static unsafe int Main(string[] args)
{
int testResult = Pass;
if (Avx.IsSupported)
{
using (TestTable <float> floatTable = new TestTable <float>(new float[8] {
22, -1, -50, 0, 22, -1, -50, 0
}, new float[8] {
22, -1, -50, 0, 22, -1, -50, 0
}, new float[8]))
using (TestTable <double> doubleTable = new TestTable <double>(new double[4] {
1, -5, 100, 0
}, new double[4] {
22, -1, -50, 0
}, new double[4]))
{
var vf1 = Unsafe.Read <Vector256 <float> >(floatTable.inArray1Ptr);
var vf2 = Unsafe.Read <Vector256 <float> >(floatTable.inArray2Ptr);
var vf3 = Avx.HorizontalAdd(vf1, vf2);
Unsafe.Write(floatTable.outArrayPtr, vf3);
if (!floatTable.CheckResult((left, right, result) =>
(left[0] + left[1] == result[0]) && (right[0] + right[1] == result[2]) &&
(left[2] + left[3] == result[1]) && (right[2] + right[3] == result[3]) &&
(left[4] + left[5] == result[4]) && (right[4] + right[5] == result[6]) &&
(left[6] + left[7] == result[5]) && (right[6] + right[7] == result[7])))
{
Console.WriteLine("Avx HorizontalAdd failed on float:");
foreach (var item in floatTable.outArray)
{
Console.Write(item + ", ");
}
Console.WriteLine();
testResult = Fail;
}
var vd1 = Unsafe.Read <Vector256 <double> >(doubleTable.inArray1Ptr);
var vd2 = Unsafe.Read <Vector256 <double> >(doubleTable.inArray2Ptr);
var vd3 = Avx.HorizontalAdd(vd1, vd2);
Unsafe.Write(doubleTable.outArrayPtr, vd3);
if (!doubleTable.CheckResult((left, right, result) =>
(left[0] + left[1] == result[0]) && (right[0] + right[1] == result[1]) &&
(left[2] + left[3] == result[2]) && (right[2] + right[3] == result[3])))
{
Console.WriteLine("Avx HorizontalAdd failed on double:");
foreach (var item in doubleTable.outArray)
{
Console.Write(item + ", ");
}
Console.WriteLine();
testResult = Fail;
}
}
}
return(testResult);
}
public static Span <Complex> Multiply(ReadOnlySpan <Complex> left, ReadOnlySpan <Complex> right)
{
var result = new Complex[Math.Min(left.Length, right.Length)].AsSpan();
var vectorRes = MemoryMarshal.Cast <Complex, Vector256 <double> >(result);
var vectorLeft = MemoryMarshal.Cast <Complex, Vector256 <double> >(left);
var vectorRight = MemoryMarshal.Cast <Complex, Vector256 <double> >(right);
for (int i = 0; i < vectorRes.Length; i++)
{
var l = vectorLeft[i];
var r = vectorRight[i];
vectorRes[i] = Avx.HorizontalAdd(
Avx.Multiply(
Avx.Multiply(l, r),
Vector256.Create(1.0, -1.0, 1.0, -1.0)),
Avx.Multiply(
l,
Avx.Permute(r, 0b0101)
));
}
for (int i = 2 * vectorRes.Length; i < result.Length; i++)
{
result[i] = left[i] * right[i];
}
return(result);
}
public static Vector256 <double> HorizontalAdd(Vector256 <double> left, Vector256 <double> right)
{
if (Avx.IsSupported)
{
return(Avx.HorizontalAdd(left, right));
}
return(HorizontalAdd_Software(left, right));
}
private Hit[] RayTraceAVXFaster(Ray ray)
{
Vector256 <double> dir = (Vector256 <double>)ray.Direction;
Vector256 <double> vert0 = (Vector256 <double>)Vert0.Position;
Vector256 <double> edge0to1 = (Vector256 <double>)Edge0to1;
Vector256 <double> edge0to2 = (Vector256 <double>)Edge0to2;
Vector256 <double> offset = Avx.Subtract((Vector256 <double>)ray.Origin, vert0);
Vector256 <double> side1 = SIMDHelpers.Cross(offset, edge0to1);
Vector256 <double> side2 = SIMDHelpers.Cross(dir, edge0to2);
