← Back to branch summary

~siretart/xine-lib/ubuntu

« back to all changes in this revision

Viewing changes to src/libffmpeg/libavcodec/libpostproc/postprocess_template.c

  • Committer: Bazaar Package Importer
  • Author(s): Martin Pitt
  • Date: 2005-12-15 13:13:45 UTC
  • mfrom: (0.1.2 upstream)
  • Revision ID: james.westby@ubuntu.com-20051215131345-8n4osv1j7fy9c1s1
* SECURITY UPDATE: Fix arbitrary code execution with crafted PNG images in
  embedded ffmpeg copy.
* src/libffmpeg/libavcodec/utils.c, avcodec_default_get_buffer(): Apply
  upstream patch to fix buffer overflow on decoding of small PIX_FMT_PAL8
  PNG files.
* References:
  CVE-2005-4048
  http://mplayerhq.hu/pipermail/ffmpeg-devel/2005-November/005333.html
  http://www1.mplayerhq.hu/cgi-bin/cvsweb.cgi/ffmpeg/libavcodec/
  utils.c.diff?r1=1.161&r2=1.162&cvsroot=FFMpeg

Show diffs side-by-side

added added

removed removed

Lines of Context:
src/libffmpeg/libavcodec/libpostproc/postprocess_template.c
22
22
 */
23
23
 
24
24
 
 
25
#ifdef ARCH_X86_64
 
26
#  define REGa  rax
 
27
#  define REGc  rcx
 
28
#  define REGd  rdx
 
29
#  define REG_a  "rax"
 
30
#  define REG_c  "rcx"
 
31
#  define REG_d  "rdx"
 
32
#  define REG_SP "rsp"
 
33
#  define ALIGN_MASK "$0xFFFFFFFFFFFFFFF8"
 
34
#else
 
35
#  define REGa  eax
 
36
#  define REGc  ecx
 
37
#  define REGd  edx
 
38
#  define REG_a  "eax"
 
39
#  define REG_c  "ecx"
 
40
#  define REG_d  "edx"
 
41
#  define REG_SP "esp"
 
42
#  define ALIGN_MASK "$0xFFFFFFF8"
 
43
#endif
 
44
 
 
45
 
25
46
#undef PAVGB
26
47
#undef PMINUB
27
48
#undef PMAXUB
28
49
 
29
50
#ifdef HAVE_MMX2
30
 
#define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
 
51
#define REAL_PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
31
52
#elif defined (HAVE_3DNOW)
32
 
#define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
 
53
#define REAL_PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
33
54
#endif
 
55
#define PAVGB(a,b)  REAL_PAVGB(a,b)
34
56
 
35
57
#ifdef HAVE_MMX2
36
58
#define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
64
86
                );
65
87
                
66
88
asm volatile(
67
 
                "leal (%2, %3), %%eax                           \n\t"
 
89
                "lea (%2, %3), %%"REG_a"                        \n\t"
68
90
//      0       1       2       3       4       5       6       7       8       9
69
91
//      %1      eax     eax+%2  eax+2%2 %1+4%2  ecx     ecx+%2  ecx+2%2 %1+8%2  ecx+4%2
70
92
 
71
93
                "movq (%2), %%mm0                               \n\t"
72
 
                "movq (%%eax), %%mm1                            \n\t"
 
94
                "movq (%%"REG_a"), %%mm1                        \n\t"
73
95
                "movq %%mm0, %%mm3                              \n\t"
74
96
                "movq %%mm0, %%mm4                              \n\t"
75
97
                PMAXUB(%%mm1, %%mm4)
78
100
                "paddb %%mm7, %%mm0                             \n\t"
79
101
                "pcmpgtb %%mm6, %%mm0                           \n\t"
80
102
 
81
 
                "movq (%%eax,%3), %%mm2                         \n\t"
 
103
                "movq (%%"REG_a",%3), %%mm2                     \n\t"
82
104
                PMAXUB(%%mm2, %%mm4)
83
105
                PMINUB(%%mm2, %%mm3, %%mm5)
84
106
                "psubb %%mm2, %%mm1                             \n\t"
86
108
                "pcmpgtb %%mm6, %%mm1                           \n\t"
87
109
                "paddb %%mm1, %%mm0                             \n\t"
88
110
 
89
 
                "movq (%%eax, %3, 2), %%mm1                     \n\t"
 
111
                "movq (%%"REG_a", %3, 2), %%mm1                 \n\t"
90
112
                PMAXUB(%%mm1, %%mm4)
91
113
                PMINUB(%%mm1, %%mm3, %%mm5)
92
114
                "psubb %%mm1, %%mm2                             \n\t"
94
116
                "pcmpgtb %%mm6, %%mm2                           \n\t"
95
117
                "paddb %%mm2, %%mm0                             \n\t"
96
118
                
97
 
                "leal (%%eax, %3, 4), %%eax                     \n\t"
 
119
                "lea (%%"REG_a", %3, 4), %%"REG_a"              \n\t"
98
120
 
99
121
                "movq (%2, %3, 4), %%mm2                        \n\t"
100
122
                PMAXUB(%%mm2, %%mm4)
104
126
                "pcmpgtb %%mm6, %%mm1                           \n\t"
105
127
                "paddb %%mm1, %%mm0                             \n\t"
106
128
 
107
 
                "movq (%%eax), %%mm1                            \n\t"
 
129
                "movq (%%"REG_a"), %%mm1                        \n\t"
108
130
                PMAXUB(%%mm1, %%mm4)
109
131
                PMINUB(%%mm1, %%mm3, %%mm5)
110
132
                "psubb %%mm1, %%mm2                             \n\t"
112
134
                "pcmpgtb %%mm6, %%mm2                           \n\t"
113
135
                "paddb %%mm2, %%mm0                             \n\t"
114
136
 
115
 
                "movq (%%eax, %3), %%mm2                        \n\t"
 
137
                "movq (%%"REG_a", %3), %%mm2                    \n\t"
116
138
                PMAXUB(%%mm2, %%mm4)
117
139
                PMINUB(%%mm2, %%mm3, %%mm5)
118
140
                "psubb %%mm2, %%mm1                             \n\t"
120
142
                "pcmpgtb %%mm6, %%mm1                           \n\t"
121
143
                "paddb %%mm1, %%mm0                             \n\t"
122
144
 
123
 
                "movq (%%eax, %3, 2), %%mm1                     \n\t"
 
145
                "movq (%%"REG_a", %3, 2), %%mm1                 \n\t"
124
146
                PMAXUB(%%mm1, %%mm4)
125
147
                PMINUB(%%mm1, %%mm3, %%mm5)
126
148
                "psubb %%mm1, %%mm2                             \n\t"
152
174
                "movd %%mm4, %1                                 \n\t"
153
175
 
154
176
                : "=r" (numEq), "=r" (dcOk)
155
 
                : "r" (src), "r" (stride), "m" (c->pQPb)
156
 
                : "%eax"
 
177
                : "r" (src), "r" ((long)stride), "m" (c->pQPb)
 
178
                : "%"REG_a
157
179
                );
158
180
 
159
181
        numEq= (-numEq) &0xFF;
194
216
                "por %%mm2, %%mm6                               \n\t"// First Line to Filter
195
217
 
196
218
                "movq (%0, %1, 8), %%mm5                        \n\t"
197
 
                "leal (%0, %1, 4), %%eax                        \n\t"
198
 
                "leal (%0, %1, 8), %%ecx                        \n\t"
199
 
                "subl %1, %%ecx                                 \n\t"
200
 
                "addl %1, %0                                    \n\t" // %0 points to line 1 not 0
 
219
                "lea (%0, %1, 4), %%"REG_a"                     \n\t"
 
220
                "lea (%0, %1, 8), %%"REG_c"                     \n\t"
 
221
                "sub %1, %%"REG_c"                              \n\t"
 
222
                "add %1, %0                                     \n\t" // %0 points to line 1 not 0
201
223
                "movq (%0, %1, 8), %%mm7                        \n\t"
202
224
                "movq %%mm5, %%mm1                              \n\t"
203
225
                "movq %%mm7, %%mm2                              \n\t"
225
247
 
226
248
                "movq (%0, %1, 4), %%mm2                        \n\t" //     1
227
249
                "movq %%mm2, %%mm5                              \n\t" //     1
228
 
                PAVGB((%%eax), %%mm2)                                 //    11  /2
 
250
                PAVGB((%%REGa), %%mm2)                                //    11  /2
229
251
                PAVGB((%0, %1, 2), %%mm2)                             //   211  /4
230
252
                "movq %%mm2, %%mm3                              \n\t" //   211  /4
231
253
                "movq (%0), %%mm4                               \n\t" // 1
237
259
                PAVGB(%%mm6, %%mm0)                                   //1 1     /2
238
260
                "movq %%mm4, %%mm3                              \n\t" // 1
239
261
                PAVGB((%0,%1,2), %%mm3)                               // 1 1    /2
240
 
                PAVGB((%%eax,%1,2), %%mm5)                            //     11 /2
241
 
                PAVGB((%%eax), %%mm5)                                 //    211 /4
 
262
                PAVGB((%%REGa,%1,2), %%mm5)                           //     11 /2
 
263
                PAVGB((%%REGa), %%mm5)                                //    211 /4
242
264
                PAVGB(%%mm5, %%mm3)                                   // 2 2211 /8
243
265
                PAVGB(%%mm0, %%mm3)                                   //4242211 /16
244
266
                "movq %%mm3, (%0,%1)                            \n\t" //  X
245
267
                // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
246
268
                PAVGB(%%mm4, %%mm6)                                   //11      /2
247
 
                "movq (%%ecx), %%mm0                            \n\t" //       1
248
 
                PAVGB((%%eax, %1, 2), %%mm0)                          //      11/2
 
269
                "movq (%%"REG_c"), %%mm0                        \n\t" //       1
 
270
                PAVGB((%%REGa, %1, 2), %%mm0)                         //      11/2
249
271
                "movq %%mm0, %%mm3                              \n\t" //      11/2
250
272
                PAVGB(%%mm1, %%mm0)                                   //  2   11/4
251
273
                PAVGB(%%mm6, %%mm0)                                   //222   11/8
253
275
                "movq (%0, %1, 2), %%mm2                        \n\t" //   1
254
276
                "movq %%mm0, (%0, %1, 2)                        \n\t" //   X
255
277
                // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
256
 
                "movq (%%eax, %1, 4), %%mm0                     \n\t" //        1
257
 
                PAVGB((%%ecx), %%mm0)                                 //       11       /2
 
278
                "movq (%%"REG_a", %1, 4), %%mm0                 \n\t" //        1
 
279
                PAVGB((%%REGc), %%mm0)                                //       11       /2
258
280
                PAVGB(%%mm0, %%mm6)                                   //11     11       /4
259
281
                PAVGB(%%mm1, %%mm4)                                   // 11             /2
260
282
                PAVGB(%%mm2, %%mm1)                                   //  11            /2
261
283
                PAVGB(%%mm1, %%mm6)                                   //1122   11       /8
262
284
                PAVGB(%%mm5, %%mm6)                                   //112242211       /16
263
 
                "movq (%%eax), %%mm5                            \n\t" //    1
264
 
                "movq %%mm6, (%%eax)                            \n\t" //    X
 
285
                "movq (%%"REG_a"), %%mm5                        \n\t" //    1
 
286
                "movq %%mm6, (%%"REG_a")                        \n\t" //    X
265
287
                // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
266
 
                "movq (%%eax, %1, 4), %%mm6                     \n\t" //        1
 
288
                "movq (%%"REG_a", %1, 4), %%mm6                 \n\t" //        1
267
289
                PAVGB(%%mm7, %%mm6)                                   //        11      /2
268
290
                PAVGB(%%mm4, %%mm6)                                   // 11     11      /4
269
291
                PAVGB(%%mm3, %%mm6)                                   // 11   2211      /8
276
298
                PAVGB(%%mm7, %%mm1)                                   //  11     2      /4
277
299
                PAVGB(%%mm4, %%mm5)                                   //    11          /2
278
300
                PAVGB(%%mm5, %%mm0)                                   //    11 11       /4
279
 
                "movq (%%eax, %1, 2), %%mm6                     \n\t" //      1
 
301
                "movq (%%"REG_a", %1, 2), %%mm6                 \n\t" //      1
280
302
                PAVGB(%%mm6, %%mm1)                                   //  11  4  2      /8
281
303
                PAVGB(%%mm0, %%mm1)                                   //  11224222      /16
282
 
                "movq %%mm1, (%%eax, %1, 2)                     \n\t" //      X
 
304
                "movq %%mm1, (%%"REG_a", %1, 2)                 \n\t" //      X
283
305
                // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
284
 
                PAVGB((%%ecx), %%mm2)                                 //   112 4        /8
285
 
                "movq (%%eax, %1, 4), %%mm0                     \n\t" //        1
 
306
                PAVGB((%%REGc), %%mm2)                                //   112 4        /8
 
307
                "movq (%%"REG_a", %1, 4), %%mm0                 \n\t" //        1
286
308
                PAVGB(%%mm0, %%mm6)                                   //      1 1       /2
287
309
                PAVGB(%%mm7, %%mm6)                                   //      1 12      /4
288
310
                PAVGB(%%mm2, %%mm6)                                   //   1122424      /4
289
 
                "movq %%mm6, (%%ecx)                            \n\t" //       X
 
311
                "movq %%mm6, (%%"REG_c")                        \n\t" //       X
290
312
                // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
291
313
                PAVGB(%%mm7, %%mm5)                                   //    11   2      /4
292
314
                PAVGB(%%mm7, %%mm5)                                   //    11   6      /8
293
315
 
294
316
                PAVGB(%%mm3, %%mm0)                                   //      112       /4
295
317
                PAVGB(%%mm0, %%mm5)                                   //    112246      /16
296
 
                "movq %%mm5, (%%eax, %1, 4)                     \n\t" //        X
297
 
                "subl %1, %0                                    \n\t"
 
318
                "movq %%mm5, (%%"REG_a", %1, 4)                 \n\t" //        X
 
319
                "sub %1, %0                                     \n\t"
298
320
 
299
321
                :
300
 
                : "r" (src), "r" (stride), "m" (c->pQPb)
301
 
                : "%eax", "%ecx"
 
322
                : "r" (src), "r" ((long)stride), "m" (c->pQPb)
 
323
                : "%"REG_a, "%"REG_c
302
324
        );
303
325
#else
304
326
        const int l1= stride;
317
339
                const int first= ABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
318
340
                const int last= ABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
319
341
 
320
 
                int sums[9];
321
 
                sums[0] = first + src[l1];
322
 
                sums[1] = src[l1] + src[l2];
323
 
                sums[2] = src[l2] + src[l3];
324
 
                sums[3] = src[l3] + src[l4];
325
 
                sums[4] = src[l4] + src[l5];
326
 
                sums[5] = src[l5] + src[l6];
327
 
                sums[6] = src[l6] + src[l7];
328
 
                sums[7] = src[l7] + src[l8];
329
 
                sums[8] = src[l8] + last;
 
342
                int sums[10];
 
343
                sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
 
344
                sums[1] = sums[0] - first  + src[l4];
 
345
                sums[2] = sums[1] - first  + src[l5];
 
346
                sums[3] = sums[2] - first  + src[l6];
 
347
                sums[4] = sums[3] - first  + src[l7];
 
348
                sums[5] = sums[4] - src[l1] + src[l8];
 
349
                sums[6] = sums[5] - src[l2] + last;
 
350
                sums[7] = sums[6] - src[l3] + last;
 
351
                sums[8] = sums[7] - src[l4] + last;
 
352
                sums[9] = sums[8] - src[l5] + last;
330
353
 
331
 
                src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
332
 
                src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
333
 
                src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
334
 
                src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
335
 
                src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
336
 
                src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
337
 
                src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4;
338
 
                src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
 
354
                src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
 
355
                src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
 
356
                src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
 
357
                src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
 
358
                src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
 
359
                src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
 
360
                src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
 
361
                src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
339
362
 
340
363
                src++;
341
364
        }
363
386
        asm volatile(
364
387
                "pxor %%mm7, %%mm7                              \n\t" // 0
365
388
                "movq "MANGLE(b80)", %%mm6                      \n\t" // MIN_SIGNED_BYTE
366
 
                "leal (%0, %1), %%eax                           \n\t"
367
 
                "leal (%%eax, %1, 4), %%ecx                     \n\t"
 
389
                "leal (%0, %1), %%"REG_a"                       \n\t"
 
390
                "leal (%%"REG_a", %1, 4), %%"REG_c"             \n\t"
368
391
//      0       1       2       3       4       5       6       7       8       9
369
392
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1 %0+8%1  ecx+4%1
370
393
                "movq "MANGLE(pQPb)", %%mm0                     \n\t" // QP,..., QP
374
397
                "pand "MANGLE(b3F)", %%mm0                      \n\t" // QP/4,..., QP/4
375
398
                "paddusb %%mm1, %%mm0                           \n\t" // QP*1.25 ...
376
399
                "movq (%0, %1, 4), %%mm2                        \n\t" // line 4
377
 
                "movq (%%ecx), %%mm3                            \n\t" // line 5
 
400
                "movq (%%"REG_c"), %%mm3                                \n\t" // line 5
378
401
                "movq %%mm2, %%mm4                              \n\t" // line 4
379
402
                "pcmpeqb %%mm5, %%mm5                           \n\t" // -1
380
403
                "pxor %%mm2, %%mm5                              \n\t" // -line 4 - 1
392
415
//              "psubb %%mm6, %%mm2                             \n\t"
393
416
                "movq %%mm2, (%0,%1, 4)                         \n\t"
394
417
 
395
 
                "movq (%%ecx), %%mm2                            \n\t"
 
418
                "movq (%%"REG_c"), %%mm2                                \n\t"
396
419
//              "paddb %%mm6, %%mm2                             \n\t" // line 5 + 0x80
397
420
                "psubb %%mm5, %%mm2                             \n\t"
398
421
//              "psubb %%mm6, %%mm2                             \n\t"
399
 
                "movq %%mm2, (%%ecx)                            \n\t"
 
422
                "movq %%mm2, (%%"REG_c")                                \n\t"
400
423
 
401
424
                "paddb %%mm6, %%mm5                             \n\t"
402
425
                "psrlw $2, %%mm5                                \n\t"
403
426
                "pand "MANGLE(b3F)", %%mm5                      \n\t"
404
427
                "psubb "MANGLE(b20)", %%mm5                     \n\t" // (l5-l4)/8
405
428
 
406
 
                "movq (%%eax, %1, 2), %%mm2                     \n\t"
 
429
                "movq (%%"REG_a", %1, 2), %%mm2                 \n\t"
407
430
                "paddb %%mm6, %%mm2                             \n\t" // line 3 + 0x80
408
431
                "paddsb %%mm5, %%mm2                            \n\t"
409
432
                "psubb %%mm6, %%mm2                             \n\t"
410
 
                "movq %%mm2, (%%eax, %1, 2)                     \n\t"
 
433
                "movq %%mm2, (%%"REG_a", %1, 2)                 \n\t"
411
434
 
412
 
                "movq (%%ecx, %1), %%mm2                        \n\t"
 
435
                "movq (%%"REG_c", %1), %%mm2                    \n\t"
413
436
                "paddb %%mm6, %%mm2                             \n\t" // line 6 + 0x80
414
437
                "psubsb %%mm5, %%mm2                            \n\t"
415
438
                "psubb %%mm6, %%mm2                             \n\t"
416
 
                "movq %%mm2, (%%ecx, %1)                        \n\t"
 
