license header consistency cosmetics
[libav.git] / libpostproc / postprocess_template.c
1 /*
2 * Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at)
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 /**
22 * @file postprocess_template.c
23 * mmx/mmx2/3dnow postprocess code.
24 */
25
26
27 #ifdef ARCH_X86_64
28 # define REGa rax
29 # define REGc rcx
30 # define REGd rdx
31 # define REG_a "rax"
32 # define REG_c "rcx"
33 # define REG_d "rdx"
34 # define REG_SP "rsp"
35 # define ALIGN_MASK "$0xFFFFFFFFFFFFFFF8"
36 #else
37 # define REGa eax
38 # define REGc ecx
39 # define REGd edx
40 # define REG_a "eax"
41 # define REG_c "ecx"
42 # define REG_d "edx"
43 # define REG_SP "esp"
44 # define ALIGN_MASK "$0xFFFFFFF8"
45 #endif
46
47
48 #undef PAVGB
49 #undef PMINUB
50 #undef PMAXUB
51
52 #ifdef HAVE_MMX2
53 #define REAL_PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
54 #elif defined (HAVE_3DNOW)
55 #define REAL_PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
56 #endif
57 #define PAVGB(a,b) REAL_PAVGB(a,b)
58
59 #ifdef HAVE_MMX2
60 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
61 #elif defined (HAVE_MMX)
62 #define PMINUB(b,a,t) \
63 "movq " #a ", " #t " \n\t"\
64 "psubusb " #b ", " #t " \n\t"\
65 "psubb " #t ", " #a " \n\t"
66 #endif
67
68 #ifdef HAVE_MMX2
69 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
70 #elif defined (HAVE_MMX)
71 #define PMAXUB(a,b) \
72 "psubusb " #a ", " #b " \n\t"\
73 "paddb " #a ", " #b " \n\t"
74 #endif
75
76 //FIXME? |255-0| = 1 (should not be a problem ...)
77 #ifdef HAVE_MMX
78 /**
79 * Check if the middle 8x8 Block in the given 8x16 block is flat
80 */
81 static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){
82 int numEq= 0, dcOk;
83 src+= stride*4; // src points to begin of the 8x8 Block
84 asm volatile(
85 "movq %0, %%mm7 \n\t"
86 "movq %1, %%mm6 \n\t"
87 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
88 );
89
90 asm volatile(
91 "lea (%2, %3), %%"REG_a" \n\t"
92 // 0 1 2 3 4 5 6 7 8 9
93 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
94
95 "movq (%2), %%mm0 \n\t"
96 "movq (%%"REG_a"), %%mm1 \n\t"
97 "movq %%mm0, %%mm3 \n\t"
98 "movq %%mm0, %%mm4 \n\t"
99 PMAXUB(%%mm1, %%mm4)
100 PMINUB(%%mm1, %%mm3, %%mm5)
101 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
102 "paddb %%mm7, %%mm0 \n\t"
103 "pcmpgtb %%mm6, %%mm0 \n\t"
104
105 "movq (%%"REG_a",%3), %%mm2 \n\t"
106 PMAXUB(%%mm2, %%mm4)
107 PMINUB(%%mm2, %%mm3, %%mm5)
108 "psubb %%mm2, %%mm1 \n\t"
109 "paddb %%mm7, %%mm1 \n\t"
110 "pcmpgtb %%mm6, %%mm1 \n\t"
111 "paddb %%mm1, %%mm0 \n\t"
112
113 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
114 PMAXUB(%%mm1, %%mm4)
115 PMINUB(%%mm1, %%mm3, %%mm5)
116 "psubb %%mm1, %%mm2 \n\t"
117 "paddb %%mm7, %%mm2 \n\t"
118 "pcmpgtb %%mm6, %%mm2 \n\t"
119 "paddb %%mm2, %%mm0 \n\t"
120
121 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
122
123 "movq (%2, %3, 4), %%mm2 \n\t"
124 PMAXUB(%%mm2, %%mm4)
125 PMINUB(%%mm2, %%mm3, %%mm5)
126 "psubb %%mm2, %%mm1 \n\t"
127 "paddb %%mm7, %%mm1 \n\t"
128 "pcmpgtb %%mm6, %%mm1 \n\t"
129 "paddb %%mm1, %%mm0 \n\t"
130
131 "movq (%%"REG_a"), %%mm1 \n\t"
132 PMAXUB(%%mm1, %%mm4)
133 PMINUB(%%mm1, %%mm3, %%mm5)
134 "psubb %%mm1, %%mm2 \n\t"
135 "paddb %%mm7, %%mm2 \n\t"
136 "pcmpgtb %%mm6, %%mm2 \n\t"
137 "paddb %%mm2, %%mm0 \n\t"
138
139 "movq (%%"REG_a", %3), %%mm2 \n\t"
140 PMAXUB(%%mm2, %%mm4)
141 PMINUB(%%mm2, %%mm3, %%mm5)
142 "psubb %%mm2, %%mm1 \n\t"
143 "paddb %%mm7, %%mm1 \n\t"
144 "pcmpgtb %%mm6, %%mm1 \n\t"
145 "paddb %%mm1, %%mm0 \n\t"
146
147 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
148 PMAXUB(%%mm1, %%mm4)
149 PMINUB(%%mm1, %%mm3, %%mm5)
150 "psubb %%mm1, %%mm2 \n\t"
151 "paddb %%mm7, %%mm2 \n\t"
152 "pcmpgtb %%mm6, %%mm2 \n\t"
153 "paddb %%mm2, %%mm0 \n\t"
154 "psubusb %%mm3, %%mm4 \n\t"
155
156 " \n\t"
157 #ifdef HAVE_MMX2
158 "pxor %%mm7, %%mm7 \n\t"
159 "psadbw %%mm7, %%mm0 \n\t"
160 #else
161 "movq %%mm0, %%mm1 \n\t"
162 "psrlw $8, %%mm0 \n\t"
163 "paddb %%mm1, %%mm0 \n\t"
164 "movq %%mm0, %%mm1 \n\t"
165 "psrlq $16, %%mm0 \n\t"
166 "paddb %%mm1, %%mm0 \n\t"
167 "movq %%mm0, %%mm1 \n\t"
168 "psrlq $32, %%mm0 \n\t"
169 "paddb %%mm1, %%mm0 \n\t"
170 #endif
171 "movq %4, %%mm7 \n\t" // QP,..., QP
172 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
173 "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0
174 "packssdw %%mm4, %%mm4 \n\t"
175 "movd %%mm0, %0 \n\t"
176 "movd %%mm4, %1 \n\t"
177
178 : "=r" (numEq), "=r" (dcOk)
179 : "r" (src), "r" ((long)stride), "m" (c->pQPb)
180 : "%"REG_a
181 );
182
183 numEq= (-numEq) &0xFF;
184 if(numEq > c->ppMode.flatnessThreshold){
185 if(dcOk) return 0;
186 else return 1;
187 }else{
188 return 2;
189 }
190 }
191 #endif //HAVE_MMX
192
193 /**
194 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
195 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
196 */
197 #ifndef HAVE_ALTIVEC
198 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
199 {
200 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
201 src+= stride*3;
202 asm volatile( //"movv %0 %1 %2\n\t"
203 "movq %2, %%mm0 \n\t" // QP,..., QP
204 "pxor %%mm4, %%mm4 \n\t"
205
206 "movq (%0), %%mm6 \n\t"
207 "movq (%0, %1), %%mm5 \n\t"
208 "movq %%mm5, %%mm1 \n\t"
209 "movq %%mm6, %%mm2 \n\t"
210 "psubusb %%mm6, %%mm5 \n\t"
211 "psubusb %%mm1, %%mm2 \n\t"
212 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
213 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
214 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
215
216 "pand %%mm2, %%mm6 \n\t"
217 "pandn %%mm1, %%mm2 \n\t"
218 "por %%mm2, %%mm6 \n\t"// First Line to Filter
219
220 "movq (%0, %1, 8), %%mm5 \n\t"
221 "lea (%0, %1, 4), %%"REG_a" \n\t"
222 "lea (%0, %1, 8), %%"REG_c" \n\t"
223 "sub %1, %%"REG_c" \n\t"
224 "add %1, %0 \n\t" // %0 points to line 1 not 0
225 "movq (%0, %1, 8), %%mm7 \n\t"
226 "movq %%mm5, %%mm1 \n\t"
227 "movq %%mm7, %%mm2 \n\t"
228 "psubusb %%mm7, %%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 %%mm4, %%mm2 \n\t" // diff <= QP -> FF
233
234 "pand %%mm2, %%mm7 \n\t"
235 "pandn %%mm1, %%mm2 \n\t"
236 "por %%mm2, %%mm7 \n\t" // First Line to Filter
237
238
239 // 1 2 3 4 5 6 7 8
240 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
241 // 6 4 2 2 1 1
242 // 6 4 4 2
243 // 6 8 2
244
245 "movq (%0, %1), %%mm0 \n\t" // 1
246 "movq %%mm0, %%mm1 \n\t" // 1
247 PAVGB(%%mm6, %%mm0) //1 1 /2
248 PAVGB(%%mm6, %%mm0) //3 1 /4
249
250 "movq (%0, %1, 4), %%mm2 \n\t" // 1
251 "movq %%mm2, %%mm5 \n\t" // 1
252 PAVGB((%%REGa), %%mm2) // 11 /2
253 PAVGB((%0, %1, 2), %%mm2) // 211 /4
254 "movq %%mm2, %%mm3 \n\t" // 211 /4
255 "movq (%0), %%mm4 \n\t" // 1
256 PAVGB(%%mm4, %%mm3) // 4 211 /8
257 PAVGB(%%mm0, %%mm3) //642211 /16
258 "movq %%mm3, (%0) \n\t" // X
259 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
260 "movq %%mm1, %%mm0 \n\t" // 1
261 PAVGB(%%mm6, %%mm0) //1 1 /2
262 "movq %%mm4, %%mm3 \n\t" // 1
263 PAVGB((%0,%1,2), %%mm3) // 1 1 /2
264 PAVGB((%%REGa,%1,2), %%mm5) // 11 /2
265 PAVGB((%%REGa), %%mm5) // 211 /4
266 PAVGB(%%mm5, %%mm3) // 2 2211 /8
267 PAVGB(%%mm0, %%mm3) //4242211 /16
268 "movq %%mm3, (%0,%1) \n\t" // X
269 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
270 PAVGB(%%mm4, %%mm6) //11 /2
271 "movq (%%"REG_c"), %%mm0 \n\t" // 1
272 PAVGB((%%REGa, %1, 2), %%mm0) // 11/2
273 "movq %%mm0, %%mm3 \n\t" // 11/2
274 PAVGB(%%mm1, %%mm0) // 2 11/4
275 PAVGB(%%mm6, %%mm0) //222 11/8
276 PAVGB(%%mm2, %%mm0) //22242211/16
277 "movq (%0, %1, 2), %%mm2 \n\t" // 1
278 "movq %%mm0, (%0, %1, 2) \n\t" // X
279 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
280 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
281 PAVGB((%%REGc), %%mm0) // 11 /2
282 PAVGB(%%mm0, %%mm6) //11 11 /4
283 PAVGB(%%mm1, %%mm4) // 11 /2
284 PAVGB(%%mm2, %%mm1) // 11 /2
285 PAVGB(%%mm1, %%mm6) //1122 11 /8
286 PAVGB(%%mm5, %%mm6) //112242211 /16
287 "movq (%%"REG_a"), %%mm5 \n\t" // 1
288 "movq %%mm6, (%%"REG_a") \n\t" // X
289 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
290 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1
291 PAVGB(%%mm7, %%mm6) // 11 /2
292 PAVGB(%%mm4, %%mm6) // 11 11 /4
293 PAVGB(%%mm3, %%mm6) // 11 2211 /8
294 PAVGB(%%mm5, %%mm2) // 11 /2
295 "movq (%0, %1, 4), %%mm4 \n\t" // 1
296 PAVGB(%%mm4, %%mm2) // 112 /4
297 PAVGB(%%mm2, %%mm6) // 112242211 /16
298 "movq %%mm6, (%0, %1, 4) \n\t" // X
299 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
300 PAVGB(%%mm7, %%mm1) // 11 2 /4
301 PAVGB(%%mm4, %%mm5) // 11 /2
302 PAVGB(%%mm5, %%mm0) // 11 11 /4
303 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1
304 PAVGB(%%mm6, %%mm1) // 11 4 2 /8
305 PAVGB(%%mm0, %%mm1) // 11224222 /16
306 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X
307 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
308 PAVGB((%%REGc), %%mm2) // 112 4 /8
309 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
310 PAVGB(%%mm0, %%mm6) // 1 1 /2
311 PAVGB(%%mm7, %%mm6) // 1 12 /4
312 PAVGB(%%mm2, %%mm6) // 1122424 /4
313 "movq %%mm6, (%%"REG_c") \n\t" // X
314 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
315 PAVGB(%%mm7, %%mm5) // 11 2 /4
316 PAVGB(%%mm7, %%mm5) // 11 6 /8
317
318 PAVGB(%%mm3, %%mm0) // 112 /4
319 PAVGB(%%mm0, %%mm5) // 112246 /16
320 "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X
321 "sub %1, %0 \n\t"
322
323 :
324 : "r" (src), "r" ((long)stride), "m" (c->pQPb)
325 : "%"REG_a, "%"REG_c
326 );
327 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
328 const int l1= stride;
329 const int l2= stride + l1;
330 const int l3= stride + l2;
331 const int l4= stride + l3;
332 const int l5= stride + l4;
333 const int l6= stride + l5;
334 const int l7= stride + l6;
335 const int l8= stride + l7;
336 const int l9= stride + l8;
337 int x;
338 src+= stride*3;
339 for(x=0; x<BLOCK_SIZE; x++)
340 {
341 const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
342 const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
343
344 int sums[10];
345 sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
346 sums[1] = sums[0] - first + src[l4];
347 sums[2] = sums[1] - first + src[l5];
348 sums[3] = sums[2] - first + src[l6];
349 sums[4] = sums[3] - first + src[l7];
350 sums[5] = sums[4] - src[l1] + src[l8];
351 sums[6] = sums[5] - src[l2] + last;
352 sums[7] = sums[6] - src[l3] + last;
353 sums[8] = sums[7] - src[l4] + last;
354 sums[9] = sums[8] - src[l5] + last;
355
356 src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
357 src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
358 src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
359 src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
360 src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
361 src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
362 src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
363 src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
364
365 src++;
366 }
367 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
368 }
369 #endif //HAVE_ALTIVEC
370
371 #if 0
372 /**
373 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
374 * values are correctly clipped (MMX2)
375 * values are wraparound (C)
376 * Conclusion: It is fast, but introduces ugly horizontal patterns
377 * if there is a continuous gradient.
