public void RunLclVarScenario_UnsafeRead()
{
var left = Unsafe.Read <Vector128 <Int32> >(_dataTable.inArray1Ptr);
var right = Unsafe.Read <Vector128 <Int32> >(_dataTable.inArray2Ptr);
var result = Sse41.PackUnsignedSaturate(left, right);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(left, right, _dataTable.outArrayPtr);
}
public void RunClassLclFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario));
var test = new HorizontalBinaryOpTest__PackUnsignedSaturateUInt16();
var result = Sse41.PackUnsignedSaturate(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunStructLclFldScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunStructLclFldScenario));
var test = TestStruct.Create();
var result = Sse41.PackUnsignedSaturate(test._fld1, test._fld2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunClsVarScenario()
{
var result = Sse41.PackUnsignedSaturate(
_clsVar1,
_clsVar2
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
public void RunBasicScenario_LoadAligned()
{
var result = Sse41.PackUnsignedSaturate(
Sse2.LoadAlignedVector128((Int32 *)(_dataTable.inArray1Ptr)),
Sse2.LoadAlignedVector128((Int32 *)(_dataTable.inArray2Ptr))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunBasicScenario_UnsafeRead()
{
var result = Sse41.PackUnsignedSaturate(
Unsafe.Read <Vector128 <Int32> >(_dataTable.inArray1Ptr),
Unsafe.Read <Vector128 <Int32> >(_dataTable.inArray2Ptr)
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunLclVarScenario_LoadAligned()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunLclVarScenario_LoadAligned));
var left = Sse2.LoadAlignedVector128((Int32 *)(_dataTable.inArray1Ptr));
var right = Sse2.LoadAlignedVector128((Int32 *)(_dataTable.inArray2Ptr));
var result = Sse41.PackUnsignedSaturate(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 <Vector128 <Int32> >(_dataTable.inArray1Ptr);
var op2 = Unsafe.Read <Vector128 <Int32> >(_dataTable.inArray2Ptr);
var result = Sse41.PackUnsignedSaturate(op1, op2);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(op1, op2, _dataTable.outArrayPtr);
}
private static uint32_t parse_eight_digits_unrolled(bytechar *chars)
{
// this actually computes *16* values so we are being wasteful.
Vector128 <sbyte> ascii0 = Vector128.Create((bytechar)'0');
Vector128 <sbyte> input = Sse2.Subtract(Sse2.LoadVector128(chars), ascii0);
Vector128 <short> t1 = Ssse3.MultiplyAddAdjacent(input.AsByte(), mul_1_10);
Vector128 <int> t2 = Sse2.MultiplyAddAdjacent(t1, mul_1_100);
Vector128 <ushort> t3 = Sse41.PackUnsignedSaturate(t2, t2);
Vector128 <int> t4 = Sse2.MultiplyAddAdjacent(t3.AsInt16(), mul_1_10000);
return(Sse2.ConvertToUInt32(t4.AsUInt32())); // only captures the sum of the first 8 digits, drop the rest
}
public void RunBasicScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunBasicScenario_Load));
var result = Sse41.PackUnsignedSaturate(
Sse2.LoadVector128((Int32 *)(_dataTable.inArray1Ptr)),
Sse2.LoadVector128((Int32 *)(_dataTable.inArray2Ptr))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_dataTable.inArray1Ptr, _dataTable.inArray2Ptr, _dataTable.outArrayPtr);