439
                "movq %%mm2, (%%"REG_c", %1)                    \n\t"
417
440
 
418
441
                :
419
 
                : "r" (src), "r" (stride)
420
 
                : "%eax", "%ecx"
 
442
                : "r" (src), "r" ((long)stride)
 
443
                : "%"REG_a, "%"REG_c
421
444
        );
422
445
#else
423
446
        const int l1= stride;
463
486
 
464
487
        asm volatile(
465
488
                "pxor %%mm7, %%mm7                              \n\t" // 0
466
 
                "leal (%0, %1), %%eax                           \n\t"
467
 
                "leal (%%eax, %1, 4), %%ecx                     \n\t"
 
489
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
490
                "lea (%%"REG_a", %1, 4), %%"REG_c"              \n\t"
468
491
//      0       1       2       3       4       5       6       7       8       9
469
492
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1 %0+8%1  ecx+4%1
470
 
                "movq (%%eax, %1, 2), %%mm0                     \n\t" // line 3
 
493
                "movq (%%"REG_a", %1, 2), %%mm0                 \n\t" // line 3
471
494
                "movq (%0, %1, 4), %%mm1                        \n\t" // line 4
472
495
                "movq %%mm1, %%mm2                              \n\t" // line 4
473
496
                "psubusb %%mm0, %%mm1                           \n\t"
474
497
                "psubusb %%mm2, %%mm0                           \n\t"
475
498
                "por %%mm1, %%mm0                               \n\t" // |l2 - l3|
476
 
                "movq (%%ecx), %%mm3                            \n\t" // line 5
477
 
                "movq (%%ecx, %1), %%mm4                        \n\t" // line 6
 
499
                "movq (%%"REG_c"), %%mm3                                \n\t" // line 5
 
500
                "movq (%%"REG_c", %1), %%mm4                    \n\t" // line 6
478
501
                "movq %%mm3, %%mm5                              \n\t" // line 5
479
502
                "psubusb %%mm4, %%mm3                           \n\t"
480
503
                "psubusb %%mm5, %%mm4                           \n\t"
506
529
                "pxor %%mm2, %%mm0                              \n\t"
507
530
                "movq %%mm0, (%0, %1, 4)                        \n\t" // line 4
508
531
 
509
 
                "movq (%%ecx), %%mm0                            \n\t" // line 5
 
532
                "movq (%%"REG_c"), %%mm0                        \n\t" // line 5
510
533
                "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
511
534
                "paddusb %%mm3, %%mm0                           \n\t"
512
535
                "pxor %%mm2, %%mm0                              \n\t"
513
 
                "movq %%mm0, (%%ecx)                            \n\t" // line 5
 
536
                "movq %%mm0, (%%"REG_c")                        \n\t" // line 5
514
537
 
515
538
                PAVGB(%%mm7, %%mm1)                                   // d/4
516
539
 
517
 
                "movq (%%eax, %1, 2), %%mm0                     \n\t" // line 3
 
540
                "movq (%%"REG_a", %1, 2), %%mm0                 \n\t" // line 3
518
541
                "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
519
542
                "psubusb %%mm1, %%mm0                           \n\t"
520
543
                "pxor %%mm2, %%mm0                              \n\t"
521
 
                "movq %%mm0, (%%eax, %1, 2)                     \n\t" // line 3
 
544
                "movq %%mm0, (%%"REG_a", %1, 2)                 \n\t" // line 3
522
545
 
523
 
                "movq (%%ecx, %1), %%mm0                        \n\t" // line 6
 
546
                "movq (%%"REG_c", %1), %%mm0                    \n\t" // line 6
524
547
                "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
525
548
                "paddusb %%mm1, %%mm0                           \n\t"
526
549
                "pxor %%mm2, %%mm0                              \n\t"
527
 
                "movq %%mm0, (%%ecx, %1)                        \n\t" // line 6
 
550
                "movq %%mm0, (%%"REG_c", %1)                    \n\t" // line 6
528
551
 
529
552
                PAVGB(%%mm7, %%mm1)                                   // d/8
530
553
 
531
 
                "movq (%%eax, %1), %%mm0                        \n\t" // line 2
 
554
                "movq (%%"REG_a", %1), %%mm0                    \n\t" // line 2
532
555
                "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
533
556
                "psubusb %%mm1, %%mm0                           \n\t"
534
557
                "pxor %%mm2, %%mm0                              \n\t"
535
 
                "movq %%mm0, (%%eax, %1)                        \n\t" // line 2
 
558
                "movq %%mm0, (%%"REG_a", %1)                    \n\t" // line 2
536
559
 
537
 
                "movq (%%ecx, %1, 2), %%mm0                     \n\t" // line 7
 
560
                "movq (%%"REG_c", %1, 2), %%mm0                 \n\t" // line 7
538
561
                "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
539
562
                "paddusb %%mm1, %%mm0                           \n\t"
540
563
                "pxor %%mm2, %%mm0                              \n\t"
541
 
                "movq %%mm0, (%%ecx, %1, 2)                     \n\t" // line 7
 
564
                "movq %%mm0, (%%"REG_c", %1, 2)                 \n\t" // line 7
542
565
 
543
566
                :
544
 
                : "r" (src), "r" (stride), "m" (co->pQPb)
545
 
                : "%eax", "%ecx"
 
567
                : "r" (src), "r" ((long)stride), "m" (co->pQPb)
 
568
                : "%"REG_a, "%"REG_c
546
569
        );
547
570
#else
548
571
 
607
630
 
608
631
#if 0 //sligtly more accurate and slightly slower
609
632
                "pxor %%mm7, %%mm7                              \n\t" // 0
610
 
                "leal (%0, %1), %%eax                           \n\t"
611
 
                "leal (%%eax, %1, 4), %%ecx                     \n\t"
 
633
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
634
                "lea (%%"REG_a", %1, 4), %%"REG_c"              \n\t"
612
635
//      0       1       2       3       4       5       6       7
613
636
//      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ecx+%1  ecx+2%1
614
637
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1
621
644
                PAVGB(%%mm1, %%mm0)                                   // ~(l2 + 2l0)/4
622
645
                PAVGB(%%mm2, %%mm0)                                   // ~(5l2 + 2l0)/8
623
646
 
624
 
                "movq (%%eax), %%mm1                            \n\t" // l1
625
 
                "movq (%%eax, %1, 2), %%mm3                     \n\t" // l3
 
647
                "movq (%%"REG_a"), %%mm1                        \n\t" // l1
 
648
                "movq (%%"REG_a", %1, 2), %%mm3                 \n\t" // l3
626
649
                "movq %%mm1, %%mm4                              \n\t" // l1
627
650
                PAVGB(%%mm7, %%mm1)                                   // ~l1/2
628
651
                PAVGB(%%mm3, %%mm1)                                   // ~(l1 + 2l3)/4
640
663
                PAVGB(%%mm2, %%mm0)                                   // ~(l4 + 2l2)/4
641
664
                PAVGB(%%mm4, %%mm0)                                   // ~(5l4 + 2l2)/8
642
665
 
643
 
                "movq (%%ecx), %%mm2                            \n\t" // l5
 
666
                "movq (%%"REG_c"), %%mm2                        \n\t" // l5
644
667
                "movq %%mm3, %%mm5                              \n\t" // l3
645
668
                PAVGB(%%mm7, %%mm3)                                   // ~l3/2
646
669
                PAVGB(%%mm2, %%mm3)                                   // ~(l3 + 2l5)/4
653
676
                "pcmpeqb %%mm7, %%mm0                           \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
654
677
// mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
655
678
 
656
 
                "movq (%%ecx, %1), %%mm6                        \n\t" // l6
 
679
                "movq (%%"REG_c", %1), %%mm6                    \n\t" // l6
657
680
                "movq %%mm6, %%mm5                              \n\t" // l6
658
681
                PAVGB(%%mm7, %%mm6)                                   // ~l6/2
659
682
                PAVGB(%%mm4, %%mm6)                                   // ~(l6 + 2l4)/4
660
683
                PAVGB(%%mm5, %%mm6)                                   // ~(5l6 + 2l4)/8
661
684
 
662
 
                "movq (%%ecx, %1, 2), %%mm5                     \n\t" // l7
 
685
                "movq (%%"REG_c", %1, 2), %%mm5                 \n\t" // l7
663
686
                "movq %%mm2, %%mm4                              \n\t" // l5
664
687
                PAVGB(%%mm7, %%mm2)                                   // ~l5/2
665
688
                PAVGB(%%mm5, %%mm2)                                   // ~(l5 + 2l7)/4
686
709
                "paddusb %%mm1, %%mm3                           \n\t"
687
710
//              "paddusb "MANGLE(b01)", %%mm3                   \n\t"
688
711
 
689
 
                "movq (%%eax, %1, 2), %%mm6                     \n\t" //l3
 
712
                "movq (%%"REG_a", %1, 2), %%mm6                 \n\t" //l3
690
713
                "movq (%0, %1, 4), %%mm5                        \n\t" //l4
691
714
                "movq (%0, %1, 4), %%mm4                        \n\t" //l4
692
715
                "psubusb %%mm6, %%mm5                           \n\t"
700
723
                "psubusb "MANGLE(b01)", %%mm3                   \n\t"
701
724
                PAVGB(%%mm7, %%mm3)
702
725
 
703
 
                "movq (%%eax, %1, 2), %%mm0                     \n\t"
 
726
                "movq (%%"REG_a", %1, 2), %%mm0                 \n\t"
704
727
                "movq (%0, %1, 4), %%mm2                        \n\t"
705
728
                "pxor %%mm6, %%mm0                              \n\t"
706
729
                "pxor %%mm6, %%mm2                              \n\t"
708
731
                "paddb %%mm3, %%mm2                             \n\t"
709
732
                "pxor %%mm6, %%mm0                              \n\t"
710
733
                "pxor %%mm6, %%mm2                              \n\t"
711
 
                "movq %%mm0, (%%eax, %1, 2)                     \n\t"
 
734
                "movq %%mm0, (%%"REG_a", %1, 2)                 \n\t"
712
735
                "movq %%mm2, (%0, %1, 4)                        \n\t"
713
736
#endif
714
737
 
715
 
                "leal (%0, %1), %%eax                           \n\t"
 
738
                "lea (%0, %1), %%"REG_a"                        \n\t"
716
739
                "pcmpeqb %%mm6, %%mm6                           \n\t" // -1
717
740
//      0       1       2       3       4       5       6       7
718
741
//      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ecx+%1  ecx+2%1
719
742
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1
720
743
 
721
744
 
722
 
                "movq (%%eax, %1, 2), %%mm1                     \n\t" // l3
 
745
                "movq (%%"REG_a", %1, 2), %%mm1                 \n\t" // l3
723
746
                "movq (%0, %1, 4), %%mm0                        \n\t" // l4
724
747
                "pxor %%mm6, %%mm1                              \n\t" // -l3-1
725
748
                PAVGB(%%mm1, %%mm0)                                   // -q+128 = (l4-l3+256)/2
726
749
// mm1=-l3-1, mm0=128-q
727
750
 
728
 
                "movq (%%eax, %1, 4), %%mm2                     \n\t" // l5
729
 
                "movq (%%eax, %1), %%mm3                        \n\t" // l2
 
751
                "movq (%%"REG_a", %1, 4), %%mm2                 \n\t" // l5
 
752
                "movq (%%"REG_a", %1), %%mm3                    \n\t" // l2
730
753
                "pxor %%mm6, %%mm2                              \n\t" // -l5-1
731
754
                "movq %%mm2, %%mm5                              \n\t" // -l5-1
732
755
                "movq "MANGLE(b80)", %%mm4                      \n\t" // 128
733
 
                "leal (%%eax, %1, 4), %%ecx                     \n\t"
 
756
                "lea (%%"REG_a", %1, 4), %%"REG_c"              \n\t"
734
757
                PAVGB(%%mm3, %%mm2)                                   // (l2-l5+256)/2
735
758
                PAVGB(%%mm0, %%mm4)                                   // ~(l4-l3)/4 + 128
736
759
                PAVGB(%%mm2, %%mm4)                                   // ~(l2-l5)/4 +(l4-l3)/8 + 128
737
760
                PAVGB(%%mm0, %%mm4)                                   // ~(l2-l5)/8 +5(l4-l3)/16 + 128
738
761
// mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
739
762
 
740
 
                "movq (%%eax), %%mm2                            \n\t" // l1
 
763
                "movq (%%"REG_a"), %%mm2                        \n\t" // l1
741
764
                "pxor %%mm6, %%mm2                              \n\t" // -l1-1
742
765
                PAVGB(%%mm3, %%mm2)                                   // (l2-l1+256)/2
743
766
                PAVGB((%0), %%mm1)                                    // (l0-l3+256)/2
747
770
                PAVGB(%%mm2, %%mm3)                                   // ~(l0-l3)/8 +5(l2-l1)/16 + 128
748
771
// mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
749
772
 
750
 
                PAVGB((%%ecx, %1), %%mm5)                             // (l6-l5+256)/2
751
 
                "movq (%%ecx, %1, 2), %%mm1                     \n\t" // l7
 
773
                PAVGB((%%REGc, %1), %%mm5)                            // (l6-l5+256)/2
 
774
                "movq (%%"REG_c", %1, 2), %%mm1                 \n\t" // l7
752
775
                "pxor %%mm6, %%mm1                              \n\t" // -l7-1
753
776
                PAVGB((%0, %1, 4), %%mm1)                             // (l4-l7+256)/2
754
777
                "movq "MANGLE(b80)", %%mm2                      \n\t" // 128
797
820
                "pxor %%mm1, %%mm7                              \n\t" // SIGN(d*q)
798
821
 
799
822
                "pand %%mm7, %%mm4                              \n\t"
800
 
                "movq (%%eax, %1, 2), %%mm0                     \n\t"
 
823
                "movq (%%"REG_a", %1, 2), %%mm0                 \n\t"
801
824
                "movq (%0, %1, 4), %%mm2                        \n\t"
802
825
                "pxor %%mm1, %%mm0                              \n\t"
803
826
                "pxor %%mm1, %%mm2                              \n\t"
805
828
                "psubb %%mm4, %%mm2                             \n\t"
806
829
                "pxor %%mm1, %%mm0                              \n\t"
807
830
                "pxor %%mm1, %%mm2                              \n\t"
808
 
                "movq %%mm0, (%%eax, %1, 2)                     \n\t"
 
831
                "movq %%mm0, (%%"REG_a", %1, 2)                 \n\t"
809
832
                "movq %%mm2, (%0, %1, 4)                        \n\t"
810
833
 
811
834
                :
812
 
                : "r" (src), "r" (stride), "m" (c->pQPb)
813
 
                : "%eax", "%ecx"
 
835
                : "r" (src), "r" ((long)stride), "m" (c->pQPb)
 
836
                : "%"REG_a, "%"REG_c
814
837
        );
815
838
 
816
839
/*
881
904
        src+= stride*4;
882
905
        asm volatile(
883
906
                "pxor %%mm7, %%mm7                              \n\t"
884
 
                "leal -40(%%esp), %%ecx                         \n\t" // make space for 4 8-byte vars
885
 
                "andl $0xFFFFFFF8, %%ecx                        \n\t" // align
 
907
                "lea -40(%%"REG_SP"), %%"REG_c"                 \n\t" // make space for 4 8-byte vars
 
908
                "and "ALIGN_MASK", %%"REG_c"                    \n\t" // align
886
909
//      0       1       2       3       4       5       6       7
887
910
//      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 edx+%1  edx+2%1
888
911
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1
893
916
                "punpckhbw %%mm7, %%mm1                         \n\t" // high part of line 0
894
917
 
895
918
                "movq (%0, %1), %%mm2                           \n\t"
896
 
                "leal (%0, %1, 2), %%eax                        \n\t"
 
919
                "lea (%0, %1, 2), %%"REG_a"                     \n\t"
897
920
                "movq %%mm2, %%mm3                              \n\t"
898
921
                "punpcklbw %%mm7, %%mm2                         \n\t" // low part of line 1
899
922
                "punpckhbw %%mm7, %%mm3                         \n\t" // high part of line 1
900
923
 
901
 
                "movq (%%eax), %%mm4                            \n\t"
 
924
                "movq (%%"REG_a"), %%mm4                        \n\t"
902
925
                "movq %%mm4, %%mm5                              \n\t"
903
926
                "punpcklbw %%mm7, %%mm4                         \n\t" // low part of line 2
904
927
                "punpckhbw %%mm7, %%mm5                         \n\t" // high part of line 2
915
938
                "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2
916
939
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2
917
940
 
918
 
                "movq (%%eax, %1), %%mm2                        \n\t"
 
941
                "movq (%%"REG_a", %1), %%mm2                    \n\t"
919
942
                "movq %%mm2, %%mm3                              \n\t"
920
943
                "punpcklbw %%mm7, %%mm2                         \n\t" // L3
921
944
                "punpckhbw %%mm7, %%mm3                         \n\t" // H3
924
947
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - H3
925
948
                "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - 2L3
926
949
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - 2H3
927
 
                "movq %%mm0, (%%ecx)                            \n\t" // 2L0 - 5L1 + 5L2 - 2L3
928
 
                "movq %%mm1, 8(%%ecx)                           \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 
950
                "movq %%mm0, (%%"REG_c")                        \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 
951
                "movq %%mm1, 8(%%"REG_c")                       \n\t" // 2H0 - 5H1 + 5H2 - 2H3
929
952
 
930
 
                "movq (%%eax, %1, 2), %%mm0                     \n\t"
 
953
                "movq (%%"REG_a", %1, 2), %%mm0                 \n\t"
931
954
                "movq %%mm0, %%mm1                              \n\t"
932
955
                "punpcklbw %%mm7, %%mm0                         \n\t" // L4
933
956
                "punpckhbw %%mm7, %%mm1                         \n\t" // H4
934
957
 
935
958
                "psubw %%mm0, %%mm2                             \n\t" // L3 - L4
936
959
                "psubw %%mm1, %%mm3                             \n\t" // H3 - H4
937
 
                "movq %%mm2, 16(%%ecx)                          \n\t" // L3 - L4
938
 
                "movq %%mm3, 24(%%ecx)                          \n\t" // H3 - H4
 
960
                "movq %%mm2, 16(%%"REG_c")                      \n\t" // L3 - L4
 
961
                "movq %%mm3, 24(%%"REG_c")                      \n\t" // H3 - H4
939
962
                "paddw %%mm4, %%mm4                             \n\t" // 2L2
940
963
                "paddw %%mm5, %%mm5                             \n\t" // 2H2
941
964
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - L3 + L4
942
965
                "psubw %%mm3, %%mm5                             \n\t" // 2H2 - H3 + H4
943
966
 
944
 
                "leal (%%eax, %1), %0                           \n\t"
 
967
                "lea (%%"REG_a", %1), %0                        \n\t"
945
968
                "psllw $2, %%mm2                                \n\t" // 4L3 - 4L4
946
969
                "psllw $2, %%mm3                                \n\t" // 4H3 - 4H4
947
970
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4
956
979
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - 2L5
957
980
                "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - 2H5
958
981
 
959
 
                "movq (%%eax, %1, 4), %%mm6                     \n\t"
 
982
                "movq (%%"REG_a", %1, 4), %%mm6                 \n\t"
960
983
                "punpcklbw %%mm7, %%mm6                         \n\t" // L6
961
984
                "psubw %%mm6, %%mm2                             \n\t" // L5 - L6
962
 
                "movq (%%eax, %1, 4), %%mm6                     \n\t"
 
