gstreamer/gst/deinterlace2/tvtime/tomsmocomp/WierdBob.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

286 lines
9.7 KiB
C++

// -*- c++ -*-
// First, get and save our possible Bob values
// Assume our pixels are layed out as follows with x the calc'd bob value
// and the other pixels are from the current field
//
// j a b c k current field
// x calculated line
// m d e f n current field
//
// we calc the bob value as:
// x2 = either avg(a,f), avg(c,d), avg(b,e), avg(j,n), or avg(k,m)
// selected for the smallest of abs(a,f), abs(c,d), or abs(b,e), etc.
#ifndef IS_C
// a,f
"movq -2(%%"XBX"), %%mm0\n\t" // value a from top left
"movq 2(%%"XBX", %%"XCX"), %%mm1\n\t" // value f from bottom right
"movq %%mm0, %%mm6\n\t"
// pavgb %%mm6, %%mm1 // avg(a,f), also best so far
V_PAVGB ("%%mm6", "%%mm1", "%%mm7", _ShiftMask) // avg(a,f), also best so far
"movq %%mm0, %%mm7\n\t"
"psubusb %%mm1, %%mm7\n\t"
"psubusb %%mm0, %%mm1\n\t"
"por %%mm1, %%mm7\n\t" // abs diff, also best so far
// c,d
"movq 2(%%"XBX"), %%mm0\n\t" // value a from top left
"movq -2(%%"XBX", %%"XCX"), %%mm1\n\t" // value f from bottom right
"movq %%mm0, %%mm2\n\t"
// pavgb %%mm2, %%mm1 // avg(c,d)
V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(c,d)
"movq %%mm0, %%mm3\n\t"
"psubusb %%mm1, %%mm3\n\t"
"psubusb %%mm0, %%mm1\n\t"
"por %%mm1, %%mm3\n\t" // abs(c,d)
"movq %%mm3, %%mm1\n\t" // keep copy
"psubusb %%mm7, %%mm3\n\t" // nonzero where new weights bigger, else 0
"pxor %%mm4, %%mm4\n\t"
"pcmpeqb %%mm4, %%mm3\n\t" // now ff where new better, else 00
"pcmpeqb %%mm3, %%mm4\n\t" // here ff where old better, else 00
"pand %%mm3, %%mm1\n\t" // keep only better new avg and abs
"pand %%mm3, %%mm2\n\t"
"pand %%mm4, %%mm6\n\t"
"pand %%mm4, %%mm7\n\t"
"por %%mm2, %%mm6\n\t" // and merge new & old vals keeping best
"por %%mm1, %%mm7\n\t"
"por "_UVMask", %%mm7\n\t" // but we know chroma is worthless so far
"pand "_YMask", %%mm5\n\t" // mask out chroma from here also
// j,n
"movq -4(%%"XBX"), %%mm0\n\t" // value j from top left
"movq 4(%%"XBX", %%"XCX"), %%mm1\n\t" // value n from bottom right
"movq %%mm0, %%mm2\n\t"
// pavgb %%mm2, %%mm1 // avg(j,n)
V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(j,n)
"movq %%mm0, %%mm3\n\t"
"psubusb %%mm1, %%mm3\n\t"
"psubusb %%mm0, %%mm1\n\t"
"por %%mm1, %%mm3\n\t" // abs(j-n)
"movq %%mm3, %%mm1\n\t" // keep copy
"psubusb %%mm7, %%mm3\n\t" // nonzero where new weights bigger, else 0
"pxor %%mm4, %%mm4\n\t"
"pcmpeqb %%mm4, %%mm3\n\t" // now ff where new better, else 00
"pcmpeqb %%mm3, %%mm4\n\t" // here ff where old better, else 00
"pand %%mm3, %%mm1\n\t" // keep only better new avg and abs
"pand %%mm2, %%mm3\n\t"
"pand %%mm4, %%mm6\n\t"
"pand %%mm4, %%mm7\n\t"
"por %%mm3, %%mm6\n\t" // and merge new & old vals keeping best
"por %%mm1, %%mm7\n\t" // "
// k, m
"movq 4(%%"XBX"), %%mm0\n\t" // value k from top right
"movq -4(%%"XBX", %%"XCX"), %%mm1\n\t" // value n from bottom left
"movq %%mm0, %%mm4\n\t"
// pavgb %%mm4, %%mm1 // avg(k,m)
V_PAVGB ("%%mm4", "%%mm1", "%%mm3", _ShiftMask) // avg(k,m)
