194
176
* Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
195
177
* using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
198
180
static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
200
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
202
asm volatile( //"movv %0 %1 %2\n\t"
203
"movq %2, %%mm0 \n\t" // QP,..., QP
204
"pxor %%mm4, %%mm4 \n\t"
206
"movq (%0), %%mm6 \n\t"
207
"movq (%0, %1), %%mm5 \n\t"
208
"movq %%mm5, %%mm1 \n\t"
209
"movq %%mm6, %%mm2 \n\t"
210
"psubusb %%mm6, %%mm5 \n\t"
211
"psubusb %%mm1, %%mm2 \n\t"
212
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
213
"psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
214
"pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
216
"pand %%mm2, %%mm6 \n\t"
217
"pandn %%mm1, %%mm2 \n\t"
218
"por %%mm2, %%mm6 \n\t"// First Line to Filter
220
"movq (%0, %1, 8), %%mm5 \n\t"
221
"lea (%0, %1, 4), %%"REG_a" \n\t"
222
"lea (%0, %1, 8), %%"REG_c" \n\t"
223
"sub %1, %%"REG_c" \n\t"
224
"add %1, %0 \n\t" // %0 points to line 1 not 0
225
"movq (%0, %1, 8), %%mm7 \n\t"
226
"movq %%mm5, %%mm1 \n\t"
227
"movq %%mm7, %%mm2 \n\t"
228
"psubusb %%mm7, %%mm5 \n\t"
229
"psubusb %%mm1, %%mm2 \n\t"
230
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
231
"psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
232
"pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
234
"pand %%mm2, %%mm7 \n\t"
235
"pandn %%mm1, %%mm2 \n\t"
236
"por %%mm2, %%mm7 \n\t" // First Line to Filter
240
// %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
245
"movq (%0, %1), %%mm0 \n\t" // 1
246
"movq %%mm0, %%mm1 \n\t" // 1
247
PAVGB(%%mm6, %%mm0) //1 1 /2
248
PAVGB(%%mm6, %%mm0) //3 1 /4
250
"movq (%0, %1, 4), %%mm2 \n\t" // 1
251
"movq %%mm2, %%mm5 \n\t" // 1
252
PAVGB((%%REGa), %%mm2) // 11 /2
253
PAVGB((%0, %1, 2), %%mm2) // 211 /4
254
"movq %%mm2, %%mm3 \n\t" // 211 /4
255
"movq (%0), %%mm4 \n\t" // 1
256
PAVGB(%%mm4, %%mm3) // 4 211 /8
257
PAVGB(%%mm0, %%mm3) //642211 /16
258
"movq %%mm3, (%0) \n\t" // X
259
// mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
260
"movq %%mm1, %%mm0 \n\t" // 1
261
PAVGB(%%mm6, %%mm0) //1 1 /2
262
"movq %%mm4, %%mm3 \n\t" // 1
263
PAVGB((%0,%1,2), %%mm3) // 1 1 /2
264
PAVGB((%%REGa,%1,2), %%mm5) // 11 /2
265
PAVGB((%%REGa), %%mm5) // 211 /4
266
PAVGB(%%mm5, %%mm3) // 2 2211 /8
267
PAVGB(%%mm0, %%mm3) //4242211 /16
268
"movq %%mm3, (%0,%1) \n\t" // X
269
// mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
270
PAVGB(%%mm4, %%mm6) //11 /2
271
"movq (%%"REG_c"), %%mm0 \n\t" // 1
272
PAVGB((%%REGa, %1, 2), %%mm0) // 11/2
273
"movq %%mm0, %%mm3 \n\t" // 11/2
274
PAVGB(%%mm1, %%mm0) // 2 11/4
275
PAVGB(%%mm6, %%mm0) //222 11/8
276
PAVGB(%%mm2, %%mm0) //22242211/16
277
"movq (%0, %1, 2), %%mm2 \n\t" // 1
278
"movq %%mm0, (%0, %1, 2) \n\t" // X
279
// mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
280
"movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
281
PAVGB((%%REGc), %%mm0) // 11 /2
282
PAVGB(%%mm0, %%mm6) //11 11 /4
283
PAVGB(%%mm1, %%mm4) // 11 /2
284
PAVGB(%%mm2, %%mm1) // 11 /2
285
PAVGB(%%mm1, %%mm6) //1122 11 /8
286
PAVGB(%%mm5, %%mm6) //112242211 /16
287
"movq (%%"REG_a"), %%mm5 \n\t" // 1
288
"movq %%mm6, (%%"REG_a") \n\t" // X
289
// mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
290
"movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1
291
PAVGB(%%mm7, %%mm6) // 11 /2
292
PAVGB(%%mm4, %%mm6) // 11 11 /4
293
PAVGB(%%mm3, %%mm6) // 11 2211 /8
294
PAVGB(%%mm5, %%mm2) // 11 /2
295
"movq (%0, %1, 4), %%mm4 \n\t" // 1
296
PAVGB(%%mm4, %%mm2) // 112 /4
297
PAVGB(%%mm2, %%mm6) // 112242211 /16
298
"movq %%mm6, (%0, %1, 4) \n\t" // X
299
// mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
300
PAVGB(%%mm7, %%mm1) // 11 2 /4
301
PAVGB(%%mm4, %%mm5) // 11 /2
302
PAVGB(%%mm5, %%mm0) // 11 11 /4
303
"movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1
304
PAVGB(%%mm6, %%mm1) // 11 4 2 /8
305
PAVGB(%%mm0, %%mm1) // 11224222 /16
306
"movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X
307
// mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
308
PAVGB((%%REGc), %%mm2) // 112 4 /8
309
"movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
310
PAVGB(%%mm0, %%mm6) // 1 1 /2
311
PAVGB(%%mm7, %%mm6) // 1 12 /4
312
PAVGB(%%mm2, %%mm6) // 1122424 /4
313
"movq %%mm6, (%%"REG_c") \n\t" // X
314
// mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
315
PAVGB(%%mm7, %%mm5) // 11 2 /4
316
PAVGB(%%mm7, %%mm5) // 11 6 /8
318
PAVGB(%%mm3, %%mm0) // 112 /4
319
PAVGB(%%mm0, %%mm5) // 112246 /16
320
"movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X
324
: "r" (src), "r" ((long)stride), "m" (c->pQPb)
327
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
328
const int l1= stride;
329
const int l2= stride + l1;
330
const int l3= stride + l2;
331
const int l4= stride + l3;
332
const int l5= stride + l4;
333
const int l6= stride + l5;
334
const int l7= stride + l6;
335
const int l8= stride + l7;
336
const int l9= stride + l8;
339
for(x=0; x<BLOCK_SIZE; x++)
341
const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
342
const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
345
sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
346
sums[1] = sums[0] - first + src[l4];
347
sums[2] = sums[1] - first + src[l5];
348
sums[3] = sums[2] - first + src[l6];
349
sums[4] = sums[3] - first + src[l7];
350
sums[5] = sums[4] - src[l1] + src[l8];
351
sums[6] = sums[5] - src[l2] + last;
352
sums[7] = sums[6] - src[l3] + last;
353
sums[8] = sums[7] - src[l4] + last;
354
sums[9] = sums[8] - src[l5] + last;
356
src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
357
src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
358
src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
359
src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
360
src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
361
src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
362
src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
363
src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
367
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
182
#if HAVE_MMX2 || HAVE_AMD3DNOW
184
__asm__ volatile( //"movv %0 %1 %2\n\t"
185
"movq %2, %%mm0 \n\t" // QP,..., QP
186
"pxor %%mm4, %%mm4 \n\t"
188
"movq (%0), %%mm6 \n\t"
189
"movq (%0, %1), %%mm5 \n\t"
190
"movq %%mm5, %%mm1 \n\t"
191
"movq %%mm6, %%mm2 \n\t"
192
"psubusb %%mm6, %%mm5 \n\t"
193
"psubusb %%mm1, %%mm2 \n\t"
194
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
195
"psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
196
"pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
198
"pand %%mm2, %%mm6 \n\t"
199
"pandn %%mm1, %%mm2 \n\t"
200
"por %%mm2, %%mm6 \n\t"// First Line to Filter
202
"movq (%0, %1, 8), %%mm5 \n\t"
203
"lea (%0, %1, 4), %%"REG_a" \n\t"
204
"lea (%0, %1, 8), %%"REG_c" \n\t"
205
"sub %1, %%"REG_c" \n\t"
206
"add %1, %0 \n\t" // %0 points to line 1 not 0
207
"movq (%0, %1, 8), %%mm7 \n\t"
208
"movq %%mm5, %%mm1 \n\t"
209
"movq %%mm7, %%mm2 \n\t"
210
"psubusb %%mm7, %%mm5 \n\t"
211
"psubusb %%mm1, %%mm2 \n\t"
212
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
213
"psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
214
"pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
216
"pand %%mm2, %%mm7 \n\t"
217
"pandn %%mm1, %%mm2 \n\t"
218
"por %%mm2, %%mm7 \n\t" // First Line to Filter
222
// %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
227
"movq (%0, %1), %%mm0 \n\t" // 1
228
"movq %%mm0, %%mm1 \n\t" // 1
229
PAVGB(%%mm6, %%mm0) //1 1 /2
230
PAVGB(%%mm6, %%mm0) //3 1 /4
232
"movq (%0, %1, 4), %%mm2 \n\t" // 1
233
"movq %%mm2, %%mm5 \n\t" // 1
234
PAVGB((%%REGa), %%mm2) // 11 /2
235
PAVGB((%0, %1, 2), %%mm2) // 211 /4
236
"movq %%mm2, %%mm3 \n\t" // 211 /4
237
"movq (%0), %%mm4 \n\t" // 1
238
PAVGB(%%mm4, %%mm3) // 4 211 /8
239
PAVGB(%%mm0, %%mm3) //642211 /16
240
"movq %%mm3, (%0) \n\t" // X
241
// mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
242
"movq %%mm1, %%mm0 \n\t" // 1
243
PAVGB(%%mm6, %%mm0) //1 1 /2
244
"movq %%mm4, %%mm3 \n\t" // 1
245
PAVGB((%0,%1,2), %%mm3) // 1 1 /2
246
PAVGB((%%REGa,%1,2), %%mm5) // 11 /2
247
PAVGB((%%REGa), %%mm5) // 211 /4
248
PAVGB(%%mm5, %%mm3) // 2 2211 /8
249
PAVGB(%%mm0, %%mm3) //4242211 /16
250
"movq %%mm3, (%0,%1) \n\t" // X
251
// mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
252
PAVGB(%%mm4, %%mm6) //11 /2
253
"movq (%%"REG_c"), %%mm0 \n\t" // 1
254
PAVGB((%%REGa, %1, 2), %%mm0) // 11/2
255
"movq %%mm0, %%mm3 \n\t" // 11/2
256
PAVGB(%%mm1, %%mm0) // 2 11/4
257
PAVGB(%%mm6, %%mm0) //222 11/8
258
PAVGB(%%mm2, %%mm0) //22242211/16
259
"movq (%0, %1, 2), %%mm2 \n\t" // 1
260
"movq %%mm0, (%0, %1, 2) \n\t" // X
261
// mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
262
"movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
263
PAVGB((%%REGc), %%mm0) // 11 /2
264
PAVGB(%%mm0, %%mm6) //11 11 /4
265
PAVGB(%%mm1, %%mm4) // 11 /2
266
PAVGB(%%mm2, %%mm1) // 11 /2
267
PAVGB(%%mm1, %%mm6) //1122 11 /8
268
PAVGB(%%mm5, %%mm6) //112242211 /16
269
"movq (%%"REG_a"), %%mm5 \n\t" // 1
270
"movq %%mm6, (%%"REG_a") \n\t" // X
271
// mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
272
"movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1
273
PAVGB(%%mm7, %%mm6) // 11 /2
274
PAVGB(%%mm4, %%mm6) // 11 11 /4
275
PAVGB(%%mm3, %%mm6) // 11 2211 /8
276
PAVGB(%%mm5, %%mm2) // 11 /2
277
"movq (%0, %1, 4), %%mm4 \n\t" // 1
278
PAVGB(%%mm4, %%mm2) // 112 /4
279
PAVGB(%%mm2, %%mm6) // 112242211 /16
280
"movq %%mm6, (%0, %1, 4) \n\t" // X
281
// mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
282
PAVGB(%%mm7, %%mm1) // 11 2 /4
283
PAVGB(%%mm4, %%mm5) // 11 /2
284
PAVGB(%%mm5, %%mm0) // 11 11 /4
285
"movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1
286
PAVGB(%%mm6, %%mm1) // 11 4 2 /8
287
PAVGB(%%mm0, %%mm1) // 11224222 /16
288
"movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X
289
// mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
290
PAVGB((%%REGc), %%mm2) // 112 4 /8
291
"movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1
292
PAVGB(%%mm0, %%mm6) // 1 1 /2
293
PAVGB(%%mm7, %%mm6) // 1 12 /4
294
PAVGB(%%mm2, %%mm6) // 1122424 /4
295
"movq %%mm6, (%%"REG_c") \n\t" // X
296
// mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
297
PAVGB(%%mm7, %%mm5) // 11 2 /4
298
PAVGB(%%mm7, %%mm5) // 11 6 /8
300
PAVGB(%%mm3, %%mm0) // 112 /4
301
PAVGB(%%mm0, %%mm5) // 112246 /16
302
"movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X
306
: "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
309
#else //HAVE_MMX2 || HAVE_AMD3DNOW
310
const int l1= stride;
311
const int l2= stride + l1;
312
const int l3= stride + l2;
313
const int l4= stride + l3;
314
const int l5= stride + l4;
315
const int l6= stride + l5;
316
const int l7= stride + l6;
317
const int l8= stride + l7;
318
const int l9= stride + l8;
321
for(x=0; x<BLOCK_SIZE; x++){
322
const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
323
const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
326
sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
327
sums[1] = sums[0] - first + src[l4];
328
sums[2] = sums[1] - first + src[l5];
329
sums[3] = sums[2] - first + src[l6];
330
sums[4] = sums[3] - first + src[l7];
331
sums[5] = sums[4] - src[l1] + src[l8];
332
sums[6] = sums[5] - src[l2] + last;
333
sums[7] = sums[6] - src[l3] + last;
334
sums[8] = sums[7] - src[l4] + last;
335
sums[9] = sums[8] - src[l5] + last;
337
src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
338
src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
339
src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
340
src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
341
src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
342
src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
343
src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
344
src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;
348
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
369
350
#endif //HAVE_ALTIVEC
384
365
static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
386
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
367
#if HAVE_MMX2 || HAVE_AMD3DNOW
388
369
// FIXME rounding
390
"pxor %%mm7, %%mm7 \n\t" // 0
391
"movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
392
"leal (%0, %1), %%"REG_a" \n\t"
393
"leal (%%"REG_a", %1, 4), %%"REG_c" \n\t"
371
"pxor %%mm7, %%mm7 \n\t" // 0
372
"movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
373
"leal (%0, %1), %%"REG_a" \n\t"
374
"leal (%%"REG_a", %1, 4), %%"REG_c" \n\t"
394
375
// 0 1 2 3 4 5 6 7 8 9
395
376
// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
396
"movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
397
"movq %%mm0, %%mm1 \n\t" // QP,..., QP
398
"paddusb "MANGLE(b02)", %%mm0 \n\t"
399
"psrlw $2, %%mm0 \n\t"
400
"pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4
401
"paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
402
"movq (%0, %1, 4), %%mm2 \n\t" // line 4
403
"movq (%%"REG_c"), %%mm3 \n\t" // line 5
404
"movq %%mm2, %%mm4 \n\t" // line 4
405
"pcmpeqb %%mm5, %%mm5 \n\t" // -1
406
"pxor %%mm2, %%mm5 \n\t" // -line 4 - 1
408
"paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2
409
"psubusb %%mm3, %%mm4 \n\t"
410
"psubusb %%mm2, %%mm3 \n\t"
411
"por %%mm3, %%mm4 \n\t" // |l4 - l5|
412
"psubusb %%mm0, %%mm4 \n\t"
413
"pcmpeqb %%mm7, %%mm4 \n\t"
414
"pand %%mm4, %%mm5 \n\t" // d/2
416
// "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80
417
"paddb %%mm5, %%mm2 \n\t"
418
// "psubb %%mm6, %%mm2 \n\t"
419
"movq %%mm2, (%0,%1, 4) \n\t"
421
"movq (%%"REG_c"), %%mm2 \n\t"
422
// "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80
423
"psubb %%mm5, %%mm2 \n\t"
424
// "psubb %%mm6, %%mm2 \n\t"
425
"movq %%mm2, (%%"REG_c") \n\t"
427
"paddb %%mm6, %%mm5 \n\t"
428
"psrlw $2, %%mm5 \n\t"
429
"pand "MANGLE(b3F)", %%mm5 \n\t"
430
"psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
432
"movq (%%"REG_a", %1, 2), %%mm2 \n\t"
433
"paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
434
"paddsb %%mm5, %%mm2 \n\t"
435
"psubb %%mm6, %%mm2 \n\t"
436
"movq %%mm2, (%%"REG_a", %1, 2) \n\t"
438
"movq (%%"REG_c", %1), %%mm2 \n\t"
439
"paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80
440
"psubsb %%mm5, %%mm2 \n\t"
441
"psubb %%mm6, %%mm2 \n\t"
442
"movq %%mm2, (%%"REG_c", %1) \n\t"
445
: "r" (src), "r" ((long)stride)
448
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
449
const int l1= stride;
450
const int l2= stride + l1;
451
const int l3= stride + l2;
452
const int l4= stride + l3;
453
const int l5= stride + l4;
454
const int l6= stride + l5;
455
// const int l7= stride + l6;
456
// const int l8= stride + l7;
457
// const int l9= stride + l8;
459
const int QP15= QP + (QP>>2);
461
for(x=0; x<BLOCK_SIZE; x++)
463
const int v = (src[x+l5] - src[x+l4]);
377
"movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
378
"movq %%mm0, %%mm1 \n\t" // QP,..., QP
379
"paddusb "MANGLE(b02)", %%mm0 \n\t"
380
"psrlw $2, %%mm0 \n\t"
381
"pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4
382
"paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
383
"movq (%0, %1, 4), %%mm2 \n\t" // line 4
384
"movq (%%"REG_c"), %%mm3 \n\t" // line 5
385
"movq %%mm2, %%mm4 \n\t" // line 4
386
"pcmpeqb %%mm5, %%mm5 \n\t" // -1
387
"pxor %%mm2, %%mm5 \n\t" // -line 4 - 1
389
"paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2
390
"psubusb %%mm3, %%mm4 \n\t"
391
"psubusb %%mm2, %%mm3 \n\t"
392
"por %%mm3, %%mm4 \n\t" // |l4 - l5|
393
"psubusb %%mm0, %%mm4 \n\t"
394
"pcmpeqb %%mm7, %%mm4 \n\t"
395
"pand %%mm4, %%mm5 \n\t" // d/2
397
// "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80
398
"paddb %%mm5, %%mm2 \n\t"
399
// "psubb %%mm6, %%mm2 \n\t"
400
"movq %%mm2, (%0,%1, 4) \n\t"
402
"movq (%%"REG_c"), %%mm2 \n\t"
403
// "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80
404
"psubb %%mm5, %%mm2 \n\t"
405
// "psubb %%mm6, %%mm2 \n\t"
406
"movq %%mm2, (%%"REG_c") \n\t"
408
"paddb %%mm6, %%mm5 \n\t"
409
"psrlw $2, %%mm5 \n\t"
410
"pand "MANGLE(b3F)", %%mm5 \n\t"
411
"psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
413
"movq (%%"REG_a", %1, 2), %%mm2 \n\t"
414
"paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
415
"paddsb %%mm5, %%mm2 \n\t"
416
"psubb %%mm6, %%mm2 \n\t"
417
"movq %%mm2, (%%"REG_a", %1, 2) \n\t"
419
"movq (%%"REG_c", %1), %%mm2 \n\t"
420
"paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80
421
"psubsb %%mm5, %%mm2 \n\t"
422
"psubb %%mm6, %%mm2 \n\t"
423
"movq %%mm2, (%%"REG_c", %1) \n\t"
426
: "r" (src), "r" ((x86_reg)stride)
429
#else //HAVE_MMX2 || HAVE_AMD3DNOW
430
const int l1= stride;
431
const int l2= stride + l1;
432
const int l3= stride + l2;
433
const int l4= stride + l3;
434
const int l5= stride + l4;
435
const int l6= stride + l5;
436
// const int l7= stride + l6;
437
// const int l8= stride + l7;
438
// const int l9= stride + l8;
440
const int QP15= QP + (QP>>2);
442
for(x=0; x<BLOCK_SIZE; x++){
443
const int v = (src[x+l5] - src[x+l4]);
474
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
452
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
479
457
* Experimental Filter 1
480
458
* will not damage linear gradients
481
* Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
459
* Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
482
460
* can only smooth blocks at the expected locations (it cannot smooth them if they did move)
483
461
* MMX2 version does correct clipping C version does not
485
463
static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
487
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
465
#if HAVE_MMX2 || HAVE_AMD3DNOW
491
"pxor %%mm7, %%mm7 \n\t" // 0
492
"lea (%0, %1), %%"REG_a" \n\t"
493
"lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
469
"pxor %%mm7, %%mm7 \n\t" // 0
470
"lea (%0, %1), %%"REG_a" \n\t"
471
"lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
494
472
// 0 1 2 3 4 5 6 7 8 9
495
473
// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
496
"movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
497
"movq (%0, %1, 4), %%mm1 \n\t" // line 4
498
"movq %%mm1, %%mm2 \n\t" // line 4
499
"psubusb %%mm0, %%mm1 \n\t"
500
"psubusb %%mm2, %%mm0 \n\t"
501
"por %%mm1, %%mm0 \n\t" // |l2 - l3|
502
"movq (%%"REG_c"), %%mm3 \n\t" // line 5
503
"movq (%%"REG_c", %1), %%mm4 \n\t" // line 6
504
"movq %%mm3, %%mm5 \n\t" // line 5
505
"psubusb %%mm4, %%mm3 \n\t"
506
"psubusb %%mm5, %%mm4 \n\t"
507
"por %%mm4, %%mm3 \n\t" // |l5 - l6|
508
PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
509
"movq %%mm2, %%mm1 \n\t" // line 4
510
"psubusb %%mm5, %%mm2 \n\t"
511
"movq %%mm2, %%mm4 \n\t"
512
"pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
513
"psubusb %%mm1, %%mm5 \n\t"
514
"por %%mm5, %%mm4 \n\t" // |l4 - l5|
515
"psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
516
"movq %%mm4, %%mm3 \n\t" // d
517
"movq %2, %%mm0 \n\t"
518
"paddusb %%mm0, %%mm0 \n\t"
519
"psubusb %%mm0, %%mm4 \n\t"
520
"pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
521
"psubusb "MANGLE(b01)", %%mm3 \n\t"
522
"pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
524
PAVGB(%%mm7, %%mm3) // d/2
525
"movq %%mm3, %%mm1 \n\t" // d/2
526
PAVGB(%%mm7, %%mm3) // d/4
527
PAVGB(%%mm1, %%mm3) // 3*d/8
529
"movq (%0, %1, 4), %%mm0 \n\t" // line 4
530
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
531
"psubusb %%mm3, %%mm0 \n\t"
532
"pxor %%mm2, %%mm0 \n\t"
533
"movq %%mm0, (%0, %1, 4) \n\t" // line 4
535
"movq (%%"REG_c"), %%mm0 \n\t" // line 5
536
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
537
"paddusb %%mm3, %%mm0 \n\t"
538
"pxor %%mm2, %%mm0 \n\t"
539
"movq %%mm0, (%%"REG_c") \n\t" // line 5
541
PAVGB(%%mm7, %%mm1) // d/4
543
"movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
544
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
545
"psubusb %%mm1, %%mm0 \n\t"
546
"pxor %%mm2, %%mm0 \n\t"
547
"movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3
549
"movq (%%"REG_c", %1), %%mm0 \n\t" // line 6
550
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
551
"paddusb %%mm1, %%mm0 \n\t"
552
"pxor %%mm2, %%mm0 \n\t"
553
"movq %%mm0, (%%"REG_c", %1) \n\t" // line 6
555
PAVGB(%%mm7, %%mm1) // d/8
557
"movq (%%"REG_a", %1), %%mm0 \n\t" // line 2
558
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
559
"psubusb %%mm1, %%mm0 \n\t"
560
"pxor %%mm2, %%mm0 \n\t"
561
"movq %%mm0, (%%"REG_a", %1) \n\t" // line 2
563
"movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7
564
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
565
"paddusb %%mm1, %%mm0 \n\t"
566
"pxor %%mm2, %%mm0 \n\t"
567
"movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7
570
: "r" (src), "r" ((long)stride), "m" (co->pQPb)
573
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
575
const int l1= stride;
576
const int l2= stride + l1;
577
const int l3= stride + l2;
578
const int l4= stride + l3;
579
const int l5= stride + l4;
580
const int l6= stride + l5;
581
const int l7= stride + l6;
582
// const int l8= stride + l7;
583
// const int l9= stride + l8;
587
for(x=0; x<BLOCK_SIZE; x++)
589
int a= src[l3] - src[l4];
590
int b= src[l4] - src[l5];
591
int c= src[l5] - src[l6];
593
int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
598
int v = d * FFSIGN(-b);
474
"movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
475
"movq (%0, %1, 4), %%mm1 \n\t" // line 4
476
"movq %%mm1, %%mm2 \n\t" // line 4
477
"psubusb %%mm0, %%mm1 \n\t"
478
"psubusb %%mm2, %%mm0 \n\t"
479
"por %%mm1, %%mm0 \n\t" // |l2 - l3|
480
"movq (%%"REG_c"), %%mm3 \n\t" // line 5
481
"movq (%%"REG_c", %1), %%mm4 \n\t" // line 6
482
"movq %%mm3, %%mm5 \n\t" // line 5
483
"psubusb %%mm4, %%mm3 \n\t"
484
"psubusb %%mm5, %%mm4 \n\t"
485
"por %%mm4, %%mm3 \n\t" // |l5 - l6|
486
PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
487
"movq %%mm2, %%mm1 \n\t" // line 4
488
"psubusb %%mm5, %%mm2 \n\t"
489
"movq %%mm2, %%mm4 \n\t"
490
"pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
491
"psubusb %%mm1, %%mm5 \n\t"
492
"por %%mm5, %%mm4 \n\t" // |l4 - l5|
493
"psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
494
"movq %%mm4, %%mm3 \n\t" // d
495
"movq %2, %%mm0 \n\t"
496
"paddusb %%mm0, %%mm0 \n\t"
497
"psubusb %%mm0, %%mm4 \n\t"
498
"pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
499
"psubusb "MANGLE(b01)", %%mm3 \n\t"
500
"pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
502
PAVGB(%%mm7, %%mm3) // d/2
503
"movq %%mm3, %%mm1 \n\t" // d/2
504
PAVGB(%%mm7, %%mm3) // d/4
505
PAVGB(%%mm1, %%mm3) // 3*d/8
507
"movq (%0, %1, 4), %%mm0 \n\t" // line 4
508
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
509
"psubusb %%mm3, %%mm0 \n\t"
510
"pxor %%mm2, %%mm0 \n\t"
511
"movq %%mm0, (%0, %1, 4) \n\t" // line 4
513
"movq (%%"REG_c"), %%mm0 \n\t" // line 5
