
2068 }
2069 
2070 /**
2071  * Transpose the given 8x8 block.
2072  */
2073 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
2074 {
2075  __asm__(
2076  "lea (%0, %1), %%"REG_a" \n\t"
2077  "lea (%%"REG_a",%1,4), %%"REG_d" \n\t"
2078 // 0 1 2 3 4 5 6 7 8 9
2079 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2080  "movq (%2), %%mm0 \n\t" // 12345678
2081  "movq 16(%2), %%mm1 \n\t" // abcdefgh
2082  "movq %%mm0, %%mm2 \n\t" // 12345678
2083  "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2084  "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2085 
2086  "movq 32(%2), %%mm1 \n\t"
2087  "movq 48(%2), %%mm3 \n\t"
2088  "movq %%mm1, %%mm4 \n\t"
2089  "punpcklbw %%mm3, %%mm1 \n\t"
2090  "punpckhbw %%mm3, %%mm4 \n\t"
2091 
2092  "movq %%mm0, %%mm3 \n\t"
2093  "punpcklwd %%mm1, %%mm0 \n\t"
2094  "punpckhwd %%mm1, %%mm3 \n\t"
2095  "movq %%mm2, %%mm1 \n\t"
2096  "punpcklwd %%mm4, %%mm2 \n\t"
2097  "punpckhwd %%mm4, %%mm1 \n\t"
2098 
2099  "movd %%mm0, (%0) \n\t"
2100  "psrlq $32, %%mm0 \n\t"
2101  "movd %%mm0, (%%"REG_a") \n\t"
2102  "movd %%mm3, (%%"REG_a", %1) \n\t"
2103  "psrlq $32, %%mm3 \n\t"
2104  "movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2105  "movd %%mm2, (%0, %1, 4) \n\t"
2106  "psrlq $32, %%mm2 \n\t"
2107  "movd %%mm2, (%%"REG_d") \n\t"
2108  "movd %%mm1, (%%"REG_d", %1) \n\t"
2109  "psrlq $32, %%mm1 \n\t"
2110  "movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2111 
2112 
2113  "movq 64(%2), %%mm0 \n\t" // 12345678
2114  "movq 80(%2), %%mm1 \n\t" // abcdefgh
2115  "movq %%mm0, %%mm2 \n\t" // 12345678
2116  "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2117  "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2118 
2119  "movq 96(%2), %%mm1 \n\t"
2120  "movq 112(%2), %%mm3 \n\t"
2121  "movq %%mm1, %%mm4 \n\t"
2122  "punpcklbw %%mm3, %%mm1 \n\t"
2123  "punpckhbw %%mm3, %%mm4 \n\t"
2124 
2125  "movq %%mm0, %%mm3 \n\t"
2126  "punpcklwd %%mm1, %%mm0 \n\t"
2127  "punpckhwd %%mm1, %%mm3 \n\t"
2128  "movq %%mm2, %%mm1 \n\t"
2129  "punpcklwd %%mm4, %%mm2 \n\t"
2130  "punpckhwd %%mm4, %%mm1 \n\t"
2131 
2132  "movd %%mm0, 4(%0) \n\t"
2133  "psrlq $32, %%mm0 \n\t"
2134  "movd %%mm0, 4(%%"REG_a") \n\t"
2135  "movd %%mm3, 4(%%"REG_a", %1) \n\t"
2136  "psrlq $32, %%mm3 \n\t"
2137  "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t"
2138  "movd %%mm2, 4(%0, %1, 4) \n\t"
2139  "psrlq $32, %%mm2 \n\t"
2140  "movd %%mm2, 4(%%"REG_d") \n\t"
2141  "movd %%mm1, 4(%%"REG_d", %1) \n\t"
2142  "psrlq $32, %%mm1 \n\t"
2143  "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t"
2144 
2145  :: "r" (dst), "r" ((x86_reg)dstStride), "r" (src)
2146  : "%"REG_a, "%"REG_d
2147  );
2148 }
2149 #endif //TEMPLATE_PP_MMX
2150 //static long test=0;
2151 
2152 #if !TEMPLATE_PP_ALTIVEC
2153 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2154  uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise)
2155 {
2156  // to save a register (FIXME do this outside of the loops)
2157  tempBlurredPast[127]= maxNoise[0];
2158  tempBlurredPast[128]= maxNoise[1];
2159  tempBlurredPast[129]= maxNoise[2];
2160 
2161 #define FAST_L2_DIFF
2162 //#define L1_DIFF //u should change the thresholds too if u try that one
2163 #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
2164  __asm__ volatile(
2165  "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride
2166  "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride
2167  "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2168 // 0 1 2 3 4 5 6 7 8 9
2169 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
2170 //FIXME reorder?
2171 #ifdef L1_DIFF //needs mmx2
2172  "movq (%0), %%mm0 \n\t" // L0
2173  "psadbw (%1), %%mm0 \n\t" // |L0-R0|
2174  "movq (%0, %2), %%mm1 \n\t" // L1
2175  "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2176  "movq (%0, %2, 2), %%mm2 \n\t" // L2
2177  "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2178  "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2179  "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3|
2180 
2181  "movq (%0, %2, 4), %%mm4 \n\t" // L4
2182  "paddw %%mm1, %%mm0 \n\t"
2183  "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2184  "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2185  "paddw %%mm2, %%mm0 \n\t"
2186  "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5|
2187  "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2188  "paddw %%mm3, %%mm0 \n\t"
2189  "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6|
2190  "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2191  "paddw %%mm4, %%mm0 \n\t"
2192  "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7|
2193  "paddw %%mm5, %%mm6 \n\t"
2194  "paddw %%mm7, %%mm6 \n\t"
2195  "paddw %%mm6, %%mm0 \n\t"
2196 #else //L1_DIFF
2197 #if defined (FAST_L2_DIFF)
2198  "pcmpeqb %%mm7, %%mm7 \n\t"
2199  "movq "MANGLE(b80)", %%mm6 \n\t"
2200  "pxor %%mm0, %%mm0 \n\t"
2201 #define REAL_L2_DIFF_CORE(a, b)\
2202  "movq " #a ", %%mm5 \n\t"\
2203  "movq " #b ", %%mm2 \n\t"\
2204  "pxor %%mm7, %%mm2 \n\t"\
2205  PAVGB(%%mm2, %%mm5)\
2206  "paddb %%mm6, %%mm5 \n\t"\
2207  "movq %%mm5, %%mm2 \n\t"\
2208  "psllw $8, %%mm5 \n\t"\
2209  "pmaddwd %%mm5, %%mm5 \n\t"\
2210  "pmaddwd %%mm2, %%mm2 \n\t"\
2211  "paddd %%mm2, %%mm5 \n\t"\
2212  "psrld $14, %%mm5 \n\t"\
2213  "paddd %%mm5, %%mm0 \n\t"
2214 
2215 #else //defined (FAST_L2_DIFF)
2216  "pxor %%mm7, %%mm7 \n\t"
2217  "pxor %%mm0, %%mm0 \n\t"
2218 #define REAL_L2_DIFF_CORE(a, b)\
2219  "movq " #a ", %%mm5 \n\t"\
2220  "movq " #b ", %%mm2 \n\t"\
2221  "movq %%mm5, %%mm1 \n\t"\
2222  "movq %%mm2, %%mm3 \n\t"\
2223  "punpcklbw %%mm7, %%mm5 \n\t"\
2224  "punpckhbw %%mm7, %%mm1 \n\t"\
2225  "punpcklbw %%mm7, %%mm2 \n\t"\
2226  "punpckhbw %%mm7, %%mm3 \n\t"\
2227  "psubw %%mm2, %%mm5 \n\t"\
2228  "psubw %%mm3, %%mm1 \n\t"\
2229  "pmaddwd %%mm5, %%mm5 \n\t"\
2230  "pmaddwd %%mm1, %%mm1 \n\t"\
2231  "paddd %%mm1, %%mm5 \n\t"\
2232  "paddd %%mm5, %%mm0 \n\t"
2233 
2234 #endif //defined (FAST_L2_DIFF)
2235 
2236 #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b)
2237 
2238 L2_DIFF_CORE((%0) , (%1))
2239 L2_DIFF_CORE((%0, %2) , (%1, %2))
2240 L2_DIFF_CORE((%0, %2, 2) , (%1, %2, 2))
2241 L2_DIFF_CORE((%0, %%REGa) , (%1, %%REGa))
2242 L2_DIFF_CORE((%0, %2, 4) , (%1, %2, 4))
2243 L2_DIFF_CORE((%0, %%REGd) , (%1, %%REGd))
2244 L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2))
2245 L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc))
2246 
2247 #endif //L1_DIFF
2248 
2249  "movq %%mm0, %%mm4 \n\t"
2250  "psrlq $32, %%mm0 \n\t"
2251  "paddd %%mm0, %%mm4 \n\t"
2252  "movd %%mm4, %%ecx \n\t"
2253  "shll $2, %%ecx \n\t"
2254  "mov %3, %%"REG_d" \n\t"
2255  "addl -4(%%"REG_d"), %%ecx \n\t"
2256  "addl 4(%%"REG_d"), %%ecx \n\t"
2257  "addl -1024(%%"REG_d"), %%ecx \n\t"
2258  "addl $4, %%ecx \n\t"
2259  "addl 1024(%%"REG_d"), %%ecx \n\t"
