daf22cb299c7f80ba572cef5ae00433d8a9e0528
[ffmpeg.git] / postproc / postprocess_template.c
1 /*
2     Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at)
3
4     This program is free software; you can redistribute it and/or modify
5     it under the terms of the GNU General Public License as published by
6     the Free Software Foundation; either version 2 of the License, or
7     (at your option) any later version.
8
9     This program is distributed in the hope that it will be useful,
10     but WITHOUT ANY WARRANTY; without even the implied warranty of
11     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12     GNU General Public License for more details.
13
14     You should have received a copy of the GNU General Public License
15     along with this program; if not, write to the Free Software
16     Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
17 */
18
19 #undef PAVGB
20 #undef PMINUB
21 #undef PMAXUB
22
23 #ifdef HAVE_MMX2
24 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
25 #elif defined (HAVE_3DNOW)
26 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
27 #endif
28
29 #ifdef HAVE_MMX2
30 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
31 #elif defined (HAVE_MMX)
32 #define PMINUB(b,a,t) \
33         "movq " #a ", " #t " \n\t"\
34         "psubusb " #b ", " #t " \n\t"\
35         "psubb " #t ", " #a " \n\t"
36 #endif
37
38 #ifdef HAVE_MMX2
39 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
40 #elif defined (HAVE_MMX)
41 #define PMAXUB(a,b) \
42         "psubusb " #a ", " #b " \n\t"\
43         "paddb " #a ", " #b " \n\t"
44 #endif
45
46
47 //FIXME? |255-0| = 1 (shouldnt be a problem ...)
48 /**
49  * Check if the middle 8x8 Block in the given 8x16 block is flat
50  */
51 static inline int RENAME(isVertDC)(uint8_t src[], int stride){
52         int numEq= 0;
53 #ifndef HAVE_MMX
54         int y;
55 #endif
56         src+= stride*4; // src points to begin of the 8x8 Block
57 #ifdef HAVE_MMX
58 asm volatile(
59                 "leal (%1, %2), %%eax                           \n\t"
60                 "leal (%%eax, %2, 4), %%ebx                     \n\t"
61 //      0       1       2       3       4       5       6       7       8       9
62 //      %1      eax     eax+%2  eax+2%2 %1+4%2  ebx     ebx+%2  ebx+2%2 %1+8%2  ebx+4%2
63                 "movq "MANGLE(mmxDCOffset)", %%mm7              \n\t" // mm7 = 0x7F
64                 "movq "MANGLE(mmxDCThreshold)", %%mm6           \n\t" // mm6 = 0x7D
65                 "movq (%1), %%mm0                               \n\t"
66                 "movq (%%eax), %%mm1                            \n\t"
67                 "psubb %%mm1, %%mm0                             \n\t" // mm0 = differnece
68                 "paddb %%mm7, %%mm0                             \n\t"
69                 "pcmpgtb %%mm6, %%mm0                           \n\t"
70
71                 "movq (%%eax,%2), %%mm2                         \n\t"
72                 "psubb %%mm2, %%mm1                             \n\t"
73                 "paddb %%mm7, %%mm1                             \n\t"
74                 "pcmpgtb %%mm6, %%mm1                           \n\t"
75                 "paddb %%mm1, %%mm0                             \n\t"
76
77                 "movq (%%eax, %2, 2), %%mm1                     \n\t"
78                 "psubb %%mm1, %%mm2                             \n\t"
79                 "paddb %%mm7, %%mm2                             \n\t"
80                 "pcmpgtb %%mm6, %%mm2                           \n\t"
81                 "paddb %%mm2, %%mm0                             \n\t"
82
83                 "movq (%1, %2, 4), %%mm2                        \n\t"
84                 "psubb %%mm2, %%mm1                             \n\t"
85                 "paddb %%mm7, %%mm1                             \n\t"
86                 "pcmpgtb %%mm6, %%mm1                           \n\t"
87                 "paddb %%mm1, %%mm0                             \n\t"
88
89                 "movq (%%ebx), %%mm1                            \n\t"
90                 "psubb %%mm1, %%mm2                             \n\t"
91                 "paddb %%mm7, %%mm2                             \n\t"
92                 "pcmpgtb %%mm6, %%mm2                           \n\t"
93                 "paddb %%mm2, %%mm0                             \n\t"
94
95                 "movq (%%ebx, %2), %%mm2                        \n\t"
96                 "psubb %%mm2, %%mm1                             \n\t"
97                 "paddb %%mm7, %%mm1                             \n\t"
98                 "pcmpgtb %%mm6, %%mm1                           \n\t"
99                 "paddb %%mm1, %%mm0                             \n\t"
100
101                 "movq (%%ebx, %2, 2), %%mm1                     \n\t"
102                 "psubb %%mm1, %%mm2                             \n\t"
103                 "paddb %%mm7, %%mm2                             \n\t"
104                 "pcmpgtb %%mm6, %%mm2                           \n\t"
105                 "paddb %%mm2, %%mm0                             \n\t"
106
107                 "                                               \n\t"
108 #ifdef HAVE_MMX2
109                 "pxor %%mm7, %%mm7                              \n\t"
110                 "psadbw %%mm7, %%mm0                            \n\t"
111 #else
112                 "movq %%mm0, %%mm1                              \n\t"
113                 "psrlw $8, %%mm0                                \n\t"
114                 "paddb %%mm1, %%mm0                             \n\t"
115                 "movq %%mm0, %%mm1                              \n\t"
116                 "psrlq $16, %%mm0                               \n\t"
117                 "paddb %%mm1, %%mm0                             \n\t"
118                 "movq %%mm0, %%mm1                              \n\t"
119                 "psrlq $32, %%mm0                               \n\t"
120                 "paddb %%mm1, %%mm0                             \n\t"
121 #endif
122                 "movd %%mm0, %0                                 \n\t"
123                 : "=r" (numEq)
124                 : "r" (src), "r" (stride)
125                 : "%ebx"
126                 );
127         numEq= (-numEq) &0xFF;
128
129 #else
130         for(y=0; y<BLOCK_SIZE-1; y++)
131         {
132                 if(((src[0] - src[0+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
133                 if(((src[1] - src[1+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
134                 if(((src[2] - src[2+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
135                 if(((src[3] - src[3+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
136                 if(((src[4] - src[4+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
137                 if(((src[5] - src[5+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
138                 if(((src[6] - src[6+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
139                 if(((src[7] - src[7+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
140                 src+= stride;
141         }
142 #endif
143 /*      if(abs(numEq - asmEq) > 0)
144         {
145                 printf("\nasm:%d  c:%d\n", asmEq, numEq);
146                 for(int y=0; y<8; y++)
147                 {
148                         for(int x=0; x<8; x++)
149                         {
150                                 printf("%d ", temp[x + y*stride]);
151                         }
152                         printf("\n");
153                 }
154         }
155 */
156 //      for(int i=0; i<numEq/8; i++) src[i]=255;
157         return (numEq > vFlatnessThreshold) ? 1 : 0;
158 }
159
160 static inline int RENAME(isVertMinMaxOk)(uint8_t src[], int stride, int QP)
161 {
162 #ifdef HAVE_MMX
163         int isOk;
164         src+= stride*3;
165         asm volatile(
166 //              "int $3 \n\t"
167                 "movq (%1, %2), %%mm0                           \n\t"
168                 "movq (%1, %2, 8), %%mm1                        \n\t"
169                 "movq %%mm0, %%mm2                              \n\t"
170                 "psubusb %%mm1, %%mm0                           \n\t"
171                 "psubusb %%mm2, %%mm1                           \n\t"
172                 "por %%mm1, %%mm0                               \n\t" // ABS Diff
173
174                 "movq "MANGLE(pQPb)", %%mm7                     \n\t" // QP,..., QP
175                 "paddusb %%mm7, %%mm7                           \n\t" // 2QP ... 2QP
176                 "psubusb %%mm7, %%mm0                           \n\t" // Diff <= 2QP -> 0
177                 "pcmpeqd "MANGLE(b00)", %%mm0                   \n\t"
178                 "psrlq $16, %%mm0                               \n\t"
179                 "pcmpeqd "MANGLE(bFF)", %%mm0                   \n\t"
180 //              "movd %%mm0, (%1, %2, 4)\n\t"
181                 "movd %%mm0, %0                                 \n\t"
182                 : "=r" (isOk)
183                 : "r" (src), "r" (stride)
184                 );
185         return isOk;
186 #else
187
188         int isOk2= 1;
189         int x;
190         src+= stride*3;
191         for(x=0; x<BLOCK_SIZE; x++)
192         {
193                 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0;
194         }
195 /*      if(isOk && !isOk2 || !isOk && isOk2)
196         {
197                 printf("\nasm:%d  c:%d QP:%d\n", isOk, isOk2, QP);
198                 for(int y=0; y<9; y++)
199                 {
200                         for(int x=0; x<8; x++)
201                         {
202                                 printf("%d ", src[x + y*stride]);
203                         }
204                         printf("\n");
205                 }
206         } */
207
208         return isOk2;
209 #endif
210
211 }
212
213 /**
214  * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
215  * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
216  */
217 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, int QP)
218 {
219 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
220         src+= stride*3;
221         asm volatile(   //"movv %0 %1 %2\n\t"
222                 "movq "MANGLE(pQPb)", %%mm0                     \n\t"  // QP,..., QP
223
224                 "movq (%0), %%mm6                               \n\t"
225                 "movq (%0, %1), %%mm5                           \n\t"
226                 "movq %%mm5, %%mm1                              \n\t"
227                 "movq %%mm6, %%mm2                              \n\t"
228                 "psubusb %%mm6, %%mm5                           \n\t"
229                 "psubusb %%mm1, %%mm2                           \n\t"
230                 "por %%mm5, %%mm2                               \n\t" // ABS Diff of lines
231                 "psubusb %%mm0, %%mm2                           \n\t" // diff <= QP -> 0
232                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // diff <= QP -> FF
233
234                 "pand %%mm2, %%mm6                              \n\t"
235                 "pandn %%mm1, %%mm2                             \n\t"
236                 "por %%mm2, %%mm6                               \n\t"// First Line to Filter
237
238                 "movq (%0, %1, 8), %%mm5                        \n\t"
239                 "leal (%0, %1, 4), %%eax                        \n\t"
240                 "leal (%0, %1, 8), %%ebx                        \n\t"
241                 "subl %1, %%ebx                                 \n\t"
242                 "addl %1, %0                                    \n\t" // %0 points to line 1 not 0
243                 "movq (%0, %1, 8), %%mm7                        \n\t"
244                 "movq %%mm5, %%mm1                              \n\t"
245                 "movq %%mm7, %%mm2                              \n\t"
246                 "psubusb %%mm7, %%mm5                           \n\t"
247                 "psubusb %%mm1, %%mm2                           \n\t"
248                 "por %%mm5, %%mm2                               \n\t" // ABS Diff of lines
249                 "psubusb %%mm0, %%mm2                           \n\t" // diff <= QP -> 0
250                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // diff <= QP -> FF
251
252                 "pand %%mm2, %%mm7                              \n\t"
253                 "pandn %%mm1, %%mm2                             \n\t"
254                 "por %%mm2, %%mm7                               \n\t" // First Line to Filter
255
256
257                 //      1       2       3       4       5       6       7       8
258                 //      %0      %0+%1   %0+2%1  eax     %0+4%1  eax+2%1 ebx     eax+4%1
259                 // 6 4 2 2 1 1
260                 // 6 4 4 2
261                 // 6 8 2
262
263                 "movq (%0, %1), %%mm0                           \n\t" //  1
264                 "movq %%mm0, %%mm1                              \n\t" //  1
265                 PAVGB(%%mm6, %%mm0)                                   //1 1     /2
266                 PAVGB(%%mm6, %%mm0)                                   //3 1     /4
267
268                 "movq (%0, %1, 4), %%mm2                        \n\t" //     1
269                 "movq %%mm2, %%mm5                              \n\t" //     1
270                 PAVGB((%%eax), %%mm2)                                 //    11  /2
271                 PAVGB((%0, %1, 2), %%mm2)                             //   211  /4
272                 "movq %%mm2, %%mm3                              \n\t" //   211  /4
273                 "movq (%0), %%mm4                               \n\t" // 1
274                 PAVGB(%%mm4, %%mm3)                                   // 4 211  /8
275                 PAVGB(%%mm0, %%mm3)                                   //642211  /16
276                 "movq %%mm3, (%0)                               \n\t" // X
277                 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
278                 "movq %%mm1, %%mm0                              \n\t" //  1
279                 PAVGB(%%mm6, %%mm0)                                   //1 1     /2
280                 "movq %%mm4, %%mm3                              \n\t" // 1
281                 PAVGB((%0,%1,2), %%mm3)                               // 1 1    /2
282                 PAVGB((%%eax,%1,2), %%mm5)                            //     11 /2
283                 PAVGB((%%eax), %%mm5)                                 //    211 /4
284                 PAVGB(%%mm5, %%mm3)                                   // 2 2211 /8
285                 PAVGB(%%mm0, %%mm3)                                   //4242211 /16
286                 "movq %%mm3, (%0,%1)                            \n\t" //  X
287                 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
288                 PAVGB(%%mm4, %%mm6)                                   //11      /2
289                 "movq (%%ebx), %%mm0                            \n\t" //       1
290                 PAVGB((%%eax, %1, 2), %%mm0)                          //      11/2
291                 "movq %%mm0, %%mm3                              \n\t" //      11/2
292                 PAVGB(%%mm1, %%mm0)                                   //  2   11/4
293                 PAVGB(%%mm6, %%mm0)                                   //222   11/8
294                 PAVGB(%%mm2, %%mm0)                                   //22242211/16
295                 "movq (%0, %1, 2), %%mm2                        \n\t" //   1
296                 "movq %%mm0, (%0, %1, 2)                        \n\t" //   X
297                 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
298                 "movq (%%eax, %1, 4), %%mm0                     \n\t" //        1
299                 PAVGB((%%ebx), %%mm0)                                 //       11       /2
300                 PAVGB(%%mm0, %%mm6)                                   //11     11       /4
301                 PAVGB(%%mm1, %%mm4)                                   // 11             /2
302                 PAVGB(%%mm2, %%mm1)                                   //  11            /2
303                 PAVGB(%%mm1, %%mm6)                                   //1122   11       /8
304                 PAVGB(%%mm5, %%mm6)                                   //112242211       /16
305                 "movq (%%eax), %%mm5                            \n\t" //    1
306                 "movq %%mm6, (%%eax)                            \n\t" //    X
307                 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
308                 "movq (%%eax, %1, 4), %%mm6                     \n\t" //        1
309                 PAVGB(%%mm7, %%mm6)                                   //        11      /2
310                 PAVGB(%%mm4, %%mm6)                                   // 11     11      /4
311                 PAVGB(%%mm3, %%mm6)                                   // 11   2211      /8
312                 PAVGB(%%mm5, %%mm2)                                   //   11           /2
313                 "movq (%0, %1, 4), %%mm4                        \n\t" //     1
314                 PAVGB(%%mm4, %%mm2)                                   //   112          /4
315                 PAVGB(%%mm2, %%mm6)                                   // 112242211      /16
316                 "movq %%mm6, (%0, %1, 4)                        \n\t" //     X
317                 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
318                 PAVGB(%%mm7, %%mm1)                                   //  11     2      /4
319                 PAVGB(%%mm4, %%mm5)                                   //    11          /2
320                 PAVGB(%%mm5, %%mm0)                                   //    11 11       /4
321                 "movq (%%eax, %1, 2), %%mm6                     \n\t" //      1
322                 PAVGB(%%mm6, %%mm1)                                   //  11  4  2      /8
323                 PAVGB(%%mm0, %%mm1)                                   //  11224222      /16
324                 "movq %%mm1, (%%eax, %1, 2)                     \n\t" //      X
325                 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
326                 PAVGB((%%ebx), %%mm2)                                 //   112 4        /8
327                 "movq (%%eax, %1, 4), %%mm0                     \n\t" //        1
328                 PAVGB(%%mm0, %%mm6)                                   //      1 1       /2
329                 PAVGB(%%mm7, %%mm6)                                   //      1 12      /4
330                 PAVGB(%%mm2, %%mm6)                                   //   1122424      /4
331                 "movq %%mm6, (%%ebx)                            \n\t" //       X
332                 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
333                 PAVGB(%%mm7, %%mm5)                                   //    11   2      /4
334                 PAVGB(%%mm7, %%mm5)                                   //    11   6      /8
335
336                 PAVGB(%%mm3, %%mm0)                                   //      112       /4
337                 PAVGB(%%mm0, %%mm5)                                   //    112246      /16
338                 "movq %%mm5, (%%eax, %1, 4)                     \n\t" //        X
339                 "subl %1, %0                                    \n\t"
340
341                 :
342                 : "r" (src), "r" (stride)
343                 : "%eax", "%ebx"
344         );
345 #else
346         const int l1= stride;
347         const int l2= stride + l1;
348         const int l3= stride + l2;
349         const int l4= stride + l3;
350         const int l5= stride + l4;
351         const int l6= stride + l5;
352         const int l7= stride + l6;
353         const int l8= stride + l7;
354         const int l9= stride + l8;
355         int x;
356         src+= stride*3;
357         for(x=0; x<BLOCK_SIZE; x++)
358         {
359                 const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1];
360                 const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8];
361
362                 int sums[9];
363                 sums[0] = first + src[l1];
364                 sums[1] = src[l1] + src[l2];
365                 sums[2] = src[l2] + src[l3];
366                 sums[3] = src[l3] + src[l4];
367                 sums[4] = src[l4] + src[l5];
368                 sums[5] = src[l5] + src[l6];
369                 sums[6] = src[l6] + src[l7];
370                 sums[7] = src[l7] + src[l8];
371                 sums[8] = src[l8] + last;
372
373                 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
374                 src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
375                 src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
376                 src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
377                 src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
378                 src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
379                 src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4;
380                 src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
381
382                 src++;
383         }
384
385 #endif
386 }
387
388 /**
389  * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
390  * values are correctly clipped (MMX2)
391  * values are wraparound (C)
392  * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient
393         0 8 16 24
394         x = 8
395         x/2 = 4
396         x/8 = 1
397         1 12 12 23
398  */
399 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
400 {
401 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
402         src+= stride*3;
403 // FIXME rounding
404         asm volatile(
405                 "pxor %%mm7, %%mm7                              \n\t" // 0
406                 "movq "MANGLE(b80)", %%mm6                      \n\t" // MIN_SIGNED_BYTE
407                 "leal (%0, %1), %%eax                           \n\t"
408                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
409 //      0       1       2       3       4       5       6       7       8       9
410 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
411                 "movq "MANGLE(pQPb)", %%mm0                     \n\t" // QP,..., QP
412                 "movq %%mm0, %%mm1                              \n\t" // QP,..., QP
413                 "paddusb "MANGLE(b02)", %%mm0                   \n\t"
414                 "psrlw $2, %%mm0                                \n\t"
415                 "pand "MANGLE(b3F)", %%mm0                      \n\t" // QP/4,..., QP/4
416                 "paddusb %%mm1, %%mm0                           \n\t" // QP*1.25 ...