514
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
515
"paddusb %%mm3, %%mm0 \n\t"
516
"pxor %%mm2, %%mm0 \n\t"
517
"movq %%mm0, (%%"REG_c") \n\t" // line 5
519
PAVGB(%%mm7, %%mm1) // d/4
521
"movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3
522
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
523
"psubusb %%mm1, %%mm0 \n\t"
524
"pxor %%mm2, %%mm0 \n\t"
525
"movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3
527
"movq (%%"REG_c", %1), %%mm0 \n\t" // line 6
528
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
529
"paddusb %%mm1, %%mm0 \n\t"
530
"pxor %%mm2, %%mm0 \n\t"
531
"movq %%mm0, (%%"REG_c", %1) \n\t" // line 6
533
PAVGB(%%mm7, %%mm1) // d/8
535
"movq (%%"REG_a", %1), %%mm0 \n\t" // line 2
536
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
537
"psubusb %%mm1, %%mm0 \n\t"
538
"pxor %%mm2, %%mm0 \n\t"
539
"movq %%mm0, (%%"REG_a", %1) \n\t" // line 2
541
"movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7
542
"pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
543
"paddusb %%mm1, %%mm0 \n\t"
544
"pxor %%mm2, %%mm0 \n\t"
545
"movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7
548
: "r" (src), "r" ((x86_reg)stride), "m" (co->pQPb)
551
#else //HAVE_MMX2 || HAVE_AMD3DNOW
553
const int l1= stride;
554
const int l2= stride + l1;
555
const int l3= stride + l2;
556
const int l4= stride + l3;
557
const int l5= stride + l4;
558
const int l6= stride + l5;
559
const int l7= stride + l6;
560
// const int l8= stride + l7;
561
// const int l9= stride + l8;
565
for(x=0; x<BLOCK_SIZE; x++){
566
int a= src[l3] - src[l4];
567
int b= src[l4] - src[l5];
568
int c= src[l5] - src[l6];
570
int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
574
int v = d * FFSIGN(-b);
610
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
585
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
614
589
static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
616
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
591
#if HAVE_MMX2 || HAVE_AMD3DNOW
619
const int l1= stride;
620
const int l2= stride + l1;
621
const int l3= stride + l2;
622
const int l4= (int)tmp - (int)src - stride*3;
623
const int l5= (int)tmp - (int)src - stride*3 + 8;
624
const int l6= stride*3 + l3;
625
const int l7= stride + l6;
626
const int l8= stride + l7;
594
const int l1= stride;
595
const int l2= stride + l1;
596
const int l3= stride + l2;
597
const int l4= (int)tmp - (int)src - stride*3;
598
const int l5= (int)tmp - (int)src - stride*3 + 8;
599
const int l6= stride*3 + l3;
600
const int l7= stride + l6;
601
const int l8= stride + l7;
628
memcpy(tmp, src+stride*7, 8);
629
memcpy(tmp+8, src+stride*8, 8);
603
memcpy(tmp, src+stride*7, 8);
604
memcpy(tmp+8, src+stride*8, 8);
634
#if 0 //sligtly more accurate and slightly slower
635
"pxor %%mm7, %%mm7 \n\t" // 0
636
"lea (%0, %1), %%"REG_a" \n\t"
637
"lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
609
#if 0 //slightly more accurate and slightly slower
610
"pxor %%mm7, %%mm7 \n\t" // 0
611
"lea (%0, %1), %%"REG_a" \n\t"
612
"lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
638
613
// 0 1 2 3 4 5 6 7
639
614
// %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
640
615
// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
643
"movq (%0, %1, 2), %%mm0 \n\t" // l2
644
"movq (%0), %%mm1 \n\t" // l0
645
"movq %%mm0, %%mm2 \n\t" // l2
646
PAVGB(%%mm7, %%mm0) // ~l2/2
647
PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
648
PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
650
"movq (%%"REG_a"), %%mm1 \n\t" // l1
651
"movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3
652
"movq %%mm1, %%mm4 \n\t" // l1
653
PAVGB(%%mm7, %%mm1) // ~l1/2
654
PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
655
PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
657
"movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
658
"psubusb %%mm1, %%mm0 \n\t"
659
"psubusb %%mm4, %%mm1 \n\t"
660
"por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
618
"movq (%0, %1, 2), %%mm0 \n\t" // l2
619
"movq (%0), %%mm1 \n\t" // l0
620
"movq %%mm0, %%mm2 \n\t" // l2
621
PAVGB(%%mm7, %%mm0) // ~l2/2
622
PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4
623
PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8
625
"movq (%%"REG_a"), %%mm1 \n\t" // l1
626
"movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3
627
"movq %%mm1, %%mm4 \n\t" // l1
628
PAVGB(%%mm7, %%mm1) // ~l1/2
629
PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4
630
PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8
632
"movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8
633
"psubusb %%mm1, %%mm0 \n\t"
634
"psubusb %%mm4, %%mm1 \n\t"
635
"por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
661
636
// mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
663
"movq (%0, %1, 4), %%mm0 \n\t" // l4
664
"movq %%mm0, %%mm4 \n\t" // l4
665
PAVGB(%%mm7, %%mm0) // ~l4/2
666
PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
667
PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
669
"movq (%%"REG_c"), %%mm2 \n\t" // l5
670
"movq %%mm3, %%mm5 \n\t" // l3
671
PAVGB(%%mm7, %%mm3) // ~l3/2
672
PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
673
PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
675
"movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
676
"psubusb %%mm3, %%mm0 \n\t"
677
"psubusb %%mm6, %%mm3 \n\t"
678
"por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
679
"pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
638
"movq (%0, %1, 4), %%mm0 \n\t" // l4
639
"movq %%mm0, %%mm4 \n\t" // l4
640
PAVGB(%%mm7, %%mm0) // ~l4/2
641
PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4
642
PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8
644
"movq (%%"REG_c"), %%mm2 \n\t" // l5
645
"movq %%mm3, %%mm5 \n\t" // l3
646
PAVGB(%%mm7, %%mm3) // ~l3/2
647
PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4
648
PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8
650
"movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8
651
"psubusb %%mm3, %%mm0 \n\t"
652
"psubusb %%mm6, %%mm3 \n\t"
653
"por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
654
"pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
680
655
// mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
682
"movq (%%"REG_c", %1), %%mm6 \n\t" // l6
683
"movq %%mm6, %%mm5 \n\t" // l6
684
PAVGB(%%mm7, %%mm6) // ~l6/2
685
PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
686
PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
688
"movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7
689
"movq %%mm2, %%mm4 \n\t" // l5
690
PAVGB(%%mm7, %%mm2) // ~l5/2
691
PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
692
PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
694
"movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
695
"psubusb %%mm2, %%mm6 \n\t"
696
"psubusb %%mm4, %%mm2 \n\t"
697
"por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
657
"movq (%%"REG_c", %1), %%mm6 \n\t" // l6
658
"movq %%mm6, %%mm5 \n\t" // l6
659
PAVGB(%%mm7, %%mm6) // ~l6/2
660
PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4
661
PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8
663
"movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7
664
"movq %%mm2, %%mm4 \n\t" // l5
665
PAVGB(%%mm7, %%mm2) // ~l5/2
666
PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4
667
PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8
669
"movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8
670
"psubusb %%mm2, %%mm6 \n\t"
671
"psubusb %%mm4, %%mm2 \n\t"
672
"por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
698
673
// mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
701
PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
702
"movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
703
"paddusb "MANGLE(b01)", %%mm4 \n\t"
704
"pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
705
"psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
706
"pand %%mm4, %%mm3 \n\t"
708
"movq %%mm3, %%mm1 \n\t"
709
// "psubusb "MANGLE(b01)", %%mm3 \n\t"
712
"paddusb %%mm1, %%mm3 \n\t"
713
// "paddusb "MANGLE(b01)", %%mm3 \n\t"
715
"movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3
716
"movq (%0, %1, 4), %%mm5 \n\t" //l4
717
"movq (%0, %1, 4), %%mm4 \n\t" //l4
718
"psubusb %%mm6, %%mm5 \n\t"
719
"psubusb %%mm4, %%mm6 \n\t"
720
"por %%mm6, %%mm5 \n\t" // |l3-l4|
721
"pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
722
"pxor %%mm6, %%mm0 \n\t"
723
"pand %%mm0, %%mm3 \n\t"
724
PMINUB(%%mm5, %%mm3, %%mm0)
726
"psubusb "MANGLE(b01)", %%mm3 \n\t"
729
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
730
"movq (%0, %1, 4), %%mm2 \n\t"
731
"pxor %%mm6, %%mm0 \n\t"
732
"pxor %%mm6, %%mm2 \n\t"
733
"psubb %%mm3, %%mm0 \n\t"
734
"paddb %%mm3, %%mm2 \n\t"
735
"pxor %%mm6, %%mm0 \n\t"
736
"pxor %%mm6, %%mm2 \n\t"
737
"movq %%mm0, (%%"REG_a", %1, 2) \n\t"
738
"movq %%mm2, (%0, %1, 4) \n\t"
676
PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8
677
"movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
678
"paddusb "MANGLE(b01)", %%mm4 \n\t"
679
"pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP
680
"psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
681
"pand %%mm4, %%mm3 \n\t"
683
"movq %%mm3, %%mm1 \n\t"
684
// "psubusb "MANGLE(b01)", %%mm3 \n\t"
687
"paddusb %%mm1, %%mm3 \n\t"
688
// "paddusb "MANGLE(b01)", %%mm3 \n\t"
690
"movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3
691
"movq (%0, %1, 4), %%mm5 \n\t" //l4
692
"movq (%0, %1, 4), %%mm4 \n\t" //l4
693
"psubusb %%mm6, %%mm5 \n\t"
694
"psubusb %%mm4, %%mm6 \n\t"
695
"por %%mm6, %%mm5 \n\t" // |l3-l4|
696
"pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4)
697
"pxor %%mm6, %%mm0 \n\t"
698
"pand %%mm0, %%mm3 \n\t"
699
PMINUB(%%mm5, %%mm3, %%mm0)
701
"psubusb "MANGLE(b01)", %%mm3 \n\t"
704
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
705
"movq (%0, %1, 4), %%mm2 \n\t"
706
"pxor %%mm6, %%mm0 \n\t"
707
"pxor %%mm6, %%mm2 \n\t"
708
"psubb %%mm3, %%mm0 \n\t"
709
"paddb %%mm3, %%mm2 \n\t"
710
"pxor %%mm6, %%mm0 \n\t"
711
"pxor %%mm6, %%mm2 \n\t"
712
"movq %%mm0, (%%"REG_a", %1, 2) \n\t"
713
"movq %%mm2, (%0, %1, 4) \n\t"
741
"lea (%0, %1), %%"REG_a" \n\t"
742
"pcmpeqb %%mm6, %%mm6 \n\t" // -1
716
"lea (%0, %1), %%"REG_a" \n\t"
717
"pcmpeqb %%mm6, %%mm6 \n\t" // -1
743
718
// 0 1 2 3 4 5 6 7
744
719
// %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1
745
720
// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
748
"movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3
749
"movq (%0, %1, 4), %%mm0 \n\t" // l4
750
"pxor %%mm6, %%mm1 \n\t" // -l3-1
751
PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
723
"movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3
724
"movq (%0, %1, 4), %%mm0 \n\t" // l4
725
"pxor %%mm6, %%mm1 \n\t" // -l3-1
726
PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2
752
727
// mm1=-l3-1, mm0=128-q
754
"movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5
755
"movq (%%"REG_a", %1), %%mm3 \n\t" // l2
756
"pxor %%mm6, %%mm2 \n\t" // -l5-1
757
"movq %%mm2, %%mm5 \n\t" // -l5-1
758
"movq "MANGLE(b80)", %%mm4 \n\t" // 128
759
"lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
760
PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
761
PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
762
PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
763
PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
729
"movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5
730
"movq (%%"REG_a", %1), %%mm3 \n\t" // l2
731
"pxor %%mm6, %%mm2 \n\t" // -l5-1
732
"movq %%mm2, %%mm5 \n\t" // -l5-1
733
"movq "MANGLE(b80)", %%mm4 \n\t" // 128
734
"lea (%%"REG_a", %1, 4), %%"REG_c" \n\t"
735
PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2
736
PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128
737
PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128
738
PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128
764
739
// mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
766
"movq (%%"REG_a"), %%mm2 \n\t" // l1
767
"pxor %%mm6, %%mm2 \n\t" // -l1-1
768
PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
769
PAVGB((%0), %%mm1) // (l0-l3+256)/2
770
"movq "MANGLE(b80)", %%mm3 \n\t" // 128
771
PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
772
PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
773
PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
741
"movq (%%"REG_a"), %%mm2 \n\t" // l1
742
"pxor %%mm6, %%mm2 \n\t" // -l1-1
743
PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2
744
PAVGB((%0), %%mm1) // (l0-l3+256)/2
745
"movq "MANGLE(b80)", %%mm3 \n\t" // 128
746
PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128
747
PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128
748
PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128
774
749
// mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
776
PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2
777
"movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7
778
"pxor %%mm6, %%mm1 \n\t" // -l7-1
779
PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
780
"movq "MANGLE(b80)", %%mm2 \n\t" // 128
781
PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
782
PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
783
PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
751
PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2
752
"movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7
753
"pxor %%mm6, %%mm1 \n\t" // -l7-1
754
PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
755
"movq "MANGLE(b80)", %%mm2 \n\t" // 128
756
PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128
757
PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128
758
PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128
784
759
// mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
786
"movq "MANGLE(b00)", %%mm1 \n\t" // 0
787
"movq "MANGLE(b00)", %%mm5 \n\t" // 0
788
"psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
789
"psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
790
PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
791
PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
792
PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
761
"movq "MANGLE(b00)", %%mm1 \n\t" // 0
762
"movq "MANGLE(b00)", %%mm5 \n\t" // 0
763
"psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16
764
"psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16
765
PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16|
766
PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16|
767
PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16
794
769
// mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
796
"movq "MANGLE(b00)", %%mm7 \n\t" // 0
797
"movq %2, %%mm2 \n\t" // QP
798
PAVGB(%%mm6, %%mm2) // 128 + QP/2
799
"psubb %%mm6, %%mm2 \n\t"
771
"movq "MANGLE(b00)", %%mm7 \n\t" // 0
772
"movq %2, %%mm2 \n\t" // QP
773
PAVGB(%%mm6, %%mm2) // 128 + QP/2
774
"psubb %%mm6, %%mm2 \n\t"
801
"movq %%mm4, %%mm1 \n\t"
802
"pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
803
"pxor %%mm1, %%mm4 \n\t"
804
"psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
805
"pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
806
"psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
776
"movq %%mm4, %%mm1 \n\t"
777
"pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy)
778
"pxor %%mm1, %%mm4 \n\t"
779
"psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16
780
"pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2
781
"psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
807
782
// mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
809
"movq %%mm4, %%mm3 \n\t" // d
810
"psubusb "MANGLE(b01)", %%mm4 \n\t"
811
PAVGB(%%mm7, %%mm4) // d/32
812
PAVGB(%%mm7, %%mm4) // (d + 32)/64
813
"paddb %%mm3, %%mm4 \n\t" // 5d/64
814
"pand %%mm2, %%mm4 \n\t"
816
"movq "MANGLE(b80)", %%mm5 \n\t" // 128
817
"psubb %%mm0, %%mm5 \n\t" // q
818
"paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
819
"pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
820
"pxor %%mm7, %%mm5 \n\t"
822
PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
823
"pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
825
"pand %%mm7, %%mm4 \n\t"
826
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
827
"movq (%0, %1, 4), %%mm2 \n\t"
828
"pxor %%mm1, %%mm0 \n\t"
829
"pxor %%mm1, %%mm2 \n\t"
830
"paddb %%mm4, %%mm0 \n\t"
831
"psubb %%mm4, %%mm2 \n\t"
832
"pxor %%mm1, %%mm0 \n\t"
833
"pxor %%mm1, %%mm2 \n\t"
834
"movq %%mm0, (%%"REG_a", %1, 2) \n\t"
835
"movq %%mm2, (%0, %1, 4) \n\t"
838
: "r" (src), "r" ((long)stride), "m" (c->pQPb)
784
"movq %%mm4, %%mm3 \n\t" // d
785
"psubusb "MANGLE(b01)", %%mm4 \n\t"
786
PAVGB(%%mm7, %%mm4) // d/32
787
PAVGB(%%mm7, %%mm4) // (d + 32)/64
788
"paddb %%mm3, %%mm4 \n\t" // 5d/64
789
"pand %%mm2, %%mm4 \n\t"
791
"movq "MANGLE(b80)", %%mm5 \n\t" // 128
792
"psubb %%mm0, %%mm5 \n\t" // q
793
"paddsb %%mm6, %%mm5 \n\t" // fix bad rounding
794
"pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q)
795
"pxor %%mm7, %%mm5 \n\t"
797
PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64)
798
"pxor %%mm1, %%mm7 \n\t" // SIGN(d*q)
800
"pand %%mm7, %%mm4 \n\t"
801
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
802
"movq (%0, %1, 4), %%mm2 \n\t"
803
"pxor %%mm1, %%mm0 \n\t"
804
"pxor %%mm1, %%mm2 \n\t"
805
"paddb %%mm4, %%mm0 \n\t"
806
"psubb %%mm4, %%mm2 \n\t"
807
"pxor %%mm1, %%mm0 \n\t"
808
"pxor %%mm1, %%mm2 \n\t"
809
"movq %%mm0, (%%"REG_a", %1, 2) \n\t"
810
"movq %%mm2, (%0, %1, 4) \n\t"
813
: "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb)
846
for(x=0; x<BLOCK_SIZE; x++)
848
const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
849
if(FFABS(middleEnergy)< 8*QP)
851
const int q=(src[l4] - src[l5])/2;
852
const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
853
const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
855
int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
859
d*= FFSIGN(-middleEnergy);
883
int d= src[x+y*stride] - tmp[x+(y-4)*8];
894
src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
900
av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias);
821
for(x=0; x<BLOCK_SIZE; x++){
822
const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
823
if(FFABS(middleEnergy)< 8*QP){
824
const int q=(src[l4] - src[l5])/2;
825
const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
826
const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
828
int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
832
d*= FFSIGN(-middleEnergy);
851
int d= src[x+y*stride] - tmp[x+(y-4)*8];
861
src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
865
if(num%1000000 == 0){
866
av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias);
906
#elif defined (HAVE_MMX)
909
"pxor %%mm7, %%mm7 \n\t"
910
"lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
911
"and "ALIGN_MASK", %%"REG_c" \n\t" // align
875
"pxor %%mm7, %%mm7 \n\t"
876
"lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
877
"and "ALIGN_MASK", %%"REG_c" \n\t" // align
912
878
// 0 1 2 3 4 5 6 7
913
879
// %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
914
880
// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
916
"movq (%0), %%mm0 \n\t"
917
"movq %%mm0, %%mm1 \n\t"
918
"punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
919
"punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
921
"movq (%0, %1), %%mm2 \n\t"
922
"lea (%0, %1, 2), %%"REG_a" \n\t"
923
"movq %%mm2, %%mm3 \n\t"
924
"punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
925
"punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
927
"movq (%%"REG_a"), %%mm4 \n\t"
928
"movq %%mm4, %%mm5 \n\t"
929
"punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
930
"punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
932
"paddw %%mm0, %%mm0 \n\t" // 2L0
933
"paddw %%mm1, %%mm1 \n\t" // 2H0
934
"psubw %%mm4, %%mm2 \n\t" // L1 - L2
935
"psubw %%mm5, %%mm3 \n\t" // H1 - H2
936
"psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
937
"psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
939
"psllw $2, %%mm2 \n\t" // 4L1 - 4L2
940
"psllw $2, %%mm3 \n\t" // 4H1 - 4H2
941
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
942
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
944
"movq (%%"REG_a", %1), %%mm2 \n\t"
945
"movq %%mm2, %%mm3 \n\t"
946
"punpcklbw %%mm7, %%mm2 \n\t" // L3
947
"punpckhbw %%mm7, %%mm3 \n\t" // H3
949
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
950
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
951
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
952
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
953
"movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
954
"movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
956
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
957
"movq %%mm0, %%mm1 \n\t"
958
"punpcklbw %%mm7, %%mm0 \n\t" // L4
959
"punpckhbw %%mm7, %%mm1 \n\t" // H4
961
"psubw %%mm0, %%mm2 \n\t" // L3 - L4
962
"psubw %%mm1, %%mm3 \n\t" // H3 - H4
963
"movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
964
"movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
965
"paddw %%mm4, %%mm4 \n\t" // 2L2
966
"paddw %%mm5, %%mm5 \n\t" // 2H2
967
"psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
968
"psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
970
"lea (%%"REG_a", %1), %0 \n\t"
971
"psllw $2, %%mm2 \n\t" // 4L3 - 4L4
972
"psllw $2, %%mm3 \n\t" // 4H3 - 4H4
973
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
974
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
882
"movq (%0), %%mm0 \n\t"
883
"movq %%mm0, %%mm1 \n\t"
884
"punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
885
"punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
887
"movq (%0, %1), %%mm2 \n\t"
888
"lea (%0, %1, 2), %%"REG_a" \n\t"
889
"movq %%mm2, %%mm3 \n\t"
890
"punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
891
"punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
893
"movq (%%"REG_a"), %%mm4 \n\t"
894
"movq %%mm4, %%mm5 \n\t"
895
"punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
896
"punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
898
"paddw %%mm0, %%mm0 \n\t" // 2L0
899
"paddw %%mm1, %%mm1 \n\t" // 2H0
900
"psubw %%mm4, %%mm2 \n\t" // L1 - L2
901
"psubw %%mm5, %%mm3 \n\t" // H1 - H2
902
"psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
903
"psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
905
"psllw $2, %%mm2 \n\t" // 4L1 - 4L2
906
"psllw $2, %%mm3 \n\t" // 4H1 - 4H2
907
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
908
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
910
"movq (%%"REG_a", %1), %%mm2 \n\t"
911
"movq %%mm2, %%mm3 \n\t"