2260  "shrl $3, %%ecx \n\t"
2261  "movl %%ecx, (%%"REG_d") \n\t"
2262 
2263 // "mov %3, %%"REG_c" \n\t"
2264 // "mov %%"REG_c", test \n\t"
2265 // "jmp 4f \n\t"
2266  "cmpl 512(%%"REG_d"), %%ecx \n\t"
2267  " jb 2f \n\t"
2268  "cmpl 516(%%"REG_d"), %%ecx \n\t"
2269  " jb 1f \n\t"
2270 
2271  "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2272  "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2273  "movq (%0), %%mm0 \n\t" // L0
2274  "movq (%0, %2), %%mm1 \n\t" // L1
2275  "movq (%0, %2, 2), %%mm2 \n\t" // L2
2276  "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2277  "movq (%0, %2, 4), %%mm4 \n\t" // L4
2278  "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2279  "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2280  "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2281  "movq %%mm0, (%1) \n\t" // L0
2282  "movq %%mm1, (%1, %2) \n\t" // L1
2283  "movq %%mm2, (%1, %2, 2) \n\t" // L2
2284  "movq %%mm3, (%1, %%"REG_a") \n\t" // L3
2285  "movq %%mm4, (%1, %2, 4) \n\t" // L4
2286  "movq %%mm5, (%1, %%"REG_d") \n\t" // L5
2287  "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6
2288  "movq %%mm7, (%1, %%"REG_c") \n\t" // L7
2289  "jmp 4f \n\t"
2290 
2291  "1: \n\t"
2292  "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2293  "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2294  "movq (%0), %%mm0 \n\t" // L0
2295  PAVGB((%1), %%mm0) // L0
2296  "movq (%0, %2), %%mm1 \n\t" // L1
2297  PAVGB((%1, %2), %%mm1) // L1
2298  "movq (%0, %2, 2), %%mm2 \n\t" // L2
2299  PAVGB((%1, %2, 2), %%mm2) // L2
2300  "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2301  PAVGB((%1, %%REGa), %%mm3) // L3
2302  "movq (%0, %2, 4), %%mm4 \n\t" // L4
2303  PAVGB((%1, %2, 4), %%mm4) // L4
2304  "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2305  PAVGB((%1, %%REGd), %%mm5) // L5
2306  "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2307  PAVGB((%1, %%REGa, 2), %%mm6) // L6
2308  "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2309  PAVGB((%1, %%REGc), %%mm7) // L7
2310  "movq %%mm0, (%1) \n\t" // R0
2311  "movq %%mm1, (%1, %2) \n\t" // R1
2312  "movq %%mm2, (%1, %2, 2) \n\t" // R2
2313  "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2314  "movq %%mm4, (%1, %2, 4) \n\t" // R4
2315  "movq %%mm5, (%1, %%"REG_d") \n\t" // R5
2316  "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6
2317  "movq %%mm7, (%1, %%"REG_c") \n\t" // R7
2318  "movq %%mm0, (%0) \n\t" // L0
2319  "movq %%mm1, (%0, %2) \n\t" // L1
2320  "movq %%mm2, (%0, %2, 2) \n\t" // L2
2321  "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2322  "movq %%mm4, (%0, %2, 4) \n\t" // L4
2323  "movq %%mm5, (%0, %%"REG_d") \n\t" // L5
2324  "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6
2325  "movq %%mm7, (%0, %%"REG_c") \n\t" // L7
2326  "jmp 4f \n\t"
2327 
2328  "2: \n\t"
2329  "cmpl 508(%%"REG_d"), %%ecx \n\t"
2330  " jb 3f \n\t"
2331 
2332  "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2333  "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2334  "movq (%0), %%mm0 \n\t" // L0
2335  "movq (%0, %2), %%mm1 \n\t" // L1
2336  "movq (%0, %2, 2), %%mm2 \n\t" // L2
2337  "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2338  "movq (%1), %%mm4 \n\t" // R0
2339  "movq (%1, %2), %%mm5 \n\t" // R1
2340  "movq (%1, %2, 2), %%mm6 \n\t" // R2
2341  "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2342  PAVGB(%%mm4, %%mm0)
2343  PAVGB(%%mm5, %%mm1)
2344  PAVGB(%%mm6, %%mm2)
2345  PAVGB(%%mm7, %%mm3)
2346  PAVGB(%%mm4, %%mm0)
2347  PAVGB(%%mm5, %%mm1)
2348  PAVGB(%%mm6, %%mm2)
2349  PAVGB(%%mm7, %%mm3)
2350  "movq %%mm0, (%1) \n\t" // R0
2351  "movq %%mm1, (%1, %2) \n\t" // R1
2352  "movq %%mm2, (%1, %2, 2) \n\t" // R2
2353  "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2354  "movq %%mm0, (%0) \n\t" // L0
2355  "movq %%mm1, (%0, %2) \n\t" // L1
2356  "movq %%mm2, (%0, %2, 2) \n\t" // L2
2357  "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2358 
2359  "movq (%0, %2, 4), %%mm0 \n\t" // L4
2360  "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2361  "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2362  "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2363  "movq (%1, %2, 4), %%mm4 \n\t" // R4
2364  "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2365  "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2366  "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2367  PAVGB(%%mm4, %%mm0)
2368  PAVGB(%%mm5, %%mm1)
2369  PAVGB(%%mm6, %%mm2)
2370  PAVGB(%%mm7, %%mm3)
2371  PAVGB(%%mm4, %%mm0)
2372  PAVGB(%%mm5, %%mm1)
2373  PAVGB(%%mm6, %%mm2)
2374  PAVGB(%%mm7, %%mm3)
2375  "movq %%mm0, (%1, %2, 4) \n\t" // R4
2376  "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2377  "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2378  "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2379  "movq %%mm0, (%0, %2, 4) \n\t" // L4
2380  "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2381  "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2382  "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2383  "jmp 4f \n\t"
2384 
2385  "3: \n\t"
2386  "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2387  "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2388  "movq (%0), %%mm0 \n\t" // L0
2389  "movq (%0, %2), %%mm1 \n\t" // L1
2390  "movq (%0, %2, 2), %%mm2 \n\t" // L2
2391  "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2392  "movq (%1), %%mm4 \n\t" // R0
2393  "movq (%1, %2), %%mm5 \n\t" // R1
2394  "movq (%1, %2, 2), %%mm6 \n\t" // R2
2395  "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2396  PAVGB(%%mm4, %%mm0)
2397  PAVGB(%%mm5, %%mm1)
2398  PAVGB(%%mm6, %%mm2)
2399  PAVGB(%%mm7, %%mm3)
2400  PAVGB(%%mm4, %%mm0)
2401  PAVGB(%%mm5, %%mm1)
2402  PAVGB(%%mm6, %%mm2)
2403  PAVGB(%%mm7, %%mm3)
2404  PAVGB(%%mm4, %%mm0)
2405  PAVGB(%%mm5, %%mm1)
2406  PAVGB(%%mm6, %%mm2)
2407  PAVGB(%%mm7, %%mm3)
2408  "movq %%mm0, (%1) \n\t" // R0
2409  "movq %%mm1, (%1, %2) \n\t" // R1
2410  "movq %%mm2, (%1, %2, 2) \n\t" // R2
2411  "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2412  "movq %%mm0, (%0) \n\t" // L0
2413  "movq %%mm1, (%0, %2) \n\t" // L1
2414  "movq %%mm2, (%0, %2, 2) \n\t" // L2
2415  "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2416 
2417  "movq (%0, %2, 4), %%mm0 \n\t" // L4
2418  "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2419  "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2420  "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2421  "movq (%1, %2, 4), %%mm4 \n\t" // R4
2422  "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2423  "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2424  "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2425  PAVGB(%%mm4, %%mm0)
2426  PAVGB(%%mm5, %%mm1)
2427  PAVGB(%%mm6, %%mm2)
2428  PAVGB(%%mm7, %%mm3)
2429  PAVGB(%%mm4, %%mm0)