// Prepare all dot products
Vector256 <double> uvTemp = Avx.Multiply(offset, side2); // u
Vector256 <double> temp = Avx.Multiply(dir, side1); // v
Vector256 <double> edge2Temp = Avx.Multiply(edge0to2, side1);
Vector256 <double> distTemp = Avx.Multiply(edge0to1, side2);
uvTemp = Avx.HorizontalAdd(uvTemp, temp);
edge2Temp = Avx.HorizontalAdd(edge2Temp, edge2Temp);
distTemp = Avx.HorizontalAdd(distTemp, distTemp);
// Complete all dot products for SSE ops
Vector128 <double> uvs = SIMDHelpers.Add2(uvTemp);
Vector128 <double> dist = SIMDHelpers.Add2(edge2Temp);
Vector128 <double> temp1 = SIMDHelpers.Add2(distTemp);
Vector128 <double> temp2;
// vec2 constants we'll be using later
Vector128 <double> ones2 = SIMDHelpers.BroadcastScalar2(1D);
Vector128 <double> zeroes2 = new Vector128 <double>();
// Reciprocal of distance along edge0to1
temp1 = Sse2.Divide(ones2, temp1);
temp2 = Sse2.CompareOrdered(temp1, temp1);
// Remove NaNs from the result, replaced with 0
Vector128 <double> distZeroed = Sse2.And(temp1, temp2);
uvs = Sse2.Multiply(uvs, distZeroed);
dist = Sse2.Multiply(dist, distZeroed);
// compare uvs < 0 and > 1, dist < 0, jump out if any of those conditions are met
temp1 = Sse2.CompareLessThan(uvs, zeroes2);
temp2 = Mirror ? uvs : Sse3.HorizontalAdd(uvs, uvs);
temp2 = Sse2.CompareGreaterThan(temp2, ones2);
temp1 = Sse2.Or(temp1, temp2);
temp2 = Sse2.CompareLessThan(dist, zeroes2);
temp1 = Sse2.Or(temp1, temp2);
if (!Avx.TestZ(temp1, temp1))
{
return(default);
public static double ReduceSum(this Vector <double> vector)
{
#if NETCOREAPP3_0
if (Avx.IsSupported)
{
Vector256 <double> a = Unsafe.As <Vector <double>, Vector256 <double> >(ref vector);
Vector256 <double> tmp = Avx.HorizontalAdd(a, a);
Vector128 <double> hi128 = tmp.GetUpper();
Vector128 <double> lo128 = tmp.GetLower();
Vector128 <double> s = Sse2.Add(lo128, hi128);
return(s.ToScalar());
}
#endif
return(Vector.Dot(Vector <double> .One, vector));
}
public static Vector256 <double> DotProduct4D(Vector256 <double> left, Vector256 <double> right)
{
if (Avx.IsSupported)
{
Vector256 <double> result = Avx.Multiply(left, right);
// We now have (X, Y, Z, 0) correctly, and want to add them together and fill with that result
result = Avx.HorizontalAdd(result, result);
// Now we have (X + Y, X + Y, Z + 0, Z + 0)
result = Avx.Add(result, Avx.Permute2x128(result, result, 0b_0000_0001));
// We switch the 2 halves, and add that to the original, getting the result in all elems
return(result);
}
return(DotProduct4D_Software(left, right));
}
public Vector256 <double> Permute(Vector256 <double> left, Vector256 <double> right)
{
Vector256 <double> mul = Avx.Multiply(left, right);
// Set W to zero
Vector256 <double> result = Avx.And(mul, MaskWDouble);
// We now have (X, Y, Z, 0) correctly, and want to add them together and fill with that result
result = Avx.HorizontalAdd(result, result);
// Now we have (X + Y, X + Y, Z + 0, Z + 0)
result = Avx.Add(result, Avx.Permute2x128(result, result, 0b_0000_0001));
// We switch the 2 halves, and add that to the original, getting the result in all elems
// Set W to zero
result = Avx.And(result, MaskWDouble);
return(result);
}
public Vector256 <double> DoubleHadd(Vector256 <double> left, Vector256 <double> right)
{