}
public void RunClsVarScenario()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClsVarScenario));
var result = Sse41.PackUnsignedSaturate(
_clsVar1,
_clsVar2
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_clsVar1, _clsVar2, _dataTable.outArrayPtr);
}
public void RunStructLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunStructLclFldScenario_Load));
var test = TestStruct.Create();
var result = Sse41.PackUnsignedSaturate(
Sse2.LoadVector128((Int32 *)(&test._fld1)),
Sse2.LoadVector128((Int32 *)(&test._fld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
public void RunStructFldScenario_Load(HorizontalBinaryOpTest__PackUnsignedSaturateUInt16 testClass)
{
fixed(Vector128 <Int32> *pFld1 = &_fld1)
fixed(Vector128 <Int32> *pFld2 = &_fld2)
{
var result = Sse41.PackUnsignedSaturate(
Sse2.LoadVector128((Int32 *)(pFld1)),
Sse2.LoadVector128((Int32 *)(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(Vector128 <Int32> *pFld1 = &_fld1)
fixed(Vector128 <Int32> *pFld2 = &_fld2)
{
var result = Sse41.PackUnsignedSaturate(
Sse2.LoadVector128((Int32 *)(pFld1)),
Sse2.LoadVector128((Int32 *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(_fld1, _fld2, _dataTable.outArrayPtr);
}
}
public void RunClassLclFldScenario_Load()
{
TestLibrary.TestFramework.BeginScenario(nameof(RunClassLclFldScenario_Load));
var test = new HorizontalBinaryOpTest__PackUnsignedSaturateUInt16();
fixed(Vector128 <Int32> *pFld1 = &test._fld1)
fixed(Vector128 <Int32> *pFld2 = &test._fld2)
{
var result = Sse41.PackUnsignedSaturate(
Sse2.LoadVector128((Int32 *)(pFld1)),
Sse2.LoadVector128((Int32 *)(pFld2))
);
Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateResult(test._fld1, test._fld2, _dataTable.outArrayPtr);
}
}
private unsafe static void BCnDecodeTileAlpha(Span <byte> output, Span <byte> rPal, ulong rI)
{
if (Avx2.IsSupported)
{
Span <Vector128 <byte> > outputAsVector128 = MemoryMarshal.Cast <byte, Vector128 <byte> >(output);
Vector128 <uint> shifts = Vector128.Create(0u, 3u, 6u, 9u);
Vector128 <uint> masks = Vector128.Create(7u);
Vector128 <byte> vClut;
fixed(byte *pRPal = rPal)
{
vClut = Sse2.LoadScalarVector128((ulong *)pRPal).AsByte();
}
Vector128 <uint> indices0 = Vector128.Create((uint)rI);
Vector128 <uint> indices1 = Vector128.Create((uint)(rI >> 24));
Vector128 <uint> indices00 = Avx2.ShiftRightLogicalVariable(indices0, shifts);
Vector128 <uint> indices10 = Avx2.ShiftRightLogicalVariable(indices1, shifts);
Vector128 <uint> indices01 = Sse2.ShiftRightLogical(indices00, 12);
Vector128 <uint> indices11 = Sse2.ShiftRightLogical(indices10, 12);
indices00 = Sse2.And(indices00, masks);
indices10 = Sse2.And(indices10, masks);
indices01 = Sse2.And(indices01, masks);
indices11 = Sse2.And(indices11, masks);
Vector128 <ushort> indicesW0 = Sse41.PackUnsignedSaturate(indices00.AsInt32(), indices01.AsInt32());
Vector128 <ushort> indicesW1 = Sse41.PackUnsignedSaturate(indices10.AsInt32(), indices11.AsInt32());
Vector128 <byte> indices = Sse2.PackUnsignedSaturate(indicesW0.AsInt16(), indicesW1.AsInt16());