378 0 8 16 24
379 x = 8
380 x/2 = 4
381 x/8 = 1
382 1 12 12 23
383 */
384 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
385 {
386 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
387 src+= stride*3;
388 // FIXME rounding
389 asm volatile(
390 "pxor %%mm7, %%mm7 \n\t" // 0
391 "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
392 "leal (%0, %1), %%"REG_a" \n\t"
393 "leal (%%"REG_a", %1, 4), %%"REG_c" \n\t"
394 // 0 1 2 3 4 5 6 7 8 9
395 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
396 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
397 "movq %%mm0, %%mm1 \n\t" // QP,..., QP
398 "paddusb "MANGLE(b02)", %%mm0 \n\t"
399 "psrlw $2, %%mm0 \n\t"
400 "pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4
401 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
402 "movq (%0, %1, 4), %%mm2 \n\t" // line 4
403 "movq (%%"REG_c"), %%mm3 \n\t" // line 5
404 "movq %%mm2, %%mm4 \n\t" // line 4
405 "pcmpeqb %%mm5, %%mm5 \n\t" // -1
406 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1
407 PAVGB(%%mm3, %%mm5)
408 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2
409 "psubusb %%mm3, %%mm4 \n\t"
410 "psubusb %%mm2, %%mm3 \n\t"
411 "por %%mm3, %%mm4 \n\t" // |l4 - l5|
412 "psubusb %%mm0, %%mm4 \n\t"
413 "pcmpeqb %%mm7, %%mm4 \n\t"
414 "pand %%mm4, %%mm5 \n\t" // d/2
415
416 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80
417 "paddb %%mm5, %%mm2 \n\t"
418 // "psubb %%mm6, %%mm2 \n\t"
419 "movq %%mm2, (%0,%1, 4) \n\t"
420
421 "movq (%%"REG_c"), %%mm2 \n\t"
422 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80
423 "psubb %%mm5, %%mm2 \n\t"
424 // "psubb %%mm6, %%mm2 \n\t"
425 "movq %%mm2, (%%"REG_c") \n\t"
426
427 "paddb %%mm6, %%mm5 \n\t"
428 "psrlw $2, %%mm5 \n\t"
429 "pand "MANGLE(b3F)", %%mm5 \n\t"
430 "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
431
432 "movq (%%"REG_a", %1, 2), %%mm2 \n\t"
433 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
434 "paddsb %%mm5, %%mm2 \n\t"
435 "psubb %%mm6, %%mm2 \n\t"
436 "movq %%mm2, (%%"REG_a", %1, 2) \n\t"
437
438 "movq (%%"REG_c", %1), %%mm2 \n\t"
439 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80
440 "psubsb %%mm5, %%mm2 \n\t"
441 "psubb %%mm6, %%mm2 \n\t"
442 "movq %%mm2, (%%"REG_c", %1) \n\t"
443
444 :
445 : "r" (src), "r" ((long)stride)
446 : "%"REG_a, "%"REG_c
447 );
448 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
449 const int l1= stride;
450 const int l2= stride + l1;
451 const int l3= stride + l2;
452 const int l4= stride + l3;
453 const int l5= stride + l4;
454 const int l6= stride + l5;
455 // const int l7= stride + l6;
456 // const int l8= stride + l7;
457 // const int l9= stride + l8;
458 int x;
459 const int QP15= QP + (QP>>2);
460 src+= stride*3;
461 for(x=0; x<BLOCK_SIZE; x++)
462 {
463 const int v = (src[x+l5] - src[x+l4]);
464 if(FFABS(v) < QP15)
465 {
466 src[x+l3] +=v>>3;
467 src[x+l4] +=v>>1;
468 src[x+l5] -=v>>1;
469 src[x+l6] -=v>>3;
470
471 }
472 }
473
474 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
475 }
476 #endif //0
477
478 /**
479 * Experimental Filter 1
480 * will not damage linear gradients
481 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
482 * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
483 * MMX2 version does correct clipping C version does not
484 */
485 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
486 {
487 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
488 src+= stride*3;
489
490 asm volatile(
491 "pxor %%mm7, %%mm7 \n\t" // 0
492 "lea (%0, %1), %%"REG_a" \n\t"
493 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
494 // 0 1 2 3 4 5 6 7 8 9
495 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
496 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
497 "movq (%0, %1, 4), %%mm1 \n\t" // line 4
498 "movq %%mm1, %%mm2 \n\t" // line 4
499 "psubusb %%mm0, %%mm1 \n\t"
500 "psubusb %%mm2, %%mm0 \n\t"
501 "por %%mm1, %%mm0 \n\t" // |l2 - l3|
502 "movq (%%"REG_c"), %%mm3 \n\t" // line 5
503 "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6
504 "movq %%mm3, %%mm5 \n\t" // line 5
505 "psubusb %%mm4, %%mm3 \n\t"
506 "psubusb %%mm5, %%mm4 \n\t"
507 "por %%mm4, %%mm3 \n\t" // |l5 - l6|
508 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
509 "movq %%mm2, %%mm1 \n\t" // line 4
510 "psubusb %%mm5, %%mm2 \n\t"
511 "movq %%mm2, %%mm4 \n\t"
512 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
513 "psubusb %%mm1, %%mm5 \n\t"
514 "por %%mm5, %%mm4 \n\t" // |l4 - l5|
515 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
516 "movq %%mm4, %%mm3 \n\t" // d
517 "movq %2, %%mm0 \n\t"
518 "paddusb %%mm0, %%mm0 \n\t"
519 "psubusb %%mm0, %%mm4 \n\t"
520 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
521 "psubusb "MANGLE(b01)", %%mm3 \n\t"
522 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
523
524 PAVGB(%%mm7, %%mm3) // d/2
525 "movq %%mm3, %%mm1 \n\t" // d/2
526 PAVGB(%%mm7, %%mm3) // d/4
527 PAVGB(%%mm1, %%mm3) // 3*d/8
528
529 "movq (%0, %1, 4), %%mm0 \n\t" // line 4
530 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
531 "psubusb %%mm3, %%mm0 \n\t"
532 "pxor %%mm2, %%mm0 \n\t"
533 "movq %%mm0, (%0, %1, 4) \n\t" // line 4
534
535 "movq (%%"REG_c"), %%mm0 \n\t" // line 5
536 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
537 "paddusb %%mm3, %%mm0 \n\t"
538 "pxor %%mm2, %%mm0 \n\t"
539 "movq %%mm0, (%%"REG_c") \n\t" // line 5
540
541 PAVGB(%%mm7, %%mm1) // d/4
542
543 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
544 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
545 "psubusb %%mm1, %%mm0 \n\t"
546 "pxor %%mm2, %%mm0 \n\t"
547 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3
548
549 "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6
550 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
551 "paddusb %%mm1, %%mm0 \n\t"
552 "pxor %%mm2, %%mm0 \n\t"
553 "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6
554
555 PAVGB(%%mm7, %%mm1) // d/8
556
557 "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2
558 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
559 "psubusb %%mm1, %%mm0 \n\t"
560 "pxor %%mm2, %%mm0 \n\t"
561 "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2
562
563 "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7
564 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
565 "paddusb %%mm1, %%mm0 \n\t"
566 "pxor %%mm2, %%mm0 \n\t"
567 "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7
568
569 :
570 : "r" (src), "r" ((long)stride), "m" (co->pQPb)
571 : "%"REG_a, "%"REG_c
572 );
573 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
574
575 const int l1= stride;
576 const int l2= stride + l1;
577 const int l3= stride + l2;
578 const int l4= stride + l3;
579 const int l5= stride + l4;
580 const int l6= stride + l5;
581 const int l7= stride + l6;
582 // const int l8= stride + l7;
583 // const int l9= stride + l8;
584 int x;
585
586 src+= stride*3;
587 for(x=0; x<BLOCK_SIZE; x++)
588 {
589 int a= src[l3] - src[l4];
590 int b= src[l4] - src[l5];
591 int c= src[l5] - src[l6];
592
593 int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
594 d= FFMAX(d, 0);
595
596 if(d < co->QP*2)
597 {
598 int v = d * FFSIGN(-b);
599
600 src[l2] +=v>>3;
601 src[l3] +=v>>2;
602 src[l4] +=(3*v)>>3;
603 src[l5] -=(3*v)>>3;
604 src[l6] -=v>>2;
605 src[l7] -=v>>3;
606
607 }
608 src++;
609 }
610 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
611 }
612
613 #ifndef HAVE_ALTIVEC
614 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
615 {
616 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
617 /*
618 uint8_t tmp[16];
619 const int l1= stride;
620 const int l2= stride + l1;
621 const int l3= stride + l2;
622 const int l4= (int)tmp - (int)src - stride*3;
623 const int l5= (int)tmp - (int)src - stride*3 + 8;
624 const int l6= stride*3 + l3;
625 const int l7= stride + l6;
626 const int l8= stride + l7;
627
628 memcpy(tmp, src+stride*7, 8);
629 memcpy(tmp+8, src+stride*8, 8);
630 */
631 src+= stride*4;
632 asm volatile(
633
634 #if 0 //sligtly more accurate and slightly slower
635 "pxor %%mm7, %%mm7 \n\t" // 0
636 "lea (%0, %1), %%"REG_a" \n\t"
637 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
638 // 0 1 2 3 4 5 6 7
639 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
640 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
641
642
643 "movq (%0, %1, 2), %%mm0 \n\t" // l2
644 "movq (%0), %%mm1 \n\t" // l0
645 "movq %%mm0, %%mm2 \n\t" // l2
646 PAVGB(%%mm7, %%mm0) // ~l2/2
647 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
648 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
649
650 "movq (%%"REG_a"), %%mm1 \n\t" // l1
651 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3
652 "movq %%mm1, %%mm4 \n\t" // l1
653 PAVGB(%%mm7, %%mm1) // ~l1/2
654 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
655 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
656
657 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
658 "psubusb %%mm1, %%mm0 \n\t"
659 "psubusb %%mm4, %%mm1 \n\t"
660 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
661 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
662
663 "movq (%0, %1, 4), %%mm0 \n\t" // l4
664 "movq %%mm0, %%mm4 \n\t" // l4
665 PAVGB(%%mm7, %%mm0) // ~l4/2
666 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
667 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
668
669 "movq (%%"REG_c"), %%mm2 \n\t" // l5
670 "movq %%mm3, %%mm5 \n\t" // l3
671 PAVGB(%%mm7, %%mm3) // ~l3/2
672 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
673 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
674
675 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
676 "psubusb %%mm3, %%mm0 \n\t"
677 "psubusb %%mm6, %%mm3 \n\t"
678 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
679 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
680 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
681
682 "movq (%%"REG_c", %1), %%mm6 \n\t" // l6
683 "movq %%mm6, %%mm5 \n\t" // l6
684 PAVGB(%%mm7, %%mm6) // ~l6/2
685 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
686 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
687
688 "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7
689 "movq %%mm2, %%mm4 \n\t" // l5
690 PAVGB(%%mm7, %%mm2) // ~l5/2
691 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
692 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
693
694 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
695 "psubusb %%mm2, %%mm6 \n\t"
696 "psubusb %%mm4, %%mm2 \n\t"
697 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
698 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
699
700
701 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
702 "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
703 "paddusb "MANGLE(b01)", %%mm4 \n\t"
704 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
705 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
706 "pand %%mm4, %%mm3 \n\t"
707
708 "movq %%mm3, %%mm1 \n\t"
709 // "psubusb "MANGLE(b01)", %%mm3 \n\t"
710 PAVGB(%%mm7, %%mm3)
711 PAVGB(%%mm7, %%mm3)
712 "paddusb %%mm1, %%mm3 \n\t"
713 // "paddusb "MANGLE(b01)", %%mm3 \n\t"
714
715 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3
716 "movq (%0, %1, 4), %%mm5 \n\t" //l4
717 "movq (%0, %1, 4), %%mm4 \n\t" //l4
718 "psubusb %%mm6, %%mm5 \n\t"
719 "psubusb %%mm4, %%mm6 \n\t"
720 "por %%mm6, %%mm5 \n\t" // |l3-l4|
721 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
722 "pxor %%mm6, %%mm0 \n\t"
723 "pand %%mm0, %%mm3 \n\t"
724 PMINUB(%%mm5, %%mm3, %%mm0)
725
726 "psubusb "MANGLE(b01)", %%mm3 \n\t"
727 PAVGB(%%mm7, %%mm3)
728
729 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
730 "movq (%0, %1, 4), %%mm2 \n\t"
731 "pxor %%mm6, %%mm0 \n\t"
732 "pxor %%mm6, %%mm2 \n\t"
733 "psubb %%mm3, %%mm0 \n\t"
734 "paddb %%mm3, %%mm2 \n\t"
735 "pxor %%mm6, %%mm0 \n\t"
736 "pxor %%mm6, %%mm2 \n\t"
737 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
738 "movq %%mm2, (%0, %1, 4) \n\t"
739 #endif //0
740
741 "lea (%0, %1), %%"REG_a" \n\t"
742 "pcmpeqb %%mm6, %%mm6 \n\t" // -1
743 // 0 1 2 3 4 5 6 7
744 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
745 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
746
747
748 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3
749 "movq (%0, %1, 4), %%mm0 \n\t" // l4
750 "pxor %%mm6, %%mm1 \n\t" // -l3-1
751 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
752 // mm1=-l3-1, mm0=128-q
753
754 "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5
755 "movq (%%"REG_a", %1), %%mm3 \n\t" // l2
756 "pxor %%mm6, %%mm2 \n\t" // -l5-1
757 "movq %%mm2, %%mm5 \n\t" // -l5-1
758 "movq "MANGLE(b80)", %%mm4 \n\t" // 128
759 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
760 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
761 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
762 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
763 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
764 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
765
766 "movq (%%"REG_a"), %%mm2 \n\t" // l1
767 "pxor %%mm6, %%mm2 \n\t" // -l1-1
768 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
769 PAVGB((%0), %%mm1) // (l0-l3+256)/2