985
                "movq (%%"REG_a", %1, 4), %%mm6                 \n\t"
963
986
                "punpckhbw %%mm7, %%mm6                         \n\t" // H6
964
987
                "psubw %%mm6, %%mm3                             \n\t" // H5 - H6
965
988
 
983
1006
                "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6 - 2L7
984
1007
                "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6 - 2H7
985
1008
 
986
 
                "movq (%%ecx), %%mm2                            \n\t" // 2L0 - 5L1 + 5L2 - 2L3
987
 
                "movq 8(%%ecx), %%mm3                           \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 
1009
                "movq (%%"REG_c"), %%mm2                        \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 
1010
                "movq 8(%%"REG_c"), %%mm3                       \n\t" // 2H0 - 5H1 + 5H2 - 2H3
988
1011
 
989
1012
#ifdef HAVE_MMX2
990
1013
                "movq %%mm7, %%mm6                              \n\t" // 0
1030
1053
                "psubw %%mm6, %%mm1                             \n\t"
1031
1054
#endif
1032
1055
 
 
1056
                "movd %2, %%mm2                                 \n\t" // QP
 
1057
                "punpcklbw %%mm7, %%mm2                         \n\t"
 
1058
 
1033
1059
                "movq %%mm7, %%mm6                              \n\t" // 0
1034
1060
                "pcmpgtw %%mm4, %%mm6                           \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1035
1061
                "pxor %%mm6, %%mm4                              \n\t"
1038
1064
                "pxor %%mm7, %%mm5                              \n\t"
1039
1065
                "psubw %%mm7, %%mm5                             \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1040
1066
// 100 opcodes
1041
 
                "movd %2, %%mm2                                 \n\t" // QP
1042
1067
                "psllw $3, %%mm2                                \n\t" // 8QP
1043
1068
                "movq %%mm2, %%mm3                              \n\t" // 8QP
1044
1069
                "pcmpgtw %%mm4, %%mm2                           \n\t"
1060
1085
                "psrlw $6, %%mm4                                \n\t"
1061
1086
                "psrlw $6, %%mm5                                \n\t"
1062
1087
 
1063
 
                "movq 16(%%ecx), %%mm0                          \n\t" // L3 - L4
1064
 
                "movq 24(%%ecx), %%mm1                          \n\t" // H3 - H4
 
1088
                "movq 16(%%"REG_c"), %%mm0                      \n\t" // L3 - L4
 
1089
                "movq 24(%%"REG_c"), %%mm1                      \n\t" // H3 - H4
1065
1090
 
1066
1091
                "pxor %%mm2, %%mm2                              \n\t"
1067
1092
                "pxor %%mm3, %%mm3                              \n\t"
1104
1129
                "movq %%mm0, (%0, %1)                           \n\t"
1105
1130
 
1106
1131
                : "+r" (src)
1107
 
                : "r" (stride), "m" (c->pQPb)
1108
 
                : "%eax", "%ecx"
 
1132
                : "r" ((long)stride), "m" (c->pQPb)
 
1133
                : "%"REG_a, "%"REG_c
1109
1134
        );
1110
1135
#else
1111
1136
        const int l1= stride;
1168
1193
                "packuswb %%mm0, %%mm0                          \n\t"
1169
1194
                "movq %%mm0, %3                                 \n\t"
1170
1195
 
1171
 
                "leal (%0, %1), %%eax                           \n\t"
1172
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
1196
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1197
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
1173
1198
                
1174
1199
//      0       1       2       3       4       5       6       7       8       9
1175
1200
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1
1176
1201
 
1177
1202
#undef FIND_MIN_MAX
1178
1203
#ifdef HAVE_MMX2
1179
 
#define FIND_MIN_MAX(addr)\
 
1204
#define REAL_FIND_MIN_MAX(addr)\
1180
1205
                "movq " #addr ", %%mm0                          \n\t"\
1181
1206
                "pminub %%mm0, %%mm7                            \n\t"\
1182
1207
                "pmaxub %%mm0, %%mm6                            \n\t"
1183
1208
#else
1184
 
#define FIND_MIN_MAX(addr)\
 
1209
#define REAL_FIND_MIN_MAX(addr)\
1185
1210
                "movq " #addr ", %%mm0                          \n\t"\
1186
1211
                "movq %%mm7, %%mm1                              \n\t"\
1187
1212
                "psubusb %%mm0, %%mm6                           \n\t"\
1189
1214
                "psubusb %%mm0, %%mm1                           \n\t"\
1190
1215
                "psubb %%mm1, %%mm7                             \n\t"
1191
1216
#endif
 
1217
#define FIND_MIN_MAX(addr)  REAL_FIND_MIN_MAX(addr)
1192
1218
 
1193
 
FIND_MIN_MAX((%%eax))
1194
 
FIND_MIN_MAX((%%eax, %1))
1195
 
FIND_MIN_MAX((%%eax, %1, 2))
 
1219
FIND_MIN_MAX((%%REGa))
 
1220
FIND_MIN_MAX((%%REGa, %1))
 
1221
FIND_MIN_MAX((%%REGa, %1, 2))
1196
1222
FIND_MIN_MAX((%0, %1, 4))
1197
 
FIND_MIN_MAX((%%edx))
1198
 
FIND_MIN_MAX((%%edx, %1))
1199
 
FIND_MIN_MAX((%%edx, %1, 2))
 
1223
FIND_MIN_MAX((%%REGd))
 
1224
FIND_MIN_MAX((%%REGd, %1))
 
1225
FIND_MIN_MAX((%%REGd, %1, 2))
1200
1226
FIND_MIN_MAX((%0, %1, 8))
1201
1227
 
1202
1228
                "movq %%mm7, %%mm4                              \n\t"
1249
1275
                "movd %%mm6, %%ecx                              \n\t"
1250
1276
                "cmpb "MANGLE(deringThreshold)", %%cl           \n\t"
1251
1277
                " jb 1f                                         \n\t"
1252
 
                "leal -24(%%esp), %%ecx                         \n\t"
1253
 
                "andl $0xFFFFFFF8, %%ecx                        \n\t" 
 
1278
                "lea -24(%%"REG_SP"), %%"REG_c"                 \n\t"
 
1279
                "and "ALIGN_MASK", %%"REG_c"                    \n\t" 
1254
1280
                PAVGB(%%mm0, %%mm7)                                   // a=(max + min)/2
1255
1281
                "punpcklbw %%mm7, %%mm7                         \n\t"
1256
1282
                "punpcklbw %%mm7, %%mm7                         \n\t"
1257
1283
                "punpcklbw %%mm7, %%mm7                         \n\t"
1258
 
                "movq %%mm7, (%%ecx)                            \n\t"
 
1284
                "movq %%mm7, (%%"REG_c")                        \n\t"
1259
1285
 
1260
1286
                "movq (%0), %%mm0                               \n\t" // L10
1261
1287
                "movq %%mm0, %%mm1                              \n\t" // L10
1280
1306
                "paddb %%mm2, %%mm0                             \n\t"
1281
1307
                "paddb %%mm3, %%mm0                             \n\t"
1282
1308
 
1283
 
                "movq (%%eax), %%mm2                            \n\t" // L11
 
1309
                "movq (%%"REG_a"), %%mm2                        \n\t" // L11
1284
1310
                "movq %%mm2, %%mm3                              \n\t" // L11
1285
1311
                "movq %%mm2, %%mm4                              \n\t" // L11
1286
1312
                "psllq $8, %%mm3                                \n\t"
1287
1313
                "psrlq $8, %%mm4                                \n\t"
1288
 
                "movd -4(%%eax), %%mm5                          \n\t"
1289
 
                "movd 8(%%eax), %%mm6                           \n\t"
 
1314
                "movd -4(%%"REG_a"), %%mm5                      \n\t"
 
1315
                "movd 8(%%"REG_a"), %%mm6                       \n\t"
1290
1316
                "psrlq $24, %%mm5                               \n\t"
1291
1317
                "psllq $56, %%mm6                               \n\t"
1292
1318
                "por %%mm5, %%mm3                               \n\t" // L01
1303
1329
                "paddb %%mm4, %%mm2                             \n\t"
1304
1330
                "paddb %%mm5, %%mm2                             \n\t"
1305
1331
// 0, 2, 3, 1
1306
 
#define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
 
1332
#define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1307
1333
                "movq " #src ", " #sx "                         \n\t" /* src[0] */\
1308
1334
                "movq " #sx ", " #lx "                          \n\t" /* src[0] */\
1309
1335
                "movq " #sx ", " #t0 "                          \n\t" /* src[0] */\
1319
1345
                PAVGB(t0, lx)                                         /* (src[-1] + src[+1])/2 */\
1320
1346
                PAVGB(sx, lx)                                 /* (src[-1] + 2src[0] + src[+1])/4 */\
1321
1347
                PAVGB(lx, pplx)                                      \
1322
 
                "movq " #lx ", 8(%%ecx)                         \n\t"\
1323
 
                "movq (%%ecx), " #lx "                          \n\t"\
 
1348
                "movq " #lx ", 8(%%"REG_c")                     \n\t"\
 
1349
                "movq (%%"REG_c"), " #lx "                      \n\t"\
1324
1350
                "psubusb " #lx ", " #t1 "                       \n\t"\
1325
1351
                "psubusb " #lx ", " #t0 "                       \n\t"\
1326
1352
                "psubusb " #lx ", " #sx "                       \n\t"\
1347
1373
                "pandn " #dst ", " #ppsx "                      \n\t"\
1348
1374
                "por " #pplx ", " #ppsx "                       \n\t"\
1349
1375
                "movq " #ppsx ", " #dst "                       \n\t"\
1350
 
                "movq 8(%%ecx), " #lx "                         \n\t"
 
1376
                "movq 8(%%"REG_c"), " #lx "                     \n\t"
1351
1377
 
 
1378
#define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
 
1379
   REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1)
1352
1380
/*
1353
1381
0000000
1354
1382
1111111
1365
1393
 
1366
1394
*/
1367
1395
//DERING_CORE(dst,src                  ,ppsx ,psx  ,sx   ,pplx ,plx  ,lx   ,t0   ,t1)
1368
 
DERING_CORE((%%eax),(%%eax, %1)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1369
 
DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1370
 
DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1371
 
DERING_CORE((%0, %1, 4),(%%edx)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1372
 
DERING_CORE((%%edx),(%%edx, %1)        ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1373
 
DERING_CORE((%%edx, %1), (%%edx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1374
 
DERING_CORE((%%edx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1375
 
DERING_CORE((%0, %1, 8),(%%edx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
 
1396
DERING_CORE((%%REGa),(%%REGa, %1)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
 
1397
DERING_CORE((%%REGa, %1),(%%REGa, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
 
1398
DERING_CORE((%%REGa, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
 
1399
DERING_CORE((%0, %1, 4),(%%REGd)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
 
1400
DERING_CORE((%%REGd),(%%REGd, %1)        ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
 
1401
DERING_CORE((%%REGd, %1), (%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
 
1402
DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
 
1403
DERING_CORE((%0, %1, 8),(%%REGd, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1376
1404
 
1377
1405
                "1:                     \n\t"
1378
 
                : : "r" (src), "r" (stride), "m" (c->pQPb), "m"(c->pQPb2)
1379
 
                : "%eax", "%edx", "%ecx"
 
1406
                : : "r" (src), "r" ((long)stride), "m" (c->pQPb), "m"(c->pQPb2)
 
1407
                : "%"REG_a, "%"REG_d, "%"REG_c
1380
1408
        );
1381
1409
#else
1382
1410
        int y;
1523
1551
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1524
1552
        src+= 4*stride;
1525
1553
        asm volatile(
1526
 
                "leal (%0, %1), %%eax                           \n\t"
1527
 
                "leal (%%eax, %1, 4), %%ecx                     \n\t"
 
1554
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1555
                "lea (%%"REG_a", %1, 4), %%"REG_c"              \n\t"
1528
1556
//      0       1       2       3       4       5       6       7       8       9
1529
1557
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1 %0+8%1  ecx+4%1
1530
1558
 
1531
1559
                "movq (%0), %%mm0                               \n\t"
1532
 
                "movq (%%eax, %1), %%mm1                        \n\t"
 
1560
                "movq (%%"REG_a", %1), %%mm1                    \n\t"
1533
1561
                PAVGB(%%mm1, %%mm0)
1534
 
                "movq %%mm0, (%%eax)                            \n\t"
 
1562
                "movq %%mm0, (%%"REG_a")                        \n\t"
1535
1563
                "movq (%0, %1, 4), %%mm0                        \n\t"
1536
1564
                PAVGB(%%mm0, %%mm1)
1537
 
                "movq %%mm1, (%%eax, %1, 2)                     \n\t"
1538
 
                "movq (%%ecx, %1), %%mm1                        \n\t"
 
1565
                "movq %%mm1, (%%"REG_a", %1, 2)                 \n\t"
 
1566
                "movq (%%"REG_c", %1), %%mm1                    \n\t"
1539
1567
                PAVGB(%%mm1, %%mm0)
1540
 
                "movq %%mm0, (%%ecx)                            \n\t"
 
1568
                "movq %%mm0, (%%"REG_c")                        \n\t"
1541
1569
                "movq (%0, %1, 8), %%mm0                        \n\t"
1542
1570
                PAVGB(%%mm0, %%mm1)
1543
 
                "movq %%mm1, (%%ecx, %1, 2)                     \n\t"
 
1571
                "movq %%mm1, (%%"REG_c", %1, 2)                 \n\t"
1544
1572
 
1545
 
                : : "r" (src), "r" (stride)
1546
 
                : "%eax", "%ecx"
 
1573
                : : "r" (src), "r" ((long)stride)
 
1574
                : "%"REG_a, "%"REG_c
1547
1575
        );
1548
1576
#else
1549
1577
        int a, b, x;
1576
1604
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1577
1605
        src+= stride*3;
1578
1606
        asm volatile(
1579
 
                "leal (%0, %1), %%eax                           \n\t"
1580
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
1581
 
                "leal (%%edx, %1, 4), %%ecx                     \n\t"
1582
 
                "addl %1, %%ecx                                 \n\t"
 
1607
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1608
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
 
1609
                "lea (%%"REG_d", %1, 4), %%"REG_c"              \n\t"
 
1610
                "add %1, %%"REG_c"                              \n\t"
1583
1611
                "pxor %%mm7, %%mm7                              \n\t"
1584
1612
//      0       1       2       3       4       5       6       7       8       9       10
1585
1613
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1 ecx
1586
1614
 
1587
 
#define DEINT_CUBIC(a,b,c,d,e)\
 
1615
#define REAL_DEINT_CUBIC(a,b,c,d,e)\
1588
1616
                "movq " #a ", %%mm0                             \n\t"\
1589
1617
                "movq " #b ", %%mm1                             \n\t"\
1590
1618
                "movq " #d ", %%mm2                             \n\t"\
1605
1633
                "psubw %%mm2, %%mm3                             \n\t"   /* H(9b + 9d - a - e)/16 */\
1606
1634
                "packuswb %%mm3, %%mm1                          \n\t"\
1607
1635
                "movq %%mm1, " #c "                             \n\t"
1608
 
 
1609
 
DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx, %1))
1610
 
DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%edx), (%%edx, %1), (%0, %1, 8))
1611
 
DEINT_CUBIC((%0, %1, 4), (%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%ecx))
1612
 
DEINT_CUBIC((%%edx, %1), (%0, %1, 8), (%%edx, %1, 4), (%%ecx), (%%ecx, %1, 2))
1613
 
 
1614
 
                : : "r" (src), "r" (stride)
1615
 
                : "%eax", "%edx", "ecx"
 
1636
#define DEINT_CUBIC(a,b,c,d,e)  REAL_DEINT_CUBIC(a,b,c,d,e)
 
1637
 
 
1638
DEINT_CUBIC((%0), (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4), (%%REGd, %1))
 
1639
DEINT_CUBIC((%%REGa, %1), (%0, %1, 4), (%%REGd), (%%REGd, %1), (%0, %1, 8))
 
1640
DEINT_CUBIC((%0, %1, 4), (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8), (%%REGc))
 
1641
DEINT_CUBIC((%%REGd, %1), (%0, %1, 8), (%%REGd, %1, 4), (%%REGc), (%%REGc, %1, 2))
 
1642
 
 
1643
                : : "r" (src), "r" ((long)stride)
 
1644
                : "%"REG_a, "%"REG_d, "%"REG_c
1616
1645
        );
1617
1646
#else
1618
1647
        int x;
1640
1669
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1641
1670
        src+= stride*4;
1642
1671
        asm volatile(
1643
 
                "leal (%0, %1), %%eax                           \n\t"
1644
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
1672
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1673
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
1645
1674
                "pxor %%mm7, %%mm7                              \n\t"
1646
1675
                "movq (%2), %%mm0                               \n\t"
1647
1676
//      0       1       2       3       4       5       6       7       8       9       10
1648
1677
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1 ecx
1649
1678
 
1650
 
#define DEINT_FF(a,b,c,d)\
 
1679
#define REAL_DEINT_FF(a,b,c,d)\
1651
1680
                "movq " #a ", %%mm1                             \n\t"\
1652
1681
                "movq " #b ", %%mm2                             \n\t"\
1653
1682
                "movq " #c ", %%mm3                             \n\t"\
1675
1704
                "packuswb %%mm4, %%mm1                          \n\t"\
1676
1705
                "movq %%mm1, " #b "                             \n\t"\
1677
1706
 
1678
 
DEINT_FF((%0)       , (%%eax)       , (%%eax, %1), (%%eax, %1, 2))
1679
 
DEINT_FF((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx)       )
1680
 
DEINT_FF((%0, %1, 4), (%%edx)       , (%%edx, %1), (%%edx, %1, 2))
1681
 
DEINT_FF((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%edx, %1, 4))
 
1707
#define DEINT_FF(a,b,c,d)  REAL_DEINT_FF(a,b,c,d)
 
1708
 
 
1709
DEINT_FF((%0)       ,  (%%REGa)       , (%%REGa, %1), (%%REGa, %1, 2))
 
1710
DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4),  (%%REGd)       )
 
1711
DEINT_FF((%0, %1, 4),  (%%REGd)       , (%%REGd, %1), (%%REGd, %1, 2))
 
1712
DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8),  (%%REGd, %1, 4))
1682
1713
 
1683
1714
                "movq %%mm0, (%2)                               \n\t"
1684
 
                : : "r" (src), "r" (stride), "r"(tmp)
1685
 
                : "%eax", "%edx"
 
1715
                : : "r" (src), "r" ((long)stride), "r"(tmp)
 
1716
                : "%"REG_a, "%"REG_d
1686
1717
        );
1687
1718
#else
1688
1719
        int x;
1718
1749
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1719
1750
        src+= stride*4;
1720
1751
        asm volatile(
1721
 
                "leal (%0, %1), %%eax                           \n\t"
1722
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
1752
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1753
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
1723
1754
                "pxor %%mm7, %%mm7                              \n\t"
1724
1755
                "movq (%2), %%mm0                               \n\t"
1725
1756
                "movq (%3), %%mm1                               \n\t"
1726
1757
//      0       1       2       3       4       5       6       7       8       9       10
1727
1758
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1 ecx
1728
1759
 
1729
 
#define DEINT_L5(t1,t2,a,b,c)\
 
1760
#define REAL_DEINT_L5(t1,t2,a,b,c)\
1730
1761
                "movq " #a ", %%mm2                             \n\t"\
1731
1762
                "movq " #b ", %%mm3                             \n\t"\
1732
1763
                "movq " #c ", %%mm4                             \n\t"\
1759
1790
                "packuswb %%mm5, %%mm2                          \n\t"\
1760
1791
                "movq %%mm2, " #a "                             \n\t"\
1761
1792
 