Vector256 <double> mul = Avx.Multiply(left, right);
// Set W to zero
Vector256 <double> result = Avx.And(mul, MaskWDouble);
// We now have (X, Y, Z, 0) correctly, and want to add them together and fill with that result
result = Avx.HorizontalAdd(result, result);
// Now we have (X + Y, X + Y, Z + 0, Z + 0)
result = Avx.Shuffle(result, result, ShuffleValues._3_1_2_0);
result = Avx.HorizontalAdd(result, result);
// We switch the 2 halves, and add that to the original, getting the result in all elems
// Set W to zero
result = Avx.And(result, MaskWDouble);
return(result);
}
public static unsafe float Sum_AVX(float[] array)
{
Vector256 <float> sum = Avx.SetZeroVector256 <float>();
fixed(float *ptr = &array[0])
{
for (int i = 0; i < array.Length; i += 8)
{
var current = Avx.LoadVector256(ptr + i);
sum = Avx.Add(current, sum);
}
}
// sum all values in __m256 (horizontal sum)
var ha = Avx.HorizontalAdd(sum, sum);
var ha2 = Avx.HorizontalAdd(ha, ha);
var lo = Avx.ExtractVector128(ha2, 1);
var resultV = Sse.Add(Avx.GetLowerHalf(ha2), lo);
return(Sse.ConvertToSingle(resultV));
}
public static Vector256 <double> DotProduct3D(Vector256 <double> left, Vector256 <double> right)
{
// We can use AVX to vectorize the multiplication
if (Avx.IsSupported)
{
Vector256 <double> mul = Avx.Multiply(left, right);
// Set W to zero
Vector256 <double> result = Avx.And(mul, DoubleConstants.MaskW);
// We now have (X, Y, Z, 0) correctly, and want to add them together and fill with that result
result = Avx.HorizontalAdd(result, result);
// Now we have (X + Y, X + Y, Z + 0, Z + 0)
result = Avx.Add(result, Avx.Permute2x128(result, result, 0b_0000_0001));
// We switch the 2 halves, and add that to the original, getting the result in all elems
return(result);
}
return(DotProduct3D_Software(left, right));
}
// This function implements Algorithm 1 in https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf.
// Compute the output value of the field-aware factorization, as the sum of the linear part and the latent part.
// The linear part is the inner product of linearWeights and featureValues.
// The latent part is the sum of all intra-field interactions in one field f, for all fields possible
public static unsafe void CalculateIntermediateVariables(int *fieldIndices, int *featureIndices, float *featureValues,
float *linearWeights, float *latentWeights, float *latentSum, float *response, int fieldCount, int latentDim, int count)
{
Contracts.Assert(Avx.IsSupported);
// The number of all possible fields.
int m = fieldCount;
int d = latentDim;
int c = count;
int * pf = fieldIndices;
int * pi = featureIndices;
float *px = featureValues;
float *pw = linearWeights;
float *pv = latentWeights;
float *pq = latentSum;
float linearResponse = 0;
float latentResponse = 0;
Unsafe.InitBlock(pq, 0, (uint)(m * m * d * sizeof(float)));
Vector256 <float> y = Vector256 <float> .Zero;
Vector256 <float> tmp = Vector256 <float> .Zero;
for (int i = 0; i < c; i++)
{
int f = pf[i];
int j = pi[i];
linearResponse += pw[j] * px[i];
Vector256 <float> x = Avx.BroadcastScalarToVector256(px + i);
Vector256 <float> xx = Avx.Multiply(x, x);
// tmp -= <v_j,f, v_j,f> * x * x
int vBias = j * m * d + f * d;
// j-th feature's latent vector in the f-th field hidden space.