770 "movq "MANGLE(b80)", %%mm3 \n\t" // 128
771 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
772 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
773 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
774 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
775
776 PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2
777 "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7
778 "pxor %%mm6, %%mm1 \n\t" // -l7-1
779 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
780 "movq "MANGLE(b80)", %%mm2 \n\t" // 128
781 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
782 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
783 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
784 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
785
786 "movq "MANGLE(b00)", %%mm1 \n\t" // 0
787 "movq "MANGLE(b00)", %%mm5 \n\t" // 0
788 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
789 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
790 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
791 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
792 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
793
794 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
795
796 "movq "MANGLE(b00)", %%mm7 \n\t" // 0
797 "movq %2, %%mm2 \n\t" // QP
798 PAVGB(%%mm6, %%mm2) // 128 + QP/2
799 "psubb %%mm6, %%mm2 \n\t"
800
801 "movq %%mm4, %%mm1 \n\t"
802 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
803 "pxor %%mm1, %%mm4 \n\t"
804 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
805 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
806 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
807 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
808
809 "movq %%mm4, %%mm3 \n\t" // d
810 "psubusb "MANGLE(b01)", %%mm4 \n\t"
811 PAVGB(%%mm7, %%mm4) // d/32
812 PAVGB(%%mm7, %%mm4) // (d + 32)/64
813 "paddb %%mm3, %%mm4 \n\t" // 5d/64
814 "pand %%mm2, %%mm4 \n\t"
815
816 "movq "MANGLE(b80)", %%mm5 \n\t" // 128
817 "psubb %%mm0, %%mm5 \n\t" // q
818 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
819 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
820 "pxor %%mm7, %%mm5 \n\t"
821
822 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
823 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
824
825 "pand %%mm7, %%mm4 \n\t"
826 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
827 "movq (%0, %1, 4), %%mm2 \n\t"
828 "pxor %%mm1, %%mm0 \n\t"
829 "pxor %%mm1, %%mm2 \n\t"
830 "paddb %%mm4, %%mm0 \n\t"
831 "psubb %%mm4, %%mm2 \n\t"
832 "pxor %%mm1, %%mm0 \n\t"
833 "pxor %%mm1, %%mm2 \n\t"
834 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
835 "movq %%mm2, (%0, %1, 4) \n\t"
836
837 :
838 : "r" (src), "r" ((long)stride), "m" (c->pQPb)
839 : "%"REG_a, "%"REG_c
840 );
841
842 /*
843 {
844 int x;
845 src-= stride;
846 for(x=0; x<BLOCK_SIZE; x++)
847 {
848 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
849 if(FFABS(middleEnergy)< 8*QP)
850 {
851 const int q=(src[l4] - src[l5])/2;
852 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
853 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
854
855 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
856 d= FFMAX(d, 0);
857
858 d= (5*d + 32) >> 6;
859 d*= FFSIGN(-middleEnergy);
860
861 if(q>0)
862 {
863 d= d<0 ? 0 : d;
864 d= d>q ? q : d;
865 }
866 else
867 {
868 d= d>0 ? 0 : d;
869 d= d<q ? q : d;
870 }
871
872 src[l4]-= d;
873 src[l5]+= d;
874 }
875 src++;
876 }
877 src-=8;
878 for(x=0; x<8; x++)
879 {
880 int y;
881 for(y=4; y<6; y++)
882 {
883 int d= src[x+y*stride] - tmp[x+(y-4)*8];
884 int ad= FFABS(d);
885 static int max=0;
886 static int sum=0;
887 static int num=0;
888 static int bias=0;
889
890 if(max<ad) max=ad;
891 sum+= ad>3 ? 1 : 0;
892 if(ad>3)
893 {
894 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
895 }
896 if(y==4) bias+=d;
897 num++;
898 if(num%1000000 == 0)
899 {
900 av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias);
901 }
902 }
903 }
904 }
905 */
906 #elif defined (HAVE_MMX)
907 src+= stride*4;
908 asm volatile(
909 "pxor %%mm7, %%mm7 \n\t"
910 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
911 "and "ALIGN_MASK", %%"REG_c" \n\t" // align
912 // 0 1 2 3 4 5 6 7
913 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
914 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
915
916 "movq (%0), %%mm0 \n\t"
917 "movq %%mm0, %%mm1 \n\t"
918 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
919 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
920
921 "movq (%0, %1), %%mm2 \n\t"
922 "lea (%0, %1, 2), %%"REG_a" \n\t"
923 "movq %%mm2, %%mm3 \n\t"
924 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
925 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
926
927 "movq (%%"REG_a"), %%mm4 \n\t"
928 "movq %%mm4, %%mm5 \n\t"
929 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
930 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
931
932 "paddw %%mm0, %%mm0 \n\t" // 2L0
933 "paddw %%mm1, %%mm1 \n\t" // 2H0
934 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
935 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
936 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
937 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
938
939 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
940 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
941 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
942 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
943
944 "movq (%%"REG_a", %1), %%mm2 \n\t"
945 "movq %%mm2, %%mm3 \n\t"
946 "punpcklbw %%mm7, %%mm2 \n\t" // L3
947 "punpckhbw %%mm7, %%mm3 \n\t" // H3
948
949 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
950 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
951 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
952 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
953 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
954 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
955
956 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
957 "movq %%mm0, %%mm1 \n\t"
958 "punpcklbw %%mm7, %%mm0 \n\t" // L4
959 "punpckhbw %%mm7, %%mm1 \n\t" // H4
960
961 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
962 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
963 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
964 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
965 "paddw %%mm4, %%mm4 \n\t" // 2L2
966 "paddw %%mm5, %%mm5 \n\t" // 2H2
967 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
968 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
969
970 "lea (%%"REG_a", %1), %0 \n\t"
971 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
972 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
973 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
974 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
975 //50 opcodes so far
976 "movq (%0, %1, 2), %%mm2 \n\t"
977 "movq %%mm2, %%mm3 \n\t"
978 "punpcklbw %%mm7, %%mm2 \n\t" // L5
979 "punpckhbw %%mm7, %%mm3 \n\t" // H5
980 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
981 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
982 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
983 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
984
985 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
986 "punpcklbw %%mm7, %%mm6 \n\t" // L6
987 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
988 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
989 "punpckhbw %%mm7, %%mm6 \n\t" // H6
990 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
991
992 "paddw %%mm0, %%mm0 \n\t" // 2L4
993 "paddw %%mm1, %%mm1 \n\t" // 2H4
994 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
995 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
996
997 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
998 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
999 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
1000 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
1001
1002 "movq (%0, %1, 4), %%mm2 \n\t"
1003 "movq %%mm2, %%mm3 \n\t"
1004 "punpcklbw %%mm7, %%mm2 \n\t" // L7
1005 "punpckhbw %%mm7, %%mm3 \n\t" // H7
1006
1007 "paddw %%mm2, %%mm2 \n\t" // 2L7
1008 "paddw %%mm3, %%mm3 \n\t" // 2H7
1009 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
1010 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
1011
1012 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
1013 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
1014
1015 #ifdef HAVE_MMX2
1016 "movq %%mm7, %%mm6 \n\t" // 0
1017 "psubw %%mm0, %%mm6 \n\t"
1018 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1019 "movq %%mm7, %%mm6 \n\t" // 0
1020 "psubw %%mm1, %%mm6 \n\t"
1021 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1022 "movq %%mm7, %%mm6 \n\t" // 0
1023 "psubw %%mm2, %%mm6 \n\t"
1024 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1025 "movq %%mm7, %%mm6 \n\t" // 0
1026 "psubw %%mm3, %%mm6 \n\t"
1027 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1028 #else
1029 "movq %%mm7, %%mm6 \n\t" // 0
1030 "pcmpgtw %%mm0, %%mm6 \n\t"
1031 "pxor %%mm6, %%mm0 \n\t"
1032 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1033 "movq %%mm7, %%mm6 \n\t" // 0
1034 "pcmpgtw %%mm1, %%mm6 \n\t"
1035 "pxor %%mm6, %%mm1 \n\t"
1036 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1037 "movq %%mm7, %%mm6 \n\t" // 0
1038 "pcmpgtw %%mm2, %%mm6 \n\t"
1039 "pxor %%mm6, %%mm2 \n\t"
1040 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1041 "movq %%mm7, %%mm6 \n\t" // 0
1042 "pcmpgtw %%mm3, %%mm6 \n\t"
1043 "pxor %%mm6, %%mm3 \n\t"
1044 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1045 #endif
1046
1047 #ifdef HAVE_MMX2
1048 "pminsw %%mm2, %%mm0 \n\t"
1049 "pminsw %%mm3, %%mm1 \n\t"
1050 #else
1051 "movq %%mm0, %%mm6 \n\t"
1052 "psubusw %%mm2, %%mm6 \n\t"
1053 "psubw %%mm6, %%mm0 \n\t"
1054 "movq %%mm1, %%mm6 \n\t"
1055 "psubusw %%mm3, %%mm6 \n\t"
1056 "psubw %%mm6, %%mm1 \n\t"
1057 #endif
1058
1059 "movd %2, %%mm2 \n\t" // QP
1060 "punpcklbw %%mm7, %%mm2 \n\t"
1061
1062 "movq %%mm7, %%mm6 \n\t" // 0
1063 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1064 "pxor %%mm6, %%mm4 \n\t"
1065 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1066 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1067 "pxor %%mm7, %%mm5 \n\t"
1068 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1069 // 100 opcodes
1070 "psllw $3, %%mm2 \n\t" // 8QP
1071 "movq %%mm2, %%mm3 \n\t" // 8QP
1072 "pcmpgtw %%mm4, %%mm2 \n\t"
1073 "pcmpgtw %%mm5, %%mm3 \n\t"
1074 "pand %%mm2, %%mm4 \n\t"
1075 "pand %%mm3, %%mm5 \n\t"
1076
1077
1078 "psubusw %%mm0, %%mm4 \n\t" // hd
1079 "psubusw %%mm1, %%mm5 \n\t" // ld
1080
1081
1082 "movq "MANGLE(w05)", %%mm2 \n\t" // 5
1083 "pmullw %%mm2, %%mm4 \n\t"
1084 "pmullw %%mm2, %%mm5 \n\t"
1085 "movq "MANGLE(w20)", %%mm2 \n\t" // 32
1086 "paddw %%mm2, %%mm4 \n\t"
1087 "paddw %%mm2, %%mm5 \n\t"
1088 "psrlw $6, %%mm4 \n\t"
1089 "psrlw $6, %%mm5 \n\t"
1090
1091 "movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4
1092 "movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4
1093
1094 "pxor %%mm2, %%mm2 \n\t"
1095 "pxor %%mm3, %%mm3 \n\t"
1096
1097 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1098 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
1099 "pxor %%mm2, %%mm0 \n\t"
1100 "pxor %%mm3, %%mm1 \n\t"
1101 "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
1102 "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
1103 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
1104 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
1105
1106 "pxor %%mm6, %%mm2 \n\t"
1107 "pxor %%mm7, %%mm3 \n\t"
1108 "pand %%mm2, %%mm4 \n\t"
1109 "pand %%mm3, %%mm5 \n\t"
1110
1111 #ifdef HAVE_MMX2
1112 "pminsw %%mm0, %%mm4 \n\t"
1113 "pminsw %%mm1, %%mm5 \n\t"
1114 #else
1115 "movq %%mm4, %%mm2 \n\t"
1116 "psubusw %%mm0, %%mm2 \n\t"
1117 "psubw %%mm2, %%mm4 \n\t"
1118 "movq %%mm5, %%mm2 \n\t"
1119 "psubusw %%mm1, %%mm2 \n\t"
1120 "psubw %%mm2, %%mm5 \n\t"
1121 #endif
1122 "pxor %%mm6, %%mm4 \n\t"
1123 "pxor %%mm7, %%mm5 \n\t"
1124 "psubw %%mm6, %%mm4 \n\t"
1125 "psubw %%mm7, %%mm5 \n\t"
1126 "packsswb %%mm5, %%mm4 \n\t"
1127 "movq (%0), %%mm0 \n\t"
1128 "paddb %%mm4, %%mm0 \n\t"
1129 "movq %%mm0, (%0) \n\t"
1130 "movq (%0, %1), %%mm0 \n\t"
1131 "psubb %%mm4, %%mm0 \n\t"
1132 "movq %%mm0, (%0, %1) \n\t"
1133
1134 : "+r" (src)
1135 : "r" ((long)stride), "m" (c->pQPb)
1136 : "%"REG_a, "%"REG_c
1137 );
1138 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1139 const int l1= stride;
1140 const int l2= stride + l1;
1141 const int l3= stride + l2;
1142 const int l4= stride + l3;
1143 const int l5= stride + l4;
1144 const int l6= stride + l5;
1145 const int l7= stride + l6;
1146 const int l8= stride + l7;
1147 // const int l9= stride + l8;
1148 int x;
1149 src+= stride*3;
1150 for(x=0; x<BLOCK_SIZE; x++)
1151 {
1152 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1153 if(FFABS(middleEnergy) < 8*c->QP)
1154 {
1155 const int q=(src[l4] - src[l5])/2;
1156 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1157 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1158
1159 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