1762
 
DEINT_L5(%%mm0, %%mm1, (%0)          , (%%eax)       , (%%eax, %1)   )
1763
 
DEINT_L5(%%mm1, %%mm0, (%%eax)       , (%%eax, %1)   , (%%eax, %1, 2))
1764
 
DEINT_L5(%%mm0, %%mm1, (%%eax, %1)   , (%%eax, %1, 2), (%0, %1, 4)   )
1765
 
DEINT_L5(%%mm1, %%mm0, (%%eax, %1, 2), (%0, %1, 4)   , (%%edx)       )
1766
 
DEINT_L5(%%mm0, %%mm1, (%0, %1, 4)   , (%%edx)       , (%%edx, %1)   )  
1767
 
DEINT_L5(%%mm1, %%mm0, (%%edx)       , (%%edx, %1)   , (%%edx, %1, 2))
1768
 
DEINT_L5(%%mm0, %%mm1, (%%edx, %1)   , (%%edx, %1, 2), (%0, %1, 8)   )
1769
 
DEINT_L5(%%mm1, %%mm0, (%%edx, %1, 2), (%0, %1, 8)   , (%%edx, %1, 4))
 
1793
#define DEINT_L5(t1,t2,a,b,c)  REAL_DEINT_L5(t1,t2,a,b,c)
 
1794
 
 
1795
DEINT_L5(%%mm0, %%mm1, (%0)           , (%%REGa)       , (%%REGa, %1)   )
 
1796
DEINT_L5(%%mm1, %%mm0, (%%REGa)       , (%%REGa, %1)   , (%%REGa, %1, 2))
 
1797
DEINT_L5(%%mm0, %%mm1, (%%REGa, %1)   , (%%REGa, %1, 2), (%0, %1, 4)   )
 
1798
DEINT_L5(%%mm1, %%mm0, (%%REGa, %1, 2), (%0, %1, 4)    , (%%REGd)       )
 
1799
DEINT_L5(%%mm0, %%mm1, (%0, %1, 4)    , (%%REGd)       , (%%REGd, %1)   )  
 
1800
DEINT_L5(%%mm1, %%mm0, (%%REGd)       , (%%REGd, %1)   , (%%REGd, %1, 2))
 
1801
DEINT_L5(%%mm0, %%mm1, (%%REGd, %1)   , (%%REGd, %1, 2), (%0, %1, 8)   )
 
1802
DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8)    , (%%REGd, %1, 4))
1770
1803
 
1771
1804
                "movq %%mm0, (%2)                               \n\t"
1772
1805
                "movq %%mm1, (%3)                               \n\t"
1773
 
                : : "r" (src), "r" (stride), "r"(tmp), "r"(tmp2)
1774
 
                : "%eax", "%edx"
 
1806
                : : "r" (src), "r" ((long)stride), "r"(tmp), "r"(tmp2)
 
1807
                : "%"REG_a, "%"REG_d
1775
1808
        );
1776
1809
#else
1777
1810
        int x;
1818
1851
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1819
1852
        src+= 4*stride;
1820
1853
        asm volatile(
1821
 
                "leal (%0, %1), %%eax                           \n\t"
1822
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
1854
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1855
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
1823
1856
//      0       1       2       3       4       5       6       7       8       9
1824
1857
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1
1825
1858
 
1826
1859
                "movq (%2), %%mm0                               \n\t" // L0
1827
 
                "movq (%%eax), %%mm1                            \n\t" // L2
 
1860
                "movq (%%"REG_a"), %%mm1                        \n\t" // L2
1828
1861
                PAVGB(%%mm1, %%mm0)                                   // L0+L2
1829
1862
                "movq (%0), %%mm2                               \n\t" // L1
1830
1863
                PAVGB(%%mm2, %%mm0)
1831
1864
                "movq %%mm0, (%0)                               \n\t"
1832
 
                "movq (%%eax, %1), %%mm0                        \n\t" // L3
 
1865
                "movq (%%"REG_a", %1), %%mm0                    \n\t" // L3
1833
1866
                PAVGB(%%mm0, %%mm2)                                   // L1+L3
1834
1867
                PAVGB(%%mm1, %%mm2)                                   // 2L2 + L1 + L3
1835
 
                "movq %%mm2, (%%eax)                            \n\t"
1836
 
                "movq (%%eax, %1, 2), %%mm2                     \n\t" // L4
 
1868
                "movq %%mm2, (%%"REG_a")                        \n\t"
 
1869
                "movq (%%"REG_a", %1, 2), %%mm2                 \n\t" // L4
1837
1870
                PAVGB(%%mm2, %%mm1)                                   // L2+L4
1838
1871
                PAVGB(%%mm0, %%mm1)                                   // 2L3 + L2 + L4
1839
 
                "movq %%mm1, (%%eax, %1)                        \n\t"
 
1872
                "movq %%mm1, (%%"REG_a", %1)                    \n\t"
1840
1873
                "movq (%0, %1, 4), %%mm1                        \n\t" // L5
1841
1874
                PAVGB(%%mm1, %%mm0)                                   // L3+L5
1842
1875
                PAVGB(%%mm2, %%mm0)                                   // 2L4 + L3 + L5
1843
 
                "movq %%mm0, (%%eax, %1, 2)                     \n\t"
1844
 
                "movq (%%edx), %%mm0                            \n\t" // L6
 
1876
                "movq %%mm0, (%%"REG_a", %1, 2)                 \n\t"
 
1877
                "movq (%%"REG_d"), %%mm0                        \n\t" // L6
1845
1878
                PAVGB(%%mm0, %%mm2)                                   // L4+L6
1846
1879
                PAVGB(%%mm1, %%mm2)                                   // 2L5 + L4 + L6
1847
1880
                "movq %%mm2, (%0, %1, 4)                        \n\t"
1848
 
                "movq (%%edx, %1), %%mm2                        \n\t" // L7
 
1881
                "movq (%%"REG_d", %1), %%mm2                    \n\t" // L7
1849
1882
                PAVGB(%%mm2, %%mm1)                                   // L5+L7
1850
1883
                PAVGB(%%mm0, %%mm1)                                   // 2L6 + L5 + L7
1851
 
                "movq %%mm1, (%%edx)                            \n\t"
1852
 
                "movq (%%edx, %1, 2), %%mm1                     \n\t" // L8
 
1884
                "movq %%mm1, (%%"REG_d")                        \n\t"
 
1885
                "movq (%%"REG_d", %1, 2), %%mm1                 \n\t" // L8
1853
1886
                PAVGB(%%mm1, %%mm0)                                   // L6+L8
1854
1887
                PAVGB(%%mm2, %%mm0)                                   // 2L7 + L6 + L8
1855
 
                "movq %%mm0, (%%edx, %1)                        \n\t"
 
1888
                "movq %%mm0, (%%"REG_d", %1)                    \n\t"
1856
1889
                "movq (%0, %1, 8), %%mm0                        \n\t" // L9
1857
1890
                PAVGB(%%mm0, %%mm2)                                   // L7+L9
1858
1891
                PAVGB(%%mm1, %%mm2)                                   // 2L8 + L7 + L9
1859
 
                "movq %%mm2, (%%edx, %1, 2)                     \n\t"
 
1892
                "movq %%mm2, (%%"REG_d", %1, 2)                 \n\t"
1860
1893
                "movq %%mm1, (%2)                               \n\t"
1861
1894
 
1862
 
                : : "r" (src), "r" (stride), "r" (tmp)
1863
 
                : "%eax", "%edx"
 
1895
                : : "r" (src), "r" ((long)stride), "r" (tmp)
 
1896
                : "%"REG_a, "%"REG_d
1864
1897
        );
1865
1898
#else
1866
1899
        int a, b, c, x;
1920
1953
        src+= 4*stride;
1921
1954
#ifdef HAVE_MMX2
1922
1955
        asm volatile(
1923
 
                "leal (%0, %1), %%eax                           \n\t"
1924
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
1956
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
1957
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
1925
1958
//      0       1       2       3       4       5       6       7       8       9
1926
1959
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1
1927
1960
 
1928
1961
                "movq (%0), %%mm0                               \n\t" //
1929
 
                "movq (%%eax, %1), %%mm2                        \n\t" //
1930
 
                "movq (%%eax), %%mm1                            \n\t" //
 
1962
                "movq (%%"REG_a", %1), %%mm2                    \n\t" //
 
1963
                "movq (%%"REG_a"), %%mm1                        \n\t" //
1931
1964
                "movq %%mm0, %%mm3                              \n\t"
1932
1965
                "pmaxub %%mm1, %%mm0                            \n\t" //
1933
1966
                "pminub %%mm3, %%mm1                            \n\t" //
1934
1967
                "pmaxub %%mm2, %%mm1                            \n\t" //
1935
1968
                "pminub %%mm1, %%mm0                            \n\t"
1936
 
                "movq %%mm0, (%%eax)                            \n\t"
 
1969
                "movq %%mm0, (%%"REG_a")                        \n\t"
1937
1970
 
1938
1971
                "movq (%0, %1, 4), %%mm0                        \n\t" //
1939
 
                "movq (%%eax, %1, 2), %%mm1                     \n\t" //
 
1972
                "movq (%%"REG_a", %1, 2), %%mm1                 \n\t" //
1940
1973
                "movq %%mm2, %%mm3                              \n\t"
1941
1974
                "pmaxub %%mm1, %%mm2                            \n\t" //
1942
1975
                "pminub %%mm3, %%mm1                            \n\t" //
1943
1976
                "pmaxub %%mm0, %%mm1                            \n\t" //
1944
1977
                "pminub %%mm1, %%mm2                            \n\t"
1945
 
                "movq %%mm2, (%%eax, %1, 2)                     \n\t"
 
1978
                "movq %%mm2, (%%"REG_a", %1, 2)                 \n\t"
1946
1979
 
1947
 
                "movq (%%edx), %%mm2                            \n\t" //
1948
 
                "movq (%%edx, %1), %%mm1                        \n\t" //
 
1980
                "movq (%%"REG_d"), %%mm2                        \n\t" //
 
1981
                "movq (%%"REG_d", %1), %%mm1                    \n\t" //
1949
1982
                "movq %%mm2, %%mm3                              \n\t"
1950
1983
                "pmaxub %%mm0, %%mm2                            \n\t" //
1951
1984
                "pminub %%mm3, %%mm0                            \n\t" //
1952
1985
                "pmaxub %%mm1, %%mm0                            \n\t" //
1953
1986
                "pminub %%mm0, %%mm2                            \n\t"
1954
 
                "movq %%mm2, (%%edx)                            \n\t"
 
1987
                "movq %%mm2, (%%"REG_d")                        \n\t"
1955
1988
 
1956
 
                "movq (%%edx, %1, 2), %%mm2                     \n\t" //
 
1989
                "movq (%%"REG_d", %1, 2), %%mm2                 \n\t" //
1957
1990
                "movq (%0, %1, 8), %%mm0                        \n\t" //
1958
1991
                "movq %%mm2, %%mm3                              \n\t"
1959
1992
                "pmaxub %%mm0, %%mm2                            \n\t" //
1960
1993
                "pminub %%mm3, %%mm0                            \n\t" //
1961
1994
                "pmaxub %%mm1, %%mm0                            \n\t" //
1962
1995
                "pminub %%mm0, %%mm2                            \n\t"
1963
 
                "movq %%mm2, (%%edx, %1, 2)                     \n\t"
1964
 
 
1965
 
 
1966
 
                : : "r" (src), "r" (stride)
1967
 
                : "%eax", "%edx"
 
1996
                "movq %%mm2, (%%"REG_d", %1, 2)                 \n\t"
 
1997
 
 
1998
 
 
1999
                : : "r" (src), "r" ((long)stride)
 
2000
                : "%"REG_a, "%"REG_d
1968
2001
        );
1969
2002
 
1970
2003
#else // MMX without MMX2
1971
2004
        asm volatile(
1972
 
                "leal (%0, %1), %%eax                           \n\t"
1973
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
2005
                "lea (%0, %1), %%"REG_a"                        \n\t"
 
2006
                "lea (%%"REG_a", %1, 4), %%"REG_d"              \n\t"
1974
2007
//      0       1       2       3       4       5       6       7       8       9
1975
2008
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1
1976
2009
                "pxor %%mm7, %%mm7                              \n\t"
1977
2010
 
1978
 
#define MEDIAN(a,b,c)\
 
2011
#define REAL_MEDIAN(a,b,c)\
1979
2012
                "movq " #a ", %%mm0                             \n\t"\
1980
2013
                "movq " #b ", %%mm2                             \n\t"\
1981
2014
                "movq " #c ", %%mm1                             \n\t"\
1998
2031
                "pand %%mm2, %%mm0                              \n\t"\
1999
2032
                "pand %%mm1, %%mm0                              \n\t"\
2000
2033
                "movq %%mm0, " #b "                             \n\t"
2001
 
 
2002
 
MEDIAN((%0), (%%eax), (%%eax, %1))
2003
 
MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
2004
 
MEDIAN((%0, %1, 4), (%%edx), (%%edx, %1))
2005
 
MEDIAN((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8))
2006
 
 
2007
 
                : : "r" (src), "r" (stride)
2008
 
                : "%eax", "%edx"
 
2034
#define MEDIAN(a,b,c)  REAL_MEDIAN(a,b,c)
 
2035
 
 
2036
MEDIAN((%0), (%%REGa), (%%REGa, %1))
 
2037
MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4))
 
2038
MEDIAN((%0, %1, 4), (%%REGd), (%%REGd, %1))
 
2039
MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8))
 
2040
 
 
2041
                : : "r" (src), "r" ((long)stride)
 
2042
                : "%"REG_a, "%"REG_d
2009
2043
        );
2010
2044
#endif // MMX
2011
2045
#else
2039
2073
static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
2040
2074
{
2041
2075
        asm(
2042
 
                "leal (%0, %1), %%eax                           \n\t"
 
2076
                "lea (%0, %1), %%"REG_a"        \n\t"
2043
2077
//      0       1       2       3       4       5       6       7       8       9
2044
2078
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1
2045
2079
                "movq (%0), %%mm0               \n\t" // 12345678
2046
 
                "movq (%%eax), %%mm1            \n\t" // abcdefgh
 
2080
                "movq (%%"REG_a"), %%mm1        \n\t" // abcdefgh
2047
2081
                "movq %%mm0, %%mm2              \n\t" // 12345678
2048
2082
                "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
2049
2083
                "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
2050
2084
 
2051
 
                "movq (%%eax, %1), %%mm1        \n\t"
2052
 
                "movq (%%eax, %1, 2), %%mm3     \n\t"
 
2085
                "movq (%%"REG_a", %1), %%mm1    \n\t"
 
2086
                "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2053
2087
                "movq %%mm1, %%mm4              \n\t"
2054
2088
                "punpcklbw %%mm3, %%mm1         \n\t"
2055
2089
                "punpckhbw %%mm3, %%mm4         \n\t"
2076
2110
                "psrlq $32, %%mm1               \n\t"
2077
2111
                "movd %%mm1, 112(%3)            \n\t"
2078
2112
 
2079
 
                "leal (%%eax, %1, 4), %%eax     \n\t"
 
2113
                "lea (%%"REG_a", %1, 4), %%"REG_a"      \n\t"
2080
2114
                
2081
2115
                "movq (%0, %1, 4), %%mm0        \n\t" // 12345678
2082
 
                "movq (%%eax), %%mm1            \n\t" // abcdefgh
 
2116
                "movq (%%"REG_a"), %%mm1        \n\t" // abcdefgh
2083
2117
                "movq %%mm0, %%mm2              \n\t" // 12345678
2084
2118
                "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
2085
2119
                "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
2086
2120
 
2087
 
                "movq (%%eax, %1), %%mm1        \n\t"
2088
 
                "movq (%%eax, %1, 2), %%mm3     \n\t"
 
2121
                "movq (%%"REG_a", %1), %%mm1    \n\t"
 
2122
                "movq (%%"REG_a", %1, 2), %%mm3 \n\t"
2089
2123
                "movq %%mm1, %%mm4              \n\t"
2090
2124
                "punpcklbw %%mm3, %%mm1         \n\t"
2091
2125
                "punpckhbw %%mm3, %%mm4         \n\t"
2113
2147
                "movd %%mm1, 116(%3)            \n\t"
2114
2148
 
2115
2149
 
2116
 
        :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)
2117
 
        : "%eax"
 
2150
        :: "r" (src), "r" ((long)srcStride), "r" (dst1), "r" (dst2)
 
2151
        : "%"REG_a
2118
2152
        );
2119
2153
}
2120
2154
 
2124
2158
static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
2125
2159
{
2126
2160
        asm(
2127
 
                "leal (%0, %1), %%eax                           \n\t"
2128
 
                "leal (%%eax, %1, 4), %%edx                     \n\t"
 
2161
                "lea (%0, %1), %%"REG_a"        \n\t"
 
2162
                "lea (%%"REG_a",%1,4), %%"REG_d"\n\t"
2129
2163
//      0       1       2       3       4       5       6       7       8       9
2130
2164
//      %0      eax     eax+%1  eax+2%1 %0+4%1  edx     edx+%1  edx+2%1 %0+8%1  edx+4%1
2131
2165
                "movq (%2), %%mm0               \n\t" // 12345678
2149
2183
 
2150
2184
                "movd %%mm0, (%0)               \n\t"
2151
2185
                "psrlq $32, %%mm0               \n\t"
2152
 
                "movd %%mm0, (%%eax)            \n\t"
2153
 
                "movd %%mm3, (%%eax, %1)        \n\t"
 
2186
                "movd %%mm0, (%%"REG_a")        \n\t"
 
2187
                "movd %%mm3, (%%"REG_a", %1)    \n\t"
2154
2188
                "psrlq $32, %%mm3               \n\t"
2155
 
                "movd %%mm3, (%%eax, %1, 2)     \n\t"
 
2189
                "movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2156
2190
                "movd %%mm2, (%0, %1, 4)        \n\t"
2157
2191
                "psrlq $32, %%mm2               \n\t"
2158
 
                "movd %%mm2, (%%edx)            \n\t"
2159
 
                "movd %%mm1, (%%edx, %1)        \n\t"
 
2192
                "movd %%mm2, (%%"REG_d")        \n\t"
 
2193
                "movd %%mm1, (%%"REG_d", %1)    \n\t"
2160
2194
                "psrlq $32, %%mm1               \n\t"
2161
 
                "movd %%mm1, (%%edx, %1, 2)     \n\t"
 
2195
                "movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2162
2196
 
2163
2197
 
2164
2198
                "movq 64(%2), %%mm0             \n\t" // 12345678
2182
2216
 
2183
2217
                "movd %%mm0, 4(%0)              \n\t"
2184
2218
                "psrlq $32, %%mm0               \n\t"
2185
 
                "movd %%mm0, 4(%%eax)           \n\t"
2186
 
                "movd %%mm3, 4(%%eax, %1)       \n\t"
 
2219
                "movd %%mm0, 4(%%"REG_a")               \n\t"
 
2220
                "movd %%mm3, 4(%%"REG_a", %1)   \n\t"
2187
2221
                "psrlq $32, %%mm3               \n\t"
2188
 
                "movd %%mm3, 4(%%eax, %1, 2)    \n\t"
 
2222
                "movd %%mm3, 4(%%"REG_a", %1, 2)        \n\t"
2189
2223
                "movd %%mm2, 4(%0, %1, 4)       \n\t"
2190
2224
                "psrlq $32, %%mm2               \n\t"
2191
 
