public void RunBasicScenario_Load()
{
var result = Avx2.And(
Avx.LoadVector256((Byte *)(_dataTable.inArray1Ptr)),
Avx.LoadVector256((Byte *)(_dataTable.inArray2Ptr))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunStructLclFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunStructLclFldScenario));
var test = TestStruct.Create();
var result = Avx2.And(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunBasicScenario_UnsafeRead()
{
var result = Avx2.And(
Unsafe.Read <Vector256 <Byte> >(_dataTable.inArray1Ptr),
Unsafe.Read <Vector256 <Byte> >(_dataTable.inArray2Ptr)
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunClassLclFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario));
var test = new SimpleBinaryOpTest__AndInt64();
var result = Avx2.And(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_LoadAligned()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_LoadAligned));
var left = Avx.LoadAlignedVector256((UInt32 *)(_dataTable.inArray1Ptr));
var right = Avx.LoadAlignedVector256((UInt32 *)(_dataTable.inArray2Ptr));
var result = Avx2.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 <Byte> >(_dataTable.inArray1Ptr);
var op2 = Unsafe.Read <Vector256 <Byte> >(_dataTable.inArray2Ptr);
var result = Avx2.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 <UInt32> >(_dataTable.inArray1Ptr);
var right = Unsafe.Read <Vector256 <UInt32> >(_dataTable.inArray2Ptr);
var result = Avx2.And(left, right);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(left, right, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_Load));
var op1 = Avx.LoadVector256((Int32 *)(_dataTable.inArray1Ptr));
var op2 = Avx.LoadVector256((Int32 *)(_dataTable.inArray2Ptr));
var result = Avx2.And(op1, op2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(op1, op2, _dataTable.outArrayPtr);
}
public void RunBasicScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunBasicScenario_Load));
var result = Avx2.And(
Avx.LoadVector256((UInt32 *)(_dataTable.inArray1Ptr)),
Avx.LoadVector256((UInt32 *)(_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 = Avx2.And(
Unsafe.Read <Vector256 <UInt16> >(_dataTable.inArray1Ptr),
Unsafe.Read <Vector256 <UInt16> >(_dataTable.inArray2Ptr)
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunClsVarScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClsVarScenario));
var result = Avx2.And(
_clsVar1,
_clsVar2
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
public static unsafe void Decode32Bytes(byte *source, byte *dest)
{
Vector256 <byte> maskA = Vector256.Create((uint)0x0000_003f).AsByte();
Vector256 <byte> maskB = Vector256.Create((uint)0x0000_3f00).AsByte();
Vector256 <byte> maskC = Vector256.Create((uint)0x003f_0000).AsByte();
Vector256 <byte> maskD = Vector256.Create((uint)0x3f00_0000).AsByte();
Vector256 <byte> offsets = Vector256.Create((sbyte)-32).AsByte();
Vector256 <byte> vecSource = Unsafe.As <byte, Vector256 <byte> >(ref source[0]);
Vector256 <byte> subtracted = Avx2.Add(vecSource, offsets);
Vector256 <byte> a = Avx2.And(subtracted, maskA);
Vector256 <byte> b = Avx2.And(subtracted, maskB);
Vector256 <byte> c = Avx2.And(subtracted, maskC);
