unsafe private static void denoiseLineSse2(byte *pcurr, byte *pprev, byte *pnext, int cb)
{
byte *ip = pcurr, pp = pprev, np = pnext;
nuint cnt = 0, end = (nuint)cb - (nuint)Vector128 <byte> .Count;
var voffset = Vector128.Create((byte)0x80);
var vthresh = Vector128.Create(denoiseThreshold);
LoopTop:
do
{
var vcurr = Sse2.LoadVector128(ip + cnt);
var vprev = Sse2.LoadVector128(pp + cnt);
var vnext = Sse2.LoadVector128(np + cnt);
var vdiffp = Sse2.Or(Sse2.SubtractSaturate(vcurr, vprev), Sse2.SubtractSaturate(vprev, vcurr));
var vmaskp = Sse2.CompareEqual(Sse2.Max(vdiffp, vthresh), vthresh);
var vdiffn = Sse2.Or(Sse2.SubtractSaturate(vcurr, vnext), Sse2.SubtractSaturate(vnext, vcurr));
var vmaskn = Sse2.CompareEqual(Sse2.Max(vdiffn, vthresh), vthresh);
var vavgp = Sse2.Average(vcurr, vprev);
var vavgn = Sse2.Average(vcurr, vnext);
var voutval = Sse2.Average(HWIntrinsics.BlendVariable(vavgn, vavgp, vmaskp), HWIntrinsics.BlendVariable(vavgp, vavgn, vmaskn));
var voutmsk = Sse2.Or(vmaskp, vmaskn);
voutval = Sse2.Average(voutval, HWIntrinsics.BlendVariable(voutval, Sse2.Average(vprev, vnext), Sse2.And(vmaskp, vmaskn)));
var vcurrs = Sse2.Xor(vcurr, voffset).AsSByte();
var vprevs = Sse2.Xor(vprev, voffset).AsSByte();
var vnexts = Sse2.Xor(vnext, voffset).AsSByte();
var vsurlt = Sse2.And(Sse2.CompareGreaterThan(vcurrs, vprevs), Sse2.CompareGreaterThan(vcurrs, vnexts));
var vsurgt = Sse2.And(Sse2.CompareGreaterThan(vprevs, vcurrs), Sse2.CompareGreaterThan(vnexts, vcurrs));
voutmsk = Sse2.And(voutmsk, Sse2.Or(vsurlt, vsurgt).AsByte());
voutval = HWIntrinsics.BlendVariable(vcurr, voutval, voutmsk);
Sse2.Store(ip + cnt, voutval);
cnt += (nuint)Vector128 <byte> .Count;
} while (cnt <= end);
if (cnt < end + (nuint)Vector128 <byte> .Count)
{
cnt = end;
goto LoopTop;
}
}
unsafe private static (uint eql, uint eqr) dedupeLineSse2(byte *pcurr, byte *pprev, byte *penc, int cb, uint bg)
{
byte *ip = pcurr, pp = pprev, op = penc;
nuint cnt = 0, end = (nuint)cb - (nuint)Vector128 <byte> .Count;
bool lfound = false;
uint eql = 0u, eqr = 0u;
var vbg = pp == (byte *)0 ? Vector128.Create(bg) : Vector128 <uint> .Zero;
LoopTop:
do
{
var vprev = pp != (byte *)0 ? Sse2.LoadVector128(pp + cnt).AsUInt32() : vbg;
var vcurr = Sse2.LoadVector128(ip + cnt).AsUInt32();
var veq = Sse2.CompareEqual(vcurr, vprev);
vcurr = HWIntrinsics.BlendVariable(vcurr, vbg, veq);
Sse2.Store(op + cnt, vcurr.AsByte());
cnt += (nuint)Vector128 <byte> .Count;
uint msk = (uint)Sse2.MoveMask(veq.AsByte());
if (msk == ushort.MinValue)
{
lfound = true;
eqr = 0u;
}
else if (msk == ushort.MaxValue)
{
if (!lfound)
{
eql += (uint)Vector128 <uint> .Count;
}
eqr += (uint)Vector128 <uint> .Count;
}
else
{
msk = ~msk;
if (!lfound)
{
eql += (uint)BitOperations.TrailingZeroCount(msk) / sizeof(uint);
lfound = true;
}
eqr = (uint)BitOperations.LeadingZeroCount(msk) / sizeof(uint);
}
} while (cnt <= end);
if (cnt < end + (nuint)Vector128 <byte> .Count)
{
uint offs = (uint)(cnt - end) / sizeof(uint);
if (!lfound)
{
eql -= offs;
}
eqr -= offs;
cnt = end;
goto LoopTop;
}
return(eql, eqr);
}