public void RunClsVarScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClsVarScenario_Load));
fixed(Vector256 <Double> *pClsVar1 = &_clsVar1)
fixed(Vector256 <Double> *pClsVar2 = &_clsVar2)
{
var result = Avx.Max(
Avx.LoadVector256((Double *)(pClsVar1)),
Avx.LoadVector256((Double *)(pClsVar2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
}
unsafe private static void greyLinearToGreyFloat(byte *ipstart, byte *opstart, int cb)
{
float *ip = (float *)ipstart, ipe = (float *)(ipstart + cb), op = (float *)opstart;
#if HWINTRINSICS
if (Avx.IsSupported)
{
var vzero = Vector256 <float> .Zero;
ipe -= Vector256 <float> .Count;
while (ip <= ipe)
{
var v = Avx.Max(vzero, Avx.LoadVector256(ip));
ip += Vector256 <float> .Count;
v = Avx.Sqrt(v);
Avx.Store(op, v);
op += Vector256 <float> .Count;
}
ipe += Vector256 <float> .Count;
}
else
#endif
{
var vzero = Vector <float> .Zero;
ipe -= VectorF.Count;
while (ip <= ipe)
{
var v = Unsafe.ReadUnaligned <VectorF>(ip);
ip += VectorF.Count;
v = Vector.SquareRoot(Vector.Max(v, vzero));
Unsafe.WriteUnaligned(op, v);
op += VectorF.Count;
}
ipe += VectorF.Count;
}
float fmin = Vector4.Zero.X;
while (ip < ipe)
{
*op++ = MaxF(*ip++, fmin).Sqrt();
}
}
public void RunClassLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario_Load));
var test = new SimpleBinaryOpTest__MaxSingle();
fixed(Vector256 <Single> *pFld1 = &test._fld1)
fixed(Vector256 <Single> *pFld2 = &test._fld2)
{
var result = Avx.Max(
Avx.LoadVector256((Single *)(pFld1)),
Avx.LoadVector256((Single *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
}
private static unsafe double[] BilinearInterpol_AVX(
double[] x,
double[] A,
double minXA,
double maxXA,
double[] B,
double minXB,
double maxXB,
double weightB)
{
double[] z = new double[outputVectorSize];
fixed(double *pX = &x[0], pA = &A[0], pB = &B[0], pZ = &z[0])
{
Vector256 <double> vWeightB = Vector256.Create(weightB);
Vector256 <double> vWeightA = Vector256.Create(1 - weightB);
Vector256 <double> vMinXA = Vector256.Create(minXA);
Vector256 <double> vMaxXA = Vector256.Create(maxXA);
Vector256 <double> vMinXB = Vector256.Create(minXB);
Vector256 <double> vMaxXB = Vector256.Create(maxXB);
double deltaA = (maxXA - minXA) / (double)(A.Length - 1);
double deltaB = (maxXB - minXB) / (double)(B.Length - 1);
Vector256 <double> vDeltaA = Vector256.Create(deltaA);
Vector256 <double> vDeltaB = Vector256.Create(deltaB);
double invDeltaA = 1.0 / deltaA;
double invDeltaB = 1.0 / deltaB;
Vector256 <double> vInvDeltaA = Vector256.Create(invDeltaA);
Vector256 <double> vInvDeltaB = Vector256.Create(invDeltaB);
Vector128 <int> ALengthMinusOne = Vector128.Create(A.Length - 1);
Vector128 <int> BLengthMinusOne = Vector128.Create(B.Length - 1);
Vector128 <int> One = Vector128.Create(1);
for (var i = 0; i < x.Length; i += Vector256 <double> .Count)
{
Vector256 <double> currentX = Avx.LoadVector256(pX + i);
// Determine the largest a, such that A[i] = f(xA) and xA <= x[i].
// This involves casting from double to int; here we use a Vector conversion.
Vector256 <double> aDouble = Avx.Multiply(Avx.Subtract(currentX, vMinXA), vInvDeltaA);
Vector128 <int> a = Avx.ConvertToVector128Int32WithTruncation(aDouble);
a = Sse41.Min(Sse41.Max(a, Vector128 <int> .Zero), ALengthMinusOne);
Vector128 <int> aPlusOne = Sse41.Min(Sse2.Add(a, One), ALengthMinusOne);
// Now, get the reference input, xA, for our index a.