Vector256 <byte> d = Avx2.And(subtracted, maskD);
a = Avx2.ShiftLeftLogical(a.AsUInt32(), 18).AsByte(); // 00000000 00000000 00000000 00aaaaaa -> 00000000 aaaaaa00 00000000 00000000
b = Avx2.ShiftLeftLogical(b.AsUInt32(), 4).AsByte(); // 00000000 00000000 00bbbbbb 00000000 -> 00000000 000000bb bbbb0000 00000000
c = Avx2.ShiftRightLogical(c.AsUInt32(), 10).AsByte(); // 00000000 00cccccc 00000000 00000000 -> 00000000 00000000 0000cccc cc000000
d = Avx2.ShiftRightLogical(d.AsUInt32(), 24).AsByte(); // 00dddddd 00000000 00000000 00000000 -> 00000000 00000000 00000000 00dddddd
// After Or: 00000000 aaaaaabb bbbbcccc ccdddddd
// byte 3 byte 1 byte 2 byte 0
// a uint: 0x00000000_00000000__00000000_00111111
// b uint: 0x00000000_00000000__00111111_00000000
// c uint: 0x00000000_00111111__00000000_00000000
// d uint: 0x00111111_00000000__00000000_00000000
a = Avx2.Or(a, b);
c = Avx2.Or(c, d);
a = Avx2.Or(a, c); // AA BB CC 00 AA BB CC 00
// a contains: [C,B,A,0, F,E,D,0, I,H,G,0, L,K,J,0]
// Shuffle bytes so that it becomes: [A,B,C, D,E,F, G,H,I, J,K,L, 0,0,0,0]
//2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, // 3, 7, 11, 15
// 18, 17, 16, 22, 21, 20, // 19
var vecShuffle = Vector256.Create(
0x02, 0x01, 0x00, 0x06, 0x05, 0x04, 0x0a, 0x09, 0x08, 0x0e, 0x0d, 0x0c,
0x80, 0x80, 0x80, 0x80, // 0x03, 0x07, 0x0b, 0x0f
0x12, 0x11, 0x10, 0x16, 0x15, 0x14, 0x1a, 0x19, 0x18, 0x1e, 0x1d, 0x1c,
0x80, 0x80, 0x80, 0x80); // 0x13, 0x17, 0x1b, 0x1f
var vecBytes2 = Avx2.Shuffle(a, vecShuffle);
Sse2.Store(dest, vecBytes2.GetLower());
Sse2.Store(dest + 12, vecBytes2.GetUpper());
}
public void RunStructLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunStructLclFldScenario_Load));
var test = TestStruct.Create();
var result = Avx2.And(
Avx.LoadVector256((UInt16 *)(&test._fld1)),
Avx.LoadVector256((UInt16 *)(&test._fld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public static Vector256 <byte> op_Multiply(Vector256 <byte> left, Vector256 <byte> right)
{
Vector256 <ushort> lowBits = Vector256.Create((ushort)0x00FF);
var lowProduct = Avx2.And(lowBits, Avx2.MultiplyLow(left.As <ushort>(), right.As <ushort>())).AsByte();
var highProduct =
Avx2.ShiftLeftLogical(
Avx2.MultiplyLow(
Avx2.ShiftRightLogical(left.As <ushort>(), 8),
Avx2.ShiftRightLogical(right.As <ushort>(), 8)
),
8).AsByte();
return(Avx2.Or(lowProduct, highProduct));
}
private static Vector256 <ulong> SpreadSIMD(Vector256 <byte> x)
{
// x = _mm256_shuffle_epi8(x, _mm256_set_epi8(
// -1, 11, -1, 10, -1, 9, -1, 8,
// -1, 3, -1, 2, -1, 1, -1, 0,
// -1, 11, -1, 10, -1, 9, -1, 8,
// -1, 3, -1, 2, -1, 1, -1, 0));
//the order of Vector256.Create is reversed of _mm256_set_epi8!
x = Avx2.Shuffle(x.AsSByte(), Vector256.Create(
0, -1, 1, -1, 2, -1, 3, -1,
8, -1, 9, -1, 10, -1, 11, -1,
0, -1, 1, -1, 2, -1, 3, -1,
8, -1, 9, -1, 10, -1, 11, -1)).AsByte();
// const __m256i lut = _mm256_set_epi8(
// 85, 84, 81, 80, 69, 68, 65, 64,
// 21, 20, 17, 16, 5, 4, 1, 0,
// 85, 84, 81, 80, 69, 68, 65, 64,
// 21, 20, 17, 16, 5, 4, 1, 0);
// Vector256<byte> lut = Vector256.Create(
// (byte)0,1,4,5,16,17,20,21,
// 64,65,68,69,80,81,84,85,
// 0,1,4,5,16,17,20,21,