outputAsVector128[0] = Ssse3.Shuffle(vClut, indices);
}
else
{
for (int i = 0; i < BlockWidth * BlockHeight; i++, rI >>= 3)
{
output[i] = rPal[(int)(rI & 7)];
}
}
}
static Vector128 <ushort> ToUshortScalar(int i)
{
var xmm = Vector128.Create(i);
return(Sse41.PackUnsignedSaturate(xmm, xmm));
}
private static Vector128 <byte> PackUnsignedSaturate(Vector128 <int> value, Vector128 <int> zero)
{
return(Sse2.PackUnsignedSaturate(Sse41.PackUnsignedSaturate(value, zero).AsInt16(), zero.AsInt16()));
}
public static Vector128 <ushort> _mm_packus_epi32(Vector128 <int> left, Vector128 <int> right)
{
return(Sse41.PackUnsignedSaturate(left, right));
}
private unsafe static void WriteNv12(ResourceManager rm, Surface input, ref OutputSurfaceConfig config, ref PlaneOffsets offsets)
{
int gobBlocksInY = 1 << config.OutBlkHeight;
bool outLinear = config.OutBlkKind == 0;
int width = Math.Min(config.OutLumaWidth + 1, input.Width);
int height = Math.Min(config.OutLumaHeight + 1, input.Height);
int yStride = GetPitch(config.OutLumaWidth + 1, 1);
int dstYIndex = rm.BufferPool.Rent((config.OutLumaHeight + 1) * yStride, out Span <byte> dstY);
if (Sse41.IsSupported)
{
Vector128 <ushort> mask = Vector128.Create(0xffffUL).AsUInt16();
int widthTrunc = width & ~0xf;
int strideGap = yStride - width;
fixed(Pixel *srcPtr = input.Data)
{
Pixel *ip = srcPtr;
fixed(byte *dstPtr = dstY)
{
byte *op = dstPtr;
for (int y = 0; y < height; y++, ip += input.Width)
{
int x = 0;
for (; x < widthTrunc; x += 16)
{
byte *baseOffset = (byte *)(ip + (ulong)(uint)x);
Vector128 <ushort> pixelp1 = Sse2.LoadVector128((ushort *)baseOffset);
Vector128 <ushort> pixelp2 = Sse2.LoadVector128((ushort *)(baseOffset + 0x10));
Vector128 <ushort> pixelp3 = Sse2.LoadVector128((ushort *)(baseOffset + 0x20));
Vector128 <ushort> pixelp4 = Sse2.LoadVector128((ushort *)(baseOffset + 0x30));
Vector128 <ushort> pixelp5 = Sse2.LoadVector128((ushort *)(baseOffset + 0x40));
Vector128 <ushort> pixelp6 = Sse2.LoadVector128((ushort *)(baseOffset + 0x50));
Vector128 <ushort> pixelp7 = Sse2.LoadVector128((ushort *)(baseOffset + 0x60));
Vector128 <ushort> pixelp8 = Sse2.LoadVector128((ushort *)(baseOffset + 0x70));
pixelp1 = Sse2.And(pixelp1, mask);
pixelp2 = Sse2.And(pixelp2, mask);
pixelp3 = Sse2.And(pixelp3, mask);
pixelp4 = Sse2.And(pixelp4, mask);
pixelp5 = Sse2.And(pixelp5, mask);
pixelp6 = Sse2.And(pixelp6, mask);
pixelp7 = Sse2.And(pixelp7, mask);
pixelp8 = Sse2.And(pixelp8, mask);
Vector128 <ushort> pixelq1 = Sse41.PackUnsignedSaturate(pixelp1.AsInt32(), pixelp2.AsInt32());
Vector128 <ushort> pixelq2 = Sse41.PackUnsignedSaturate(pixelp3.AsInt32(), pixelp4.AsInt32());
Vector128 <ushort> pixelq3 = Sse41.PackUnsignedSaturate(pixelp5.AsInt32(), pixelp6.AsInt32());
Vector128 <ushort> pixelq4 = Sse41.PackUnsignedSaturate(pixelp7.AsInt32(), pixelp8.AsInt32());
pixelq1 = Sse41.PackUnsignedSaturate(pixelq1.AsInt32(), pixelq2.AsInt32());
pixelq2 = Sse41.PackUnsignedSaturate(pixelq3.AsInt32(), pixelq4.AsInt32());
pixelq1 = Sse2.ShiftRightLogical(pixelq1, 2);
pixelq2 = Sse2.ShiftRightLogical(pixelq2, 2);