"movq %%mm0, %%mm3\n\t"
"psubusb %%mm1, %%mm3\n\t"
"psubusb %%mm0, %%mm1\n\t"
"por %%mm1, %%mm3\n\t" // abs(k,m)
"movq %%mm3, %%mm1\n\t" // keep copy
"movq %%mm4, %%mm2\n\t" // avg(k,m)
"psubusb %%mm7, %%mm3\n\t" // nonzero where new weights bigger, else 0
"pxor %%mm4, %%mm4\n\t"
"pcmpeqb %%mm4, %%mm3\n\t" // now ff where new better, else 00
"pcmpeqb %%mm3, %%mm4\n\t" // here ff where old better, else 00
"pand %%mm3, %%mm1\n\t" // keep only better new avg and abs
"pand %%mm2, %%mm3\n\t"
"pand %%mm4, %%mm6\n\t"
"pand %%mm4, %%mm7\n\t"
"por %%mm3, %%mm6\n\t" // and merge new & old vals keeping best
"por %%mm1, %%mm7\n\t" // "
// b,e
"movq (%%"XBX"), %%mm0\n\t" // value b from top
"movq (%%"XBX", %%"XCX"), %%mm1\n\t" // value e from bottom
// We will also calc here the max/min values to later limit comb
// so the max excursion will not exceed the Max_Comb constant
#ifdef SKIP_SEARCH
"movq %%mm0, %%mm2\n\t"
// pminub %%mm2, %%mm1
V_PMINUB ("%%mm2", "%%mm1", "%%mm4")
// pmaxub %%mm6, %%mm2 // clip our current results so far to be above this
V_PMAXUB ("%%mm6", "%%mm2")
"movq %%mm0, %%mm2\n\t"
V_PMAXUB ("%%mm2", "%%mm1")
// pminub %%mm6, %%mm2 // clip our current results so far to be below this
V_PMINUB ("%%mm6", "%%mm2", "%%mm4")
#else
"movq %%mm0, %%mm2\n\t"
"movq (%%"XAX"), %%mm4\n\t"
"psubusb %%mm4, %%mm2\n\t"
"psubusb %%mm0, %%mm4\n\t"
"por %%mm2, %%mm4\n\t" // abs diff
"movq %%mm1, %%mm2\n\t"
"movq (%%"XAX", %%"XCX"), %%mm3\n\t"
"psubusb %%mm3, %%mm2\n\t"
"psubusb %%mm1, %%mm3\n\t"
"por %%mm2, %%mm3\n\t" // abs diff
// pmaxub %%mm3, %%mm4 // top or bottom pixel moved most
V_PMAXUB ("%%mm3", "%%mm4") // top or bottom pixel moved most
"psubusb "_Max_Mov", %%mm3\n\t" // moved more than allowed? or goes to 0?
"pxor %%mm4, %%mm4\n\t"
"pcmpeqb %%mm4, %%mm3\n\t" // now ff where low motion, else high motion
"movq %%mm0, %%mm2\n\t"
// pminub %%mm2, %%mm1
V_PMINUB ("%%mm2", "%%mm1", "%%mm4")
// pmaxub %%mm6, %%mm2 // clip our current results so far to be above this
V_PMAXUB ("%%mm6", "%%mm2")
"psubusb %%mm3, %%mm2\n\t" // maybe decrease it to 0000.. if no surround motion
"movq %%mm2, "_Min_Vals"\n\t"
"movq %%mm0, %%mm2\n\t"
V_PMAXUB ("%%mm2", "%%mm1")
// pminub %%mm6, %%mm2 // clip our current results so far to be below this
V_PMINUB ("%%mm6", "%%mm2", "%%mm4")
"paddusb %%mm3, %%mm2\n\t" // maybe increase it to ffffff if no surround motion
"movq %%mm2, "_Max_Vals"\n\t"
#endif
"movq %%mm0, %%mm2\n\t"
// pavgb %%mm2, %%mm1 // avg(b,e)
V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(b,e)
"movq %%mm0, %%mm3\n\t"
"psubusb %%mm1, %%mm3\n\t"
"psubusb %%mm0, %%mm1\n\t"
"por %%mm1, %%mm3\n\t" // abs(c,d)
"movq %%mm3, %%mm1\n\t" // keep copy of diffs
"pxor %%mm4, %%mm4\n\t"
"psubusb %%mm7, %%mm3\n\t" // nonzero where new weights bigger, else 0
"pcmpeqb %%mm4, %%mm3\n\t" // now ff where new better, else 00
"pcmpeqb %%mm3, %%mm4\n\t" // here ff where old better, else 00
"pand %%mm3, %%mm1\n\t"
"pand %%mm3, %%mm2\n\t"
"pand %%mm4, %%mm6\n\t"
"pand %%mm4, %%mm7\n\t"