// 64,65,68,69,80,81,84,85
// );
Vector256 <byte> lut = Vector256.Create(
(byte)0b00000000, 0b00000001, 0b00000100, 0b00000101, 0b00010000, 0b00010001, 0b00010100, 0b00010101,
0b01000000, 0b01000001, 0b01000100, 0b01000101, 0b01010000, 0b01010001, 0b01010100, 0b01010101,
0b00000000, 0b00000001, 0b00000100, 0b00000101, 0b00010000, 0b00010001, 0b00010100, 0b00010101,
0b01000000, 0b01000001, 0b01000100, 0b01000101, 0b01010000, 0b01010001, 0b01010100, 0b01010101
);
// __m256i lo = _mm256_and_si256(x, _mm256_set1_epi8(0xf));
Vector256 <byte> lo = Avx2.And(x, Vector256.Create((byte)0x0f));
// lo = _mm256_shuffle_epi8(lut, lo);
lo = Avx2.Shuffle(lut, lo);
// __m256i hi = _mm256_and_si256(x, _mm256_set1_epi8(0xf0));
var hi = Avx2.And(x, Vector256.Create((byte)0xf0));
// hi = _mm256_shuffle_epi8(lut, _mm256_srli_epi64(hi, 4));
hi = Avx2.Shuffle(lut, Avx2.ShiftRightLogical(hi.AsUInt64(), 4).AsByte());
// x = _mm256_or_si256(lo, _mm256_slli_epi64(hi, 8));
x = Avx2.Or(lo, Avx2.ShiftRightLogical(hi.AsUInt64(), 8).AsByte());
return(x.AsUInt64());
}
public void RunStructFldScenario_Load(SimpleBinaryOpTest__AndUInt16 testClass)
{
fixed(Vector256 <UInt16> *pFld1 = &_fld1)
fixed(Vector256 <UInt16> *pFld2 = &_fld2)
{
var result = Avx2.And(
Avx.LoadVector256((UInt16 *)(pFld1)),
Avx.LoadVector256((UInt16 *)(pFld2))
);
Unsafe.Write(testClass._dataTable.outArrayPtr, result);
testClass.ValidateResult(_fld1, _fld2, testClass._dataTable.outArrayPtr);
}
}
/// <summary>
/// Absolute error bounded by 1e-4.
/// </summary>
public static Vector256 <float> Log(Vector256 <float> x)
{
Vector256 <float> exp, addcst, val;
exp = Avx2.ConvertToVector256Single(Avx2.ShiftRightArithmetic(x.As <float, int>(), 23));
// According to BenchmarkDotNet, isolating all the constants up-front
// yield nearly 10% speed-up.
const float bf0 = -89.970756366f;
const float bf1 = float.NaN; // behavior of MathF.Log() on negative numbers
const float bf2 = 3.529304993f;
const float bf3 = -2.461222105f;
const float bf4 = 1.130626167f;
const float bf5 = -0.288739945f;
const float bf6 = 3.110401639e-2f;
const float bf7 = 0.6931471805f;
const int bi0 = 0x7FFFFF;
const int bi1 = 0x3F800000;
//addcst = val > 0 ? -89.970756366f : -(float)INFINITY;
addcst = Avx.BlendVariable(Vector256.Create(bf0),
Vector256.Create(bf1),
Avx.Compare(x, Vector256 <float> .Zero, FloatComparisonMode.OrderedLessThanNonSignaling));
val = Avx2.Or(Avx2.And(
x.As <float, int>(),
Vector256.Create(bi0)),
Vector256.Create(bi1)).As <int, float>();
/* x * (3.529304993f +
* x * (-2.461222105f +
* x * (1.130626167f +
* x * (-0.288739945f +
* x * 3.110401639e-2f))))
+ (addcst + 0.6931471805f*exp); */
return(Avx2.Add(
Avx2.Multiply(val, Avx2.Add(Vector256.Create(bf2),
Avx2.Multiply(val, Avx2.Add(Vector256.Create(bf3),
Avx2.Multiply(val, Avx2.Add(Vector256.Create(bf4),
Avx2.Multiply(val, Avx2.Add(Vector256.Create(bf5),
Avx2.Multiply(val, Vector256.Create(bf6)))))))))),
Avx.Add(addcst,
Avx2.Multiply(Vector256.Create(bf7), exp))));
}
unsafe private static void denoiseLineAvx2(byte *pcurr, byte *pprev, byte *pnext, int cb)
{
byte *ip = pcurr, pp = pprev, np = pnext;
nuint cnt = 0, end = (nuint)cb - (nuint)Vector256 <byte> .Count;