417                 "movq (%0, %1, 4), %%mm2                        \n\t" // line 4
418                 "movq (%%ebx), %%mm3                            \n\t" // line 5
419                 "movq %%mm2, %%mm4                              \n\t" // line 4
420                 "pcmpeqb %%mm5, %%mm5                           \n\t" // -1
421                 "pxor %%mm2, %%mm5                              \n\t" // -line 4 - 1
422                 PAVGB(%%mm3, %%mm5)
423                 "paddb %%mm6, %%mm5                             \n\t" // (l5-l4)/2
424                 "psubusb %%mm3, %%mm4                           \n\t"
425                 "psubusb %%mm2, %%mm3                           \n\t"
426                 "por %%mm3, %%mm4                               \n\t" // |l4 - l5|
427                 "psubusb %%mm0, %%mm4                           \n\t"
428                 "pcmpeqb %%mm7, %%mm4                           \n\t"
429                 "pand %%mm4, %%mm5                              \n\t" // d/2
430
431 //              "paddb %%mm6, %%mm2                             \n\t" // line 4 + 0x80
432                 "paddb %%mm5, %%mm2                             \n\t"
433 //              "psubb %%mm6, %%mm2                             \n\t"
434                 "movq %%mm2, (%0,%1, 4)                         \n\t"
435
436                 "movq (%%ebx), %%mm2                            \n\t"
437 //              "paddb %%mm6, %%mm2                             \n\t" // line 5 + 0x80
438                 "psubb %%mm5, %%mm2                             \n\t"
439 //              "psubb %%mm6, %%mm2                             \n\t"
440                 "movq %%mm2, (%%ebx)                            \n\t"
441
442                 "paddb %%mm6, %%mm5                             \n\t"
443                 "psrlw $2, %%mm5                                \n\t"
444                 "pand "MANGLE(b3F)", %%mm5                      \n\t"
445                 "psubb "MANGLE(b20)", %%mm5                     \n\t" // (l5-l4)/8
446
447                 "movq (%%eax, %1, 2), %%mm2                     \n\t"
448                 "paddb %%mm6, %%mm2                             \n\t" // line 3 + 0x80
449                 "paddsb %%mm5, %%mm2                            \n\t"
450                 "psubb %%mm6, %%mm2                             \n\t"
451                 "movq %%mm2, (%%eax, %1, 2)                     \n\t"
452
453                 "movq (%%ebx, %1), %%mm2                        \n\t"
454                 "paddb %%mm6, %%mm2                             \n\t" // line 6 + 0x80
455                 "psubsb %%mm5, %%mm2                            \n\t"
456                 "psubb %%mm6, %%mm2                             \n\t"
457                 "movq %%mm2, (%%ebx, %1)                        \n\t"
458
459                 :
460                 : "r" (src), "r" (stride)
461                 : "%eax", "%ebx"
462         );
463 #else
464         const int l1= stride;
465         const int l2= stride + l1;
466         const int l3= stride + l2;
467         const int l4= stride + l3;
468         const int l5= stride + l4;
469         const int l6= stride + l5;
470 //      const int l7= stride + l6;
471 //      const int l8= stride + l7;
472 //      const int l9= stride + l8;
473         int x;
474         const int QP15= QP + (QP>>2);
475         src+= stride*3;
476         for(x=0; x<BLOCK_SIZE; x++)
477         {
478                 const int v = (src[x+l5] - src[x+l4]);
479                 if(ABS(v) < QP15)
480                 {
481                         src[x+l3] +=v>>3;
482                         src[x+l4] +=v>>1;
483                         src[x+l5] -=v>>1;
484                         src[x+l6] -=v>>3;
485
486                 }
487         }
488
489 #endif
490 }
491
492 /**
493  * Experimental Filter 1
494  * will not damage linear gradients
495  * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
496  * can only smooth blocks at the expected locations (it cant smooth them if they did move)
497  * MMX2 version does correct clipping C version doesnt
498  */
499 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, int QP)
500 {
501 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
502         src+= stride*3;
503
504         asm volatile(
505                 "pxor %%mm7, %%mm7                              \n\t" // 0
506 //              "movq "MANGLE(b80)", %%mm6                      \n\t" // MIN_SIGNED_BYTE
507                 "leal (%0, %1), %%eax                           \n\t"
508                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
509 //      0       1       2       3       4       5       6       7       8       9
510 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
511                 "movq (%%eax, %1, 2), %%mm0                     \n\t" // line 3
512                 "movq (%0, %1, 4), %%mm1                        \n\t" // line 4
513                 "movq %%mm1, %%mm2                              \n\t" // line 4
514                 "psubusb %%mm0, %%mm1                           \n\t"
515                 "psubusb %%mm2, %%mm0                           \n\t"
516                 "por %%mm1, %%mm0                               \n\t" // |l2 - l3|
517                 "movq (%%ebx), %%mm3                            \n\t" // line 5
518                 "movq (%%ebx, %1), %%mm4                                \n\t" // line 6
519                 "movq %%mm3, %%mm5                              \n\t" // line 5
520                 "psubusb %%mm4, %%mm3                           \n\t"
521                 "psubusb %%mm5, %%mm4                           \n\t"
522                 "por %%mm4, %%mm3                               \n\t" // |l5 - l6|
523                 PAVGB(%%mm3, %%mm0)                                   // (|l2 - l3| + |l5 - l6|)/2
524                 "movq %%mm2, %%mm1                              \n\t" // line 4
525                 "psubusb %%mm5, %%mm2                           \n\t"
526                 "movq %%mm2, %%mm4                              \n\t"
527                 "pcmpeqb %%mm7, %%mm2                           \n\t" // (l4 - l5) <= 0 ? -1 : 0
528                 "psubusb %%mm1, %%mm5                           \n\t"
529                 "por %%mm5, %%mm4                               \n\t" // |l4 - l5|
530                 "psubusb %%mm0, %%mm4           \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
531                 "movq %%mm4, %%mm3                              \n\t" // d
532                 "psubusb "MANGLE(pQPb)", %%mm4                  \n\t"
533                 "pcmpeqb %%mm7, %%mm4                           \n\t" // d <= QP ? -1 : 0
534                 "psubusb "MANGLE(b01)", %%mm3                   \n\t"
535                 "pand %%mm4, %%mm3                              \n\t" // d <= QP ? d : 0
536
537                 PAVGB(%%mm7, %%mm3)                                   // d/2
538                 "movq %%mm3, %%mm1                              \n\t" // d/2
539                 PAVGB(%%mm7, %%mm3)                                   // d/4
540                 PAVGB(%%mm1, %%mm3)                                   // 3*d/8
541
542                 "movq (%0, %1, 4), %%mm0                        \n\t" // line 4
543                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
544                 "psubusb %%mm3, %%mm0                           \n\t"
545                 "pxor %%mm2, %%mm0                              \n\t"
546                 "movq %%mm0, (%0, %1, 4)                        \n\t" // line 4
547
548                 "movq (%%ebx), %%mm0                            \n\t" // line 5
549                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
550                 "paddusb %%mm3, %%mm0                           \n\t"
551                 "pxor %%mm2, %%mm0                              \n\t"
552                 "movq %%mm0, (%%ebx)                            \n\t" // line 5
553
554                 PAVGB(%%mm7, %%mm1)                                   // d/4
555
556                 "movq (%%eax, %1, 2), %%mm0                     \n\t" // line 3
557                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
558                 "psubusb %%mm1, %%mm0                           \n\t"
559                 "pxor %%mm2, %%mm0                              \n\t"
560                 "movq %%mm0, (%%eax, %1, 2)                     \n\t" // line 3
561
562                 "movq (%%ebx, %1), %%mm0                        \n\t" // line 6
563                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
564                 "paddusb %%mm1, %%mm0                           \n\t"
565                 "pxor %%mm2, %%mm0                              \n\t"
566                 "movq %%mm0, (%%ebx, %1)                        \n\t" // line 6
567
568                 PAVGB(%%mm7, %%mm1)                                   // d/8
569
570                 "movq (%%eax, %1), %%mm0                        \n\t" // line 2
571                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
572                 "psubusb %%mm1, %%mm0                           \n\t"
573                 "pxor %%mm2, %%mm0                              \n\t"
574                 "movq %%mm0, (%%eax, %1)                        \n\t" // line 2
575
576                 "movq (%%ebx, %1, 2), %%mm0                     \n\t" // line 7
577                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
578                 "paddusb %%mm1, %%mm0                           \n\t"
579                 "pxor %%mm2, %%mm0                              \n\t"
580                 "movq %%mm0, (%%ebx, %1, 2)                     \n\t" // line 7
581
582                 :
583                 : "r" (src), "r" (stride)
584                 : "%eax", "%ebx"
585         );
586 #else
587
588         const int l1= stride;
589         const int l2= stride + l1;
590         const int l3= stride + l2;
591         const int l4= stride + l3;
592         const int l5= stride + l4;
593         const int l6= stride + l5;
594         const int l7= stride + l6;
595 //      const int l8= stride + l7;
596 //      const int l9= stride + l8;
597         int x;
598
599         src+= stride*3;
600         for(x=0; x<BLOCK_SIZE; x++)
601         {
602                 int a= src[l3] - src[l4];
603                 int b= src[l4] - src[l5];
604                 int c= src[l5] - src[l6];
605
606                 int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
607                 d= MAX(d, 0);
608
609                 if(d < QP)
610                 {
611                         int v = d * SIGN(-b);
612
613                         src[l2] +=v>>3;
614                         src[l3] +=v>>2;
615                         src[l4] +=(3*v)>>3;
616                         src[l5] -=(3*v)>>3;
617                         src[l6] -=v>>2;
618                         src[l7] -=v>>3;
619
620                 }
621                 src++;
622         }
623         /*
624         const int l1= stride;
625         const int l2= stride + l1;
626         const int l3= stride + l2;
627         const int l4= stride + l3;
628         const int l5= stride + l4;
629         const int l6= stride + l5;
630         const int l7= stride + l6;
631         const int l8= stride + l7;
632         const int l9= stride + l8;
633         for(int x=0; x<BLOCK_SIZE; x++)
634         {
635                 int v2= src[l2];
636                 int v3= src[l3];
637                 int v4= src[l4];
638                 int v5= src[l5];
639                 int v6= src[l6];
640                 int v7= src[l7];
641
642                 if(ABS(v4-v5)<QP &&  ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 )
643                 {
644                         src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6         )/16;
645                         src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7  )/16;
646                         src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16;
647                         src[l6] = (       1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16;
648                 }
649                 src++;
650         }
651 */
652 #endif
653 }
654
655 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, int QP)
656 {
657 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
658 /*
659         uint8_t tmp[16];
660         const int l1= stride;
661         const int l2= stride + l1;
662         const int l3= stride + l2;
663         const int l4= (int)tmp - (int)src - stride*3;
664         const int l5= (int)tmp - (int)src - stride*3 + 8;
665         const int l6= stride*3 + l3;
666         const int l7= stride + l6;
667         const int l8= stride + l7;
668
669         memcpy(tmp, src+stride*7, 8);
670         memcpy(tmp+8, src+stride*8, 8);
671 */
672         src+= stride*4;
673         asm volatile(
674
675 #if 0 //sligtly more accurate and slightly slower
676                 "pxor %%mm7, %%mm7                              \n\t" // 0
677                 "leal (%0, %1), %%eax                           \n\t"
678                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
679 //      0       1       2       3       4       5       6       7
680 //      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ebx+%1  ebx+2%1
681 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1
682
683
684                 "movq (%0, %1, 2), %%mm0                        \n\t" // l2
685                 "movq (%0), %%mm1                               \n\t" // l0
686                 "movq %%mm0, %%mm2                              \n\t" // l2
687                 PAVGB(%%mm7, %%mm0)                                   // ~l2/2
688                 PAVGB(%%mm1, %%mm0)                                   // ~(l2 + 2l0)/4
689                 PAVGB(%%mm2, %%mm0)                                   // ~(5l2 + 2l0)/8
690
691                 "movq (%%eax), %%mm1                            \n\t" // l1
692                 "movq (%%eax, %1, 2), %%mm3                     \n\t" // l3
693                 "movq %%mm1, %%mm4                              \n\t" // l1
694                 PAVGB(%%mm7, %%mm1)                                   // ~l1/2
695                 PAVGB(%%mm3, %%mm1)                                   // ~(l1 + 2l3)/4
696                 PAVGB(%%mm4, %%mm1)                                   // ~(5l1 + 2l3)/8
697
698                 "movq %%mm0, %%mm4                              \n\t" // ~(5l2 + 2l0)/8
699                 "psubusb %%mm1, %%mm0                           \n\t"
700                 "psubusb %%mm4, %%mm1                           \n\t"
701                 "por %%mm0, %%mm1                               \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
702 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
703
704                 "movq (%0, %1, 4), %%mm0                        \n\t" // l4
705                 "movq %%mm0, %%mm4                              \n\t" // l4
706                 PAVGB(%%mm7, %%mm0)                                   // ~l4/2
707                 PAVGB(%%mm2, %%mm0)                                   // ~(l4 + 2l2)/4
708                 PAVGB(%%mm4, %%mm0)                                   // ~(5l4 + 2l2)/8
709
710                 "movq (%%ebx), %%mm2                            \n\t" // l5
711                 "movq %%mm3, %%mm5                              \n\t" // l3
712                 PAVGB(%%mm7, %%mm3)                                   // ~l3/2
713                 PAVGB(%%mm2, %%mm3)                                   // ~(l3 + 2l5)/4
714                 PAVGB(%%mm5, %%mm3)                                   // ~(5l3 + 2l5)/8
715
716                 "movq %%mm0, %%mm6                              \n\t" // ~(5l4 + 2l2)/8
717                 "psubusb %%mm3, %%mm0                           \n\t"
718                 "psubusb %%mm6, %%mm3                           \n\t"
719                 "por %%mm0, %%mm3                               \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
720                 "pcmpeqb %%mm7, %%mm0                           \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
721 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
722
723                 "movq (%%ebx, %1), %%mm6                        \n\t" // l6
724                 "movq %%mm6, %%mm5                              \n\t" // l6
725                 PAVGB(%%mm7, %%mm6)                                   // ~l6/2
726                 PAVGB(%%mm4, %%mm6)                                   // ~(l6 + 2l4)/4
727                 PAVGB(%%mm5, %%mm6)                                   // ~(5l6 + 2l4)/8
728
729                 "movq (%%ebx, %1, 2), %%mm5                     \n\t" // l7
730                 "movq %%mm2, %%mm4                              \n\t" // l5
731                 PAVGB(%%mm7, %%mm2)                                   // ~l5/2
732                 PAVGB(%%mm5, %%mm2)                                   // ~(l5 + 2l7)/4
733                 PAVGB(%%mm4, %%mm2)                                   // ~(5l5 + 2l7)/8
734
735                 "movq %%mm6, %%mm4                              \n\t" // ~(5l6 + 2l4)/8
736                 "psubusb %%mm2, %%mm6                           \n\t"
737                 "psubusb %%mm4, %%mm2                           \n\t"
738                 "por %%mm6, %%mm2                               \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
739 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
740
741
742                 PMINUB(%%mm2, %%mm1, %%mm4)                           // MIN(|lenergy|,|renergy|)/8
743                 "movq "MANGLE(pQPb)", %%mm4                     \n\t" // QP //FIXME QP+1 ?