912
"punpcklbw %%mm7, %%mm2 \n\t" // L3
913
"punpckhbw %%mm7, %%mm3 \n\t" // H3
915
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
916
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
917
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
918
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
919
"movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
920
"movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
922
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
923
"movq %%mm0, %%mm1 \n\t"
924
"punpcklbw %%mm7, %%mm0 \n\t" // L4
925
"punpckhbw %%mm7, %%mm1 \n\t" // H4
927
"psubw %%mm0, %%mm2 \n\t" // L3 - L4
928
"psubw %%mm1, %%mm3 \n\t" // H3 - H4
929
"movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
930
"movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
931
"paddw %%mm4, %%mm4 \n\t" // 2L2
932
"paddw %%mm5, %%mm5 \n\t" // 2H2
933
"psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
934
"psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
936
"lea (%%"REG_a", %1), %0 \n\t"
937
"psllw $2, %%mm2 \n\t" // 4L3 - 4L4
938
"psllw $2, %%mm3 \n\t" // 4H3 - 4H4
939
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
940
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
975
941
//50 opcodes so far
976
"movq (%0, %1, 2), %%mm2 \n\t"
977
"movq %%mm2, %%mm3 \n\t"
978
"punpcklbw %%mm7, %%mm2 \n\t" // L5
979
"punpckhbw %%mm7, %%mm3 \n\t" // H5
980
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
981
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
982
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
983
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
985
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
986
"punpcklbw %%mm7, %%mm6 \n\t" // L6
987
"psubw %%mm6, %%mm2 \n\t" // L5 - L6
988
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
989
"punpckhbw %%mm7, %%mm6 \n\t" // H6
990
"psubw %%mm6, %%mm3 \n\t" // H5 - H6
992
"paddw %%mm0, %%mm0 \n\t" // 2L4
993
"paddw %%mm1, %%mm1 \n\t" // 2H4
994
"psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
995
"psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
997
"psllw $2, %%mm2 \n\t" // 4L5 - 4L6
998
"psllw $2, %%mm3 \n\t" // 4H5 - 4H6
999
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
1000
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
1002
"movq (%0, %1, 4), %%mm2 \n\t"
1003
"movq %%mm2, %%mm3 \n\t"
1004
"punpcklbw %%mm7, %%mm2 \n\t" // L7
1005
"punpckhbw %%mm7, %%mm3 \n\t" // H7
1007
"paddw %%mm2, %%mm2 \n\t" // 2L7
1008
"paddw %%mm3, %%mm3 \n\t" // 2H7
1009
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
1010
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
1012
"movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
1013
"movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
1016
"movq %%mm7, %%mm6 \n\t" // 0
1017
"psubw %%mm0, %%mm6 \n\t"
1018
"pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1019
"movq %%mm7, %%mm6 \n\t" // 0
1020
"psubw %%mm1, %%mm6 \n\t"
1021
"pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1022
"movq %%mm7, %%mm6 \n\t" // 0
1023
"psubw %%mm2, %%mm6 \n\t"
1024
"pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1025
"movq %%mm7, %%mm6 \n\t" // 0
1026
"psubw %%mm3, %%mm6 \n\t"
1027
"pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1029
"movq %%mm7, %%mm6 \n\t" // 0
1030
"pcmpgtw %%mm0, %%mm6 \n\t"
1031
"pxor %%mm6, %%mm0 \n\t"
1032
"psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1033
"movq %%mm7, %%mm6 \n\t" // 0
1034
"pcmpgtw %%mm1, %%mm6 \n\t"
1035
"pxor %%mm6, %%mm1 \n\t"
1036
"psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1037
"movq %%mm7, %%mm6 \n\t" // 0
1038
"pcmpgtw %%mm2, %%mm6 \n\t"
1039
"pxor %%mm6, %%mm2 \n\t"
1040
"psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1041
"movq %%mm7, %%mm6 \n\t" // 0
1042
"pcmpgtw %%mm3, %%mm6 \n\t"
1043
"pxor %%mm6, %%mm3 \n\t"
1044
"psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1048
"pminsw %%mm2, %%mm0 \n\t"
1049
"pminsw %%mm3, %%mm1 \n\t"
1051
"movq %%mm0, %%mm6 \n\t"
1052
"psubusw %%mm2, %%mm6 \n\t"
1053
"psubw %%mm6, %%mm0 \n\t"
1054
"movq %%mm1, %%mm6 \n\t"
1055
"psubusw %%mm3, %%mm6 \n\t"
1056
"psubw %%mm6, %%mm1 \n\t"
1059
"movd %2, %%mm2 \n\t" // QP
1060
"punpcklbw %%mm7, %%mm2 \n\t"
1062
"movq %%mm7, %%mm6 \n\t" // 0
1063
"pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1064
"pxor %%mm6, %%mm4 \n\t"
1065
"psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1066
"pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1067
"pxor %%mm7, %%mm5 \n\t"
1068
"psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
942
"movq (%0, %1, 2), %%mm2 \n\t"
943
"movq %%mm2, %%mm3 \n\t"
944
"punpcklbw %%mm7, %%mm2 \n\t" // L5
945
"punpckhbw %%mm7, %%mm3 \n\t" // H5
946
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
947
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
948
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
949
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
951
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
952
"punpcklbw %%mm7, %%mm6 \n\t" // L6
953
"psubw %%mm6, %%mm2 \n\t" // L5 - L6
954
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
955
"punpckhbw %%mm7, %%mm6 \n\t" // H6
956
"psubw %%mm6, %%mm3 \n\t" // H5 - H6
958
"paddw %%mm0, %%mm0 \n\t" // 2L4
959
"paddw %%mm1, %%mm1 \n\t" // 2H4
960
"psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
961
"psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
963
"psllw $2, %%mm2 \n\t" // 4L5 - 4L6
964
"psllw $2, %%mm3 \n\t" // 4H5 - 4H6
965
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
966
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
968
"movq (%0, %1, 4), %%mm2 \n\t"
969
"movq %%mm2, %%mm3 \n\t"
970
"punpcklbw %%mm7, %%mm2 \n\t" // L7
971
"punpckhbw %%mm7, %%mm3 \n\t" // H7
973
"paddw %%mm2, %%mm2 \n\t" // 2L7
974
"paddw %%mm3, %%mm3 \n\t" // 2H7
975
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
976
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
978
"movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
979
"movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
982
"movq %%mm7, %%mm6 \n\t" // 0
983
"psubw %%mm0, %%mm6 \n\t"
984
"pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
985
"movq %%mm7, %%mm6 \n\t" // 0
986
"psubw %%mm1, %%mm6 \n\t"
987
"pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
988
"movq %%mm7, %%mm6 \n\t" // 0
989
"psubw %%mm2, %%mm6 \n\t"
990
"pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
991
"movq %%mm7, %%mm6 \n\t" // 0
992
"psubw %%mm3, %%mm6 \n\t"
993
"pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
995
"movq %%mm7, %%mm6 \n\t" // 0
996
"pcmpgtw %%mm0, %%mm6 \n\t"
997
"pxor %%mm6, %%mm0 \n\t"
998
"psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
999
"movq %%mm7, %%mm6 \n\t" // 0
1000
"pcmpgtw %%mm1, %%mm6 \n\t"
1001
"pxor %%mm6, %%mm1 \n\t"
1002
"psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1003
"movq %%mm7, %%mm6 \n\t" // 0
1004
"pcmpgtw %%mm2, %%mm6 \n\t"
1005
"pxor %%mm6, %%mm2 \n\t"
1006
"psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1007
"movq %%mm7, %%mm6 \n\t" // 0
1008
"pcmpgtw %%mm3, %%mm6 \n\t"
1009
"pxor %%mm6, %%mm3 \n\t"
1010
"psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1014
"pminsw %%mm2, %%mm0 \n\t"
1015
"pminsw %%mm3, %%mm1 \n\t"
1017
"movq %%mm0, %%mm6 \n\t"
1018
"psubusw %%mm2, %%mm6 \n\t"
1019
"psubw %%mm6, %%mm0 \n\t"
1020
"movq %%mm1, %%mm6 \n\t"
1021
"psubusw %%mm3, %%mm6 \n\t"
1022
"psubw %%mm6, %%mm1 \n\t"
1025
"movd %2, %%mm2 \n\t" // QP
1026
"punpcklbw %%mm7, %%mm2 \n\t"
1028
"movq %%mm7, %%mm6 \n\t" // 0
1029
"pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1030
"pxor %%mm6, %%mm4 \n\t"
1031
"psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1032
"pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1033
"pxor %%mm7, %%mm5 \n\t"
1034
"psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1070
"psllw $3, %%mm2 \n\t" // 8QP
1071
"movq %%mm2, %%mm3 \n\t" // 8QP
1072
"pcmpgtw %%mm4, %%mm2 \n\t"
1073
"pcmpgtw %%mm5, %%mm3 \n\t"
1074
"pand %%mm2, %%mm4 \n\t"
1075
"pand %%mm3, %%mm5 \n\t"
1078
"psubusw %%mm0, %%mm4 \n\t" // hd
1079
"psubusw %%mm1, %%mm5 \n\t" // ld
1082
"movq "MANGLE(w05)", %%mm2 \n\t" // 5
1083
"pmullw %%mm2, %%mm4 \n\t"
1084
"pmullw %%mm2, %%mm5 \n\t"
1085
"movq "MANGLE(w20)", %%mm2 \n\t" // 32
1086
"paddw %%mm2, %%mm4 \n\t"
1087
"paddw %%mm2, %%mm5 \n\t"
1088
"psrlw $6, %%mm4 \n\t"
1089
"psrlw $6, %%mm5 \n\t"
1091
"movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4
1092
"movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4
1094
"pxor %%mm2, %%mm2 \n\t"
1095
"pxor %%mm3, %%mm3 \n\t"
1097
"pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1098
"pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
1099
"pxor %%mm2, %%mm0 \n\t"
1100
"pxor %%mm3, %%mm1 \n\t"
1101
"psubw %%mm2, %%mm0 \n\t" // |L3-L4|
1102
"psubw %%mm3, %%mm1 \n\t" // |H3-H4|
1103
"psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
1104
"psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
1106
"pxor %%mm6, %%mm2 \n\t"
1107
"pxor %%mm7, %%mm3 \n\t"
1108
"pand %%mm2, %%mm4 \n\t"
1109
"pand %%mm3, %%mm5 \n\t"
1112
"pminsw %%mm0, %%mm4 \n\t"
1113
"pminsw %%mm1, %%mm5 \n\t"
1036
"psllw $3, %%mm2 \n\t" // 8QP
1037
"movq %%mm2, %%mm3 \n\t" // 8QP
1038
"pcmpgtw %%mm4, %%mm2 \n\t"
1039
"pcmpgtw %%mm5, %%mm3 \n\t"
1040
"pand %%mm2, %%mm4 \n\t"
1041
"pand %%mm3, %%mm5 \n\t"
1044
"psubusw %%mm0, %%mm4 \n\t" // hd
1045
"psubusw %%mm1, %%mm5 \n\t" // ld
1048
"movq "MANGLE(w05)", %%mm2 \n\t" // 5
1049
"pmullw %%mm2, %%mm4 \n\t"
1050
"pmullw %%mm2, %%mm5 \n\t"
1051
"movq "MANGLE(w20)", %%mm2 \n\t" // 32
1052
"paddw %%mm2, %%mm4 \n\t"
1053
"paddw %%mm2, %%mm5 \n\t"
1054
"psrlw $6, %%mm4 \n\t"
1055
"psrlw $6, %%mm5 \n\t"
1057
"movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4
1058
"movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4
1060
"pxor %%mm2, %%mm2 \n\t"
1061
"pxor %%mm3, %%mm3 \n\t"
1063
"pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
1064
"pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
1065
"pxor %%mm2, %%mm0 \n\t"
1066
"pxor %%mm3, %%mm1 \n\t"
1067
"psubw %%mm2, %%mm0 \n\t" // |L3-L4|
1068
"psubw %%mm3, %%mm1 \n\t" // |H3-H4|
1069
"psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
1070
"psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
1072
"pxor %%mm6, %%mm2 \n\t"
1073
"pxor %%mm7, %%mm3 \n\t"
1074
"pand %%mm2, %%mm4 \n\t"
1075
"pand %%mm3, %%mm5 \n\t"
1078
"pminsw %%mm0, %%mm4 \n\t"
1079
"pminsw %%mm1, %%mm5 \n\t"
1115
"movq %%mm4, %%mm2 \n\t"
1116
"psubusw %%mm0, %%mm2 \n\t"
1117
"psubw %%mm2, %%mm4 \n\t"
1118
"movq %%mm5, %%mm2 \n\t"
1119
"psubusw %%mm1, %%mm2 \n\t"
1120
"psubw %%mm2, %%mm5 \n\t"
1081
"movq %%mm4, %%mm2 \n\t"
1082
"psubusw %%mm0, %%mm2 \n\t"
1083
"psubw %%mm2, %%mm4 \n\t"
1084
"movq %%mm5, %%mm2 \n\t"
1085
"psubusw %%mm1, %%mm2 \n\t"
1086
"psubw %%mm2, %%mm5 \n\t"
1122
"pxor %%mm6, %%mm4 \n\t"
1123
"pxor %%mm7, %%mm5 \n\t"
1124
"psubw %%mm6, %%mm4 \n\t"
1125
"psubw %%mm7, %%mm5 \n\t"
1126
"packsswb %%mm5, %%mm4 \n\t"
1127
"movq (%0), %%mm0 \n\t"
1128
"paddb %%mm4, %%mm0 \n\t"
1129
"movq %%mm0, (%0) \n\t"
1130
"movq (%0, %1), %%mm0 \n\t"
1131
"psubb %%mm4, %%mm0 \n\t"
1132
"movq %%mm0, (%0, %1) \n\t"
1135
: "r" ((long)stride), "m" (c->pQPb)
1136
: "%"REG_a, "%"REG_c
1138
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1139
const int l1= stride;
1140
const int l2= stride + l1;
1141
const int l3= stride + l2;
1142
const int l4= stride + l3;
1143
const int l5= stride + l4;
1144
const int l6= stride + l5;
1145
const int l7= stride + l6;
1146
const int l8= stride + l7;
1147
// const int l9= stride + l8;
1150
for(x=0; x<BLOCK_SIZE; x++)
1152
const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1153
if(FFABS(middleEnergy) < 8*c->QP)
1155
const int q=(src[l4] - src[l5])/2;
1156
const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1157
const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1159
int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
1163
d*= FFSIGN(-middleEnergy);
1088
"pxor %%mm6, %%mm4 \n\t"
1089
"pxor %%mm7, %%mm5 \n\t"
1090
"psubw %%mm6, %%mm4 \n\t"
1091
"psubw %%mm7, %%mm5 \n\t"
1092
"packsswb %%mm5, %%mm4 \n\t"
1093
"movq (%0), %%mm0 \n\t"
1094
"paddb %%mm4, %%mm0 \n\t"
1095
"movq %%mm0, (%0) \n\t"
1096
"movq (%0, %1), %%mm0 \n\t"
1097
"psubb %%mm4, %%mm0 \n\t"
1098
"movq %%mm0, (%0, %1) \n\t"
1101
: "r" ((x86_reg)stride), "m" (c->pQPb)
1102
: "%"REG_a, "%"REG_c
1104
#else //HAVE_MMX2 || HAVE_AMD3DNOW
1105
const int l1= stride;
1106
const int l2= stride + l1;
1107
const int l3= stride + l2;
1108
const int l4= stride + l3;
1109
const int l5= stride + l4;
1110
const int l6= stride + l5;
1111
const int l7= stride + l6;
1112
const int l8= stride + l7;
1113
// const int l9= stride + l8;
1116
for(x=0; x<BLOCK_SIZE; x++){
1117
const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1118
if(FFABS(middleEnergy) < 8*c->QP){
1119
const int q=(src[l4] - src[l5])/2;
1120
const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1121
const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1123
int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
1127
d*= FFSIGN(-middleEnergy);
1181
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1142
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
1183
1144
#endif //HAVE_ALTIVEC
1185
#ifndef HAVE_ALTIVEC
1186
1147
static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
1188
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1190
"pxor %%mm6, %%mm6 \n\t"
1191
"pcmpeqb %%mm7, %%mm7 \n\t"
1192
"movq %2, %%mm0 \n\t"
1193
"punpcklbw %%mm6, %%mm0 \n\t"
1194
"psrlw $1, %%mm0 \n\t"
1195
"psubw %%mm7, %%mm0 \n\t"
1196
"packuswb %%mm0, %%mm0 \n\t"
1197
"movq %%mm0, %3 \n\t"
1149
#if HAVE_MMX2 || HAVE_AMD3DNOW
1151
"pxor %%mm6, %%mm6 \n\t"
1152
"pcmpeqb %%mm7, %%mm7 \n\t"
1153
"movq %2, %%mm0 \n\t"
1154
"punpcklbw %%mm6, %%mm0 \n\t"
1155
"psrlw $1, %%mm0 \n\t"
1156
"psubw %%mm7, %%mm0 \n\t"
1157
"packuswb %%mm0, %%mm0 \n\t"
1158
"movq %%mm0, %3 \n\t"
1199
"lea (%0, %1), %%"REG_a" \n\t"
1200
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1160
"lea (%0, %1), %%"REG_a" \n\t"
1161
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1202
1163
// 0 1 2 3 4 5 6 7 8 9
1203
1164
// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1205
1166
#undef FIND_MIN_MAX
1207
1168
#define REAL_FIND_MIN_MAX(addr)\
1208
"movq " #addr ", %%mm0 \n\t"\
1209
"pminub %%mm0, %%mm7 \n\t"\
1210
"pmaxub %%mm0, %%mm6 \n\t"
1169
"movq " #addr ", %%mm0 \n\t"\
1170
"pminub %%mm0, %%mm7 \n\t"\
1171
"pmaxub %%mm0, %%mm6 \n\t"
1212
1173
#define REAL_FIND_MIN_MAX(addr)\
1213
"movq " #addr ", %%mm0 \n\t"\
1214
"movq %%mm7, %%mm1 \n\t"\
1215
"psubusb %%mm0, %%mm6 \n\t"\
1216
"paddb %%mm0, %%mm6 \n\t"\
1217
"psubusb %%mm0, %%mm1 \n\t"\
1218
"psubb %%mm1, %%mm7 \n\t"
1174
"movq " #addr ", %%mm0 \n\t"\
1175
"movq %%mm7, %%mm1 \n\t"\
1176
"psubusb %%mm0, %%mm6 \n\t"\
1177
"paddb %%mm0, %%mm6 \n\t"\
1178
"psubusb %%mm0, %%mm1 \n\t"\
1179
"psubb %%mm1, %%mm7 \n\t"
1220
1181
#define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr)
1228
1189
FIND_MIN_MAX((%%REGd, %1, 2))
1229
1190
FIND_MIN_MAX((%0, %1, 8))
1231
"movq %%mm7, %%mm4 \n\t"
1232
"psrlq $8, %%mm7 \n\t"
1234
"pminub %%mm4, %%mm7 \n\t" // min of pixels
1235
"pshufw $0xF9, %%mm7, %%mm4 \n\t"
1236
"pminub %%mm4, %%mm7 \n\t" // min of pixels
1237
"pshufw $0xFE, %%mm7, %%mm4 \n\t"
1238
"pminub %%mm4, %%mm7 \n\t"
1240
"movq %%mm7, %%mm1 \n\t"
1241
"psubusb %%mm4, %%mm1 \n\t"
1242
"psubb %%mm1, %%mm7 \n\t"
1243
"movq %%mm7, %%mm4 \n\t"
1244
"psrlq $16, %%mm7 \n\t"
1245
"movq %%mm7, %%mm1 \n\t"
1246
"psubusb %%mm4, %%mm1 \n\t"
1247
"psubb %%mm1, %%mm7 \n\t"
1248
"movq %%mm7, %%mm4 \n\t"
1249
"psrlq $32, %%mm7 \n\t"
1250
"movq %%mm7, %%mm1 \n\t"
1251
"psubusb %%mm4, %%mm1 \n\t"
1252
"psubb %%mm1, %%mm7 \n\t"
1256
"movq %%mm6, %%mm4 \n\t"
1257
"psrlq $8, %%mm6 \n\t"
1259
"pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1260
"pshufw $0xF9, %%mm6, %%mm4 \n\t"
1261
"pmaxub %%mm4, %%mm6 \n\t"
1262
"pshufw $0xFE, %%mm6, %%mm4 \n\t"
1263
"pmaxub %%mm4, %%mm6 \n\t"
1265
"psubusb %%mm4, %%mm6 \n\t"
1266
"paddb %%mm4, %%mm6 \n\t"
1267
"movq %%mm6, %%mm4 \n\t"
1268
"psrlq $16, %%mm6 \n\t"
1269
"psubusb %%mm4, %%mm6 \n\t"
1270
"paddb %%mm4, %%mm6 \n\t"
1271
"movq %%mm6, %%mm4 \n\t"
1272
"psrlq $32, %%mm6 \n\t"
1273
"psubusb %%mm4, %%mm6 \n\t"
1274
"paddb %%mm4, %%mm6 \n\t"
1276
"movq %%mm6, %%mm0 \n\t" // max
1277
"psubb %%mm7, %%mm6 \n\t" // max - min
1278
"movd %%mm6, %%ecx \n\t"
1279
"cmpb "MANGLE(deringThreshold)", %%cl \n\t"
1281
"lea -24(%%"REG_SP"), %%"REG_c" \n\t"
1282
"and "ALIGN_MASK", %%"REG_c" \n\t"
1283
PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1284
"punpcklbw %%mm7, %%mm7 \n\t"
1285
"punpcklbw %%mm7, %%mm7 \n\t"
1286
"punpcklbw %%mm7, %%mm7 \n\t"
1287
"movq %%mm7, (%%"REG_c") \n\t"
1289
"movq (%0), %%mm0 \n\t" // L10
1290
"movq %%mm0, %%mm1 \n\t" // L10
1291
"movq %%mm0, %%mm2 \n\t" // L10
1292
"psllq $8, %%mm1 \n\t"
1293
"psrlq $8, %%mm2 \n\t"
1294
"movd -4(%0), %%mm3 \n\t"
1295
"movd 8(%0), %%mm4 \n\t"
1296
"psrlq $24, %%mm3 \n\t"
1297
"psllq $56, %%mm4 \n\t"
1298
"por %%mm3, %%mm1 \n\t" // L00
1299
"por %%mm4, %%mm2 \n\t" // L20
1300
"movq %%mm1, %%mm3 \n\t" // L00
1301
PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1302
PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1303
"psubusb %%mm7, %%mm0 \n\t"
1304
"psubusb %%mm7, %%mm2 \n\t"
1305
"psubusb %%mm7, %%mm3 \n\t"
1306
"pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1307
"pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1308
"pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1309
"paddb %%mm2, %%mm0 \n\t"
1310
"paddb %%mm3, %%mm0 \n\t"
1312
"movq (%%"REG_a"), %%mm2 \n\t" // L11
1313
"movq %%mm2, %%mm3 \n\t" // L11
1314
"movq %%mm2, %%mm4 \n\t" // L11
1315
"psllq $8, %%mm3 \n\t"
1316
"psrlq $8, %%mm4 \n\t"
1317
"movd -4(%%"REG_a"), %%mm5 \n\t"
1318
"movd 8(%%"REG_a"), %%mm6 \n\t"
1319
"psrlq $24, %%mm5 \n\t"
1320
"psllq $56, %%mm6 \n\t"
1321
"por %%mm5, %%mm3 \n\t" // L01
1322
"por %%mm6, %%mm4 \n\t" // L21
1323
"movq %%mm3, %%mm5 \n\t" // L01
1324
PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1325
PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1326
"psubusb %%mm7, %%mm2 \n\t"
1327
"psubusb %%mm7, %%mm4 \n\t"
1328
"psubusb %%mm7, %%mm5 \n\t"
1329
"pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1330
"pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1331
"pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1332
"paddb %%mm4, %%mm2 \n\t"
1333
"paddb %%mm5, %%mm2 \n\t"
1192
"movq %%mm7, %%mm4 \n\t"
1193
"psrlq $8, %%mm7 \n\t"
1195
"pminub %%mm4, %%mm7 \n\t" // min of pixels
1196
"pshufw $0xF9, %%mm7, %%mm4 \n\t"
1197
"pminub %%mm4, %%mm7 \n\t" // min of pixels
1198
"pshufw $0xFE, %%mm7, %%mm4 \n\t"
1199
"pminub %%mm4, %%mm7 \n\t"
1201
"movq %%mm7, %%mm1 \n\t"
1202
"psubusb %%mm4, %%mm1 \n\t"
1203
"psubb %%mm1, %%mm7 \n\t"
1204
"movq %%mm7, %%mm4 \n\t"
1205
"psrlq $16, %%mm7 \n\t"
1206
"movq %%mm7, %%mm1 \n\t"
1207
"psubusb %%mm4, %%mm1 \n\t"
1208
"psubb %%mm1, %%mm7 \n\t"
1209
"movq %%mm7, %%mm4 \n\t"
1210
"psrlq $32, %%mm7 \n\t"
1211
"movq %%mm7, %%mm1 \n\t"
1212
"psubusb %%mm4, %%mm1 \n\t"
1213
"psubb %%mm1, %%mm7 \n\t"
1217
"movq %%mm6, %%mm4 \n\t"
1218
"psrlq $8, %%mm6 \n\t"
1220
"pmaxub %%mm4, %%mm6 \n\t" // max of pixels
1221
"pshufw $0xF9, %%mm6, %%mm4 \n\t"
1222
"pmaxub %%mm4, %%mm6 \n\t"
1223
"pshufw $0xFE, %%mm6, %%mm4 \n\t"
1224
"pmaxub %%mm4, %%mm6 \n\t"
1226
"psubusb %%mm4, %%mm6 \n\t"
1227
"paddb %%mm4, %%mm6 \n\t"
1228
"movq %%mm6, %%mm4 \n\t"
1229
"psrlq $16, %%mm6 \n\t"
1230
"psubusb %%mm4, %%mm6 \n\t"
1231
"paddb %%mm4, %%mm6 \n\t"
1232
"movq %%mm6, %%mm4 \n\t"
1233
"psrlq $32, %%mm6 \n\t"
1234
"psubusb %%mm4, %%mm6 \n\t"
1235
"paddb %%mm4, %%mm6 \n\t"
1237
"movq %%mm6, %%mm0 \n\t" // max
1238
"psubb %%mm7, %%mm6 \n\t" // max - min
1239
"movd %%mm6, %%ecx \n\t"
1240
"cmpb "MANGLE(deringThreshold)", %%cl \n\t"
1242
"lea -24(%%"REG_SP"), %%"REG_c" \n\t"
1243
"and "ALIGN_MASK", %%"REG_c" \n\t"
1244
PAVGB(%%mm0, %%mm7) // a=(max + min)/2
1245
"punpcklbw %%mm7, %%mm7 \n\t"
1246
"punpcklbw %%mm7, %%mm7 \n\t"
1247
"punpcklbw %%mm7, %%mm7 \n\t"
1248
"movq %%mm7, (%%"REG_c") \n\t"
1250
"movq (%0), %%mm0 \n\t" // L10
1251
"movq %%mm0, %%mm1 \n\t" // L10
1252
"movq %%mm0, %%mm2 \n\t" // L10
1253
"psllq $8, %%mm1 \n\t"
1254
"psrlq $8, %%mm2 \n\t"
1255
"movd -4(%0), %%mm3 \n\t"
1256
"movd 8(%0), %%mm4 \n\t"
1257
"psrlq $24, %%mm3 \n\t"
1258
"psllq $56, %%mm4 \n\t"
1259
"por %%mm3, %%mm1 \n\t" // L00
1260
"por %%mm4, %%mm2 \n\t" // L20
1261
"movq %%mm1, %%mm3 \n\t" // L00
1262
PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
1263
PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
1264
"psubusb %%mm7, %%mm0 \n\t"
1265
"psubusb %%mm7, %%mm2 \n\t"
1266
"psubusb %%mm7, %%mm3 \n\t"
1267
"pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
1268
"pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
1269
"pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
1270
"paddb %%mm2, %%mm0 \n\t"
1271
"paddb %%mm3, %%mm0 \n\t"
1273
"movq (%%"REG_a"), %%mm2 \n\t" // L11
1274
"movq %%mm2, %%mm3 \n\t" // L11
1275
"movq %%mm2, %%mm4 \n\t" // L11
1276
"psllq $8, %%mm3 \n\t"
1277
"psrlq $8, %%mm4 \n\t"
1278
"movd -4(%%"REG_a"), %%mm5 \n\t"
1279
"movd 8(%%"REG_a"), %%mm6 \n\t"
1280
"psrlq $24, %%mm5 \n\t"
1281
"psllq $56, %%mm6 \n\t"
1282
"por %%mm5, %%mm3 \n\t" // L01
1283
"por %%mm6, %%mm4 \n\t" // L21
1284
"movq %%mm3, %%mm5 \n\t" // L01
1285
PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
1286
PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
1287
"psubusb %%mm7, %%mm2 \n\t"
1288
"psubusb %%mm7, %%mm4 \n\t"
1289
"psubusb %%mm7, %%mm5 \n\t"
1290
"pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
1291
"pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
1292
"pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
1293
"paddb %%mm4, %%mm2 \n\t"
1294
"paddb %%mm5, %%mm2 \n\t"
1335
1296
#define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1336
"movq " #src ", " #sx " \n\t" /* src[0] */\
1337
"movq " #sx ", " #lx " \n\t" /* src[0] */\
1338
"movq " #sx ", " #t0 " \n\t" /* src[0] */\
1339
"psllq $8, " #lx " \n\t"\
1340
"psrlq $8, " #t0 " \n\t"\
1341
"movd -4" #src ", " #t1 " \n\t"\
1342
"psrlq $24, " #t1 " \n\t"\
1343
"por " #t1 ", " #lx " \n\t" /* src[-1] */\
1344
"movd 8" #src ", " #t1 " \n\t"\
1345
"psllq $56, " #t1 " \n\t"\
1346
"por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1347
"movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1348
PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1349
PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1351
"movq " #lx ", 8(%%"REG_c") \n\t"\
1352
"movq (%%"REG_c"), " #lx " \n\t"\
1353
"psubusb " #lx ", " #t1 " \n\t"\