var voffset = Vector256.Create((byte)0x80);
var vthresh = Vector256.Create(denoiseThreshold);
LoopTop:
do
{
var vcurr = Avx.LoadVector256(ip + cnt);
var vprev = Avx.LoadVector256(pp + cnt);
var vnext = Avx.LoadVector256(np + cnt);
var vdiffp = Avx2.Or(Avx2.SubtractSaturate(vcurr, vprev), Avx2.SubtractSaturate(vprev, vcurr));
var vmaskp = Avx2.CompareEqual(Avx2.Max(vdiffp, vthresh), vthresh);
var vdiffn = Avx2.Or(Avx2.SubtractSaturate(vcurr, vnext), Avx2.SubtractSaturate(vnext, vcurr));
var vmaskn = Avx2.CompareEqual(Avx2.Max(vdiffn, vthresh), vthresh);
var vavgp = Avx2.Average(vcurr, vprev);
var vavgn = Avx2.Average(vcurr, vnext);
var voutval = Avx2.Average(Avx2.BlendVariable(vavgn, vavgp, vmaskp), Avx2.BlendVariable(vavgp, vavgn, vmaskn));
var voutmsk = Avx2.Or(vmaskp, vmaskn);
voutval = Avx2.Average(voutval, Avx2.BlendVariable(voutval, Avx2.Average(vprev, vnext), Avx2.And(vmaskp, vmaskn)));
var vcurrs = Avx2.Xor(vcurr, voffset).AsSByte();
var vprevs = Avx2.Xor(vprev, voffset).AsSByte();
var vnexts = Avx2.Xor(vnext, voffset).AsSByte();
var vsurlt = Avx2.And(Avx2.CompareGreaterThan(vcurrs, vprevs), Avx2.CompareGreaterThan(vcurrs, vnexts));
var vsurgt = Avx2.And(Avx2.CompareGreaterThan(vprevs, vcurrs), Avx2.CompareGreaterThan(vnexts, vcurrs));
voutmsk = Avx2.And(voutmsk, Avx2.Or(vsurlt, vsurgt).AsByte());
voutval = Avx2.BlendVariable(vcurr, voutval, voutmsk);
Avx.Store(ip + cnt, voutval);
cnt += (nuint)Vector256 <byte> .Count;
} while (cnt <= end);
if (cnt < end + (nuint)Vector256 <byte> .Count)
{
cnt = end;
goto LoopTop;
}
}
public void RunClassFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassFldScenario_Load));
fixed(Vector256 <UInt16> *pFld1 = &_fld1)
fixed(Vector256 <UInt16> *pFld2 = &_fld2)
{
var result = Avx2.And(
Avx.LoadVector256((UInt16 *)(pFld1)),
Avx.LoadVector256((UInt16 *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_fld1, _fld2, _dataTable.outArrayPtr);
}
}
public void RunClsVarScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClsVarScenario_Load));
fixed(Vector256 <Int32> *pClsVar1 = &_clsVar1)
fixed(Vector256 <Int32> *pClsVar2 = &_clsVar2)
{
var result = Avx2.And(
Avx.LoadVector256((Int32 *)(pClsVar1)),
Avx.LoadVector256((Int32 *)(pClsVar2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
}
public unsafe void Test_AVX_BitsToBytes()
{
uint x = 0b0000_0001__0010_0011__0100_0101__0110_0111u;
uint y = 0b1000_1001__1010_1011__1100_1101__1110_1111u;
Vector256 <byte> mask1, mask2, zero = Vector256 <byte> .Zero, one, ff;
byte[] mask1_bytes = new byte[]
{
0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
};
byte[] mask2_bytes = new byte[]
{
0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
};
fixed(byte *ptr = mask1_bytes) mask1 = Avx2.LoadVector256(ptr);
fixed(byte *ptr = mask2_bytes) mask2 = Avx2.LoadVector256(ptr);
byte one_byte = 1;
one = Avx2.BroadcastScalarToVector256(&one_byte);
byte ff_byte = 0xff;
ff = Avx2.BroadcastScalarToVector256(&ff_byte);
// ***** load **** //
Vector256 <uint> ux = Avx2.BroadcastScalarToVector256(&y);
Vector256 <byte> bx = ux.AsByte();
Vector256 <byte> shuffled_x = Avx2.Shuffle(bx, mask1);