744                 "paddusb "MANGLE(b01)", %%mm4                   \n\t"
745                 "pcmpgtb %%mm3, %%mm4                           \n\t" // |menergy|/8 < QP
746                 "psubusb %%mm1, %%mm3                           \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
747                 "pand %%mm4, %%mm3                              \n\t"
748
749                 "movq %%mm3, %%mm1                              \n\t"
750 //              "psubusb "MANGLE(b01)", %%mm3                   \n\t"
751                 PAVGB(%%mm7, %%mm3)
752                 PAVGB(%%mm7, %%mm3)
753                 "paddusb %%mm1, %%mm3                           \n\t"
754 //              "paddusb "MANGLE(b01)", %%mm3                   \n\t"
755
756                 "movq (%%eax, %1, 2), %%mm6                     \n\t" //l3
757                 "movq (%0, %1, 4), %%mm5                        \n\t" //l4
758                 "movq (%0, %1, 4), %%mm4                        \n\t" //l4
759                 "psubusb %%mm6, %%mm5                           \n\t"
760                 "psubusb %%mm4, %%mm6                           \n\t"
761                 "por %%mm6, %%mm5                               \n\t" // |l3-l4|
762                 "pcmpeqb %%mm7, %%mm6                           \n\t" // SIGN(l3-l4)
763                 "pxor %%mm6, %%mm0                              \n\t"
764                 "pand %%mm0, %%mm3                              \n\t"
765                 PMINUB(%%mm5, %%mm3, %%mm0)
766
767                 "psubusb "MANGLE(b01)", %%mm3                   \n\t"
768                 PAVGB(%%mm7, %%mm3)
769
770                 "movq (%%eax, %1, 2), %%mm0                     \n\t"
771                 "movq (%0, %1, 4), %%mm2                        \n\t"
772                 "pxor %%mm6, %%mm0                              \n\t"
773                 "pxor %%mm6, %%mm2                              \n\t"
774                 "psubb %%mm3, %%mm0                             \n\t"
775                 "paddb %%mm3, %%mm2                             \n\t"
776                 "pxor %%mm6, %%mm0                              \n\t"
777                 "pxor %%mm6, %%mm2                              \n\t"
778                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
779                 "movq %%mm2, (%0, %1, 4)                        \n\t"
780 #endif
781
782                 "leal (%0, %1), %%eax                           \n\t"
783                 "pcmpeqb %%mm6, %%mm6                           \n\t" // -1
784 //      0       1       2       3       4       5       6       7
785 //      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ebx+%1  ebx+2%1
786 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1
787
788
789                 "movq (%%eax, %1, 2), %%mm1                     \n\t" // l3
790                 "movq (%0, %1, 4), %%mm0                        \n\t" // l4
791                 "pxor %%mm6, %%mm1                              \n\t" // -l3-1
792                 PAVGB(%%mm1, %%mm0)                                   // -q+128 = (l4-l3+256)/2
793 // mm1=-l3-1, mm0=128-q
794
795                 "movq (%%eax, %1, 4), %%mm2                     \n\t" // l5
796                 "movq (%%eax, %1), %%mm3                        \n\t" // l2
797                 "pxor %%mm6, %%mm2                              \n\t" // -l5-1
798                 "movq %%mm2, %%mm5                              \n\t" // -l5-1
799                 "movq "MANGLE(b80)", %%mm4                      \n\t" // 128
800                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
801                 PAVGB(%%mm3, %%mm2)                                   // (l2-l5+256)/2
802                 PAVGB(%%mm0, %%mm4)                                   // ~(l4-l3)/4 + 128
803                 PAVGB(%%mm2, %%mm4)                                   // ~(l2-l5)/4 +(l4-l3)/8 + 128
804                 PAVGB(%%mm0, %%mm4)                                   // ~(l2-l5)/8 +5(l4-l3)/16 + 128
805 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
806
807                 "movq (%%eax), %%mm2                            \n\t" // l1
808                 "pxor %%mm6, %%mm2                              \n\t" // -l1-1
809                 PAVGB(%%mm3, %%mm2)                                   // (l2-l1+256)/2
810                 PAVGB((%0), %%mm1)                                    // (l0-l3+256)/2
811                 "movq "MANGLE(b80)", %%mm3                      \n\t" // 128
812                 PAVGB(%%mm2, %%mm3)                                   // ~(l2-l1)/4 + 128
813                 PAVGB(%%mm1, %%mm3)                                   // ~(l0-l3)/4 +(l2-l1)/8 + 128
814                 PAVGB(%%mm2, %%mm3)                                   // ~(l0-l3)/8 +5(l2-l1)/16 + 128
815 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
816
817                 PAVGB((%%ebx, %1), %%mm5)                             // (l6-l5+256)/2
818                 "movq (%%ebx, %1, 2), %%mm1                     \n\t" // l7
819                 "pxor %%mm6, %%mm1                              \n\t" // -l7-1
820                 PAVGB((%0, %1, 4), %%mm1)                             // (l4-l7+256)/2
821                 "movq "MANGLE(b80)", %%mm2                      \n\t" // 128
822                 PAVGB(%%mm5, %%mm2)                                   // ~(l6-l5)/4 + 128
823                 PAVGB(%%mm1, %%mm2)                                   // ~(l4-l7)/4 +(l6-l5)/8 + 128
824                 PAVGB(%%mm5, %%mm2)                                   // ~(l4-l7)/8 +5(l6-l5)/16 + 128
825 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
826
827                 "movq "MANGLE(b00)", %%mm1                      \n\t" // 0
828                 "movq "MANGLE(b00)", %%mm5                      \n\t" // 0
829                 "psubb %%mm2, %%mm1                             \n\t" // 128 - renergy/16
830                 "psubb %%mm3, %%mm5                             \n\t" // 128 - lenergy/16
831                 PMAXUB(%%mm1, %%mm2)                                  // 128 + |renergy/16|
832                 PMAXUB(%%mm5, %%mm3)                                  // 128 + |lenergy/16|
833                 PMINUB(%%mm2, %%mm3, %%mm1)                           // 128 + MIN(|lenergy|,|renergy|)/16
834
835 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
836
837                 "movq "MANGLE(b00)", %%mm7                      \n\t" // 0
838                 "movq "MANGLE(pQPb)", %%mm2                     \n\t" // QP
839                 PAVGB(%%mm6, %%mm2)                                   // 128 + QP/2
840                 "psubb %%mm6, %%mm2                             \n\t"
841
842                 "movq %%mm4, %%mm1                              \n\t"
843                 "pcmpgtb %%mm7, %%mm1                           \n\t" // SIGN(menergy)
844                 "pxor %%mm1, %%mm4                              \n\t"
845                 "psubb %%mm1, %%mm4                             \n\t" // 128 + |menergy|/16
846                 "pcmpgtb %%mm4, %%mm2                           \n\t" // |menergy|/16 < QP/2
847                 "psubusb %%mm3, %%mm4                           \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
848 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
849
850                 "movq %%mm4, %%mm3                              \n\t" // d
851                 "psubusb "MANGLE(b01)", %%mm4                   \n\t"
852                 PAVGB(%%mm7, %%mm4)                                   // d/32
853                 PAVGB(%%mm7, %%mm4)                                   // (d + 32)/64
854                 "paddb %%mm3, %%mm4                             \n\t" // 5d/64
855                 "pand %%mm2, %%mm4                              \n\t"
856
857                 "movq "MANGLE(b80)", %%mm5                      \n\t" // 128
858                 "psubb %%mm0, %%mm5                             \n\t" // q
859                 "paddsb %%mm6, %%mm5                            \n\t" // fix bad rounding
860                 "pcmpgtb %%mm5, %%mm7                           \n\t" // SIGN(q)
861                 "pxor %%mm7, %%mm5                              \n\t"
862
863                 PMINUB(%%mm5, %%mm4, %%mm3)                           // MIN(|q|, 5d/64)
864                 "pxor %%mm1, %%mm7                              \n\t" // SIGN(d*q)
865
866                 "pand %%mm7, %%mm4                              \n\t"
867                 "movq (%%eax, %1, 2), %%mm0                     \n\t"
868                 "movq (%0, %1, 4), %%mm2                        \n\t"
869                 "pxor %%mm1, %%mm0                              \n\t"
870                 "pxor %%mm1, %%mm2                              \n\t"
871                 "paddb %%mm4, %%mm0                             \n\t"
872                 "psubb %%mm4, %%mm2                             \n\t"
873                 "pxor %%mm1, %%mm0                              \n\t"
874                 "pxor %%mm1, %%mm2                              \n\t"
875                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
876                 "movq %%mm2, (%0, %1, 4)                        \n\t"
877
878                 :
879                 : "r" (src), "r" (stride)
880                 : "%eax", "%ebx"
881         );
882
883 /*
884         {
885         int x;
886         src-= stride;
887         for(x=0; x<BLOCK_SIZE; x++)
888         {
889                 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
890                 if(ABS(middleEnergy)< 8*QP)
891                 {
892                         const int q=(src[l4] - src[l5])/2;
893                         const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
894                         const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
895
896                         int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
897                         d= MAX(d, 0);
898
899                         d= (5*d + 32) >> 6;
900                         d*= SIGN(-middleEnergy);
901
902                         if(q>0)
903                         {
904                                 d= d<0 ? 0 : d;
905                                 d= d>q ? q : d;
906                         }
907                         else
908                         {
909                                 d= d>0 ? 0 : d;
910                                 d= d<q ? q : d;
911                         }
912
913                         src[l4]-= d;
914                         src[l5]+= d;
915                 }
916                 src++;
917         }
918 src-=8;
919         for(x=0; x<8; x++)
920         {
921                 int y;
922                 for(y=4; y<6; y++)
923                 {
924                         int d= src[x+y*stride] - tmp[x+(y-4)*8];
925                         int ad= ABS(d);
926                         static int max=0;
927                         static int sum=0;
928                         static int num=0;
929                         static int bias=0;
930
931                         if(max<ad) max=ad;
932                         sum+= ad>3 ? 1 : 0;
933                         if(ad>3)
934                         {
935                                 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
936                         }
937                         if(y==4) bias+=d;
938                         num++;
939                         if(num%1000000 == 0)
940                         {
941                                 printf(" %d %d %d %d\n", num, sum, max, bias);
942                         }
943                 }
944         }
945 }
946 */
947 #elif defined (HAVE_MMX)
948         src+= stride*4;
949
950         asm volatile(
951                 "pxor %%mm7, %%mm7                              \n\t"
952                 "leal (%0, %1), %%eax                           \n\t"
953                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
954 //      0       1       2       3       4       5       6       7
955 //      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ebx+%1  ebx+2%1
956 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1
957
958                 "movq (%0), %%mm0                               \n\t"
959                 "movq %%mm0, %%mm1                              \n\t"
960                 "punpcklbw %%mm7, %%mm0                         \n\t" // low part of line 0
961                 "punpckhbw %%mm7, %%mm1                         \n\t" // high part of line 0
962
963                 "movq (%%eax), %%mm2                            \n\t"
964                 "movq %%mm2, %%mm3                              \n\t"
965                 "punpcklbw %%mm7, %%mm2                         \n\t" // low part of line 1
966                 "punpckhbw %%mm7, %%mm3                         \n\t" // high part of line 1
967
968                 "movq (%%eax, %1), %%mm4                        \n\t"
969                 "movq %%mm4, %%mm5                              \n\t"
970                 "punpcklbw %%mm7, %%mm4                         \n\t" // low part of line 2
971                 "punpckhbw %%mm7, %%mm5                         \n\t" // high part of line 2
972
973                 "paddw %%mm0, %%mm0                             \n\t" // 2L0
974                 "paddw %%mm1, %%mm1                             \n\t" // 2H0
975                 "psubw %%mm4, %%mm2                             \n\t" // L1 - L2
976                 "psubw %%mm5, %%mm3                             \n\t" // H1 - H2
977                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - L1 + L2
978                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - H1 + H2
979
980                 "psllw $2, %%mm2                                \n\t" // 4L1 - 4L2
981                 "psllw $2, %%mm3                                \n\t" // 4H1 - 4H2
982                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2
983                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2
984
985                 "movq (%%eax, %1, 2), %%mm2                     \n\t"
986                 "movq %%mm2, %%mm3                              \n\t"
987                 "punpcklbw %%mm7, %%mm2                         \n\t" // L3
988                 "punpckhbw %%mm7, %%mm3                         \n\t" // H3
989
990                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - L3
991                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - H3
992                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - 2L3
993                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - 2H3
994                 "movq %%mm0, "MANGLE(temp0)"                    \n\t" // 2L0 - 5L1 + 5L2 - 2L3
995                 "movq %%mm1, "MANGLE(temp1)"                    \n\t" // 2H0 - 5H1 + 5H2 - 2H3
996
997                 "movq (%0, %1, 4), %%mm0                        \n\t"
998                 "movq %%mm0, %%mm1                              \n\t"
999                 "punpcklbw %%mm7, %%mm0                         \n\t" // L4
1000                 "punpckhbw %%mm7, %%mm1                         \n\t" // H4
1001
1002                 "psubw %%mm0, %%mm2                             \n\t" // L3 - L4
1003                 "psubw %%mm1, %%mm3                             \n\t" // H3 - H4
1004                 "movq %%mm2, "MANGLE(temp2)"                    \n\t" // L3 - L4
1005                 "movq %%mm3, "MANGLE(temp3)"                    \n\t" // H3 - H4
1006                 "paddw %%mm4, %%mm4                             \n\t" // 2L2
1007                 "paddw %%mm5, %%mm5                             \n\t" // 2H2
1008                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - L3 + L4
1009                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - H3 + H4
1010
1011                 "psllw $2, %%mm2                                \n\t" // 4L3 - 4L4
1012                 "psllw $2, %%mm3                                \n\t" // 4H3 - 4H4
1013                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4
1014                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4
1015 //50 opcodes so far
1016                 "movq (%%ebx), %%mm2                            \n\t"
1017                 "movq %%mm2, %%mm3                              \n\t"
1018                 "punpcklbw %%mm7, %%mm2                         \n\t" // L5
1019                 "punpckhbw %%mm7, %%mm3                         \n\t" // H5
1020                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - L5
1021                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - H5
1022                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - 2L5
1023                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - 2H5
1024
1025                 "movq (%%ebx, %1), %%mm6                        \n\t"
1026                 "punpcklbw %%mm7, %%mm6                         \n\t" // L6
1027                 "psubw %%mm6, %%mm2                             \n\t" // L5 - L6
1028                 "movq (%%ebx, %1), %%mm6                        \n\t"
1029                 "punpckhbw %%mm7, %%mm6                         \n\t" // H6
1030                 "psubw %%mm6, %%mm3                             \n\t" // H5 - H6
1031
1032                 "paddw %%mm0, %%mm0                             \n\t" // 2L4
1033                 "paddw %%mm1, %%mm1                             \n\t" // 2H4
1034                 "psubw %%mm2, %%mm0                             \n\t" // 2L4 - L5 + L6
1035                 "psubw %%mm3, %%mm1                             \n\t" // 2H4 - H5 + H6
1036
1037                 "psllw $2, %%mm2                                \n\t" // 4L5 - 4L6
1038                 "psllw $2, %%mm3                                \n\t" // 4H5 - 4H6
1039                 "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6
1040                 "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6
1041
1042                 "movq (%%ebx, %1, 2), %%mm2                     \n\t"
1043                 "movq %%mm2, %%mm3                              \n\t"
1044                 "punpcklbw %%mm7, %%mm2                         \n\t" // L7
1045                 "punpckhbw %%mm7, %%mm3                         \n\t" // H7
1046
1047                 "paddw %%mm2, %%mm2                             \n\t" // 2L7
1048                 "paddw %%mm3, %%mm3                             \n\t" // 2H7
1049                 "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6 - 2L7
1050                 "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6 - 2H7