float *vjf = pv + vBias;
for (int k = 0; k + 8 <= d; k += 8)
{
Vector256 <float> vjfBuffer = Avx.LoadVector256(vjf + k);
tmp = MultiplyAddNegated(Avx.Multiply(vjfBuffer, vjfBuffer), xx, tmp);
}
for (int fprime = 0; fprime < m; fprime++)
{
vBias = j * m * d + fprime * d;
int qBias = f * m * d + fprime * d;
float *vjfprime = pv + vBias;
float *qffprime = pq + qBias;
// q_f,f' += v_j,f' * x
for (int k = 0; k + 8 <= d; k += 8)
{
Vector256 <float> vjfprimeBuffer = Avx.LoadVector256(vjfprime + k);
Vector256 <float> q = Avx.LoadVector256(qffprime + k);
q = MultiplyAdd(vjfprimeBuffer, x, q);
Avx.Store(qffprime + k, q);
}
}
}
for (int f = 0; f < m; f++)
{
// tmp += <q_f,f, q_f,f>
float *qff = pq + f * m * d + f * d;
for (int k = 0; k + 8 <= d; k += 8)
{
Vector256 <float> qffBuffer = Avx.LoadVector256(qff + k);
// Intra-field interactions.
tmp = MultiplyAdd(qffBuffer, qffBuffer, tmp);
}
// y += <q_f,f', q_f',f>, f != f'
// Whis loop handles inter - field interactions because f != f'.
for (int fprime = f + 1; fprime < m; fprime++)
{
float *qffprime = pq + f * m * d + fprime * d;
float *qfprimef = pq + fprime * m * d + f * d;
for (int k = 0; k + 8 <= d; k += 8)
{
// Inter-field interaction.
Vector256 <float> qffprimeBuffer = Avx.LoadVector256(qffprime + k);
Vector256 <float> qfprimefBuffer = Avx.LoadVector256(qfprimef + k);
y = MultiplyAdd(qffprimeBuffer, qfprimefBuffer, y);
}
}
}
y = MultiplyAdd(_point5, tmp, y);
tmp = Avx.Add(y, Avx.Permute2x128(y, y, 1));
tmp = Avx.HorizontalAdd(tmp, tmp);
y = Avx.HorizontalAdd(tmp, tmp);
Sse.StoreScalar(&latentResponse, y.GetLower()); // The lowest slot is the response value.
*response = linearResponse + latentResponse;
}
private static float SumVector256(Vector256 <float> v)
{
v = Avx.HorizontalAdd(v, v); //0+1, 2+3, .., .., 4+5, 6+7, .., ..
v = Avx.HorizontalAdd(v, v); //0+1+2+3, .., .., .., 4+5+6+7, .., .., ..
return(v.GetUpper().ToScalar() + v.GetLower().ToScalar());
}
public static unsafe float GetScribnerBoardFeetPerAcre(Trees trees)
{
// for now, assume all trees are of the same species
if (trees.Species != FiaCode.PseudotsugaMenziesii)
{
throw new NotSupportedException();
}
if (trees.Units != Units.English)
{
throw new NotSupportedException();
}
// Douglas-fir
#if DEBUG
Vector128 <float> v6p8 = AvxExtensions.BroadcastScalarToVector128(6.8F);
Vector128 <float> v10k = AvxExtensions.BroadcastScalarToVector128(10.0F * 1000.0F);
#endif
// constants
Vector128 <float> forestersEnglish = AvxExtensions.BroadcastScalarToVector128(Constant.ForestersEnglish);
Vector128 <float> one = AvxExtensions.BroadcastScalarToVector128(1.0F);
Vector128 <float> six = AvxExtensions.BroadcastScalarToVector128(6.0F);
Vector128 <float> vm3p21809 = AvxExtensions.BroadcastScalarToVector128(-3.21809F); // b4
Vector128 <float> v0p04948 = AvxExtensions.BroadcastScalarToVector128(0.04948F);
Vector128 <float> vm0p15664 = AvxExtensions.BroadcastScalarToVector128(-0.15664F);
Vector128 <float> v2p02132 = AvxExtensions.BroadcastScalarToVector128(2.02132F);