Vector256 <byte> result_x = Avx2.And(shuffled_x, mask2);
result_x = Avx2.Min(result_x, one);
// ***** store **** //
Vector256 <byte> reverse_x = Avx2.CompareEqual(result_x, zero);
reverse_x = Avx2.AndNot(reverse_x, ff);
uint reversed_x = (uint)Avx2.MoveMask(reverse_x);
Assert.AreEqual(reversed_x, y);
}
private unsafe static void BCnDecodeTileRgb(Span <uint> clut, Span <byte> output, ReadOnlySpan <byte> input)
{
if (Avx2.IsSupported)
{
Span <Vector256 <uint> > outputAsVector256 = MemoryMarshal.Cast <byte, Vector256 <uint> >(output);
Vector256 <uint> shifts0 = Vector256.Create(0u, 2u, 4u, 6u, 8u, 10u, 12u, 14u);
Vector256 <uint> shifts1 = Vector256.Create(16u, 18u, 20u, 22u, 24u, 26u, 28u, 30u);
Vector256 <uint> masks = Vector256.Create(3u);
Vector256 <uint> vClut;
fixed(uint *pClut = &clut[0])
{
vClut = Sse2.LoadVector128(pClut).ToVector256Unsafe();
}
Vector256 <uint> indices0;
fixed(byte *pInput = input)
{
indices0 = Avx2.BroadcastScalarToVector256((uint *)(pInput + 4));
}
Vector256 <uint> indices1 = indices0;
indices0 = Avx2.ShiftRightLogicalVariable(indices0, shifts0);
indices1 = Avx2.ShiftRightLogicalVariable(indices1, shifts1);
indices0 = Avx2.And(indices0, masks);
indices1 = Avx2.And(indices1, masks);
outputAsVector256[0] = Avx2.PermuteVar8x32(vClut, indices0);
outputAsVector256[1] = Avx2.PermuteVar8x32(vClut, indices1);
}
else
{
Span <uint> outputAsUint = MemoryMarshal.Cast <byte, uint>(output);
uint indices = BinaryPrimitives.ReadUInt32LittleEndian(input.Slice(4));
for (int i = 0; i < BlockWidth * BlockHeight; i++, indices >>= 2)
{
outputAsUint[i] = clut[(int)(indices & 3)];
}
}
}
// when 0xED is found, next byte must be no larger than 0x9F
// when 0xF4 is found, next byte must be no larger than 0x8F
// next byte must be continuation, ie sign bit is set, so signed < is ok
static void avxcheckFirstContinuationMax(Vector256 <byte> current_bytes,
Vector256 <byte> off1_current_bytes,
ref Vector256 <byte> has_error)
{
Vector256 <byte> maskED =
Avx2.CompareEqual(off1_current_bytes, Vector256.Create((byte)0xED));
Vector256 <byte> maskF4 =
Avx2.CompareEqual(off1_current_bytes, Vector256.Create((byte)0xF4));
Vector256 <byte> badfollowED = Avx2.And(
Avx2.CompareGreaterThan(current_bytes.AsSByte(), Vector256.Create((byte)0x9F).AsSByte()).AsByte(), maskED);
Vector256 <byte> badfollowF4 = Avx2.And(
Avx2.CompareGreaterThan(current_bytes.AsSByte(), Vector256.Create((byte)0x8F).AsSByte()).AsByte(), maskF4);
has_error =
Avx2.Or(has_error, Avx2.Or(badfollowED, badfollowF4));
}
public void RunClassLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario_Load));
var test = new SimpleBinaryOpTest__AndInt32();
fixed(Vector256 <Int32> *pFld1 = &test._fld1)
fixed(Vector256 <Int32> *pFld2 = &test._fld2)
{
var result = Avx2.And(
Avx.LoadVector256((Int32 *)(pFld1)),
Avx.LoadVector256((Int32 *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
}
private static void denoiseLineAvx2(byte *pcurr, byte *pprev, byte *pnext, nint cb)
{
byte *ip = pcurr, pp = pprev, np = pnext;
nint cnt = 0, end = cb - Vector256 <byte> .Count;
var vthresh = Vector256.Create(denoiseThreshold);
var vones = Avx2.CompareEqual(vthresh, vthresh);
LoopTop:
do
{
var vcurr = Avx.LoadVector256(ip + cnt);
var vprev = Avx.LoadVector256(pp + cnt);
var vnext = Avx.LoadVector256(np + cnt);
var vdiffp = Avx2.Or(Avx2.SubtractSaturate(vcurr, vprev), Avx2.SubtractSaturate(vprev, vcurr));
var vmaskp = Avx2.CompareEqual(Avx2.Max(vdiffp, vthresh), vthresh);
var vdiffn = Avx2.Or(Avx2.SubtractSaturate(vcurr, vnext), Avx2.SubtractSaturate(vnext, vcurr));
var vmaskn = Avx2.CompareEqual(Avx2.Max(vdiffn, vthresh), vthresh);
var vavgp = Avx2.Average(vcurr, vprev);
var vavgn = Avx2.Average(vcurr, vnext);
var voutval = Avx2.Average(Avx2.BlendVariable(vavgn, vavgp, vmaskp), Avx2.BlendVariable(vavgp, vavgn, vmaskn));
var voutmsk = Avx2.Or(vmaskp, vmaskn);
voutval = Avx2.Average(voutval, Avx2.BlendVariable(voutval, Avx2.Average(vprev, vnext), Avx2.And(vmaskp, vmaskn)));
var vsurlt = Avx2.Xor(Avx2.CompareEqual(Avx2.Min(Avx2.Max(vprev, vnext), vcurr), vcurr), vones);
var vsurgt = Avx2.Xor(Avx2.CompareEqual(Avx2.Max(Avx2.Min(vprev, vnext), vcurr), vcurr), vones);
voutmsk = Avx2.And(voutmsk, Avx2.Or(vsurlt, vsurgt));
voutval = Avx2.BlendVariable(vcurr, voutval, voutmsk);
Avx.Store(ip + cnt, voutval);
cnt += Vector256 <byte> .Count;
} while (cnt <= end);
if (cnt < end + Vector256 <byte> .Count)
{
cnt = end;
goto LoopTop;
}
}
private static unsafe (Vector256 <UInt32>[] hi, Vector256 <UInt32>[] lo) Mul(Vector256 <UInt32>[] v, UInt32 n)
{
Vector256 <UInt32>[] w_hi = new Vector256 <UInt32> [v.Length], w_lo = new Vector256 <UInt32> [v.Length];
Vector256 <UInt32> u = Avx2.ConvertToVector256Int64(Vector128.Create(n)).AsUInt32();
Vector256 <UInt32> mask = lower_mask;
fixed(Vector256 <UInt32> *pv = v, pw_hi = w_hi, pw_lo = w_lo)
{
for (int i = 0; i < v.Length; i++)
{
Vector256 <UInt32> c = Avx2.Multiply(pv[i], u).AsUInt32();
pw_hi[i] = Avx2.And(Avx2.Shuffle(c, MM_PERM_CDAB), mask);
pw_lo[i] = Avx2.And(c, mask);
}
}
return(w_hi, w_lo);
}
public static unsafe void ReverseBits(this Span <int> span)
{
var intsReversed = 0;
if (Avx2.IsSupported)
{
fixed(int *ptr = span)
{
var vectorCount = span.Length / 8;
for (int i = 0; i < vectorCount; i++)
{
var vector = Avx.LoadVector256((ptr + intsReversed));
var vector2 = Avx2.And(Avx2.And(vector, Vector256.Create(0xFF00FF)), Vector256.Create(-16711936));
vector =
Avx2.Add(
Avx2.Or(
Avx2.ShiftRightLogical(vector, 8),
Avx2.ShiftLeftLogical(vector, 24)
),
Avx2.Or(
Avx2.ShiftLeftLogical(vector2, 8),
Avx2.ShiftRightLogical(vector2, 24)
)
);
Avx.Store(ptr + intsReversed, vector);
intsReversed += 8;
}
}
}
for (int i = intsReversed; i < span.Length; i++)
{
span[i] = BinaryPrimitives.ReverseEndianness(span[i]);
}
fixed(void *ptr = span)
{
new Span <byte>(ptr, span.Length * 4).ReverseBits();
}
}
public static unsafe byte[] ArgbToAlpha32(byte[] argbBytes)
{
var result = new byte[argbBytes.Length];
var resultPtr = result.ToPtr();
var bytePtr = argbBytes.ToPtr();
for (var ptrIndex = 0; ptrIndex < argbBytes.Length; ptrIndex += 16)
{
var vector128 = Avx2.LoadVector128(bytePtr + ptrIndex).AsUInt32();
vector128 = Avx2.And(vector128, alphaOnlyVector128);
vector128 = Avx2.Or(vector128, Avx2.ShiftRightLogical128BitLane(vector128, 1));