1051
1052                 "movq "MANGLE(temp0)", %%mm2                    \n\t" // 2L0 - 5L1 + 5L2 - 2L3
1053                 "movq "MANGLE(temp1)", %%mm3                    \n\t" // 2H0 - 5H1 + 5H2 - 2H3
1054
1055 #ifdef HAVE_MMX2
1056                 "movq %%mm7, %%mm6                              \n\t" // 0
1057                 "psubw %%mm0, %%mm6                             \n\t"
1058                 "pmaxsw %%mm6, %%mm0                            \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1059                 "movq %%mm7, %%mm6                              \n\t" // 0
1060                 "psubw %%mm1, %%mm6                             \n\t"
1061                 "pmaxsw %%mm6, %%mm1                            \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1062                 "movq %%mm7, %%mm6                              \n\t" // 0
1063                 "psubw %%mm2, %%mm6                             \n\t"
1064                 "pmaxsw %%mm6, %%mm2                            \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1065                 "movq %%mm7, %%mm6                              \n\t" // 0
1066                 "psubw %%mm3, %%mm6                             \n\t"
1067                 "pmaxsw %%mm6, %%mm3                            \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1068 #else
1069                 "movq %%mm7, %%mm6                              \n\t" // 0
1070                 "pcmpgtw %%mm0, %%mm6                           \n\t"
1071                 "pxor %%mm6, %%mm0                              \n\t"
1072                 "psubw %%mm6, %%mm0                             \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1073                 "movq %%mm7, %%mm6                              \n\t" // 0
1074                 "pcmpgtw %%mm1, %%mm6                           \n\t"
1075                 "pxor %%mm6, %%mm1                              \n\t"
1076                 "psubw %%mm6, %%mm1                             \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1077                 "movq %%mm7, %%mm6                              \n\t" // 0
1078                 "pcmpgtw %%mm2, %%mm6                           \n\t"
1079                 "pxor %%mm6, %%mm2                              \n\t"
1080                 "psubw %%mm6, %%mm2                             \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1081                 "movq %%mm7, %%mm6                              \n\t" // 0
1082                 "pcmpgtw %%mm3, %%mm6                           \n\t"
1083                 "pxor %%mm6, %%mm3                              \n\t"
1084                 "psubw %%mm6, %%mm3                             \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1085 #endif
1086
1087 #ifdef HAVE_MMX2
1088                 "pminsw %%mm2, %%mm0                            \n\t"
1089                 "pminsw %%mm3, %%mm1                            \n\t"
1090 #else
1091                 "movq %%mm0, %%mm6                              \n\t"
1092                 "psubusw %%mm2, %%mm6                           \n\t"
1093                 "psubw %%mm6, %%mm0                             \n\t"
1094                 "movq %%mm1, %%mm6                              \n\t"
1095                 "psubusw %%mm3, %%mm6                           \n\t"
1096                 "psubw %%mm6, %%mm1                             \n\t"
1097 #endif
1098
1099                 "movq %%mm7, %%mm6                              \n\t" // 0
1100                 "pcmpgtw %%mm4, %%mm6                           \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1101                 "pxor %%mm6, %%mm4                              \n\t"
1102                 "psubw %%mm6, %%mm4                             \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1103                 "pcmpgtw %%mm5, %%mm7                           \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1104                 "pxor %%mm7, %%mm5                              \n\t"
1105                 "psubw %%mm7, %%mm5                             \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1106 // 100 opcodes
1107                 "movd %2, %%mm2                                 \n\t" // QP
1108                 "punpcklwd %%mm2, %%mm2                         \n\t"
1109                 "punpcklwd %%mm2, %%mm2                         \n\t"
1110                 "psllw $3, %%mm2                                \n\t" // 8QP
1111                 "movq %%mm2, %%mm3                              \n\t" // 8QP
1112                 "pcmpgtw %%mm4, %%mm2                           \n\t"
1113                 "pcmpgtw %%mm5, %%mm3                           \n\t"
1114                 "pand %%mm2, %%mm4                              \n\t"
1115                 "pand %%mm3, %%mm5                              \n\t"
1116
1117
1118                 "psubusw %%mm0, %%mm4                           \n\t" // hd
1119                 "psubusw %%mm1, %%mm5                           \n\t" // ld
1120
1121
1122                 "movq "MANGLE(w05)", %%mm2                      \n\t" // 5
1123                 "pmullw %%mm2, %%mm4                            \n\t"
1124                 "pmullw %%mm2, %%mm5                            \n\t"
1125                 "movq "MANGLE(w20)", %%mm2                      \n\t" // 32
1126                 "paddw %%mm2, %%mm4                             \n\t"
1127                 "paddw %%mm2, %%mm5                             \n\t"
1128                 "psrlw $6, %%mm4                                \n\t"
1129                 "psrlw $6, %%mm5                                \n\t"
1130
1131 /*
1132                 "movq w06, %%mm2                                \n\t" // 6
1133                 "paddw %%mm2, %%mm4                             \n\t"
1134                 "paddw %%mm2, %%mm5                             \n\t"
1135                 "movq w1400, %%mm2                              \n\t" // 1400h = 5120 = 5/64*2^16
1136 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120
1137                 "pmulhw %%mm2, %%mm4                            \n\t" // hd/13
1138                 "pmulhw %%mm2, %%mm5                            \n\t" // ld/13
1139 */
1140
1141                 "movq "MANGLE(temp2)", %%mm0                    \n\t" // L3 - L4
1142                 "movq "MANGLE(temp3)", %%mm1                    \n\t" // H3 - H4
1143
1144                 "pxor %%mm2, %%mm2                              \n\t"
1145                 "pxor %%mm3, %%mm3                              \n\t"
1146
1147                 "pcmpgtw %%mm0, %%mm2                           \n\t" // sign (L3-L4)
1148                 "pcmpgtw %%mm1, %%mm3                           \n\t" // sign (H3-H4)
1149                 "pxor %%mm2, %%mm0                              \n\t"
1150                 "pxor %%mm3, %%mm1                              \n\t"
1151                 "psubw %%mm2, %%mm0                             \n\t" // |L3-L4|
1152                 "psubw %%mm3, %%mm1                             \n\t" // |H3-H4|
1153                 "psrlw $1, %%mm0                                \n\t" // |L3 - L4|/2
1154                 "psrlw $1, %%mm1                                \n\t" // |H3 - H4|/2
1155
1156                 "pxor %%mm6, %%mm2                              \n\t"
1157                 "pxor %%mm7, %%mm3                              \n\t"
1158                 "pand %%mm2, %%mm4                              \n\t"
1159                 "pand %%mm3, %%mm5                              \n\t"
1160
1161 #ifdef HAVE_MMX2
1162                 "pminsw %%mm0, %%mm4                            \n\t"
1163                 "pminsw %%mm1, %%mm5                            \n\t"
1164 #else
1165                 "movq %%mm4, %%mm2                              \n\t"
1166                 "psubusw %%mm0, %%mm2                           \n\t"
1167                 "psubw %%mm2, %%mm4                             \n\t"
1168                 "movq %%mm5, %%mm2                              \n\t"
1169                 "psubusw %%mm1, %%mm2                           \n\t"
1170                 "psubw %%mm2, %%mm5                             \n\t"
1171 #endif
1172                 "pxor %%mm6, %%mm4                              \n\t"
1173                 "pxor %%mm7, %%mm5                              \n\t"
1174                 "psubw %%mm6, %%mm4                             \n\t"
1175                 "psubw %%mm7, %%mm5                             \n\t"
1176                 "packsswb %%mm5, %%mm4                          \n\t"
1177                 "movq (%%eax, %1, 2), %%mm0                     \n\t"
1178                 "paddb   %%mm4, %%mm0                           \n\t"
1179                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
1180                 "movq (%0, %1, 4), %%mm0                        \n\t"
1181                 "psubb %%mm4, %%mm0                             \n\t"
1182                 "movq %%mm0, (%0, %1, 4)                        \n\t"
1183
1184                 :
1185                 : "r" (src), "r" (stride), "r" (QP)
1186                 : "%eax", "%ebx"
1187         );
1188 #else
1189         const int l1= stride;
1190         const int l2= stride + l1;
1191         const int l3= stride + l2;
1192         const int l4= stride + l3;
1193         const int l5= stride + l4;
1194         const int l6= stride + l5;
1195         const int l7= stride + l6;
1196         const int l8= stride + l7;
1197 //      const int l9= stride + l8;
1198         int x;
1199         src+= stride*3;
1200         for(x=0; x<BLOCK_SIZE; x++)
1201         {
1202                 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1203                 if(ABS(middleEnergy) < 8*QP)
1204                 {
1205                         const int q=(src[l4] - src[l5])/2;
1206                         const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1207                         const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1208
1209                         int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
1210                         d= MAX(d, 0);
1211
1212                         d= (5*d + 32) >> 6;
1213                         d*= SIGN(-middleEnergy);
1214
1215                         if(q>0)
1216                         {
1217                                 d= d<0 ? 0 : d;
1218                                 d= d>q ? q : d;
1219                         }
1220                         else
1221                         {
1222                                 d= d>0 ? 0 : d;
1223                                 d= d<q ? q : d;
1224                         }
1225
1226                         src[l4]-= d;
1227                         src[l5]+= d;
1228                 }
1229                 src++;
1230         }
1231 #endif
1232 }
1233
1234 static inline void RENAME(dering)(uint8_t src[], int stride, int QP)
1235 {
1236 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1237         asm volatile(
1238                 "movq "MANGLE(pQPb)", %%mm0                     \n\t"
1239                 "paddusb %%mm0, %%mm0                           \n\t"
1240                 "movq %%mm0, "MANGLE(pQPb2)"                    \n\t"
1241
1242                 "leal (%0, %1), %%eax                           \n\t"
1243                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1244 //      0       1       2       3       4       5       6       7       8       9
1245 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1246
1247                 "pcmpeqb %%mm7, %%mm7                           \n\t"
1248                 "pxor %%mm6, %%mm6                              \n\t"
1249 #undef FIND_MIN_MAX
1250 #ifdef HAVE_MMX2
1251 #define FIND_MIN_MAX(addr)\
1252                 "movq " #addr ", %%mm0                          \n\t"\
1253                 "pminub %%mm0, %%mm7                            \n\t"\
1254                 "pmaxub %%mm0, %%mm6                            \n\t"
1255 #else
1256 #define FIND_MIN_MAX(addr)\
1257                 "movq " #addr ", %%mm0                          \n\t"\
1258                 "movq %%mm7, %%mm1                              \n\t"\
1259                 "psubusb %%mm0, %%mm6                           \n\t"\
1260                 "paddb %%mm0, %%mm6                             \n\t"\
1261                 "psubusb %%mm0, %%mm1                           \n\t"\
1262                 "psubb %%mm1, %%mm7                             \n\t"
1263 #endif
1264
1265 FIND_MIN_MAX((%%eax))
1266 FIND_MIN_MAX((%%eax, %1))
1267 FIND_MIN_MAX((%%eax, %1, 2))
1268 FIND_MIN_MAX((%0, %1, 4))
1269 FIND_MIN_MAX((%%ebx))
1270 FIND_MIN_MAX((%%ebx, %1))
1271 FIND_MIN_MAX((%%ebx, %1, 2))
1272 FIND_MIN_MAX((%0, %1, 8))
1273
1274                 "movq %%mm7, %%mm4                              \n\t"
1275                 "psrlq $8, %%mm7                                \n\t"
1276 #ifdef HAVE_MMX2
1277                 "pminub %%mm4, %%mm7                            \n\t" // min of pixels
1278                 "pshufw $0xF9, %%mm7, %%mm4                     \n\t"
1279                 "pminub %%mm4, %%mm7                            \n\t" // min of pixels
1280                 "pshufw $0xFE, %%mm7, %%mm4                     \n\t"
1281                 "pminub %%mm4, %%mm7                            \n\t"
1282 #else
1283                 "movq %%mm7, %%mm1                              \n\t"
1284                 "psubusb %%mm4, %%mm1                           \n\t"
1285                 "psubb %%mm1, %%mm7                             \n\t"
1286                 "movq %%mm7, %%mm4                              \n\t"
1287                 "psrlq $16, %%mm7                               \n\t"
1288                 "movq %%mm7, %%mm1                              \n\t"
1289                 "psubusb %%mm4, %%mm1                           \n\t"
1290                 "psubb %%mm1, %%mm7                             \n\t"
1291                 "movq %%mm7, %%mm4                              \n\t"
1292                 "psrlq $32, %%mm7                               \n\t"
1293                 "movq %%mm7, %%mm1                              \n\t"
1294                 "psubusb %%mm4, %%mm1                           \n\t"
1295                 "psubb %%mm1, %%mm7                             \n\t"
1296 #endif
1297
1298
1299                 "movq %%mm6, %%mm4                              \n\t"
1300                 "psrlq $8, %%mm6                                \n\t"
1301 #ifdef HAVE_MMX2
1302                 "pmaxub %%mm4, %%mm6                            \n\t" // max of pixels
1303                 "pshufw $0xF9, %%mm6, %%mm4                     \n\t"
1304                 "pmaxub %%mm4, %%mm6                            \n\t"
1305                 "pshufw $0xFE, %%mm6, %%mm4                     \n\t"
1306                 "pmaxub %%mm4, %%mm6                            \n\t"
1307 #else
1308                 "psubusb %%mm4, %%mm6                           \n\t"
1309                 "paddb %%mm4, %%mm6                             \n\t"
1310                 "movq %%mm6, %%mm4                              \n\t"
1311                 "psrlq $16, %%mm6                               \n\t"
1312                 "psubusb %%mm4, %%mm6                           \n\t"
1313                 "paddb %%mm4, %%mm6                             \n\t"
1314                 "movq %%mm6, %%mm4                              \n\t"
1315                 "psrlq $32, %%mm6                               \n\t"
1316                 "psubusb %%mm4, %%mm6                           \n\t"
1317                 "paddb %%mm4, %%mm6                             \n\t"
1318 #endif
1319                 "movq %%mm6, %%mm0                              \n\t" // max
1320                 "psubb %%mm7, %%mm6                             \n\t" // max - min
1321                 "movd %%mm6, %%ecx                              \n\t"
1322                 "cmpb "MANGLE(deringThreshold)", %%cl           \n\t"
1323                 " jb 1f                                         \n\t"
1324                 PAVGB(%%mm0, %%mm7)                                   // a=(max + min)/2
1325                 "punpcklbw %%mm7, %%mm7                         \n\t"
1326                 "punpcklbw %%mm7, %%mm7                         \n\t"
1327                 "punpcklbw %%mm7, %%mm7                         \n\t"
1328                 "movq %%mm7, "MANGLE(temp0)"                    \n\t"
1329
1330                 "movq (%0), %%mm0                               \n\t" // L10
1331                 "movq %%mm0, %%mm1                              \n\t" // L10
1332                 "movq %%mm0, %%mm2                              \n\t" // L10
1333                 "psllq $8, %%mm1                                \n\t"
1334                 "psrlq $8, %%mm2                                \n\t"
1335                 "movd -4(%0), %%mm3                             \n\t"
1336                 "movd 8(%0), %%mm4                              \n\t"
1337                 "psrlq $24, %%mm3                               \n\t"
1338                 "psllq $56, %%mm4                               \n\t"
1339                 "por %%mm3, %%mm1                               \n\t" // L00
1340                 "por %%mm4, %%mm2                               \n\t" // L20
1341                 "movq %%mm1, %%mm3                              \n\t" // L00
1342                 PAVGB(%%mm2, %%mm1)                                   // (L20 + L00)/2
1343                 PAVGB(%%mm0, %%mm1)                                   // (L20 + L00 + 2L10)/4
1344                 "psubusb %%mm7, %%mm0                           \n\t"
1345                 "psubusb %%mm7, %%mm2                           \n\t"
1346                 "psubusb %%mm7, %%mm3                           \n\t"
1347                 "pcmpeqb "MANGLE(b00)", %%mm0                   \n\t" // L10 > a ? 0 : -1
1348                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // L20 > a ? 0 : -1
1349                 "pcmpeqb "MANGLE(b00)", %%mm3                   \n\t" // L00 > a ? 0 : -1
1350                 "paddb %%mm2, %%mm0                             \n\t"
1351                 "paddb %%mm3, %%mm0                             \n\t"
1352
1353                 "movq (%%eax), %%mm2                            \n\t" // L11
1354                 "movq %%mm2, %%mm3                              \n\t" // L11
1355                 "movq %%mm2, %%mm4                              \n\t" // L11
1356                 "psllq $8, %%mm3                                \n\t"
1357                 "psrlq $8, %%mm4                                \n\t"