Vector128 <float> v1p63408 = AvxExtensions.BroadcastScalarToVector128(1.63408F);
Vector128 <float> vm0p16184 = AvxExtensions.BroadcastScalarToVector128(-0.16184F);
Vector128 <float> v1p033 = AvxExtensions.BroadcastScalarToVector128(1.033F);
Vector128 <float> v1p382937 = AvxExtensions.BroadcastScalarToVector128(1.382937F);
Vector128 <float> vm0p4015292 = AvxExtensions.BroadcastScalarToVector128(-0.4015292F);
Vector128 <float> v0p087266 = AvxExtensions.BroadcastScalarToVector128(0.087266F);
Vector128 <float> vm0p174533 = AvxExtensions.BroadcastScalarToVector128(-0.174533F);
Vector128 <float> vm0p6896598794 = AvxExtensions.BroadcastScalarToVector128(-0.6896598794F); // rc6-rs632
Vector128 <float> v0p993 = AvxExtensions.BroadcastScalarToVector128(0.993F);
Vector128 <float> v0p174439 = AvxExtensions.BroadcastScalarToVector128(0.174439F);
Vector128 <float> v0p117594 = AvxExtensions.BroadcastScalarToVector128(0.117594F);
Vector128 <float> vm8p210585 = AvxExtensions.BroadcastScalarToVector128(-8.210585F);
Vector128 <float> v0p236693 = AvxExtensions.BroadcastScalarToVector128(0.236693F);
Vector128 <float> v0p00001345 = AvxExtensions.BroadcastScalarToVector128(0.00001345F);
Vector128 <float> v0p00001937 = AvxExtensions.BroadcastScalarToVector128(0.00001937F);
Vector128 <float> v1p001491 = AvxExtensions.BroadcastScalarToVector128(1.001491F);
Vector128 <float> vm6p924097 = AvxExtensions.BroadcastScalarToVector128(-6.924097F);
Vector128 <float> v0p912733 = AvxExtensions.BroadcastScalarToVector128(0.912733F);
Vector128 <float> v0p00001351 = AvxExtensions.BroadcastScalarToVector128(0.00001351F);
fixed(float *dbh = &trees.Dbh[0], expansionFactors = &trees.LiveExpansionFactor[0], height = &trees.Height[0])
{
Vector128 <float> standBoardFeetPerAcre = Vector128 <float> .Zero;
for (int treeIndex = 0; treeIndex < trees.Count; treeIndex += Constant.Simd128x4.Width)
{
Vector128 <float> dbhInInches = Avx.LoadVector128(dbh + treeIndex);
Vector128 <float> heightInFeet = Avx.LoadVector128(height + treeIndex);
Vector128 <float> logDbhInInches = MathV.Log10(dbhInInches);
Vector128 <float> logHeightInFeet = MathV.Log10(heightInFeet);
// FiaCode.PseudotsugaMenziesii => -3.21809F + 0.04948F * logHeightInFeet * logDbhInInches - 0.15664F * logDbhInInches * logDbhInInches +
// 2.02132F * logDbhInInches + 1.63408F * logHeightInFeet - 0.16184F * logHeightInFeet * logHeightInFeet,
Vector128 <float> cvtsl = Avx.Add(vm3p21809, Avx.Multiply(v0p04948, Avx.Multiply(logHeightInFeet, logDbhInInches)));
cvtsl = Avx.Add(cvtsl, Avx.Multiply(vm0p15664, Avx.Multiply(logDbhInInches, logDbhInInches)));
cvtsl = Avx.Add(cvtsl, Avx.Multiply(v2p02132, logDbhInInches));
cvtsl = Avx.Add(cvtsl, Avx.Multiply(v1p63408, logHeightInFeet));
cvtsl = Avx.Add(cvtsl, Avx.Multiply(vm0p16184, Avx.Multiply(logHeightInFeet, logHeightInFeet)));
Vector128 <float> cubicFeet = MathV.Exp10(cvtsl);