2430  PAVGB(%%mm5, %%mm1)
2431  PAVGB(%%mm6, %%mm2)
2432  PAVGB(%%mm7, %%mm3)
2433  PAVGB(%%mm4, %%mm0)
2434  PAVGB(%%mm5, %%mm1)
2435  PAVGB(%%mm6, %%mm2)
2436  PAVGB(%%mm7, %%mm3)
2437  "movq %%mm0, (%1, %2, 4) \n\t" // R4
2438  "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2439  "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2440  "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2441  "movq %%mm0, (%0, %2, 4) \n\t" // L4
2442  "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2443  "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2444  "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2445 
2446  "4: \n\t"
2447 
2448  :: "r" (src), "r" (tempBlurred), "r"((x86_reg)stride), "m" (tempBlurredPast)
2449  : "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2450  );
2451 #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
2452 {
2453  int y;
2454  int d=0;
2455 // int sysd=0;
2456  int i;
2457 
2458  for(y=0; y<8; y++){
2459  int x;
2460  for(x=0; x<8; x++){
2461  int ref= tempBlurred[ x + y*stride ];
2462  int cur= src[ x + y*stride ];
2463  int d1=ref - cur;
2464 // if(x==0 || x==7) d1+= d1>>1;
2465 // if(y==0 || y==7) d1+= d1>>1;
2466 // d+= FFABS(d1);
2467  d+= d1*d1;
2468 // sysd+= d1;
2469  }
2470  }
2471  i=d;
2472  d= (
2473  4*d
2474  +(*(tempBlurredPast-256))
2475  +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1))
2476  +(*(tempBlurredPast+256))
2477  +4)>>3;
2478  *tempBlurredPast=i;
2479 // ((*tempBlurredPast)*3 + d + 2)>>2;
2480 
2481 /*
2482 Switch between
2483  1 0 0 0 0 0 0 (0)
2484 64 32 16 8 4 2 1 (1)
2485 64 48 36 27 20 15 11 (33) (approx)
2486 64 56 49 43 37 33 29 (200) (approx)
2487 */
2488  if(d > maxNoise[1]){
2489  if(d < maxNoise[2]){
2490  for(y=0; y<8; y++){
2491  int x;
2492  for(x=0; x<8; x++){
2493  int ref= tempBlurred[ x + y*stride ];
2494  int cur= src[ x + y*stride ];
2495  tempBlurred[ x + y*stride ]=
2496  src[ x + y*stride ]=
2497  (ref + cur + 1)>>1;
2498  }
2499  }
2500  }else{
2501  for(y=0; y<8; y++){
2502  int x;
2503  for(x=0; x<8; x++){
2504  tempBlurred[ x + y*stride ]= src[ x + y*stride ];
2505  }
2506  }
2507  }
2508  }else{
2509  if(d < maxNoise[0]){
2510  for(y=0; y<8; y++){
2511  int x;
2512  for(x=0; x<8; x++){
2513  int ref= tempBlurred[ x + y*stride ];
2514  int cur= src[ x + y*stride ];
2515  tempBlurred[ x + y*stride ]=
2516  src[ x + y*stride ]=
2517  (ref*7 + cur + 4)>>3;
2518  }
2519  }
2520  }else{
2521  for(y=0; y<8; y++){
2522  int x;
2523  for(x=0; x<8; x++){
2524  int ref= tempBlurred[ x + y*stride ];
2525  int cur= src[ x + y*stride ];
2526  tempBlurred[ x + y*stride ]=
2527  src[ x + y*stride ]=
2528  (ref*3 + cur + 2)>>2;
2529  }
2530  }
2531  }
2532  }
2533 }
2534 #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW
2535 }
2536 #endif //TEMPLATE_PP_ALTIVEC
2537 
2538 #if TEMPLATE_PP_MMX
2539 /**
2540  * accurate deblock filter
2541  */
2542 static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){
2543  int64_t dc_mask, eq_mask, both_masks;
2544  int64_t sums[10*8*2];
2545  src+= step*3; // src points to begin of the 8x8 Block
2546  //{ START_TIMER
2547  __asm__ volatile(
2548  "movq %0, %%mm7 \n\t"
2549  "movq %1, %%mm6 \n\t"
2550  : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
2551  );
2552 
2553  __asm__ volatile(
2554  "lea (%2, %3), %%"REG_a" \n\t"
2555 // 0 1 2 3 4 5 6 7 8 9
2556 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
2557 
2558  "movq (%2), %%mm0 \n\t"
2559  "movq (%%"REG_a"), %%mm1 \n\t"
2560  "movq %%mm1, %%mm3 \n\t"
2561  "movq %%mm1, %%mm4 \n\t"
2562  "psubb %%mm1, %%mm0 \n\t" // mm0 = difference
2563  "paddb %%mm7, %%mm0 \n\t"
2564  "pcmpgtb %%mm6, %%mm0 \n\t"
2565 
2566  "movq (%%"REG_a",%3), %%mm2 \n\t"
2567  PMAXUB(%%mm2, %%mm4)
2568  PMINUB(%%mm2, %%mm3, %%mm5)
2569  "psubb %%mm2, %%mm1 \n\t"
2570  "paddb %%mm7, %%mm1 \n\t"
2571  "pcmpgtb %%mm6, %%mm1 \n\t"
2572  "paddb %%mm1, %%mm0 \n\t"
2573 
2574  "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2575  PMAXUB(%%mm1, %%mm4)
2576  PMINUB(%%mm1, %%mm3, %%mm5)
2577  "psubb %%mm1, %%mm2 \n\t"
2578  "paddb %%mm7, %%mm2 \n\t"
2579  "pcmpgtb %%mm6, %%mm2 \n\t"
2580  "paddb %%mm2, %%mm0 \n\t"
2581 
2582  "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
2583 
2584  "movq (%2, %3, 4), %%mm2 \n\t"
2585  PMAXUB(%%mm2, %%mm4)
2586  PMINUB(%%mm2, %%mm3, %%mm5)
2587  "psubb %%mm2, %%mm1 \n\t"
2588  "paddb %%mm7, %%mm1 \n\t"
2589  "pcmpgtb %%mm6, %%mm1 \n\t"
2590  "paddb %%mm1, %%mm0 \n\t"
2591 
2592  "movq (%%"REG_a"), %%mm1 \n\t"
2593  PMAXUB(%%mm1, %%mm4)
2594  PMINUB(%%mm1, %%mm3, %%mm5)
2595  "psubb %%mm1, %%mm2 \n\t"
2596  "paddb %%mm7, %%mm2 \n\t"
2597  "pcmpgtb %%mm6, %%mm2 \n\t"
2598  "paddb %%mm2, %%mm0 \n\t"
2599 
2600  "movq (%%"REG_a", %3), %%mm2 \n\t"
2601  PMAXUB(%%mm2, %%mm4)
2602  PMINUB(%%mm2, %%mm3, %%mm5)
2603  "psubb %%mm2, %%mm1 \n\t"
2604  "paddb %%mm7, %%mm1 \n\t"
2605  "pcmpgtb %%mm6, %%mm1 \n\t"
2606  "paddb %%mm1, %%mm0 \n\t"
2607 
2608  "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2609  PMAXUB(%%mm1, %%mm4)
2610  PMINUB(%%mm1, %%mm3, %%mm5)
2611  "psubb %%mm1, %%mm2 \n\t"
2612  "paddb %%mm7, %%mm2 \n\t"
2613  "pcmpgtb %%mm6, %%mm2 \n\t"
2614  "paddb %%mm2, %%mm0 \n\t"
2615 
2616  "movq (%2, %3, 8), %%mm2 \n\t"
2617  PMAXUB(%%mm2, %%mm4)
2618  PMINUB(%%mm2, %%mm3, %%mm5)
2619  "psubb %%mm2, %%mm1 \n\t"
2620  "paddb %%mm7, %%mm1 \n\t"
2621  "pcmpgtb %%mm6, %%mm1 \n\t"
2622  "paddb %%mm1, %%mm0 \n\t"
2623 
2624  "movq (%%"REG_a", %3, 4), %%mm1 \n\t"
2625  "psubb %%mm1, %%mm2 \n\t"
2626  "paddb %%mm7, %%mm2 \n\t"
2627  "pcmpgtb %%mm6, %%mm2 \n\t"
2628  "paddb %%mm2, %%mm0 \n\t"
2629  "psubusb %%mm3, %%mm4 \n\t"
2630 
2631  "pxor %%mm6, %%mm6 \n\t"
2632  "movq %4, %%mm7 \n\t" // QP,..., QP
2633  "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
2634  "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0
2635  "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2636  "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2637  "movq %%mm7, %1 \n\t"
2638 
2639  "movq %5, %%mm7 \n\t"
2640  "punpcklbw %%mm7, %%mm7 \n\t"
2641  "punpcklbw %%mm7, %%mm7 \n\t"
2642  "punpcklbw %%mm7, %%mm7 \n\t"
2643  "psubb %%mm0, %%mm6 \n\t"
2644  "pcmpgtb %%mm7, %%mm6 \n\t"
2645  "movq %%mm6, %0 \n\t"
2646 
2647  : "=m" (eq_mask), "=m" (dc_mask)
2648  : "r" (src), "r" ((x86_reg)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
2649  : "%"REG_a
2650  );
2651 
2652  both_masks = dc_mask & eq_mask;
2653 
2654  if(both_masks){
2655  x86_reg offset= -8*step;
2656  int64_t *temp_sums= sums;
2657 
2658  __asm__ volatile(
2659  "movq %2, %%mm0 \n\t" // QP,..., QP
2660  "pxor %%mm4, %%mm4 \n\t"
2661 
2662  "movq (%0), %%mm6 \n\t"
2663  "movq (%0, %1), %%mm5 \n\t"
2664  "movq %%mm5, %%mm1 \n\t"
2665  "movq %%mm6, %%mm2 \n\t"
2666  "psubusb %%mm6, %%mm5 \n\t"
2667  "psubusb %%mm1, %%mm2 \n\t"