1358                 "movd -4(%%eax), %%mm5                          \n\t"
1359                 "movd 8(%%eax), %%mm6                           \n\t"
1360                 "psrlq $24, %%mm5                               \n\t"
1361                 "psllq $56, %%mm6                               \n\t"
1362                 "por %%mm5, %%mm3                               \n\t" // L01
1363                 "por %%mm6, %%mm4                               \n\t" // L21
1364                 "movq %%mm3, %%mm5                              \n\t" // L01
1365                 PAVGB(%%mm4, %%mm3)                                   // (L21 + L01)/2
1366                 PAVGB(%%mm2, %%mm3)                                   // (L21 + L01 + 2L11)/4
1367                 "psubusb %%mm7, %%mm2                           \n\t"
1368                 "psubusb %%mm7, %%mm4                           \n\t"
1369                 "psubusb %%mm7, %%mm5                           \n\t"
1370                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // L11 > a ? 0 : -1
1371                 "pcmpeqb "MANGLE(b00)", %%mm4                   \n\t" // L21 > a ? 0 : -1
1372                 "pcmpeqb "MANGLE(b00)", %%mm5                   \n\t" // L01 > a ? 0 : -1
1373                 "paddb %%mm4, %%mm2                             \n\t"
1374                 "paddb %%mm5, %%mm2                             \n\t"
1375 // 0, 2, 3, 1
1376 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1377                 "movq " #src ", " #sx "                         \n\t" /* src[0] */\
1378                 "movq " #sx ", " #lx "                          \n\t" /* src[0] */\
1379                 "movq " #sx ", " #t0 "                          \n\t" /* src[0] */\
1380                 "psllq $8, " #lx "                              \n\t"\
1381                 "psrlq $8, " #t0 "                              \n\t"\
1382                 "movd -4" #src ", " #t1 "                       \n\t"\
1383                 "psrlq $24, " #t1 "                             \n\t"\
1384                 "por " #t1 ", " #lx "                           \n\t" /* src[-1] */\
1385                 "movd 8" #src ", " #t1 "                        \n\t"\
1386                 "psllq $56, " #t1 "                             \n\t"\
1387                 "por " #t1 ", " #t0 "                           \n\t" /* src[+1] */\
1388                 "movq " #lx ", " #t1 "                          \n\t" /* src[-1] */\
1389                 PAVGB(t0, lx)                                         /* (src[-1] + src[+1])/2 */\
1390                 PAVGB(sx, lx)                                 /* (src[-1] + 2src[0] + src[+1])/4 */\
1391                 PAVGB(lx, pplx)                                      \
1392                 "movq " #lx ", "MANGLE(temp1)"                  \n\t"\
1393                 "movq "MANGLE(temp0)", " #lx "                  \n\t"\
1394                 "psubusb " #lx ", " #t1 "                       \n\t"\
1395                 "psubusb " #lx ", " #t0 "                       \n\t"\
1396                 "psubusb " #lx ", " #sx "                       \n\t"\
1397                 "movq "MANGLE(b00)", " #lx "                    \n\t"\
1398                 "pcmpeqb " #lx ", " #t1 "                       \n\t" /* src[-1] > a ? 0 : -1*/\
1399                 "pcmpeqb " #lx ", " #t0 "                       \n\t" /* src[+1] > a ? 0 : -1*/\
1400                 "pcmpeqb " #lx ", " #sx "                       \n\t" /* src[0]  > a ? 0 : -1*/\
1401                 "paddb " #t1 ", " #t0 "                         \n\t"\
1402                 "paddb " #t0 ", " #sx "                         \n\t"\
1403 \
1404                 PAVGB(plx, pplx)                                      /* filtered */\
1405                 "movq " #dst ", " #t0 "                         \n\t" /* dst */\
1406                 "movq " #t0 ", " #t1 "                          \n\t" /* dst */\
1407                 "psubusb "MANGLE(pQPb2)", " #t0 "               \n\t"\
1408                 "paddusb "MANGLE(pQPb2)", " #t1 "               \n\t"\
1409                 PMAXUB(t0, pplx)\
1410                 PMINUB(t1, pplx, t0)\
1411                 "paddb " #sx ", " #ppsx "                       \n\t"\
1412                 "paddb " #psx ", " #ppsx "                      \n\t"\
1413                 "#paddb "MANGLE(b02)", " #ppsx "                \n\t"\
1414                 "pand "MANGLE(b08)", " #ppsx "                  \n\t"\
1415                 "pcmpeqb " #lx ", " #ppsx "                     \n\t"\
1416                 "pand " #ppsx ", " #pplx "                      \n\t"\
1417                 "pandn " #dst ", " #ppsx "                      \n\t"\
1418                 "por " #pplx ", " #ppsx "                       \n\t"\
1419                 "movq " #ppsx ", " #dst "                       \n\t"\
1420                 "movq "MANGLE(temp1)", " #lx "                  \n\t"
1421
1422 /*
1423 0000000
1424 1111111
1425
1426 1111110
1427 1111101
1428 1111100
1429 1111011
1430 1111010
1431 1111001
1432
1433 1111000
1434 1110111
1435
1436 */
1437 //DERING_CORE(dst,src                  ,ppsx ,psx  ,sx   ,pplx ,plx  ,lx   ,t0   ,t1)
1438 DERING_CORE((%%eax),(%%eax, %1)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1439 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1440 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1441 DERING_CORE((%0, %1, 4),(%%ebx)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1442 DERING_CORE((%%ebx),(%%ebx, %1)        ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1443 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1444 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1445 DERING_CORE((%0, %1, 8),(%%ebx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1446
1447                 "1:                     \n\t"
1448                 : : "r" (src), "r" (stride), "r" (QP)
1449                 : "%eax", "%ebx", "%ecx"
1450         );
1451 #else
1452         int y;
1453         int min=255;
1454         int max=0;
1455         int avg;
1456         uint8_t *p;
1457         int s[10];
1458
1459         for(y=1; y<9; y++)
1460         {
1461                 int x;
1462                 p= src + stride*y;
1463                 for(x=1; x<9; x++)
1464                 {
1465                         p++;
1466                         if(*p > max) max= *p;
1467                         if(*p < min) min= *p;
1468                 }
1469         }
1470         avg= (min + max + 1)/2;
1471
1472         if(max - min <deringThreshold) return;
1473
1474         for(y=0; y<10; y++)
1475         {
1476                 int x;
1477                 int t = 0;
1478                 p= src + stride*y;
1479                 for(x=0; x<10; x++)
1480                 {
1481                         if(*p > avg) t |= (1<<x);
1482                         p++;
1483                 }
1484                 t |= (~t)<<16;
1485                 t &= (t<<1) & (t>>1);
1486                 s[y] = t;
1487         }
1488
1489         for(y=1; y<9; y++)
1490         {
1491                 int x;
1492                 int t = s[y-1] & s[y] & s[y+1];
1493                 t|= t>>16;
1494
1495                 p= src + stride*y;
1496                 for(x=1; x<9; x++)
1497                 {
1498                         p++;
1499                         if(t & (1<<x))
1500                         {
1501                                 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1502                                       +2*(*(p     -1)) + 4*(*p         ) + 2*(*(p     +1))
1503                                       +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1504                                 f= (f + 8)>>4;
1505
1506 #ifdef DEBUG_DERING_THRESHOLD
1507                                 asm volatile("emms\n\t":);
1508                                 {
1509                                 static long long numPixels=0;
1510                                 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1511 //                              if((max-min)<20 || (max-min)*QP<200)
1512 //                              if((max-min)*QP < 500)
1513 //                              if(max-min<QP/2)
1514                                 if(max-min < 20)
1515                                 {
1516                                         static int numSkiped=0;
1517                                         static int errorSum=0;
1518                                         static int worstQP=0;
1519                                         static int worstRange=0;
1520                                         static int worstDiff=0;
1521                                         int diff= (f - *p);
1522                                         int absDiff= ABS(diff);
1523                                         int error= diff*diff;
1524
1525                                         if(x==1 || x==8 || y==1 || y==8) continue;
1526
1527                                         numSkiped++;
1528                                         if(absDiff > worstDiff)
1529                                         {
1530                                                 worstDiff= absDiff;
1531                                                 worstQP= QP;
1532                                                 worstRange= max-min;
1533                                         }
1534                                         errorSum+= error;
1535
1536                                         if(1024LL*1024LL*1024LL % numSkiped == 0)
1537                                         {
1538                                                 printf( "sum:%1.3f, skip:%d, wQP:%d, "
1539                                                         "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1540                                                         (float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
1541                                                         worstDiff, (float)numSkiped/numPixels);
1542                                         }
1543                                 }
1544                                 }
1545 #endif
1546                                 if     (*p + 2*QP < f) *p= *p + 2*QP;
1547                                 else if(*p - 2*QP > f) *p= *p - 2*QP;
1548                                 else *p=f;
1549                         }
1550                 }
1551         }
1552 #ifdef DEBUG_DERING_THRESHOLD
1553         if(max-min < 20)
1554         {
1555                 for(y=1; y<9; y++)
1556                 {
1557                         int x;
1558                         int t = 0;
1559                         p= src + stride*y;
1560                         for(x=1; x<9; x++)
1561                         {
1562                                 p++;
1563                                 *p = MIN(*p + 20, 255);
1564                         }
1565                 }
1566 //              src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1567         }
1568 #endif
1569 #endif
1570 }
1571
1572 /**
1573  * Deinterlaces the given block
1574  * will be called for every 8x8 block and can read & write from line 4-15
1575  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1576  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1577  */
1578 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1579 {
1580 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1581         src+= 4*stride;
1582         asm volatile(
1583                 "leal (%0, %1), %%eax                           \n\t"
1584                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1585 //      0       1       2       3       4       5       6       7       8       9
1586 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1587
1588                 "movq (%0), %%mm0                               \n\t"
1589                 "movq (%%eax, %1), %%mm1                        \n\t"
1590                 PAVGB(%%mm1, %%mm0)
1591                 "movq %%mm0, (%%eax)                            \n\t"
1592                 "movq (%0, %1, 4), %%mm0                        \n\t"
1593                 PAVGB(%%mm0, %%mm1)
1594                 "movq %%mm1, (%%eax, %1, 2)                     \n\t"
1595                 "movq (%%ebx, %1), %%mm1                        \n\t"
1596                 PAVGB(%%mm1, %%mm0)
1597                 "movq %%mm0, (%%ebx)                            \n\t"
1598                 "movq (%0, %1, 8), %%mm0                        \n\t"
1599                 PAVGB(%%mm0, %%mm1)
1600                 "movq %%mm1, (%%ebx, %1, 2)                     \n\t"
1601
1602                 : : "r" (src), "r" (stride)
1603                 : "%eax", "%ebx"
1604         );
1605 #else
1606         int x;
1607         src+= 4*stride;
1608         for(x=0; x<8; x++)
1609         {
1610                 src[stride]   = (src[0]        + src[stride*2])>>1;
1611                 src[stride*3] = (src[stride*2] + src[stride*4])>>1;
1612                 src[stride*5] = (src[stride*4] + src[stride*6])>>1;
1613                 src[stride*7] = (src[stride*6] + src[stride*8])>>1;
1614                 src++;
1615         }
1616 #endif
1617 }
1618
1619 /**
1620  * Deinterlaces the given block
1621  * will be called for every 8x8 block and can read & write from line 4-15
1622  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1623  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1624  * this filter will read lines 3-15 and write 7-13
1625  * no cliping in C version
1626  */
1627 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1628 {
1629 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1630         src+= stride*3;
1631         asm volatile(
1632                 "leal (%0, %1), %%eax                           \n\t"
1633                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1634                 "leal (%%ebx, %1, 4), %%ecx                     \n\t"
1635                 "addl %1, %%ecx                                 \n\t"
1636                 "pxor %%mm7, %%mm7                              \n\t"
1637 //      0       1       2       3       4       5       6       7       8       9       10
1638 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1 ecx
1639
1640 #define DEINT_CUBIC(a,b,c,d,e)\
1641                 "movq " #a ", %%mm0                             \n\t"\
1642                 "movq " #b ", %%mm1                             \n\t"\
1643                 "movq " #d ", %%mm2                             \n\t"\
1644                 "movq " #e ", %%mm3                             \n\t"\
1645                 PAVGB(%%mm2, %%mm1)                                     /* (b+d) /2 */\
1646                 PAVGB(%%mm3, %%mm0)                                     /* a(a+e) /2 */\
1647                 "movq %%mm0, %%mm2                              \n\t"\
1648                 "punpcklbw %%mm7, %%mm0                         \n\t"\
1649                 "punpckhbw %%mm7, %%mm2                         \n\t"\
1650                 "movq %%mm1, %%mm3                              \n\t"\
1651                 "punpcklbw %%mm7, %%mm1                         \n\t"\
1652                 "punpckhbw %%mm7, %%mm3                         \n\t"\
1653                 "psubw %%mm1, %%mm0                             \n\t"   /* L(a+e - (b+d))/2 */\
1654                 "psubw %%mm3, %%mm2                             \n\t"   /* H(a+e - (b+d))/2 */\
1655                 "psraw $3, %%mm0                                \n\t"   /* L(a+e - (b+d))/16 */\
1656                 "psraw $3, %%mm2                                \n\t"   /* H(a+e - (b+d))/16 */\
1657                 "psubw %%mm0, %%mm1                             \n\t"   /* L(9b + 9d - a - e)/16 */\
1658                 "psubw %%mm2, %%mm3                             \n\t"   /* H(9b + 9d - a - e)/16 */\
1659                 "packuswb %%mm3, %%mm1                          \n\t"\
1660                 "movq %%mm1, " #c "                             \n\t"
1661
1662 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1))
1663 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8))
1664 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx))
1665 DEINT_CUBIC((%%ebx, %1), (%0, %1, 8), (%%ebx, %1, 4), (%%ecx), (%%ecx, %1, 2))
1666
1667                 : : "r" (src), "r" (stride)
1668                 : "%eax", "%ebx", "ecx"
1669         );
1670 #else
1671         int x;
1672         src+= stride*3;
1673         for(x=0; x<8; x++)
1674         {
1675                 src[stride*3] = (-src[0]        + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4;
1676                 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4;
1677                 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4;
1678                 src[stride*9] = (-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4;
1679                 src++;
1680         }
1681 #endif
1682 }
1683
1684 /**
1685  * Deinterlaces the given block
1686  * will be called for every 8x8 block and can read & write from line 4-15
1687  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1688  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1689  * will shift the image up by 1 line (FIXME if this is a problem)
1690  * this filter will read lines 4-13 and write 4-11
1691  */
1692 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride)
1693 {
1694 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1695         src+= 4*stride;
1696         asm volatile(
1697                 "leal (%0, %1), %%eax                           \n\t"
1698                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1699 //      0       1       2       3       4       5       6       7       8       9
1700 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1701
1702                 "movq (%0), %%mm0                               \n\t" // L0
1703                 "movq (%%eax, %1), %%mm1                        \n\t" // L2
1704                 PAVGB(%%mm1, %%mm0)                                   // L0+L2
1705                 "movq (%%eax), %%mm2                            \n\t" // L1
1706                 PAVGB(%%mm2, %%mm0)
1707                 "movq %%mm0, (%0)                               \n\t"
1708                 "movq (%%eax, %1, 2), %%mm0                     \n\t" // L3
1709                 PAVGB(%%mm0, %%mm2)                                   // L1+L3
1710                 PAVGB(%%mm1, %%mm2)                                   // 2L2 + L1 + L3
1711                 "movq %%mm2, (%%eax)                            \n\t"
1712                 "movq (%0, %1, 4), %%mm2                        \n\t" // L4