// This involves casting from int to double.
Vector256 <double> xA = Avx.Add(Avx.Multiply(Avx.ConvertToVector256Double(a), vDeltaA), vMinXA);
// Now, compute the lambda for our A reference point.
Vector256 <double> currentXNormA = Avx.Max(vMinXA, Avx.Min(currentX, vMaxXA));
Vector256 <double> lambdaA = Avx.Multiply(Avx.Subtract(currentXNormA, xA), vInvDeltaA);
// Now, we need to load up our reference points using Vector Gather operations.
Vector256 <double> AVector = Avx2.GatherVector256(pA, a, 8);
Vector256 <double> AVectorPlusOne = Avx2.GatherVector256(pA, aPlusOne, 8);
// Now, do the all of the above for our B reference point.
Vector256 <double> bDouble = Avx.Multiply(Avx.Subtract(currentX, vMinXB), vInvDeltaB);
Vector128 <int> b = Avx.ConvertToVector128Int32WithTruncation(bDouble);
b = Sse41.Min(Sse41.Max(b, Vector128 <int> .Zero), BLengthMinusOne);
Vector128 <int> bPlusOne = Sse41.Min(Sse2.Add(b, One), BLengthMinusOne);
Vector256 <double> xB = Avx.Add(Avx.Multiply(Avx.ConvertToVector256Double(b), vDeltaB), vMinXB);
Vector256 <double> currentXNormB = Avx.Max(vMinXB, Avx.Min(currentX, vMaxXB));
Vector256 <double> lambdaB = Avx.Multiply(Avx.Subtract(currentXNormB, xB), vInvDeltaB);
Vector256 <double> BVector = Avx2.GatherVector256(pB, b, 8);
Vector256 <double> BVectorPlusOne = Avx2.GatherVector256(pB, bPlusOne, 8);
Vector256 <double> newZ = Avx.Add(Avx.Multiply(vWeightA, Avx.Add(AVector, Avx.Multiply(lambdaA, Avx.Subtract(AVectorPlusOne, AVector)))),
Avx.Multiply(vWeightB, Avx.Add(BVector, Avx.Multiply(lambdaB, Avx.Subtract(BVectorPlusOne, BVector)))));
Avx.Store(pZ + i, newZ);
}
}
return(z);
}
public static f32 Max_f32(f32 a, f32 b)
{
return(Avx.Max(a, b));
}
unsafe public static void ConvertFloat3A(byte *ipstart, byte *opstart, float *lutstart, int lutmax, int cb)
{
Debug.Assert(ipstart == opstart);
float *ip = (float *)ipstart, ipe = (float *)(ipstart + cb);
float *lp = lutstart;
#if HWINTRINSICS
if (Avx2.IsSupported)
{
var vgmsk = Avx.BroadcastVector128ToVector256((float *)Unsafe.AsPointer(ref MemoryMarshal.GetReference(HWIntrinsics.GatherMask3x)));
var vgmax = Vector256.Create((float)lutmax);
var vzero = Vector256 <float> .Zero;
var vfone = Vector256.Create(1f);
var vione = Vector256.Create(1);
ipe -= Vector256 <float> .Count;
while (ip <= ipe)
{
var vf = Avx.Max(vzero, Avx.LoadVector256(ip));
var va = Avx.Shuffle(vf, vf, HWIntrinsics.ShuffleMaskAlpha);
vf = Avx.Multiply(vf, Avx.Multiply(vgmax, Avx.Reciprocal(va)));
vf = Avx.Min(vf, vgmax);
var vi = Avx.ConvertToVector256Int32WithTruncation(vf);
var vfi = Avx.ConvertToVector256Single(vi);
var vl = Avx2.GatherMaskVector256(vfone, lp, vi, vgmsk, sizeof(float));