                "movd %%mm2, 4(%%edx)           \n\t"
2192
 
                "movd %%mm1, 4(%%edx, %1)       \n\t"
 
2225
                "movd %%mm2, 4(%%"REG_d")               \n\t"
 
2226
                "movd %%mm1, 4(%%"REG_d", %1)   \n\t"
2193
2227
                "psrlq $32, %%mm1               \n\t"
2194
 
                "movd %%mm1, 4(%%edx, %1, 2)    \n\t"
 
2228
                "movd %%mm1, 4(%%"REG_d", %1, 2)        \n\t"
2195
2229
 
2196
 
        :: "r" (dst), "r" (dstStride), "r" (src)
2197
 
        : "%eax", "%edx"
 
2230
        :: "r" (dst), "r" ((long)dstStride), "r" (src)
 
2231
        : "%"REG_a, "%"REG_d
2198
2232
        );
2199
2233
}
2200
2234
#endif
2201
 
//static int test=0;
 
2235
//static long test=0;
2202
2236
 
 
2237
#ifndef HAVE_ALTIVEC
2203
2238
static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2204
2239
                                    uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2205
2240
{
2212
2247
//#define L1_DIFF //u should change the thresholds too if u try that one
2213
2248
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2214
2249
        asm volatile(
2215
 
                "leal (%2, %2, 2), %%eax                        \n\t" // 3*stride
2216
 
                "leal (%2, %2, 4), %%edx                        \n\t" // 5*stride
2217
 
                "leal (%%edx, %2, 2), %%ecx                     \n\t" // 7*stride
 
2250
                "lea (%2, %2, 2), %%"REG_a"                     \n\t" // 3*stride
 
2251
                "lea (%2, %2, 4), %%"REG_d"                     \n\t" // 5*stride
 
2252
                "lea (%%"REG_d", %2, 2), %%"REG_c"              \n\t" // 7*stride
2218
2253
//      0       1       2       3       4       5       6       7       8       9
2219
2254
//      %x      %x+%2   %x+2%2  %x+eax  %x+4%2  %x+edx  %x+2eax %x+ecx  %x+8%2
2220
2255
//FIXME reorder?
2225
2260
                "psadbw (%1, %2), %%mm1                         \n\t" // |L1-R1|
2226
2261
                "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2227
2262
                "psadbw (%1, %2, 2), %%mm2                      \n\t" // |L2-R2|
2228
 
                "movq (%0, %%eax), %%mm3                        \n\t" // L3
2229
 
                "psadbw (%1, %%eax), %%mm3                      \n\t" // |L3-R3|
 
2263
                "movq (%0, %%"REG_a"), %%mm3                    \n\t" // L3
 
2264
                "psadbw (%1, %%"REG_a"), %%mm3                  \n\t" // |L3-R3|
2230
2265
 
2231
2266
                "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2232
2267
                "paddw %%mm1, %%mm0                             \n\t"
2233
2268
                "psadbw (%1, %2, 4), %%mm4                      \n\t" // |L4-R4|
2234
 
                "movq (%0, %%edx), %%mm5                        \n\t" // L5
 
2269
                "movq (%0, %%"REG_d"), %%mm5                    \n\t" // L5
2235
2270
                "paddw %%mm2, %%mm0                             \n\t"
2236
 
                "psadbw (%1, %%edx), %%mm5                      \n\t" // |L5-R5|
2237
 
                "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
 
2271
                "psadbw (%1, %%"REG_d"), %%mm5                  \n\t" // |L5-R5|
 
2272
                "movq (%0, %%"REG_a", 2), %%mm6                 \n\t" // L6
2238
2273
                "paddw %%mm3, %%mm0                             \n\t"
2239
 
                "psadbw (%1, %%eax, 2), %%mm6                   \n\t" // |L6-R6|
2240
 
                "movq (%0, %%ecx), %%mm7                        \n\t" // L7
 
2274
                "psadbw (%1, %%"REG_a", 2), %%mm6               \n\t" // |L6-R6|
 
2275
                "movq (%0, %%"REG_c"), %%mm7                    \n\t" // L7
2241
2276
                "paddw %%mm4, %%mm0                             \n\t"
2242
 
                "psadbw (%1, %%ecx), %%mm7                      \n\t" // |L7-R7|
 
2277
                "psadbw (%1, %%"REG_c"), %%mm7                  \n\t" // |L7-R7|
2243
2278
                "paddw %%mm5, %%mm6                             \n\t"
2244
2279
                "paddw %%mm7, %%mm6                             \n\t"
2245
2280
                "paddw %%mm6, %%mm0                             \n\t"
2246
 
#elif defined (FAST_L2_DIFF)
 
2281
#else
 
2282
#if defined (FAST_L2_DIFF)
2247
2283
                "pcmpeqb %%mm7, %%mm7                           \n\t"
2248
2284
                "movq "MANGLE(b80)", %%mm6                      \n\t"
2249
2285
                "pxor %%mm0, %%mm0                              \n\t"
2250
 
#define L2_DIFF_CORE(a, b)\
 
2286
#define REAL_L2_DIFF_CORE(a, b)\
2251
2287
                "movq " #a ", %%mm5                             \n\t"\
2252
2288
                "movq " #b ", %%mm2                             \n\t"\
2253
2289
                "pxor %%mm7, %%mm2                              \n\t"\
2261
2297
                "psrld $14, %%mm5                               \n\t"\
2262
2298
                "paddd %%mm5, %%mm0                             \n\t"
2263
2299
 
2264
 
L2_DIFF_CORE((%0), (%1))
2265
 
L2_DIFF_CORE((%0, %2), (%1, %2))
2266
 
L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2267
 
L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2268
 
L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2269
 
L2_DIFF_CORE((%0, %%edx), (%1, %%edx))
2270
 
L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2271
 
L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2272
 
 
2273
2300
#else
2274
2301
                "pxor %%mm7, %%mm7                              \n\t"
2275
2302
                "pxor %%mm0, %%mm0                              \n\t"
2276
 
#define L2_DIFF_CORE(a, b)\
 
2303
#define REAL_L2_DIFF_CORE(a, b)\
2277
2304
                "movq " #a ", %%mm5                             \n\t"\
2278
2305
                "movq " #b ", %%mm2                             \n\t"\
2279
2306
                "movq %%mm5, %%mm1                              \n\t"\
2289
2316
                "paddd %%mm1, %%mm5                             \n\t"\
2290
2317
                "paddd %%mm5, %%mm0                             \n\t"
2291
2318
 
 
2319
#endif
 
2320
 
 
2321
#define L2_DIFF_CORE(a, b)  REAL_L2_DIFF_CORE(a, b)
 
2322
 
2292
2323
L2_DIFF_CORE((%0), (%1))
2293
2324
L2_DIFF_CORE((%0, %2), (%1, %2))
2294
2325
L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2295
 
L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
 
2326
L2_DIFF_CORE((%0, %%REGa), (%1, %%REGa))
2296
2327
L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2297
 
L2_DIFF_CORE((%0, %%edx), (%1, %%edx))
2298
 
L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2299
 
L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
 
2328
L2_DIFF_CORE((%0, %%REGd), (%1, %%REGd))
 
2329
L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2))
 
2330
L2_DIFF_CORE((%0, %%REGc), (%1, %%REGc))
2300
2331
 
2301
2332
#endif
2302
2333
 
2305
2336
                "paddd %%mm0, %%mm4                             \n\t"
2306
2337
                "movd %%mm4, %%ecx                              \n\t"
2307
2338
                "shll $2, %%ecx                                 \n\t"
2308
 
                "movl %3, %%edx                                 \n\t"
2309
 
                "addl -4(%%edx), %%ecx                          \n\t"
2310
 
                "addl 4(%%edx), %%ecx                           \n\t"
2311
 
                "addl -1024(%%edx), %%ecx                       \n\t"
 
2339
                "mov %3, %%"REG_d"                              \n\t"
 
2340
                "addl -4(%%"REG_d"), %%ecx                      \n\t"
 
2341
                "addl 4(%%"REG_d"), %%ecx                       \n\t"
 
2342
                "addl -1024(%%"REG_d"), %%ecx                   \n\t"
2312
2343
                "addl $4, %%ecx                                 \n\t"
2313
 
                "addl 1024(%%edx), %%ecx                        \n\t"
 
2344
                "addl 1024(%%"REG_d"), %%ecx                    \n\t"
2314
2345
                "shrl $3, %%ecx                                 \n\t"
2315
 
                "movl %%ecx, (%%edx)                            \n\t"
 
2346
                "movl %%ecx, (%%"REG_d")                        \n\t"
2316
2347
 
2317
 
//              "movl %3, %%ecx                                 \n\t"
2318
 
//              "movl %%ecx, test                               \n\t"
 
2348
//              "mov %3, %%"REG_c"                              \n\t"
 
2349
//              "mov %%"REG_c", test                            \n\t"
2319
2350
//              "jmp 4f \n\t"
2320
 
                "cmpl 512(%%edx), %%ecx                         \n\t"
 
2351
                "cmpl 512(%%"REG_d"), %%ecx                     \n\t"
2321
2352
                " jb 2f                                         \n\t"
2322
 
                "cmpl 516(%%edx), %%ecx                         \n\t"
 
2353
                "cmpl 516(%%"REG_d"), %%ecx                     \n\t"
2323
2354
                " jb 1f                                         \n\t"
2324
2355
 
2325
 
                "leal (%%eax, %2, 2), %%edx                     \n\t" // 5*stride
2326
 
                "leal (%%edx, %2, 2), %%ecx                     \n\t" // 7*stride
 
2356
                "lea (%%"REG_a", %2, 2), %%"REG_d"              \n\t" // 5*stride
 
2357
                "lea (%%"REG_d", %2, 2), %%"REG_c"              \n\t" // 7*stride
2327
2358
                "movq (%0), %%mm0                               \n\t" // L0
2328
2359
                "movq (%0, %2), %%mm1                           \n\t" // L1
2329
2360
                "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2330
 
                "movq (%0, %%eax), %%mm3                        \n\t" // L3
 
2361
                "movq (%0, %%"REG_a"), %%mm3                    \n\t" // L3
2331
2362
                "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2332
 
                "movq (%0, %%edx), %%mm5                        \n\t" // L5
2333
 
                "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2334
 
                "movq (%0, %%ecx), %%mm7                        \n\t" // L7
 
2363
                "movq (%0, %%"REG_d"), %%mm5                    \n\t" // L5
 
2364
                "movq (%0, %%"REG_a", 2), %%mm6                 \n\t" // L6
 
2365
                "movq (%0, %%"REG_c"), %%mm7                    \n\t" // L7
2335
2366
                "movq %%mm0, (%1)                               \n\t" // L0
2336
2367
                "movq %%mm1, (%1, %2)                           \n\t" // L1
2337
2368
                "movq %%mm2, (%1, %2, 2)                        \n\t" // L2
2338
 
                "movq %%mm3, (%1, %%eax)                        \n\t" // L3
 
2369
                "movq %%mm3, (%1, %%"REG_a")                    \n\t" // L3
2339
2370
                "movq %%mm4, (%1, %2, 4)                        \n\t" // L4
2340
 
                "movq %%mm5, (%1, %%edx)                        \n\t" // L5
2341
 
                "movq %%mm6, (%1, %%eax, 2)                     \n\t" // L6
2342
 
                "movq %%mm7, (%1, %%ecx)                        \n\t" // L7
 
2371
                "movq %%mm5, (%1, %%"REG_d")                    \n\t" // L5
 
2372
                "movq %%mm6, (%1, %%"REG_a", 2)                 \n\t" // L6
 
2373
                "movq %%mm7, (%1, %%"REG_c")                    \n\t" // L7
2343
2374
                "jmp 4f                                         \n\t"
2344
2375
 
2345
2376
                "1:                                             \n\t"
2346
 
                "leal (%%eax, %2, 2), %%edx                     \n\t" // 5*stride
2347
 
                "leal (%%edx, %2, 2), %%ecx                     \n\t" // 7*stride
 
2377
                "lea (%%"REG_a", %2, 2), %%"REG_d"              \n\t" // 5*stride
 
2378
                "lea (%%"REG_d", %2, 2), %%"REG_c"              \n\t" // 7*stride
2348
2379
                "movq (%0), %%mm0                               \n\t" // L0
2349
2380
                PAVGB((%1), %%mm0)                                    // L0
2350
2381
                "movq (%0, %2), %%mm1                           \n\t" // L1
2351
2382
                PAVGB((%1, %2), %%mm1)                                // L1
2352
2383
                "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2353
2384
                PAVGB((%1, %2, 2), %%mm2)                             // L2
2354
 
                "movq (%0, %%eax), %%mm3                        \n\t" // L3
2355
 
                PAVGB((%1, %%eax), %%mm3)                             // L3
 
2385
                "movq (%0, %%"REG_a"), %%mm3                    \n\t" // L3
 
2386
                PAVGB((%1, %%REGa), %%mm3)                            // L3
2356
2387
                "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2357
2388
                PAVGB((%1, %2, 4), %%mm4)                             // L4
2358
 
                "movq (%0, %%edx), %%mm5                        \n\t" // L5
2359
 
                PAVGB((%1, %%edx), %%mm5)                             // L5
2360
 
                "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2361
 
                PAVGB((%1, %%eax, 2), %%mm6)                          // L6
2362
 
                "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2363
 
                PAVGB((%1, %%ecx), %%mm7)                             // L7
 
2389
                "movq (%0, %%"REG_d"), %%mm5                    \n\t" // L5
 
2390
                PAVGB((%1, %%REGd), %%mm5)                            // L5
 
2391
                "movq (%0, %%"REG_a", 2), %%mm6                 \n\t" // L6
 
2392
                PAVGB((%1, %%REGa, 2), %%mm6)                         // L6
 
2393
                "movq (%0, %%"REG_c"), %%mm7                    \n\t" // L7
 
2394
                PAVGB((%1, %%REGc), %%mm7)                            // L7
2364
2395
                "movq %%mm0, (%1)                               \n\t" // R0
2365
2396
                "movq %%mm1, (%1, %2)                           \n\t" // R1
2366
2397
                "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2367
 
                "movq %%mm3, (%1, %%eax)                        \n\t" // R3
 
2398
                "movq %%mm3, (%1, %%"REG_a")                    \n\t" // R3
2368
2399
                "movq %%mm4, (%1, %2, 4)                        \n\t" // R4
2369
 
                "movq %%mm5, (%1, %%edx)                        \n\t" // R5
2370
 
                "movq %%mm6, (%1, %%eax, 2)                     \n\t" // R6
2371
 
                "movq %%mm7, (%1, %%ecx)                        \n\t" // R7
 
2400
                "movq %%mm5, (%1, %%"REG_d")                    \n\t" // R5
 
2401
                "movq %%mm6, (%1, %%"REG_a", 2)                 \n\t" // R6
 
2402
                "movq %%mm7, (%1, %%"REG_c")                    \n\t" // R7
2372
2403
                "movq %%mm0, (%0)                               \n\t" // L0
2373
2404
                "movq %%mm1, (%0, %2)                           \n\t" // L1
2374
2405
                "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2375
 
                "movq %%mm3, (%0, %%eax)                        \n\t" // L3
 
2406
                "movq %%mm3, (%0, %%"REG_a")                    \n\t" // L3
2376
2407
                "movq %%mm4, (%0, %2, 4)                        \n\t" // L4
2377
 
                "movq %%mm5, (%0, %%edx)                        \n\t" // L5
2378
 
                "movq %%mm6, (%0, %%eax, 2)                     \n\t" // L6
2379
 
                "movq %%mm7, (%0, %%ecx)                        \n\t" // L7
 
2408
                "movq %%mm5, (%0, %%"REG_d")                    \n\t" // L5
 
2409
                "movq %%mm6, (%0, %%"REG_a", 2)                 \n\t" // L6
 
2410
                "movq %%mm7, (%0, %%"REG_c")                    \n\t" // L7
2380
2411
                "jmp 4f                                         \n\t"
2381
2412
 
2382
2413
                "2:                                             \n\t"
2383
 
                "cmpl 508(%%edx), %%ecx                         \n\t"
 
2414
                "cmpl 508(%%"REG_d"), %%ecx                     \n\t"
2384
2415
                " jb 3f                                         \n\t"
2385
2416
 
2386
 
                "leal (%%eax, %2, 2), %%edx                     \n\t" // 5*stride
2387
 
                "leal (%%edx, %2, 2), %%ecx                     \n\t" // 7*stride
 
2417
                "lea (%%"REG_a", %2, 2), %%"REG_d"              \n\t" // 5*stride
 
2418
                "lea (%%"REG_d", %2, 2), %%"REG_c"              \n\t" // 7*stride
2388
2419
                "movq (%0), %%mm0                               \n\t" // L0
2389
2420
                "movq (%0, %2), %%mm1                           \n\t" // L1
2390
2421
                "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2391
 
                "movq (%0, %%eax), %%mm3                        \n\t" // L3
 
2422
                "movq (%0, %%"REG_a"), %%mm3                    \n\t" // L3
2392
2423
                "movq (%1), %%mm4                               \n\t" // R0
2393
2424
                "movq (%1, %2), %%mm5                           \n\t" // R1
2394
2425
                "movq (%1, %2, 2), %%mm6                        \n\t" // R2
2395
 
                "movq (%1, %%eax), %%mm7                        \n\t" // R3
 
2426
                "movq (%1, %%"REG_a"), %%mm7                    \n\t" // R3
2396
2427
                PAVGB(%%mm4, %%mm0)
2397
2428
                PAVGB(%%mm5, %%mm1)
2398
2429
                PAVGB(%%mm6, %%mm2)
2404
2435
                "movq %%mm0, (%1)                               \n\t" // R0
2405
2436
                "movq %%mm1, (%1, %2)                           \n\t" // R1
2406
2437
                "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2407
 
                "movq %%mm3, (%1, %%eax)                        \n\t" // R3
 
2438
                "movq %%mm3, (%1, %%"REG_a")                    \n\t" // R3
2408
2439
                "movq %%mm0, (%0)                               \n\t" // L0
2409
2440
                "movq %%mm1, (%0, %2)                           \n\t" // L1
2410
2441
                "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2411
 
                "movq %%mm3, (%0, %%eax)                        \n\t" // L3
 
2442
                "movq %%mm3, (%0, %%"REG_a")                    \n\t" // L3
2412
2443
 
2413
2444
                "movq (%0, %2, 4), %%mm0                        \n\t" // L4
2414
 
                "movq (%0, %%edx), %%mm1                        \n\t" // L5
2415
 
                "movq (%0, %%eax, 2), %%mm2                     \n\t" // L6
2416
 
                "movq (%0, %%ecx), %%mm3                        \n\t" // L7
 
2445
                "movq (%0, %%"REG_d"), %%mm1                    \n\t" // L5
 
2446
                "movq (%0, %%"REG_a", 2), %%mm2                 \n\t" // L6
 
2447
                "movq (%0, %%"REG_c"), %%mm3                    \n\t" // L7
2417
2448
                "movq (%1, %2, 4), %%mm4                        \n\t" // R4
2418
 
                "movq (%1, %%edx), %%mm5                        \n\t" // R5
2419
 
                "movq (%1, %%eax, 2), %%mm6                     \n\t" // R6
2420
 
                "movq (%1, %%ecx), %%mm7                        \n\t" // R7
 
2449
                "movq (%1, %%"REG_d"), %%mm5                    \n\t" // R5
 
2450
                "movq (%1, %%"REG_a", 2), %%mm6                 \n\t" // R6
 
2451
                "movq (%1, %%"REG_c"), %%mm7                    \n\t" // R7
2421
2452
                PAVGB(%%mm4, %%mm0)
2422
2453
                PAVGB(%%mm5, %%mm1)
2423
2454
                PAVGB(%%mm6, %%mm2)
2427
2458
                PAVGB(%%mm6, %%mm2)
2428
2459
                PAVGB(%%mm7, %%mm3)
2429
2460
                "movq %%mm0, (%1, %2, 4)                        \n\t" // R4
2430
 