1713                 PAVGB(%%mm2, %%mm1)                                   // L2+L4
1714                 PAVGB(%%mm0, %%mm1)                                   // 2L3 + L2 + L4
1715                 "movq %%mm1, (%%eax, %1)                        \n\t"
1716                 "movq (%%ebx), %%mm1                            \n\t" // L5
1717                 PAVGB(%%mm1, %%mm0)                                   // L3+L5
1718                 PAVGB(%%mm2, %%mm0)                                   // 2L4 + L3 + L5
1719                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
1720                 "movq (%%ebx, %1), %%mm0                        \n\t" // L6
1721                 PAVGB(%%mm0, %%mm2)                                   // L4+L6
1722                 PAVGB(%%mm1, %%mm2)                                   // 2L5 + L4 + L6
1723                 "movq %%mm2, (%0, %1, 4)                        \n\t"
1724                 "movq (%%ebx, %1, 2), %%mm2                     \n\t" // L7
1725                 PAVGB(%%mm2, %%mm1)                                   // L5+L7
1726                 PAVGB(%%mm0, %%mm1)                                   // 2L6 + L5 + L7
1727                 "movq %%mm1, (%%ebx)                            \n\t"
1728                 "movq (%0, %1, 8), %%mm1                        \n\t" // L8
1729                 PAVGB(%%mm1, %%mm0)                                   // L6+L8
1730                 PAVGB(%%mm2, %%mm0)                                   // 2L7 + L6 + L8
1731                 "movq %%mm0, (%%ebx, %1)                        \n\t"
1732                 "movq (%%ebx, %1, 4), %%mm0                     \n\t" // L9
1733                 PAVGB(%%mm0, %%mm2)                                   // L7+L9
1734                 PAVGB(%%mm1, %%mm2)                                   // 2L8 + L7 + L9
1735                 "movq %%mm2, (%%ebx, %1, 2)                     \n\t"
1736
1737
1738                 : : "r" (src), "r" (stride)
1739                 : "%eax", "%ebx"
1740         );
1741 #else
1742         int x;
1743         src+= 4*stride;
1744         for(x=0; x<8; x++)
1745         {
1746                 src[0       ] = (src[0       ] + 2*src[stride  ] + src[stride*2])>>2;
1747                 src[stride  ] = (src[stride  ] + 2*src[stride*2] + src[stride*3])>>2;
1748                 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
1749                 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
1750                 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
1751                 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
1752                 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
1753                 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
1754                 src++;
1755         }
1756 #endif
1757 }
1758
1759 /**
1760  * Deinterlaces the given block
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 allready, but can be read too
1763  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1764  */
1765 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1766 {
1767 #ifdef HAVE_MMX
1768         src+= 4*stride;
1769 #ifdef HAVE_MMX2
1770         asm volatile(
1771                 "leal (%0, %1), %%eax                           \n\t"
1772                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1773 //      0       1       2       3       4       5       6       7       8       9
1774 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1775
1776                 "movq (%0), %%mm0                               \n\t" //
1777                 "movq (%%eax, %1), %%mm2                        \n\t" //
1778                 "movq (%%eax), %%mm1                            \n\t" //
1779                 "movq %%mm0, %%mm3                              \n\t"
1780                 "pmaxub %%mm1, %%mm0                            \n\t" //
1781                 "pminub %%mm3, %%mm1                            \n\t" //
1782                 "pmaxub %%mm2, %%mm1                            \n\t" //
1783                 "pminub %%mm1, %%mm0                            \n\t"
1784                 "movq %%mm0, (%%eax)                            \n\t"
1785
1786                 "movq (%0, %1, 4), %%mm0                        \n\t" //
1787                 "movq (%%eax, %1, 2), %%mm1                     \n\t" //
1788                 "movq %%mm2, %%mm3                              \n\t"
1789                 "pmaxub %%mm1, %%mm2                            \n\t" //
1790                 "pminub %%mm3, %%mm1                            \n\t" //
1791                 "pmaxub %%mm0, %%mm1                            \n\t" //
1792                 "pminub %%mm1, %%mm2                            \n\t"
1793                 "movq %%mm2, (%%eax, %1, 2)                     \n\t"
1794
1795                 "movq (%%ebx), %%mm2                            \n\t" //
1796                 "movq (%%ebx, %1), %%mm1                        \n\t" //
1797                 "movq %%mm2, %%mm3                              \n\t"
1798                 "pmaxub %%mm0, %%mm2                            \n\t" //
1799                 "pminub %%mm3, %%mm0                            \n\t" //
1800                 "pmaxub %%mm1, %%mm0                            \n\t" //
1801                 "pminub %%mm0, %%mm2                            \n\t"
1802                 "movq %%mm2, (%%ebx)                            \n\t"
1803
1804                 "movq (%%ebx, %1, 2), %%mm2                     \n\t" //
1805                 "movq (%0, %1, 8), %%mm0                        \n\t" //
1806                 "movq %%mm2, %%mm3                              \n\t"
1807                 "pmaxub %%mm0, %%mm2                            \n\t" //
1808                 "pminub %%mm3, %%mm0                            \n\t" //
1809                 "pmaxub %%mm1, %%mm0                            \n\t" //
1810                 "pminub %%mm0, %%mm2                            \n\t"
1811                 "movq %%mm2, (%%ebx, %1, 2)                     \n\t"
1812
1813
1814                 : : "r" (src), "r" (stride)
1815                 : "%eax", "%ebx"
1816         );
1817
1818 #else // MMX without MMX2
1819         asm volatile(
1820                 "leal (%0, %1), %%eax                           \n\t"
1821                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1822 //      0       1       2       3       4       5       6       7       8       9
1823 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1824                 "pxor %%mm7, %%mm7                              \n\t"
1825
1826 #define MEDIAN(a,b,c)\
1827                 "movq " #a ", %%mm0                             \n\t"\
1828                 "movq " #b ", %%mm2                             \n\t"\
1829                 "movq " #c ", %%mm1                             \n\t"\
1830                 "movq %%mm0, %%mm3                              \n\t"\
1831                 "movq %%mm1, %%mm4                              \n\t"\
1832                 "movq %%mm2, %%mm5                              \n\t"\
1833                 "psubusb %%mm1, %%mm3                           \n\t"\
1834                 "psubusb %%mm2, %%mm4                           \n\t"\
1835                 "psubusb %%mm0, %%mm5                           \n\t"\
1836                 "pcmpeqb %%mm7, %%mm3                           \n\t"\
1837                 "pcmpeqb %%mm7, %%mm4                           \n\t"\
1838                 "pcmpeqb %%mm7, %%mm5                           \n\t"\
1839                 "movq %%mm3, %%mm6                              \n\t"\
1840                 "pxor %%mm4, %%mm3                              \n\t"\
1841                 "pxor %%mm5, %%mm4                              \n\t"\
1842                 "pxor %%mm6, %%mm5                              \n\t"\
1843                 "por %%mm3, %%mm1                               \n\t"\
1844                 "por %%mm4, %%mm2                               \n\t"\
1845                 "por %%mm5, %%mm0                               \n\t"\
1846                 "pand %%mm2, %%mm0                              \n\t"\
1847                 "pand %%mm1, %%mm0                              \n\t"\
1848                 "movq %%mm0, " #b "                             \n\t"
1849
1850 MEDIAN((%0), (%%eax), (%%eax, %1))
1851 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
1852 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1))
1853 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8))
1854
1855                 : : "r" (src), "r" (stride)
1856                 : "%eax", "%ebx"
1857         );
1858 #endif // MMX
1859 #else
1860         //FIXME
1861         int x;
1862         src+= 4*stride;
1863         for(x=0; x<8; x++)
1864         {
1865                 src[0       ] = (src[0       ] + 2*src[stride  ] + src[stride*2])>>2;
1866                 src[stride  ] = (src[stride  ] + 2*src[stride*2] + src[stride*3])>>2;
1867                 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
1868                 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
1869                 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
1870                 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
1871                 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
1872                 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
1873                 src++;
1874         }
1875 #endif
1876 }
1877
1878 #ifdef HAVE_MMX
1879 /**
1880  * transposes and shift the given 8x8 Block into dst1 and dst2
1881  */
1882 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
1883 {
1884         asm(
1885                 "leal (%0, %1), %%eax                           \n\t"
1886                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1887 //      0       1       2       3       4       5       6       7       8       9
1888 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1889                 "movq (%0), %%mm0               \n\t" // 12345678
1890                 "movq (%%eax), %%mm1            \n\t" // abcdefgh
1891                 "movq %%mm0, %%mm2              \n\t" // 12345678
1892                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
1893                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
1894
1895                 "movq (%%eax, %1), %%mm1        \n\t"
1896                 "movq (%%eax, %1, 2), %%mm3     \n\t"
1897                 "movq %%mm1, %%mm4              \n\t"
1898                 "punpcklbw %%mm3, %%mm1         \n\t"
1899                 "punpckhbw %%mm3, %%mm4         \n\t"
1900
1901                 "movq %%mm0, %%mm3              \n\t"
1902                 "punpcklwd %%mm1, %%mm0         \n\t"
1903                 "punpckhwd %%mm1, %%mm3         \n\t"
1904                 "movq %%mm2, %%mm1              \n\t"
1905                 "punpcklwd %%mm4, %%mm2         \n\t"
1906                 "punpckhwd %%mm4, %%mm1         \n\t"
1907
1908                 "movd %%mm0, 128(%2)            \n\t"
1909                 "psrlq $32, %%mm0               \n\t"
1910                 "movd %%mm0, 144(%2)            \n\t"
1911                 "movd %%mm3, 160(%2)            \n\t"
1912                 "psrlq $32, %%mm3               \n\t"
1913                 "movd %%mm3, 176(%2)            \n\t"
1914                 "movd %%mm3, 48(%3)             \n\t"
1915                 "movd %%mm2, 192(%2)            \n\t"
1916                 "movd %%mm2, 64(%3)             \n\t"
1917                 "psrlq $32, %%mm2               \n\t"
1918                 "movd %%mm2, 80(%3)             \n\t"
1919                 "movd %%mm1, 96(%3)             \n\t"
1920                 "psrlq $32, %%mm1               \n\t"
1921                 "movd %%mm1, 112(%3)            \n\t"
1922
1923                 "movq (%0, %1, 4), %%mm0        \n\t" // 12345678
1924                 "movq (%%ebx), %%mm1            \n\t" // abcdefgh
1925                 "movq %%mm0, %%mm2              \n\t" // 12345678
1926                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
1927                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
1928
1929                 "movq (%%ebx, %1), %%mm1        \n\t"
1930                 "movq (%%ebx, %1, 2), %%mm3     \n\t"
1931                 "movq %%mm1, %%mm4              \n\t"
1932                 "punpcklbw %%mm3, %%mm1         \n\t"
1933                 "punpckhbw %%mm3, %%mm4         \n\t"
1934
1935                 "movq %%mm0, %%mm3              \n\t"
1936                 "punpcklwd %%mm1, %%mm0         \n\t"
1937                 "punpckhwd %%mm1, %%mm3         \n\t"
1938                 "movq %%mm2, %%mm1              \n\t"
1939                 "punpcklwd %%mm4, %%mm2         \n\t"
1940                 "punpckhwd %%mm4, %%mm1         \n\t"
1941
1942                 "movd %%mm0, 132(%2)            \n\t"
1943                 "psrlq $32, %%mm0               \n\t"
1944                 "movd %%mm0, 148(%2)            \n\t"
1945                 "movd %%mm3, 164(%2)            \n\t"
1946                 "psrlq $32, %%mm3               \n\t"
1947                 "movd %%mm3, 180(%2)            \n\t"
1948                 "movd %%mm3, 52(%3)             \n\t"
1949                 "movd %%mm2, 196(%2)            \n\t"
1950                 "movd %%mm2, 68(%3)             \n\t"
1951                 "psrlq $32, %%mm2               \n\t"
1952                 "movd %%mm2, 84(%3)             \n\t"
1953                 "movd %%mm1, 100(%3)            \n\t"
1954                 "psrlq $32, %%mm1               \n\t"
1955                 "movd %%mm1, 116(%3)            \n\t"
1956
1957
1958         :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)
1959         : "%eax", "%ebx"
1960         );
1961 }
1962
1963 /**
1964  * transposes the given 8x8 block
1965  */
1966 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
1967 {
1968         asm(
1969                 "leal (%0, %1), %%eax                           \n\t"
1970                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1971 //      0       1       2       3       4       5       6       7       8       9
1972 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1973                 "movq (%2), %%mm0               \n\t" // 12345678
1974                 "movq 16(%2), %%mm1             \n\t" // abcdefgh
1975                 "movq %%mm0, %%mm2              \n\t" // 12345678
1976                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
1977                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
1978
1979                 "movq 32(%2), %%mm1             \n\t"
1980                 "movq 48(%2), %%mm3             \n\t"
1981                 "movq %%mm1, %%mm4              \n\t"
1982                 "punpcklbw %%mm3, %%mm1         \n\t"
1983                 "punpckhbw %%mm3, %%mm4         \n\t"
1984
1985                 "movq %%mm0, %%mm3              \n\t"
1986                 "punpcklwd %%mm1, %%mm0         \n\t"
1987                 "punpckhwd %%mm1, %%mm3         \n\t"
1988                 "movq %%mm2, %%mm1              \n\t"
1989                 "punpcklwd %%mm4, %%mm2         \n\t"
1990                 "punpckhwd %%mm4, %%mm1         \n\t"
1991
1992                 "movd %%mm0, (%0)               \n\t"
1993                 "psrlq $32, %%mm0               \n\t"
1994                 "movd %%mm0, (%%eax)            \n\t"
1995                 "movd %%mm3, (%%eax, %1)        \n\t"
1996                 "psrlq $32, %%mm3               \n\t"
1997                 "movd %%mm3, (%%eax, %1, 2)     \n\t"
1998                 "movd %%mm2, (%0, %1, 4)        \n\t"
1999                 "psrlq $32, %%mm2               \n\t"
2000                 "movd %%mm2, (%%ebx)            \n\t"
2001                 "movd %%mm1, (%%ebx, %1)        \n\t"
2002                 "psrlq $32, %%mm1               \n\t"
2003                 "movd %%mm1, (%%ebx, %1, 2)     \n\t"
2004
2005
2006                 "movq 64(%2), %%mm0             \n\t" // 12345678
2007                 "movq 80(%2), %%mm1             \n\t" // abcdefgh
2008                 "movq %%mm0, %%mm2              \n\t" // 12345678
2009                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
2010                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
2011
2012                 "movq 96(%2), %%mm1             \n\t"
2013                 "movq 112(%2), %%mm3            \n\t"
2014                 "movq %%mm1, %%mm4              \n\t"
2015                 "punpcklbw %%mm3, %%mm1         \n\t"
2016                 "punpckhbw %%mm3, %%mm4         \n\t"
2017
2018                 "movq %%mm0, %%mm3              \n\t"
2019                 "punpcklwd %%mm1, %%mm0         \n\t"
2020                 "punpckhwd %%mm1, %%mm3         \n\t"
2021                 "movq %%mm2, %%mm1              \n\t"
2022                 "punpcklwd %%mm4, %%mm2         \n\t"
2023                 "punpckhwd %%mm4, %%mm1         \n\t"
2024
2025                 "movd %%mm0, 4(%0)              \n\t"
2026                 "psrlq $32, %%mm0               \n\t"
2027                 "movd %%mm0, 4(%%eax)           \n\t"
2028                 "movd %%mm3, 4(%%eax, %1)       \n\t"
2029                 "psrlq $32, %%mm3               \n\t"
2030                 "movd %%mm3, 4(%%eax, %1, 2)    \n\t"
2031                 "movd %%mm2, 4(%0, %1, 4)       \n\t"
2032                 "psrlq $32, %%mm2               \n\t"
2033                 "movd %%mm2, 4(%%ebx)           \n\t"
2034                 "movd %%mm1, 4(%%ebx, %1)       \n\t"
2035                 "psrlq $32, %%mm1               \n\t"
2036                 "movd %%mm1, 4(%%ebx, %1, 2)    \n\t"
2037
2038         :: "r" (dst), "r" (dstStride), "r" (src)
2039         : "%eax", "%ebx"
2040         );
2041 }
2042 #endif
2043 //static int test=0;
2044
2045 static void inline RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2046                                     uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2047 {
2048 #define FAST_L2_DIFF
2049 //#define L1_DIFF //u should change the thresholds too if u try that one
2050 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2051         asm volatile(
2052                 "leal (%2, %2, 2), %%eax                        \n\t" // 3*stride
2053                 "leal (%2, %2, 4), %%ebx                        \n\t" // 5*stride
2054                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2055 //      0       1       2       3       4       5       6       7       8       9
2056 //      %x      %x+%2   %x+2%2  %x+eax  %x+4%2  %x+ebx  %x+2eax %x+ecx  %x+8%2
2057 //FIXME reorder?