1354
"psubusb " #lx ", " #t0 " \n\t"\
1355
"psubusb " #lx ", " #sx " \n\t"\
1356
"movq "MANGLE(b00)", " #lx " \n\t"\
1357
"pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1358
"pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1359
"pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1360
"paddb " #t1 ", " #t0 " \n\t"\
1361
"paddb " #t0 ", " #sx " \n\t"\
1297
"movq " #src ", " #sx " \n\t" /* src[0] */\
1298
"movq " #sx ", " #lx " \n\t" /* src[0] */\
1299
"movq " #sx ", " #t0 " \n\t" /* src[0] */\
1300
"psllq $8, " #lx " \n\t"\
1301
"psrlq $8, " #t0 " \n\t"\
1302
"movd -4" #src ", " #t1 " \n\t"\
1303
"psrlq $24, " #t1 " \n\t"\
1304
"por " #t1 ", " #lx " \n\t" /* src[-1] */\
1305
"movd 8" #src ", " #t1 " \n\t"\
1306
"psllq $56, " #t1 " \n\t"\
1307
"por " #t1 ", " #t0 " \n\t" /* src[+1] */\
1308
"movq " #lx ", " #t1 " \n\t" /* src[-1] */\
1309
PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
1310
PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
1312
"movq " #lx ", 8(%%"REG_c") \n\t"\
1313
"movq (%%"REG_c"), " #lx " \n\t"\
1314
"psubusb " #lx ", " #t1 " \n\t"\
1315
"psubusb " #lx ", " #t0 " \n\t"\
1316
"psubusb " #lx ", " #sx " \n\t"\
1317
"movq "MANGLE(b00)", " #lx " \n\t"\
1318
"pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
1319
"pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
1320
"pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
1321
"paddb " #t1 ", " #t0 " \n\t"\
1322
"paddb " #t0 ", " #sx " \n\t"\
1363
PAVGB(plx, pplx) /* filtered */\
1364
"movq " #dst ", " #t0 " \n\t" /* dst */\
1365
"movq " #t0 ", " #t1 " \n\t" /* dst */\
1366
"psubusb %3, " #t0 " \n\t"\
1367
"paddusb %3, " #t1 " \n\t"\
1369
PMINUB(t1, pplx, t0)\
1370
"paddb " #sx ", " #ppsx " \n\t"\
1371
"paddb " #psx ", " #ppsx " \n\t"\
1372
"#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1373
"pand "MANGLE(b08)", " #ppsx " \n\t"\
1374
"pcmpeqb " #lx ", " #ppsx " \n\t"\
1375
"pand " #ppsx ", " #pplx " \n\t"\
1376
"pandn " #dst ", " #ppsx " \n\t"\
1377
"por " #pplx ", " #ppsx " \n\t"\
1378
"movq " #ppsx ", " #dst " \n\t"\
1379
"movq 8(%%"REG_c"), " #lx " \n\t"
1324
PAVGB(plx, pplx) /* filtered */\
1325
"movq " #dst ", " #t0 " \n\t" /* dst */\
1326
"movq " #t0 ", " #t1 " \n\t" /* dst */\
1327
"psubusb %3, " #t0 " \n\t"\
1328
"paddusb %3, " #t1 " \n\t"\
1330
PMINUB(t1, pplx, t0)\
1331
"paddb " #sx ", " #ppsx " \n\t"\
1332
"paddb " #psx ", " #ppsx " \n\t"\
1333
"#paddb "MANGLE(b02)", " #ppsx " \n\t"\
1334
"pand "MANGLE(b08)", " #ppsx " \n\t"\
1335
"pcmpeqb " #lx ", " #ppsx " \n\t"\
1336
"pand " #ppsx ", " #pplx " \n\t"\
1337
"pandn " #dst ", " #ppsx " \n\t"\
1338
"por " #pplx ", " #ppsx " \n\t"\
1339
"movq " #ppsx ", " #dst " \n\t"\
1340
"movq 8(%%"REG_c"), " #lx " \n\t"
1381
1342
#define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1382
1343
REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1)
1804
1750
DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) )
1805
1751
DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4))
1807
"movq %%mm0, (%2) \n\t"
1808
"movq %%mm1, (%3) \n\t"
1809
: : "r" (src), "r" ((long)stride), "r"(tmp), "r"(tmp2)
1810
: "%"REG_a, "%"REG_d
1812
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1821
src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
1823
src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
1825
src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
1827
src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
1829
src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
1831
src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
1833
src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
1835
src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);
1842
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1753
"movq %%mm0, (%2) \n\t"
1754
"movq %%mm1, (%3) \n\t"
1755
: : "r" (src), "r" ((x86_reg)stride), "r"(tmp), "r"(tmp2)
1756
: "%"REG_a, "%"REG_d
1758
#else //HAVE_MMX2 || HAVE_AMD3DNOW
1766
src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
1768
src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
1770
src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
1772
src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
1774
src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
1776
src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
1778
src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
1780
src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);
1787
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
1846
1791
* Deinterlaces the given block by filtering all lines with a (1 2 1) filter.
1847
1792
* will be called for every 8x8 block and can read & write from line 4-15
1848
* lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1793
* lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1849
1794
* lines 4-12 will be read into the deblocking filter and should be deinterlaced
1850
1795
* this filter will read lines 4-13 and write 4-11
1852
1797
static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp)
1854
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1857
"lea (%0, %1), %%"REG_a" \n\t"
1858
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1799
#if HAVE_MMX2 || HAVE_AMD3DNOW
1802
"lea (%0, %1), %%"REG_a" \n\t"
1803
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1859
1804
// 0 1 2 3 4 5 6 7 8 9
1860
1805
// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1862
"movq (%2), %%mm0 \n\t" // L0
1863
"movq (%%"REG_a"), %%mm1 \n\t" // L2
1864
PAVGB(%%mm1, %%mm0) // L0+L2
1865
"movq (%0), %%mm2 \n\t" // L1
1867
"movq %%mm0, (%0) \n\t"
1868
"movq (%%"REG_a", %1), %%mm0 \n\t" // L3
1869
PAVGB(%%mm0, %%mm2) // L1+L3
1870
PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1871
"movq %%mm2, (%%"REG_a") \n\t"
1872
"movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4
1873
PAVGB(%%mm2, %%mm1) // L2+L4
1874
PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1875
"movq %%mm1, (%%"REG_a", %1) \n\t"
1876
"movq (%0, %1, 4), %%mm1 \n\t" // L5
1877
PAVGB(%%mm1, %%mm0) // L3+L5
1878
PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1879
"movq %%mm0, (%%"REG_a", %1, 2) \n\t"
1880
"movq (%%"REG_d"), %%mm0 \n\t" // L6
1881
PAVGB(%%mm0, %%mm2) // L4+L6
1882
PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1883
"movq %%mm2, (%0, %1, 4) \n\t"
1884
"movq (%%"REG_d", %1), %%mm2 \n\t" // L7
1885
PAVGB(%%mm2, %%mm1) // L5+L7
1886
PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1887
"movq %%mm1, (%%"REG_d") \n\t"
1888
"movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8
1889
PAVGB(%%mm1, %%mm0) // L6+L8
1890
PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1891
"movq %%mm0, (%%"REG_d", %1) \n\t"
1892
"movq (%0, %1, 8), %%mm0 \n\t" // L9
1893
PAVGB(%%mm0, %%mm2) // L7+L9
1894
PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1895
"movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1896
"movq %%mm1, (%2) \n\t"
1898
: : "r" (src), "r" ((long)stride), "r" (tmp)
1899
: "%"REG_a, "%"REG_d
1901
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1906
a= *(uint32_t*)&tmp[stride*0];
1907
b= *(uint32_t*)&src[stride*0];
1908
c= *(uint32_t*)&src[stride*1];
1909
a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1910
*(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1912
a= *(uint32_t*)&src[stride*2];
1913
b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1914
*(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1916
b= *(uint32_t*)&src[stride*3];
1917
c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1918
*(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1920
c= *(uint32_t*)&src[stride*4];
1921
a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1922
*(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1924
a= *(uint32_t*)&src[stride*5];
1925
b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1926
*(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1928
b= *(uint32_t*)&src[stride*6];
1929
c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1930
*(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1932
c= *(uint32_t*)&src[stride*7];
1933
a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1934
*(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1936
a= *(uint32_t*)&src[stride*8];
1937
b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1938
*(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1940
*(uint32_t*)&tmp[stride*0]= c;
1944
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1807
"movq (%2), %%mm0 \n\t" // L0
1808
"movq (%%"REG_a"), %%mm1 \n\t" // L2
1809
PAVGB(%%mm1, %%mm0) // L0+L2
1810
"movq (%0), %%mm2 \n\t" // L1
1812
"movq %%mm0, (%0) \n\t"
1813
"movq (%%"REG_a", %1), %%mm0 \n\t" // L3
1814
PAVGB(%%mm0, %%mm2) // L1+L3
1815
PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
1816
"movq %%mm2, (%%"REG_a") \n\t"
1817
"movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4
1818
PAVGB(%%mm2, %%mm1) // L2+L4
1819
PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
1820
"movq %%mm1, (%%"REG_a", %1) \n\t"
1821
"movq (%0, %1, 4), %%mm1 \n\t" // L5
1822
PAVGB(%%mm1, %%mm0) // L3+L5
1823
PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
1824
"movq %%mm0, (%%"REG_a", %1, 2) \n\t"
1825
"movq (%%"REG_d"), %%mm0 \n\t" // L6
1826
PAVGB(%%mm0, %%mm2) // L4+L6
1827
PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
1828
"movq %%mm2, (%0, %1, 4) \n\t"
1829
"movq (%%"REG_d", %1), %%mm2 \n\t" // L7
1830
PAVGB(%%mm2, %%mm1) // L5+L7
1831
PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
1832
"movq %%mm1, (%%"REG_d") \n\t"
1833
"movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8
1834
PAVGB(%%mm1, %%mm0) // L6+L8
1835
PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
1836
"movq %%mm0, (%%"REG_d", %1) \n\t"
1837
"movq (%0, %1, 8), %%mm0 \n\t" // L9
1838
PAVGB(%%mm0, %%mm2) // L7+L9
1839
PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
1840
"movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1841
"movq %%mm1, (%2) \n\t"
1843
: : "r" (src), "r" ((x86_reg)stride), "r" (tmp)
1844
: "%"REG_a, "%"REG_d
1846
#else //HAVE_MMX2 || HAVE_AMD3DNOW
1851
a= *(uint32_t*)&tmp[stride*0];
1852
b= *(uint32_t*)&src[stride*0];
1853
c= *(uint32_t*)&src[stride*1];
1854
a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1855
*(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1857
a= *(uint32_t*)&src[stride*2];
1858
b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1859
*(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1861
b= *(uint32_t*)&src[stride*3];
1862
c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1863
*(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1865
c= *(uint32_t*)&src[stride*4];
1866
a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1867
*(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1869
a= *(uint32_t*)&src[stride*5];
1870
b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1871
*(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1873
b= *(uint32_t*)&src[stride*6];
1874
c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
1875
*(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);
1877
c= *(uint32_t*)&src[stride*7];
1878
a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
1879
*(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
1881
a= *(uint32_t*)&src[stride*8];
1882
b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
1883
*(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);
1885
*(uint32_t*)&tmp[stride*0]= c;
1889
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
1948
1893
* Deinterlaces the given block by applying a median filter to every second line.
1949
1894
* will be called for every 8x8 block and can read & write from line 4-15,
1950
* lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1895
* lines 0-3 have been passed through the deblock / dering filters already, but can be read, too.
1951
1896
* lines 4-12 will be read into the deblocking filter and should be deinterlaced
1953
1898
static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1959
"lea (%0, %1), %%"REG_a" \n\t"
1960
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1904
"lea (%0, %1), %%"REG_a" \n\t"
1905
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1961
1906
// 0 1 2 3 4 5 6 7 8 9
1962
1907
// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
1964
"movq (%0), %%mm0 \n\t" //
1965
"movq (%%"REG_a", %1), %%mm2 \n\t" //
1966
"movq (%%"REG_a"), %%mm1 \n\t" //
1967
"movq %%mm0, %%mm3 \n\t"
1968
"pmaxub %%mm1, %%mm0 \n\t" //
1969
"pminub %%mm3, %%mm1 \n\t" //
1970
"pmaxub %%mm2, %%mm1 \n\t" //
1971
"pminub %%mm1, %%mm0 \n\t"
1972
"movq %%mm0, (%%"REG_a") \n\t"
1974
"movq (%0, %1, 4), %%mm0 \n\t" //
1975
"movq (%%"REG_a", %1, 2), %%mm1 \n\t" //
1976
"movq %%mm2, %%mm3 \n\t"
1977
"pmaxub %%mm1, %%mm2 \n\t" //
1978
"pminub %%mm3, %%mm1 \n\t" //
1979
"pmaxub %%mm0, %%mm1 \n\t" //
1980
"pminub %%mm1, %%mm2 \n\t"
1981
"movq %%mm2, (%%"REG_a", %1, 2) \n\t"
1983
"movq (%%"REG_d"), %%mm2 \n\t" //
1984
"movq (%%"REG_d", %1), %%mm1 \n\t" //
1985
"movq %%mm2, %%mm3 \n\t"
1986
"pmaxub %%mm0, %%mm2 \n\t" //
1987
"pminub %%mm3, %%mm0 \n\t" //
1988
"pmaxub %%mm1, %%mm0 \n\t" //
1989
"pminub %%mm0, %%mm2 \n\t"
1990
"movq %%mm2, (%%"REG_d") \n\t"
1992
"movq (%%"REG_d", %1, 2), %%mm2 \n\t" //
1993
"movq (%0, %1, 8), %%mm0 \n\t" //
1994
"movq %%mm2, %%mm3 \n\t"
1995
"pmaxub %%mm0, %%mm2 \n\t" //
1996
"pminub %%mm3, %%mm0 \n\t" //
1997
"pmaxub %%mm1, %%mm0 \n\t" //
1998
"pminub %%mm0, %%mm2 \n\t"
1999
"movq %%mm2, (%%"REG_d", %1, 2) \n\t"
2002
: : "r" (src), "r" ((long)stride)
2003
: "%"REG_a, "%"REG_d
1909
"movq (%0), %%mm0 \n\t" //
1910
"movq (%%"REG_a", %1), %%mm2 \n\t" //
1911
"movq (%%"REG_a"), %%mm1 \n\t" //
1912
"movq %%mm0, %%mm3 \n\t"
1913
"pmaxub %%mm1, %%mm0 \n\t" //
1914
"pminub %%mm3, %%mm1 \n\t" //
1915
"pmaxub %%mm2, %%mm1 \n\t" //
1916
"pminub %%mm1, %%mm0 \n\t"
1917
"movq %%mm0, (%%"REG_a") \n\t"
1919
"movq (%0, %1, 4), %%mm0 \n\t" //
1920
"movq (%%"REG_a", %1, 2), %%mm1 \n\t" //
1921
"movq %%mm2, %%mm3 \n\t"
1922
"pmaxub %%mm1, %%mm2 \n\t" //
1923
"pminub %%mm3, %%mm1 \n\t" //
1924
"pmaxub %%mm0, %%mm1 \n\t" //
1925
"pminub %%mm1, %%mm2 \n\t"
1926
"movq %%mm2, (%%"REG_a", %1, 2) \n\t"
1928
"movq (%%"REG_d"), %%mm2 \n\t" //
1929
"movq (%%"REG_d", %1), %%mm1 \n\t" //
1930
"movq %%mm2, %%mm3 \n\t"
1931
"pmaxub %%mm0, %%mm2 \n\t" //
1932
"pminub %%mm3, %%mm0 \n\t" //
1933
"pmaxub %%mm1, %%mm0 \n\t" //
1934
"pminub %%mm0, %%mm2 \n\t"
1935
"movq %%mm2, (%%"REG_d") \n\t"
1937
"movq (%%"REG_d", %1, 2), %%mm2 \n\t" //
1938
"movq (%0, %1, 8), %%mm0 \n\t" //
1939
"movq %%mm2, %%mm3 \n\t"
1940
"pmaxub %%mm0, %%mm2 \n\t" //
1941
"pminub %%mm3, %%mm0 \n\t" //
1942
"pmaxub %%mm1, %%mm0 \n\t" //
1943
"pminub %%mm0, %%mm2 \n\t"
1944
"movq %%mm2, (%%"REG_d", %1, 2) \n\t"
1947
: : "r" (src), "r" ((x86_reg)stride)
1948
: "%"REG_a, "%"REG_d
2006
1951
#else // MMX without MMX2
2008
"lea (%0, %1), %%"REG_a" \n\t"
2009
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
1953
"lea (%0, %1), %%"REG_a" \n\t"
1954
"lea (%%"REG_a", %1, 4), %%"REG_d" \n\t"
2010
1955
// 0 1 2 3 4 5 6 7 8 9
2011
1956
// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2012
"pxor %%mm7, %%mm7 \n\t"
1957
"pxor %%mm7, %%mm7 \n\t"
2014
1959
#define REAL_MEDIAN(a,b,c)\
2015
"movq " #a ", %%mm0 \n\t"\
2016
"movq " #b ", %%mm2 \n\t"\
2017
"movq " #c ", %%mm1 \n\t"\
2018
"movq %%mm0, %%mm3 \n\t"\
2019
"movq %%mm1, %%mm4 \n\t"\
2020
"movq %%mm2, %%mm5 \n\t"\
2021
"psubusb %%mm1, %%mm3 \n\t"\
2022
"psubusb %%mm2, %%mm4 \n\t"\
2023
"psubusb %%mm0, %%mm5 \n\t"\
2024
"pcmpeqb %%mm7, %%mm3 \n\t"\
2025
"pcmpeqb %%mm7, %%mm4 \n\t"\
2026
"pcmpeqb %%mm7, %%mm5 \n\t"\
2027
"movq %%mm3, %%mm6 \n\t"\
2028
"pxor %%mm4, %%mm3 \n\t"\
2029
"pxor %%mm5, %%mm4 \n\t"\
2030
"pxor %%mm6, %%mm5 \n\t"\
2031
"por %%mm3, %%mm1 \n\t"\
2032
"por %%mm4, %%mm2 \n\t"\
2033
"por %%mm5, %%mm0 \n\t"\
2034
"pand %%mm2, %%mm0 \n\t"\
2035
"pand %%mm1, %%mm0 \n\t"\
2036
"movq %%mm0, " #b " \n\t"
1960
"movq " #a ", %%mm0 \n\t"\
1961
"movq " #b ", %%mm2 \n\t"\
1962
"movq " #c ", %%mm1 \n\t"\
1963
"movq %%mm0, %%mm3 \n\t"\
1964
"movq %%mm1, %%mm4 \n\t"\
1965
"movq %%mm2, %%mm5 \n\t"\
1966
"psubusb %%mm1, %%mm3 \n\t"\
1967
"psubusb %%mm2, %%mm4 \n\t"\
1968
"psubusb %%mm0, %%mm5 \n\t"\
1969
"pcmpeqb %%mm7, %%mm3 \n\t"\
1970
"pcmpeqb %%mm7, %%mm4 \n\t"\
1971
"pcmpeqb %%mm7, %%mm5 \n\t"\
1972
"movq %%mm3, %%mm6 \n\t"\
1973
"pxor %%mm4, %%mm3 \n\t"\
1974
"pxor %%mm5, %%mm4 \n\t"\
1975
"pxor %%mm6, %%mm5 \n\t"\
1976
"por %%mm3, %%mm1 \n\t"\
1977
"por %%mm4, %%mm2 \n\t"\
1978
"por %%mm5, %%mm0 \n\t"\
1979
"pand %%mm2, %%mm0 \n\t"\
1980
"pand %%mm1, %%mm0 \n\t"\
1981
"movq %%mm0, " #b " \n\t"
2037
1982
#define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c)
2039
1984
MEDIAN((%0) , (%%REGa) , (%%REGa, %1))
2161
2104
static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
2164
"lea (%0, %1), %%"REG_a" \n\t"
2165
"lea (%%"REG_a",%1,4), %%"REG_d" \n\t"
2107
"lea (%0, %1), %%"REG_a" \n\t"
2108
"lea (%%"REG_a",%1,4), %%"REG_d" \n\t"
2166
2109
// 0 1 2 3 4 5 6 7 8 9
2167
2110
// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
2168
"movq (%2), %%mm0 \n\t" // 12345678
2169
"movq 16(%2), %%mm1 \n\t" // abcdefgh
2170
"movq %%mm0, %%mm2 \n\t" // 12345678
2171
"punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2172
"punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2174
"movq 32(%2), %%mm1 \n\t"
2175
"movq 48(%2), %%mm3 \n\t"
2176
"movq %%mm1, %%mm4 \n\t"
2177
"punpcklbw %%mm3, %%mm1 \n\t"
2178
"punpckhbw %%mm3, %%mm4 \n\t"
2180
"movq %%mm0, %%mm3 \n\t"
2181
"punpcklwd %%mm1, %%mm0 \n\t"
2182
"punpckhwd %%mm1, %%mm3 \n\t"
2183
"movq %%mm2, %%mm1 \n\t"
2184
"punpcklwd %%mm4, %%mm2 \n\t"
2185
"punpckhwd %%mm4, %%mm1 \n\t"
2187
"movd %%mm0, (%0) \n\t"
2188
"psrlq $32, %%mm0 \n\t"
2189
"movd %%mm0, (%%"REG_a") \n\t"
2190
"movd %%mm3, (%%"REG_a", %1) \n\t"
2191
"psrlq $32, %%mm3 \n\t"
2192
"movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2193
"movd %%mm2, (%0, %1, 4) \n\t"
2194
"psrlq $32, %%mm2 \n\t"
2195
"movd %%mm2, (%%"REG_d") \n\t"
2196
"movd %%mm1, (%%"REG_d", %1) \n\t"
2197
"psrlq $32, %%mm1 \n\t"
2198
"movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2201
"movq 64(%2), %%mm0 \n\t" // 12345678
2202
"movq 80(%2), %%mm1 \n\t" // abcdefgh
2203
"movq %%mm0, %%mm2 \n\t" // 12345678
2204
"punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2205
"punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2207
"movq 96(%2), %%mm1 \n\t"
2208
"movq 112(%2), %%mm3 \n\t"
2209
"movq %%mm1, %%mm4 \n\t"
2210
"punpcklbw %%mm3, %%mm1 \n\t"
2211
"punpckhbw %%mm3, %%mm4 \n\t"
2213
"movq %%mm0, %%mm3 \n\t"
2214
"punpcklwd %%mm1, %%mm0 \n\t"
2215
"punpckhwd %%mm1, %%mm3 \n\t"
2216
"movq %%mm2, %%mm1 \n\t"
2217
"punpcklwd %%mm4, %%mm2 \n\t"
2218
"punpckhwd %%mm4, %%mm1 \n\t"
2220
"movd %%mm0, 4(%0) \n\t"
2221
"psrlq $32, %%mm0 \n\t"
2222
"movd %%mm0, 4(%%"REG_a") \n\t"
2223
"movd %%mm3, 4(%%"REG_a", %1) \n\t"
2224
"psrlq $32, %%mm3 \n\t"
2225
"movd %%mm3, 4(%%"REG_a", %1, 2) \n\t"
2226
"movd %%mm2, 4(%0, %1, 4) \n\t"
2227
"psrlq $32, %%mm2 \n\t"
2228
"movd %%mm2, 4(%%"REG_d") \n\t"
2229
"movd %%mm1, 4(%%"REG_d", %1) \n\t"
2230
"psrlq $32, %%mm1 \n\t"
2231
"movd %%mm1, 4(%%"REG_d", %1, 2) \n\t"
2233
:: "r" (dst), "r" ((long)dstStride), "r" (src)
2111
"movq (%2), %%mm0 \n\t" // 12345678
2112
"movq 16(%2), %%mm1 \n\t" // abcdefgh
2113
"movq %%mm0, %%mm2 \n\t" // 12345678
2114
"punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2115
"punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2117
"movq 32(%2), %%mm1 \n\t"
2118
"movq 48(%2), %%mm3 \n\t"
2119
"movq %%mm1, %%mm4 \n\t"
2120
"punpcklbw %%mm3, %%mm1 \n\t"
2121
"punpckhbw %%mm3, %%mm4 \n\t"
2123
"movq %%mm0, %%mm3 \n\t"
2124
"punpcklwd %%mm1, %%mm0 \n\t"
2125
"punpckhwd %%mm1, %%mm3 \n\t"
2126
"movq %%mm2, %%mm1 \n\t"
2127
"punpcklwd %%mm4, %%mm2 \n\t"
2128
"punpckhwd %%mm4, %%mm1 \n\t"
2130
"movd %%mm0, (%0) \n\t"
2131
"psrlq $32, %%mm0 \n\t"
2132
"movd %%mm0, (%%"REG_a") \n\t"
2133
"movd %%mm3, (%%"REG_a", %1) \n\t"
2134
"psrlq $32, %%mm3 \n\t"
2135
"movd %%mm3, (%%"REG_a", %1, 2) \n\t"
2136
"movd %%mm2, (%0, %1, 4) \n\t"
2137
"psrlq $32, %%mm2 \n\t"
2138
"movd %%mm2, (%%"REG_d") \n\t"
2139
"movd %%mm1, (%%"REG_d", %1) \n\t"
2140
"psrlq $32, %%mm1 \n\t"
2141
"movd %%mm1, (%%"REG_d", %1, 2) \n\t"
2144
"movq 64(%2), %%mm0 \n\t" // 12345678
2145
"movq 80(%2), %%mm1 \n\t" // abcdefgh
2146
"movq %%mm0, %%mm2 \n\t" // 12345678
2147
"punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
2148
"punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
2150
"movq 96(%2), %%mm1 \n\t"
2151
"movq 112(%2), %%mm3 \n\t"
2152
"movq %%mm1, %%mm4 \n\t"
2153
"punpcklbw %%mm3, %%mm1 \n\t"
2154
"punpckhbw %%mm3, %%mm4 \n\t"
2156
"movq %%mm0, %%mm3 \n\t"
2157
"punpcklwd %%mm1, %%mm0 \n\t"
2158
"punpckhwd %%mm1, %%mm3 \n\t"
2159
"movq %%mm2, %%mm1 \n\t"
2160
"punpcklwd %%mm4, %%mm2 \n\t"
2161
"punpckhwd %%mm4, %%mm1 \n\t"
2163
"movd %%mm0, 4(%0) \n\t"
2164
"psrlq $32, %%mm0 \n\t"
2165
"movd %%mm0, 4(%%"REG_a") \n\t"
2166
"movd %%mm3, 4(%%"REG_a", %1) \n\t"
2167
"psrlq $32, %%mm3 \n\t"
2168
"movd %%mm3, 4(%%"REG_a", %1, 2) \n\t"
2169
"movd %%mm2, 4(%0, %1, 4) \n\t"
2170
"psrlq $32, %%mm2 \n\t"
2171
"movd %%mm2, 4(%%"REG_d") \n\t"
2172
"movd %%mm1, 4(%%"REG_d", %1) \n\t"
2173
"psrlq $32, %%mm1 \n\t"
2174
"movd %%mm1, 4(%%"REG_d", %1, 2) \n\t"
2176
:: "r" (dst), "r" ((x86_reg)dstStride), "r" (src)
2234
2177
: "%"REG_a, "%"REG_d
2237
2180
#endif //HAVE_MMX
2238
2181
//static long test=0;
2240
#ifndef HAVE_ALTIVEC
2241
2184
static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2242
uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2185
uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise)
2244
// to save a register (FIXME do this outside of the loops)
2245
tempBluredPast[127]= maxNoise[0];
2246
tempBluredPast[128]= maxNoise[1];
2247
tempBluredPast[129]= maxNoise[2];
2187
// to save a register (FIXME do this outside of the loops)
2188
tempBlurredPast[127]= maxNoise[0];
2189
tempBlurredPast[128]= maxNoise[1];
2190
tempBlurredPast[129]= maxNoise[2];
2249
2192
#define FAST_L2_DIFF
2250
2193
//#define L1_DIFF //u should change the thresholds too if u try that one
2251
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2253
"lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride
2254
"lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride
2255
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2194
#if HAVE_MMX2 || HAVE_AMD3DNOW
2196
"lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride
2197
"lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride
2198
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2256
2199
// 0 1 2 3 4 5 6 7 8 9
2257
2200
// %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
2258
2201
//FIXME reorder?