2668  "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2669  "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2670  "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2671 
2672  "pxor %%mm6, %%mm1 \n\t"
2673  "pand %%mm0, %%mm1 \n\t"
2674  "pxor %%mm1, %%mm6 \n\t"
2675  // 0:QP 6:First
2676 
2677  "movq (%0, %1, 8), %%mm5 \n\t"
2678  "add %1, %0 \n\t" // %0 points to line 1 not 0
2679  "movq (%0, %1, 8), %%mm7 \n\t"
2680  "movq %%mm5, %%mm1 \n\t"
2681  "movq %%mm7, %%mm2 \n\t"
2682  "psubusb %%mm7, %%mm5 \n\t"
2683  "psubusb %%mm1, %%mm2 \n\t"
2684  "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2685  "movq %2, %%mm0 \n\t" // QP,..., QP
2686  "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2687  "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2688 
2689  "pxor %%mm7, %%mm1 \n\t"
2690  "pand %%mm0, %%mm1 \n\t"
2691  "pxor %%mm1, %%mm7 \n\t"
2692 
2693  "movq %%mm6, %%mm5 \n\t"
2694  "punpckhbw %%mm4, %%mm6 \n\t"
2695  "punpcklbw %%mm4, %%mm5 \n\t"
2696  // 4:0 5/6:First 7:Last
2697 
2698  "movq %%mm5, %%mm0 \n\t"
2699  "movq %%mm6, %%mm1 \n\t"
2700  "psllw $2, %%mm0 \n\t"
2701  "psllw $2, %%mm1 \n\t"
2702  "paddw "MANGLE(w04)", %%mm0 \n\t"
2703  "paddw "MANGLE(w04)", %%mm1 \n\t"
2704 
2705 #define NEXT\
2706  "movq (%0), %%mm2 \n\t"\
2707  "movq (%0), %%mm3 \n\t"\
2708  "add %1, %0 \n\t"\
2709  "punpcklbw %%mm4, %%mm2 \n\t"\
2710  "punpckhbw %%mm4, %%mm3 \n\t"\
2711  "paddw %%mm2, %%mm0 \n\t"\
2712  "paddw %%mm3, %%mm1 \n\t"
2713 
2714 #define PREV\
2715  "movq (%0), %%mm2 \n\t"\
2716  "movq (%0), %%mm3 \n\t"\
2717  "add %1, %0 \n\t"\
2718  "punpcklbw %%mm4, %%mm2 \n\t"\
2719  "punpckhbw %%mm4, %%mm3 \n\t"\
2720  "psubw %%mm2, %%mm0 \n\t"\
2721  "psubw %%mm3, %%mm1 \n\t"
2722 
2723 
2724  NEXT //0
2725  NEXT //1
2726  NEXT //2
2727  "movq %%mm0, (%3) \n\t"
2728  "movq %%mm1, 8(%3) \n\t"
2729 
2730  NEXT //3
2731  "psubw %%mm5, %%mm0 \n\t"
2732  "psubw %%mm6, %%mm1 \n\t"
2733  "movq %%mm0, 16(%3) \n\t"
2734  "movq %%mm1, 24(%3) \n\t"
2735 
2736  NEXT //4
2737  "psubw %%mm5, %%mm0 \n\t"
2738  "psubw %%mm6, %%mm1 \n\t"
2739  "movq %%mm0, 32(%3) \n\t"
2740  "movq %%mm1, 40(%3) \n\t"
2741 
2742  NEXT //5
2743  "psubw %%mm5, %%mm0 \n\t"
2744  "psubw %%mm6, %%mm1 \n\t"
2745  "movq %%mm0, 48(%3) \n\t"
2746  "movq %%mm1, 56(%3) \n\t"
2747 
2748  NEXT //6
2749  "psubw %%mm5, %%mm0 \n\t"
2750  "psubw %%mm6, %%mm1 \n\t"
2751  "movq %%mm0, 64(%3) \n\t"
2752  "movq %%mm1, 72(%3) \n\t"
2753 
2754  "movq %%mm7, %%mm6 \n\t"
2755  "punpckhbw %%mm4, %%mm7 \n\t"
2756  "punpcklbw %%mm4, %%mm6 \n\t"
2757 
2758  NEXT //7
2759  "mov %4, %0 \n\t"
2760  "add %1, %0 \n\t"
2761  PREV //0
2762  "movq %%mm0, 80(%3) \n\t"
2763  "movq %%mm1, 88(%3) \n\t"
2764 
2765  PREV //1
2766  "paddw %%mm6, %%mm0 \n\t"
2767  "paddw %%mm7, %%mm1 \n\t"
2768  "movq %%mm0, 96(%3) \n\t"
2769  "movq %%mm1, 104(%3) \n\t"
2770 
2771  PREV //2
2772  "paddw %%mm6, %%mm0 \n\t"
2773  "paddw %%mm7, %%mm1 \n\t"
2774  "movq %%mm0, 112(%3) \n\t"
2775  "movq %%mm1, 120(%3) \n\t"
2776 
2777  PREV //3
2778  "paddw %%mm6, %%mm0 \n\t"
2779  "paddw %%mm7, %%mm1 \n\t"
2780  "movq %%mm0, 128(%3) \n\t"
2781  "movq %%mm1, 136(%3) \n\t"
2782 
2783  PREV //4
2784  "paddw %%mm6, %%mm0 \n\t"
2785  "paddw %%mm7, %%mm1 \n\t"
2786  "movq %%mm0, 144(%3) \n\t"
2787  "movq %%mm1, 152(%3) \n\t"
2788 
2789  "mov %4, %0 \n\t" //FIXME
2790 
2791  : "+&r"(src)
2792  : "r" ((x86_reg)step), "m" (c->pQPb), "r"(sums), "g"(src)
2793  );
2794 
2795  src+= step; // src points to begin of the 8x8 Block
2796 
2797  __asm__ volatile(
2798  "movq %4, %%mm6 \n\t"
2799  "pcmpeqb %%mm5, %%mm5 \n\t"
2800  "pxor %%mm6, %%mm5 \n\t"
2801  "pxor %%mm7, %%mm7 \n\t"
2802 
2803  "1: \n\t"
2804  "movq (%1), %%mm0 \n\t"
2805  "movq 8(%1), %%mm1 \n\t"
2806  "paddw 32(%1), %%mm0 \n\t"
2807  "paddw 40(%1), %%mm1 \n\t"
2808  "movq (%0, %3), %%mm2 \n\t"
2809  "movq %%mm2, %%mm3 \n\t"
2810  "movq %%mm2, %%mm4 \n\t"
2811  "punpcklbw %%mm7, %%mm2 \n\t"
2812  "punpckhbw %%mm7, %%mm3 \n\t"
2813  "paddw %%mm2, %%mm0 \n\t"
2814  "paddw %%mm3, %%mm1 \n\t"
2815  "paddw %%mm2, %%mm0 \n\t"
2816  "paddw %%mm3, %%mm1 \n\t"
2817  "psrlw $4, %%mm0 \n\t"
2818  "psrlw $4, %%mm1 \n\t"
2819  "packuswb %%mm1, %%mm0 \n\t"
2820  "pand %%mm6, %%mm0 \n\t"
2821  "pand %%mm5, %%mm4 \n\t"
2822  "por %%mm4, %%mm0 \n\t"
2823  "movq %%mm0, (%0, %3) \n\t"
2824  "add $16, %1 \n\t"
2825  "add %2, %0 \n\t"
2826  " js 1b \n\t"
2827 
2828  : "+r"(offset), "+r"(temp_sums)
2829  : "r" ((x86_reg)step), "r"(src - offset), "m"(both_masks)
2830  );
2831  }else
2832  src+= step; // src points to begin of the 8x8 Block
2833 
2834  if(eq_mask != -1LL){
2835  uint8_t *temp_src= src;
2836  DECLARE_ALIGNED(8, uint64_t, tmp)[4]; // make space for 4 8-byte vars
2837  __asm__ volatile(
2838  "pxor %%mm7, %%mm7 \n\t"
2839 // 0 1 2 3 4 5 6 7 8 9
2840 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1
2841 
2842  "movq (%0), %%mm0 \n\t"
2843  "movq %%mm0, %%mm1 \n\t"
2844  "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
2845  "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
2846 
2847  "movq (%0, %1), %%mm2 \n\t"
2848  "lea (%0, %1, 2), %%"REG_a" \n\t"
2849  "movq %%mm2, %%mm3 \n\t"
2850  "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
2851  "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
2852 
2853  "movq (%%"REG_a"), %%mm4 \n\t"
2854  "movq %%mm4, %%mm5 \n\t"
2855  "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
2856  "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
2857 
2858  "paddw %%mm0, %%mm0 \n\t" // 2L0
2859  "paddw %%mm1, %%mm1 \n\t" // 2H0
2860  "psubw %%mm4, %%mm2 \n\t" // L1 - L2
2861  "psubw %%mm5, %%mm3 \n\t" // H1 - H2
2862  "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
2863  "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
2864 
2865  "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
2866  "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
2867  "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
2868  "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
2869 
2870  "movq (%%"REG_a", %1), %%mm2 \n\t"
2871  "movq %%mm2, %%mm3 \n\t"
2872  "punpcklbw %%mm7, %%mm2 \n\t" // L3
2873  "punpckhbw %%mm7, %%mm3 \n\t" // H3
2874 
2875  "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
2876  "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
2877  "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2878  "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2879  "movq %%mm0, (%4) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2880  "movq %%mm1, 8(%4) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2881 
2882  "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
2883  "movq %%mm0, %%mm1 \n\t"
2884  "punpcklbw %%mm7, %%mm0 \n\t" // L4
2885  "punpckhbw %%mm7, %%mm1 \n\t" // H4
2886 
2887  "psubw %%mm0, %%mm2 \n\t" // L3 - L4