1160 d= FFMAX(d, 0);
1161
1162 d= (5*d + 32) >> 6;
1163 d*= FFSIGN(-middleEnergy);
1164
1165 if(q>0)
1166 {
1167 d= d<0 ? 0 : d;
1168 d= d>q ? q : d;
1169 }
1170 else
1171 {
1172 d= d>0 ? 0 : d;
1173 d= d<q ? q : d;
1174 }
1175
1176 src[l4]-= d;
1177 src[l5]+= d;
1178 }
1179 src++;
1180 }
1181 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1182 }
1183 #endif //HAVE_ALTIVEC
1184
1185 #ifndef HAVE_ALTIVEC
1186 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
1187 {
1188 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1189 asm volatile(
1190 "pxor %%mm6, %%mm6 \n\t"
1191 "pcmpeqb %%mm7, %%mm7 \n\t"
1192 "movq %2, %%mm0 \n\t"
1193 "punpcklbw %%mm6, %%mm0 \n\t"
1194 "psrlw $1, %%mm0 \n\t"
1195 "psubw %%mm7, %%mm0 \n\t"
1196 "packuswb %%mm0, %%mm0 \n\t"
1197 "movq %%mm0, %3 \n\t"
1198
1199 "lea (%0, %1), %%"REG_a" \n\t"
1200 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1201
1202 // 0 1 2 3 4 5 6 7 8 9
1203 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1204
1205 #undef FIND_MIN_MAX
1206 #ifdef HAVE_MMX2
1207 #define REAL_FIND_MIN_MAX(addr)\
1208 "movq " #addr ", %%mm0 \n\t"\
1209 "pminub %%mm0, %%mm7 \n\t"\
1210 "pmaxub %%mm0, %%mm6 \n\t"
1211 #else
1212 #define REAL_FIND_MIN_MAX(addr)\
1213 "movq " #addr ", %%mm0 \n\t"\
1214 "movq %%mm7, %%mm1 \n\t"\
1215 "psubusb %%mm0, %%mm6 \n\t"\
1216 "paddb %%mm0, %%mm6 \n\t"\
1217 "psubusb %%mm0, %%mm1 \n\t"\
1218 "psubb %%mm1, %%mm7 \n\t"
1219 #endif
1220 #define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr)
1221
1222 FIND_MIN_MAX((%%REGa))
1223 FIND_MIN_MAX((%%REGa, %1))
1224 FIND_MIN_MAX((%%REGa, %1, 2))
1225 FIND_MIN_MAX((%0, %1, 4))
1226 FIND_MIN_MAX((%%REGd))
1227 FIND_MIN_MAX((%%REGd, %1))
1228 FIND_MIN_MAX((%%REGd, %1, 2))
1229 FIND_MIN_MAX((%0, %1, 8))
1230
1231 "movq %%mm7, %%mm4 \n\t"
1232 "psrlq $8, %%mm7 \n\t"
1233 #ifdef HAVE_MMX2
1234 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1235 "pshufw $0xF9, %%mm7, %%mm4 \n\t"
1236 "pminub %%mm4, %%mm7 \n\t" // min of pixels
1237 "pshufw $0xFE, %%mm7, %%mm4 \n\t"
1238 "pminub %%mm4, %%mm7 \n\t"
1239 #else
1240 "movq %%mm7, %%mm1 \n\t"
1241 "psubusb %%mm4, %%mm1 \n\t"
1242 "psubb %%mm1, %%mm7 \n\t"
1243 "movq %%mm7, %%mm4 \n\t"
1244 "psrlq $16, %%mm7 \n\t"
1245 "movq %%mm7, %%mm1 \n\t"
1246 "psubusb %%mm4, %%mm1 \n\t"
1247 "psubb %%mm1, %%mm7 \n\t"
1248 "movq %%mm7, %%mm4 \n\t"
1249 "psrlq $32, %%mm7 \n\t"
1250 "movq %%mm7, %%mm1 \n\t"
1251 "psubusb %%mm4, %%mm1 \n\t"
1252 "psubb %%mm1, %%mm7 \n\t"
1253 #endif
1254
1255
1256 "movq %%mm6, %%mm4 \n\t"
1257 "psrlq $8, %%mm6 \n\t"
1258 #ifdef HAVE_MMX2
1259 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1260 "pshufw $0xF9, %%mm6, %%mm4 \n\t"
1261 "pmaxub %%mm4, %%mm6 \n\t"
1262 "pshufw $0xFE, %%mm6, %%mm4 \n\t"
1263 "pmaxub %%mm4, %%mm6 \n\t"
1264 #else
1265 "psubusb %%mm4, %%mm6 \n\t"
1266 "paddb %%mm4, %%mm6 \n\t"
1267 "movq %%mm6, %%mm4 \n\t"
1268 "psrlq $16, %%mm6 \n\t"
1269 "psubusb %%mm4, %%mm6 \n\t"
1270 "paddb %%mm4, %%mm6 \n\t"
1271 "movq %%mm6, %%mm4 \n\t"
1272 "psrlq $32, %%mm6 \n\t"
1273 "psubusb %%mm4, %%mm6 \n\t"
1274 "paddb %%mm4, %%mm6 \n\t"
1275 #endif
1276 "movq %%mm6, %%mm0 \n\t" // max
1277 "psubb %%mm7, %%mm6 \n\t" // max - min
1278 "movd %%mm6, %%ecx \n\t"
1279 "cmpb "MANGLE(deringThreshold)", %%cl \n\t"
1280 " jb 1f \n\t"
1281 "lea -24(%%"REG_SP"), %%"REG_c" \n\t"
1282 "and "ALIGN_MASK", %%"REG_c" \n\t"
1283 PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1284 "punpcklbw %%mm7, %%mm7 \n\t"
1285 "punpcklbw %%mm7, %%mm7 \n\t"
1286 "punpcklbw %%mm7, %%mm7 \n\t"
1287 "movq %%mm7, (%%"REG_c") \n\t"
1288
1289 "movq (%0), %%mm0 \n\t" // L10
1290 "movq %%mm0, %%mm1 \n\t" // L10
1291 "movq %%mm0, %%mm2 \n\t" // L10
1292 "psllq $8, %%mm1 \n\t"
1293 "psrlq $8, %%mm2 \n\t"
1294 "movd -4(%0), %%mm3 \n\t"
1295 "movd 8(%0), %%mm4 \n\t"
1296 "psrlq $24, %%mm3 \n\t"
1297 "psllq $56, %%mm4 \n\t"
1298 "por %%mm3, %%mm1 \n\t" // L00
1299 "por %%mm4, %%mm2 \n\t" // L20
1300 "movq %%mm1, %%mm3 \n\t" // L00
1301 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1302 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1303 "psubusb %%mm7, %%mm0 \n\t"
1304 "psubusb %%mm7, %%mm2 \n\t"
1305 "psubusb %%mm7, %%mm3 \n\t"
1306 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1307 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1308 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1309 "paddb %%mm2, %%mm0 \n\t"
1310 "paddb %%mm3, %%mm0 \n\t"
1311
1312 "movq (%%"REG_a"), %%mm2 \n\t" // L11
1313 "movq %%mm2, %%mm3 \n\t" // L11
1314 "movq %%mm2, %%mm4 \n\t" // L11
1315 "psllq $8, %%mm3 \n\t"
1316 "psrlq $8, %%mm4 \n\t"
1317 "movd -4(%%"REG_a"), %%mm5 \n\t"
1318 "movd 8(%%"REG_a"), %%mm6 \n\t"
1319 "psrlq $24, %%mm5 \n\t"
1320 "psllq $56, %%mm6 \n\t"
1321 "por %%mm5, %%mm3 \n\t" // L01
1322 "por %%mm6, %%mm4 \n\t" // L21
1323 "movq %%mm3, %%mm5 \n\t" // L01
1324 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1325 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1326 "psubusb %%mm7, %%mm2 \n\t"
1327 "psubusb %%mm7, %%mm4 \n\t"
1328 "psubusb %%mm7, %%mm5 \n\t"
1329 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1330 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1331 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1332 "paddb %%mm4, %%mm2 \n\t"
1333 "paddb %%mm5, %%mm2 \n\t"
1334 // 0, 2, 3, 1
1335 #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1336 "movq " #src ", " #sx " \n\t" /* src[0] */\
1337 "movq " #sx ", " #lx " \n\t" /* src[0] */\
1338 "movq " #sx ", " #t0 " \n\t" /* src[0] */\
1339 "psllq $8, " #lx " \n\t"\
1340 "psrlq $8, " #t0 " \n\t"\
1341 "movd -4" #src ", " #t1 " \n\t"\
1342 "psrlq $24, " #t1 " \n\t"\
1343 "por " #t1 ", " #lx " \n\t" /* src[-1] */\
1344 "movd 8" #src ", " #t1 " \n\t"\
1345 "psllq $56, " #t1 " \n\t"\
1346 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1347 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1348 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1349 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1350 PAVGB(lx, pplx) \
1351 "movq " #lx ", 8(%%"REG_c") \n\t"\
1352 "movq (%%"REG_c"), " #lx " \n\t"\
1353 "psubusb " #lx ", " #t1 " \n\t"\
1354 "psubusb " #lx ", " #t0 " \n\t"\
1355 "psubusb " #lx ", " #sx " \n\t"\
1356 "movq "MANGLE(b00)", " #lx " \n\t"\
1357 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1358 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1359 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1360 "paddb " #t1 ", " #t0 " \n\t"\
1361 "paddb " #t0 ", " #sx " \n\t"\
1362 \
1363 PAVGB(plx, pplx) /* filtered */\
1364 "movq " #dst ", " #t0 " \n\t" /* dst */\
1365 "movq " #t0 ", " #t1 " \n\t" /* dst */\
1366 "psubusb %3, " #t0 " \n\t"\
1367 "paddusb %3, " #t1 " \n\t"\
1368 PMAXUB(t0, pplx)\
1369 PMINUB(t1, pplx, t0)\
1370 "paddb " #sx ", " #ppsx " \n\t"\
1371 "paddb " #psx ", " #ppsx " \n\t"\
1372 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1373 "pand "MANGLE(b08)", " #ppsx " \n\t"\
1374 "pcmpeqb " #lx ", " #ppsx " \n\t"\
1375 "pand " #ppsx ", " #pplx " \n\t"\
1376 "pandn " #dst ", " #ppsx " \n\t"\
1377 "por " #pplx ", " #ppsx " \n\t"\
1378 "movq " #ppsx ", " #dst " \n\t"\
1379 "movq 8(%%"REG_c"), " #lx " \n\t"
1380
1381 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1382 REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1)
1383 /*
1384 0000000
1385 1111111
1386
1387 1111110
1388 1111101
1389 1111100
1390 1111011
1391 1111010
1392 1111001
1393
1394 1111000
1395 1110111
1396
1397 */
1398 //DERING_CORE(dst ,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
1399 DERING_CORE((%%REGa) ,(%%REGa, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1400 DERING_CORE((%%REGa, %1) ,(%%REGa, %1, 2),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1401 DERING_CORE((%%REGa, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1402 DERING_CORE((%0, %1, 4) ,(%%REGd) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1403 DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1404 DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1405 DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1406 DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1407
1408 "1: \n\t"
1409 : : "r" (src), "r" ((long)stride), "m" (c->pQPb), "m"(c->pQPb2)
1410 : "%"REG_a, "%"REG_d, "%"REG_c
1411 );
1412 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1413 int y;
1414 int min=255;
1415 int max=0;
1416 int avg;
1417 uint8_t *p;
1418 int s[10];
1419 const int QP2= c->QP/2 + 1;
1420
1421 for(y=1; y<9; y++)
1422 {
1423 int x;
1424 p= src + stride*y;
1425 for(x=1; x<9; x++)
1426 {
1427 p++;
1428 if(*p > max) max= *p;
1429 if(*p < min) min= *p;
1430 }
1431 }
1432 avg= (min + max + 1)>>1;
1433
1434 if(max - min <deringThreshold) return;
1435
1436 for(y=0; y<10; y++)
1437 {
1438 int t = 0;
1439
1440 if(src[stride*y + 0] > avg) t+= 1;
1441 if(src[stride*y + 1] > avg) t+= 2;
1442 if(src[stride*y + 2] > avg) t+= 4;
1443 if(src[stride*y + 3] > avg) t+= 8;
1444 if(src[stride*y + 4] > avg) t+= 16;
1445 if(src[stride*y + 5] > avg) t+= 32;
1446 if(src[stride*y + 6] > avg) t+= 64;
1447 if(src[stride*y + 7] > avg) t+= 128;
1448 if(src[stride*y + 8] > avg) t+= 256;
1449 if(src[stride*y + 9] > avg) t+= 512;
1450
1451 t |= (~t)<<16;
1452 t &= (t<<1) & (t>>1);
1453 s[y] = t;
1454 }
1455
1456 for(y=1; y<9; y++)
1457 {
1458 int t = s[y-1] & s[y] & s[y+1];
1459 t|= t>>16;
1460 s[y-1]= t;
1461 }
1462
1463 for(y=1; y<9; y++)
1464 {
1465 int x;
1466 int t = s[y-1];
1467
1468 p= src + stride*y;
1469 for(x=1; x<9; x++)
1470 {
1471 p++;
1472 if(t & (1<<x))
1473 {
1474 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1475 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
1476 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1477 f= (f + 8)>>4;
1478
1479 #ifdef DEBUG_DERING_THRESHOLD
1480 asm volatile("emms\n\t":);
1481 {
1482 static long long numPixels=0;
1483 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1484 // if((max-min)<20 || (max-min)*QP<200)
1485 // if((max-min)*QP < 500)
1486 // if(max-min<QP/2)
1487 if(max-min < 20)
1488 {
1489 static int numSkiped=0;
1490 static int errorSum=0;
1491 static int worstQP=0;
1492 static int worstRange=0;
1493 static int worstDiff=0;
1494 int diff= (f - *p);
1495 int absDiff= FFABS(diff);
1496 int error= diff*diff;
1497
1498 if(x==1 || x==8 || y==1 || y==8) continue;
1499
1500 numSkiped++;
1501 if(absDiff > worstDiff)
1502 {
1503 worstDiff= absDiff;
1504 worstQP= QP;
1505 worstRange= max-min;
1506 }
1507 errorSum+= error;
1508
1509 if(1024LL*1024LL*1024LL % numSkiped == 0)
1510 {
1511 av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, "
1512 "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1513 (float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
1514 worstDiff, (float)numSkiped/numPixels);
1515 }
1516 }
1517 }
1518 #endif
1519 if (*p + QP2 < f) *p= *p + QP2;
1520 else if(*p - QP2 > f) *p= *p - QP2;
1521 else *p=f;
1522 }
1523 }
1524 }
1525 #ifdef DEBUG_DERING_THRESHOLD
1526 if(max-min < 20)
1527 {
1528 for(y=1; y<9; y++)
1529 {
1530 int x;
1531 int t = 0;
1532 p= src + stride*y;
1533 for(x=1; x<9; x++)
1534 {
1535 p++;
1536 *p = FFMIN(*p + 20, 255);
1537 }
1538 }
1539 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1540 }
1541 #endif
1542 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1543 }
1544 #endif //HAVE_ALTIVEC
1545
1546 /**
1547 * Deinterlaces the given block by linearly interpolating every second line.
1548 * will be called for every 8x8 block and can read & write from line 4-15
1549 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1550 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1551 */
1552 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1553 {
1554 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1555 src+= 4*stride;
1556 asm volatile(
1557 "lea (%0, %1), %%"REG_a" \n\t"
1558 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
1559 // 0 1 2 3 4 5 6 7 8 9
1560 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
1561
1562 "movq (%0), %%mm0 \n\t"
1563 "movq (%%"REG_a", %1), %%mm1 \n\t"
1564 PAVGB(%%mm1, %%mm0)
1565 "movq %%mm0, (%%"REG_a") \n\t"
1566 "movq (%0, %1, 4), %%mm0 \n\t"
1567 PAVGB(%%mm0, %%mm1)
1568 "movq %%mm1, (%%"REG_a", %1, 2) \n\t"
1569 "movq (%%"REG_c", %1), %%mm1 \n\t"
1570 PAVGB(%%mm1, %%mm0)
1571 "movq %%mm0, (%%"REG_c") \n\t"
1572 "movq (%0, %1, 8), %%mm0 \n\t"
1573 PAVGB(%%mm0, %%mm1)
1574 "movq %%mm1, (%%"REG_c", %1, 2) \n\t"
1575
1576 : : "r" (src), "r" ((long)stride)
1577 : "%"REG_a, "%"REG_c
1578 );
1579 #else
1580 int a, b, x;
1581 src+= 4*stride;
1582
1583 for(x=0; x<2; x++){
1584 a= *(uint32_t*)&src[stride*0];
1585 b= *(uint32_t*)&src[stride*2];
1586 *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1587 a= *(uint32_t*)&src[stride*4];
1588 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1589 b= *(uint32_t*)&src[stride*6];
1590 *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1591 a= *(uint32_t*)&src[stride*8];
1592 *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1593 src += 4;
1594 }
1595 #endif
1596 }
1597
1598 /**
1599 * Deinterlaces the given block by cubic interpolating every second line.