2058 #ifdef L1_DIFF //needs mmx2
2059                 "movq (%0), %%mm0                               \n\t" // L0
2060                 "psadbw (%1), %%mm0                             \n\t" // |L0-R0|
2061                 "movq (%0, %2), %%mm1                           \n\t" // L1
2062                 "psadbw (%1, %2), %%mm1                         \n\t" // |L1-R1|
2063                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2064                 "psadbw (%1, %2, 2), %%mm2                      \n\t" // |L2-R2|
2065                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2066                 "psadbw (%1, %%eax), %%mm3                      \n\t" // |L3-R3|
2067
2068                 "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2069                 "paddw %%mm1, %%mm0                             \n\t"
2070                 "psadbw (%1, %2, 4), %%mm4                      \n\t" // |L4-R4|
2071                 "movq (%0, %%ebx), %%mm5                        \n\t" // L5
2072                 "paddw %%mm2, %%mm0                             \n\t"
2073                 "psadbw (%1, %%ebx), %%mm5                      \n\t" // |L5-R5|
2074                 "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2075                 "paddw %%mm3, %%mm0                             \n\t"
2076                 "psadbw (%1, %%eax, 2), %%mm6                   \n\t" // |L6-R6|
2077                 "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2078                 "paddw %%mm4, %%mm0                             \n\t"
2079                 "psadbw (%1, %%ecx), %%mm7                      \n\t" // |L7-R7|
2080                 "paddw %%mm5, %%mm6                             \n\t"
2081                 "paddw %%mm7, %%mm6                             \n\t"
2082                 "paddw %%mm6, %%mm0                             \n\t"
2083 #elif defined (FAST_L2_DIFF)
2084                 "pcmpeqb %%mm7, %%mm7                           \n\t"
2085                 "movq "MANGLE(b80)", %%mm6                      \n\t"
2086                 "pxor %%mm0, %%mm0                              \n\t"
2087 #define L2_DIFF_CORE(a, b)\
2088                 "movq " #a ", %%mm5                             \n\t"\
2089                 "movq " #b ", %%mm2                             \n\t"\
2090                 "pxor %%mm7, %%mm2                              \n\t"\
2091                 PAVGB(%%mm2, %%mm5)\
2092                 "paddb %%mm6, %%mm5                             \n\t"\
2093                 "movq %%mm5, %%mm2                              \n\t"\
2094                 "psllw $8, %%mm5                                \n\t"\
2095                 "pmaddwd %%mm5, %%mm5                           \n\t"\
2096                 "pmaddwd %%mm2, %%mm2                           \n\t"\
2097                 "paddd %%mm2, %%mm5                             \n\t"\
2098                 "psrld $14, %%mm5                               \n\t"\
2099                 "paddd %%mm5, %%mm0                             \n\t"
2100
2101 L2_DIFF_CORE((%0), (%1))
2102 L2_DIFF_CORE((%0, %2), (%1, %2))
2103 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2104 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2105 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2106 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx))
2107 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2108 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2109
2110 #else
2111                 "pxor %%mm7, %%mm7                              \n\t"
2112                 "pxor %%mm0, %%mm0                              \n\t"
2113 #define L2_DIFF_CORE(a, b)\
2114                 "movq " #a ", %%mm5                             \n\t"\
2115                 "movq " #b ", %%mm2                             \n\t"\
2116                 "movq %%mm5, %%mm1                              \n\t"\
2117                 "movq %%mm2, %%mm3                              \n\t"\
2118                 "punpcklbw %%mm7, %%mm5                         \n\t"\
2119                 "punpckhbw %%mm7, %%mm1                         \n\t"\
2120                 "punpcklbw %%mm7, %%mm2                         \n\t"\
2121                 "punpckhbw %%mm7, %%mm3                         \n\t"\
2122                 "psubw %%mm2, %%mm5                             \n\t"\
2123                 "psubw %%mm3, %%mm1                             \n\t"\
2124                 "pmaddwd %%mm5, %%mm5                           \n\t"\
2125                 "pmaddwd %%mm1, %%mm1                           \n\t"\
2126                 "paddd %%mm1, %%mm5                             \n\t"\
2127                 "paddd %%mm5, %%mm0                             \n\t"
2128
2129 L2_DIFF_CORE((%0), (%1))
2130 L2_DIFF_CORE((%0, %2), (%1, %2))
2131 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2132 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2133 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2134 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx))
2135 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2136 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2137
2138 #endif
2139
2140                 "movq %%mm0, %%mm4                              \n\t"
2141                 "psrlq $32, %%mm0                               \n\t"
2142                 "paddd %%mm0, %%mm4                             \n\t"
2143                 "movd %%mm4, %%ecx                              \n\t"
2144                 "shll $2, %%ecx                                 \n\t"
2145                 "movl %3, %%ebx                                 \n\t"
2146                 "addl -4(%%ebx), %%ecx                          \n\t"
2147                 "addl 4(%%ebx), %%ecx                           \n\t"
2148                 "addl -1024(%%ebx), %%ecx                       \n\t"
2149                 "addl $4, %%ecx                                 \n\t"
2150                 "addl 1024(%%ebx), %%ecx                        \n\t"
2151                 "shrl $3, %%ecx                                 \n\t"
2152                 "movl %%ecx, (%%ebx)                            \n\t"
2153                 "leal (%%eax, %2, 2), %%ebx                     \n\t" // 5*stride
2154
2155 //              "movl %3, %%ecx                                 \n\t"
2156 //              "movl %%ecx, test                               \n\t"
2157 //              "jmp 4f \n\t"
2158                 "cmpl 4+"MANGLE(maxTmpNoise)", %%ecx            \n\t"
2159                 " jb 2f                                         \n\t"
2160                 "cmpl 8+"MANGLE(maxTmpNoise)", %%ecx            \n\t"
2161                 " jb 1f                                         \n\t"
2162
2163                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2164                 "movq (%0), %%mm0                               \n\t" // L0
2165                 "movq (%0, %2), %%mm1                           \n\t" // L1
2166                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2167                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2168                 "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2169                 "movq (%0, %%ebx), %%mm5                        \n\t" // L5
2170                 "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2171                 "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2172                 "movq %%mm0, (%1)                               \n\t" // L0
2173                 "movq %%mm1, (%1, %2)                           \n\t" // L1
2174                 "movq %%mm2, (%1, %2, 2)                        \n\t" // L2
2175                 "movq %%mm3, (%1, %%eax)                        \n\t" // L3
2176                 "movq %%mm4, (%1, %2, 4)                        \n\t" // L4
2177                 "movq %%mm5, (%1, %%ebx)                        \n\t" // L5
2178                 "movq %%mm6, (%1, %%eax, 2)                     \n\t" // L6
2179                 "movq %%mm7, (%1, %%ecx)                        \n\t" // L7
2180                 "jmp 4f                                         \n\t"
2181
2182                 "1:                                             \n\t"
2183                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2184                 "movq (%0), %%mm0                               \n\t" // L0
2185                 "pavgb (%1), %%mm0                              \n\t" // L0
2186                 "movq (%0, %2), %%mm1                           \n\t" // L1
2187                 "pavgb (%1, %2), %%mm1                          \n\t" // L1
2188                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2189                 "pavgb (%1, %2, 2), %%mm2                       \n\t" // L2
2190                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2191                 "pavgb (%1, %%eax), %%mm3                       \n\t" // L3
2192                 "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2193                 "pavgb (%1, %2, 4), %%mm4                       \n\t" // L4
2194                 "movq (%0, %%ebx), %%mm5                        \n\t" // L5
2195                 "pavgb (%1, %%ebx), %%mm5                       \n\t" // L5
2196                 "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2197                 "pavgb (%1, %%eax, 2), %%mm6                    \n\t" // L6
2198                 "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2199                 "pavgb (%1, %%ecx), %%mm7                       \n\t" // L7
2200                 "movq %%mm0, (%1)                               \n\t" // R0
2201                 "movq %%mm1, (%1, %2)                           \n\t" // R1
2202                 "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2203                 "movq %%mm3, (%1, %%eax)                        \n\t" // R3
2204                 "movq %%mm4, (%1, %2, 4)                        \n\t" // R4
2205                 "movq %%mm5, (%1, %%ebx)                        \n\t" // R5
2206                 "movq %%mm6, (%1, %%eax, 2)                     \n\t" // R6
2207                 "movq %%mm7, (%1, %%ecx)                        \n\t" // R7
2208                 "movq %%mm0, (%0)                               \n\t" // L0
2209                 "movq %%mm1, (%0, %2)                           \n\t" // L1
2210                 "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2211                 "movq %%mm3, (%0, %%eax)                        \n\t" // L3
2212                 "movq %%mm4, (%0, %2, 4)                        \n\t" // L4
2213                 "movq %%mm5, (%0, %%ebx)                        \n\t" // L5
2214                 "movq %%mm6, (%0, %%eax, 2)                     \n\t" // L6
2215                 "movq %%mm7, (%0, %%ecx)                        \n\t" // L7
2216                 "jmp 4f                                         \n\t"
2217
2218                 "2:                                             \n\t"
2219                 "cmpl "MANGLE(maxTmpNoise)", %%ecx              \n\t"
2220                 " jb 3f                                         \n\t"
2221
2222                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2223                 "movq (%0), %%mm0                               \n\t" // L0
2224                 "movq (%0, %2), %%mm1                           \n\t" // L1
2225                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2226                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2227                 "movq (%1), %%mm4                               \n\t" // R0
2228                 "movq (%1, %2), %%mm5                           \n\t" // R1
2229                 "movq (%1, %2, 2), %%mm6                        \n\t" // R2
2230                 "movq (%1, %%eax), %%mm7                        \n\t" // R3
2231                 PAVGB(%%mm4, %%mm0)
2232                 PAVGB(%%mm5, %%mm1)
2233                 PAVGB(%%mm6, %%mm2)
2234                 PAVGB(%%mm7, %%mm3)
2235                 PAVGB(%%mm4, %%mm0)
2236                 PAVGB(%%mm5, %%mm1)
2237                 PAVGB(%%mm6, %%mm2)
2238                 PAVGB(%%mm7, %%mm3)
2239                 "movq %%mm0, (%1)                               \n\t" // R0
2240                 "movq %%mm1, (%1, %2)                           \n\t" // R1
2241                 "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2242                 "movq %%mm3, (%1, %%eax)                        \n\t" // R3
2243                 "movq %%mm0, (%0)                               \n\t" // L0
2244                 "movq %%mm1, (%0, %2)                           \n\t" // L1
2245                 "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2246                 "movq %%mm3, (%0, %%eax)                        \n\t" // L3
2247
2248                 "movq (%0, %2, 4), %%mm0                        \n\t" // L4
2249                 "movq (%0, %%ebx), %%mm1                        \n\t" // L5
2250                 "movq (%0, %%eax, 2), %%mm2                     \n\t" // L6
2251                 "movq (%0, %%ecx), %%mm3                        \n\t" // L7
2252                 "movq (%1, %2, 4), %%mm4                        \n\t" // R4
2253                 "movq (%1, %%ebx), %%mm5                        \n\t" // R5
2254                 "movq (%1, %%eax, 2), %%mm6                     \n\t" // R6
2255                 "movq (%1, %%ecx), %%mm7                        \n\t" // R7
2256                 PAVGB(%%mm4, %%mm0)
2257                 PAVGB(%%mm5, %%mm1)
2258                 PAVGB(%%mm6, %%mm2)
2259                 PAVGB(%%mm7, %%mm3)
2260                 PAVGB(%%mm4, %%mm0)
2261                 PAVGB(%%mm5, %%mm1)
2262                 PAVGB(%%mm6, %%mm2)
2263                 PAVGB(%%mm7, %%mm3)
2264                 "movq %%mm0, (%1, %2, 4)                        \n\t" // R4
2265                 "movq %%mm1, (%1, %%ebx)                        \n\t" // R5
2266                 "movq %%mm2, (%1, %%eax, 2)                     \n\t" // R6
2267                 "movq %%mm3, (%1, %%ecx)                        \n\t" // R7
2268                 "movq %%mm0, (%0, %2, 4)                        \n\t" // L4
2269                 "movq %%mm1, (%0, %%ebx)                        \n\t" // L5
2270                 "movq %%mm2, (%0, %%eax, 2)                     \n\t" // L6
2271                 "movq %%mm3, (%0, %%ecx)                        \n\t" // L7
2272                 "jmp 4f                                         \n\t"
2273
2274                 "3:                                             \n\t"
2275                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2276                 "movq (%0), %%mm0                               \n\t" // L0
2277                 "movq (%0, %2), %%mm1                           \n\t" // L1
2278                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2279                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2280                 "movq (%1), %%mm4                               \n\t" // R0
2281                 "movq (%1, %2), %%mm5                           \n\t" // R1
2282                 "movq (%1, %2, 2), %%mm6                        \n\t" // R2
2283                 "movq (%1, %%eax), %%mm7                        \n\t" // R3
2284                 PAVGB(%%mm4, %%mm0)
2285                 PAVGB(%%mm5, %%mm1)
2286                 PAVGB(%%mm6, %%mm2)
2287                 PAVGB(%%mm7, %%mm3)
2288                 PAVGB(%%mm4, %%mm0)
2289                 PAVGB(%%mm5, %%mm1)
2290                 PAVGB(%%mm6, %%mm2)
2291                 PAVGB(%%mm7, %%mm3)
2292                 PAVGB(%%mm4, %%mm0)
2293                 PAVGB(%%mm5, %%mm1)
2294                 PAVGB(%%mm6, %%mm2)
2295                 PAVGB(%%mm7, %%mm3)
2296                 "movq %%mm0, (%1)                               \n\t" // R0
2297                 "movq %%mm1, (%1, %2)                           \n\t" // R1
2298                 "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2299                 "movq %%mm3, (%1, %%eax)                        \n\t" // R3
2300                 "movq %%mm0, (%0)                               \n\t" // L0
2301                 "movq %%mm1, (%0, %2)                           \n\t" // L1
2302                 "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2303                 "movq %%mm3, (%0, %%eax)                        \n\t" // L3
2304
2305                 "movq (%0, %2, 4), %%mm0                        \n\t" // L4
2306                 "movq (%0, %%ebx), %%mm1                        \n\t" // L5
2307                 "movq (%0, %%eax, 2), %%mm2                     \n\t" // L6
2308                 "movq (%0, %%ecx), %%mm3                        \n\t" // L7
2309                 "movq (%1, %2, 4), %%mm4                        \n\t" // R4
2310                 "movq (%1, %%ebx), %%mm5                        \n\t" // R5
2311                 "movq (%1, %%eax, 2), %%mm6                     \n\t" // R6
2312                 "movq (%1, %%ecx), %%mm7                        \n\t" // R7
2313                 PAVGB(%%mm4, %%mm0)
2314                 PAVGB(%%mm5, %%mm1)
2315                 PAVGB(%%mm6, %%mm2)
2316                 PAVGB(%%mm7, %%mm3)
2317                 PAVGB(%%mm4, %%mm0)
2318                 PAVGB(%%mm5, %%mm1)
2319                 PAVGB(%%mm6, %%mm2)
2320                 PAVGB(%%mm7, %%mm3)
2321                 PAVGB(%%mm4, %%mm0)
2322                 PAVGB(%%mm5, %%mm1)
2323                 PAVGB(%%mm6, %%mm2)
2324                 PAVGB(%%mm7, %%mm3)
2325                 "movq %%mm0, (%1, %2, 4)                        \n\t" // R4
2326                 "movq %%mm1, (%1, %%ebx)                        \n\t" // R5
2327                 "movq %%mm2, (%1, %%eax, 2)                     \n\t" // R6
2328                 "movq %%mm3, (%1, %%ecx)                        \n\t" // R7
2329                 "movq %%mm0, (%0, %2, 4)                        \n\t" // L4
2330                 "movq %%mm1, (%0, %%ebx)                        \n\t" // L5
2331                 "movq %%mm2, (%0, %%eax, 2)                     \n\t" // L6
2332                 "movq %%mm3, (%0, %%ecx)                        \n\t" // L7
2333
2334                 "4:                                             \n\t"
2335
2336                 :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast)
2337                 : "%eax", "%ebx", "%ecx", "memory"
2338                 );
2339 //printf("%d\n", test);
2340 #else
2341         int y;
2342         int d=0;
2343         int sysd=0;
2344         int i;
2345
2346         for(y=0; y<8; y++)
2347         {
2348                 int x;
2349                 for(x=0; x<8; x++)
2350                 {
2351                         int ref= tempBlured[ x + y*stride ];
2352                         int cur= src[ x + y*stride ];
2353                         int d1=ref - cur;
2354 //                      if(x==0 || x==7) d1+= d1>>1;
2355 //                      if(y==0 || y==7) d1+= d1>>1;
2356 //                      d+= ABS(d1);
2357                         d+= d1*d1;
2358                         sysd+= d1;
2359                 }
2360         }
2361         i=d;
2362         d=      (
2363                 4*d
2364                 +(*(tempBluredPast-256))
2365                 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
2366                 +(*(tempBluredPast+256))
2367                 +4)>>3;
2368         *tempBluredPast=i;
2369 //      ((*tempBluredPast)*3 + d + 2)>>2;
2370
2371 //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]);
2372 /*
2373 Switch between
2374  1  0  0  0  0  0  0  (0)
2375 64 32 16  8  4  2  1  (1)
2376 64 48 36 27 20 15 11 (33) (approx)
2377 64 56 49 43 37 33 29 (200) (approx)
2378 */
2379         if(d > maxNoise[1])
2380         {
2381                 if(d < maxNoise[2])
2382                 {
2383                         for(y=0; y<8; y++)
2384                         {
2385                                 int x;
2386                                 for(x=0; x<8; x++)
2387                                 {
2388                                         int ref= tempBlured[ x + y*stride ];
2389                                         int cur= src[ x + y*stride ];
2390                                         tempBlured[ x + y*stride ]=