"por %%mm2, %%mm6\n\t" // our x2 value
"por %%mm1, %%mm7\n\t" // our x2 diffs
"movq %%mm7, %%mm4\n\t" // save as bob uncertainty indicator
#else
// a,f
best[0] = (pBob[-2] + pBob[src_pitch2 + 2]) / 2;
diff[0] = ABS (pBob[-2] - pBob[src_pitch2 + 2]);
best[1] = (pBob[-1] + pBob[src_pitch2 + 3]) / 2;
diff[1] = ABS (pBob[-1] - pBob[src_pitch2 + 3]);
// c,d
if (ABS (pBob[2] - pBob[src_pitch2 - 2]) < diff[0]) {
best[0] = (pBob[2] + pBob[src_pitch2 - 2]) / 2;
diff[0] = ABS (pBob[2] - pBob[src_pitch2 - 2]);
}
if (ABS (pBob[3] - pBob[src_pitch2 - 1]) < diff[1]) {
best[1] = (pBob[3] + pBob[src_pitch2 - 1]) / 2;
diff[1] = ABS (pBob[3] - pBob[src_pitch2 - 1]);
}
// j,n
if (ABS (pBob[-4] - pBob[src_pitch2 + 4]) < diff[0]) {
best[0] = (pBob[-4] + pBob[src_pitch2 + 4]) / 2;
diff[0] = ABS (pBob[-4] - pBob[src_pitch2 + 4]);
}
if (ABS (pBob[-3] - pBob[src_pitch2 + 5]) < diff[1]) {
best[1] = (pBob[-3] + pBob[src_pitch2 + 5]) / 2;
diff[1] = ABS (pBob[-3] - pBob[src_pitch2 + 5]);
}
// k,m
if (ABS (pBob[4] - pBob[src_pitch2 - 4]) < diff[0]) {
best[0] = (pBob[4] + pBob[src_pitch2 - 4]) / 2;
diff[0] = ABS (pBob[-4] - pBob[src_pitch2 - 4]);
}
if (ABS (pBob[5] - pBob[src_pitch2 - 3]) < diff[1]) {
best[1] = (pBob[5] + pBob[src_pitch2 - 3]) / 2;
diff[1] = ABS (pBob[-3] - pBob[src_pitch2 - 3]);
}
// k,m
if (ABS (pBob[4] - pBob[src_pitch2 - 4]) < diff[0]) {
best[0] = (pBob[4] + pBob[src_pitch2 - 4]) / 2;
diff[0] = ABS (pBob[-4] - pBob[src_pitch2 - 4]);
}
if (ABS (pBob[5] - pBob[src_pitch2 - 3]) < diff[1]) {
best[1] = (pBob[5] + pBob[src_pitch2 - 3]) / 2;
diff[1] = ABS (pBob[-3] - pBob[src_pitch2 - 3]);
}
// We will also calc here the max/min values to later limit comb
// so the max excursion will not exceed the Max_Comb constant
#ifdef SKIP_SEARCH
best[0] = CLAMP (best[0], MIN (pBob[src_pitch2], pBob[0]), MAX (pBob[src_pitch2], pBob[0]));
best[1] = CLAMP (best[1], MIN (pBob[src_pitch2 + 1], pBob[1]), MAX (pBob[src_pitch2 + 1], pBob[1]));
#else
mov[0] = MAX (ABS (pBob[0] - pBobP[0]), ABS (pBob[src_pitch2] - pBobP[src_pitch2]));
mov[1] = MAX (ABS (pBob[1] - pBobP[1]), ABS (pBob[src_pitch2 + 1] - pBobP[src_pitch2 + 1]));
MinVals[0] = 0;
MinVals[1] = 0;
MaxVals[0] = 255;
MaxVals[1] = 255;
if (mov[0] > Max_Mov[0]) {
MinVals[0] = MAX (MIN (pBob[0], pBob[src_pitch2]), best[0]);
MaxVals[0] = MIN (MAX (pBob[0], pBob[src_pitch2]), best[0]);
}
if (mov[1] > Max_Mov[1]) {
MinVals[1] = MAX (MIN (pBob[1], pBob[src_pitch2 + 1]), best[1]);
MaxVals[1] = MIN (MAX (pBob[1], pBob[src_pitch2 + 1]), best[1]);
}
best[0] = CLAMP (best[0], MIN (pBob[src_pitch2], pBob[0]), MAX (pBob[src_pitch2], pBob[0]));
best[1] = CLAMP (best[1], MIN (pBob[src_pitch2 + 1], pBob[1]), MAX (pBob[src_pitch2 + 1], pBob[1]));
#endif
avg[0] = (pBob[src_pitch2] + pBob[0]) / 2;
avg[1] = (pBob[src_pitch2 + 1] + pBob[1]) / 2;
diff2[0] = ABS (pBob[src_pitch2] - pBob[0]);
diff2[1] = ABS (pBob[src_pitch2 + 1] - pBob[1]);
if (diff2[0] < diff[0]) {
best[0] = avg[0];
diff[0] = diff2[0];
}
if (diff2[1] < diff[1]) {
best[1] = avg[1];
diff[1] = diff2[1];
}
#endif