1600 * will be called for every 8x8 block and can read & write from line 4-15
1601 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1602 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1603 * this filter will read lines 3-15 and write 7-13
1604 */
1605 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1606 {
1607 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1608 src+= stride*3;
1609 asm volatile(
1610 "lea (%0, %1), %%"REG_a" \n\t"
1611 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1612 "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t"
1613 "add %1, %%"REG_c" \n\t"
1614 "pxor %%mm7, %%mm7 \n\t"
1615 // 0 1 2 3 4 5 6 7 8 9 10
1616 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1617
1618 #define REAL_DEINT_CUBIC(a,b,c,d,e)\
1619 "movq " #a ", %%mm0 \n\t"\
1620 "movq " #b ", %%mm1 \n\t"\
1621 "movq " #d ", %%mm2 \n\t"\
1622 "movq " #e ", %%mm3 \n\t"\
1623 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\
1624 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\
1625 "movq %%mm0, %%mm2 \n\t"\
1626 "punpcklbw %%mm7, %%mm0 \n\t"\
1627 "punpckhbw %%mm7, %%mm2 \n\t"\
1628 "movq %%mm1, %%mm3 \n\t"\
1629 "punpcklbw %%mm7, %%mm1 \n\t"\
1630 "punpckhbw %%mm7, %%mm3 \n\t"\
1631 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\
1632 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\
1633 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\
1634 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\
1635 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\
1636 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\
1637 "packuswb %%mm3, %%mm1 \n\t"\
1638 "movq %%mm1, " #c " \n\t"
1639 #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e)
1640
1641 DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1))
1642 DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8))
1643 DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc))
1644 DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2))
1645
1646 : : "r" (src), "r" ((long)stride)
1647 : "%"REG_a, "%"REG_d, "%"REG_c
1648 );
1649 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1650 int x;
1651 src+= stride*3;
1652 for(x=0; x<8; x++)
1653 {
1654 src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4);
1655 src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4);
1656 src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4);
1657 src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4);
1658 src++;
1659 }
1660 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1661 }
1662
1663 /**
1664 * Deinterlaces the given block by filtering every second line with a (-1 4 2 4 -1) filter.
1665 * will be called for every 8x8 block and can read & write from line 4-15
1666 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1667 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1668 * this filter will read lines 4-13 and write 5-11
1669 */
1670 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
1671 {
1672 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1673 src+= stride*4;
1674 asm volatile(
1675 "lea (%0, %1), %%"REG_a" \n\t"
1676 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1677 "pxor %%mm7, %%mm7 \n\t"
1678 "movq (%2), %%mm0 \n\t"
1679 // 0 1 2 3 4 5 6 7 8 9 10
1680 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1681
1682 #define REAL_DEINT_FF(a,b,c,d)\
1683 "movq " #a ", %%mm1 \n\t"\
1684 "movq " #b ", %%mm2 \n\t"\
1685 "movq " #c ", %%mm3 \n\t"\
1686 "movq " #d ", %%mm4 \n\t"\
1687 PAVGB(%%mm3, %%mm1) \
1688 PAVGB(%%mm4, %%mm0) \
1689 "movq %%mm0, %%mm3 \n\t"\
1690 "punpcklbw %%mm7, %%mm0 \n\t"\
1691 "punpckhbw %%mm7, %%mm3 \n\t"\
1692 "movq %%mm1, %%mm4 \n\t"\
1693 "punpcklbw %%mm7, %%mm1 \n\t"\
1694 "punpckhbw %%mm7, %%mm4 \n\t"\
1695 "psllw $2, %%mm1 \n\t"\
1696 "psllw $2, %%mm4 \n\t"\
1697 "psubw %%mm0, %%mm1 \n\t"\
1698 "psubw %%mm3, %%mm4 \n\t"\
1699 "movq %%mm2, %%mm5 \n\t"\
1700 "movq %%mm2, %%mm0 \n\t"\
1701 "punpcklbw %%mm7, %%mm2 \n\t"\
1702 "punpckhbw %%mm7, %%mm5 \n\t"\
1703 "paddw %%mm2, %%mm1 \n\t"\
1704 "paddw %%mm5, %%mm4 \n\t"\
1705 "psraw $2, %%mm1 \n\t"\
1706 "psraw $2, %%mm4 \n\t"\
1707 "packuswb %%mm4, %%mm1 \n\t"\
1708 "movq %%mm1, " #b " \n\t"\
1709
1710 #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d)
1711
1712 DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2))
1713 DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1714 DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2))
1715 DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1716
1717 "movq %%mm0, (%2) \n\t"
1718 : : "r" (src), "r" ((long)stride), "r"(tmp)
1719 : "%"REG_a, "%"REG_d
1720 );
1721 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1722 int x;
1723 src+= stride*4;
1724 for(x=0; x<8; x++)
1725 {
1726 int t1= tmp[x];
1727 int t2= src[stride*1];
1728
1729 src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3);
1730 t1= src[stride*4];
1731 src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3);
1732 t2= src[stride*6];
1733 src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3);
1734 t1= src[stride*8];
1735 src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3);
1736 tmp[x]= t1;
1737
1738 src++;
1739 }
1740 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1741 }
1742
1743 /**
1744 * Deinterlaces the given block by filtering every line with a (-1 2 6 2 -1) filter.
1745 * will be called for every 8x8 block and can read & write from line 4-15
1746 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1747 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1748 * this filter will read lines 4-13 and write 4-11
1749 */
1750 static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2)
1751 {
1752 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1753 src+= stride*4;
1754 asm volatile(
1755 "lea (%0, %1), %%"REG_a" \n\t"
1756 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1757 "pxor %%mm7, %%mm7 \n\t"
1758 "movq (%2), %%mm0 \n\t"
1759 "movq (%3), %%mm1 \n\t"
1760 // 0 1 2 3 4 5 6 7 8 9 10
1761 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
1762
1763 #define REAL_DEINT_L5(t1,t2,a,b,c)\
1764 "movq " #a ", %%mm2 \n\t"\
1765 "movq " #b ", %%mm3 \n\t"\
1766 "movq " #c ", %%mm4 \n\t"\
1767 PAVGB(t2, %%mm3) \
1768 PAVGB(t1, %%mm4) \
1769 "movq %%mm2, %%mm5 \n\t"\
1770 "movq %%mm2, " #t1 " \n\t"\
1771 "punpcklbw %%mm7, %%mm2 \n\t"\
1772 "punpckhbw %%mm7, %%mm5 \n\t"\
1773 "movq %%mm2, %%mm6 \n\t"\
1774 "paddw %%mm2, %%mm2 \n\t"\
1775 "paddw %%mm6, %%mm2 \n\t"\
1776 "movq %%mm5, %%mm6 \n\t"\
1777 "paddw %%mm5, %%mm5 \n\t"\
1778 "paddw %%mm6, %%mm5 \n\t"\
1779 "movq %%mm3, %%mm6 \n\t"\
1780 "punpcklbw %%mm7, %%mm3 \n\t"\
1781 "punpckhbw %%mm7, %%mm6 \n\t"\
1782 "paddw %%mm3, %%mm3 \n\t"\
1783 "paddw %%mm6, %%mm6 \n\t"\
1784 "paddw %%mm3, %%mm2 \n\t"\
1785 "paddw %%mm6, %%mm5 \n\t"\
1786 "movq %%mm4, %%mm6 \n\t"\
1787 "punpcklbw %%mm7, %%mm4 \n\t"\
1788 "punpckhbw %%mm7, %%mm6 \n\t"\
1789 "psubw %%mm4, %%mm2 \n\t"\
1790 "psubw %%mm6, %%mm5 \n\t"\
1791 "psraw $2, %%mm2 \n\t"\
1792 "psraw $2, %%mm5 \n\t"\
1793 "packuswb %%mm5, %%mm2 \n\t"\
1794 "movq %%mm2, " #a " \n\t"\
1795
1796 #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c)
1797
1798 DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) )
1799 DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2))
1800 DEINT_L5(%%mm0, %%mm1, (%%REGa, %1) , (%%REGa, %1, 2), (%0, %1, 4) )
1801 DEINT_L5(%%mm1, %%mm0, (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) )
1802 DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) )
1803 DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2))
1804 DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) )
1805 DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1806
1807 "movq %%mm0, (%2) \n\t"
1808 "movq %%mm1, (%3) \n\t"
1809 : : "r" (src), "r" ((long)stride), "r"(tmp), "r"(tmp2)
1810 : "%"REG_a, "%"REG_d
1811 );
1812 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1813 int x;
1814 src+= stride*4;
1815 for(x=0; x<8; x++)
1816 {
1817 int t1= tmp[x];
1818 int t2= tmp2[x];
1819 int t3= src[0];
1820
1821 src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
1822 t1= src[stride*1];
1823 src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
1824 t2= src[stride*2];
1825 src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
1826 t3= src[stride*3];
1827 src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
1828 t1= src[stride*4];
1829 src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
1830 t2= src[stride*5];
1831 src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
1832 t3= src[stride*6];
1833 src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
1834 t1= src[stride*7];
1835 src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);
1836
1837 tmp[x]= t3;
1838 tmp2[x]= t1;
1839
1840 src++;
1841 }
1842 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1843 }
1844
1845 /**
1846 * Deinterlaces the given block by filtering all lines with a (1 2 1) filter.
1847 * will be called for every 8x8 block and can read & write from line 4-15
1848 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1849 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1850 * this filter will read lines 4-13 and write 4-11
1851 */
1852 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp)
1853 {
1854 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1855 src+= 4*stride;
1856 asm volatile(
1857 "lea (%0, %1), %%"REG_a" \n\t"
1858 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1859 // 0 1 2 3 4 5 6 7 8 9
1860 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1861
1862 "movq (%2), %%mm0 \n\t" // L0
1863 "movq (%%"REG_a"), %%mm1 \n\t" // L2
1864 PAVGB(%%mm1, %%mm0) // L0+L2
1865 "movq (%0), %%mm2 \n\t" // L1
1866 PAVGB(%%mm2, %%mm0)
1867 "movq %%mm0, (%0) \n\t"
1868 "movq (%%"REG_a", %1), %%mm0 \n\t" // L3
1869 PAVGB(%%mm0, %%mm2) // L1+L3
1870 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1871 "movq %%mm2, (%%"REG_a") \n\t"
1872 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4
1873 PAVGB(%%mm2, %%mm1) // L2+L4
1874 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1875 "movq %%mm1, (%%"REG_a", %1) \n\t"
1876 "movq (%0, %1, 4), %%mm1 \n\t" // L5
1877 PAVGB(%%mm1, %%mm0) // L3+L5
1878 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1879 "movq %%mm0, (%%"REG_a", %1, 2) \n\t"
1880 "movq (%%"REG_d"), %%mm0 \n\t" // L6
1881 PAVGB(%%mm0, %%mm2) // L4+L6
1882 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1883 "movq %%mm2, (%0, %1, 4) \n\t"
1884 "movq (%%"REG_d", %1), %%mm2 \n\t" // L7
1885 PAVGB(%%mm2, %%mm1) // L5+L7
1886 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1887 "movq %%mm1, (%%"REG_d") \n\t"
1888 "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8
1889 PAVGB(%%mm1, %%mm0) // L6+L8
1890 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1891 "movq %%mm0, (%%"REG_d", %1) \n\t"
1892 "movq (%0, %1, 8), %%mm0 \n\t" // L9
1893 PAVGB(%%mm0, %%mm2) // L7+L9
1894 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1895 "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1896 "movq %%mm1, (%2) \n\t"
1897
1898 : : "r" (src), "r" ((long)stride), "r" (tmp)
1899 : "%"REG_a, "%"REG_d
1900 );
1901 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1902 int a, b, c, x;
1903 src+= 4*stride;
1904
1905 for(x=0; x<2; x++){
1906 a= *(uint32_t*)&tmp[stride*0];
1907 b= *(uint32_t*)&src[stride*0];
1908 c= *(uint32_t*)&src[stride*1];
1909 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1910 *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1911
1912 a= *(uint32_t*)&src[stride*2];
1913 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1914 *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1915
1916 b= *(uint32_t*)&src[stride*3];
1917 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1918 *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1919
1920 c= *(uint32_t*)&src[stride*4];
1921 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1922 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1923
1924 a= *(uint32_t*)&src[stride*5];
1925 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1926 *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1927
1928 b= *(uint32_t*)&src[stride*6];
1929 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1930 *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1931
1932 c= *(uint32_t*)&src[stride*7];
1933 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1934 *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1935
1936 a= *(uint32_t*)&src[stride*8];
1937 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1938 *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1939
1940 *(uint32_t*)&tmp[stride*0]= c;
1941 src += 4;
1942 tmp += 4;
1943 }
1944 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1945 }
1946
1947 /**
1948 * Deinterlaces the given block by applying a median filter to every second line.
1949 * will be called for every 8x8 block and can read & write from line 4-15,
1950 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1951 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1952 */
1953 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1954 {
1955 #ifdef HAVE_MMX
1956 src+= 4*stride;
1957 #ifdef HAVE_MMX2
1958 asm volatile(
1959 "lea (%0, %1), %%"REG_a" \n\t"
1960 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1961 // 0 1 2 3 4 5 6 7 8 9
1962 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1963
1964 "movq (%0), %%mm0 \n\t" //
1965 "movq (%%"REG_a", %1), %%mm2 \n\t" //
1966 "movq (%%"REG_a"), %%mm1 \n\t" //
1967 "movq %%mm0, %%mm3 \n\t"
1968 "pmaxub %%mm1, %%mm0 \n\t" //
1969 "pminub %%mm3, %%mm1 \n\t" //
1970 "pmaxub %%mm2, %%mm1 \n\t" //
1971 "pminub %%mm1, %%mm0 \n\t"
1972 "movq %%mm0, (%%"REG_a") \n\t"
1973
1974 "movq (%0, %1, 4), %%mm0 \n\t" //
1975 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" //
1976 "movq %%mm2, %%mm3 \n\t"
1977 "pmaxub %%mm1, %%mm2 \n\t" //
1978 "pminub %%mm3, %%mm1 \n\t" //
1979 "pmaxub %%mm0, %%mm1 \n\t" //
1980 "pminub %%mm1, %%mm2 \n\t"
1981 "movq %%mm2, (%%"REG_a", %1, 2) \n\t"
1982
1983 "movq (%%"REG_d"), %%mm2 \n\t" //
1984 "movq (%%"REG_d", %1), %%mm1 \n\t" //
1985 "movq %%mm2, %%mm3 \n\t"
1986 "pmaxub %%mm0, %%mm2 \n\t" //
1987 "pminub %%mm3, %%mm0 \n\t" //
1988 "pmaxub %%mm1, %%mm0 \n\t" //
1989 "pminub %%mm0, %%mm2 \n\t"
1990 "movq %%mm2, (%%"REG_d") \n\t"
1991
1992 "movq (%%"REG_d", %1, 2), %%mm2 \n\t" //
1993 "movq (%0, %1, 8), %%mm0 \n\t" //
1994 "movq %%mm2, %%mm3 \n\t"
1995 "pmaxub %%mm0, %%mm2 \n\t" //
1996 "pminub %%mm3, %%mm0 \n\t" //
1997 "pmaxub %%mm1, %%mm0 \n\t" //
1998 "pminub %%mm0, %%mm2 \n\t"
1999 "movq %%mm2, (%%"REG_d", %1, 2) \n\t"
2000
2001
2002 : : "r" (src), "r" ((long)stride)
2003 : "%"REG_a, "%"REG_d
2004 );
2005
2006 #else // MMX without MMX2
2007 asm volatile(
2008 "lea (%0, %1), %%"REG_a" \n\t"
2009 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
2010 // 0 1 2 3 4 5 6 7 8 9
2011 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2012 "pxor %%mm7, %%mm7 \n\t"
2013
2014 #define REAL_MEDIAN(a,b,c)\
2015 "movq " #a ", %%mm0 \n\t"\
2016 "movq " #b ", %%mm2 \n\t"\
2017 "movq " #c ", %%mm1 \n\t"\
2018 "movq %%mm0, %%mm3 \n\t"\
2019 "movq %%mm1, %%mm4 \n\t"\
2020 "movq %%mm2, %%mm5 \n\t"\
2021 "psubusb %%mm1, %%mm3 \n\t"\
2022 "psubusb %%mm2, %%mm4 \n\t"\
2023 "psubusb %%mm0, %%mm5 \n\t"\
2024 "pcmpeqb %%mm7, %%mm3 \n\t"\
2025 "pcmpeqb %%mm7, %%mm4 \n\t"\
2026 "pcmpeqb %%mm7, %%mm5 \n\t"\
2027 "movq %%mm3, %%mm6 \n\t"\
2028 "pxor %%mm4, %%mm3 \n\t"\
2029 "pxor %%mm5, %%mm4 \n\t"\
2030 "pxor %%mm6, %%mm5 \n\t"\
2031 "por %%mm3, %%mm1 \n\t"\
2032 "por %%mm4, %%mm2 \n\t"\
2033 "por %%mm5, %%mm0 \n\t"\
2034 "pand %%mm2, %%mm0 \n\t"\
2035 "pand %%mm1, %%mm0 \n\t"\
2036 "movq %%mm0, " #b " \n\t"
2037 #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c)
2038
2039 MEDIAN((%0) , (%%REGa) , (%%REGa, %1))
2040 MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4))
2041 MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1))
2042 MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8))
2043
2044 : : "r" (src), "r" ((long)stride)
2045 : "%"REG_a, "%"REG_d
2046 );
2047 #endif //HAVE_MMX2
2048 #else //HAVE_MMX
2049 int x, y;
2050 src+= 4*stride;
2051 // FIXME - there should be a way to do a few columns in parallel like w/mmx
2052 for(x=0; x<8; x++)
2053 {
2054 uint8_t *colsrc = src;
2055 for (y=0; y<4; y++)
2056 {
2057 int a, b, c, d, e, f;
2058 a = colsrc[0 ];
2059 b = colsrc[stride ];
2060 c = colsrc[stride*2];
2061 d = (a-b)>>31;
2062 e = (b-c)>>31;
2063 f = (c-a)>>31;
2064 colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f));
2065 colsrc += stride*2;
2066 }
2067 src++;
2068 }
2069 #endif //HAVE_MMX
2070 }
2071
2072 #ifdef HAVE_MMX
2073 /**
2074 * transposes and shift the given 8x8 Block into dst1 and dst2
2075 */
2076 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
2077 {
2078 asm(
2079 "lea (%0, %1), %%"REG_a" \n\t"
2080 // 0 1 2 3 4 5 6 7 8 9
2081 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2082 "movq (%0), %%mm0 \n\t" // 12345678
2083 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
2084 "movq %%mm0, %%mm2 \n\t" // 12345678
2085 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2086 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2087
2088 "movq (%%"REG_a", %1), %%mm1 \n\t"
2089 "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2090 "movq %%mm1, %%mm4 \n\t"
2091 "punpcklbw %%mm3, %%mm1 \n\t"
2092 "punpckhbw %%mm3, %%mm4 \n\t"
2093
2094 "movq %%mm0, %%mm3 \n\t"
2095 "punpcklwd %%mm1, %%mm0 \n\t"
2096 "punpckhwd %%mm1, %%mm3 \n\t"
2097 "movq %%mm2, %%mm1 \n\t"
2098 "punpcklwd %%mm4, %%mm2 \n\t"
2099 "punpckhwd %%mm4, %%mm1 \n\t"
2100
2101 "movd %%mm0, 128(%2) \n\t"
2102 "psrlq $32, %%mm0 \n\t"
2103 "movd %%mm0, 144(%2) \n\t"
2104 "movd %%mm3, 160(%2) \n\t"
2105 "psrlq $32, %%mm3 \n\t"
2106 "movd %%mm3, 176(%2) \n\t"
2107 "movd %%mm3, 48(%3) \n\t"
2108 "movd %%mm2, 192(%2) \n\t"
2109 "movd %%mm2, 64(%3) \n\t"
2110 "psrlq $32, %%mm2 \n\t"
2111 "movd %%mm2, 80(%3) \n\t"
2112 "movd %%mm1, 96(%3) \n\t"
2113 "psrlq $32, %%mm1 \n\t"
2114 "movd %%mm1, 112(%3) \n\t"
2115
2116 "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t"
2117
2118 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
2119 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh
2120 "movq %%mm0, %%mm2 \n\t" // 12345678
2121 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2122 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2123
2124 "movq (%%"REG_a", %1), %%mm1 \n\t"
2125 "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2126 "movq %%mm1, %%mm4 \n\t"
2127 "punpcklbw %%mm3, %%mm1 \n\t"
2128 "punpckhbw %%mm3, %%mm4 \n\t"
2129
2130 "movq %%mm0, %%mm3 \n\t"
2131 "punpcklwd %%mm1, %%mm0 \n\t"
2132 "punpckhwd %%mm1, %%mm3 \n\t"
2133 "movq %%mm2, %%mm1 \n\t"
2134 "punpcklwd %%mm4, %%mm2 \n\t"
2135 "punpckhwd %%mm4, %%mm1 \n\t"
2136
2137 "movd %%mm0, 132(%2) \n\t"
2138 "psrlq $32, %%mm0 \n\t"
2139 "movd %%mm0, 148(%2) \n\t"
2140 "movd %%mm3, 164(%2) \n\t"
2141 "psrlq $32, %%mm3 \n\t"
2142 "movd %%mm3, 180(%2) \n\t"
2143 "movd %%mm3, 52(%3) \n\t"
2144 "movd %%mm2, 196(%2) \n\t"
2145 "movd %%mm2, 68(%3) \n\t"
2146 "psrlq $32, %%mm2 \n\t"
2147 "movd %%mm2, 84(%3) \n\t"
2148 "movd %%mm1, 100(%3) \n\t"
2149 "psrlq $32, %%mm1 \n\t"
2150 "movd %%mm1, 116(%3) \n\t"
2151
2152
2153 :: "r" (src), "r" ((long)srcStride), "r" (dst1), "r" (dst2)
2154 : "%"REG_a
2155 );
2156 }
2157
2158 /**
2159 * transposes the given 8x8 block
2160 */
2161 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
2162 {
2163 asm(
2164 "lea (%0, %1), %%"REG_a" \n\t"
2165 "lea (%%"REG_a",%1,4), %%"REG_d" \n\t"
2166 // 0 1 2 3 4 5 6 7 8 9
2167 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2168 "movq (%2), %%mm0 \n\t" // 12345678
2169 "movq 16(%2), %%mm1 \n\t" // abcdefgh
2170 "movq %%mm0, %%mm2 \n\t" // 12345678
2171 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2172 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2173
2174 "movq 32(%2), %%mm1 \n\t"
2175 "movq 48(%2), %%mm3 \n\t"
2176 "movq %%mm1, %%mm4 \n\t"
2177 "punpcklbw %%mm3, %%mm1 \n\t"
2178 "punpckhbw %%mm3, %%mm4 \n\t"
2179
2180 "movq %%mm0, %%mm3 \n\t"
2181 "punpcklwd %%mm1, %%mm0 \n\t"
2182 "punpckhwd %%mm1, %%mm3 \n\t"
2183 "movq %%mm2, %%mm1 \n\t"
2184 "punpcklwd %%mm4, %%mm2 \n\t"
2185 "punpckhwd %%mm4, %%mm1 \n\t"
2186
2187 "movd %%mm0, (%0) \n\t"
2188 "psrlq $32, %%mm0 \n\t"
2189 "movd %%mm0, (%%"REG_a") \n\t"
2190 "movd %%mm3, (%%"REG_a", %1) \n\t"
2191 "psrlq $32, %%mm3 \n\t"
2192 "movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2193 "movd %%mm2, (%0, %1, 4) \n\t"
2194 "psrlq $32, %%mm2 \n\t"
2195 "movd %%mm2, (%%"REG_d") \n\t"
2196 "movd %%mm1, (%%"REG_d", %1) \n\t"
2197 "psrlq $32, %%mm1 \n\t"
2198 "movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2199
2200
2201 "movq 64(%2), %%mm0 \n\t" // 12345678
2202 "movq 80(%2), %%mm1 \n\t" // abcdefgh
2203 "movq %%mm0, %%mm2 \n\t" // 12345678
2204 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2205 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2206
2207 "movq 96(%2), %%mm1 \n\t"
2208 "movq 112(%2), %%mm3 \n\t"
2209 "movq %%mm1, %%mm4 \n\t"
2210 "punpcklbw %%mm3, %%mm1 \n\t"
2211 "punpckhbw %%mm3, %%mm4 \n\t"
2212
2213 "movq %%mm0, %%mm3 \n\t"
2214 "punpcklwd %%mm1, %%mm0 \n\t"
2215 "punpckhwd %%mm1, %%mm3 \n\t"
2216 "movq %%mm2, %%mm1 \n\t"
2217 "punpcklwd %%mm4, %%mm2 \n\t"
2218 "punpckhwd %%mm4, %%mm1 \n\t"
2219
2220 "movd %%mm0, 4(%0) \n\t"
2221 "psrlq $32, %%mm0 \n\t"
2222 "movd %%mm0, 4(%%"REG_a") \n\t"
2223 "movd %%mm3, 4(%%"REG_a", %1) \n\t"
2224 "psrlq $32, %%mm3 \n\t"
2225 "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t"
2226 "movd %%mm2, 4(%0, %1, 4) \n\t"
2227 "psrlq $32, %%mm2 \n\t"
2228 "movd %%mm2, 4(%%"REG_d") \n\t"
2229 "movd %%mm1, 4(%%"REG_d", %1) \n\t"
2230 "psrlq $32, %%mm1 \n\t"
2231 "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t"
2232
2233 :: "r" (dst), "r" ((long)dstStride), "r" (src)
2234 : "%"REG_a, "%"REG_d
2235 );
2236 }
2237 #endif //HAVE_MMX
2238 //static long test=0;
2239
2240 #ifndef HAVE_ALTIVEC
2241 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2242 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2243 {
2244 // to save a register (FIXME do this outside of the loops)
2245 tempBluredPast[127]= maxNoise[0];
2246 tempBluredPast[128]= maxNoise[1];
2247 tempBluredPast[129]= maxNoise[2];
2248
2249 #define FAST_L2_DIFF
2250 //#define L1_DIFF //u should change the thresholds too if u try that one
2251 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2252 asm volatile(
2253 "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride
2254 "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride
2255 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2256 // 0 1 2 3 4 5 6 7 8 9
2257 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
2258 //FIXME reorder?