2888  "psubw %%mm1, %%mm3 \n\t" // H3 - H4
2889  "movq %%mm2, 16(%4) \n\t" // L3 - L4
2890  "movq %%mm3, 24(%4) \n\t" // H3 - H4
2891  "paddw %%mm4, %%mm4 \n\t" // 2L2
2892  "paddw %%mm5, %%mm5 \n\t" // 2H2
2893  "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
2894  "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
2895 
2896  "lea (%%"REG_a", %1), %0 \n\t"
2897  "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
2898  "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
2899  "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
2900  "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
2901 //50 opcodes so far
2902  "movq (%0, %1, 2), %%mm2 \n\t"
2903  "movq %%mm2, %%mm3 \n\t"
2904  "punpcklbw %%mm7, %%mm2 \n\t" // L5
2905  "punpckhbw %%mm7, %%mm3 \n\t" // H5
2906  "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
2907  "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
2908  "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
2909  "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
2910 
2911  "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
2912  "punpcklbw %%mm7, %%mm6 \n\t" // L6
2913  "psubw %%mm6, %%mm2 \n\t" // L5 - L6
2914  "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
2915  "punpckhbw %%mm7, %%mm6 \n\t" // H6
2916  "psubw %%mm6, %%mm3 \n\t" // H5 - H6
2917 
2918  "paddw %%mm0, %%mm0 \n\t" // 2L4
2919  "paddw %%mm1, %%mm1 \n\t" // 2H4
2920  "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
2921  "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
2922 
2923  "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
2924  "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
2925  "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
2926  "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
2927 
2928  "movq (%0, %1, 4), %%mm2 \n\t"
2929  "movq %%mm2, %%mm3 \n\t"
2930  "punpcklbw %%mm7, %%mm2 \n\t" // L7
2931  "punpckhbw %%mm7, %%mm3 \n\t" // H7
2932 
2933  "paddw %%mm2, %%mm2 \n\t" // 2L7
2934  "paddw %%mm3, %%mm3 \n\t" // 2H7
2935  "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
2936  "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
2937 
2938  "movq (%4), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2939  "movq 8(%4), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2940 
2941 #if TEMPLATE_PP_MMXEXT
2942  "movq %%mm7, %%mm6 \n\t" // 0
2943  "psubw %%mm0, %%mm6 \n\t"
2944  "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
2945  "movq %%mm7, %%mm6 \n\t" // 0
2946  "psubw %%mm1, %%mm6 \n\t"
2947  "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
2948  "movq %%mm7, %%mm6 \n\t" // 0
2949  "psubw %%mm2, %%mm6 \n\t"
2950  "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
2951  "movq %%mm7, %%mm6 \n\t" // 0
2952  "psubw %%mm3, %%mm6 \n\t"
2953  "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
2954 #else
2955  "movq %%mm7, %%mm6 \n\t" // 0
2956  "pcmpgtw %%mm0, %%mm6 \n\t"
2957  "pxor %%mm6, %%mm0 \n\t"
2958  "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
2959  "movq %%mm7, %%mm6 \n\t" // 0
2960  "pcmpgtw %%mm1, %%mm6 \n\t"
2961  "pxor %%mm6, %%mm1 \n\t"
2962  "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
2963  "movq %%mm7, %%mm6 \n\t" // 0
2964  "pcmpgtw %%mm2, %%mm6 \n\t"
2965  "pxor %%mm6, %%mm2 \n\t"
2966  "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
2967  "movq %%mm7, %%mm6 \n\t" // 0
2968  "pcmpgtw %%mm3, %%mm6 \n\t"
2969  "pxor %%mm6, %%mm3 \n\t"
2970  "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
2971 #endif
2972 
2973 #if TEMPLATE_PP_MMXEXT
2974  "pminsw %%mm2, %%mm0 \n\t"
2975  "pminsw %%mm3, %%mm1 \n\t"
2976 #else
2977  "movq %%mm0, %%mm6 \n\t"
2978  "psubusw %%mm2, %%mm6 \n\t"
2979  "psubw %%mm6, %%mm0 \n\t"
2980  "movq %%mm1, %%mm6 \n\t"
2981  "psubusw %%mm3, %%mm6 \n\t"
2982  "psubw %%mm6, %%mm1 \n\t"
2983 #endif
2984 
2985  "movd %2, %%mm2 \n\t" // QP
2986  "punpcklbw %%mm7, %%mm2 \n\t"
2987 
2988  "movq %%mm7, %%mm6 \n\t" // 0
2989  "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
2990  "pxor %%mm6, %%mm4 \n\t"
2991  "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
2992  "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
2993  "pxor %%mm7, %%mm5 \n\t"
2994  "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
2995 // 100 opcodes
2996  "psllw $3, %%mm2 \n\t" // 8QP
2997  "movq %%mm2, %%mm3 \n\t" // 8QP
2998  "pcmpgtw %%mm4, %%mm2 \n\t"
2999  "pcmpgtw %%mm5, %%mm3 \n\t"
3000  "pand %%mm2, %%mm4 \n\t"
3001  "pand %%mm3, %%mm5 \n\t"
3002 
3003 
3004  "psubusw %%mm0, %%mm4 \n\t" // hd
3005  "psubusw %%mm1, %%mm5 \n\t" // ld
3006 
3007 
3008  "movq "MANGLE(w05)", %%mm2 \n\t" // 5
3009  "pmullw %%mm2, %%mm4 \n\t"
3010  "pmullw %%mm2, %%mm5 \n\t"
3011  "movq "MANGLE(w20)", %%mm2 \n\t" // 32
3012  "paddw %%mm2, %%mm4 \n\t"
3013  "paddw %%mm2, %%mm5 \n\t"
3014  "psrlw $6, %%mm4 \n\t"
3015  "psrlw $6, %%mm5 \n\t"
3016 
3017  "movq 16(%4), %%mm0 \n\t" // L3 - L4
3018  "movq 24(%4), %%mm1 \n\t" // H3 - H4
3019 
3020  "pxor %%mm2, %%mm2 \n\t"
3021  "pxor %%mm3, %%mm3 \n\t"
3022 
3023  "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
3024  "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
3025  "pxor %%mm2, %%mm0 \n\t"
3026  "pxor %%mm3, %%mm1 \n\t"
3027  "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
3028  "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
3029  "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
3030  "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
3031 
3032  "pxor %%mm6, %%mm2 \n\t"
3033  "pxor %%mm7, %%mm3 \n\t"
3034  "pand %%mm2, %%mm4 \n\t"
3035  "pand %%mm3, %%mm5 \n\t"
3036 
3037 #if TEMPLATE_PP_MMXEXT
3038  "pminsw %%mm0, %%mm4 \n\t"
3039  "pminsw %%mm1, %%mm5 \n\t"
3040 #else
3041  "movq %%mm4, %%mm2 \n\t"
3042  "psubusw %%mm0, %%mm2 \n\t"
3043  "psubw %%mm2, %%mm4 \n\t"
3044  "movq %%mm5, %%mm2 \n\t"
3045  "psubusw %%mm1, %%mm2 \n\t"
3046  "psubw %%mm2, %%mm5 \n\t"
3047 #endif
3048  "pxor %%mm6, %%mm4 \n\t"
3049  "pxor %%mm7, %%mm5 \n\t"
3050  "psubw %%mm6, %%mm4 \n\t"
3051  "psubw %%mm7, %%mm5 \n\t"
3052  "packsswb %%mm5, %%mm4 \n\t"
3053  "movq %3, %%mm1 \n\t"
3054  "pandn %%mm4, %%mm1 \n\t"
3055  "movq (%0), %%mm0 \n\t"
3056  "paddb %%mm1, %%mm0 \n\t"
3057  "movq %%mm0, (%0) \n\t"
3058  "movq (%0, %1), %%mm0 \n\t"
3059  "psubb %%mm1, %%mm0 \n\t"
3060  "movq %%mm0, (%0, %1) \n\t"
3061 
3062  : "+r" (temp_src)
3063  : "r" ((x86_reg)step), "m" (c->pQPb), "m"(eq_mask), "r"(tmp)
3064  : "%"REG_a
3065  );
3066  }
3067 /*if(step==16){
3068  STOP_TIMER("step16")
3069 }else{
3070  STOP_TIMER("stepX")
3071 }
3072  } */
3073 }
3074 #endif //TEMPLATE_PP_MMX
3075 
3076 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
3077  const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c);
3078 
3079 /**
3080  * Copy a block from src to dst and fixes the blacklevel.