2391                                         src[ x + y*stride ]=
2392                                                 (ref + cur + 1)>>1;
2393                                 }
2394                         }
2395                 }
2396                 else
2397                 {
2398                         for(y=0; y<8; y++)
2399                         {
2400                                 int x;
2401                                 for(x=0; x<8; x++)
2402                                 {
2403                                         tempBlured[ x + y*stride ]= src[ x + y*stride ];
2404                                 }
2405                         }
2406                 }
2407         }
2408         else
2409         {
2410                 if(d < maxNoise[0])
2411                 {
2412                         for(y=0; y<8; y++)
2413                         {
2414                                 int x;
2415                                 for(x=0; x<8; x++)
2416                                 {
2417                                         int ref= tempBlured[ x + y*stride ];
2418                                         int cur= src[ x + y*stride ];
2419                                         tempBlured[ x + y*stride ]=
2420                                         src[ x + y*stride ]=
2421                                                 (ref*7 + cur + 4)>>3;
2422                                 }
2423                         }
2424                 }
2425                 else
2426                 {
2427                         for(y=0; y<8; y++)
2428                         {
2429                                 int x;
2430                                 for(x=0; x<8; x++)
2431                                 {
2432                                         int ref= tempBlured[ x + y*stride ];
2433                                         int cur= src[ x + y*stride ];
2434                                         tempBlured[ x + y*stride ]=
2435                                         src[ x + y*stride ]=
2436                                                 (ref*3 + cur + 2)>>2;
2437                                 }
2438                         }
2439                 }
2440         }
2441 #endif
2442 }
2443
2444 static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2445         QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
2446
2447 /**
2448  * Copies a block from src to dst and fixes the blacklevel
2449  * numLines must be a multiple of 4
2450  * levelFix == 0 -> dont touch the brighness & contrast
2451  */
2452 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
2453         int levelFix)
2454 {
2455 #ifndef HAVE_MMX
2456         int i;
2457 #endif
2458         if(levelFix)
2459         {
2460 #ifdef HAVE_MMX
2461                                         asm volatile(
2462                                                 "leal (%0,%2), %%eax    \n\t"
2463                                                 "leal (%1,%3), %%ebx    \n\t"
2464                                                 "movq "MANGLE(packedYOffset)", %%mm2\n\t"
2465                                                 "movq "MANGLE(packedYScale)", %%mm3\n\t"
2466                                                 "pxor %%mm4, %%mm4      \n\t"
2467 #ifdef HAVE_MMX2
2468 #define SCALED_CPY(src1, src2, dst1, dst2)                                      \
2469                                                 "movq " #src1 ", %%mm0  \n\t"\
2470                                                 "movq " #src1 ", %%mm5  \n\t"\
2471                                                 "movq " #src2 ", %%mm1  \n\t"\
2472                                                 "movq " #src2 ", %%mm6  \n\t"\
2473                                                 "punpcklbw %%mm0, %%mm0 \n\t"\
2474                                                 "punpckhbw %%mm5, %%mm5 \n\t"\
2475                                                 "punpcklbw %%mm1, %%mm1 \n\t"\
2476                                                 "punpckhbw %%mm6, %%mm6 \n\t"\
2477                                                 "pmulhuw %%mm3, %%mm0   \n\t"\
2478                                                 "pmulhuw %%mm3, %%mm5   \n\t"\
2479                                                 "pmulhuw %%mm3, %%mm1   \n\t"\
2480                                                 "pmulhuw %%mm3, %%mm6   \n\t"\
2481                                                 "psubw %%mm2, %%mm0     \n\t"\
2482                                                 "psubw %%mm2, %%mm5     \n\t"\
2483                                                 "psubw %%mm2, %%mm1     \n\t"\
2484                                                 "psubw %%mm2, %%mm6     \n\t"\
2485                                                 "packuswb %%mm5, %%mm0  \n\t"\
2486                                                 "packuswb %%mm6, %%mm1  \n\t"\
2487                                                 "movq %%mm0, " #dst1 "  \n\t"\
2488                                                 "movq %%mm1, " #dst2 "  \n\t"\
2489
2490 #else //HAVE_MMX2
2491 #define SCALED_CPY(src1, src2, dst1, dst2)                                      \
2492                                                 "movq " #src1 ", %%mm0  \n\t"\
2493                                                 "movq " #src1 ", %%mm5  \n\t"\
2494                                                 "punpcklbw %%mm4, %%mm0 \n\t"\
2495                                                 "punpckhbw %%mm4, %%mm5 \n\t"\
2496                                                 "psubw %%mm2, %%mm0     \n\t"\
2497                                                 "psubw %%mm2, %%mm5     \n\t"\
2498                                                 "movq " #src2 ", %%mm1  \n\t"\
2499                                                 "psllw $6, %%mm0        \n\t"\
2500                                                 "psllw $6, %%mm5        \n\t"\
2501                                                 "pmulhw %%mm3, %%mm0    \n\t"\
2502                                                 "movq " #src2 ", %%mm6  \n\t"\
2503                                                 "pmulhw %%mm3, %%mm5    \n\t"\
2504                                                 "punpcklbw %%mm4, %%mm1 \n\t"\
2505                                                 "punpckhbw %%mm4, %%mm6 \n\t"\
2506                                                 "psubw %%mm2, %%mm1     \n\t"\
2507                                                 "psubw %%mm2, %%mm6     \n\t"\
2508                                                 "psllw $6, %%mm1        \n\t"\
2509                                                 "psllw $6, %%mm6        \n\t"\
2510                                                 "pmulhw %%mm3, %%mm1    \n\t"\
2511                                                 "pmulhw %%mm3, %%mm6    \n\t"\
2512                                                 "packuswb %%mm5, %%mm0  \n\t"\
2513                                                 "packuswb %%mm6, %%mm1  \n\t"\
2514                                                 "movq %%mm0, " #dst1 "  \n\t"\
2515                                                 "movq %%mm1, " #dst2 "  \n\t"\
2516
2517 #endif //!HAVE_MMX2
2518
2519 SCALED_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
2520 SCALED_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2))
2521 SCALED_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4))
2522                                                 "leal (%%eax,%2,4), %%eax       \n\t"
2523                                                 "leal (%%ebx,%3,4), %%ebx       \n\t"
2524 SCALED_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2))
2525
2526
2527                                                 : : "r"(src),
2528                                                 "r"(dst),
2529                                                 "r" (srcStride),
2530                                                 "r" (dstStride)
2531                                                 : "%eax", "%ebx"
2532                                         );
2533 #else
2534                                 for(i=0; i<8; i++)
2535                                         memcpy( &(dst[dstStride*i]),
2536                                                 &(src[srcStride*i]), BLOCK_SIZE);
2537 #endif
2538         }
2539         else
2540         {
2541 #ifdef HAVE_MMX
2542                                         asm volatile(
2543                                                 "leal (%0,%2), %%eax    \n\t"
2544                                                 "leal (%1,%3), %%ebx    \n\t"
2545
2546 #define SIMPLE_CPY(src1, src2, dst1, dst2)                              \
2547                                                 "movq " #src1 ", %%mm0  \n\t"\
2548                                                 "movq " #src2 ", %%mm1  \n\t"\
2549                                                 "movq %%mm0, " #dst1 "  \n\t"\
2550                                                 "movq %%mm1, " #dst2 "  \n\t"\
2551
2552 SIMPLE_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
2553 SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2))
2554 SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4))
2555                                                 "leal (%%eax,%2,4), %%eax       \n\t"
2556                                                 "leal (%%ebx,%3,4), %%ebx       \n\t"
2557 SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2))
2558
2559                                                 : : "r" (src),
2560                                                 "r" (dst),
2561                                                 "r" (srcStride),
2562                                                 "r" (dstStride)
2563                                                 : "%eax", "%ebx"
2564                                         );
2565 #else
2566                                 for(i=0; i<8; i++)
2567                                         memcpy( &(dst[dstStride*i]),
2568                                                 &(src[srcStride*i]), BLOCK_SIZE);
2569 #endif
2570         }
2571 }
2572
2573
2574 /**
2575  * Filters array of bytes (Y or U or V values)
2576  */
2577 static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2578         QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode)
2579 {
2580         int x,y;
2581 #ifdef COMPILE_TIME_MODE
2582         const int mode= COMPILE_TIME_MODE;
2583 #else
2584         const int mode= isColor ? ppMode->chromMode : ppMode->lumMode;
2585 #endif
2586         /* we need 64bit here otherwise we´ll going to have a problem
2587            after watching a black picture for 5 hours*/
2588         static uint64_t *yHistogram= NULL;
2589         int black=0, white=255; // blackest black and whitest white in the picture
2590         int QPCorrecture= 256;
2591
2592         /* Temporary buffers for handling the last row(s) */
2593         static uint8_t *tempDst= NULL;
2594         static uint8_t *tempSrc= NULL;
2595
2596         /* Temporary buffers for handling the last block */
2597         static uint8_t *tempDstBlock= NULL;
2598         static uint8_t *tempSrcBlock= NULL;
2599
2600         /* Temporal noise reducing buffers */
2601         static uint8_t *tempBlured[3]= {NULL,NULL,NULL};
2602         static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL};
2603
2604         int copyAhead;
2605
2606 #ifdef PP_FUNNY_STRIDE
2607         uint8_t *dstBlockPtrBackup;
2608         uint8_t *srcBlockPtrBackup;
2609 #endif
2610
2611 #ifdef MORE_TIMING
2612         long long T0, T1, diffTime=0;
2613 #endif
2614 #ifdef TIMING
2615         long long memcpyTime=0, vertTime=0, horizTime=0, sumTime;
2616         sumTime= rdtsc();
2617 #endif
2618
2619         dcOffset= ppMode->maxDcDiff;
2620         dcThreshold= ppMode->maxDcDiff*2 + 1;
2621
2622 #ifdef HAVE_MMX
2623         maxTmpNoise[0]= ppMode->maxTmpNoise[0];
2624         maxTmpNoise[1]= ppMode->maxTmpNoise[1];
2625         maxTmpNoise[2]= ppMode->maxTmpNoise[2];
2626         
2627         mmxDCOffset= 0x7F - dcOffset;
2628         mmxDCThreshold= 0x7F - dcThreshold;
2629
2630         mmxDCOffset*= 0x0101010101010101LL;
2631         mmxDCThreshold*= 0x0101010101010101LL;
2632 #endif
2633
2634         if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
2635         else if(mode & LINEAR_BLEND_DEINT_FILTER) copyAhead=14;
2636         else if(   (mode & V_DEBLOCK)
2637                 || (mode & LINEAR_IPOL_DEINT_FILTER)
2638                 || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
2639         else if(mode & V_X1_FILTER) copyAhead=11;
2640         else if(mode & V_RK1_FILTER) copyAhead=10;
2641         else if(mode & DERING) copyAhead=9;
2642         else copyAhead=8;
2643
2644         copyAhead-= 8;
2645
2646         if(tempDst==NULL)
2647         {
2648                 tempDst= (uint8_t*)memalign(8, 1024*24);
2649                 tempSrc= (uint8_t*)memalign(8, 1024*24);
2650                 tempDstBlock= (uint8_t*)memalign(8, 1024*24);
2651                 tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
2652         }
2653
2654         if(tempBlured[isColor]==NULL && (mode & TEMP_NOISE_FILTER))
2655         {
2656 //              printf("%d %d %d\n", isColor, dstStride, height);
2657                 //FIXME works only as long as the size doesnt increase
2658                 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end
2659                 tempBlured[isColor]= (uint8_t*)memalign(8, dstStride*((height+7)&(~7)) + 17*1024);
2660                 tempBluredPast[isColor]= (uint32_t*)memalign(8, 256*((height+7)&(~7))/2 + 17*1024);
2661
2662                 memset(tempBlured[isColor], 0, dstStride*((height+7)&(~7)) + 17*1024);
2663                 memset(tempBluredPast[isColor], 0, 256*((height+7)&(~7))/2 + 17*1024);
2664         }
2665
2666         if(!yHistogram)
2667         {
2668                 int i;
2669                 yHistogram= (uint64_t*)malloc(8*256);
2670                 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;
2671
2672                 if(mode & FULL_Y_RANGE)
2673                 {
2674                         ppMode->maxAllowedY=255;
2675                         ppMode->minAllowedY=0;
2676                 }
2677         }
2678
2679         if(!isColor)
2680         {
2681                 uint64_t sum= 0;
2682                 int i;
2683                 static int framenum= -1;
2684                 uint64_t maxClipped;
2685                 uint64_t clipped;
2686                 double scale;
2687
2688                 framenum++;
2689                 if(framenum == 1) yHistogram[0]= width*height/64*15/256;
2690
2691                 for(i=0; i<256; i++)
2692                 {
2693                         sum+= yHistogram[i];
2694 //                      printf("%d ", yHistogram[i]);
2695                 }
2696 //              printf("\n\n");
2697
2698                 /* we allways get a completly black picture first */
2699                 maxClipped= (uint64_t)(sum * maxClippedThreshold);
2700
2701                 clipped= sum;
2702                 for(black=255; black>0; black--)
2703                 {
2704                         if(clipped < maxClipped) break;
2705                         clipped-= yHistogram[black];
2706                 }
2707
2708                 clipped= sum;
2709                 for(white=0; white<256; white++)
2710                 {
2711                         if(clipped < maxClipped) break;
2712                         clipped-= yHistogram[white];
2713                 }
2714
2715                 scale= (double)(ppMode->maxAllowedY - ppMode->minAllowedY) / (double)(white-black);
2716
2717 #ifdef HAVE_MMX2
2718                 packedYScale= (uint16_t)(scale*256.0 + 0.5);
2719                 packedYOffset= (((black*packedYScale)>>8) - ppMode->minAllowedY) & 0xFFFF;
2720 #else
2721                 packedYScale= (uint16_t)(scale*1024.0 + 0.5);
2722                 packedYOffset= (black - ppMode->minAllowedY) & 0xFFFF;
2723 #endif
2724
2725                 packedYOffset|= packedYOffset<<32;
2726                 packedYOffset|= packedYOffset<<16;
2727
2728                 packedYScale|= packedYScale<<32;
2729                 packedYScale|= packedYScale<<16;
2730         }
2731         else
2732         {
2733                 packedYScale= 0x0100010001000100LL;
2734                 packedYOffset= 0;
2735         }
2736
2737         if(mode & LEVEL_FIX)    QPCorrecture= packedYScale &0xFFFF;
2738         else                    QPCorrecture= 256;
2739
2740         /* copy & deinterlace first row of blocks */
2741         y=-BLOCK_SIZE;
2742         {
2743                 //1% speedup if these are here instead of the inner loop
2744                 uint8_t *srcBlock= &(src[y*srcStride]);
2745                 uint8_t *dstBlock= &(dst[y*dstStride]);
2746
2747                 dstBlock= tempDst + dstStride;
2748
2749                 // From this point on it is guranteed that we can read and write 16 lines downward
2750                 // finish 1 block before the next otherwise we´ll might have a problem
2751                 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
2752                 for(x=0; x<width; x+=BLOCK_SIZE)
2753                 {
2754
2755 #ifdef HAVE_MMX2
2756 /*
2757                         prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
2758                         prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
2759                         prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
2760                         prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
2761 */
2762
2763                         asm(
2764                                 "movl %4, %%eax                 \n\t"
2765                                 "shrl $2, %%eax                 \n\t"
2766                                 "andl $6, %%eax                 \n\t"
2767                                 "addl %5, %%eax                 \n\t"
2768                                 "movl %%eax, %%ebx              \n\t"
2769                                 "imul %1, %%eax                 \n\t"
2770                                 "imul %3, %%ebx                 \n\t"
2771                                 "prefetchnta 32(%%eax, %0)      \n\t"
2772                                 "prefetcht0 32(%%ebx, %2)       \n\t"
2773                                 "addl %1, %%eax                 \n\t"
2774                                 "addl %3, %%ebx                 \n\t"
2775                                 "prefetchnta 32(%%eax, %0)      \n\t"
2776                                 "prefetcht0 32(%%ebx, %2)       \n\t"
2777                         :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
2778                         "m" (x), "m" (copyAhead)
2779                         : "%eax", "%ebx"
2780                         );
2781
2782 #elif defined(HAVE_3DNOW)
2783 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
2784 /*                      prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
2785                         prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
2786                         prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
2787                         prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
2788 */
2789 #endif
2790
2791                         RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
2792                                 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX);
2793
2794                         if(mode & LINEAR_IPOL_DEINT_FILTER)
2795                                 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
2796                         else if(mode & LINEAR_BLEND_DEINT_FILTER)
2797                                 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
2798                         else if(mode & MEDIAN_DEINT_FILTER)
2799                                 RENAME(deInterlaceMedian)(dstBlock, dstStride);
2800                         else if(mode & CUBIC_IPOL_DEINT_FILTER)
2801                                 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
2802 /*                      else if(mode & CUBIC_BLEND_DEINT_FILTER)
2803                                 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
2804 */