                "movq %%mm1, (%1, %%edx)                        \n\t" // R5
2431
 
                "movq %%mm2, (%1, %%eax, 2)                     \n\t" // R6
2432
 
                "movq %%mm3, (%1, %%ecx)                        \n\t" // R7
 
2461
                "movq %%mm1, (%1, %%"REG_d")                    \n\t" // R5
 
2462
                "movq %%mm2, (%1, %%"REG_a", 2)                 \n\t" // R6
 
2463
                "movq %%mm3, (%1, %%"REG_c")                    \n\t" // R7
2433
2464
                "movq %%mm0, (%0, %2, 4)                        \n\t" // L4
2434
 
                "movq %%mm1, (%0, %%edx)                        \n\t" // L5
2435
 
                "movq %%mm2, (%0, %%eax, 2)                     \n\t" // L6
2436
 
                "movq %%mm3, (%0, %%ecx)                        \n\t" // L7
 
2465
                "movq %%mm1, (%0, %%"REG_d")                    \n\t" // L5
 
2466
                "movq %%mm2, (%0, %%"REG_a", 2)                 \n\t" // L6
 
2467
                "movq %%mm3, (%0, %%"REG_c")                    \n\t" // L7
2437
2468
                "jmp 4f                                         \n\t"
2438
2469
 
2439
2470
                "3:                                             \n\t"
2440
 
                "leal (%%eax, %2, 2), %%edx                     \n\t" // 5*stride
2441
 
                "leal (%%edx, %2, 2), %%ecx                     \n\t" // 7*stride
 
2471
                "lea (%%"REG_a", %2, 2), %%"REG_d"              \n\t" // 5*stride
 
2472
                "lea (%%"REG_d", %2, 2), %%"REG_c"              \n\t" // 7*stride
2442
2473
                "movq (%0), %%mm0                               \n\t" // L0
2443
2474
                "movq (%0, %2), %%mm1                           \n\t" // L1
2444
2475
                "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2445
 
                "movq (%0, %%eax), %%mm3                        \n\t" // L3
 
2476
                "movq (%0, %%"REG_a"), %%mm3                    \n\t" // L3
2446
2477
                "movq (%1), %%mm4                               \n\t" // R0
2447
2478
                "movq (%1, %2), %%mm5                           \n\t" // R1
2448
2479
                "movq (%1, %2, 2), %%mm6                        \n\t" // R2
2449
 
                "movq (%1, %%eax), %%mm7                        \n\t" // R3
 
2480
                "movq (%1, %%"REG_a"), %%mm7                    \n\t" // R3
2450
2481
                PAVGB(%%mm4, %%mm0)
2451
2482
                PAVGB(%%mm5, %%mm1)
2452
2483
                PAVGB(%%mm6, %%mm2)
2462
2493
                "movq %%mm0, (%1)                               \n\t" // R0
2463
2494
                "movq %%mm1, (%1, %2)                           \n\t" // R1
2464
2495
                "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2465
 
                "movq %%mm3, (%1, %%eax)                        \n\t" // R3
 
2496
                "movq %%mm3, (%1, %%"REG_a")                    \n\t" // R3
2466
2497
                "movq %%mm0, (%0)                               \n\t" // L0
2467
2498
                "movq %%mm1, (%0, %2)                           \n\t" // L1
2468
2499
                "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2469
 
                "movq %%mm3, (%0, %%eax)                        \n\t" // L3
 
2500
                "movq %%mm3, (%0, %%"REG_a")                    \n\t" // L3
2470
2501
 
2471
2502
                "movq (%0, %2, 4), %%mm0                        \n\t" // L4
2472
 
                "movq (%0, %%edx), %%mm1                        \n\t" // L5
2473
 
                "movq (%0, %%eax, 2), %%mm2                     \n\t" // L6
2474
 
                "movq (%0, %%ecx), %%mm3                        \n\t" // L7
 
2503
                "movq (%0, %%"REG_d"), %%mm1                    \n\t" // L5
 
2504
                "movq (%0, %%"REG_a", 2), %%mm2                 \n\t" // L6
 
2505
                "movq (%0, %%"REG_c"), %%mm3                    \n\t" // L7
2475
2506
                "movq (%1, %2, 4), %%mm4                        \n\t" // R4
2476
 
                "movq (%1, %%edx), %%mm5                        \n\t" // R5
2477
 
                "movq (%1, %%eax, 2), %%mm6                     \n\t" // R6
2478
 
                "movq (%1, %%ecx), %%mm7                        \n\t" // R7
 
2507
                "movq (%1, %%"REG_d"), %%mm5                    \n\t" // R5
 
2508
                "movq (%1, %%"REG_a", 2), %%mm6                 \n\t" // R6
 
2509
                "movq (%1, %%"REG_c"), %%mm7                    \n\t" // R7
2479
2510
                PAVGB(%%mm4, %%mm0)
2480
2511
                PAVGB(%%mm5, %%mm1)
2481
2512
                PAVGB(%%mm6, %%mm2)
2489
2520
                PAVGB(%%mm6, %%mm2)
2490
2521
                PAVGB(%%mm7, %%mm3)
2491
2522
                "movq %%mm0, (%1, %2, 4)                        \n\t" // R4
2492
 
                "movq %%mm1, (%1, %%edx)                        \n\t" // R5
2493
 
                "movq %%mm2, (%1, %%eax, 2)                     \n\t" // R6
2494
 
                "movq %%mm3, (%1, %%ecx)                        \n\t" // R7
 
2523
                "movq %%mm1, (%1, %%"REG_d")                    \n\t" // R5
 
2524
                "movq %%mm2, (%1, %%"REG_a", 2)                 \n\t" // R6
 
2525
                "movq %%mm3, (%1, %%"REG_c")                    \n\t" // R7
2495
2526
                "movq %%mm0, (%0, %2, 4)                        \n\t" // L4
2496
 
                "movq %%mm1, (%0, %%edx)                        \n\t" // L5
2497
 
                "movq %%mm2, (%0, %%eax, 2)                     \n\t" // L6
2498
 
                "movq %%mm3, (%0, %%ecx)                        \n\t" // L7
 
2527
                "movq %%mm1, (%0, %%"REG_d")                    \n\t" // L5
 
2528
                "movq %%mm2, (%0, %%"REG_a", 2)                 \n\t" // L6
 
2529
                "movq %%mm3, (%0, %%"REG_c")                    \n\t" // L7
2499
2530
 
2500
2531
                "4:                                             \n\t"
2501
2532
 
2502
 
                :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast)
2503
 
                : "%eax", "%edx", "%ecx", "memory"
 
2533
                :: "r" (src), "r" (tempBlured), "r"((long)stride), "m" (tempBluredPast)
 
2534
                : "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2504
2535
                );
2505
2536
//printf("%d\n", test);
2506
2537
#else
2507
2538
{
2508
2539
        int y;
2509
2540
        int d=0;
2510
 
        int sysd=0;
 
2541
//      int sysd=0;
2511
2542
        int i;
2512
2543
 
2513
2544
        for(y=0; y<8; y++)
2522
2553
//                      if(y==0 || y==7) d1+= d1>>1;
2523
2554
//                      d+= ABS(d1);
2524
2555
                        d+= d1*d1;
2525
 
                        sysd+= d1;
 
2556
//                      sysd+= d1;
2526
2557
                }
2527
2558
        }
2528
2559
        i=d;
2608
2639
}
2609
2640
#endif
2610
2641
}
 
2642
#endif //HAVE_ALTIVEC
 
2643
 
 
2644
#ifdef HAVE_MMX
 
2645
/**
 
2646
 * accurate deblock filter
 
2647
 */
 
2648
static always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){
 
2649
        int64_t dc_mask, eq_mask, both_masks;
 
2650
        int64_t sums[10*8*2];
 
2651
        src+= step*3; // src points to begin of the 8x8 Block
 
2652
//START_TIMER
 
2653
asm volatile(
 
2654
                "movq %0, %%mm7                                 \n\t" 
 
2655
                "movq %1, %%mm6                                 \n\t" 
 
2656
                : : "m" (c->mmxDcOffset[c->nonBQP]),  "m" (c->mmxDcThreshold[c->nonBQP])
 
2657
                );
 
2658
                
 
2659
asm volatile(
 
2660
                "lea (%2, %3), %%"REG_a"                        \n\t"
 
2661
//      0       1       2       3       4       5       6       7       8       9
 
2662
//      %1      eax     eax+%2  eax+2%2 %1+4%2  ecx     ecx+%2  ecx+2%2 %1+8%2  ecx+4%2
 
2663
 
 
2664
                "movq (%2), %%mm0                               \n\t"
 
2665
                "movq (%%"REG_a"), %%mm1                        \n\t"
 
2666
                "movq %%mm1, %%mm3                              \n\t"
 
2667
                "movq %%mm1, %%mm4                              \n\t"
 
2668
                "psubb %%mm1, %%mm0                             \n\t" // mm0 = differnece
 
2669
                "paddb %%mm7, %%mm0                             \n\t"
 
2670
                "pcmpgtb %%mm6, %%mm0                           \n\t"
 
2671
 
 
2672
                "movq (%%"REG_a",%3), %%mm2                     \n\t"
 
2673
                PMAXUB(%%mm2, %%mm4)
 
2674
                PMINUB(%%mm2, %%mm3, %%mm5)
 
2675
                "psubb %%mm2, %%mm1                             \n\t"
 
2676
                "paddb %%mm7, %%mm1                             \n\t"
 
2677
                "pcmpgtb %%mm6, %%mm1                           \n\t"
 
2678
                "paddb %%mm1, %%mm0                             \n\t"
 
2679
 
 
2680
                "movq (%%"REG_a", %3, 2), %%mm1                 \n\t"
 
2681
                PMAXUB(%%mm1, %%mm4)
 
2682
                PMINUB(%%mm1, %%mm3, %%mm5)
 
2683
                "psubb %%mm1, %%mm2                             \n\t"
 
2684
                "paddb %%mm7, %%mm2                             \n\t"
 
2685
                "pcmpgtb %%mm6, %%mm2                           \n\t"
 
2686
                "paddb %%mm2, %%mm0                             \n\t"
 
2687
                
 
2688
                "lea (%%"REG_a", %3, 4), %%"REG_a"              \n\t"
 
2689
 
 
2690
                "movq (%2, %3, 4), %%mm2                        \n\t"
 
2691
                PMAXUB(%%mm2, %%mm4)
 
2692
                PMINUB(%%mm2, %%mm3, %%mm5)
 
2693
                "psubb %%mm2, %%mm1                             \n\t"
 
2694
                "paddb %%mm7, %%mm1                             \n\t"
 
2695
                "pcmpgtb %%mm6, %%mm1                           \n\t"
 
2696
                "paddb %%mm1, %%mm0                             \n\t"
 
2697
 
 
2698
                "movq (%%"REG_a"), %%mm1                        \n\t"
 
2699
                PMAXUB(%%mm1, %%mm4)
 
2700
                PMINUB(%%mm1, %%mm3, %%mm5)
 
2701
                "psubb %%mm1, %%mm2                             \n\t"
 
2702
                "paddb %%mm7, %%mm2                             \n\t"
 
2703
                "pcmpgtb %%mm6, %%mm2                           \n\t"
 
2704
                "paddb %%mm2, %%mm0                             \n\t"
 
2705
 
 
2706
                "movq (%%"REG_a", %3), %%mm2                    \n\t"
 
2707
                PMAXUB(%%mm2, %%mm4)
 
2708
                PMINUB(%%mm2, %%mm3, %%mm5)
 
2709
                "psubb %%mm2, %%mm1                             \n\t"
 
2710
                "paddb %%mm7, %%mm1                             \n\t"
 
2711
                "pcmpgtb %%mm6, %%mm1                           \n\t"
 
2712
                "paddb %%mm1, %%mm0                             \n\t"
 
2713
 
 
2714
                "movq (%%"REG_a", %3, 2), %%mm1                 \n\t"
 
2715
                PMAXUB(%%mm1, %%mm4)
 
2716
                PMINUB(%%mm1, %%mm3, %%mm5)
 
2717
                "psubb %%mm1, %%mm2                             \n\t"
 
2718
                "paddb %%mm7, %%mm2                             \n\t"
 
2719
                "pcmpgtb %%mm6, %%mm2                           \n\t"
 
2720
                "paddb %%mm2, %%mm0                             \n\t"
 
2721
 
 
2722
                "movq (%2, %3, 8), %%mm2                        \n\t"
 
2723
                PMAXUB(%%mm2, %%mm4)
 
2724
                PMINUB(%%mm2, %%mm3, %%mm5)
 
2725
                "psubb %%mm2, %%mm1                             \n\t"
 
2726
                "paddb %%mm7, %%mm1                             \n\t"
 
2727
                "pcmpgtb %%mm6, %%mm1                           \n\t"
 
2728
                "paddb %%mm1, %%mm0                             \n\t"
 
2729
 
 
2730
                "movq (%%"REG_a", %3, 4), %%mm1                 \n\t"
 
2731
                "psubb %%mm1, %%mm2                             \n\t"
 
2732
                "paddb %%mm7, %%mm2                             \n\t"
 
2733
                "pcmpgtb %%mm6, %%mm2                           \n\t"
 
2734
                "paddb %%mm2, %%mm0                             \n\t"
 
2735
                "psubusb %%mm3, %%mm4                           \n\t"
 
2736
 
 
2737
                "pxor %%mm6, %%mm6                              \n\t"
 
2738
                "movq %4, %%mm7                                 \n\t" // QP,..., QP
 
2739
                "paddusb %%mm7, %%mm7                           \n\t" // 2QP ... 2QP
 
2740
                "psubusb %%mm4, %%mm7                           \n\t" // Diff >=2QP -> 0
 
2741
                "pcmpeqb %%mm6, %%mm7                           \n\t" // Diff < 2QP -> 0
 
2742
                "pcmpeqb %%mm6, %%mm7                           \n\t" // Diff < 2QP -> 0
 
2743
                "movq %%mm7, %1                                 \n\t"
 
2744
 
 
2745
                "movq %5, %%mm7                                 \n\t"
 
2746
                "punpcklbw %%mm7, %%mm7                         \n\t"
 
2747
                "punpcklbw %%mm7, %%mm7                         \n\t"
 
2748
                "punpcklbw %%mm7, %%mm7                         \n\t"
 
2749
                "psubb %%mm0, %%mm6                             \n\t"
 
2750
                "pcmpgtb %%mm7, %%mm6                           \n\t"
 
2751
                "movq %%mm6, %0                                 \n\t"
 
2752
 
 
2753
                : "=m" (eq_mask), "=m" (dc_mask)
 
2754
                : "r" (src), "r" ((long)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
 
2755
                : "%"REG_a
 
2756
                );
 
2757
 
 
2758
        both_masks = dc_mask & eq_mask;
 
2759
 
 
2760
        if(both_masks){
 
2761
                long offset= -8*step;
 
2762
                int64_t *temp_sums= sums;
 
2763
 
 
2764
                asm volatile(
 
2765
                "movq %2, %%mm0                                 \n\t"  // QP,..., QP
 
2766
                "pxor %%mm4, %%mm4                              \n\t"
 
2767
 
 
2768
                "movq (%0), %%mm6                               \n\t"
 
2769
                "movq (%0, %1), %%mm5                           \n\t"
 
2770
                "movq %%mm5, %%mm1                              \n\t"
 
2771
                "movq %%mm6, %%mm2                              \n\t"
 
2772
                "psubusb %%mm6, %%mm5                           \n\t"
 
2773
                "psubusb %%mm1, %%mm2                           \n\t"
 
2774
                "por %%mm5, %%mm2                               \n\t" // ABS Diff of lines
 
2775
                "psubusb %%mm2, %%mm0                           \n\t" // diff >= QP -> 0
 
2776
                "pcmpeqb %%mm4, %%mm0                           \n\t" // diff >= QP -> FF
 
2777
 
 
2778
                "pxor %%mm6, %%mm1                              \n\t"
 
2779
                "pand %%mm0, %%mm1                              \n\t"
 
2780
                "pxor %%mm1, %%mm6                              \n\t"
 
2781
                // 0:QP  6:First
 
2782
 
 
2783
                "movq (%0, %1, 8), %%mm5                        \n\t"
 
2784
                "add %1, %0                                     \n\t" // %0 points to line 1 not 0
 
2785
                "movq (%0, %1, 8), %%mm7                        \n\t"
 
2786
                "movq %%mm5, %%mm1                              \n\t"
 
2787
                "movq %%mm7, %%mm2                              \n\t"
 
2788
                "psubusb %%mm7, %%mm5                           \n\t"
 
2789
                "psubusb %%mm1, %%mm2                           \n\t"
 
2790
                "por %%mm5, %%mm2                               \n\t" // ABS Diff of lines
 
2791
                "movq %2, %%mm0                                 \n\t"  // QP,..., QP
 
2792
                "psubusb %%mm2, %%mm0                           \n\t" // diff >= QP -> 0
 
2793
                "pcmpeqb %%mm4, %%mm0                           \n\t" // diff >= QP -> FF
 
2794
 
 
2795
                "pxor %%mm7, %%mm1                              \n\t"
 
2796
                "pand %%mm0, %%mm1                              \n\t"
 
2797
                "pxor %%mm1, %%mm7                              \n\t"
 
2798
                
 
2799
                "movq %%mm6, %%mm5                              \n\t"
 
2800
                "punpckhbw %%mm4, %%mm6                         \n\t"
 
2801
                "punpcklbw %%mm4, %%mm5                         \n\t"
 
2802
                // 4:0 5/6:First 7:Last
 
2803
 
 
2804
                "movq %%mm5, %%mm0                              \n\t"
 
2805
                "movq %%mm6, %%mm1                              \n\t"
 
2806
                "psllw $2, %%mm0                                \n\t"
 
2807
                "psllw $2, %%mm1                                \n\t"
 
2808
                "paddw "MANGLE(w04)", %%mm0                     \n\t"
 
2809
                "paddw "MANGLE(w04)", %%mm1                     \n\t"
 
2810
 
 
2811
#define NEXT\
 
2812
                "movq (%0), %%mm2                               \n\t"\
 
2813
                "movq (%0), %%mm3                               \n\t"\
 
2814
                "add %1, %0                                     \n\t"\
 
2815
                "punpcklbw %%mm4, %%mm2                         \n\t"\
 
2816
                "punpckhbw %%mm4, %%mm3                         \n\t"\
 
2817
                "paddw %%mm2, %%mm0                             \n\t"\
 
2818
                "paddw %%mm3, %%mm1                             \n\t"
 
2819
 
 
2820
#define PREV\
 
2821
                "movq (%0), %%mm2                               \n\t"\
 
2822
                "movq (%0), %%mm3                               \n\t"\
 
2823
                "add %1, %0                                     \n\t"\
 
2824
                "punpcklbw %%mm4, %%mm2                         \n\t"\
 
2825
                "punpckhbw %%mm4, %%mm3                         \n\t"\
 
2826
                "psubw %%mm2, %%mm0                             \n\t"\
 
2827
                "psubw %%mm3, %%mm1                             \n\t"
 
2828
 
 
2829
                                
 
2830
                NEXT //0
 
2831
                NEXT //1
 
2832
                NEXT //2
 
2833
                "movq %%mm0, (%3)                               \n\t"
 