3081  * levelFix == 0 -> do not touch the brightness & contrast
3082  */
3083 #undef REAL_SCALED_CPY
3084 #undef SCALED_CPY
3085 
3086 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, const uint8_t src[], int srcStride,
3087  int levelFix, int64_t *packedOffsetAndScale)
3088 {
3089 #if !TEMPLATE_PP_MMX
3090  int i;
3091 #endif
3092  if(levelFix){
3093 #if TEMPLATE_PP_MMX
3094  __asm__ volatile(
3095  "movq (%%"REG_a"), %%mm2 \n\t" // packedYOffset
3096  "movq 8(%%"REG_a"), %%mm3 \n\t" // packedYScale
3097  "lea (%2,%4), %%"REG_a" \n\t"
3098  "lea (%3,%5), %%"REG_d" \n\t"
3099  "pxor %%mm4, %%mm4 \n\t"
3100 #if TEMPLATE_PP_MMXEXT
3101 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \
3102  "movq " #src1 ", %%mm0 \n\t"\
3103  "movq " #src1 ", %%mm5 \n\t"\
3104  "movq " #src2 ", %%mm1 \n\t"\
3105  "movq " #src2 ", %%mm6 \n\t"\
3106  "punpcklbw %%mm0, %%mm0 \n\t"\
3107  "punpckhbw %%mm5, %%mm5 \n\t"\
3108  "punpcklbw %%mm1, %%mm1 \n\t"\
3109  "punpckhbw %%mm6, %%mm6 \n\t"\
3110  "pmulhuw %%mm3, %%mm0 \n\t"\
3111  "pmulhuw %%mm3, %%mm5 \n\t"\
3112  "pmulhuw %%mm3, %%mm1 \n\t"\
3113  "pmulhuw %%mm3, %%mm6 \n\t"\
3114  "psubw %%mm2, %%mm0 \n\t"\
3115  "psubw %%mm2, %%mm5 \n\t"\
3116  "psubw %%mm2, %%mm1 \n\t"\
3117  "psubw %%mm2, %%mm6 \n\t"\
3118  "packuswb %%mm5, %%mm0 \n\t"\
3119  "packuswb %%mm6, %%mm1 \n\t"\
3120  "movq %%mm0, " #dst1 " \n\t"\
3121  "movq %%mm1, " #dst2 " \n\t"\
3122 
3123 #else //TEMPLATE_PP_MMXEXT
3124 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \
3125  "movq " #src1 ", %%mm0 \n\t"\
3126  "movq " #src1 ", %%mm5 \n\t"\
3127  "punpcklbw %%mm4, %%mm0 \n\t"\
3128  "punpckhbw %%mm4, %%mm5 \n\t"\
3129  "psubw %%mm2, %%mm0 \n\t"\
3130  "psubw %%mm2, %%mm5 \n\t"\
3131  "movq " #src2 ", %%mm1 \n\t"\
3132  "psllw $6, %%mm0 \n\t"\
3133  "psllw $6, %%mm5 \n\t"\
3134  "pmulhw %%mm3, %%mm0 \n\t"\
3135  "movq " #src2 ", %%mm6 \n\t"\
3136  "pmulhw %%mm3, %%mm5 \n\t"\
3137  "punpcklbw %%mm4, %%mm1 \n\t"\
3138  "punpckhbw %%mm4, %%mm6 \n\t"\
3139  "psubw %%mm2, %%mm1 \n\t"\
3140  "psubw %%mm2, %%mm6 \n\t"\
3141  "psllw $6, %%mm1 \n\t"\
3142  "psllw $6, %%mm6 \n\t"\
3143  "pmulhw %%mm3, %%mm1 \n\t"\
3144  "pmulhw %%mm3, %%mm6 \n\t"\
3145  "packuswb %%mm5, %%mm0 \n\t"\
3146  "packuswb %%mm6, %%mm1 \n\t"\
3147  "movq %%mm0, " #dst1 " \n\t"\
3148  "movq %%mm1, " #dst2 " \n\t"\
3149 
3150 #endif //TEMPLATE_PP_MMXEXT
3151 #define SCALED_CPY(src1, src2, dst1, dst2)\
3152  REAL_SCALED_CPY(src1, src2, dst1, dst2)
3153 
3154 SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5))
3155 SCALED_CPY((%2, %4, 2), (%%REGa, %4, 2), (%3, %5, 2), (%%REGd, %5, 2))
3156 SCALED_CPY((%2, %4, 4), (%%REGa, %4, 4), (%3, %5, 4), (%%REGd, %5, 4))
3157  "lea (%%"REG_a",%4,4), %%"REG_a" \n\t"
3158  "lea (%%"REG_d",%5,4), %%"REG_d" \n\t"
3159 SCALED_CPY((%%REGa, %4), (%%REGa, %4, 2), (%%REGd, %5), (%%REGd, %5, 2))
3160 
3161 
3162  : "=&a" (packedOffsetAndScale)
3163  : "0" (packedOffsetAndScale),
3164  "r"(src),
3165  "r"(dst),
3166  "r" ((x86_reg)srcStride),
3167  "r" ((x86_reg)dstStride)
3168  : "%"REG_d
3169  );
3170 #else //TEMPLATE_PP_MMX
3171  for(i=0; i<8; i++)
3172  memcpy( &(dst[dstStride*i]),
3173  &(src[srcStride*i]), BLOCK_SIZE);
3174 #endif //TEMPLATE_PP_MMX
3175  }else{
3176 #if TEMPLATE_PP_MMX
3177  __asm__ volatile(
3178  "lea (%0,%2), %%"REG_a" \n\t"
3179  "lea (%1,%3), %%"REG_d" \n\t"
3180 
3181 #define REAL_SIMPLE_CPY(src1, src2, dst1, dst2) \
3182  "movq " #src1 ", %%mm0 \n\t"\
3183  "movq " #src2 ", %%mm1 \n\t"\
3184  "movq %%mm0, " #dst1 " \n\t"\
3185  "movq %%mm1, " #dst2 " \n\t"\
3186 
3187 #define SIMPLE_CPY(src1, src2, dst1, dst2)\
3188  REAL_SIMPLE_CPY(src1, src2, dst1, dst2)
3189 
3190 SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3))
3191 SIMPLE_CPY((%0, %2, 2), (%%REGa, %2, 2), (%1, %3, 2), (%%REGd, %3, 2))
3192 SIMPLE_CPY((%0, %2, 4), (%%REGa, %2, 4), (%1, %3, 4), (%%REGd, %3, 4))
3193  "lea (%%"REG_a",%2,4), %%"REG_a" \n\t"
3194  "lea (%%"REG_d",%3,4), %%"REG_d" \n\t"
3195 SIMPLE_CPY((%%REGa, %2), (%%REGa, %2, 2), (%%REGd, %3), (%%REGd, %3, 2))
3196 
3197  : : "r" (src),
3198  "r" (dst),
3199  "r" ((x86_reg)srcStride),
3200  "r" ((x86_reg)dstStride)
3201  : "%"REG_a, "%"REG_d
3202  );
3203 #else //TEMPLATE_PP_MMX
3204  for(i=0; i<8; i++)
3205  memcpy( &(dst[dstStride*i]),
3206  &(src[srcStride*i]), BLOCK_SIZE);
3207 #endif //TEMPLATE_PP_MMX
3208  }
3209 }
3210 
3211 /**
3212  * Duplicate the given 8 src pixels ? times upward
3213  */
3214 static inline void RENAME(duplicate)(uint8_t src[], int stride)
3215 {
3216 #if TEMPLATE_PP_MMX
3217  __asm__ volatile(
3218  "movq (%0), %%mm0 \n\t"
3219  "movq %%mm0, (%0, %1, 4) \n\t"
3220  "add %1, %0 \n\t"
3221  "movq %%mm0, (%0) \n\t"
3222  "movq %%mm0, (%0, %1) \n\t"
3223  "movq %%mm0, (%0, %1, 2) \n\t"
3224  "movq %%mm0, (%0, %1, 4) \n\t"
3225  : "+r" (src)
3226  : "r" ((x86_reg)-stride)
3227  );
3228 #else
3229  int i;
3230  uint8_t *p=src;
3231  for(i=0; i<5; i++){
3232  p-= stride;
3233  memcpy(p, src, 8);
3234  }
3235 #endif
3236 }
3237 
3238 /**
3239  * Filter array of bytes (Y or U or V values)
3240  */
3241 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
3242  const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)
3243 {
3244  DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access
3245  int x,y;
3246 #ifdef TEMPLATE_PP_TIME_MODE
3247  const int mode= TEMPLATE_PP_TIME_MODE;
3248 #else
3249  const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;
3250 #endif
3251  int black=0, white=255; // blackest black and whitest white in the picture
3252  int QPCorrecture= 256*256;
3253 
3254  int copyAhead;
3255 #if TEMPLATE_PP_MMX
3256  int i;
3257 #endif
3258 
3259  const int qpHShift= isColor ? 4-c.hChromaSubSample : 4;
3260  const int qpVShift= isColor ? 4-c.vChromaSubSample : 4;
3261 
3262  //FIXME remove
3263  uint64_t * const yHistogram= c.yHistogram;
3264  uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride;
3265  uint8_t * const tempDst= (dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride) + 32;
3266  //const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
3267 
3268 #if TEMPLATE_PP_MMX
3269  for(i=0; i<57; i++){
3270  int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;
3271  int threshold= offset*2 + 1;
3272  c.mmxDcOffset[i]= 0x7F - offset;
3273  c.mmxDcThreshold[i]= 0x7F - threshold;
3274  c.mmxDcOffset[i]*= 0x0101010101010101LL;
3275  c.mmxDcThreshold[i]*= 0x0101010101010101LL;
3276  }
3277 #endif
3278 
3279  if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
3280  else if( (mode & LINEAR_BLEND_DEINT_FILTER)
3281  || (mode & FFMPEG_DEINT_FILTER)
3282  || (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14;
3283  else if( (mode & V_DEBLOCK)
3284  || (mode & LINEAR_IPOL_DEINT_FILTER)
3285  || (mode & MEDIAN_DEINT_FILTER)
3286  || (mode & V_A_DEBLOCK)) copyAhead=13;
3287  else if(mode & V_X1_FILTER) copyAhead=11;
3288 // else if(mode & V_RK1_FILTER) copyAhead=10;
3289  else if(mode & DERING) copyAhead=9;
3290  else copyAhead=8;
3291 
3292  copyAhead-= 8;
3293 
3294  if(!isColor){
3295  uint64_t sum= 0;