2259
2202
#ifdef L1_DIFF //needs mmx2
2260
"movq (%0), %%mm0 \n\t" // L0
2261
"psadbw (%1), %%mm0 \n\t" // |L0-R0|
2262
"movq (%0, %2), %%mm1 \n\t" // L1
2263
"psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2264
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2265
"psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2266
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2267
"psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3|
2203
"movq (%0), %%mm0 \n\t" // L0
2204
"psadbw (%1), %%mm0 \n\t" // |L0-R0|
2205
"movq (%0, %2), %%mm1 \n\t" // L1
2206
"psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
2207
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2208
"psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
2209
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2210
"psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3|
2269
"movq (%0, %2, 4), %%mm4 \n\t" // L4
2270
"paddw %%mm1, %%mm0 \n\t"
2271
"psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2272
"movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2273
"paddw %%mm2, %%mm0 \n\t"
2274
"psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5|
2275
"movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2276
"paddw %%mm3, %%mm0 \n\t"
2277
"psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6|
2278
"movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2279
"paddw %%mm4, %%mm0 \n\t"
2280
"psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7|
2281
"paddw %%mm5, %%mm6 \n\t"
2282
"paddw %%mm7, %%mm6 \n\t"
2283
"paddw %%mm6, %%mm0 \n\t"
2212
"movq (%0, %2, 4), %%mm4 \n\t" // L4
2213
"paddw %%mm1, %%mm0 \n\t"
2214
"psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
2215
"movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2216
"paddw %%mm2, %%mm0 \n\t"
2217
"psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5|
2218
"movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2219
"paddw %%mm3, %%mm0 \n\t"
2220
"psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6|
2221
"movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2222
"paddw %%mm4, %%mm0 \n\t"
2223
"psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7|
2224
"paddw %%mm5, %%mm6 \n\t"
2225
"paddw %%mm7, %%mm6 \n\t"
2226
"paddw %%mm6, %%mm0 \n\t"
2284
2227
#else //L1_DIFF
2285
2228
#if defined (FAST_L2_DIFF)
2286
"pcmpeqb %%mm7, %%mm7 \n\t"
2287
"movq "MANGLE(b80)", %%mm6 \n\t"
2288
"pxor %%mm0, %%mm0 \n\t"
2229
"pcmpeqb %%mm7, %%mm7 \n\t"
2230
"movq "MANGLE(b80)", %%mm6 \n\t"
2231
"pxor %%mm0, %%mm0 \n\t"
2289
2232
#define REAL_L2_DIFF_CORE(a, b)\
2290
"movq " #a ", %%mm5 \n\t"\
2291
"movq " #b ", %%mm2 \n\t"\
2292
"pxor %%mm7, %%mm2 \n\t"\
2293
PAVGB(%%mm2, %%mm5)\
2294
"paddb %%mm6, %%mm5 \n\t"\
2295
"movq %%mm5, %%mm2 \n\t"\
2296
"psllw $8, %%mm5 \n\t"\
2297
"pmaddwd %%mm5, %%mm5 \n\t"\
2298
"pmaddwd %%mm2, %%mm2 \n\t"\
2299
"paddd %%mm2, %%mm5 \n\t"\
2300
"psrld $14, %%mm5 \n\t"\
2301
"paddd %%mm5, %%mm0 \n\t"
2233
"movq " #a ", %%mm5 \n\t"\
2234
"movq " #b ", %%mm2 \n\t"\
2235
"pxor %%mm7, %%mm2 \n\t"\
2236
PAVGB(%%mm2, %%mm5)\
2237
"paddb %%mm6, %%mm5 \n\t"\
2238
"movq %%mm5, %%mm2 \n\t"\
2239
"psllw $8, %%mm5 \n\t"\
2240
"pmaddwd %%mm5, %%mm5 \n\t"\
2241
"pmaddwd %%mm2, %%mm2 \n\t"\
2242
"paddd %%mm2, %%mm5 \n\t"\
2243
"psrld $14, %%mm5 \n\t"\
2244
"paddd %%mm5, %%mm0 \n\t"
2303
2246
#else //defined (FAST_L2_DIFF)
2304
"pxor %%mm7, %%mm7 \n\t"
2305
"pxor %%mm0, %%mm0 \n\t"
2247
"pxor %%mm7, %%mm7 \n\t"
2248
"pxor %%mm0, %%mm0 \n\t"
2306
2249
#define REAL_L2_DIFF_CORE(a, b)\
2307
"movq " #a ", %%mm5 \n\t"\
2308
"movq " #b ", %%mm2 \n\t"\
2309
"movq %%mm5, %%mm1 \n\t"\
2310
"movq %%mm2, %%mm3 \n\t"\
2311
"punpcklbw %%mm7, %%mm5 \n\t"\
2312
"punpckhbw %%mm7, %%mm1 \n\t"\
2313
"punpcklbw %%mm7, %%mm2 \n\t"\
2314
"punpckhbw %%mm7, %%mm3 \n\t"\
2315
"psubw %%mm2, %%mm5 \n\t"\
2316
"psubw %%mm3, %%mm1 \n\t"\
2317
"pmaddwd %%mm5, %%mm5 \n\t"\
2318
"pmaddwd %%mm1, %%mm1 \n\t"\
2319
"paddd %%mm1, %%mm5 \n\t"\
2320
"paddd %%mm5, %%mm0 \n\t"
2250
"movq " #a ", %%mm5 \n\t"\
2251
"movq " #b ", %%mm2 \n\t"\
2252
"movq %%mm5, %%mm1 \n\t"\
2253
"movq %%mm2, %%mm3 \n\t"\
2254
"punpcklbw %%mm7, %%mm5 \n\t"\
2255
"punpckhbw %%mm7, %%mm1 \n\t"\
2256
"punpcklbw %%mm7, %%mm2 \n\t"\
2257
"punpckhbw %%mm7, %%mm3 \n\t"\
2258
"psubw %%mm2, %%mm5 \n\t"\
2259
"psubw %%mm3, %%mm1 \n\t"\
2260
"pmaddwd %%mm5, %%mm5 \n\t"\
2261
"pmaddwd %%mm1, %%mm1 \n\t"\
2262
"paddd %%mm1, %%mm5 \n\t"\
2263
"paddd %%mm5, %%mm0 \n\t"
2322
2265
#endif //defined (FAST_L2_DIFF)
2335
2278
#endif //L1_DIFF
2337
"movq %%mm0, %%mm4 \n\t"
2338
"psrlq $32, %%mm0 \n\t"
2339
"paddd %%mm0, %%mm4 \n\t"
2340
"movd %%mm4, %%ecx \n\t"
2341
"shll $2, %%ecx \n\t"
2342
"mov %3, %%"REG_d" \n\t"
2343
"addl -4(%%"REG_d"), %%ecx \n\t"
2344
"addl 4(%%"REG_d"), %%ecx \n\t"
2345
"addl -1024(%%"REG_d"), %%ecx \n\t"
2346
"addl $4, %%ecx \n\t"
2347
"addl 1024(%%"REG_d"), %%ecx \n\t"
2348
"shrl $3, %%ecx \n\t"
2349
"movl %%ecx, (%%"REG_d") \n\t"
2351
// "mov %3, %%"REG_c" \n\t"
2352
// "mov %%"REG_c", test \n\t"
2354
"cmpl 512(%%"REG_d"), %%ecx \n\t"
2356
"cmpl 516(%%"REG_d"), %%ecx \n\t"
2359
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2360
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2361
"movq (%0), %%mm0 \n\t" // L0
2362
"movq (%0, %2), %%mm1 \n\t" // L1
2363
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2364
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2365
"movq (%0, %2, 4), %%mm4 \n\t" // L4
2366
"movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2367
"movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2368
"movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2369
"movq %%mm0, (%1) \n\t" // L0
2370
"movq %%mm1, (%1, %2) \n\t" // L1
2371
"movq %%mm2, (%1, %2, 2) \n\t" // L2
2372
"movq %%mm3, (%1, %%"REG_a") \n\t" // L3
2373
"movq %%mm4, (%1, %2, 4) \n\t" // L4
2374
"movq %%mm5, (%1, %%"REG_d") \n\t" // L5
2375
"movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6
2376
"movq %%mm7, (%1, %%"REG_c") \n\t" // L7
2380
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2381
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2382
"movq (%0), %%mm0 \n\t" // L0
2383
PAVGB((%1), %%mm0) // L0
2384
"movq (%0, %2), %%mm1 \n\t" // L1
2385
PAVGB((%1, %2), %%mm1) // L1
2386
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2387
PAVGB((%1, %2, 2), %%mm2) // L2
2388
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2389
PAVGB((%1, %%REGa), %%mm3) // L3
2390
"movq (%0, %2, 4), %%mm4 \n\t" // L4
2391
PAVGB((%1, %2, 4), %%mm4) // L4
2392
"movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2393
PAVGB((%1, %%REGd), %%mm5) // L5
2394
"movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2395
PAVGB((%1, %%REGa, 2), %%mm6) // L6
2396
"movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2397
PAVGB((%1, %%REGc), %%mm7) // L7
2398
"movq %%mm0, (%1) \n\t" // R0
2399
"movq %%mm1, (%1, %2) \n\t" // R1
2400
"movq %%mm2, (%1, %2, 2) \n\t" // R2
2401
"movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2402
"movq %%mm4, (%1, %2, 4) \n\t" // R4
2403
"movq %%mm5, (%1, %%"REG_d") \n\t" // R5
2404
"movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6
2405
"movq %%mm7, (%1, %%"REG_c") \n\t" // R7
2406
"movq %%mm0, (%0) \n\t" // L0
2407
"movq %%mm1, (%0, %2) \n\t" // L1
2408
"movq %%mm2, (%0, %2, 2) \n\t" // L2
2409
"movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2410
"movq %%mm4, (%0, %2, 4) \n\t" // L4
2411
"movq %%mm5, (%0, %%"REG_d") \n\t" // L5
2412
"movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6
2413
"movq %%mm7, (%0, %%"REG_c") \n\t" // L7
2417
"cmpl 508(%%"REG_d"), %%ecx \n\t"
2420
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2421
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2422
"movq (%0), %%mm0 \n\t" // L0
2423
"movq (%0, %2), %%mm1 \n\t" // L1
2424
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2425
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2426
"movq (%1), %%mm4 \n\t" // R0
2427
"movq (%1, %2), %%mm5 \n\t" // R1
2428
"movq (%1, %2, 2), %%mm6 \n\t" // R2
2429
"movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2438
"movq %%mm0, (%1) \n\t" // R0
2439
"movq %%mm1, (%1, %2) \n\t" // R1
2440
"movq %%mm2, (%1, %2, 2) \n\t" // R2
2441
"movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2442
"movq %%mm0, (%0) \n\t" // L0
2443
"movq %%mm1, (%0, %2) \n\t" // L1
2444
"movq %%mm2, (%0, %2, 2) \n\t" // L2
2445
"movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2447
"movq (%0, %2, 4), %%mm0 \n\t" // L4
2448
"movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2449
"movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2450
"movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2451
"movq (%1, %2, 4), %%mm4 \n\t" // R4
2452
"movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2453
"movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2454
"movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2463
"movq %%mm0, (%1, %2, 4) \n\t" // R4
2464
"movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2465
"movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2466
"movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2467
"movq %%mm0, (%0, %2, 4) \n\t" // L4
2468
"movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2469
"movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2470
"movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2474
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2475
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2476
"movq (%0), %%mm0 \n\t" // L0
2477
"movq (%0, %2), %%mm1 \n\t" // L1
2478
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2479
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2480
"movq (%1), %%mm4 \n\t" // R0
2481
"movq (%1, %2), %%mm5 \n\t" // R1
2482
"movq (%1, %2, 2), %%mm6 \n\t" // R2
2483
"movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2496
"movq %%mm0, (%1) \n\t" // R0
2497
"movq %%mm1, (%1, %2) \n\t" // R1
2498
"movq %%mm2, (%1, %2, 2) \n\t" // R2
2499
"movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2500
"movq %%mm0, (%0) \n\t" // L0
2501
"movq %%mm1, (%0, %2) \n\t" // L1
2502
"movq %%mm2, (%0, %2, 2) \n\t" // L2
2503
"movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2505
"movq (%0, %2, 4), %%mm0 \n\t" // L4
2506
"movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2507
"movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2508
"movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2509
"movq (%1, %2, 4), %%mm4 \n\t" // R4
2510
"movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2511
"movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2512
"movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2525
"movq %%mm0, (%1, %2, 4) \n\t" // R4
2526
"movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2527
"movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2528
"movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2529
"movq %%mm0, (%0, %2, 4) \n\t" // L4
2530
"movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2531
"movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2532
"movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2536
:: "r" (src), "r" (tempBlured), "r"((long)stride), "m" (tempBluredPast)
2537
: "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2539
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2280
"movq %%mm0, %%mm4 \n\t"
2281
"psrlq $32, %%mm0 \n\t"
2282
"paddd %%mm0, %%mm4 \n\t"
2283
"movd %%mm4, %%ecx \n\t"
2284
"shll $2, %%ecx \n\t"
2285
"mov %3, %%"REG_d" \n\t"
2286
"addl -4(%%"REG_d"), %%ecx \n\t"
2287
"addl 4(%%"REG_d"), %%ecx \n\t"
2288
"addl -1024(%%"REG_d"), %%ecx \n\t"
2289
"addl $4, %%ecx \n\t"
2290
"addl 1024(%%"REG_d"), %%ecx \n\t"
2291
"shrl $3, %%ecx \n\t"
2292
"movl %%ecx, (%%"REG_d") \n\t"
2294
// "mov %3, %%"REG_c" \n\t"
2295
// "mov %%"REG_c", test \n\t"
2297
"cmpl 512(%%"REG_d"), %%ecx \n\t"
2299
"cmpl 516(%%"REG_d"), %%ecx \n\t"
2302
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2303
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2304
"movq (%0), %%mm0 \n\t" // L0
2305
"movq (%0, %2), %%mm1 \n\t" // L1
2306
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2307
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2308
"movq (%0, %2, 4), %%mm4 \n\t" // L4
2309
"movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2310
"movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2311
"movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2312
"movq %%mm0, (%1) \n\t" // L0
2313
"movq %%mm1, (%1, %2) \n\t" // L1
2314
"movq %%mm2, (%1, %2, 2) \n\t" // L2
2315
"movq %%mm3, (%1, %%"REG_a") \n\t" // L3
2316
"movq %%mm4, (%1, %2, 4) \n\t" // L4
2317
"movq %%mm5, (%1, %%"REG_d") \n\t" // L5
2318
"movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6
2319
"movq %%mm7, (%1, %%"REG_c") \n\t" // L7
2323
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2324
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2325
"movq (%0), %%mm0 \n\t" // L0
2326
PAVGB((%1), %%mm0) // L0
2327
"movq (%0, %2), %%mm1 \n\t" // L1
2328
PAVGB((%1, %2), %%mm1) // L1
2329
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2330
PAVGB((%1, %2, 2), %%mm2) // L2
2331
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2332
PAVGB((%1, %%REGa), %%mm3) // L3
2333
"movq (%0, %2, 4), %%mm4 \n\t" // L4
2334
PAVGB((%1, %2, 4), %%mm4) // L4
2335
"movq (%0, %%"REG_d"), %%mm5 \n\t" // L5
2336
PAVGB((%1, %%REGd), %%mm5) // L5
2337
"movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6
2338
PAVGB((%1, %%REGa, 2), %%mm6) // L6
2339
"movq (%0, %%"REG_c"), %%mm7 \n\t" // L7
2340
PAVGB((%1, %%REGc), %%mm7) // L7
2341
"movq %%mm0, (%1) \n\t" // R0
2342
"movq %%mm1, (%1, %2) \n\t" // R1
2343
"movq %%mm2, (%1, %2, 2) \n\t" // R2
2344
"movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2345
"movq %%mm4, (%1, %2, 4) \n\t" // R4
2346
"movq %%mm5, (%1, %%"REG_d") \n\t" // R5
2347
"movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6
2348
"movq %%mm7, (%1, %%"REG_c") \n\t" // R7
2349
"movq %%mm0, (%0) \n\t" // L0
2350
"movq %%mm1, (%0, %2) \n\t" // L1
2351
"movq %%mm2, (%0, %2, 2) \n\t" // L2
2352
"movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2353
"movq %%mm4, (%0, %2, 4) \n\t" // L4
2354
"movq %%mm5, (%0, %%"REG_d") \n\t" // L5
2355
"movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6
2356
"movq %%mm7, (%0, %%"REG_c") \n\t" // L7
2360
"cmpl 508(%%"REG_d"), %%ecx \n\t"
2363
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2364
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2365
"movq (%0), %%mm0 \n\t" // L0
2366
"movq (%0, %2), %%mm1 \n\t" // L1
2367
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2368
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2369
"movq (%1), %%mm4 \n\t" // R0
2370
"movq (%1, %2), %%mm5 \n\t" // R1
2371
"movq (%1, %2, 2), %%mm6 \n\t" // R2
2372
"movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2381
"movq %%mm0, (%1) \n\t" // R0
2382
"movq %%mm1, (%1, %2) \n\t" // R1
2383
"movq %%mm2, (%1, %2, 2) \n\t" // R2
2384
"movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2385
"movq %%mm0, (%0) \n\t" // L0
2386
"movq %%mm1, (%0, %2) \n\t" // L1
2387
"movq %%mm2, (%0, %2, 2) \n\t" // L2
2388
"movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2390
"movq (%0, %2, 4), %%mm0 \n\t" // L4
2391
"movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2392
"movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2393
"movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2394
"movq (%1, %2, 4), %%mm4 \n\t" // R4
2395
"movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2396
"movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2397
"movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2406
"movq %%mm0, (%1, %2, 4) \n\t" // R4
2407
"movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2408
"movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2409
"movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2410
"movq %%mm0, (%0, %2, 4) \n\t" // L4
2411
"movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2412
"movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2413
"movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2417
"lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride
2418
"lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride
2419
"movq (%0), %%mm0 \n\t" // L0
2420
"movq (%0, %2), %%mm1 \n\t" // L1
2421
"movq (%0, %2, 2), %%mm2 \n\t" // L2
2422
"movq (%0, %%"REG_a"), %%mm3 \n\t" // L3
2423
"movq (%1), %%mm4 \n\t" // R0
2424
"movq (%1, %2), %%mm5 \n\t" // R1
2425
"movq (%1, %2, 2), %%mm6 \n\t" // R2
2426
"movq (%1, %%"REG_a"), %%mm7 \n\t" // R3
2439
"movq %%mm0, (%1) \n\t" // R0
2440
"movq %%mm1, (%1, %2) \n\t" // R1
2441
"movq %%mm2, (%1, %2, 2) \n\t" // R2
2442
"movq %%mm3, (%1, %%"REG_a") \n\t" // R3
2443
"movq %%mm0, (%0) \n\t" // L0
2444
"movq %%mm1, (%0, %2) \n\t" // L1
2445
"movq %%mm2, (%0, %2, 2) \n\t" // L2
2446
"movq %%mm3, (%0, %%"REG_a") \n\t" // L3
2448
"movq (%0, %2, 4), %%mm0 \n\t" // L4
2449
"movq (%0, %%"REG_d"), %%mm1 \n\t" // L5
2450
"movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6
2451
"movq (%0, %%"REG_c"), %%mm3 \n\t" // L7
2452
"movq (%1, %2, 4), %%mm4 \n\t" // R4
2453
"movq (%1, %%"REG_d"), %%mm5 \n\t" // R5
2454
"movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6
2455
"movq (%1, %%"REG_c"), %%mm7 \n\t" // R7
2468
"movq %%mm0, (%1, %2, 4) \n\t" // R4
2469
"movq %%mm1, (%1, %%"REG_d") \n\t" // R5
2470
"movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6
2471
"movq %%mm3, (%1, %%"REG_c") \n\t" // R7
2472
"movq %%mm0, (%0, %2, 4) \n\t" // L4
2473
"movq %%mm1, (%0, %%"REG_d") \n\t" // L5
2474
"movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6
2475
"movq %%mm3, (%0, %%"REG_c") \n\t" // L7
2479
:: "r" (src), "r" (tempBlurred), "r"((x86_reg)stride), "m" (tempBlurredPast)
2480
: "%"REG_a, "%"REG_d, "%"REG_c, "memory"
2482
#else //HAVE_MMX2 || HAVE_AMD3DNOW
2551
int ref= tempBlured[ x + y*stride ];
2552
int cur= src[ x + y*stride ];
2554
// if(x==0 || x==7) d1+= d1>>1;
2555
// if(y==0 || y==7) d1+= d1>>1;
2492
int ref= tempBlurred[ x + y*stride ];
2493
int cur= src[ x + y*stride ];
2495
// if(x==0 || x==7) d1+= d1>>1;
2496
// if(y==0 || y==7) d1+= d1>>1;
2564
+(*(tempBluredPast-256))
2565
+(*(tempBluredPast-1))+ (*(tempBluredPast+1))
2566
+(*(tempBluredPast+256))
2569
// ((*tempBluredPast)*3 + d + 2)>>2;
2505
+(*(tempBlurredPast-256))
2506
+(*(tempBlurredPast-1))+ (*(tempBlurredPast+1))
2507
+(*(tempBlurredPast+256))
2510
// ((*tempBlurredPast)*3 + d + 2)>>2;
2575
2516
64 48 36 27 20 15 11 (33) (approx)
2576
2517
64 56 49 43 37 33 29 (200) (approx)
2587
int ref= tempBlured[ x + y*stride ];
2588
int cur= src[ x + y*stride ];
2589
tempBlured[ x + y*stride ]=
2590
src[ x + y*stride ]=
2602
tempBlured[ x + y*stride ]= src[ x + y*stride ];
2616
int ref= tempBlured[ x + y*stride ];
2617
int cur= src[ x + y*stride ];
2618
tempBlured[ x + y*stride ]=
2619
src[ x + y*stride ]=
2620
(ref*7 + cur + 4)>>3;
2631
int ref= tempBlured[ x + y*stride ];
2632
int cur= src[ x + y*stride ];
2633
tempBlured[ x + y*stride ]=
2634
src[ x + y*stride ]=
2635
(ref*3 + cur + 2)>>2;
2519
if(d > maxNoise[1]){
2520
if(d < maxNoise[2]){
2524
int ref= tempBlurred[ x + y*stride ];
2525
int cur= src[ x + y*stride ];
2526
tempBlurred[ x + y*stride ]=
2527
src[ x + y*stride ]=
2535
tempBlurred[ x + y*stride ]= src[ x + y*stride ];