Vector128 <byte> pixel = Sse2.PackUnsignedSaturate(pixelq1.AsInt16(), pixelq2.AsInt16());
Sse2.Store(op, pixel);
op += 0x10;
}
for (; x < width; x++)
{
Pixel *px = ip + (uint)x;
*op++ = Downsample(px->R);
}
op += strideGap;
}
}
}
}
else
{
for (int y = 0; y < height; y++)
{
for (int x = 0; x < width; x++)
{
dstY[y * yStride + x] = Downsample(input.GetR(x, y));
}
}
}
WriteBuffer(
rm,
dstY,
offsets.LumaOffset,
outLinear,
config.OutLumaWidth + 1,
config.OutLumaHeight + 1,
1,
gobBlocksInY);
rm.BufferPool.Return(dstYIndex);
int uvWidth = Math.Min(config.OutChromaWidth + 1, (width + 1) >> 1);
int uvHeight = Math.Min(config.OutChromaHeight + 1, (height + 1) >> 1);
int uvStride = GetPitch(config.OutChromaWidth + 1, 2);
int dstUvIndex = rm.BufferPool.Rent((config.OutChromaHeight + 1) * uvStride, out Span <byte> dstUv);
if (Sse2.IsSupported)
{
int widthTrunc = uvWidth & ~7;
int strideGap = uvStride - uvWidth * 2;
fixed(Pixel *srcPtr = input.Data)
{
Pixel *ip = srcPtr;
fixed(byte *dstPtr = dstUv)
{
byte *op = dstPtr;
for (int y = 0; y < uvHeight; y++, ip += input.Width * 2)
{
int x = 0;
for (; x < widthTrunc; x += 8)
{
byte *baseOffset = (byte *)ip + (ulong)(uint)x * 16;
Vector128 <uint> pixel1 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x02));
Vector128 <uint> pixel2 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x12));
Vector128 <uint> pixel3 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x22));
Vector128 <uint> pixel4 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x32));
Vector128 <uint> pixel5 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x42));
Vector128 <uint> pixel6 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x52));
Vector128 <uint> pixel7 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x62));
Vector128 <uint> pixel8 = Sse2.LoadScalarVector128((uint *)(baseOffset + 0x72));
Vector128 <uint> pixel12 = Sse2.UnpackLow(pixel1, pixel2);
Vector128 <uint> pixel34 = Sse2.UnpackLow(pixel3, pixel4);
Vector128 <uint> pixel56 = Sse2.UnpackLow(pixel5, pixel6);
Vector128 <uint> pixel78 = Sse2.UnpackLow(pixel7, pixel8);
Vector128 <ulong> pixel1234 = Sse2.UnpackLow(pixel12.AsUInt64(), pixel34.AsUInt64());
Vector128 <ulong> pixel5678 = Sse2.UnpackLow(pixel56.AsUInt64(), pixel78.AsUInt64());
pixel1234 = Sse2.ShiftRightLogical(pixel1234, 2);
pixel5678 = Sse2.ShiftRightLogical(pixel5678, 2);
Vector128 <byte> pixel = Sse2.PackUnsignedSaturate(pixel1234.AsInt16(), pixel5678.AsInt16());
Sse2.Store(op, pixel);
op += 0x10;
}
for (; x < uvWidth; x++)
{
Pixel *px = ip + (uint)(x << 1);
*op++ = Downsample(px->G);
*op++ = Downsample(px->B);
}
op += strideGap;
}
}
}
}
else
{
for (int y = 0; y < uvHeight; y++)
{
for (int x = 0; x < uvWidth; x++)
{
int xx = x << 1;
int yy = y << 1;
int uvOffs = y * uvStride + xx;
dstUv[uvOffs + 0] = Downsample(input.GetG(xx, yy));
dstUv[uvOffs + 1] = Downsample(input.GetB(xx, yy));
}
}
}
WriteBuffer(
rm,
dstUv,
offsets.ChromaUOffset,
outLinear,
config.OutChromaWidth + 1,
config.OutChromaHeight + 1, 2,
gobBlocksInY);
rm.BufferPool.Return(dstUvIndex);
}
