Mercurial > libpostproc.hg
comparison postprocess_template.c @ 129:c78bbb57f0b9 libpostproc
Remove disabled code cruft.
author | diego |
---|---|
date | Wed, 19 Aug 2009 08:15:32 +0000 |
parents | 1500ae6cf66c |
children | 51571e34b760 |
comparison
equal
deleted
inserted
replaced
128:1afabe715e63 | 129:c78bbb57f0b9 |
---|---|
346 src++; | 346 src++; |
347 } | 347 } |
348 #endif //HAVE_MMX2 || HAVE_AMD3DNOW | 348 #endif //HAVE_MMX2 || HAVE_AMD3DNOW |
349 } | 349 } |
350 #endif //HAVE_ALTIVEC | 350 #endif //HAVE_ALTIVEC |
351 | |
352 #if 0 | |
353 /** | |
354 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
355 * values are correctly clipped (MMX2) | |
356 * values are wraparound (C) | |
357 * Conclusion: It is fast, but introduces ugly horizontal patterns | |
358 * if there is a continuous gradient. | |
359 0 8 16 24 | |
360 x = 8 | |
361 x/2 = 4 | |
362 x/8 = 1 | |
363 1 12 12 23 | |
364 */ | |
365 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP) | |
366 { | |
367 #if HAVE_MMX2 || HAVE_AMD3DNOW | |
368 src+= stride*3; | |
369 // FIXME rounding | |
370 __asm__ volatile( | |
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" | |
375 // 0 1 2 3 4 5 6 7 8 9 | |
376 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 | |
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 | |
388 PAVGB(%%mm3, %%mm5) | |
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 | |
396 | |
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" | |
401 | |
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" | |
407 | |
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 | |
412 | |
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" | |
418 | |
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" | |
424 | |
425 : | |
426 : "r" (src), "r" ((x86_reg)stride) | |
427 : "%"REG_a, "%"REG_c | |
428 ); | |
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; | |
439 int x; | |
440 const int QP15= QP + (QP>>2); | |
441 src+= stride*3; | |
442 for(x=0; x<BLOCK_SIZE; x++){ | |
443 const int v = (src[x+l5] - src[x+l4]); | |
444 if(FFABS(v) < QP15){ | |
445 src[x+l3] +=v>>3; | |
446 src[x+l4] +=v>>1; | |
447 src[x+l5] -=v>>1; | |
448 src[x+l6] -=v>>3; | |
449 } | |
450 } | |
451 | |
452 #endif //HAVE_MMX2 || HAVE_AMD3DNOW | |
453 } | |
454 #endif //0 | |
455 | 351 |
456 /** | 352 /** |
457 * Experimental Filter 1 | 353 * Experimental Filter 1 |
458 * will not damage linear gradients | 354 * will not damage linear gradients |
459 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | 355 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter |