public void RunStructFldScenario(SimpleBinaryOpTest__AndSingle testClass)
{
var result = Avx.And(_fld1, _fld2);
Unsafe.Write(testClass._dataTable.outArrayPtr, result);
testClass.ValidateResult(_fld1, _fld2, testClass._dataTable.outArrayPtr);
}
private static void AbsAvx(ReadOnlySpan <float> a, Span <float> s)
{
var remainder = a.Length & 7;
var length = a.Length - remainder;
var mask = Avx2.ShiftRightLogical(Vector256.Create(-1), 1).AsSingle();
fixed(float *ptr = a)
{
fixed(float *ptrS = s)
{
for (var i = 0; i < length; i += 8)
{
var j = Avx.LoadVector256(ptr + i);
Avx.Store(ptrS + i, Avx.And(mask, j));
}
}
}
if (remainder != 0)
{
AbsNaive(a, s, length, a.Length);
}
}
public static Vector256 <T> And <T>(Vector256 <T> left, Vector256 <T> right) where T : struct
{
if (typeof(T) == typeof(float))
{
if (Avx.IsSupported)
{
return(Avx.And(left.AsSingle(), right.AsSingle()).As <float, T>());
}
}
if (typeof(T) == typeof(double))
{
if (Avx.IsSupported)
{
return(Avx.And(left.AsDouble(), right.AsDouble()).As <double, T>());
}
}
if (Avx.IsSupported)
{
return(Avx.And(left.AsSingle(), right.AsSingle()).As <float, T>());
}
return(SoftwareFallbacks.And_Software(left, right));
}
public void RunFldScenario()
{
var result = Avx.And(_fld1, _fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_fld1, _fld2, _dataTable.outArrayPtr);
}
public void RunLclFldScenario()
{
var test = new SimpleBinaryOpTest__AndDouble();
var result = Avx.And(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunStructLclFldScenario()
{
var test = TestStruct.Create();
var result = Avx.And(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunClassFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassFldScenario));
var result = Avx.And(_fld1, _fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_fld1, _fld2, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_LoadAligned()
{
var left = Avx.LoadAlignedVector256((Double *)(_dataTable.inArray1Ptr));
var right = Avx.LoadAlignedVector256((Double *)(_dataTable.inArray2Ptr));
var result = Avx.And(left, right);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(left, right, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_UnsafeRead()
{
var left = Unsafe.Read <Vector256 <Double> >(_dataTable.inArray1Ptr);
var right = Unsafe.Read <Vector256 <Double> >(_dataTable.inArray2Ptr);
var result = Avx.And(left, right);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(left, right, _dataTable.outArrayPtr);
}
public void RunBasicScenario_LoadAligned()
{
var result = Avx.And(
Avx.LoadAlignedVector256((Double *)(_dataTable.inArray1Ptr)),
Avx.LoadAlignedVector256((Double *)(_dataTable.inArray2Ptr))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunClsVarScenario()
{
var result = Avx.And(
_clsVar1,
_clsVar2
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
public void RunStructLclFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunStructLclFldScenario));
var test = TestStruct.Create();
var result = Avx.And(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunClassLclFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario));
var test = new SimpleBinaryOpTest__AndSingle();
var result = Avx.And(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunBasicScenario_UnsafeRead()
{
var result = Avx.And(
Unsafe.Read <Vector256 <Double> >(_dataTable.inArray1Ptr),
Unsafe.Read <Vector256 <Double> >(_dataTable.inArray2Ptr)
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public static Vector256 <double> CrossProduct3D(Vector256 <double> left, Vector256 <double> right)
{
if (Avx2.IsSupported)
{
#region Comments
/* Cross product of A(x, y, z, _) and B(x, y, z, _) is
* 0 1 2 3 0 1 2 3
*
* '(X = (Ay * Bz) - (Az * By), Y = (Az * Bx) - (Ax * Bz), Z = (Ax * By) - (Ay * Bx)'
* 1 2 1 2 1 2
* So we can do (Ay, Az, Ax, _) * (Bz, Bx, By, _) (last elem is irrelevant, as this is for Vector3)
* which leaves us with a of the first subtraction element for each (marked 1 above)
* Then we repeat with the right hand of subtractions (Az, Ax, Ay, _) * (By, Bz, Bx, _)
* which leaves us with the right hand sides (marked 2 above)
* Then we subtract them to get the correct vector
* We then mask out W to zero, because that is required for the Vector3 representation
*
* We perform the first 2 multiplications by shuffling the vectors and then multiplying them
* Helpers.Shuffle is the same as the C++ macro _MM_SHUFFLE, and you provide the order you wish the elements
* to be in *reversed* (no clue why), so here (3, 0, 2, 1) means you have the 2nd elem (1, 0 indexed) in the first slot,
* the 3rd elem (2) in the next one, the 1st elem (0) in the next one, and the 4th (3, W/_, unused here) in the last reg
*/
#endregion
/*
* lhs1 goes from x, y, z, _ to y, z, x, _
* rhs1 goes from x, y, z, _ to z, x, y, _
*/
Vector256 <double> leftHandSide1 = Avx2.Permute4x64(left, ShuffleValues.YZXW);
Vector256 <double> rightHandSide1 = Avx2.Permute4x64(right, ShuffleValues.ZXYW);
/*
* lhs2 goes from x, y, z, _ to z, x, y, _
* rhs2 goes from x, y, z, _ to y, z, x, _
*/
Vector256 <double> leftHandSide2 = Avx2.Permute4x64(left, ShuffleValues.ZXYW);
Vector256 <double> rightHandSide2 = Avx2.Permute4x64(right, ShuffleValues.YZXW);
Vector256 <double> mul1 = Avx.Multiply(leftHandSide1, rightHandSide1);
Vector256 <double> mul2 = Avx.Multiply(leftHandSide2, rightHandSide2);
Vector256 <double> resultNonMaskedW = Avx.Subtract(mul1, mul2);
return(Avx.And(resultNonMaskedW, DoubleConstants.MaskW));
// TODO reuse vectors (minimal register usage) - potentially prevent any stack spilling
}
return(CrossProduct3D_Software(left, right));
}
public void RunLclVarScenario_LoadAligned()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_LoadAligned));
var left = Avx.LoadAlignedVector256((Single *)(_dataTable.inArray1Ptr));
var right = Avx.LoadAlignedVector256((Single *)(_dataTable.inArray2Ptr));
var result = Avx.And(left, right);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(left, right, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_UnsafeRead()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_UnsafeRead));
var op1 = Unsafe.Read <Vector256 <Double> >(_dataTable.inArray1Ptr);
var op2 = Unsafe.Read <Vector256 <Double> >(_dataTable.inArray2Ptr);
var result = Avx.And(op1, op2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(op1, op2, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_Load));
var op1 = Avx.LoadVector256((Single *)(_dataTable.inArray1Ptr));
var op2 = Avx.LoadVector256((Single *)(_dataTable.inArray2Ptr));
var result = Avx.And(op1, op2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(op1, op2, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_UnsafeRead()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_UnsafeRead));
var left = Unsafe.Read <Vector256 <Single> >(_dataTable.inArray1Ptr);
var right = Unsafe.Read <Vector256 <Single> >(_dataTable.inArray2Ptr);
var result = Avx.And(left, right);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(left, right, _dataTable.outArrayPtr);
}
public void RunBasicScenario_LoadAligned()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunBasicScenario_LoadAligned));
var result = Avx.And(
Avx.LoadAlignedVector256((Double *)(_dataTable.inArray1Ptr)),
Avx.LoadAlignedVector256((Double *)(_dataTable.inArray2Ptr))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunBasicScenario_UnsafeRead()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunBasicScenario_UnsafeRead));
var result = Avx.And(
Unsafe.Read <Vector256 <Single> >(_dataTable.inArray1Ptr),
Unsafe.Read <Vector256 <Single> >(_dataTable.inArray2Ptr)
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunStructLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunStructLclFldScenario_Load));
var test = TestStruct.Create();
var result = Avx.And(
Avx.LoadVector256((Single *)(&test._fld1)),
Avx.LoadVector256((Single *)(&test._fld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunStructFldScenario_Load(SimpleBinaryOpTest__AndSingle testClass)
{
fixed(Vector256 <Single> *pFld1 = &_fld1)
fixed(Vector256 <Single> *pFld2 = &_fld2)
{
var result = Avx.And(
Avx.LoadVector256((Single *)(pFld1)),
Avx.LoadVector256((Single *)(pFld2))
);
Unsafe.Write(testClass._dataTable.outArrayPtr, result);
testClass.ValidateResult(_fld1, _fld2, testClass._dataTable.outArrayPtr);
}
}
public void RunClassFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassFldScenario_Load));
fixed(Vector256 <Single> *pFld1 = &_fld1)
fixed(Vector256 <Single> *pFld2 = &_fld2)
{
var result = Avx.And(
Avx.LoadVector256((Single *)(pFld1)),
