gstreamer/gst/deinterlace2/tvtime/tomsmocomp/SearchLoopBottom.inc
Sebastian Dröge 81f8895cb4 gst/deinterlace2/tvtime/tomsmocomp/: Unroll the loop to handle two bytes at once. This should give a small speedup an...
Original commit message from CVS:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopBottom.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopTop.inc:
* gst/deinterlace2/tvtime/tomsmocomp/StrangeBob.inc:
* gst/deinterlace2/tvtime/tomsmocomp/WierdBob.inc:
Unroll the loop to handle two bytes at once. This should give
a small speedup and makes it possible to handle chroma and luma
different which is needed later.
2008-08-26 12:33:16 +00:00

174 lines
5.4 KiB
C++

// -*- c++ -*-
// Version for non-SSE2
#ifndef IS_C
#ifdef SKIP_SEARCH
"movq %%mm6, %%mm0\n\t" // just use the results of our wierd bob
#else
// JA 9/Dec/2002
// failed experiment
// but leave in placeholder for me to play about
#ifdef DONT_USE_STRANGE_BOB
// Use the best weave if diffs less than 10 as that
// means the image is still or moving cleanly
// if there is motion we will clip which will catch anything
"psubusb "_FOURS", %%mm7\n\t" // sets bits to zero if weave diff < 4
"pxor %%mm0, %%mm0\n\t"
"pcmpeqb %%mm0, %%mm7\n\t" // all ff where weave better, else 00
"pcmpeqb %%mm7, %%mm0\n\t" // all ff where bob better, else 00
"pand %%mm6, %%mm0\n\t" // use bob for these pixel values
"pand %%mm5, %%mm7\n\t" // use weave for these
"por %%mm7, %%mm0\n\t" // combine both
#else
// Use the better of bob or weave
// pminub mm4, TENS // the most we care about
V_PMINUB ("%%mm4", _TENS, "%%mm0") // the most we care about
"psubusb %%mm4, %%mm7\n\t" // foregive that much from weave est?
"psubusb "_FOURS", %%mm7\n\t" // bias it a bit toward weave
"pxor %%mm0, %%mm0\n\t"
"pcmpeqb %%mm0, %%mm7\n\t" // all ff where weave better, else 00
"pcmpeqb %%mm7, %%mm0\n\t" // all ff where bob better, else 00
"pand %%mm6, %%mm0\n\t" // use bob for these pixel values
"pand %%mm5, %%mm7\n\t" // use weave for these
"por %%mm7, %%mm0\n\t" // combine both
#endif
// pminub mm0, Max_Vals // but clip to catch the stray error
V_PMINUB ("%%mm0", _Max_Vals, "%%mm1") // but clip to catch the stray error
// pmaxub mm0, Min_Vals
V_PMAXUB ("%%mm0", _Min_Vals)
#endif
MOVX" "_pDest", %%"XAX"\n\t"
#ifdef USE_VERTICAL_FILTER
"movq %%mm0, %%mm1\n\t"
// pavgb mm0, qword ptr["XBX"]
V_PAVGB ("%%mm0", "(%%"XBX")", "%%mm2", _ShiftMask)
// movntq qword ptr["XAX"+"XDX"], mm0
V_MOVNTQ ("(%"XAX", %%"XDX")", "%%mm0")
// pavgb mm1, qword ptr["XBX"+"XCX"]
V_PAVGB ("%%mm1", "(%%"XBX", %%"XCX")", "%%mm2", _ShiftMask)
//FIXME: XDX or XAX!!
"addq "_dst_pitchw", %%"XBX
// movntq qword ptr["XAX"+"XDX"], mm1
V_MOVNTQ ("(%%"XAX", %%"XDX")", "%%mm1")
#else
// movntq qword ptr["XAX"+"XDX"], mm0
V_MOVNTQ ("(%%"XAX", %%"XDX")", "%%mm0")
#endif
LEAX" 8(%%"XDX"), %%"XDX"\n\t" // bump offset pointer
CMPX" "_Last8", %%"XDX"\n\t" // done with line?
"jb 1b\n\t" // y
MOVX" "_oldbx", %%"XBX"\n\t"
: /* no outputs */
: "m"(pBob),
"m"(src_pitch2),
"m"(ShiftMask),
"m"(pDest),
"m"(dst_pitchw),
"m"(Last8),
"m"(pSrc),
"m"(pSrcP),
"m"(pBobP),
"m"(DiffThres),
"m"(Min_Vals),
"m"(Max_Vals),
"m"(FOURS),
"m"(TENS),
"m"(ONES),
"m"(UVMask),
"m"(Max_Mov),
"m"(YMask),
"m"(oldbx)
: XAX, XCX, XDX, XSI, XDI,
"st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",
#ifdef __MMX__
"mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
#endif
"memory", "cc"
);
// adjust for next line
pSrc += src_pitch2;
pSrcP += src_pitch2;
pDest += dst_pitch2;
pBob += src_pitch2;
pBobP += src_pitch2;
}
return 0;
#else
#ifdef SKIP_SEARCH
out[0] = best[0]; // just use the results of our wierd bob
out[1] = best[1];
#else
diff[0] = diff[0] - MIN (diff[0], 10) - 4;
diff[1] = diff[1] - MIN (diff[1] - 10) - 4;
if (diff[0] < 0)
out[0] = weave[0];
else
out[0] = best[0];
if (diff[1] < 0)
out[1] = weave[1];
else
out[1] = best[1];
out[0] = CLAMP (out[0], MinVals[0], MaxVals[0]);
out[1] = CLAMP (out[1], MinVals[1], MaxVals[1]);
#endif
#ifdef USE_VERTICAL_FILTER
pDest[x] = (out[0] + pBob[0]) / 2;
pDest[x + dst_pitchw] = (pBob[src_pitch2] + out[0]) / 2;
pDest[x + 1] = (out[1] + pBob[1]) / 2;
pDest[x + 1 + dst_pitchw] = (pBob[src_pitch2 + 1] + out[1]) / 2;
#else
pDest[x] = out[0];
pDest[x+1] = out[1];
#endif
pBob += 2;
pBobP += 2;
pSrc += 2;
pSrcP += 2;
}
// adjust for next line
pSrc = src_pitch2 * (y+1) + pWeaveSrc;
pSrcP = src_pitch2 * (y+1) + pWeaveSrcP;
pDest = dst_pitch2 * (y+1) + pWeaveDest + dst_pitch2;
if (TopFirst)
{
pBob = pCopySrc + src_pitch2;
pBobP = pCopySrcP + src_pitch2;
}
else
{
pBob = pCopySrc;
pBobP = pCopySrcP;
}
pBob += src_pitch2 * (y+1);
pBobP += src_pitch2 * (y+1);
}
return 0;
#endif