2540
if(d < maxNoise[0]){
2544
int ref= tempBlurred[ x + y*stride ];
2545
int cur= src[ x + y*stride ];
2546
tempBlurred[ x + y*stride ]=
2547
src[ x + y*stride ]=
2548
(ref*7 + cur + 4)>>3;
2555
int ref= tempBlurred[ x + y*stride ];
2556
int cur= src[ x + y*stride ];
2557
tempBlurred[ x + y*stride ]=
2558
src[ x + y*stride ]=
2559
(ref*3 + cur + 2)>>2;
2641
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2565
#endif //HAVE_MMX2 || HAVE_AMD3DNOW
2643
2567
#endif //HAVE_ALTIVEC
2647
2571
* accurate deblock filter
2649
2573
static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){
2650
int64_t dc_mask, eq_mask, both_masks;
2651
int64_t sums[10*8*2];
2652
src+= step*3; // src points to begin of the 8x8 Block
2574
int64_t dc_mask, eq_mask, both_masks;
2575
int64_t sums[10*8*2];
2576
src+= step*3; // src points to begin of the 8x8 Block
2655
"movq %0, %%mm7 \n\t"
2656
"movq %1, %%mm6 \n\t"
2657
: : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
2579
"movq %0, %%mm7 \n\t"
2580
"movq %1, %%mm6 \n\t"
2581
: : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
2661
"lea (%2, %3), %%"REG_a" \n\t"
2585
"lea (%2, %3), %%"REG_a" \n\t"
2662
2586
// 0 1 2 3 4 5 6 7 8 9
2663
2587
// %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
2665
"movq (%2), %%mm0 \n\t"
2666
"movq (%%"REG_a"), %%mm1 \n\t"
2667
"movq %%mm1, %%mm3 \n\t"
2668
"movq %%mm1, %%mm4 \n\t"
2669
"psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
2670
"paddb %%mm7, %%mm0 \n\t"
2671
"pcmpgtb %%mm6, %%mm0 \n\t"
2673
"movq (%%"REG_a",%3), %%mm2 \n\t"
2674
PMAXUB(%%mm2, %%mm4)
2675
PMINUB(%%mm2, %%mm3, %%mm5)
2676
"psubb %%mm2, %%mm1 \n\t"
2677
"paddb %%mm7, %%mm1 \n\t"
2678
"pcmpgtb %%mm6, %%mm1 \n\t"
2679
"paddb %%mm1, %%mm0 \n\t"
2681
"movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2682
PMAXUB(%%mm1, %%mm4)
2683
PMINUB(%%mm1, %%mm3, %%mm5)
2684
"psubb %%mm1, %%mm2 \n\t"
2685
"paddb %%mm7, %%mm2 \n\t"
2686
"pcmpgtb %%mm6, %%mm2 \n\t"
2687
"paddb %%mm2, %%mm0 \n\t"
2689
"lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
2691
"movq (%2, %3, 4), %%mm2 \n\t"
2692
PMAXUB(%%mm2, %%mm4)
2693
PMINUB(%%mm2, %%mm3, %%mm5)
2694
"psubb %%mm2, %%mm1 \n\t"
2695
"paddb %%mm7, %%mm1 \n\t"
2696
"pcmpgtb %%mm6, %%mm1 \n\t"
2697
"paddb %%mm1, %%mm0 \n\t"
2699
"movq (%%"REG_a"), %%mm1 \n\t"
2700
PMAXUB(%%mm1, %%mm4)
2701
PMINUB(%%mm1, %%mm3, %%mm5)
2702
"psubb %%mm1, %%mm2 \n\t"
2703
"paddb %%mm7, %%mm2 \n\t"
2704
"pcmpgtb %%mm6, %%mm2 \n\t"
2705
"paddb %%mm2, %%mm0 \n\t"
2707
"movq (%%"REG_a", %3), %%mm2 \n\t"
2708
PMAXUB(%%mm2, %%mm4)
2709
PMINUB(%%mm2, %%mm3, %%mm5)
2710
"psubb %%mm2, %%mm1 \n\t"
2711
"paddb %%mm7, %%mm1 \n\t"
2712
"pcmpgtb %%mm6, %%mm1 \n\t"
2713
"paddb %%mm1, %%mm0 \n\t"
2715
"movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2716
PMAXUB(%%mm1, %%mm4)
2717
PMINUB(%%mm1, %%mm3, %%mm5)
2718
"psubb %%mm1, %%mm2 \n\t"
2719
"paddb %%mm7, %%mm2 \n\t"
2720
"pcmpgtb %%mm6, %%mm2 \n\t"
2721
"paddb %%mm2, %%mm0 \n\t"
2723
"movq (%2, %3, 8), %%mm2 \n\t"
2724
PMAXUB(%%mm2, %%mm4)
2725
PMINUB(%%mm2, %%mm3, %%mm5)
2726
"psubb %%mm2, %%mm1 \n\t"
2727
"paddb %%mm7, %%mm1 \n\t"
2728
"pcmpgtb %%mm6, %%mm1 \n\t"
2729
"paddb %%mm1, %%mm0 \n\t"
2731
"movq (%%"REG_a", %3, 4), %%mm1 \n\t"
2732
"psubb %%mm1, %%mm2 \n\t"
2733
"paddb %%mm7, %%mm2 \n\t"
2734
"pcmpgtb %%mm6, %%mm2 \n\t"
2735
"paddb %%mm2, %%mm0 \n\t"
2736
"psubusb %%mm3, %%mm4 \n\t"
2738
"pxor %%mm6, %%mm6 \n\t"
2739
"movq %4, %%mm7 \n\t" // QP,..., QP
2740
"paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
2741
"psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0
2742
"pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2743
"pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2744
"movq %%mm7, %1 \n\t"
2746
"movq %5, %%mm7 \n\t"
2747
"punpcklbw %%mm7, %%mm7 \n\t"
2748
"punpcklbw %%mm7, %%mm7 \n\t"
2749
"punpcklbw %%mm7, %%mm7 \n\t"
2750
"psubb %%mm0, %%mm6 \n\t"
2751
"pcmpgtb %%mm7, %%mm6 \n\t"
2752
"movq %%mm6, %0 \n\t"
2754
: "=m" (eq_mask), "=m" (dc_mask)
2755
: "r" (src), "r" ((long)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
2759
both_masks = dc_mask & eq_mask;
2762
long offset= -8*step;
2763
int64_t *temp_sums= sums;
2766
"movq %2, %%mm0 \n\t" // QP,..., QP
2767
"pxor %%mm4, %%mm4 \n\t"
2769
"movq (%0), %%mm6 \n\t"
2770
"movq (%0, %1), %%mm5 \n\t"
2771
"movq %%mm5, %%mm1 \n\t"
2772
"movq %%mm6, %%mm2 \n\t"
2773
"psubusb %%mm6, %%mm5 \n\t"
2774
"psubusb %%mm1, %%mm2 \n\t"
2775
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2776
"psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2777
"pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2779
"pxor %%mm6, %%mm1 \n\t"
2780
"pand %%mm0, %%mm1 \n\t"
2781
"pxor %%mm1, %%mm6 \n\t"
2784
"movq (%0, %1, 8), %%mm5 \n\t"
2785
"add %1, %0 \n\t" // %0 points to line 1 not 0
2786
"movq (%0, %1, 8), %%mm7 \n\t"
2787
"movq %%mm5, %%mm1 \n\t"
2788
"movq %%mm7, %%mm2 \n\t"
2789
"psubusb %%mm7, %%mm5 \n\t"
2790
"psubusb %%mm1, %%mm2 \n\t"
2791
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2792
"movq %2, %%mm0 \n\t" // QP,..., QP
2793
"psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2794
"pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2796
"pxor %%mm7, %%mm1 \n\t"
2797
"pand %%mm0, %%mm1 \n\t"
2798
"pxor %%mm1, %%mm7 \n\t"
2800
"movq %%mm6, %%mm5 \n\t"
2801
"punpckhbw %%mm4, %%mm6 \n\t"
2802
"punpcklbw %%mm4, %%mm5 \n\t"
2803
// 4:0 5/6:First 7:Last
2805
"movq %%mm5, %%mm0 \n\t"
2806
"movq %%mm6, %%mm1 \n\t"
2807
"psllw $2, %%mm0 \n\t"
2808
"psllw $2, %%mm1 \n\t"
2809
"paddw "MANGLE(w04)", %%mm0 \n\t"
2810
"paddw "MANGLE(w04)", %%mm1 \n\t"
2589
"movq (%2), %%mm0 \n\t"
2590
"movq (%%"REG_a"), %%mm1 \n\t"
2591
"movq %%mm1, %%mm3 \n\t"
2592
"movq %%mm1, %%mm4 \n\t"
2593
"psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
2594
"paddb %%mm7, %%mm0 \n\t"
2595
"pcmpgtb %%mm6, %%mm0 \n\t"
2597
"movq (%%"REG_a",%3), %%mm2 \n\t"
2598
PMAXUB(%%mm2, %%mm4)
2599
PMINUB(%%mm2, %%mm3, %%mm5)
2600
"psubb %%mm2, %%mm1 \n\t"
2601
"paddb %%mm7, %%mm1 \n\t"
2602
"pcmpgtb %%mm6, %%mm1 \n\t"
2603
"paddb %%mm1, %%mm0 \n\t"
2605
"movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2606
PMAXUB(%%mm1, %%mm4)
2607
PMINUB(%%mm1, %%mm3, %%mm5)
2608
"psubb %%mm1, %%mm2 \n\t"
2609
"paddb %%mm7, %%mm2 \n\t"
2610
"pcmpgtb %%mm6, %%mm2 \n\t"
2611
"paddb %%mm2, %%mm0 \n\t"
2613
"lea (%%"REG_a", %3, 4), %%"REG_a" \n\t"
2615
"movq (%2, %3, 4), %%mm2 \n\t"
2616
PMAXUB(%%mm2, %%mm4)
2617
PMINUB(%%mm2, %%mm3, %%mm5)
2618
"psubb %%mm2, %%mm1 \n\t"
2619
"paddb %%mm7, %%mm1 \n\t"
2620
"pcmpgtb %%mm6, %%mm1 \n\t"
2621
"paddb %%mm1, %%mm0 \n\t"
2623
"movq (%%"REG_a"), %%mm1 \n\t"
2624
PMAXUB(%%mm1, %%mm4)
2625
PMINUB(%%mm1, %%mm3, %%mm5)
2626
"psubb %%mm1, %%mm2 \n\t"
2627
"paddb %%mm7, %%mm2 \n\t"
2628
"pcmpgtb %%mm6, %%mm2 \n\t"
2629
"paddb %%mm2, %%mm0 \n\t"
2631
"movq (%%"REG_a", %3), %%mm2 \n\t"
2632
PMAXUB(%%mm2, %%mm4)
2633
PMINUB(%%mm2, %%mm3, %%mm5)
2634
"psubb %%mm2, %%mm1 \n\t"
2635
"paddb %%mm7, %%mm1 \n\t"
2636
"pcmpgtb %%mm6, %%mm1 \n\t"
2637
"paddb %%mm1, %%mm0 \n\t"
2639
"movq (%%"REG_a", %3, 2), %%mm1 \n\t"
2640
PMAXUB(%%mm1, %%mm4)
2641
PMINUB(%%mm1, %%mm3, %%mm5)
2642
"psubb %%mm1, %%mm2 \n\t"
2643
"paddb %%mm7, %%mm2 \n\t"
2644
"pcmpgtb %%mm6, %%mm2 \n\t"
2645
"paddb %%mm2, %%mm0 \n\t"
2647
"movq (%2, %3, 8), %%mm2 \n\t"
2648
PMAXUB(%%mm2, %%mm4)
2649
PMINUB(%%mm2, %%mm3, %%mm5)
2650
"psubb %%mm2, %%mm1 \n\t"
2651
"paddb %%mm7, %%mm1 \n\t"
2652
"pcmpgtb %%mm6, %%mm1 \n\t"
2653
"paddb %%mm1, %%mm0 \n\t"
2655
"movq (%%"REG_a", %3, 4), %%mm1 \n\t"
2656
"psubb %%mm1, %%mm2 \n\t"
2657
"paddb %%mm7, %%mm2 \n\t"
2658
"pcmpgtb %%mm6, %%mm2 \n\t"
2659
"paddb %%mm2, %%mm0 \n\t"
2660
"psubusb %%mm3, %%mm4 \n\t"
2662
"pxor %%mm6, %%mm6 \n\t"
2663
"movq %4, %%mm7 \n\t" // QP,..., QP
2664
"paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
2665
"psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0
2666
"pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2667
"pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0
2668
"movq %%mm7, %1 \n\t"
2670
"movq %5, %%mm7 \n\t"
2671
"punpcklbw %%mm7, %%mm7 \n\t"
2672
"punpcklbw %%mm7, %%mm7 \n\t"
2673
"punpcklbw %%mm7, %%mm7 \n\t"
2674
"psubb %%mm0, %%mm6 \n\t"
2675
"pcmpgtb %%mm7, %%mm6 \n\t"
2676
"movq %%mm6, %0 \n\t"
2678
: "=m" (eq_mask), "=m" (dc_mask)
2679
: "r" (src), "r" ((x86_reg)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold)
2683
both_masks = dc_mask & eq_mask;
2686
x86_reg offset= -8*step;
2687
int64_t *temp_sums= sums;
2690
"movq %2, %%mm0 \n\t" // QP,..., QP
2691
"pxor %%mm4, %%mm4 \n\t"
2693
"movq (%0), %%mm6 \n\t"
2694
"movq (%0, %1), %%mm5 \n\t"
2695
"movq %%mm5, %%mm1 \n\t"
2696
"movq %%mm6, %%mm2 \n\t"
2697
"psubusb %%mm6, %%mm5 \n\t"
2698
"psubusb %%mm1, %%mm2 \n\t"
2699
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2700
"psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2701
"pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2703
"pxor %%mm6, %%mm1 \n\t"
2704
"pand %%mm0, %%mm1 \n\t"
2705
"pxor %%mm1, %%mm6 \n\t"
2708
"movq (%0, %1, 8), %%mm5 \n\t"
2709
"add %1, %0 \n\t" // %0 points to line 1 not 0
2710
"movq (%0, %1, 8), %%mm7 \n\t"
2711
"movq %%mm5, %%mm1 \n\t"
2712
"movq %%mm7, %%mm2 \n\t"
2713
"psubusb %%mm7, %%mm5 \n\t"
2714
"psubusb %%mm1, %%mm2 \n\t"
2715
"por %%mm5, %%mm2 \n\t" // ABS Diff of lines
2716
"movq %2, %%mm0 \n\t" // QP,..., QP
2717
"psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0
2718
"pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF
2720
"pxor %%mm7, %%mm1 \n\t"
2721
"pand %%mm0, %%mm1 \n\t"
2722
"pxor %%mm1, %%mm7 \n\t"
2724
"movq %%mm6, %%mm5 \n\t"
2725
"punpckhbw %%mm4, %%mm6 \n\t"
2726
"punpcklbw %%mm4, %%mm5 \n\t"
2727
// 4:0 5/6:First 7:Last
2729
"movq %%mm5, %%mm0 \n\t"
2730
"movq %%mm6, %%mm1 \n\t"
2731
"psllw $2, %%mm0 \n\t"
2732
"psllw $2, %%mm1 \n\t"
2733
"paddw "MANGLE(w04)", %%mm0 \n\t"
2734
"paddw "MANGLE(w04)", %%mm1 \n\t"
2813
"movq (%0), %%mm2 \n\t"\
2814
"movq (%0), %%mm3 \n\t"\
2816
"punpcklbw %%mm4, %%mm2 \n\t"\
2817
"punpckhbw %%mm4, %%mm3 \n\t"\
2818
"paddw %%mm2, %%mm0 \n\t"\
2819
"paddw %%mm3, %%mm1 \n\t"
2737
"movq (%0), %%mm2 \n\t"\
2738
"movq (%0), %%mm3 \n\t"\
2740
"punpcklbw %%mm4, %%mm2 \n\t"\
2741
"punpckhbw %%mm4, %%mm3 \n\t"\
2742
"paddw %%mm2, %%mm0 \n\t"\
2743
"paddw %%mm3, %%mm1 \n\t"
2822
"movq (%0), %%mm2 \n\t"\
2823
"movq (%0), %%mm3 \n\t"\
2825
"punpcklbw %%mm4, %%mm2 \n\t"\
2826
"punpckhbw %%mm4, %%mm3 \n\t"\
2827
"psubw %%mm2, %%mm0 \n\t"\
2828
"psubw %%mm3, %%mm1 \n\t"
2834
"movq %%mm0, (%3) \n\t"
2835
"movq %%mm1, 8(%3) \n\t"
2838
"psubw %%mm5, %%mm0 \n\t"
2839
"psubw %%mm6, %%mm1 \n\t"
2840
"movq %%mm0, 16(%3) \n\t"
2841
"movq %%mm1, 24(%3) \n\t"
2844
"psubw %%mm5, %%mm0 \n\t"
2845
"psubw %%mm6, %%mm1 \n\t"
2846
"movq %%mm0, 32(%3) \n\t"
2847
"movq %%mm1, 40(%3) \n\t"
2850
"psubw %%mm5, %%mm0 \n\t"
2851
"psubw %%mm6, %%mm1 \n\t"
2852
"movq %%mm0, 48(%3) \n\t"
2853
"movq %%mm1, 56(%3) \n\t"
2856
"psubw %%mm5, %%mm0 \n\t"
2857
"psubw %%mm6, %%mm1 \n\t"
2858
"movq %%mm0, 64(%3) \n\t"
2859
"movq %%mm1, 72(%3) \n\t"
2861
"movq %%mm7, %%mm6 \n\t"
2862
"punpckhbw %%mm4, %%mm7 \n\t"
2863
"punpcklbw %%mm4, %%mm6 \n\t"
2869
"movq %%mm0, 80(%3) \n\t"
2870
"movq %%mm1, 88(%3) \n\t"
2873
"paddw %%mm6, %%mm0 \n\t"
2874
"paddw %%mm7, %%mm1 \n\t"
2875
"movq %%mm0, 96(%3) \n\t"
2876
"movq %%mm1, 104(%3) \n\t"
2879
"paddw %%mm6, %%mm0 \n\t"
2880
"paddw %%mm7, %%mm1 \n\t"
2881
"movq %%mm0, 112(%3) \n\t"
2882
"movq %%mm1, 120(%3) \n\t"
2885
"paddw %%mm6, %%mm0 \n\t"
2886
"paddw %%mm7, %%mm1 \n\t"
2887
"movq %%mm0, 128(%3) \n\t"
2888
"movq %%mm1, 136(%3) \n\t"
2891
"paddw %%mm6, %%mm0 \n\t"
2892
"paddw %%mm7, %%mm1 \n\t"
2893
"movq %%mm0, 144(%3) \n\t"
2894
"movq %%mm1, 152(%3) \n\t"
2896
"mov %4, %0 \n\t" //FIXME
2899
: "r" ((long)step), "m" (c->pQPb), "r"(sums), "g"(src)
2902
src+= step; // src points to begin of the 8x8 Block
2905
"movq %4, %%mm6 \n\t"
2906
"pcmpeqb %%mm5, %%mm5 \n\t"
2907
"pxor %%mm6, %%mm5 \n\t"
2908
"pxor %%mm7, %%mm7 \n\t"
2911
"movq (%1), %%mm0 \n\t"
2912
"movq 8(%1), %%mm1 \n\t"
2913
"paddw 32(%1), %%mm0 \n\t"
2914
"paddw 40(%1), %%mm1 \n\t"
2915
"movq (%0, %3), %%mm2 \n\t"
2916
"movq %%mm2, %%mm3 \n\t"
2917
"movq %%mm2, %%mm4 \n\t"
2918
"punpcklbw %%mm7, %%mm2 \n\t"
2919
"punpckhbw %%mm7, %%mm3 \n\t"
2920
"paddw %%mm2, %%mm0 \n\t"
2921
"paddw %%mm3, %%mm1 \n\t"
2922
"paddw %%mm2, %%mm0 \n\t"
2923
"paddw %%mm3, %%mm1 \n\t"
2924
"psrlw $4, %%mm0 \n\t"
2925
"psrlw $4, %%mm1 \n\t"
2926
"packuswb %%mm1, %%mm0 \n\t"
2927
"pand %%mm6, %%mm0 \n\t"
2928
"pand %%mm5, %%mm4 \n\t"
2929
"por %%mm4, %%mm0 \n\t"
2930
"movq %%mm0, (%0, %3) \n\t"
2935
: "+r"(offset), "+r"(temp_sums)
2936
: "r" ((long)step), "r"(src - offset), "m"(both_masks)
2939
src+= step; // src points to begin of the 8x8 Block
2941
if(eq_mask != -1LL){
2942
uint8_t *temp_src= src;
2944
"pxor %%mm7, %%mm7 \n\t"
2945
"lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
2946
"and "ALIGN_MASK", %%"REG_c" \n\t" // align
2746
"movq (%0), %%mm2 \n\t"\
2747
"movq (%0), %%mm3 \n\t"\
2749
"punpcklbw %%mm4, %%mm2 \n\t"\
2750
"punpckhbw %%mm4, %%mm3 \n\t"\
2751
"psubw %%mm2, %%mm0 \n\t"\
2752
"psubw %%mm3, %%mm1 \n\t"
2758
"movq %%mm0, (%3) \n\t"
2759
"movq %%mm1, 8(%3) \n\t"
2762
"psubw %%mm5, %%mm0 \n\t"
2763
"psubw %%mm6, %%mm1 \n\t"
2764
"movq %%mm0, 16(%3) \n\t"
2765
"movq %%mm1, 24(%3) \n\t"
2768
"psubw %%mm5, %%mm0 \n\t"
2769
"psubw %%mm6, %%mm1 \n\t"
2770
"movq %%mm0, 32(%3) \n\t"
2771
"movq %%mm1, 40(%3) \n\t"
2774
"psubw %%mm5, %%mm0 \n\t"
2775
"psubw %%mm6, %%mm1 \n\t"
2776
"movq %%mm0, 48(%3) \n\t"
2777
"movq %%mm1, 56(%3) \n\t"
2780
"psubw %%mm5, %%mm0 \n\t"
2781
"psubw %%mm6, %%mm1 \n\t"
2782
"movq %%mm0, 64(%3) \n\t"
2783
"movq %%mm1, 72(%3) \n\t"
2785
"movq %%mm7, %%mm6 \n\t"
2786
"punpckhbw %%mm4, %%mm7 \n\t"
2787
"punpcklbw %%mm4, %%mm6 \n\t"
2793
"movq %%mm0, 80(%3) \n\t"
2794
"movq %%mm1, 88(%3) \n\t"
2797
"paddw %%mm6, %%mm0 \n\t"
2798
"paddw %%mm7, %%mm1 \n\t"
2799
"movq %%mm0, 96(%3) \n\t"
2800
"movq %%mm1, 104(%3) \n\t"
2803
"paddw %%mm6, %%mm0 \n\t"
2804
"paddw %%mm7, %%mm1 \n\t"
2805
"movq %%mm0, 112(%3) \n\t"
2806
"movq %%mm1, 120(%3) \n\t"
2809
"paddw %%mm6, %%mm0 \n\t"
2810
"paddw %%mm7, %%mm1 \n\t"
2811
"movq %%mm0, 128(%3) \n\t"
2812
"movq %%mm1, 136(%3) \n\t"
2815
"paddw %%mm6, %%mm0 \n\t"
2816
"paddw %%mm7, %%mm1 \n\t"
2817
"movq %%mm0, 144(%3) \n\t"
2818
"movq %%mm1, 152(%3) \n\t"
2820
"mov %4, %0 \n\t" //FIXME
2823
: "r" ((x86_reg)step), "m" (c->pQPb), "r"(sums), "g"(src)
2826
src+= step; // src points to begin of the 8x8 Block
2829
"movq %4, %%mm6 \n\t"
2830
"pcmpeqb %%mm5, %%mm5 \n\t"
2831
"pxor %%mm6, %%mm5 \n\t"
2832
"pxor %%mm7, %%mm7 \n\t"
2835
"movq (%1), %%mm0 \n\t"
2836
"movq 8(%1), %%mm1 \n\t"
2837
"paddw 32(%1), %%mm0 \n\t"
2838
"paddw 40(%1), %%mm1 \n\t"
2839
"movq (%0, %3), %%mm2 \n\t"
2840
"movq %%mm2, %%mm3 \n\t"
2841
"movq %%mm2, %%mm4 \n\t"
2842
"punpcklbw %%mm7, %%mm2 \n\t"
2843
"punpckhbw %%mm7, %%mm3 \n\t"
2844
"paddw %%mm2, %%mm0 \n\t"
2845
"paddw %%mm3, %%mm1 \n\t"
2846
"paddw %%mm2, %%mm0 \n\t"
2847
"paddw %%mm3, %%mm1 \n\t"
2848
"psrlw $4, %%mm0 \n\t"
2849
"psrlw $4, %%mm1 \n\t"
2850
"packuswb %%mm1, %%mm0 \n\t"
2851
"pand %%mm6, %%mm0 \n\t"
2852
"pand %%mm5, %%mm4 \n\t"
2853
"por %%mm4, %%mm0 \n\t"
2854
"movq %%mm0, (%0, %3) \n\t"
2859
: "+r"(offset), "+r"(temp_sums)
2860
: "r" ((x86_reg)step), "r"(src - offset), "m"(both_masks)
2863
src+= step; // src points to begin of the 8x8 Block
2865
if(eq_mask != -1LL){
2866
uint8_t *temp_src= src;
2868
"pxor %%mm7, %%mm7 \n\t"
2869
"lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars
2870
"and "ALIGN_MASK", %%"REG_c" \n\t" // align
2947
2871
// 0 1 2 3 4 5 6 7 8 9
2948
2872
// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1
2950
"movq (%0), %%mm0 \n\t"
2951
"movq %%mm0, %%mm1 \n\t"
2952
"punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
2953
"punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
2955
"movq (%0, %1), %%mm2 \n\t"
2956
"lea (%0, %1, 2), %%"REG_a" \n\t"
2957
"movq %%mm2, %%mm3 \n\t"
2958
"punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
2959
"punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
2961
"movq (%%"REG_a"), %%mm4 \n\t"
2962
"movq %%mm4, %%mm5 \n\t"
2963
"punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
2964
"punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
2966
"paddw %%mm0, %%mm0 \n\t" // 2L0
2967
"paddw %%mm1, %%mm1 \n\t" // 2H0
2968
"psubw %%mm4, %%mm2 \n\t" // L1 - L2
2969
"psubw %%mm5, %%mm3 \n\t" // H1 - H2
2970
"psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
2971
"psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
2973
"psllw $2, %%mm2 \n\t" // 4L1 - 4L2
2974
"psllw $2, %%mm3 \n\t" // 4H1 - 4H2
2975
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
2976
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
2978
"movq (%%"REG_a", %1), %%mm2 \n\t"
2979
"movq %%mm2, %%mm3 \n\t"
2980
"punpcklbw %%mm7, %%mm2 \n\t" // L3
2981
"punpckhbw %%mm7, %%mm3 \n\t" // H3
2983
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
2984
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
2985
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2986
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2987
"movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2988
"movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2990
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
2991
"movq %%mm0, %%mm1 \n\t"
2992
"punpcklbw %%mm7, %%mm0 \n\t" // L4
2993
"punpckhbw %%mm7, %%mm1 \n\t" // H4
2995
"psubw %%mm0, %%mm2 \n\t" // L3 - L4
2996
"psubw %%mm1, %%mm3 \n\t" // H3 - H4
2997
"movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
2998
"movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
2999
"paddw %%mm4, %%mm4 \n\t" // 2L2
3000
"paddw %%mm5, %%mm5 \n\t" // 2H2
3001
"psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
3002
"psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
3004
"lea (%%"REG_a", %1), %0 \n\t"
3005
"psllw $2, %%mm2 \n\t" // 4L3 - 4L4
3006
"psllw $2, %%mm3 \n\t" // 4H3 - 4H4
3007
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
3008
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
2874
"movq (%0), %%mm0 \n\t"
2875
"movq %%mm0, %%mm1 \n\t"
2876
"punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
2877
"punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
2879
"movq (%0, %1), %%mm2 \n\t"
2880
"lea (%0, %1, 2), %%"REG_a" \n\t"
2881
"movq %%mm2, %%mm3 \n\t"
2882
"punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
2883
"punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
2885
"movq (%%"REG_a"), %%mm4 \n\t"
2886
"movq %%mm4, %%mm5 \n\t"
2887
"punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
2888
"punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
2890
"paddw %%mm0, %%mm0 \n\t" // 2L0
2891
"paddw %%mm1, %%mm1 \n\t" // 2H0
2892
"psubw %%mm4, %%mm2 \n\t" // L1 - L2
2893
"psubw %%mm5, %%mm3 \n\t" // H1 - H2
2894
"psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
2895
"psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
2897
"psllw $2, %%mm2 \n\t" // 4L1 - 4L2
2898
"psllw $2, %%mm3 \n\t" // 4H1 - 4H2
2899
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
2900
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
2902
"movq (%%"REG_a", %1), %%mm2 \n\t"
2903
"movq %%mm2, %%mm3 \n\t"
2904
"punpcklbw %%mm7, %%mm2 \n\t" // L3
2905