2834
                "movq %%mm1, 8(%3)                              \n\t"
 
2835
 
 
2836
                NEXT //3
 
2837
                "psubw %%mm5, %%mm0                             \n\t"
 
2838
                "psubw %%mm6, %%mm1                             \n\t"
 
2839
                "movq %%mm0, 16(%3)                             \n\t"
 
2840
                "movq %%mm1, 24(%3)                             \n\t"
 
2841
 
 
2842
                NEXT //4
 
2843
                "psubw %%mm5, %%mm0                             \n\t"
 
2844
                "psubw %%mm6, %%mm1                             \n\t"
 
2845
                "movq %%mm0, 32(%3)                             \n\t"
 
2846
                "movq %%mm1, 40(%3)                             \n\t"
 
2847
 
 
2848
                NEXT //5
 
2849
                "psubw %%mm5, %%mm0                             \n\t"
 
2850
                "psubw %%mm6, %%mm1                             \n\t"
 
2851
                "movq %%mm0, 48(%3)                             \n\t"
 
2852
                "movq %%mm1, 56(%3)                             \n\t"
 
2853
 
 
2854
                NEXT //6
 
2855
                "psubw %%mm5, %%mm0                             \n\t"
 
2856
                "psubw %%mm6, %%mm1                             \n\t"
 
2857
                "movq %%mm0, 64(%3)                             \n\t"
 
2858
                "movq %%mm1, 72(%3)                             \n\t"
 
2859
 
 
2860
                "movq %%mm7, %%mm6                              \n\t"
 
2861
                "punpckhbw %%mm4, %%mm7                         \n\t"
 
2862
                "punpcklbw %%mm4, %%mm6                         \n\t"
 
2863
                
 
2864
                NEXT //7
 
2865
                "mov %4, %0                                     \n\t"
 
2866
                "add %1, %0                                     \n\t"
 
2867
                PREV //0
 
2868
                "movq %%mm0, 80(%3)                             \n\t"
 
2869
                "movq %%mm1, 88(%3)                             \n\t"
 
2870
 
 
2871
                PREV //1
 
2872
                "paddw %%mm6, %%mm0                             \n\t"
 
2873
                "paddw %%mm7, %%mm1                             \n\t"
 
2874
                "movq %%mm0, 96(%3)                             \n\t"
 
2875
                "movq %%mm1, 104(%3)                            \n\t"
 
2876
                
 
2877
                PREV //2
 
2878
                "paddw %%mm6, %%mm0                             \n\t"
 
2879
                "paddw %%mm7, %%mm1                             \n\t"
 
2880
                "movq %%mm0, 112(%3)                            \n\t"
 
2881
                "movq %%mm1, 120(%3)                            \n\t"
 
2882
 
 
2883
                PREV //3
 
2884
                "paddw %%mm6, %%mm0                             \n\t"
 
2885
                "paddw %%mm7, %%mm1                             \n\t"
 
2886
                "movq %%mm0, 128(%3)                            \n\t"
 
2887
                "movq %%mm1, 136(%3)                            \n\t"
 
2888
 
 
2889
                PREV //4
 
2890
                "paddw %%mm6, %%mm0                             \n\t"
 
2891
                "paddw %%mm7, %%mm1                             \n\t"
 
2892
                "movq %%mm0, 144(%3)                            \n\t"
 
2893
                "movq %%mm1, 152(%3)                            \n\t"
 
2894
 
 
2895
                "mov %4, %0                                     \n\t" //FIXME
 
2896
 
 
2897
                : "+&r"(src)
 
2898
                : "r" ((long)step), "m" (c->pQPb), "r"(sums), "g"(src)
 
2899
                );
 
2900
 
 
2901
                src+= step; // src points to begin of the 8x8 Block
 
2902
 
 
2903
                asm volatile(
 
2904
                "movq %4, %%mm6                                 \n\t"
 
2905
                "pcmpeqb %%mm5, %%mm5                           \n\t"
 
2906
                "pxor %%mm6, %%mm5                              \n\t"
 
2907
                "pxor %%mm7, %%mm7                              \n\t"
 
2908
 
 
2909
                "1:                                             \n\t"
 
2910
                "movq (%1), %%mm0                               \n\t"
 
2911
                "movq 8(%1), %%mm1                              \n\t"
 
2912
                "paddw 32(%1), %%mm0                            \n\t"
 
2913
                "paddw 40(%1), %%mm1                            \n\t"
 
2914
                "movq (%0, %3), %%mm2                           \n\t"
 
2915
                "movq %%mm2, %%mm3                              \n\t"
 
2916
                "movq %%mm2, %%mm4                              \n\t"
 
2917
                "punpcklbw %%mm7, %%mm2                         \n\t"
 
2918
                "punpckhbw %%mm7, %%mm3                         \n\t"
 
2919
                "paddw %%mm2, %%mm0                             \n\t"
 
2920
                "paddw %%mm3, %%mm1                             \n\t"
 
2921
                "paddw %%mm2, %%mm0                             \n\t"
 
2922
                "paddw %%mm3, %%mm1                             \n\t"
 
2923
                "psrlw $4, %%mm0                                \n\t"
 
2924
                "psrlw $4, %%mm1                                \n\t"
 
2925
                "packuswb %%mm1, %%mm0                          \n\t"
 
2926
                "pand %%mm6, %%mm0                              \n\t"
 
2927
                "pand %%mm5, %%mm4                              \n\t"
 
2928
                "por %%mm4, %%mm0                               \n\t"
 
2929
                "movq %%mm0, (%0, %3)                           \n\t"
 
2930
                "add $16, %1                                    \n\t"
 
2931
                "add %2, %0                                     \n\t"
 
2932
                " js 1b                                         \n\t"
 
2933
 
 
2934
                : "+r"(offset), "+r"(temp_sums)
 
2935
                : "r" ((long)step), "r"(src - offset), "m"(both_masks)
 
2936
                );
 
2937
        }else
 
2938
                src+= step; // src points to begin of the 8x8 Block
 
2939
 
 
2940
        if(eq_mask != -1LL){
 
2941
                uint8_t *temp_src= src;
 
2942
                asm volatile(
 
2943
                "pxor %%mm7, %%mm7                              \n\t"
 
2944
                "lea -40(%%"REG_SP"), %%"REG_c"                 \n\t" // make space for 4 8-byte vars
 
2945
                "and "ALIGN_MASK", %%"REG_c"                    \n\t" // align
 
2946
//      0       1       2       3       4       5       6       7       8       9
 
2947
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1 %1+8%1  ecx+4%1
 
2948
 
 
2949
                "movq (%0), %%mm0                               \n\t"
 
2950
                "movq %%mm0, %%mm1                              \n\t"
 
2951
                "punpcklbw %%mm7, %%mm0                         \n\t" // low part of line 0
 
2952
                "punpckhbw %%mm7, %%mm1                         \n\t" // high part of line 0
 
2953
 
 
2954
                "movq (%0, %1), %%mm2                           \n\t"
 
2955
                "lea (%0, %1, 2), %%"REG_a"                     \n\t"
 
2956
                "movq %%mm2, %%mm3                              \n\t"
 
2957
                "punpcklbw %%mm7, %%mm2                         \n\t" // low part of line 1
 
2958
                "punpckhbw %%mm7, %%mm3                         \n\t" // high part of line 1
 
2959
 
 
2960
                "movq (%%"REG_a"), %%mm4                        \n\t"
 
2961
                "movq %%mm4, %%mm5                              \n\t"
 
2962
                "punpcklbw %%mm7, %%mm4                         \n\t" // low part of line 2
 
2963
                "punpckhbw %%mm7, %%mm5                         \n\t" // high part of line 2
 
2964
 
 
2965
                "paddw %%mm0, %%mm0                             \n\t" // 2L0
 
2966
                "paddw %%mm1, %%mm1                             \n\t" // 2H0
 
2967
                "psubw %%mm4, %%mm2                             \n\t" // L1 - L2
 
2968
                "psubw %%mm5, %%mm3                             \n\t" // H1 - H2
 
2969
                "psubw %%mm2, %%mm0                             \n\t" // 2L0 - L1 + L2
 
2970
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - H1 + H2
 
2971
 
 
2972
                "psllw $2, %%mm2                                \n\t" // 4L1 - 4L2
 
2973
                "psllw $2, %%mm3                                \n\t" // 4H1 - 4H2
 
2974
                "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2
 
2975
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2
 
2976
 
 
2977
                "movq (%%"REG_a", %1), %%mm2                    \n\t"
 
2978
                "movq %%mm2, %%mm3                              \n\t"
 
2979
                "punpcklbw %%mm7, %%mm2                         \n\t" // L3
 
2980
                "punpckhbw %%mm7, %%mm3                         \n\t" // H3
 
2981
 
 
2982
                "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - L3
 
2983
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - H3
 
2984
                "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 
2985
                "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 
2986
                "movq %%mm0, (%%"REG_c")                        \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 
2987
                "movq %%mm1, 8(%%"REG_c")                       \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 
2988
 
 
2989
                "movq (%%"REG_a", %1, 2), %%mm0                 \n\t"
 
2990
                "movq %%mm0, %%mm1                              \n\t"
 
2991
                "punpcklbw %%mm7, %%mm0                         \n\t" // L4
 
2992
                "punpckhbw %%mm7, %%mm1                         \n\t" // H4
 
2993
 
 
2994
                "psubw %%mm0, %%mm2                             \n\t" // L3 - L4
 
2995
                "psubw %%mm1, %%mm3                             \n\t" // H3 - H4
 
2996
                "movq %%mm2, 16(%%"REG_c")                      \n\t" // L3 - L4
 
2997
                "movq %%mm3, 24(%%"REG_c")                      \n\t" // H3 - H4
 
2998
                "paddw %%mm4, %%mm4                             \n\t" // 2L2
 
2999
                "paddw %%mm5, %%mm5                             \n\t" // 2H2
 
3000
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - L3 + L4
 
3001
                "psubw %%mm3, %%mm5                             \n\t" // 2H2 - H3 + H4
 
3002
 
 
3003
                "lea (%%"REG_a", %1), %0                        \n\t"
 
3004
                "psllw $2, %%mm2                                \n\t" // 4L3 - 4L4
 
3005
                "psllw $2, %%mm3                                \n\t" // 4H3 - 4H4
 
3006
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4
 
3007
                "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4
 
3008
//50 opcodes so far
 
3009
                "movq (%0, %1, 2), %%mm2                        \n\t"
 
3010
                "movq %%mm2, %%mm3                              \n\t"
 
3011
                "punpcklbw %%mm7, %%mm2                         \n\t" // L5
 
3012
                "punpckhbw %%mm7, %%mm3                         \n\t" // H5
 
3013
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - L5
 
3014
                "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - H5
 
3015
                "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - 2L5
 
3016
                "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - 2H5
 
3017
 
 
3018
                "movq (%%"REG_a", %1, 4), %%mm6                 \n\t"
 
3019
                "punpcklbw %%mm7, %%mm6                         \n\t" // L6
 
3020
                "psubw %%mm6, %%mm2                             \n\t" // L5 - L6
 
3021
                "movq (%%"REG_a", %1, 4), %%mm6                 \n\t"
 
3022
                "punpckhbw %%mm7, %%mm6                         \n\t" // H6
 
3023
                "psubw %%mm6, %%mm3                             \n\t" // H5 - H6
 
3024
 
 
3025
                "paddw %%mm0, %%mm0                             \n\t" // 2L4
 
3026
                "paddw %%mm1, %%mm1                             \n\t" // 2H4
 
3027
                "psubw %%mm2, %%mm0                             \n\t" // 2L4 - L5 + L6
 
3028
                "psubw %%mm3, %%mm1                             \n\t" // 2H4 - H5 + H6
 
3029
 
 
3030
                "psllw $2, %%mm2                                \n\t" // 4L5 - 4L6
 
3031
                "psllw $2, %%mm3                                \n\t" // 4H5 - 4H6
 
3032
                "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6
 
3033
                "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6
 
3034
 
 
3035
                "movq (%0, %1, 4), %%mm2                        \n\t"
 
3036
                "movq %%mm2, %%mm3                              \n\t"
 
3037
                "punpcklbw %%mm7, %%mm2                         \n\t" // L7
 
3038
                "punpckhbw %%mm7, %%mm3                         \n\t" // H7
 
3039
 
 
3040
                "paddw %%mm2, %%mm2                             \n\t" // 2L7
 
3041
                "paddw %%mm3, %%mm3                             \n\t" // 2H7
 
3042
                "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6 - 2L7
 
3043
                "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6 - 2H7
 
3044
 
 
3045
                "movq (%%"REG_c"), %%mm2                        \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 
3046
                "movq 8(%%"REG_c"), %%mm3                       \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 
3047
 
 
3048
#ifdef HAVE_MMX2
 
3049
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3050
                "psubw %%mm0, %%mm6                             \n\t"
 
3051
                "pmaxsw %%mm6, %%mm0                            \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
 
3052
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3053
                "psubw %%mm1, %%mm6                             \n\t"
 
3054
                "pmaxsw %%mm6, %%mm1                            \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
 
3055
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3056
                "psubw %%mm2, %%mm6                             \n\t"
 
3057
                "pmaxsw %%mm6, %%mm2                            \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
 
3058
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3059
                "psubw %%mm3, %%mm6                             \n\t"
 
3060
                "pmaxsw %%mm6, %%mm3                            \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
 
3061
#else
 
3062
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3063
                "pcmpgtw %%mm0, %%mm6                           \n\t"
 
3064
                "pxor %%mm6, %%mm0                              \n\t"
 
3065
                "psubw %%mm6, %%mm0                             \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
 
3066
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3067
                "pcmpgtw %%mm1, %%mm6                           \n\t"
 
3068
                "pxor %%mm6, %%mm1                              \n\t"
 
3069
                "psubw %%mm6, %%mm1                             \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
 
3070
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3071
                "pcmpgtw %%mm2, %%mm6                           \n\t"
 
3072
                "pxor %%mm6, %%mm2                              \n\t"
 
3073
                "psubw %%mm6, %%mm2                             \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
 
3074
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3075
                "pcmpgtw %%mm3, %%mm6                           \n\t"
 
3076
                "pxor %%mm6, %%mm3                              \n\t"
 
3077
                "psubw %%mm6, %%mm3                             \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
 
3078
#endif
 
3079
 
 
3080
#ifdef HAVE_MMX2
 
3081
                "pminsw %%mm2, %%mm0                            \n\t"
 
3082
                "pminsw %%mm3, %%mm1                            \n\t"
 
3083
#else
 
3084
                "movq %%mm0, %%mm6                              \n\t"
 
3085
                "psubusw %%mm2, %%mm6                           \n\t"
 
3086
                "psubw %%mm6, %%mm0                             \n\t"
 
3087
                "movq %%mm1, %%mm6                              \n\t"
 
3088
                "psubusw %%mm3, %%mm6                           \n\t"
 
3089
                "psubw %%mm6, %%mm1                             \n\t"
 
3090
#endif
 
3091
 
 
3092
                "movd %2, %%mm2                                 \n\t" // QP
 
3093
                "punpcklbw %%mm7, %%mm2                         \n\t"
 
3094
 
 
3095
                "movq %%mm7, %%mm6                              \n\t" // 0
 
3096
                "pcmpgtw %%mm4, %%mm6                           \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
 
3097
                "pxor %%mm6, %%mm4                              \n\t"
 
3098
                "psubw %%mm6, %%mm4                             \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
 
3099
                "pcmpgtw %%mm5, %%mm7                           \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
 
3100
                "pxor %%mm7, %%mm5                              \n\t"
 
3101
                "psubw %%mm7, %%mm5                             \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
 
3102
// 100 opcodes
 
3103
                "psllw $3, %%mm2                                \n\t" // 8QP
 
3104
                "movq %%mm2, %%mm3                              \n\t" // 8QP
 
3105
                "pcmpgtw %%mm4, %%mm2                           \n\t"
 
3106
                "pcmpgtw %%mm5, %%mm3                           \n\t"
 
3107
                "pand %%mm2, %%mm4                              \n\t"
 
3108
                "pand %%mm3, %%mm5                              \n\t"
 
3109
 
 
3110
 
 
3111
                "psubusw %%mm0, %%mm4                           \n\t" // hd
 
3112
                "psubusw %%mm1, %%mm5                           \n\t" // ld
 
3113
 
 
3114
 
 
3115
                "movq "MANGLE(w05)", %%mm2                      \n\t" // 5
 
3116
                "pmullw %%mm2, %%mm4                            \n\t"
 
3117
                "pmullw %%mm2, %%mm5                            \n\t"
 
3118
                "movq "MANGLE(w20)", %%mm2                      \n\t" // 32
 
3119
                "paddw %%mm2, %%mm4                             \n\t"
 
3120
                "paddw %%mm2, %%mm5                             \n\t"
 
3121
                "psrlw $6, %%mm4                                \n\t"
 
3122
                "psrlw $6, %%mm5                                \n\t"
 
3123
 
 
3124
                "movq 16(%%"REG_c"), %%mm0                      \n\t" // L3 - L4
 
3125
                "movq 24(%%"REG_c"), %%mm1                      \n\t" // H3 - H4
 
3126
 
 
3127
                "pxor %%mm2, %%mm2                              \n\t"
 
3128
                "pxor %%mm3, %%mm3                              \n\t"
 
3129
 
 
3130
                "pcmpgtw %%mm0, %%mm2                           \n\t" // sign (L3-L4)
 
3131
                "pcmpgtw %%mm1, %%mm3                           \n\t" // sign (H3-H4)
 
3132
                "pxor %%mm2, %%mm0                              \n\t"
 
3133
                "pxor %%mm3, %%mm1                              \n\t"
 
3134
                "psubw %%mm2, %%mm0                             \n\t" // |L3-L4|
 
3135
                "psubw %%mm3, %%mm1                             \n\t" // |H3-H4|
 
3136
                "psrlw $1, %%mm0                                \n\t" // |L3 - L4|/2
 
3137
                "psrlw $1, %%mm1                                \n\t" // |H3 - H4|/2
 
3138
 
 
3139
                "pxor %%mm6, %%mm2                              \n\t"
 
3140
                "pxor %%mm7, %%mm3                              \n\t"
 
3141
                "pand %%mm2, %%mm4                              \n\t"
 
3142
                "pand %%mm3, %%mm5                              \n\t"
 
3143
 
 
3144
#ifdef HAVE_MMX2
 
3145
                "pminsw %%mm0, %%mm4                            \n\t"
 
3146
                "pminsw %%mm1, %%mm5                            \n\t"
 
3147
#else
 
3148
                "movq %%mm4, %%mm2                              \n\t"
 
3149
                "psubusw %%mm0, %%mm2                           \n\t"
 
3150
                "psubw %%mm2, %%mm4                             \n\t"
 
3151
                "movq %%mm5, %%mm2                              \n\t"
 
3152
                "psubusw %%mm1, %%mm2                           \n\t"
 
3153
                "psubw %%mm2, %%mm5                             \n\t"
 
3154
#endif
 
3155
                "pxor %%mm6, %%mm4                              \n\t"
 
3156
                "pxor %%mm7, %%mm5                              \n\t"
 
3157
                "psubw %%mm6, %%mm4                             \n\t"
 
3158
                "psubw %%mm7, %%mm5                             \n\t"
 
3159
                "packsswb %%mm5, %%mm4                          \n\t"
 
3160
                "movq %3, %%mm1                                 \n\t"
 
3161
                "pandn %%mm4, %%mm1                             \n\t"
 