3296  int i;
3297  uint64_t maxClipped;
3298  uint64_t clipped;
3299  double scale;
3300 
3301  c.frameNum++;
3302  // first frame is fscked so we ignore it
3303  if(c.frameNum == 1) yHistogram[0]= width*(uint64_t)height/64*15/256;
3304 
3305  for(i=0; i<256; i++){
3306  sum+= yHistogram[i];
3307  }
3308 
3309  /* We always get a completely black picture first. */
3310  maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold);
3311 
3312  clipped= sum;
3313  for(black=255; black>0; black--){
3314  if(clipped < maxClipped) break;
3315  clipped-= yHistogram[black];
3316  }
3317 
3318  clipped= sum;
3319  for(white=0; white<256; white++){
3320  if(clipped < maxClipped) break;
3321  clipped-= yHistogram[white];
3322  }
3323 
3324  scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);
3325 
3326 #if TEMPLATE_PP_MMXEXT
3327  c.packedYScale= (uint16_t)(scale*256.0 + 0.5);
3328  c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;
3329 #else
3330  c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);
3331  c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;
3332 #endif
3333 
3334  c.packedYOffset|= c.packedYOffset<<32;
3335  c.packedYOffset|= c.packedYOffset<<16;
3336 
3337  c.packedYScale|= c.packedYScale<<32;
3338  c.packedYScale|= c.packedYScale<<16;
3339 
3340  if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5);
3341  else QPCorrecture= 256*256;
3342  }else{
3343  c.packedYScale= 0x0100010001000100LL;
3344  c.packedYOffset= 0;
3345  QPCorrecture= 256*256;
3346  }
3347 
3348  /* copy & deinterlace first row of blocks */
3349  y=-BLOCK_SIZE;
3350  {
3351  const uint8_t *srcBlock= &(src[y*srcStride]);
3352  uint8_t *dstBlock= tempDst + dstStride;
3353 
3354  // From this point on it is guaranteed that we can read and write 16 lines downward
3355  // finish 1 block before the next otherwise we might have a problem
3356  // with the L1 Cache of the P4 ... or only a few blocks at a time or something
3357  for(x=0; x<width; x+=BLOCK_SIZE){
3358 
3359 #if TEMPLATE_PP_MMXEXT
3360 /*
3361  prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3362  prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3363  prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3364  prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3365 */
3366 
3367  __asm__(
3368  "mov %4, %%"REG_a" \n\t"
3369  "shr $2, %%"REG_a" \n\t"
3370  "and $6, %%"REG_a" \n\t"
3371  "add %5, %%"REG_a" \n\t"
3372  "mov %%"REG_a", %%"REG_d" \n\t"
3373  "imul %1, %%"REG_a" \n\t"
3374  "imul %3, %%"REG_d" \n\t"
3375  "prefetchnta 32(%%"REG_a", %0) \n\t"
3376  "prefetcht0 32(%%"REG_d", %2) \n\t"
3377  "add %1, %%"REG_a" \n\t"
3378  "add %3, %%"REG_d" \n\t"
3379  "prefetchnta 32(%%"REG_a", %0) \n\t"
3380  "prefetcht0 32(%%"REG_d", %2) \n\t"
3381  :: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride),
3382  "g" ((x86_reg)x), "g" ((x86_reg)copyAhead)
3383  : "%"REG_a, "%"REG_d
3384  );
3385 
3386 #elif TEMPLATE_PP_3DNOW
3387 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3388 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3389  prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3390  prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3391  prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3392 */
3393 #endif
3394 
3395  RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
3396  srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3397 
3398  RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
3399 
3400  if(mode & LINEAR_IPOL_DEINT_FILTER)
3401  RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3402  else if(mode & LINEAR_BLEND_DEINT_FILTER)
3403  RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3404  else if(mode & MEDIAN_DEINT_FILTER)
3405  RENAME(deInterlaceMedian)(dstBlock, dstStride);
3406  else if(mode & CUBIC_IPOL_DEINT_FILTER)
3407  RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3408  else if(mode & FFMPEG_DEINT_FILTER)
3409  RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3410  else if(mode & LOWPASS5_DEINT_FILTER)
3411  RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3412 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3413  RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3414 */
3415  dstBlock+=8;
3416  srcBlock+=8;
3417  }
3418  if(width==FFABS(dstStride))
3419  linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride);
3420  else{
3421  int i;
3422  for(i=0; i<copyAhead; i++){
3423  memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width);
3424  }
3425  }
3426  }
3427 
3428  for(y=0; y<height; y+=BLOCK_SIZE){
3429  //1% speedup if these are here instead of the inner loop
3430  const uint8_t *srcBlock= &(src[y*srcStride]);
3431  uint8_t *dstBlock= &(dst[y*dstStride]);
3432 #if TEMPLATE_PP_MMX
3433  uint8_t *tempBlock1= c.tempBlocks;
3434  uint8_t *tempBlock2= c.tempBlocks + 8;
3435 #endif
3436  const int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
3437  int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)];
3438  int QP=0;
3439  /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
3440  if not than use a temporary buffer */
3441  if(y+15 >= height){
3442  int i;
3443  /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
3444  blockcopy to dst later */
3445  linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
3446  FFMAX(height-y-copyAhead, 0), srcStride);
3447 
3448  /* duplicate last line of src to fill the void up to line (copyAhead+7) */
3449  for(i=FFMAX(height-y, 8); i<copyAhead+8; i++)
3450  memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), FFABS(srcStride));
3451 
3452  /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
3453  linecpy(tempDst, dstBlock - dstStride, FFMIN(height-y+1, copyAhead+1), dstStride);
3454 
3455  /* duplicate last line of dst to fill the void up to line (copyAhead) */
3456  for(i=height-y+1; i<=copyAhead; i++)
3457  memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), FFABS(dstStride));
3458 
3459  dstBlock= tempDst + dstStride;
3460  srcBlock= tempSrc;
3461  }
3462 
3463  // From this point on it is guaranteed that we can read and write 16 lines downward
3464  // finish 1 block before the next otherwise we might have a problem
3465  // with the L1 Cache of the P4 ... or only a few blocks at a time or something
3466  for(x=0; x<width; x+=BLOCK_SIZE){
3467  const int stride= dstStride;
3468 #if TEMPLATE_PP_MMX
3469  uint8_t *tmpXchg;
3470 #endif
3471  if(isColor){
3472  QP= QPptr[x>>qpHShift];
3473  c.nonBQP= nonBQPptr[x>>qpHShift];
3474  }else{
3475  QP= QPptr[x>>4];
3476  QP= (QP* QPCorrecture + 256*128)>>16;
3477  c.nonBQP= nonBQPptr[x>>4];
3478  c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
3479  yHistogram[ srcBlock[srcStride*12 + 4] ]++;
3480  }
3481  c.QP= QP;
3482 #if TEMPLATE_PP_MMX
3483  __asm__ volatile(
3484  "movd %1, %%mm7 \n\t"
3485  "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
3486  "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
3487  "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
3488  "movq %%mm7, %0 \n\t"
3489  : "=m" (c.pQPb)
3490  : "r" (QP)
3491  );