var vh = Avx2.GatherMaskVector256(vfone, lp, Avx2.Add(vi, vione), vgmsk, sizeof(float));
vf = HWIntrinsics.Lerp(vl, vh, Avx.Subtract(vf, vfi));
vf = Avx.Multiply(vf, va);
Avx.Store(ip, vf);
ip += Vector256 <float> .Count;
}
ipe += Vector256 <float> .Count;
}
#endif
{
var vlmax = new Vector4(lutmax);
var vzero = Vector4.Zero;
float famin = new Vector4(1 / 1024f).X;
while (ip < ipe)
{
var vf = Unsafe.ReadUnaligned <Vector4>(ip);
float f3 = vf.W;
if (f3 < famin)
{
Unsafe.WriteUnaligned(ip, vzero);
}
else
{
vf = (vf * vlmax / f3).Clamp(vzero, vlmax);
float f0 = vf.X;
float f1 = vf.Y;
float f2 = vf.Z;
uint i0 = (uint)f0;
uint i1 = (uint)f1;
uint i2 = (uint)f2;
ip[0] = Lerp(lp[i0], lp[i0 + 1], f0 - (int)i0) * f3;
ip[1] = Lerp(lp[i1], lp[i1 + 1], f1 - (int)i1) * f3;
ip[2] = Lerp(lp[i2], lp[i2 + 1], f2 - (int)i2) * f3;
}
ip += 4;
}
}
}
unsafe public static void ConvertFloat(byte *ipstart, byte *opstart, float *lutstart, int lutmax, int cb)
{
Debug.Assert(ipstart == opstart);
float *ip = (float *)ipstart, ipe = (float *)(ipstart + cb);
float *lp = lutstart;
#if HWINTRINSICS
if (Avx2.IsSupported)
{
var vlmax = Vector256.Create((float)lutmax);
var vzero = Vector256 <float> .Zero;
var vione = Vector256.Create(1);
ipe -= Vector256 <float> .Count;
while (ip <= ipe)
{
var vf = Avx.Multiply(vlmax, Avx.LoadVector256(ip));
vf = Avx.Min(Avx.Max(vzero, vf), vlmax);
var vi = Avx.ConvertToVector256Int32WithTruncation(vf);
var vp = Avx.ConvertToVector256Single(vi);
var vl = Avx2.GatherVector256(lp, vi, sizeof(float));
var vh = Avx2.GatherVector256(lp, Avx2.Add(vi, vione), sizeof(float));
vf = HWIntrinsics.Lerp(vl, vh, Avx.Subtract(vf, vp));
Avx.Store(ip, vf);
ip += Vector256 <float> .Count;
}
ipe += Vector256 <float> .Count;
float fmin = vzero.ToScalar(), flmax = vlmax.ToScalar();
while (ip < ipe)
{
float f = (*ip * flmax).Clamp(fmin, flmax);
uint i = (uint)f;
*ip++ = Lerp(lp[i], lp[i + 1], f - i);
}
}
else
#endif
{
var vlmax = new Vector4(lutmax);
var vzero = Vector4.Zero;
ipe -= 4;
while (ip <= ipe)
{
var vf = (Unsafe.ReadUnaligned <Vector4>(ip) * vlmax).Clamp(vzero, vlmax);
float f0 = vf.X;
float f1 = vf.Y;
float f2 = vf.Z;
float f3 = vf.W;
uint i0 = (uint)f0;
uint i1 = (uint)f1;
uint i2 = (uint)f2;
uint i3 = (uint)f3;
ip[0] = Lerp(lp[i0], lp[i0 + 1], f0 - (int)i0);
ip[1] = Lerp(lp[i1], lp[i1 + 1], f1 - (int)i1);
ip[2] = Lerp(lp[i2], lp[i2 + 1], f2 - (int)i2);
ip[3] = Lerp(lp[i3], lp[i3 + 1], f3 - (int)i3);
ip += 4;
}
ipe += 4;
float fmin = vzero.X, flmax = vlmax.X;
while (ip < ipe)
{
float f = (*ip * flmax).Clamp(fmin, flmax);
uint i = (uint)f;
*ip++ = Lerp(lp[i], lp[i + 1], f - i);
}
}
}
unsafe void IConversionProcessor.ConvertLine(byte *ipstart, byte *opstart, int cb)
{
float *ip = (float *)ipstart, ipe = (float *)(ipstart + cb);
byte * op = opstart;
#if HWINTRINSICS
if (Avx2.IsSupported)
{
var vzero = Vector256 <float> .Zero;
var vmin = Vector256.Create(0.5f / byte.MaxValue);
var vscale = Vector256.Create((float)byte.MaxValue);
var vmaskp = Avx.LoadVector256((int *)Unsafe.AsPointer(ref MemoryMarshal.GetReference(HWIntrinsics.PermuteMaskDeinterleave8x32)));
ipe -= Vector256 <byte> .Count;
while (ip <= ipe)
{
var vf0 = Avx.LoadVector256(ip);
var vf1 = Avx.LoadVector256(ip + Vector256 <float> .Count);
var vf2 = Avx.LoadVector256(ip + Vector256 <float> .Count * 2);
var vf3 = Avx.LoadVector256(ip + Vector256 <float> .Count * 3);
ip += Vector256 <byte> .Count;
var vfa0 = Avx.Shuffle(vf0, vf0, HWIntrinsics.ShuffleMaskAlpha);
var vfa1 = Avx.Shuffle(vf1, vf1, HWIntrinsics.ShuffleMaskAlpha);
var vfa2 = Avx.Shuffle(vf2, vf2, HWIntrinsics.ShuffleMaskAlpha);
var vfa3 = Avx.Shuffle(vf3, vf3, HWIntrinsics.ShuffleMaskAlpha);
vfa0 = Avx.Max(vfa0, vmin);
vfa1 = Avx.Max(vfa1, vmin);
vfa2 = Avx.Max(vfa2, vmin);
vfa3 = Avx.Max(vfa3, vmin);
vf0 = Avx.Multiply(vf0, Avx.Reciprocal(vfa0));
vf1 = Avx.Multiply(vf1, Avx.Reciprocal(vfa1));
vf2 = Avx.Multiply(vf2, Avx.Reciprocal(vfa2));
vf3 = Avx.Multiply(vf3, Avx.Reciprocal(vfa3));
vf0 = Avx.Blend(vf0, vfa0, HWIntrinsics.BlendMaskAlpha);
vf1 = Avx.Blend(vf1, vfa1, HWIntrinsics.BlendMaskAlpha);
vf2 = Avx.Blend(vf2, vfa2, HWIntrinsics.BlendMaskAlpha);
vf3 = Avx.Blend(vf3, vfa3, HWIntrinsics.BlendMaskAlpha);
vf0 = Avx.BlendVariable(vf0, vzero, HWIntrinsics.AvxCompareEqual(vfa0, vmin));
vf1 = Avx.BlendVariable(vf1, vzero, HWIntrinsics.AvxCompareEqual(vfa1, vmin));
vf2 = Avx.BlendVariable(vf2, vzero, HWIntrinsics.AvxCompareEqual(vfa2, vmin));
vf3 = Avx.BlendVariable(vf3, vzero, HWIntrinsics.AvxCompareEqual(vfa3, vmin));
vf0 = Avx.Multiply(vf0, vscale);
vf1 = Avx.Multiply(vf1, vscale);
vf2 = Avx.Multiply(vf2, vscale);
vf3 = Avx.Multiply(vf3, vscale);
var vi0 = Avx.ConvertToVector256Int32(vf0);
var vi1 = Avx.ConvertToVector256Int32(vf1);
var vi2 = Avx.ConvertToVector256Int32(vf2);
var vi3 = Avx.ConvertToVector256Int32(vf3);
var vs0 = Avx2.PackSignedSaturate(vi0, vi1);
var vs1 = Avx2.PackSignedSaturate(vi2, vi3);
var vb0 = Avx2.PackUnsignedSaturate(vs0, vs1);
vb0 = Avx2.PermuteVar8x32(vb0.AsInt32(), vmaskp).AsByte();
Avx.Store(op, vb0);
op += Vector256 <byte> .Count;
}
ipe += Vector256 <byte> .Count;
}
else if (Sse41.IsSupported)
{
var vzero = Vector128 <float> .Zero;
var vmin = Vector128.Create(0.5f / byte.MaxValue);
var vscale = Vector128.Create((float)byte.MaxValue);
ipe -= Vector128 <byte> .Count;
while (ip <= ipe)
{
var vf0 = Sse.LoadVector128(ip);
var vf1 = Sse.LoadVector128(ip + Vector128 <float> .Count);
var vf2 = Sse.LoadVector128(ip + Vector128 <float> .Count * 2);