3162
                "movq (%0), %%mm0                               \n\t"
 
3163
                "paddb   %%mm1, %%mm0                           \n\t"
 
3164
                "movq %%mm0, (%0)                               \n\t"
 
3165
                "movq (%0, %1), %%mm0                           \n\t"
 
3166
                "psubb %%mm1, %%mm0                             \n\t"
 
3167
                "movq %%mm0, (%0, %1)                           \n\t"
 
3168
 
 
3169
                : "+r" (temp_src)
 
3170
                : "r" ((long)step), "m" (c->pQPb), "m"(eq_mask)
 
3171
                : "%"REG_a, "%"REG_c
 
3172
                );
 
3173
        }
 
3174
/*if(step==16){
 
3175
    STOP_TIMER("step16")
 
3176
}else{
 
3177
    STOP_TIMER("stepX")
 
3178
}*/
 
3179
}
 
3180
#endif //HAVE_MMX
2611
3181
 
2612
3182
static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2613
3183
        QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c);
2628
3198
        {
2629
3199
#ifdef HAVE_MMX
2630
3200
                                        asm volatile(
2631
 
                                                "movq (%%eax), %%mm2    \n\t" // packedYOffset
2632
 
                                                "movq 8(%%eax), %%mm3   \n\t" // packedYScale
2633
 
                                                "leal (%2,%4), %%eax    \n\t"
2634
 
                                                "leal (%3,%5), %%edx    \n\t"
 
3201
                                                "movq (%%"REG_a"), %%mm2        \n\t" // packedYOffset
 
3202
                                                "movq 8(%%"REG_a"), %%mm3       \n\t" // packedYScale
 
3203
                                                "lea (%2,%4), %%"REG_a" \n\t"
 
3204
                                                "lea (%3,%5), %%"REG_d" \n\t"
2635
3205
                                                "pxor %%mm4, %%mm4      \n\t"
2636
3206
#ifdef HAVE_MMX2
2637
 
#define SCALED_CPY(src1, src2, dst1, dst2)                                      \
 
3207
#define REAL_SCALED_CPY(src1, src2, dst1, dst2)                                 \
2638
3208
                                                "movq " #src1 ", %%mm0  \n\t"\
2639
3209
                                                "movq " #src1 ", %%mm5  \n\t"\
2640
3210
                                                "movq " #src2 ", %%mm1  \n\t"\
2657
3227
                                                "movq %%mm1, " #dst2 "  \n\t"\
2658
3228
 
2659
3229
#else //HAVE_MMX2
2660
 
#define SCALED_CPY(src1, src2, dst1, dst2)                                      \
 
3230
#define REAL_SCALED_CPY(src1, src2, dst1, dst2)                                 \
2661
3231
                                                "movq " #src1 ", %%mm0  \n\t"\
2662
3232
                                                "movq " #src1 ", %%mm5  \n\t"\
2663
3233
                                                "punpcklbw %%mm4, %%mm0 \n\t"\
2684
3254
                                                "movq %%mm1, " #dst2 "  \n\t"\
2685
3255
 
2686
3256
#endif //!HAVE_MMX2
 
3257
#define SCALED_CPY(src1, src2, dst1, dst2)\
 
3258
   REAL_SCALED_CPY(src1, src2, dst1, dst2)
2687
3259
 
2688
3260
SCALED_CPY((%2)       , (%2, %4)      , (%3)       , (%3, %5))
2689
 
SCALED_CPY((%2, %4, 2), (%%eax, %4, 2), (%3, %5, 2), (%%edx, %5, 2))
2690
 
SCALED_CPY((%2, %4, 4), (%%eax, %4, 4), (%3, %5, 4), (%%edx, %5, 4))
2691
 
                                                "leal (%%eax,%4,4), %%eax       \n\t"
2692
 
                                                "leal (%%edx,%5,4), %%edx       \n\t"
2693
 
SCALED_CPY((%%eax, %4), (%%eax, %4, 2), (%%edx, %5), (%%edx, %5, 2))
 
3261
SCALED_CPY((%2, %4, 2), (%%REGa, %4, 2), (%3, %5, 2), (%%REGd, %5, 2))
 
3262
SCALED_CPY((%2, %4, 4), (%%REGa, %4, 4), (%3, %5, 4), (%%REGd, %5, 4))
 
3263
                                                "lea (%%"REG_a",%4,4), %%"REG_a"        \n\t"
 
3264
                                                "lea (%%"REG_d",%5,4), %%"REG_d"        \n\t"
 
3265
SCALED_CPY((%%REGa, %4), (%%REGa, %4, 2), (%%REGd, %5), (%%REGd, %5, 2))
2694
3266
 
2695
3267
 
2696
3268
                                                : "=&a" (packedOffsetAndScale)
2697
3269
                                                : "0" (packedOffsetAndScale),
2698
3270
                                                "r"(src),
2699
3271
                                                "r"(dst),
2700
 
                                                "r" (srcStride),
2701
 
                                                "r" (dstStride)
2702
 
                                                : "%edx"
 
3272
                                                "r" ((long)srcStride),
 
3273
                                                "r" ((long)dstStride)
 
3274
                                                : "%"REG_d
2703
3275
                                        );
2704
3276
#else
2705
3277
                                for(i=0; i<8; i++)
2711
3283
        {
2712
3284
#ifdef HAVE_MMX
2713
3285
                                        asm volatile(
2714
 
                                                "leal (%0,%2), %%eax    \n\t"
2715
 
                                                "leal (%1,%3), %%edx    \n\t"
 
3286
                                                "lea (%0,%2), %%"REG_a" \n\t"
 
3287
                                                "lea (%1,%3), %%"REG_d" \n\t"
2716
3288
 
2717
 
#define SIMPLE_CPY(src1, src2, dst1, dst2)                              \
 
3289
#define REAL_SIMPLE_CPY(src1, src2, dst1, dst2)                         \
2718
3290
                                                "movq " #src1 ", %%mm0  \n\t"\
2719
3291
                                                "movq " #src2 ", %%mm1  \n\t"\
2720
3292
                                                "movq %%mm0, " #dst1 "  \n\t"\
2721
3293
                                                "movq %%mm1, " #dst2 "  \n\t"\
2722
3294
 
 
3295
#define SIMPLE_CPY(src1, src2, dst1, dst2)\
 
3296
   REAL_SIMPLE_CPY(src1, src2, dst1, dst2)
 
3297
 
2723
3298
SIMPLE_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
2724
 
SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%edx, %3, 2))
2725
 
SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%edx, %3, 4))
2726
 
                                                "leal (%%eax,%2,4), %%eax       \n\t"
2727
 
                                                "leal (%%edx,%3,4), %%edx       \n\t"
2728
 
SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%edx, %3), (%%edx, %3, 2))
 
3299
SIMPLE_CPY((%0, %2, 2), (%%REGa, %2, 2), (%1, %3, 2), (%%REGd, %3, 2))
 
3300
SIMPLE_CPY((%0, %2, 4), (%%REGa, %2, 4), (%1, %3, 4), (%%REGd, %3, 4))
 
3301
                                                "lea (%%"REG_a",%2,4), %%"REG_a"        \n\t"
 
3302
                                                "lea (%%"REG_d",%3,4), %%"REG_d"        \n\t"
 
3303
SIMPLE_CPY((%%REGa, %2), (%%REGa, %2, 2), (%%REGd, %3), (%%REGd, %3, 2))
2729
3304
 
2730
3305
                                                : : "r" (src),
2731
3306
                                                "r" (dst),
2732
 
                                                "r" (srcStride),
2733
 
                                                "r" (dstStride)
2734
 
                                                : "%eax", "%edx"
 
3307
                                                "r" ((long)srcStride),
 
3308
                                                "r" ((long)dstStride)
 
3309
                                                : "%"REG_a, "%"REG_d
2735
3310
                                        );
2736
3311
#else
2737
3312
                                for(i=0; i<8; i++)
2749
3324
#ifdef HAVE_MMX
2750
3325
        asm volatile(
2751
3326
                "movq (%0), %%mm0               \n\t"
2752
 
                "addl %1, %0                    \n\t"
 
3327
                "add %1, %0                     \n\t"
2753
3328
                "movq %%mm0, (%0)               \n\t"
2754
3329
                "movq %%mm0, (%0, %1)           \n\t"
2755
3330
                "movq %%mm0, (%0, %1, 2)        \n\t"
2756
3331
                : "+r" (src)
2757
 
                : "r" (-stride)
 
3332
                : "r" ((long)-stride)
2758
3333
        );
2759
3334
#else
2760
3335
        int i;
2793
3368
 
2794
3369
        //FIXME remove
2795
3370
        uint64_t * const yHistogram= c.yHistogram;
2796
 
        uint8_t * const tempSrc= c.tempSrc;
2797
 
        uint8_t * const tempDst= c.tempDst;
 
3371
        uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride;
 
3372
        uint8_t * const tempDst= dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride;
2798
3373
        //const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
2799
3374
 
2800
3375
#ifdef HAVE_MMX
2814
3389
                || (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14;
2815
3390
        else if(   (mode & V_DEBLOCK)
2816
3391
                || (mode & LINEAR_IPOL_DEINT_FILTER)
2817
 
                || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
 
3392
                || (mode & MEDIAN_DEINT_FILTER)
 
3393
                || (mode & V_A_DEBLOCK)) copyAhead=13;
2818
3394
        else if(mode & V_X1_FILTER) copyAhead=11;
2819
3395
//      else if(mode & V_RK1_FILTER) copyAhead=10;
2820
3396
        else if(mode & DERING) copyAhead=9;
2891
3467
                uint8_t *dstBlock= tempDst + dstStride;
2892
3468
 
2893
3469
                // From this point on it is guranteed that we can read and write 16 lines downward
2894
 
                // finish 1 block before the next otherwise we�ll might have a problem
 
3470
                // finish 1 block before the next otherwise we might have a problem
2895
3471
                // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
2896
3472
                for(x=0; x<width; x+=BLOCK_SIZE)
2897
3473
                {
2905
3481
*/
2906
3482
 
2907
3483
                        asm(
2908
 
                                "movl %4, %%eax                 \n\t"
2909
 
                                "shrl $2, %%eax                 \n\t"
2910
 
                                "andl $6, %%eax                 \n\t"
2911
 
                                "addl %5, %%eax                 \n\t"
2912
 
                                "movl %%eax, %%edx              \n\t"
2913
 
                                "imul %1, %%eax                 \n\t"
2914
 
                                "imul %3, %%edx                 \n\t"
2915
 
                                "prefetchnta 32(%%eax, %0)      \n\t"
2916
 
                                "prefetcht0 32(%%edx, %2)       \n\t"
2917
 
                                "addl %1, %%eax                 \n\t"
2918
 
                                "addl %3, %%edx                 \n\t"
2919
 
                                "prefetchnta 32(%%eax, %0)      \n\t"
2920
 
                                "prefetcht0 32(%%edx, %2)       \n\t"
2921
 
                        :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
2922
 
                        "m" (x), "m" (copyAhead)
2923
 
                        : "%eax", "%edx"
 
3484
                                "mov %4, %%"REG_a"              \n\t"
 
3485
                                "shr $2, %%"REG_a"              \n\t"
 
3486
                                "and $6, %%"REG_a"              \n\t"
 
3487
                                "add %5, %%"REG_a"              \n\t"
 
3488
                                "mov %%"REG_a", %%"REG_d"       \n\t"
 
3489
                                "imul %1, %%"REG_a"             \n\t"
 
3490
                                "imul %3, %%"REG_d"             \n\t"
 
3491
                                "prefetchnta 32(%%"REG_a", %0)  \n\t"
 
3492
                                "prefetcht0 32(%%"REG_d", %2)   \n\t"
 
3493
                                "add %1, %%"REG_a"              \n\t"
 
3494
                                "add %3, %%"REG_d"              \n\t"
 
3495
                                "prefetchnta 32(%%"REG_a", %0)  \n\t"
 
3496
                                "prefetcht0 32(%%"REG_d", %2)   \n\t"
 
3497
                        :: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride),
 
3498
                        "g" ((long)x), "g" ((long)copyAhead)
 
3499
                        : "%"REG_a, "%"REG_d
2924
3500
                        );
2925
3501
 
2926
3502
#elif defined(HAVE_3DNOW)
2955
3531
                        dstBlock+=8;
2956
3532
                        srcBlock+=8;
2957
3533
                }
2958
 
                if(width==dstStride)
2959
 
                        memcpy(dst, tempDst + 9*dstStride, copyAhead*dstStride);
 
3534
                if(width==ABS(dstStride))
 
3535
                        linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride);
2960
3536
                else
2961
3537
                {
2962
3538
                        int i;
2978
3554
                uint8_t *tempBlock2= c.tempBlocks + 8;
2979
3555
#endif
2980
3556
                int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
2981
 
                int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*QPStride];
 
3557
                int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*ABS(QPStride)];
2982
3558
                int QP=0;
2983
3559
                /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
2984
3560
                   if not than use a temporary buffer */
2987
3563
                        int i;
2988
3564
                        /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
2989
3565
                           blockcopy to dst later */
2990
 
                        memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
2991
 
                                srcStride*MAX(height-y-copyAhead, 0) );
 
3566
                        linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
 
3567
                                MAX(height-y-copyAhead, 0), srcStride);
2992
3568
 
2993
3569
                        /* duplicate last line of src to fill the void upto line (copyAhead+7) */
2994
3570
                        for(i=MAX(height-y, 8); i<copyAhead+8; i++)
2995
 
                                memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);
 
3571
                                memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), ABS(srcStride));
2996
3572
 
2997
3573
                        /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
2998
 
                        memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );
 
3574
                        linecpy(tempDst, dstBlock - dstStride, MIN(height-y+1, copyAhead+1), dstStride);
2999
3575
 
3000
3576
                        /* duplicate last line of dst to fill the void upto line (copyAhead) */
3001
3577
                        for(i=height-y+1; i<=copyAhead; i++)
3002
 
                                memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);
 
3578
                                memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), ABS(dstStride));
3003
3579
 
3004
3580
                        dstBlock= tempDst + dstStride;
3005
3581
                        srcBlock= tempSrc;
3007
3583
//printf("\n");
3008
3584
 
3009
3585
                // From this point on it is guranteed that we can read and write 16 lines downward
3010
 
                // finish 1 block before the next otherwise we�ll might have a problem
 
3586
                // finish 1 block before the next otherwise we might have a problem
3011
3587
                // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3012
3588
                for(x=0; x<width; x+=BLOCK_SIZE)
3013
3589
                {
3051
3627
*/
3052
3628
 
3053
3629
                        asm(
3054
 
                                "movl %4, %%eax                 \n\t"
3055
 
                                "shrl $2, %%eax                 \n\t"
3056
 
                                "andl $6, %%eax                 \n\t"
3057
 
                                "addl %5, %%eax                 \n\t"
3058
 
                                "movl %%eax, %%edx              \n\t"
3059
 
                                "imul %1, %%eax                 \n\t"
3060
 
                                "imul %3, %%edx                 \n\t"
3061
 
                                "prefetchnta 32(%%eax, %0)      \n\t"
3062
 
                                "prefetcht0 32(%%edx, %2)       \n\t"
3063
 
                                "addl %1, %%eax                 \n\t"
3064
 
                                "addl %3, %%edx                 \n\t"
3065
 
                                "prefetchnta 32(%%eax, %0)      \n\t"
3066
 
                                "prefetcht0 32(%%edx, %2)       \n\t"
3067
 
                        :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
3068
 
                        "m" (x), "m" (copyAhead)
3069
 
                        : "%eax", "%edx"
 
3630
                                "mov %4, %%"REG_a"              \n\t"
 
3631
                                "shr $2, %%"REG_a"              \n\t"
 
3632
                                "and $6, %%"REG_a"              \n\t"
 
3633
                                "add %5, %%"REG_a"              \n\t"
 
3634
                                "mov %%"REG_a", %%"REG_d"       \n\t"
 
3635
                                "imul %1, %%"REG_a"             \n\t"
 
3636
                                "imul %3, %%"REG_d"             \n\t"
 
3637
                                "prefetchnta 32(%%"REG_a", %0)  \n\t"
 
3638
                                "prefetcht0 32(%%"REG_d", %2)   \n\t"
 
3639
                                "add %1, %%"REG_a"              \n\t"
 
3640
                                "add %3, %%"REG_d"              \n\t"
 
3641
                                "prefetchnta 32(%%"REG_a", %0)  \n\t"
 
3642
                                "prefetcht0 32(%%"REG_d", %2)   \n\t"
 
3643
                        :: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride),
 
3644
                         "g" ((long)x), "g" ((long)copyAhead)
 
3645
                        : "%"REG_a, "%"REG_d
3070
3646
                        );
3071
3647
 
3072
3648
#elif defined(HAVE_3DNOW)
3110
3686
                                                RENAME(doVertLowPass)(dstBlock, stride, &c);
3111
3687
                                        else if(t==2)
3112
3688
                                                RENAME(doVertDefFilter)(dstBlock, stride, &c);
 
3689
                                }else if(mode & V_A_DEBLOCK){
 
3690
                                        RENAME(do_a_deblock)(dstBlock, stride, 1, &c);
3113
3691
                                }
3114
3692
                        }
3115
3693
 
3131
3709
                                                RENAME(doVertLowPass)(tempBlock1, 16, &c);
3132
3710
                                        else if(t==2)
3133
3711
                                                RENAME(doVertDefFilter)(tempBlock1, 16, &c);
 
3712
                                }else if(mode & H_A_DEBLOCK){
 
3713
                                        RENAME(do_a_deblock)(tempBlock1, 16, 1, &c);
3134
3714
                                }
3135
3715
 
3136
3716
                                RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
3140
3720
                                        horizX1Filter(dstBlock-4, stride, QP);
3141
3721
                                else if(mode & H_DEBLOCK)
3142
3722
                                {
 
3723
#ifdef HAVE_ALTIVEC
 
3724
                                        unsigned char __attribute__ ((aligned(16))) tempBlock[272];
 
3725
                                        transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride);
 
3726
 
 
3727
                                        const int t=vertClassify_altivec(tempBlock-48, 16, &c);
 
3728
                                        if(t==1) {
 
3729
                                                doVertLowPass_altivec(tempBlock-48, 16, &c);
 
3730
                                                transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
 
3731
                                        }
 
3732
                                        else if(t==2) {
 
3733
                                                doVertDefFilter_altivec(tempBlock-48, 16, &c);
 
3734
                                                transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
 
3735
                                        }
 
3736
#else
3143
3737
                                        const int t= RENAME(horizClassify)(dstBlock-4, stride, &c);
3144
3738
 
3145
3739
                                        if(t==1)
3146
3740
                                                RENAME(doHorizLowPass)(dstBlock-4, stride, &c);
3147
3741
                                        else if(t==2)
3148
3742
                                                RENAME(doHorizDefFilter)(dstBlock-4, stride, &c);
 
3743
#endif
 
3744
                                }else if(mode & H_A_DEBLOCK){
 
3745
                                        RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c);
3149
3746
                                }
3150
3747
#endif
3151
3748
                                if(mode & DERING)
3190
3787
                if(y+15 >= height)
3191
3788
                {
3192
3789
                        uint8_t *dstBlock= &(dst[y*dstStride]);
3193
 
                        if(width==dstStride)
3194
 
                                memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y));
 
3790
                        if(width==ABS(dstStride))
 
3791
                                linecpy(dstBlock, tempDst + dstStride, height-y, dstStride);
3195
3792
                        else
3196
3793
                        {
3197
3794
                                int i;