Vector128 <float> dbhSquared = Avx.Multiply(dbhInInches, dbhInInches); // could be consolidated by merging other scaling constants with Forester's constant for basal area
Vector128 <float> basalAreaInSquareFeet = Avx.Multiply(forestersEnglish, dbhSquared);
// b4 = cubicFeet / (1.033F * (1.0F + 1.382937F * MathV.Exp(-4.015292F * dbhInInches / 10.0F)) * (basalAreaInSquareFeet + 0.087266F) - 0.174533F);
Vector128 <float> b4 = Avx.Divide(cubicFeet, Avx.Add(Avx.Multiply(v1p033,
Avx.Multiply(Avx.Add(one, Avx.Multiply(v1p382937,
MathV.Exp(Avx.Multiply(vm0p4015292,
dbhInInches)))),
Avx.Add(basalAreaInSquareFeet, v0p087266))),
vm0p174533));
Vector128 <float> cv4 = Avx.Multiply(b4, Avx.Subtract(basalAreaInSquareFeet, v0p087266));
// conversion to Scribner volumes for 32 foot trees
// Waddell 2014:32
// rc6 = 0.993F * (1.0F - MathF.Pow(0.62F, dbhInInches - 6.0F));
Vector128 <float> rc6 = Avx.Multiply(v0p993, Avx.Subtract(one, MathV.Exp(Avx.Multiply(vm0p6896598794, Avx.Subtract(dbhInInches, six))))); // log2(0.62) = -0.6896598794
Vector128 <float> cv6 = Avx.Multiply(rc6, cv4);
Vector128 <float> logB4 = MathV.Log10(b4);
// float rs616 = MathF.Pow(10.0F, 0.174439F + 0.117594F * logDbhInInches * logB4 - 8.210585F / (dbhInInches * dbhInInches) + 0.236693F * logB4 - 0.00001345F * b4 * b4 - 0.00001937F * dbhInInches * dbhInInches);
Vector128 <float> rs616l = Avx.Add(v0p174439, Avx.Multiply(v0p117594, Avx.Multiply(logDbhInInches, logB4)));
rs616l = Avx.Add(rs616l, Avx.Divide(vm8p210585, dbhSquared));
rs616l = Avx.Add(rs616l, Avx.Multiply(v0p236693, logB4));
rs616l = Avx.Subtract(rs616l, Avx.Multiply(v0p00001345, Avx.Multiply(b4, b4)));
rs616l = Avx.Subtract(rs616l, Avx.Multiply(v0p00001937, dbhSquared));
Vector128 <float> rs616 = MathV.Exp10(rs616l);
Vector128 <float> sv616 = Avx.Multiply(rs616, cv6); // Scribner board foot volume to a 6 inch top for 16 foot logs
// float rs632 = 1.001491F - 6.924097F / tarif + 0.00001351F * dbhInInches * dbhInInches;
Vector128 <float> rs632 = Avx.Add(v1p001491, Avx.Divide(vm6p924097, Avx.Multiply(v0p912733, b4)));
rs632 = Avx.Add(rs632, Avx.Multiply(v0p00001351, dbhSquared));
Vector128 <float> zeroVolumeMask = Avx.CompareLessThanOrEqual(dbhInInches, six);
Vector128 <float> sv632 = Avx.Multiply(rs632, sv616); // Scribner board foot volume to a 6 inch top for 32 foot logs
sv632 = Avx.BlendVariable(sv632, Vector128 <float> .Zero, zeroVolumeMask);
#if DEBUG
DebugV.Assert(Avx.CompareGreaterThanOrEqual(Avx.BlendVariable(rc6, Vector128 <float> .Zero, zeroVolumeMask), Vector128 <float> .Zero));
DebugV.Assert(Avx.CompareLessThanOrEqual(rc6, one));
DebugV.Assert(Avx.CompareGreaterThanOrEqual(Avx.BlendVariable(rs616, one, zeroVolumeMask), one));
DebugV.Assert(Avx.CompareLessThanOrEqual(Avx.BlendVariable(rs616, Vector128 <float> .Zero, zeroVolumeMask), v6p8));
DebugV.Assert(Avx.CompareGreaterThanOrEqual(Avx.BlendVariable(rs632, Vector128 <float> .Zero, zeroVolumeMask), Vector128 <float> .Zero));
DebugV.Assert(Avx.CompareLessThanOrEqual(Avx.BlendVariable(rs632, Vector128 <float> .Zero, zeroVolumeMask), one));