Avx.LoadVector256((Single *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_fld1, _fld2, _dataTable.outArrayPtr);
}
}
public void RunClsVarScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClsVarScenario_Load));
fixed(Vector256 <Double> *pClsVar1 = &_clsVar1)
fixed(Vector256 <Double> *pClsVar2 = &_clsVar2)
{
var result = Avx.And(
Avx.LoadVector256((Double *)(pClsVar1)),
Avx.LoadVector256((Double *)(pClsVar2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
}
internal static unsafe (UnsafeMemory <BitState> bits, bool isValidBinary) ToBitStates(ReadOnlySpan <byte> valueText, BitAllocator bitAlloc)
{
UnsafeMemory <BitState> bitsMem = bitAlloc.GetBits(valueText.Length);
Span <BitState> bits = bitsMem.Span;
ulong isValidBinary = 0;
int index = 0;
if (Ssse3.IsSupported && bits.Length >= Vector128 <byte> .Count)
{
int vecBitCount = bits.Length / Vector128 <byte> .Count;
fixed(BitState *bitsPtr = bits)
{
fixed(byte *textPtr = valueText)
{
Vector128 <ulong> isValidBin = Vector128 <ulong> .Zero;
for (; index < vecBitCount; index++)
{
var charText = Avx.LoadVector128(textPtr + index * Vector128 <byte> .Count);
var byteText = Avx.Shuffle(charText, shuffleIdxs);
var firstBit = Avx.And(onlyFirstBit, Avx.Or(byteText, Avx.ShiftRightLogical(byteText.AsInt32(), 1).AsByte()));
var secondBit = Avx.And(onlySecondBit, Avx.ShiftRightLogical(byteText.AsInt32(), 5).AsByte());
var bytesAsBitStates = Avx.Or(firstBit, secondBit);
Avx.Store((byte *)bitsPtr + bits.Length - (index + 1) * Vector128 <byte> .Count, bytesAsBitStates);
isValidBin = Avx.Or(isValidBin, secondBit.AsUInt64());
}
isValidBinary = isValidBin.GetElement(0) | isValidBin.GetElement(1);
}
}
index *= Vector128 <byte> .Count;
}
for (; index < bits.Length; index++)
{
BitState bit = ToBitState(valueText[index]);
bits[bits.Length - index - 1] = bit;
isValidBinary |= (uint)bit & 0b10;
}
return(bitsMem, isValidBinary == 0);
}
public void RunClassLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario_Load));
var test = new SimpleBinaryOpTest__AndSingle();
fixed(Vector256 <Single> *pFld1 = &test._fld1)
fixed(Vector256 <Single> *pFld2 = &test._fld2)
{
var result = Avx.And(
Avx.LoadVector256((Single *)(pFld1)),
Avx.LoadVector256((Single *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
}
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 (double near, double far) IntersectAVX(Ray ray)
{
Vector256 <double> origin = (Vector256 <double>)ray.Origin;
Vector256 <double> direction = (Vector256 <double>)ray.Direction;
Vector256 <double> zeroes = new Vector256 <double>();
Vector256 <double> min = (Vector256 <double>)Minimum;
Vector256 <double> max = (Vector256 <double>)Maximum;
// Replace slabs that won't be checked (0 direction axis) with infinity so that NaN doesn't propagate
Vector256 <double> dirInfMask = Avx.And(
Avx.Compare(direction, zeroes, FloatComparisonMode.OrderedEqualNonSignaling),
Avx.And(
Avx.Compare(origin, min, FloatComparisonMode.OrderedGreaterThanOrEqualNonSignaling),
Avx.Compare(origin, max, FloatComparisonMode.OrderedLessThanOrEqualNonSignaling)));
min = Avx.BlendVariable(min, SIMDHelpers.BroadcastScalar4(double.NegativeInfinity), dirInfMask);
max = Avx.BlendVariable(max, SIMDHelpers.BroadcastScalar4(double.PositiveInfinity), dirInfMask);
// Flip slabs in direction axes that are negative (using direction as mask takes the most significant bit, the sign.. probably includes -0)
Vector256 <double> minMasked = Avx.BlendVariable(min, max, direction);
Vector256 <double> maxMasked = Avx.BlendVariable(max, min, direction);
direction = Avx.Divide(Vector256.Create(1D), direction);
Vector256 <double> near4 = Avx.Multiply(Avx.Subtract(minMasked, origin), direction);
Vector256 <double> far4 = Avx.Multiply(Avx.Subtract(maxMasked, origin), direction);
Vector128 <double> near2 = Sse2.Max(near4.GetLower(), near4.GetUpper());
near2 = Sse2.MaxScalar(near2, SIMDHelpers.Swap(near2));
Vector128 <double> far2 = Sse2.Min(far4.GetLower(), far4.GetUpper());
far2 = Sse2.MinScalar(far2, SIMDHelpers.Swap(far2));
if (Sse2.CompareScalarOrderedGreaterThan(near2, far2) | Sse2.CompareScalarOrderedLessThan(far2, new Vector128 <double>()))
{
return(double.NaN, double.NaN);
}
return(near2.ToScalar(), far2.ToScalar());
}
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);
}