vector128 = Avx2.Or(vector128, Avx2.ShiftRightLogical128BitLane(vector128, 2));
Avx2.Store(resultPtr + ptrIndex, vector128.AsByte());
}
return(result);
}
//[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static __m256 exp256_ps(__m256 V)
{
__m256 x = V;
__m256 tmp = __m256.Zero;
__m256 one = SET(1.0f);
x = Avx2.Min(x, exp_hi);
x = Avx2.Max(x, exp_lo);
__m256 fx = Avx2.Multiply(x, cLOG2EF);
fx = Avx2.Add(fx, SET(0.5f));
tmp = Avx2.Floor(fx);
var mask = Avx2.Compare(tmp, fx, FloatComparisonMode.OrderedGreaterThanSignaling);
mask = Avx2.And(mask, one);
fx = Avx2.Subtract(tmp, mask);
tmp = Avx2.Multiply(fx, cexp_C1);
__m256 z = Avx2.Multiply(fx, cexp_C2);
x = Avx2.Subtract(x, tmp);
x = Avx2.Subtract(x, z);
z = Avx2.Multiply(x, x);
__m256 y = cexp_p0;
y = Fma.MultiplyAdd(y, x, cexp_p1);
y = Fma.MultiplyAdd(y, x, cexp_p2);
y = Fma.MultiplyAdd(y, x, cexp_p3);
y = Fma.MultiplyAdd(y, x, cexp_p4);
y = Fma.MultiplyAdd(y, x, cexp_p5);
y = Fma.MultiplyAdd(y, z, x);
y = Avx2.Add(y, one);
var imm0 = Avx2.ConvertToVector256Int32(fx);
var F7 = Vector256.Create((int)0x7f);
imm0 = Avx2.Add(imm0, F7);
imm0 = Avx2.ShiftLeftLogical(imm0, 23);
__m256 pow2n = Vector256.AsSingle(imm0);
y = Avx2.Multiply(y, pow2n);
return(y);
}
public static void CollectColorBlueTransforms(Span <uint> bgra, int stride, int tileWidth, int tileHeight, int greenToBlue, int redToBlue, Span <int> histo)
{
#if SUPPORTS_RUNTIME_INTRINSICS
if (Avx2.IsSupported && tileWidth >= 16)
{
const int span = 16;
Span <ushort> values = stackalloc ushort[span];
var multsr = Vector256.Create(LosslessUtils.Cst5b(redToBlue));
var multsg = Vector256.Create(LosslessUtils.Cst5b(greenToBlue));
for (int y = 0; y < tileHeight; y++)
{
Span <uint> srcSpan = bgra.Slice(y * stride);
ref uint inputRef = ref MemoryMarshal.GetReference(srcSpan);
for (nint x = 0; x <= tileWidth - span; x += span)
{
nint input0Idx = x;
nint input1Idx = x + (span / 2);
Vector256 <byte> input0 = Unsafe.As <uint, Vector256 <uint> >(ref Unsafe.Add(ref inputRef, input0Idx)).AsByte();
Vector256 <byte> input1 = Unsafe.As <uint, Vector256 <uint> >(ref Unsafe.Add(ref inputRef, input1Idx)).AsByte();
Vector256 <byte> r0 = Avx2.Shuffle(input0, CollectColorBlueTransformsShuffleLowMask256);
Vector256 <byte> r1 = Avx2.Shuffle(input1, CollectColorBlueTransformsShuffleHighMask256);
Vector256 <byte> r = Avx2.Or(r0, r1);
Vector256 <byte> gb0 = Avx2.And(input0, CollectColorBlueTransformsGreenBlueMask256);
Vector256 <byte> gb1 = Avx2.And(input1, CollectColorBlueTransformsGreenBlueMask256);
Vector256 <ushort> gb = Avx2.PackUnsignedSaturate(gb0.AsInt32(), gb1.AsInt32());
Vector256 <byte> g = Avx2.And(gb.AsByte(), CollectColorBlueTransformsGreenMask256);
Vector256 <short> a = Avx2.MultiplyHigh(r.AsInt16(), multsr);
Vector256 <short> b = Avx2.MultiplyHigh(g.AsInt16(), multsg);
Vector256 <byte> c = Avx2.Subtract(gb.AsByte(), b.AsByte());
Vector256 <byte> d = Avx2.Subtract(c, a.AsByte());
Vector256 <byte> e = Avx2.And(d, CollectColorBlueTransformsBlueMask256);
ref ushort outputRef = ref MemoryMarshal.GetReference(values);
Unsafe.As <ushort, Vector256 <ushort> >(ref outputRef) = e.AsUInt16();
for (int i = 0; i < span; i++)
{
++histo[values[i]];
}
}
}