DebugV.Assert(Avx.CompareGreaterThanOrEqual(Avx.BlendVariable(sv632, Vector128 <float> .Zero, zeroVolumeMask), Vector128 <float> .Zero));
DebugV.Assert(Avx.CompareLessThanOrEqual(Avx.BlendVariable(sv632, Vector128 <float> .Zero, zeroVolumeMask), v10k));
#endif
Vector128 <float> expansionFactor = Avx.LoadVector128(expansionFactors + treeIndex);
standBoardFeetPerAcre = Avx.Add(standBoardFeetPerAcre, Avx.Multiply(expansionFactor, sv632));
}
standBoardFeetPerAcre = Avx.HorizontalAdd(standBoardFeetPerAcre, standBoardFeetPerAcre);
standBoardFeetPerAcre = Avx.HorizontalAdd(standBoardFeetPerAcre, standBoardFeetPerAcre);
return(standBoardFeetPerAcre.ToScalar());
}
}
public Intro()
{
var middleVector = Vector128.Create(1.0f); // middleVector = <1,1,1,1>
middleVector = Vector128.CreateScalar(-1.0f); // middleVector = <-1,0,0,0>
var floatBytes = Vector64.AsByte(Vector64.Create(1.0f, -1.0f)); // floatBytes = <0, 0, 128, 63, 0, 0, 128, 191>
if (Avx.IsSupported)
{
var left = Vector256.Create(-2.5f); // <-2.5, -2.5, -2.5, -2.5, -2.5, -2.5, -2.5, -2.5>
var right = Vector256.Create(5.0f); // <5, 5, 5, 5, 5, 5, 5, 5>
Vector256 <float> result = Avx.AddSubtract(left, right); // result = <-7.5, 2.5, -7.5, 2.5, -7.5, 2.5, -7.5, 2.5>xit
left = Vector256.Create(-1.0f, -2.0f, -3.0f, -4.0f, -50.0f, -60.0f, -70.0f, -80.0f);
right = Vector256.Create(0.0f, 2.0f, 3.0f, 4.0f, 50.0f, 60.0f, 70.0f, 80.0f);
result = Avx.UnpackHigh(left, right); // result = <-3, 3, -4, 4, -70, 70, -80, 80>
result = Avx.UnpackLow(left, right); // result = <-1, 1, -2, 2, -50, 50, -60, 60>
result = Avx.DotProduct(left, right, 0b1111_0001); // result = <-30, 0, 0, 0, -17400, 0, 0, 0>
bool testResult = Avx.TestC(left, right); // testResult = true
testResult = Avx.TestC(right, left); // testResult = false
Vector256 <float> result1 = Avx.Divide(left, right);
var plusOne = Vector256.Create(1.0f);
result = Avx.Compare(right, result1, FloatComparisonMode.OrderedGreaterThanNonSignaling);
result = Avx.Compare(right, result1, FloatComparisonMode.UnorderedNotLessThanNonSignaling);
left = Vector256.Create(0.0f, 3.0f, -3.0f, 4.0f, -50.0f, 60.0f, -70.0f, 80.0f);
right = Vector256.Create(0.0f, 2.0f, 3.0f, 2.0f, 50.0f, -60.0f, 70.0f, -80.0f);
Vector256 <float> nanInFirstPosition = Avx.Divide(left, right);
left = Vector256.Create(1.1f, 3.3333333f, -3.0f, 4.22f, -50.0f, 60.0f, -70.0f, 80.0f);
Vector256 <float> InfInFirstPosition = Avx.Divide(left, right);
left = Vector256.Create(-1.1f, 3.0f, 1.0f / 3.0f, MathF.PI, -50.0f, 60.0f, -70.0f, 80.0f);
right = Vector256.Create(0.0f, 2.0f, 3.1f, 2.0f, 50.0f, -60.0f, 70.0f, -80.0f);
Vector256 <float> compareResult = Avx.Compare(left, right, FloatComparisonMode.OrderedGreaterThanNonSignaling); // compareResult = <0, NaN, 0, NaN, 0, NaN, 0, NaN>
Vector256 <float> mixed = Avx.BlendVariable(left, right, compareResult); // mixed = <-1, 2, -3, 2, -50, -60, -70, -80>
//left = Vector256.Create(-1.0f, 1.0f, -1.0f, 1.0f, -1.0f, 1.0f, -1.0f, 1.0f);