"punpckhbw %%mm7, %%mm3 \n\t" // H3
2907
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
2908
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
2909
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2910
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2911
"movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2912
"movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2914
"movq (%%"REG_a", %1, 2), %%mm0 \n\t"
2915
"movq %%mm0, %%mm1 \n\t"
2916
"punpcklbw %%mm7, %%mm0 \n\t" // L4
2917
"punpckhbw %%mm7, %%mm1 \n\t" // H4
2919
"psubw %%mm0, %%mm2 \n\t" // L3 - L4
2920
"psubw %%mm1, %%mm3 \n\t" // H3 - H4
2921
"movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4
2922
"movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4
2923
"paddw %%mm4, %%mm4 \n\t" // 2L2
2924
"paddw %%mm5, %%mm5 \n\t" // 2H2
2925
"psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
2926
"psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
2928
"lea (%%"REG_a", %1), %0 \n\t"
2929
"psllw $2, %%mm2 \n\t" // 4L3 - 4L4
2930
"psllw $2, %%mm3 \n\t" // 4H3 - 4H4
2931
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
2932
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
3009
2933
//50 opcodes so far
3010
"movq (%0, %1, 2), %%mm2 \n\t"
3011
"movq %%mm2, %%mm3 \n\t"
3012
"punpcklbw %%mm7, %%mm2 \n\t" // L5
3013
"punpckhbw %%mm7, %%mm3 \n\t" // H5
3014
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
3015
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
3016
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
3017
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
3019
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
3020
"punpcklbw %%mm7, %%mm6 \n\t" // L6
3021
"psubw %%mm6, %%mm2 \n\t" // L5 - L6
3022
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
3023
"punpckhbw %%mm7, %%mm6 \n\t" // H6
3024
"psubw %%mm6, %%mm3 \n\t" // H5 - H6
3026
"paddw %%mm0, %%mm0 \n\t" // 2L4
3027
"paddw %%mm1, %%mm1 \n\t" // 2H4
3028
"psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
3029
"psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
3031
"psllw $2, %%mm2 \n\t" // 4L5 - 4L6
3032
"psllw $2, %%mm3 \n\t" // 4H5 - 4H6
3033
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
3034
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
3036
"movq (%0, %1, 4), %%mm2 \n\t"
3037
"movq %%mm2, %%mm3 \n\t"
3038
"punpcklbw %%mm7, %%mm2 \n\t" // L7
3039
"punpckhbw %%mm7, %%mm3 \n\t" // H7
3041
"paddw %%mm2, %%mm2 \n\t" // 2L7
3042
"paddw %%mm3, %%mm3 \n\t" // 2H7
3043
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
3044
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
3046
"movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
3047
"movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
3050
"movq %%mm7, %%mm6 \n\t" // 0
3051
"psubw %%mm0, %%mm6 \n\t"
3052
"pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
3053
"movq %%mm7, %%mm6 \n\t" // 0
3054
"psubw %%mm1, %%mm6 \n\t"
3055
"pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
3056
"movq %%mm7, %%mm6 \n\t" // 0
3057
"psubw %%mm2, %%mm6 \n\t"
3058
"pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
3059
"movq %%mm7, %%mm6 \n\t" // 0
3060
"psubw %%mm3, %%mm6 \n\t"
3061
"pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3063
"movq %%mm7, %%mm6 \n\t" // 0
3064
"pcmpgtw %%mm0, %%mm6 \n\t"
3065
"pxor %%mm6, %%mm0 \n\t"
3066
"psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
3067
"movq %%mm7, %%mm6 \n\t" // 0
3068
"pcmpgtw %%mm1, %%mm6 \n\t"
3069
"pxor %%mm6, %%mm1 \n\t"
3070
"psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
3071
"movq %%mm7, %%mm6 \n\t" // 0
3072
"pcmpgtw %%mm2, %%mm6 \n\t"
3073
"pxor %%mm6, %%mm2 \n\t"
3074
"psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
3075
"movq %%mm7, %%mm6 \n\t" // 0
3076
"pcmpgtw %%mm3, %%mm6 \n\t"
3077
"pxor %%mm6, %%mm3 \n\t"
3078
"psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3082
"pminsw %%mm2, %%mm0 \n\t"
3083
"pminsw %%mm3, %%mm1 \n\t"
3085
"movq %%mm0, %%mm6 \n\t"
3086
"psubusw %%mm2, %%mm6 \n\t"
3087
"psubw %%mm6, %%mm0 \n\t"
3088
"movq %%mm1, %%mm6 \n\t"
3089
"psubusw %%mm3, %%mm6 \n\t"
3090
"psubw %%mm6, %%mm1 \n\t"
3093
"movd %2, %%mm2 \n\t" // QP
3094
"punpcklbw %%mm7, %%mm2 \n\t"
3096
"movq %%mm7, %%mm6 \n\t" // 0
3097
"pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
3098
"pxor %%mm6, %%mm4 \n\t"
3099
"psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
3100
"pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
3101
"pxor %%mm7, %%mm5 \n\t"
3102
"psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
2934
"movq (%0, %1, 2), %%mm2 \n\t"
2935
"movq %%mm2, %%mm3 \n\t"
2936
"punpcklbw %%mm7, %%mm2 \n\t" // L5
2937
"punpckhbw %%mm7, %%mm3 \n\t" // H5
2938
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
2939
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
2940
"psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
2941
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
2943
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
2944
"punpcklbw %%mm7, %%mm6 \n\t" // L6
2945
"psubw %%mm6, %%mm2 \n\t" // L5 - L6
2946
"movq (%%"REG_a", %1, 4), %%mm6 \n\t"
2947
"punpckhbw %%mm7, %%mm6 \n\t" // H6
2948
"psubw %%mm6, %%mm3 \n\t" // H5 - H6
2950
"paddw %%mm0, %%mm0 \n\t" // 2L4
2951
"paddw %%mm1, %%mm1 \n\t" // 2H4
2952
"psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
2953
"psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
2955
"psllw $2, %%mm2 \n\t" // 4L5 - 4L6
2956
"psllw $2, %%mm3 \n\t" // 4H5 - 4H6
2957
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
2958
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
2960
"movq (%0, %1, 4), %%mm2 \n\t"
2961
"movq %%mm2, %%mm3 \n\t"
2962
"punpcklbw %%mm7, %%mm2 \n\t" // L7
2963
"punpckhbw %%mm7, %%mm3 \n\t" // H7
2965
"paddw %%mm2, %%mm2 \n\t" // 2L7
2966
"paddw %%mm3, %%mm3 \n\t" // 2H7
2967
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
2968
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
2970
"movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
2971
"movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
2974
"movq %%mm7, %%mm6 \n\t" // 0
2975
"psubw %%mm0, %%mm6 \n\t"
2976
"pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
2977
"movq %%mm7, %%mm6 \n\t" // 0
2978
"psubw %%mm1, %%mm6 \n\t"
2979
"pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
2980
"movq %%mm7, %%mm6 \n\t" // 0
2981
"psubw %%mm2, %%mm6 \n\t"
2982
"pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
2983
"movq %%mm7, %%mm6 \n\t" // 0
2984
"psubw %%mm3, %%mm6 \n\t"
2985
"pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
2987
"movq %%mm7, %%mm6 \n\t" // 0
2988
"pcmpgtw %%mm0, %%mm6 \n\t"
2989
"pxor %%mm6, %%mm0 \n\t"
2990
"psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
2991
"movq %%mm7, %%mm6 \n\t" // 0
2992
"pcmpgtw %%mm1, %%mm6 \n\t"
2993
"pxor %%mm6, %%mm1 \n\t"
2994
"psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
2995
"movq %%mm7, %%mm6 \n\t" // 0
2996
"pcmpgtw %%mm2, %%mm6 \n\t"
2997
"pxor %%mm6, %%mm2 \n\t"
2998
"psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
2999
"movq %%mm7, %%mm6 \n\t" // 0
3000
"pcmpgtw %%mm3, %%mm6 \n\t"
3001
"pxor %%mm6, %%mm3 \n\t"
3002
"psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
3006
"pminsw %%mm2, %%mm0 \n\t"
3007
"pminsw %%mm3, %%mm1 \n\t"
3009
"movq %%mm0, %%mm6 \n\t"
3010
"psubusw %%mm2, %%mm6 \n\t"
3011
"psubw %%mm6, %%mm0 \n\t"
3012
"movq %%mm1, %%mm6 \n\t"
3013
"psubusw %%mm3, %%mm6 \n\t"
3014
"psubw %%mm6, %%mm1 \n\t"
3017
"movd %2, %%mm2 \n\t" // QP
3018
"punpcklbw %%mm7, %%mm2 \n\t"
3020
"movq %%mm7, %%mm6 \n\t" // 0
3021
"pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
3022
"pxor %%mm6, %%mm4 \n\t"
3023
"psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
3024
"pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
3025
"pxor %%mm7, %%mm5 \n\t"
3026
"psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
3104
"psllw $3, %%mm2 \n\t" // 8QP
3105
"movq %%mm2, %%mm3 \n\t" // 8QP
3106
"pcmpgtw %%mm4, %%mm2 \n\t"
3107
"pcmpgtw %%mm5, %%mm3 \n\t"
3108
"pand %%mm2, %%mm4 \n\t"
3109
"pand %%mm3, %%mm5 \n\t"
3112
"psubusw %%mm0, %%mm4 \n\t" // hd
3113
"psubusw %%mm1, %%mm5 \n\t" // ld
3116
"movq "MANGLE(w05)", %%mm2 \n\t" // 5
3117
"pmullw %%mm2, %%mm4 \n\t"
3118
"pmullw %%mm2, %%mm5 \n\t"
3119
"movq "MANGLE(w20)", %%mm2 \n\t" // 32
3120
"paddw %%mm2, %%mm4 \n\t"
3121
"paddw %%mm2, %%mm5 \n\t"
3122
"psrlw $6, %%mm4 \n\t"
3123
"psrlw $6, %%mm5 \n\t"
3125
"movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4
3126
"movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4
3128
"pxor %%mm2, %%mm2 \n\t"
3129
"pxor %%mm3, %%mm3 \n\t"
3131
"pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
3132
"pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
3133
"pxor %%mm2, %%mm0 \n\t"
3134
"pxor %%mm3, %%mm1 \n\t"
3135
"psubw %%mm2, %%mm0 \n\t" // |L3-L4|
3136
"psubw %%mm3, %%mm1 \n\t" // |H3-H4|
3137
"psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
3138
"psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
3140
"pxor %%mm6, %%mm2 \n\t"
3141
"pxor %%mm7, %%mm3 \n\t"
3142
"pand %%mm2, %%mm4 \n\t"
3143
"pand %%mm3, %%mm5 \n\t"
3146
"pminsw %%mm0, %%mm4 \n\t"
3147
"pminsw %%mm1, %%mm5 \n\t"
3028
"psllw $3, %%mm2 \n\t" // 8QP
3029
"movq %%mm2, %%mm3 \n\t" // 8QP
3030
"pcmpgtw %%mm4, %%mm2 \n\t"
3031
"pcmpgtw %%mm5, %%mm3 \n\t"
3032
"pand %%mm2, %%mm4 \n\t"
3033
"pand %%mm3, %%mm5 \n\t"
3036
"psubusw %%mm0, %%mm4 \n\t" // hd
3037
"psubusw %%mm1, %%mm5 \n\t" // ld
3040
"movq "MANGLE(w05)", %%mm2 \n\t" // 5
3041
"pmullw %%mm2, %%mm4 \n\t"
3042
"pmullw %%mm2, %%mm5 \n\t"
3043
"movq "MANGLE(w20)", %%mm2 \n\t" // 32
3044
"paddw %%mm2, %%mm4 \n\t"
3045
"paddw %%mm2, %%mm5 \n\t"
3046
"psrlw $6, %%mm4 \n\t"
3047
"psrlw $6, %%mm5 \n\t"
3049
"movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4
3050
"movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4
3052
"pxor %%mm2, %%mm2 \n\t"
3053
"pxor %%mm3, %%mm3 \n\t"
3055
"pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
3056
"pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
3057
"pxor %%mm2, %%mm0 \n\t"
3058
"pxor %%mm3, %%mm1 \n\t"
3059
"psubw %%mm2, %%mm0 \n\t" // |L3-L4|
3060
"psubw %%mm3, %%mm1 \n\t" // |H3-H4|
3061
"psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
3062
"psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
3064
"pxor %%mm6, %%mm2 \n\t"
3065
"pxor %%mm7, %%mm3 \n\t"
3066
"pand %%mm2, %%mm4 \n\t"
3067
"pand %%mm3, %%mm5 \n\t"
3070
"pminsw %%mm0, %%mm4 \n\t"
3071
"pminsw %%mm1, %%mm5 \n\t"
3149
"movq %%mm4, %%mm2 \n\t"
3150
"psubusw %%mm0, %%mm2 \n\t"
3151
"psubw %%mm2, %%mm4 \n\t"
3152
"movq %%mm5, %%mm2 \n\t"
3153
"psubusw %%mm1, %%mm2 \n\t"
3154
"psubw %%mm2, %%mm5 \n\t"
3073
"movq %%mm4, %%mm2 \n\t"
3074
"psubusw %%mm0, %%mm2 \n\t"
3075
"psubw %%mm2, %%mm4 \n\t"
3076
"movq %%mm5, %%mm2 \n\t"
3077
"psubusw %%mm1, %%mm2 \n\t"
3078
"psubw %%mm2, %%mm5 \n\t"
3156
"pxor %%mm6, %%mm4 \n\t"
3157
"pxor %%mm7, %%mm5 \n\t"
3158
"psubw %%mm6, %%mm4 \n\t"
3159
"psubw %%mm7, %%mm5 \n\t"
3160
"packsswb %%mm5, %%mm4 \n\t"
3161
"movq %3, %%mm1 \n\t"
3162
"pandn %%mm4, %%mm1 \n\t"
3163
"movq (%0), %%mm0 \n\t"
3164
"paddb %%mm1, %%mm0 \n\t"
3165
"movq %%mm0, (%0) \n\t"
3166
"movq (%0, %1), %%mm0 \n\t"
3167
"psubb %%mm1, %%mm0 \n\t"
3168
"movq %%mm0, (%0, %1) \n\t"
3080
"pxor %%mm6, %%mm4 \n\t"
3081
"pxor %%mm7, %%mm5 \n\t"
3082
"psubw %%mm6, %%mm4 \n\t"
3083
"psubw %%mm7, %%mm5 \n\t"
3084
"packsswb %%mm5, %%mm4 \n\t"
3085
"movq %3, %%mm1 \n\t"
3086
"pandn %%mm4, %%mm1 \n\t"
3087
"movq (%0), %%mm0 \n\t"
3088
"paddb %%mm1, %%mm0 \n\t"
3089
"movq %%mm0, (%0) \n\t"
3090
"movq (%0, %1), %%mm0 \n\t"
3091
"psubb %%mm1, %%mm0 \n\t"
3092
"movq %%mm0, (%0, %1) \n\t"
3171
: "r" ((long)step), "m" (c->pQPb), "m"(eq_mask)
3172
: "%"REG_a, "%"REG_c
3095
: "r" ((x86_reg)step), "m" (c->pQPb), "m"(eq_mask)
3096
: "%"REG_a, "%"REG_c
3175
3099
/*if(step==16){
3176
3100
STOP_TIMER("step16")
3323
3244
static inline void RENAME(duplicate)(uint8_t src[], int stride)
3327
"movq (%0), %%mm0 \n\t"
3329
"movq %%mm0, (%0) \n\t"
3330
"movq %%mm0, (%0, %1) \n\t"
3331
"movq %%mm0, (%0, %1, 2) \n\t"
3333
: "r" ((long)-stride)
3248
"movq (%0), %%mm0 \n\t"
3250
"movq %%mm0, (%0) \n\t"
3251
"movq %%mm0, (%0, %1) \n\t"
3252
"movq %%mm0, (%0, %1, 2) \n\t"
3254
: "r" ((x86_reg)-stride)
3347
3267
* Filters array of bytes (Y or U or V values)
3349
static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
3350
QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)
3269
static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
3270
const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)
3352
DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access
3272
DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access
3354
3274
#ifdef COMPILE_TIME_MODE
3355
const int mode= COMPILE_TIME_MODE;
3275
const int mode= COMPILE_TIME_MODE;
3357
const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;
3359
int black=0, white=255; // blackest black and whitest white in the picture
3360
int QPCorrecture= 256*256;
3277
const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;
3279
int black=0, white=255; // blackest black and whitest white in the picture
3280
int QPCorrecture= 256*256;
3287
const int qpHShift= isColor ? 4-c.hChromaSubSample : 4;
3288
const int qpVShift= isColor ? 4-c.vChromaSubSample : 4;
3291
uint64_t * const yHistogram= c.yHistogram;
3292
uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride;
3293
uint8_t * const tempDst= dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride;
3294
//const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
3297
for(i=0; i<57; i++){
3298
int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;
3299
int threshold= offset*2 + 1;
3300
c.mmxDcOffset[i]= 0x7F - offset;
3301
c.mmxDcThreshold[i]= 0x7F - threshold;
3302
c.mmxDcOffset[i]*= 0x0101010101010101LL;
3303
c.mmxDcThreshold[i]*= 0x0101010101010101LL;
3307
if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
3308
else if( (mode & LINEAR_BLEND_DEINT_FILTER)
3309
|| (mode & FFMPEG_DEINT_FILTER)
3310
|| (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14;
3311
else if( (mode & V_DEBLOCK)
3312
|| (mode & LINEAR_IPOL_DEINT_FILTER)
3313
|| (mode & MEDIAN_DEINT_FILTER)
3314
|| (mode & V_A_DEBLOCK)) copyAhead=13;
3315
else if(mode & V_X1_FILTER) copyAhead=11;
3316
// else if(mode & V_RK1_FILTER) copyAhead=10;
3317
else if(mode & DERING) copyAhead=9;
3367
const int qpHShift= isColor ? 4-c.hChromaSubSample : 4;
3368
const int qpVShift= isColor ? 4-c.vChromaSubSample : 4;
3371
uint64_t * const yHistogram= c.yHistogram;
3372
uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride;
3373
uint8_t * const tempDst= dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride;
3374
//const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
3377
for(i=0; i<57; i++){
3378
int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;
3379
int threshold= offset*2 + 1;
3380
c.mmxDcOffset[i]= 0x7F - offset;
3381
c.mmxDcThreshold[i]= 0x7F - threshold;
3382
c.mmxDcOffset[i]*= 0x0101010101010101LL;
3383
c.mmxDcThreshold[i]*= 0x0101010101010101LL;
3387
if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
3388
else if( (mode & LINEAR_BLEND_DEINT_FILTER)
3389
|| (mode & FFMPEG_DEINT_FILTER)
3390
|| (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14;
3391
else if( (mode & V_DEBLOCK)
3392
|| (mode & LINEAR_IPOL_DEINT_FILTER)
3393
|| (mode & MEDIAN_DEINT_FILTER)
3394
|| (mode & V_A_DEBLOCK)) copyAhead=13;
3395
else if(mode & V_X1_FILTER) copyAhead=11;
3396
// else if(mode & V_RK1_FILTER) copyAhead=10;
3397
else if(mode & DERING) copyAhead=9;
3406
uint64_t maxClipped;
3411
// first frame is fscked so we ignore it
3412
if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256;
3414
for(i=0; i<256; i++)
3416
sum+= yHistogram[i];
3419
/* We always get a completely black picture first. */
3420
maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold);
3423
for(black=255; black>0; black--)
3425
if(clipped < maxClipped) break;
3426
clipped-= yHistogram[black];
3430
for(white=0; white<256; white++)
3432
if(clipped < maxClipped) break;
3433
clipped-= yHistogram[white];
3436
scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);
3439
c.packedYScale= (uint16_t)(scale*256.0 + 0.5);
3440
c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;
3325
uint64_t maxClipped;
3330
// first frame is fscked so we ignore it
3331
if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256;
3333
for(i=0; i<256; i++){
3334
sum+= yHistogram[i];
3337
/* We always get a completely black picture first. */
3338
maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold);
3341
for(black=255; black>0; black--){
3342
if(clipped < maxClipped) break;
3343
clipped-= yHistogram[black];
3347
for(white=0; white<256; white++){
3348
if(clipped < maxClipped) break;
3349
clipped-= yHistogram[white];
3352
scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);
3355
c.packedYScale= (uint16_t)(scale*256.0 + 0.5);
3356
c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;
3442
c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);
3443
c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;
3446
c.packedYOffset|= c.packedYOffset<<32;
3447
c.packedYOffset|= c.packedYOffset<<16;
3449
c.packedYScale|= c.packedYScale<<32;
3450
c.packedYScale|= c.packedYScale<<16;
3452
if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5);
3453
else QPCorrecture= 256*256;
3457
c.packedYScale= 0x0100010001000100LL;
3459
QPCorrecture= 256*256;
3462
/* copy & deinterlace first row of blocks */
3465
uint8_t *srcBlock= &(src[y*srcStride]);
3466
uint8_t *dstBlock= tempDst + dstStride;
3468
// From this point on it is guaranteed that we can read and write 16 lines downward
3469
// finish 1 block before the next otherwise we might have a problem
3470
// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3471
for(x=0; x<width; x+=BLOCK_SIZE)
3476
prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3477
prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3478
prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3479
prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3483
"mov %4, %%"REG_a" \n\t"
3484
"shr $2, %%"REG_a" \n\t"
3485
"and $6, %%"REG_a" \n\t"
3486
"add %5, %%"REG_a" \n\t"
3487
"mov %%"REG_a", %%"REG_d" \n\t"
3488
"imul %1, %%"REG_a" \n\t"
3489
"imul %3, %%"REG_d" \n\t"
3490
"prefetchnta 32(%%"REG_a", %0) \n\t"
3491
"prefetcht0 32(%%"REG_d", %2) \n\t"
3492
"add %1, %%"REG_a" \n\t"
3493
"add %3, %%"REG_d" \n\t"
3494
"prefetchnta 32(%%"REG_a", %0) \n\t"
3495
"prefetcht0 32(%%"REG_d", %2) \n\t"
3496
:: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride),
3497
"g" ((long)x), "g" ((long)copyAhead)
3498
: "%"REG_a, "%"REG_d
3501
#elif defined(HAVE_3DNOW)
3502
//FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3503
/* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3504
prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3505
prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3506
prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3510
RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
3511
srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3513
RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
3515
if(mode & LINEAR_IPOL_DEINT_FILTER)
3516
RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3517
else if(mode & LINEAR_BLEND_DEINT_FILTER)
3518
RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3519
else if(mode & MEDIAN_DEINT_FILTER)
3520
RENAME(deInterlaceMedian)(dstBlock, dstStride);
3521
else if(mode & CUBIC_IPOL_DEINT_FILTER)
3522
RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3523
else if(mode & FFMPEG_DEINT_FILTER)
3524
RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3525
else if(mode & LOWPASS5_DEINT_FILTER)
3526
RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3527
/* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3528
RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3533
if(width==FFABS(dstStride))
3534
linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride);
3538
for(i=0; i<copyAhead; i++)
3540
memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width);
3545
for(y=0; y<height; y+=BLOCK_SIZE)
3547
//1% speedup if these are here instead of the inner loop
3548
uint8_t *srcBlock= &(src[y*srcStride]);
3549
uint8_t *dstBlock= &(dst[y*dstStride]);
3551
uint8_t *tempBlock1= c.tempBlocks;
3552
uint8_t *tempBlock2= c.tempBlocks + 8;
3554
int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
3555
int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)];
3557
/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
3558
if not than use a temporary buffer */
3562
/* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
3563
blockcopy to dst later */
3564
linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
3565
FFMAX(height-y-copyAhead, 0), srcStride);
3567
/* duplicate last line of src to fill the void upto line (copyAhead+7) */
3568
for(i=FFMAX(height-y, 8); i<copyAhead+8; i++)
3569
memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), FFABS(srcStride));
3571
/* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
3572
linecpy(tempDst, dstBlock - dstStride, FFMIN(height-y+1, copyAhead+1), dstStride);
3574
/* duplicate last line of dst to fill the void upto line (copyAhead) */
3575
for(i=height-y+1; i<=copyAhead; i++)
3576
memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), FFABS(dstStride));
3578
dstBlock= tempDst + dstStride;
3582
// From this point on it is guaranteed that we can read and write 16 lines downward
3583
// finish 1 block before the next otherwise we might have a problem
3584
// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3585
for(x=0; x<width; x+=BLOCK_SIZE)
3587
const int stride= dstStride;
3593
QP= QPptr[x>>qpHShift];
3594
c.nonBQP= nonBQPptr[x>>qpHShift];
3599
QP= (QP* QPCorrecture + 256*128)>>16;
3600
c.nonBQP= nonBQPptr[x>>4];
3601
c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
3602
yHistogram[ srcBlock[srcStride*12 + 4] ]++;
3607
"movd %1, %%mm7 \n\t"
3608
"packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
3609
"packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
3610
"packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
3611
"movq %%mm7, %0 \n\t"
3620
prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3621
prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3622
prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3623
prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3627
"mov %4, %%"REG_a" \n\t"
3628
"shr $2, %%"REG_a" \n\t"
3629
"and $6, %%"REG_a" \n\t"
3630
"add %5, %%"REG_a" \n\t"
3631
"mov %%"REG_a", %%"REG_d" \n\t"
3632
"imul %1, %%"REG_a" \n\t"
3633
"imul %3, %%"REG_d" \n\t"
3634
"prefetchnta 32(%%"REG_a", %0) \n\t"
3635
"prefetcht0 32(%%"REG_d", %2) \n\t"
3636
"add %1, %%"REG_a" \n\t"
3637
"add %3, %%"REG_d" \n\t"
3638
"prefetchnta 32(%%"REG_a", %0) \n\t"
3639
"prefetcht0 32(%%"REG_d", %2) \n\t"
3640
:: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride),
3641
"g" ((long)x), "g" ((long)copyAhead)
3642
: "%"REG_a, "%"REG_d
3645
#elif defined(HAVE_3DNOW)
3646
//FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3647
/* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3648
prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3649
prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3650
prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3654
RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
3655
srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3657
if(mode & LINEAR_IPOL_DEINT_FILTER)
3658
RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3659
else if(mode & LINEAR_BLEND_DEINT_FILTER)
3660
RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3661
else if(mode & MEDIAN_DEINT_FILTER)
3662
RENAME(deInterlaceMedian)(dstBlock, dstStride);
3663
else if(mode & CUBIC_IPOL_DEINT_FILTER)
3664
RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3665
else if(mode & FFMPEG_DEINT_FILTER)
3666
RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3667
else if(mode & LOWPASS5_DEINT_FILTER)
3668
RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3669
/* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3670
RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3673
/* only deblock if we have 2 blocks */
3676
if(mode & V_X1_FILTER)
3677
RENAME(vertX1Filter)(dstBlock, stride, &c);
3678
else if(mode & V_DEBLOCK)
3680
const int t= RENAME(vertClassify)(dstBlock, stride, &c);
3683
RENAME(doVertLowPass)(dstBlock, stride, &c);
3685
RENAME(doVertDefFilter)(dstBlock, stride, &c);
3686
}else if(mode & V_A_DEBLOCK){
3687
RENAME(do_a_deblock)(dstBlock, stride, 1, &c);
3692
RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
3694
/* check if we have a previous block to deblock it with dstBlock */
3698
if(mode & H_X1_FILTER)
3699
RENAME(vertX1Filter)(tempBlock1, 16, &c);
3700
else if(mode & H_DEBLOCK)
3358
c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);
3359
c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;
3362
c.packedYOffset|= c.packedYOffset<<32;
3363
c.packedYOffset|= c.packedYOffset<<16;
3365
c.packedYScale|= c.packedYScale<<32;
3366
c.packedYScale|= c.packedYScale<<16;
3368
if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5);
3369
else QPCorrecture= 256*256;
3371
c.packedYScale= 0x0100010001000100LL;
3373
QPCorrecture= 256*256;
3376
/* copy & deinterlace first row of blocks */
3379
const uint8_t *srcBlock= &(src[y*srcStride]);
3380
uint8_t *dstBlock= tempDst + dstStride;
3382
// From this point on it is guaranteed that we can read and write 16 lines downward
3383
// finish 1 block before the next otherwise we might have a problem
3384
// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3385
for(x=0; x<width; x+=BLOCK_SIZE){
3389
prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3390
prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3391
prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3392
prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3396
"mov %4, %%"REG_a" \n\t"
3397
"shr $2, %%"REG_a" \n\t"
3398
"and $6, %%"REG_a" \n\t"
3399
"add %5, %%"REG_a" \n\t"
3400
"mov %%"REG_a", %%"REG_d" \n\t"
3401
"imul %1, %%"REG_a" \n\t"
3402
"imul %3, %%"REG_d" \n\t"
3403
"prefetchnta 32(%%"REG_a", %0) \n\t"
3404
"prefetcht0 32(%%"REG_d", %2) \n\t"
3405
"add %1, %%"REG_a" \n\t"
3406
"add %3, %%"REG_d" \n\t"
3407
"prefetchnta 32(%%"REG_a", %0) \n\t"
3408
"prefetcht0 32(%%"REG_d", %2) \n\t"
3409
:: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride),
3410
"g" ((x86_reg)x), "g" ((x86_reg)copyAhead)
3411
: "%"REG_a, "%"REG_d
3415
//FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3416
/* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3417
prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3418
prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3419
prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3423
RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
3424
srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3426
RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
3428
if(mode & LINEAR_IPOL_DEINT_FILTER)
3429
RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3430
else if(mode & LINEAR_BLEND_DEINT_FILTER)
3431
RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3432
else if(mode & MEDIAN_DEINT_FILTER)
3433
RENAME(deInterlaceMedian)(dstBlock, dstStride);
3434
else if(mode & CUBIC_IPOL_DEINT_FILTER)
3435
RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3436
else if(mode & FFMPEG_DEINT_FILTER)
3437
RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3438
else if(mode & LOWPASS5_DEINT_FILTER)
3439
RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3440
/* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3441
RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3446
if(width==FFABS(dstStride))
3447
linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride);
3450
for(i=0; i<copyAhead; i++){
3451
memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width);
3456
for(y=0; y<height; y+=BLOCK_SIZE){
3457
//1% speedup if these are here instead of the inner loop
3458
const uint8_t *srcBlock= &(src[y*srcStride]);
3459
uint8_t *dstBlock= &(dst[y*dstStride]);
3461
uint8_t *tempBlock1= c.tempBlocks;
3462
uint8_t *tempBlock2= c.tempBlocks + 8;
3464
const int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
3465
int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)];
3467
/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
3468
if not than use a temporary buffer */
3471
/* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
3472
blockcopy to dst later */
3473
linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
3474
FFMAX(height-y-copyAhead, 0), srcStride);
3476
/* duplicate last line of src to fill the void upto line (copyAhead+7) */
3477
for(i=FFMAX(height-y, 8); i<copyAhead+8; i++)
3478
memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), FFABS(srcStride));
3480
/* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
3481
linecpy(tempDst, dstBlock - dstStride, FFMIN(height-y+1, copyAhead+1), dstStride);
3483
/* duplicate last line of dst to fill the void upto line (copyAhead) */
3484
for(i=height-y+1; i<=copyAhead; i++)
3485
memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), FFABS(dstStride));
3487
dstBlock= tempDst + dstStride;
3491
// From this point on it is guaranteed that we can read and write 16 lines downward
3492
// finish 1 block before the next otherwise we might have a problem
3493
// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
3494
for(x=0; x<width; x+=BLOCK_SIZE){
3495
const int stride= dstStride;
3500
QP= QPptr[x>>qpHShift];
3501
c.nonBQP= nonBQPptr[x>>qpHShift];
3504
QP= (QP* QPCorrecture + 256*128)>>16;
3505
c.nonBQP= nonBQPptr[x>>4];
3506
c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
3507
yHistogram[ srcBlock[srcStride*12 + 4] ]++;
3512
"movd %1, %%mm7 \n\t"
3513
"packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
3514
"packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
3515
"packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
3516
"movq %%mm7, %0 \n\t"
3525
prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3526
prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3527
prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3528
prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3532
"mov %4, %%"REG_a" \n\t"
3533
"shr $2, %%"REG_a" \n\t"
3534
"and $6, %%"REG_a" \n\t"
3535
"add %5, %%"REG_a" \n\t"
3536
"mov %%"REG_a", %%"REG_d" \n\t"
3537
"imul %1, %%"REG_a" \n\t"
3538
"imul %3, %%"REG_d" \n\t"
3539
"prefetchnta 32(%%"REG_a", %0) \n\t"
3540
"prefetcht0 32(%%"REG_d", %2) \n\t"
3541
"add %1, %%"REG_a" \n\t"
3542
"add %3, %%"REG_d" \n\t"
3543
"prefetchnta 32(%%"REG_a", %0) \n\t"
3544
"prefetcht0 32(%%"REG_d", %2) \n\t"
3545
:: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride),
3546
"g" ((x86_reg)x), "g" ((x86_reg)copyAhead)
3547
: "%"REG_a, "%"REG_d
3551
//FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ...
3552
/* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3553
prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3554
prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3555
prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3559
RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
3560
srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
3562
if(mode & LINEAR_IPOL_DEINT_FILTER)
3563
RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
3564
else if(mode & LINEAR_BLEND_DEINT_FILTER)
3565
RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x);
3566
else if(mode & MEDIAN_DEINT_FILTER)
3567
RENAME(deInterlaceMedian)(dstBlock, dstStride);
3568
else if(mode & CUBIC_IPOL_DEINT_FILTER)
3569
RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
3570
else if(mode & FFMPEG_DEINT_FILTER)
3571
RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
3572
else if(mode & LOWPASS5_DEINT_FILTER)
3573
RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);
3574
/* else if(mode & CUBIC_BLEND_DEINT_FILTER)
3575
RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
3578
/* only deblock if we have 2 blocks */
3580
if(mode & V_X1_FILTER)
3581
RENAME(vertX1Filter)(dstBlock, stride, &c);
3582
else if(mode & V_DEBLOCK){
3583
const int t= RENAME(vertClassify)(dstBlock, stride, &c);
3586
RENAME(doVertLowPass)(dstBlock, stride, &c);
3588
RENAME(doVertDefFilter)(dstBlock, stride, &c);
3589
}else if(mode & V_A_DEBLOCK){
3590
RENAME(do_a_deblock)(dstBlock, stride, 1, &c);
3595
RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
3597
/* check if we have a previous block to deblock it with dstBlock */
3600
if(mode & H_X1_FILTER)
3601
RENAME(vertX1Filter)(tempBlock1, 16, &c);
3602
else if(mode & H_DEBLOCK){
3703
const int t= RENAME(vertClassify)(tempBlock1, 16, &c);
3604
const int t= RENAME(vertClassify)(tempBlock1, 16, &c);
3704
3605
//STOP_TIMER("dc & minmax")
3706
RENAME(doVertLowPass)(tempBlock1, 16, &c);
3708
RENAME(doVertDefFilter)(tempBlock1, 16, &c);
3709
}else if(mode & H_A_DEBLOCK){
3710
RENAME(do_a_deblock)(tempBlock1, 16, 1, &c);
3713
RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
3716
if(mode & H_X1_FILTER)
3717
horizX1Filter(dstBlock-4, stride, QP);
3718
else if(mode & H_DEBLOCK)
3721
DECLARE_ALIGNED(16, unsigned char, tempBlock[272]);
3722
transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride);
3724
const int t=vertClassify_altivec(tempBlock-48, 16, &c);
3726
doVertLowPass_altivec(tempBlock-48, 16, &c);
3727
transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3730
doVertDefFilter_altivec(tempBlock-48, 16, &c);
3731
transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3734
const int t= RENAME(horizClassify)(dstBlock-4, stride, &c);
3737
RENAME(doHorizLowPass)(dstBlock-4, stride, &c);
3739
RENAME(doHorizDefFilter)(dstBlock-4, stride, &c);
3607
RENAME(doVertLowPass)(tempBlock1, 16, &c);
3609
RENAME(doVertDefFilter)(tempBlock1, 16, &c);
3610
}else if(mode & H_A_DEBLOCK){
3611
RENAME(do_a_deblock)(tempBlock1, 16, 1, &c);
3614
RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
3617
if(mode & H_X1_FILTER)
3618
horizX1Filter(dstBlock-4, stride, QP);
3619
else if(mode & H_DEBLOCK){
3621
DECLARE_ALIGNED(16, unsigned char, tempBlock[272]);
3622
transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride);
3624
const int t=vertClassify_altivec(tempBlock-48, 16, &c);
3626
doVertLowPass_altivec(tempBlock-48, 16, &c);
3627
transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3630
doVertDefFilter_altivec(tempBlock-48, 16, &c);
3631
transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride);
3634
const int t= RENAME(horizClassify)(dstBlock-4, stride, &c);
3637
RENAME(doHorizLowPass)(dstBlock-4, stride, &c);
3639
RENAME(doHorizDefFilter)(dstBlock-4, stride, &c);
3741
}else if(mode & H_A_DEBLOCK){
3742
RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c);
3641
}else if(mode & H_A_DEBLOCK){
3642
RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c);
3744
3644
#endif //HAVE_MMX
3747
//FIXME filter first line
3748
if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);
3751
if(mode & TEMP_NOISE_FILTER)
3753
RENAME(tempNoiseReducer)(dstBlock-8, stride,
3754
c.tempBlured[isColor] + y*dstStride + x,
3755
c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
3756
c.ppMode.maxTmpNoise);
3764
tmpXchg= tempBlock1;
3765
tempBlock1= tempBlock2;
3766
tempBlock2 = tmpXchg;
3646
//FIXME filter first line
3647
if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);
3650
if(mode & TEMP_NOISE_FILTER)
3652
RENAME(tempNoiseReducer)(dstBlock-8, stride,
3653
c.tempBlurred[isColor] + y*dstStride + x,
3654
c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3),
3655
c.ppMode.maxTmpNoise);
3663
tmpXchg= tempBlock1;
3664
tempBlock1= tempBlock2;
3665
tempBlock2 = tmpXchg;
3772
if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);
3775
if((mode & TEMP_NOISE_FILTER))
3777
RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
3778
c.tempBlured[isColor] + y*dstStride + x,
3779
c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
3780
c.ppMode.maxTmpNoise);
3783
/* did we use a tmp buffer for the last lines*/
3786
uint8_t *dstBlock= &(dst[y*dstStride]);
3787
if(width==FFABS(dstStride))
3788
linecpy(dstBlock, tempDst + dstStride, height-y, dstStride);
3792
for(i=0; i<height-y; i++)
3794
memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width);
3670
if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);
3673
if((mode & TEMP_NOISE_FILTER)){
3674
RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
3675
c.tempBlurred[isColor] + y*dstStride + x,
3676
c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3),
3677
c.ppMode.maxTmpNoise);
3680
/* did we use a tmp buffer for the last lines*/
3682
uint8_t *dstBlock= &(dst[y*dstStride]);
3683
if(width==FFABS(dstStride))
3684
linecpy(dstBlock, tempDst + dstStride, height-y, dstStride);
3687
for(i=0; i<height-y; i++){
3688
memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width);
3799
for(x=0; x<width; x+=32)
3802
i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
3803
+ dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
3804
+ dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
3805
// + dstBlock[x +13*dstStride]
3806
// + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
3810
asm volatile("femms");
3811
#elif defined (HAVE_MMX)
3812
asm volatile("emms");
3693
for(x=0; x<width; x+=32){
3695
i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
3696
+ dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
3697
+ dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
3698
+ dstBlock[x +13*dstStride]
3699
+ dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
3703
__asm__ volatile("femms");
3705
__asm__ volatile("emms");
3815
3708
#ifdef DEBUG_BRIGHTNESS
3820
for(i=0; i<256; i++)
3821
if(yHistogram[i] > max) max=yHistogram[i];
3823
for(i=1; i<256; i++)
3826
int start=yHistogram[i-1]/(max/256+1);
3827
int end=yHistogram[i]/(max/256+1);
3828
int inc= end > start ? 1 : -1;
3829
for(x=start; x!=end+inc; x+=inc)
3830
dst[ i*dstStride + x]+=128;
3833
for(i=0; i<100; i+=2)
3835
dst[ (white)*dstStride + i]+=128;
3836
dst[ (black)*dstStride + i]+=128;
3712
for(i=0; i<256; i++)
3713
if(yHistogram[i] > max) max=yHistogram[i];
3715
for(i=1; i<256; i++){
3717
int start=yHistogram[i-1]/(max/256+1);
3718
int end=yHistogram[i]/(max/256+1);
3719
int inc= end > start ? 1 : -1;
3720
for(x=start; x!=end+inc; x+=inc)
3721
dst[ i*dstStride + x]+=128;
3724
for(i=0; i<100; i+=2){
3725
dst[ (white)*dstStride + i]+=128;
3726
dst[ (black)*dstStride + i]+=128;
3842
*c2= c; //copy local context back
3731
*c2= c; //copy local context back