2259 #ifdef L1_DIFF //needs mmx2
2260 "movq (%0), %%mm0 \n\t" // L0
2261 "psadbw (%1), %%mm0 \n\t" // |L0-R0|
2262 "movq (%0, %2), %%mm1 \n\t" // L1
2263 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2264 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2265 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2266 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2267 "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3|
2268
2269 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2270 "paddw %%mm1, %%mm0 \n\t"
2271 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2272 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2273 "paddw %%mm2, %%mm0 \n\t"
2274 "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5|
2275 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2276 "paddw %%mm3, %%mm0 \n\t"
2277 "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6|
2278 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2279 "paddw %%mm4, %%mm0 \n\t"
2280 "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7|
2281 "paddw %%mm5, %%mm6 \n\t"
2282 "paddw %%mm7, %%mm6 \n\t"
2283 "paddw %%mm6, %%mm0 \n\t"
2284 #else //L1_DIFF
2285 #if defined (FAST_L2_DIFF)
2286 "pcmpeqb %%mm7, %%mm7 \n\t"
2287 "movq "MANGLE(b80)", %%mm6 \n\t"
2288 "pxor %%mm0, %%mm0 \n\t"
2289 #define REAL_L2_DIFF_CORE(a, b)\
2290 "movq " #a ", %%mm5 \n\t"\
2291 "movq " #b ", %%mm2 \n\t"\
2292 "pxor %%mm7, %%mm2 \n\t"\
2293 PAVGB(%%mm2, %%mm5)\
2294 "paddb %%mm6, %%mm5 \n\t"\
2295 "movq %%mm5, %%mm2 \n\t"\
2296 "psllw $8, %%mm5 \n\t"\
2297 "pmaddwd %%mm5, %%mm5 \n\t"\
2298 "pmaddwd %%mm2, %%mm2 \n\t"\
2299 "paddd %%mm2, %%mm5 \n\t"\
2300 "psrld $14, %%mm5 \n\t"\
2301 "paddd %%mm5, %%mm0 \n\t"
2302
2303 #else //defined (FAST_L2_DIFF)
2304 "pxor %%mm7, %%mm7 \n\t"
2305 "pxor %%mm0, %%mm0 \n\t"
2306 #define REAL_L2_DIFF_CORE(a, b)\
2307 "movq " #a ", %%mm5 \n\t"\
2308 "movq " #b ", %%mm2 \n\t"\
2309 "movq %%mm5, %%mm1 \n\t"\
2310 "movq %%mm2, %%mm3 \n\t"\
2311 "punpcklbw %%mm7, %%mm5 \n\t"\
2312 "punpckhbw %%mm7, %%mm1 \n\t"\
2313 "punpcklbw %%mm7, %%mm2 \n\t"\
2314 "punpckhbw %%mm7, %%mm3 \n\t"\
2315 "psubw %%mm2, %%mm5 \n\t"\
2316 "psubw %%mm3, %%mm1 \n\t"\
2317 "pmaddwd %%mm5, %%mm5 \n\t"\
2318 "pmaddwd %%mm1, %%mm1 \n\t"\
2319 "paddd %%mm1, %%mm5 \n\t"\
2320 "paddd %%mm5, %%mm0 \n\t"
2321
2322 #endif //defined (FAST_L2_DIFF)
2323
2324 #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b)
2325
2326 L2_DIFF_CORE((%0) , (%1))
2327 L2_DIFF_CORE((%0, %2) , (%1, %2))
2328 L2_DIFF_CORE((%0, %2, 2) , (%1, %2, 2))
2329 L2_DIFF_CORE((%0, %%REGa) , (%1, %%REGa))
2330 L2_DIFF_CORE((%0, %2, 4) , (%1, %2, 4))
2331 L2_DIFF_CORE((%0, %%REGd) , (%1, %%REGd))
2332 L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2))
2333 L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc))
2334
2335 #endif //L1_DIFF
2336
2337 "movq %%mm0, %%mm4 \n\t"
2338 "psrlq $32, %%mm0 \n\t"
2339 "paddd %%mm0, %%mm4 \n\t"
2340 "movd %%mm4, %%ecx \n\t"
2341 "shll $2, %%ecx \n\t"
2342 "mov %3, %%"REG_d" \n\t"
2343 "addl -4(%%"REG_d"), %%ecx \n\t"
2344 "addl 4(%%"REG_d"), %%ecx \n\t"
2345 "addl -1024(%%"REG_d"), %%ecx \n\t"
2346 "addl $4, %%ecx \n\t"
2347 "addl 1024(%%"REG_d"), %%ecx \n\t"
2348 "shrl $3, %%ecx \n\t"
2349 "movl %%ecx, (%%"REG_d") \n\t"
2350
2351 // "mov %3, %%"REG_c" \n\t"
2352 // "mov %%"REG_c", test \n\t"
2353 // "jmp 4f \n\t"
2354 "cmpl 512(%%"REG_d"), %%ecx \n\t"
2355 " jb 2f \n\t"
2356 "cmpl 516(%%"REG_d"), %%ecx \n\t"
2357 " jb 1f \n\t"
2358
2359 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2360 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2361 "movq (%0), %%mm0 \n\t" // L0
2362 "movq (%0, %2), %%mm1 \n\t" // L1
2363 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2364 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2365 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2366 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2367 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2368 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2369 "movq %%mm0, (%1) \n\t" // L0
2370 "movq %%mm1, (%1, %2) \n\t" // L1
2371 "movq %%mm2, (%1, %2, 2) \n\t" // L2
2372 "movq %%mm3, (%1, %%"REG_a") \n\t" // L3
2373 "movq %%mm4, (%1, %2, 4) \n\t" // L4
2374 "movq %%mm5, (%1, %%"REG_d") \n\t" // L5
2375 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6
2376 "movq %%mm7, (%1, %%"REG_c") \n\t" // L7
2377 "jmp 4f \n\t"
2378
2379 "1: \n\t"
2380 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2381 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2382 "movq (%0), %%mm0 \n\t" // L0
2383 PAVGB((%1), %%mm0) // L0
2384 "movq (%0, %2), %%mm1 \n\t" // L1
2385 PAVGB((%1, %2), %%mm1) // L1
2386 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2387 PAVGB((%1, %2, 2), %%mm2) // L2
2388 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2389 PAVGB((%1, %%REGa), %%mm3) // L3
2390 "movq (%0, %2, 4), %%mm4 \n\t" // L4
2391 PAVGB((%1, %2, 4), %%mm4) // L4
2392 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2393 PAVGB((%1, %%REGd), %%mm5) // L5
2394 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2395 PAVGB((%1, %%REGa, 2), %%mm6) // L6
2396 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2397 PAVGB((%1, %%REGc), %%mm7) // L7
2398 "movq %%mm0, (%1) \n\t" // R0
2399 "movq %%mm1, (%1, %2) \n\t" // R1
2400 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2401 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2402 "movq %%mm4, (%1, %2, 4) \n\t" // R4
2403 "movq %%mm5, (%1, %%"REG_d") \n\t" // R5
2404 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6
2405 "movq %%mm7, (%1, %%"REG_c") \n\t" // R7
2406 "movq %%mm0, (%0) \n\t" // L0
2407 "movq %%mm1, (%0, %2) \n\t" // L1
2408 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2409 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2410 "movq %%mm4, (%0, %2, 4) \n\t" // L4
2411 "movq %%mm5, (%0, %%"REG_d") \n\t" // L5
2412 "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6
2413 "movq %%mm7, (%0, %%"REG_c") \n\t" // L7
2414 "jmp 4f \n\t"
2415
2416 "2: \n\t"
2417 "cmpl 508(%%"REG_d"), %%ecx \n\t"
2418 " jb 3f \n\t"
2419
2420 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2421 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2422 "movq (%0), %%mm0 \n\t" // L0
2423 "movq (%0, %2), %%mm1 \n\t" // L1
2424 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2425 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2426 "movq (%1), %%mm4 \n\t" // R0
2427 "movq (%1, %2), %%mm5 \n\t" // R1
2428 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2429 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2430 PAVGB(%%mm4, %%mm0)
2431 PAVGB(%%mm5, %%mm1)
2432 PAVGB(%%mm6, %%mm2)
2433 PAVGB(%%mm7, %%mm3)
2434 PAVGB(%%mm4, %%mm0)
2435 PAVGB(%%mm5, %%mm1)
2436 PAVGB(%%mm6, %%mm2)
2437 PAVGB(%%mm7, %%mm3)
2438 "movq %%mm0, (%1) \n\t" // R0
2439 "movq %%mm1, (%1, %2) \n\t" // R1
2440 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2441 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2442 "movq %%mm0, (%0) \n\t" // L0
2443 "movq %%mm1, (%0, %2) \n\t" // L1
2444 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2445 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2446
2447 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2448 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2449 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2450 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2451 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2452 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2453 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2454 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2455 PAVGB(%%mm4, %%mm0)
2456 PAVGB(%%mm5, %%mm1)
2457 PAVGB(%%mm6, %%mm2)
2458 PAVGB(%%mm7, %%mm3)
2459 PAVGB(%%mm4, %%mm0)
2460 PAVGB(%%mm5, %%mm1)
2461 PAVGB(%%mm6, %%mm2)
2462 PAVGB(%%mm7, %%mm3)
2463 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2464 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2465 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2466 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2467 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2468 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2469 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2470 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2471 "jmp 4f \n\t"
2472
2473 "3: \n\t"
2474 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2475 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2476 "movq (%0), %%mm0 \n\t" // L0
2477 "movq (%0, %2), %%mm1 \n\t" // L1
2478 "movq (%0, %2, 2), %%mm2 \n\t" // L2
2479 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2480 "movq (%1), %%mm4 \n\t" // R0
2481 "movq (%1, %2), %%mm5 \n\t" // R1
2482 "movq (%1, %2, 2), %%mm6 \n\t" // R2
2483 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2484 PAVGB(%%mm4, %%mm0)
2485 PAVGB(%%mm5, %%mm1)
2486 PAVGB(%%mm6, %%mm2)
2487 PAVGB(%%mm7, %%mm3)
2488 PAVGB(%%mm4, %%mm0)
2489 PAVGB(%%mm5, %%mm1)
2490 PAVGB(%%mm6, %%mm2)
2491 PAVGB(%%mm7, %%mm3)
2492 PAVGB(%%mm4, %%mm0)
2493 PAVGB(%%mm5, %%mm1)
2494 PAVGB(%%mm6, %%mm2)
2495 PAVGB(%%mm7, %%mm3)
2496 "movq %%mm0, (%1) \n\t" // R0
2497 "movq %%mm1, (%1, %2) \n\t" // R1
2498 "movq %%mm2, (%1, %2, 2) \n\t" // R2
2499 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2500 "movq %%mm0, (%0) \n\t" // L0
2501 "movq %%mm1, (%0, %2) \n\t" // L1
2502 "movq %%mm2, (%0, %2, 2) \n\t" // L2
2503 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2504
2505 "movq (%0, %2, 4), %%mm0 \n\t" // L4
2506 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2507 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2508 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2509 "movq (%1, %2, 4), %%mm4 \n\t" // R4
2510 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2511 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2512 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2513 PAVGB(%%mm4, %%mm0)
2514 PAVGB(%%mm5, %%mm1)
2515 PAVGB(%%mm6, %%mm2)
2516 PAVGB(%%mm7, %%mm3)
2517 PAVGB(%%mm4, %%mm0)
2518 PAVGB(%%mm5, %%mm1)
2519 PAVGB(%%mm6, %%mm2)
2520 PAVGB(%%mm7, %%mm3)
2521 PAVGB(%%mm4, %%mm0)
2522 PAVGB(%%mm5, %%mm1)
2523 PAVGB(%%mm6, %%mm2)
2524 PAVGB(%%mm7, %%mm3)
2525 "movq %%mm0, (%1, %2, 4) \n\t" // R4
2526 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2527 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2528 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2529 "movq %%mm0, (%0, %2, 4) \n\t" // L4
2530 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2531 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2532 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2533
2534 "4: \n\t"
2535
2536 :: "r" (src), "r" (tempBlured), "r"((long)stride), "m" (tempBluredPast)
2537 : "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2538 );
2539 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2540 {
2541 int y;
2542 int d=0;
2543 // int sysd=0;
2544 int i;
2545
2546 for(y=0; y<8; y++)
2547 {
2548 int x;
2549 for(x=0; x<8; x++)
2550 {
2551 int ref= tempBlured[ x + y*stride ];
2552 int cur= src[ x + y*stride ];
2553 int d1=ref - cur;
2554 // if(x==0 || x==7) d1+= d1>>1;
2555 // if(y==0 || y==7) d1+= d1>>1;
2556 // d+= FFABS(d1);
2557 d+= d1*d1;
2558 // sysd+= d1;
2559 }
2560 }
2561 i=d;
2562 d= (
2563 4*d
2564 +(*(tempBluredPast-256))
2565 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
2566 +(*(tempBluredPast+256))
2567 +4)>>3;
2568 *tempBluredPast=i;
2569 // ((*tempBluredPast)*3 + d + 2)>>2;
2570
2571 /*
2572 Switch between
2573 1 0 0 0 0 0 0 (0)
2574 64 32 16 8 4 2 1 (1)
2575 64 48 36 27 20 15 11 (33) (approx)
2576 64 56 49 43 37 33 29 (200) (approx)
2577 */
2578 if(d > maxNoise[1])
2579 {
2580 if(d < maxNoise[2])
2581 {
2582 for(y=0; y<8; y++)
2583 {
2584 int x;
2585 for(x=0; x<8; x++)
2586 {
2587 int ref= tempBlured[ x + y*stride ];
2588 int cur= src[ x + y*stride ];
2589 tempBlured[ x + y*stride ]=
2590 src[ x + y*stride ]=
2591 (ref + cur + 1)>>1;
2592 }
2593 }
2594 }
2595 else
2596 {
2597 for(y=0; y<8; y++)
2598 {
2599 int x;
2600 for(x=0; x<8; x++)
2601 {
2602 tempBlured[ x + y*stride ]= src[ x + y*stride ];
2603 }
2604 }
2605 }
2606 }
2607 else
2608 {
2609 if(d < maxNoise[0])
2610 {
2611 for(y=0; y<8; y++)
2612 {
2613 int x;
2614 for(x=0; x<8; x++)
2615 {
2616 int ref= tempBlured[ x + y*stride ];
2617 int cur= src[ x + y*stride ];
2618 tempBlured[ x + y*stride ]=
2619 src[ x + y*stride ]=
2620 (ref*7 + cur + 4)>>3;
2621 }
2622 }
2623 }
2624 else
2625 {
2626 for(y=0; y<8; y++)
2627 {
2628 int x;
2629 for(x=0; x<8; x++)
2630 {
2631 int ref= tempBlured[ x + y*stride ];
2632 int cur= src[ x + y*stride ];
2633 tempBlured[ x + y*stride ]=
2634 src[ x + y*stride ]=
2635 (ref*3 + cur + 2)>>2;
2636 }
2637 }
2638 }
2639 }
2640 }
2641 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2642 }
2643 #endif //HAVE_ALTIVEC
2644
2645 #ifdef HAVE_MMX
2646 /**
2647 * accurate deblock filter
2648 */
2649 static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){
2650 int64_t dc_mask, eq_mask, both_masks;
2651 int64_t sums[10*8*2];
2652 src+= step*3; // src points to begin of the 8x8 Block
2653 //START_TIMER
2654 asm volatile(
2655 "movq %0, %%mm7 \n\t"
2656 "movq %1, %%mm6 \n\t"
2657 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
2658 );
2659
2660 asm volatile(
2661 "lea (%2, %3), %%"REG_a" \n\t"
2662 // 0 1 2 3 4 5 6 7 8 9
2663 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
2664
2665 "movq (%2), %%mm0 \n\t"
2666 "movq (%%"REG_a"), %%mm1 \n\t"
2667 "movq %%mm1, %%mm3 \n\t"
2668 "movq %%mm1, %%mm4 \n\t"