3492 #endif
3493 
3494 
3495 #if TEMPLATE_PP_MMXEXT
3496 /*
3497  prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3498  prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3499  prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3500  prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3501 */
3502 
3503  __asm__(
3504  "mov %4, %%"REG_a" \n\t"
3505  "shr $2, %%"REG_a" \n\t"
3506  "and $6, %%"REG_a" \n\t"
3507  "add %5, %%"REG_a" \n\t"
3508  "mov %%"REG_a", %%"REG_d" \n\t"
3509  "imul %1, %%"REG_a" \n\t"
3510  "imul %3, %%"REG_d" \n\t"
3511  "prefetchnta 32(%%"REG_a", %0) \n\t"
3512  "prefetcht0 32(%%"REG_d", %2) \n\t"
3513  "add %1, %%"REG_a" \n\t"
3514  "add %3, %%"REG_d" \n\t"
3515  "prefetchnta 32(%%"REG_a", %0) \n\t"
3516  "prefetcht0 32(%%"REG_d", %2) \n\t"
3517  :: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride),
3518  "g" ((x86_reg)x), "g" ((x86_reg)copyAhead)
3519  : "%"REG_a, "%"REG_d
3520  );
3521 
3522 #elif TEMPLATE_PP_3DNOW
3523 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3524 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3525  prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3526  prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3527  prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3528 */
3529 #endif
3530 
3531  RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
3532  srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3533 
3534  if(mode & LINEAR_IPOL_DEINT_FILTER)
3535  RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3536  else if(mode & LINEAR_BLEND_DEINT_FILTER)
3537  RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3538  else if(mode & MEDIAN_DEINT_FILTER)
3539  RENAME(deInterlaceMedian)(dstBlock, dstStride);
3540  else if(mode & CUBIC_IPOL_DEINT_FILTER)
3541  RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3542  else if(mode & FFMPEG_DEINT_FILTER)
3543  RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3544  else if(mode & LOWPASS5_DEINT_FILTER)
3545  RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3546 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3547  RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3548 */
3549 
3550  /* only deblock if we have 2 blocks */
3551  if(y + 8 < height){
3552  if(mode & V_X1_FILTER)
3553  RENAME(vertX1Filter)(dstBlock, stride, &c);
3554  else if(mode & V_DEBLOCK){
3555  const int t= RENAME(vertClassify)(dstBlock, stride, &c);
3556 
3557  if(t==1)
3558  RENAME(doVertLowPass)(dstBlock, stride, &c);
3559  else if(t==2)
3560  RENAME(doVertDefFilter)(dstBlock, stride, &c);
3561  }else if(mode & V_A_DEBLOCK){
3562  RENAME(do_a_deblock)(dstBlock, stride, 1, &c);
3563  }
3564  }
3565 
3566 #if TEMPLATE_PP_MMX
3567  RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
3568 #endif
3569  /* check if we have a previous block to deblock it with dstBlock */
3570  if(x - 8 >= 0){
3571 #if TEMPLATE_PP_MMX
3572  if(mode & H_X1_FILTER)
3573  RENAME(vertX1Filter)(tempBlock1, 16, &c);
3574  else if(mode & H_DEBLOCK){
3575 //START_TIMER
3576  const int t= RENAME(vertClassify)(tempBlock1, 16, &c);
3577 //STOP_TIMER("dc & minmax")
3578  if(t==1)
3579  RENAME(doVertLowPass)(tempBlock1, 16, &c);
3580  else if(t==2)
3581  RENAME(doVertDefFilter)(tempBlock1, 16, &c);
3582  }else if(mode & H_A_DEBLOCK){
3583  RENAME(do_a_deblock)(tempBlock1, 16, 1, &c);
3584  }
3585 
3586  RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
3587 
3588 #else
3589  if(mode & H_X1_FILTER)
3590  horizX1Filter(dstBlock-4, stride, QP);
3591  else if(mode & H_DEBLOCK){
3592 #if TEMPLATE_PP_ALTIVEC
3593  DECLARE_ALIGNED(16, unsigned char, tempBlock)[272];
3594  int t;
3595  transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride);
3596 
3597  t = vertClassify_altivec(tempBlock-48, 16, &c);
3598  if(t==1) {
3599  doVertLowPass_altivec(tempBlock-48, 16, &c);
3600  transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3601  }
3602  else if(t==2) {
3603  doVertDefFilter_altivec(tempBlock-48, 16, &c);
3604  transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3605  }
3606 #else
3607  const int t= RENAME(horizClassify)(dstBlock-4, stride, &c);
3608 
3609  if(t==1)
3610  RENAME(doHorizLowPass)(dstBlock-4, stride, &c);
3611  else if(t==2)
3612  RENAME(doHorizDefFilter)(dstBlock-4, stride, &c);
3613 #endif
3614  }else if(mode & H_A_DEBLOCK){
3615  RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c);
3616  }
3617 #endif //TEMPLATE_PP_MMX
3618  if(mode & DERING){
3619  //FIXME filter first line
3620  if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);
3621  }
3622 
3623  if(mode & TEMP_NOISE_FILTER)
3624  {
3625  RENAME(tempNoiseReducer)(dstBlock-8, stride,
3626  c.tempBlurred[isColor] + y*dstStride + x,
3627  c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3) + 256,
3628  c.ppMode.maxTmpNoise);
3629  }
3630  }
3631 
3632  dstBlock+=8;
3633  srcBlock+=8;
3634 
3635 #if TEMPLATE_PP_MMX
3636  tmpXchg= tempBlock1;
3637  tempBlock1= tempBlock2;
3638  tempBlock2 = tmpXchg;
3639 #endif
3640  }
3641 
3642  if(mode & DERING){
3643  if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);
3644  }
3645 
3646  if((mode & TEMP_NOISE_FILTER)){
3647  RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
3648  c.tempBlurred[isColor] + y*dstStride + x,
3649  c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3) + 256,
3650  c.ppMode.maxTmpNoise);
3651  }
3652 
3653  /* did we use a tmp buffer for the last lines*/
3654  if(y+15 >= height){
3655  uint8_t *dstBlock= &(dst[y*dstStride]);
3656  if(width==FFABS(dstStride))
3657  linecpy(dstBlock, tempDst + dstStride, height-y, dstStride);
3658  else{
3659  int i;
3660  for(i=0; i<height-y; i++){
3661  memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width);
3662  }
3663  }
3664  }
3665 /*
3666  for(x=0; x<width; x+=32){
3667  volatile int i;
3668  i+= dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
3669  + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
3670  + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
3671  + dstBlock[x +13*dstStride]
3672  + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
3673  }*/
3674  }
3675 #if TEMPLATE_PP_3DNOW
3676  __asm__ volatile("femms");
3677 #elif TEMPLATE_PP_MMX
3678  __asm__ volatile("emms");
3679 #endif
3680 
3681 #ifdef DEBUG_BRIGHTNESS
3682  if(!isColor){
3683  int max=1;
3684  int i;
3685  for(i=0; i<256; i++)
3686  if(yHistogram[i] > max) max=yHistogram[i];
3687 
3688  for(i=1; i<256; i++){
3689  int x;
3690  int start=yHistogram[i-1]/(max/256+1);
3691  int end=yHistogram[i]/(max/256+1);
3692  int inc= end > start ? 1 : -1;
3693  for(x=start; x!=end+inc; x+=inc)
3694  dst[ i*dstStride + x]+=128;
3695  }
3696 
3697  for(i=0; i<100; i+=2){
3698  dst[ (white)*dstStride + i]+=128;
3699  dst[ (black)*dstStride + i]+=128;
3700  }
3701  }
3702 #endif
3703 
3704  *c2= c; //copy local context back
3705 
3706 }
3707 
3708 #undef RENAME
3709 #undef TEMPLATE_PP_C
3710 #undef TEMPLATE_PP_ALTIVEC
3711 #undef TEMPLATE_PP_MMX
3712 #undef TEMPLATE_PP_MMXEXT
3713 #undef TEMPLATE_PP_3DNOW
3714 #undef TEMPLATE_PP_SSE2