var vf3 = Sse.LoadVector128(ip + Vector128 <float> .Count * 3);
ip += Vector128 <byte> .Count;
var vfa0 = Sse.Shuffle(vf0, vf0, HWIntrinsics.ShuffleMaskAlpha);
var vfa1 = Sse.Shuffle(vf1, vf1, HWIntrinsics.ShuffleMaskAlpha);
var vfa2 = Sse.Shuffle(vf2, vf2, HWIntrinsics.ShuffleMaskAlpha);
var vfa3 = Sse.Shuffle(vf3, vf3, HWIntrinsics.ShuffleMaskAlpha);
vfa0 = Sse.Max(vfa0, vmin);
vfa1 = Sse.Max(vfa1, vmin);
vfa2 = Sse.Max(vfa2, vmin);
vfa3 = Sse.Max(vfa3, vmin);
vf0 = Sse.Multiply(vf0, Sse.Reciprocal(vfa0));
vf1 = Sse.Multiply(vf1, Sse.Reciprocal(vfa1));
vf2 = Sse.Multiply(vf2, Sse.Reciprocal(vfa2));
vf3 = Sse.Multiply(vf3, Sse.Reciprocal(vfa3));
vf0 = Sse41.Blend(vf0, vfa0, HWIntrinsics.BlendMaskAlpha);
vf1 = Sse41.Blend(vf1, vfa1, HWIntrinsics.BlendMaskAlpha);
vf2 = Sse41.Blend(vf2, vfa2, HWIntrinsics.BlendMaskAlpha);
vf3 = Sse41.Blend(vf3, vfa3, HWIntrinsics.BlendMaskAlpha);
vf0 = Sse41.BlendVariable(vf0, vzero, Sse.CompareEqual(vfa0, vmin));
vf1 = Sse41.BlendVariable(vf1, vzero, Sse.CompareEqual(vfa1, vmin));
vf2 = Sse41.BlendVariable(vf2, vzero, Sse.CompareEqual(vfa2, vmin));
vf3 = Sse41.BlendVariable(vf3, vzero, Sse.CompareEqual(vfa3, vmin));
vf0 = Sse.Multiply(vf0, vscale);
vf1 = Sse.Multiply(vf1, vscale);
vf2 = Sse.Multiply(vf2, vscale);
vf3 = Sse.Multiply(vf3, vscale);
var vi0 = Sse2.ConvertToVector128Int32(vf0);
var vi1 = Sse2.ConvertToVector128Int32(vf1);
var vi2 = Sse2.ConvertToVector128Int32(vf2);
var vi3 = Sse2.ConvertToVector128Int32(vf3);
var vs0 = Sse2.PackSignedSaturate(vi0, vi1);
var vs1 = Sse2.PackSignedSaturate(vi2, vi3);
var vb0 = Sse2.PackUnsignedSaturate(vs0, vs1);
Sse2.Store(op, vb0);
op += Vector128 <byte> .Count;
}
ipe += Vector128 <byte> .Count;
}
#endif
float fmax = new Vector4(byte.MaxValue).X, fround = new Vector4(0.5f).X, fmin = fround / fmax;
while (ip < ipe)
{
float f3 = ip[3];
if (f3 < fmin)
{
*(uint *)op = 0;
}
else
{
float f3i = fmax / f3;
byte o0 = ClampToByte((int)(ip[0] * f3i + fround));
byte o1 = ClampToByte((int)(ip[1] * f3i + fround));
byte o2 = ClampToByte((int)(ip[2] * f3i + fround));
byte o3 = ClampToByte((int)(f3 * fmax + fround));
op[0] = o0;
op[1] = o1;
op[2] = o2;
op[3] = o3;
}
ip += 4;
op += 4;
}
}
unsafe void IConvolver.SharpenLine(byte *cstart, byte *ystart, byte *bstart, byte *ostart, int ox, int ow, float amt, float thresh, bool gamma)
{
float *ip = (float *)cstart + (uint)ox * channels, yp = (float *)ystart + (uint)ox, bp = (float *)bstart, op = (float *)ostart;
float *ipe = ip + (uint)ow * channels;
bool threshold = thresh > 0f;
if (Avx.IsSupported && ip <= ipe - VectorAvx.Count)
{
var vthresh = Vector256.Create(threshold ? thresh : -1f);
var vmsk = Vector256.Create(0x7fffffff).AsSingle();