2669 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
2670 "paddb %%mm7, %%mm0 \n\t"
2671 "pcmpgtb %%mm6, %%mm0 \n\t"
2672
2673 "movq (%%"REG_a",%3), %%mm2 \n\t"
2674 PMAXUB(%%mm2, %%mm4)
2675 PMINUB(%%mm2, %%mm3, %%mm5)
2676 "psubb %%mm2, %%mm1 \n\t"
2677 "paddb %%mm7, %%mm1 \n\t"
2678 "pcmpgtb %%mm6, %%mm1 \n\t"
2679 "paddb %%mm1, %%mm0 \n\t"
2680
2681 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2682 PMAXUB(%%mm1, %%mm4)
2683 PMINUB(%%mm1, %%mm3, %%mm5)
2684 "psubb %%mm1, %%mm2 \n\t"
2685 "paddb %%mm7, %%mm2 \n\t"
2686 "pcmpgtb %%mm6, %%mm2 \n\t"
2687 "paddb %%mm2, %%mm0 \n\t"
2688
2689 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
2690
2691 "movq (%2, %3, 4), %%mm2 \n\t"
2692 PMAXUB(%%mm2, %%mm4)
2693 PMINUB(%%mm2, %%mm3, %%mm5)
2694 "psubb %%mm2, %%mm1 \n\t"
2695 "paddb %%mm7, %%mm1 \n\t"
2696 "pcmpgtb %%mm6, %%mm1 \n\t"
2697 "paddb %%mm1, %%mm0 \n\t"
2698
2699 "movq (%%"REG_a"), %%mm1 \n\t"
2700 PMAXUB(%%mm1, %%mm4)
2701 PMINUB(%%mm1, %%mm3, %%mm5)
2702 "psubb %%mm1, %%mm2 \n\t"
2703 "paddb %%mm7, %%mm2 \n\t"
2704 "pcmpgtb %%mm6, %%mm2 \n\t"
2705 "paddb %%mm2, %%mm0 \n\t"
2706
2707 "movq (%%"REG_a", %3), %%mm2 \n\t"
2708 PMAXUB(%%mm2, %%mm4)
2709 PMINUB(%%mm2, %%mm3, %%mm5)
2710 "psubb %%mm2, %%mm1 \n\t"
2711 "paddb %%mm7, %%mm1 \n\t"
2712 "pcmpgtb %%mm6, %%mm1 \n\t"
2713 "paddb %%mm1, %%mm0 \n\t"
2714
2715 "movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2716 PMAXUB(%%mm1, %%mm4)
2717 PMINUB(%%mm1, %%mm3, %%mm5)
2718 "psubb %%mm1, %%mm2 \n\t"
2719 "paddb %%mm7, %%mm2 \n\t"
2720 "pcmpgtb %%mm6, %%mm2 \n\t"
2721 "paddb %%mm2, %%mm0 \n\t"
2722
2723 "movq (%2, %3, 8), %%mm2 \n\t"
2724 PMAXUB(%%mm2, %%mm4)
2725 PMINUB(%%mm2, %%mm3, %%mm5)
2726 "psubb %%mm2, %%mm1 \n\t"
2727 "paddb %%mm7, %%mm1 \n\t"
2728 "pcmpgtb %%mm6, %%mm1 \n\t"
2729 "paddb %%mm1, %%mm0 \n\t"
2730
2731 "movq (%%"REG_a", %3, 4), %%mm1 \n\t"
2732 "psubb %%mm1, %%mm2 \n\t"
2733 "paddb %%mm7, %%mm2 \n\t"
2734 "pcmpgtb %%mm6, %%mm2 \n\t"
2735 "paddb %%mm2, %%mm0 \n\t"
2736 "psubusb %%mm3, %%mm4 \n\t"
2737
2738 "pxor %%mm6, %%mm6 \n\t"
2739 "movq %4, %%mm7 \n\t" // QP,..., QP
2740 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
2741 "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0
2742 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2743 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2744 "movq %%mm7, %1 \n\t"
2745
2746 "movq %5, %%mm7 \n\t"
2747 "punpcklbw %%mm7, %%mm7 \n\t"
2748 "punpcklbw %%mm7, %%mm7 \n\t"
2749 "punpcklbw %%mm7, %%mm7 \n\t"
2750 "psubb %%mm0, %%mm6 \n\t"
2751 "pcmpgtb %%mm7, %%mm6 \n\t"
2752 "movq %%mm6, %0 \n\t"
2753
2754 : "=m" (eq_mask), "=m" (dc_mask)
2755 : "r" (src), "r" ((long)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
2756 : "%"REG_a
2757 );
2758
2759 both_masks = dc_mask & eq_mask;
2760
2761 if(both_masks){
2762 long offset= -8*step;
2763 int64_t *temp_sums= sums;
2764
2765 asm volatile(
2766 "movq %2, %%mm0 \n\t" // QP,..., QP
2767 "pxor %%mm4, %%mm4 \n\t"
2768
2769 "movq (%0), %%mm6 \n\t"
2770 "movq (%0, %1), %%mm5 \n\t"
2771 "movq %%mm5, %%mm1 \n\t"
2772 "movq %%mm6, %%mm2 \n\t"
2773 "psubusb %%mm6, %%mm5 \n\t"
2774 "psubusb %%mm1, %%mm2 \n\t"
2775 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2776 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2777 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2778
2779 "pxor %%mm6, %%mm1 \n\t"
2780 "pand %%mm0, %%mm1 \n\t"
2781 "pxor %%mm1, %%mm6 \n\t"
2782 // 0:QP 6:First
2783
2784 "movq (%0, %1, 8), %%mm5 \n\t"
2785 "add %1, %0 \n\t" // %0 points to line 1 not 0
2786 "movq (%0, %1, 8), %%mm7 \n\t"
2787 "movq %%mm5, %%mm1 \n\t"
2788 "movq %%mm7, %%mm2 \n\t"
2789 "psubusb %%mm7, %%mm5 \n\t"
2790 "psubusb %%mm1, %%mm2 \n\t"
2791 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2792 "movq %2, %%mm0 \n\t" // QP,..., QP
2793 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2794 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2795
2796 "pxor %%mm7, %%mm1 \n\t"
2797 "pand %%mm0, %%mm1 \n\t"
2798 "pxor %%mm1, %%mm7 \n\t"
2799
2800 "movq %%mm6, %%mm5 \n\t"
2801 "punpckhbw %%mm4, %%mm6 \n\t"
2802 "punpcklbw %%mm4, %%mm5 \n\t"
2803 // 4:0 5/6:First 7:Last
2804
2805 "movq %%mm5, %%mm0 \n\t"
2806 "movq %%mm6, %%mm1 \n\t"
2807 "psllw $2, %%mm0 \n\t"
2808 "psllw $2, %%mm1 \n\t"
2809 "paddw "MANGLE(w04)", %%mm0 \n\t"
2810 "paddw "MANGLE(w04)", %%mm1 \n\t"
2811
2812 #define NEXT\
2813 "movq (%0), %%mm2 \n\t"\
2814 "movq (%0), %%mm3 \n\t"\
2815 "add %1, %0 \n\t"\
2816 "punpcklbw %%mm4, %%mm2 \n\t"\
2817 "punpckhbw %%mm4, %%mm3 \n\t"\
2818 "paddw %%mm2, %%mm0 \n\t"\
2819 "paddw %%mm3, %%mm1 \n\t"
2820
2821 #define PREV\
2822 "movq (%0), %%mm2 \n\t"\
2823 "movq (%0), %%mm3 \n\t"\
2824 "add %1, %0 \n\t"\
2825 "punpcklbw %%mm4, %%mm2 \n\t"\
2826 "punpckhbw %%mm4, %%mm3 \n\t"\
2827 "psubw %%mm2, %%mm0 \n\t"\
2828 "psubw %%mm3, %%mm1 \n\t"
2829
2830
2831 NEXT //0
2832 NEXT //1
2833 NEXT //2
2834 "movq %%mm0, (%3) \n\t"
2835 "movq %%mm1, 8(%3) \n\t"
2836
2837 NEXT //3
2838 "psubw %%mm5, %%mm0 \n\t"
2839 "psubw %%mm6, %%mm1 \n\t"
2840 "movq %%mm0, 16(%3) \n\t"
2841 "movq %%mm1, 24(%3) \n\t"
2842
2843 NEXT //4
2844 "psubw %%mm5, %%mm0 \n\t"
2845 "psubw %%mm6, %%mm1 \n\t"
2846 "movq %%mm0, 32(%3) \n\t"
2847 "movq %%mm1, 40(%3) \n\t"
2848
2849 NEXT //5
2850 "psubw %%mm5, %%mm0 \n\t"
2851 "psubw %%mm6, %%mm1 \n\t"
2852 "movq %%mm0, 48(%3) \n\t"
2853 "movq %%mm1, 56(%3) \n\t"
2854
2855 NEXT //6
2856 "psubw %%mm5, %%mm0 \n\t"
2857 "psubw %%mm6, %%mm1 \n\t"
2858 "movq %%mm0, 64(%3) \n\t"
2859 "movq %%mm1, 72(%3) \n\t"
2860
2861 "movq %%mm7, %%mm6 \n\t"
2862 "punpckhbw %%mm4, %%mm7 \n\t"
2863 "punpcklbw %%mm4, %%mm6 \n\t"
2864
2865 NEXT //7
2866 "mov %4, %0 \n\t"
2867 "add %1, %0 \n\t"
2868 PREV //0
2869 "movq %%mm0, 80(%3) \n\t"
2870 "movq %%mm1, 88(%3) \n\t"
2871
2872 PREV //1
2873 "paddw %%mm6, %%mm0 \n\t"
2874 "paddw %%mm7, %%mm1 \n\t"
2875 "movq %%mm0, 96(%3) \n\t"
2876 "movq %%mm1, 104(%3) \n\t"
2877
2878 PREV //2
2879 "paddw %%mm6, %%mm0 \n\t"
2880 "paddw %%mm7, %%mm1 \n\t"
2881 "movq %%mm0, 112(%3) \n\t"
2882 "movq %%mm1, 120(%3) \n\t"
2883
2884 PREV //3
2885 "paddw %%mm6, %%mm0 \n\t"
2886 "paddw %%mm7, %%mm1 \n\t"
2887 "movq %%mm0, 128(%3) \n\t"
2888 "movq %%mm1, 136(%3) \n\t"
2889
2890 PREV //4
2891 "paddw %%mm6, %%mm0 \n\t"
2892 "paddw %%mm7, %%mm1 \n\t"
2893 "movq %%mm0, 144(%3) \n\t"
2894 "movq %%mm1, 152(%3) \n\t"
2895
2896 "mov %4, %0 \n\t" //FIXME
2897
2898 : "+&r"(src)
2899 : "r" ((long)step), "m" (c->pQPb), "r"(sums), "g"(src)
2900 );
2901
2902 src+= step; // src points to begin of the 8x8 Block
2903
2904 asm volatile(
2905 "movq %4, %%mm6 \n\t"
2906 "pcmpeqb %%mm5, %%mm5 \n\t"
2907 "pxor %%mm6, %%mm5 \n\t"
2908 "pxor %%mm7, %%mm7 \n\t"
2909
2910 "1: \n\t"
2911 "movq (%1), %%mm0 \n\t"
2912 "movq 8(%1), %%mm1 \n\t"
2913 "paddw 32(%1), %%mm0 \n\t"
2914 "paddw 40(%1), %%mm1 \n\t"
2915 "movq (%0, %3), %%mm2 \n\t"
2916 "movq %%mm2, %%mm3 \n\t"
2917 "movq %%mm2, %%mm4 \n\t"
2918 "punpcklbw %%mm7, %%mm2 \n\t"
2919 "punpckhbw %%mm7, %%mm3 \n\t"
2920 "paddw %%mm2, %%mm0 \n\t"
2921 "paddw %%mm3, %%mm1 \n\t"
2922 "paddw %%mm2, %%mm0 \n\t"
2923 "paddw %%mm3, %%mm1 \n\t"
2924 "psrlw $4, %%mm0 \n\t"
2925 "psrlw $4, %%mm1 \n\t"
2926 "packuswb %%mm1, %%mm0 \n\t"
2927 "pand %%mm6, %%mm0 \n\t"
2928 "pand %%mm5, %%mm4 \n\t"
2929 "por %%mm4, %%mm0 \n\t"
2930 "movq %%mm0, (%0, %3) \n\t"
2931 "add $16, %1 \n\t"
2932 "add %2, %0 \n\t"
2933 " js 1b \n\t"
2934
2935 : "+r"(offset), "+r"(temp_sums)
2936 : "r" ((long)step), "r"(src - offset), "m"(both_masks)
2937 );
2938 }else
2939 src+= step; // src points to begin of the 8x8 Block
2940
2941 if(eq_mask != -1LL){
2942 uint8_t *temp_src= src;
2943 asm volatile(
2944 "pxor %%mm7, %%mm7 \n\t"
2945 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
2946 "and "ALIGN_MASK", %%"REG_c" \n\t" // align
2947 // 0 1 2 3 4 5 6 7 8 9
2948 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1
2949
2950 "movq (%0), %%mm0 \n\t"
2951 "movq %%mm0, %%mm1 \n\t"
2952 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
2953 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
2954
2955 "movq (%0, %1), %%mm2 \n\t"
2956 "lea (%0, %1, 2), %%"REG_a" \n\t"
2957 "movq %%mm2, %%mm3 \n\t"
2958 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
2959 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
2960
2961 "movq (%%"REG_a"), %%mm4 \n\t"
2962 "movq %%mm4, %%mm5 \n\t"
2963 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
2964 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
2965
2966 "paddw %%mm0, %%mm0 \n\t" // 2L0
2967 "paddw %%mm1, %%mm1 \n\t" // 2H0
2968 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
2969 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
2970 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
2971 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
2972
2973 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
2974 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
2975 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
2976 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
2977
2978 "movq (%%"REG_a", %1), %%mm2 \n\t"
2979 "movq %%mm2, %%mm3 \n\t"
2980 "punpcklbw %%mm7, %%mm2 \n\t" // L3
2981 "punpckhbw %%mm7, %%mm3 \n\t" // H3
2982
2983 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
2984 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
2985 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2986 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2987 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2988 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2989
2990 "movq (%%"REG_a", %1, 2), %%mm0 \n\t"
2991 "movq %%mm0, %%mm1 \n\t"
2992 "punpcklbw %%mm7, %%mm0 \n\t" // L4
2993 "punpckhbw %%mm7, %%mm1 \n\t" // H4
2994
2995 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
2996 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
2997 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
2998 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
2999 "paddw %%mm4, %%mm4 \n\t" // 2L2
3000 "paddw %%mm5, %%mm5 \n\t" // 2H2
3001 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
3002 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
3003
3004 "lea (%%"REG_a", %1), %0 \n\t"
3005 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
3006 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
3007 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
3008 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
3009 //50 opcodes so far
3010 "movq (%0, %1, 2), %%mm2 \n\t"
3011 "movq %%mm2, %%mm3 \n\t"
3012 "punpcklbw %%mm7, %%mm2 \n\t" // L5
3013 "punpckhbw %%mm7, %%mm3 \n\t" // H5
3014 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
3015 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
3016 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
3017 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
3018
3019 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
3020 "punpcklbw %%mm7, %%mm6 \n\t" // L6
3021 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
3022 "movq (%%"REG_a", %1, 4), %%mm6 \n\t"
3023 "punpckhbw %%mm7, %%mm6 \n\t" // H6
3024 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
3025
3026 "paddw %%mm0, %%mm0 \n\t" // 2L4
3027 "paddw %%mm1, %%mm1 \n\t" // 2H4
3028 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
3029 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
3030
3031 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
3032 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
3033 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
3034 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
3035
3036 "movq (%0, %1, 4), %%mm2 \n\t"
3037 "movq %%mm2, %%mm3 \n\t"
3038 "punpcklbw %%mm7, %%mm2 \n\t" // L7
3039 "punpckhbw %%mm7, %%mm3 \n\t" // H7
3040
3041 "paddw %%mm2, %%mm2 \n\t" // 2L7
3042 "paddw %%mm3, %%mm3 \n\t" // 2H7
3043 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
3044 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
3045
3046 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
3047 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
3048
3049 #ifdef HAVE_MMX2
3050 "movq %%mm7, %%mm6 \n\t" // 0
3051 "psubw %%mm0, %%mm6 \n\t"
3052 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
3053 "movq %%mm7, %%mm6 \n\t" // 0
3054 "psubw %%mm1, %%mm6 \n\t"
3055 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
3056 "movq %%mm7, %%mm6 \n\t" // 0
3057 "psubw %%mm2, %%mm6 \n\t"
3058 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
3059 "movq %%mm7, %%mm6 \n\t" // 0
3060 "psubw %%mm3, %%mm6 \n\t"
3061 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3062 #else
3063 "movq %%mm7, %%mm6 \n\t" // 0
3064 "pcmpgtw %%mm0, %%mm6 \n\t"
3065 "pxor %%mm6, %%mm0 \n\t"
3066 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
3067 "movq %%mm7, %%mm6 \n\t" // 0
3068 "pcmpgtw %%mm1, %%mm6 \n\t"
3069 "pxor %%mm6, %%mm1 \n\t"
3070 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
3071 "movq %%mm7, %%mm6 \n\t" // 0
3072 "pcmpgtw %%mm2, %%mm6 \n\t"
3073 "pxor %%mm6, %%mm2 \n\t"
3074 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
3075 "movq %%mm7, %%mm6 \n\t" // 0
3076 "pcmpgtw %%mm3, %%mm6 \n\t"
3077 "pxor %%mm6, %%mm3 \n\t"
3078 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3079 #endif
3080
3081 #ifdef HAVE_MMX2
3082 "pminsw %%mm2, %%mm0 \n\t"
3083 "pminsw %%mm3, %%mm1 \n\t"
3084 #else
3085 "movq %%mm0, %%mm6 \n\t"
3086 "psubusw %%mm2, %%mm6 \n\t"
3087 "psubw %%mm6, %%mm0 \n\t"
3088 "movq %%mm1, %%mm6 \n\t"
3089 "psubusw %%mm3, %%mm6 \n\t"
3090 "psubw %%mm6, %%mm1 \n\t"
3091 #endif
3092
3093 "movd %2, %%mm2 \n\t" // QP
3094