//right = Vector256.Create(1.0f, 1.0f, -1.0f, 1.0f, 1.0f, 1.0f, -1.0f, 1.0f);
Vector256 <float> other = right = Vector256.Create(0.0f, 2.0f, 3.0f, 2.0f, 50.0f, -60.0f, 70.0f, -80.0f);
bool bRes = Avx.TestZ(plusOne, compareResult);
bool bRes2 = Avx.TestC(plusOne, compareResult);
bool allTrue = !Avx.TestZ(compareResult, compareResult);
compareResult = Avx.Compare(nanInFirstPosition, right, FloatComparisonMode.OrderedEqualNonSignaling); // compareResult = <0, NaN, 0, NaN, 0, NaN, 0, NaN>
compareResult = Avx.Compare(nanInFirstPosition, right, FloatComparisonMode.UnorderedEqualNonSignaling);
compareResult = Avx.Compare(InfInFirstPosition, right, FloatComparisonMode.UnorderedNotLessThanOrEqualNonSignaling);
compareResult = Avx.Compare(InfInFirstPosition, right, FloatComparisonMode.OrderedGreaterThanNonSignaling);
var left128 = Vector128.Create(1.0f, 2.0f, 3.0f, 4.0f);
var right128 = Vector128.Create(2.0f, 3.0f, 4.0f, 5.0f);
Vector128 <float> compResult128 = Sse.CompareGreaterThan(left128, right128); // compResult128 = <0, 0, 0, 0>
int res = Avx.MoveMask(compareResult);
if (Fma.IsSupported)
{
Vector256 <float> resultFma = Fma.MultiplyAdd(left, right, other); // = left * right + other for each element
resultFma = Fma.MultiplyAddNegated(left, right, other); // = -(left * right + other) for each element
resultFma = Fma.MultiplySubtract(left, right, other); // = left * right - other for each element
Fma.MultiplyAddSubtract(left, right, other); // even elements (0, 2, ...) like MultiplyAdd, odd elements like MultiplySubtract
}
result = Avx.DotProduct(left, right, 0b1010_0001); // result = <-20, 0, 0, 0, -10000, 0, 0, 0>
result = Avx.Floor(left); // result = <-3, -3, -3, -3, -3, -3, -3, -3>
result = Avx.Add(left, right); // result = <2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5>
result = Avx.Ceiling(left); // result = <-2, -2, -2, -2, -2, -2, -2, -2>
result = Avx.Multiply(left, right); // result = <-12.5, -12.5, -12.5, -12.5, -12.5, -12.5, -12.5, -12.5>
result = Avx.HorizontalAdd(left, right); // result = <-5, -5, 10, 10, -5, -5, 10, 10>
result = Avx.HorizontalSubtract(left, right); // result = <0, 0, 0, 0, 0, 0, 0, 0>
double[] someDoubles = new double[] { 1.0, 3.0, -2.5, 7.5, 10.8, 0.33333 };
double[] someOtherDoubles = new double[] { 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 };
double[] someResult = new double[someDoubles.Length];
float[] someFloats = new float[] { 1, 2, 3, 4, 10, 20, 30, 40, 0 };
float[] someOtherFloats = new float[] { 1, 1, 1, 1, 1, 1, 1, 1 };
unsafe
{
fixed(double *ptr = &someDoubles[1])
{
fixed(double *ptr2 = &someResult[0])
{
Vector256 <double> res2 = Avx.LoadVector256(ptr); // res2 = <3, -2.5, 7.5, 10.8>
Avx.Store(ptr2, res2);
}
}
fixed(float *ptr = &someFloats[0])
{
fixed(float *ptr2 = &someOtherFloats[0])
{
Vector256 <float> res2 = Avx.DotProduct(Avx.LoadVector256(ptr), Avx.LoadVector256(ptr2), 0b0001_0001);
//Avx.Store(ptr2, res2);
}
}
}
}
}