var vamt = Vector256.Create(amt);
var vmin = VectorAvx.Zero;
ipe -= VectorAvx.Count;
do
{
var vd = Avx.Subtract(Avx.LoadVector256(yp), Avx.LoadVector256(bp));
yp += VectorAvx.Count;
bp += VectorAvx.Count;
if (threshold)
{
var sm = HWIntrinsics.AvxCompareGreaterThan(Avx.And(vd, vmsk), vthresh);
vd = Avx.And(vd, sm);
}
vd = Avx.Multiply(vd, vamt);
var v0 = Avx.LoadVector256(ip);
ip += VectorAvx.Count;
if (gamma)
{
v0 = Avx.Max(v0, vmin);
v0 = Avx.Multiply(v0, Avx.ReciprocalSqrt(v0));
v0 = Avx.Add(v0, vd);
v0 = Avx.Max(v0, vmin);
v0 = Avx.Multiply(v0, v0);
}
else
{
v0 = Avx.Add(v0, vd);
}
Avx.Store(op, v0);
op += VectorAvx.Count;
} while (ip <= ipe);
ipe += VectorAvx.Count;
}
else if (ip <= ipe - VectorSse.Count)
{
var vthresh = Vector128.Create(threshold ? thresh : -1f);
var vmsk = Vector128.Create(0x7fffffff).AsSingle();
var vamt = Vector128.Create(amt);
var vmin = VectorSse.Zero;
ipe -= VectorSse.Count;
do
{
var vd = Sse.Subtract(Sse.LoadVector128(yp), Sse.LoadVector128(bp));
yp += VectorSse.Count;
bp += VectorSse.Count;
if (threshold)
{
var sm = Sse.CompareGreaterThan(Sse.And(vd, vmsk), vthresh);
vd = Sse.And(vd, sm);
}
vd = Sse.Multiply(vd, vamt);
var v0 = Sse.LoadVector128(ip);
ip += VectorSse.Count;
if (gamma)
{
v0 = Sse.Max(v0, vmin);
v0 = Sse.Multiply(v0, Sse.ReciprocalSqrt(v0));
v0 = Sse.Add(v0, vd);
v0 = Sse.Max(v0, vmin);
v0 = Sse.Multiply(v0, v0);
}
else
{
v0 = Sse.Add(v0, vd);
}
Sse.Store(op, v0);
op += VectorSse.Count;
} while (ip <= ipe);
ipe += VectorSse.Count;
}
float fmin = VectorSse.Zero.ToScalar();
while (ip < ipe)
{
float dif = *yp++ - *bp++;
float c0 = *ip++;
if (!threshold || Math.Abs(dif) > thresh)
{
dif *= amt;
if (gamma)
{
c0 = MathUtil.MaxF(c0, fmin).Sqrt();
c0 = MathUtil.MaxF(c0 + dif, fmin);
c0 *= c0;
}
else
{
c0 += dif;
}
}
*op++ = c0;
}
}
public static unsafe float Max(this Matrix <float> matrix)
{
var i = 0;
fixed(float *ptr = matrix.GetArray())
{
var span = new Span <float>(ptr, matrix.Length);
var maxScalar = span[0];
if (Avx.IsSupported)
{
var maxValues = stackalloc float[8]
{
span[0],
span[0],
span[0],
span[0],
span[0],
span[0],
span[0],
span[0]
};
var max = Avx.LoadVector256(maxValues);
while (i < span.Length - 8)
{
var vector256 = Avx.LoadVector256(ptr + i);
max = Avx.Max(vector256, max);
i += 8;
}
maxScalar = max.MaxVector256(8);
}
else if (Sse.IsSupported)
{
var maxValues = stackalloc float[4]
{
span[0],
span[0],
span[0],
span[0]
};
var max = Sse.LoadVector128(maxValues);
while (i < span.Length - 4)
{
var vector128 = Sse.LoadVector128(ptr + i);
max = Sse.Max(vector128, max);
i += 4;
}
maxScalar = max.MaxVector128(4);
}
while (i < span.Length)
{
if (maxScalar < span[i])
{
maxScalar = span[i];
}
i++;
}
return(maxScalar);
}
}