public override MemoryHandle Pin(int byteOffset = 0)
{
unsafe
{
Retain(); // this checks IsDisposed
try
{
if ((IntPtr.Size == 4 && (uint)byteOffset > (uint)_array.Length * (uint)Unsafe.SizeOf <T>()) ||
(IntPtr.Size != 4 && (ulong)byteOffset > (uint)_array.Length * (ulong)Unsafe.SizeOf <T>()))
{
throw new ArgumentOutOfRangeException(nameof(byteOffset));
}
var handle = GCHandle.Alloc(_array, GCHandleType.Pinned);
return(new MemoryHandle(this, Unsafe.Add <byte>((void *)handle.AddrOfPinnedObject(), _offset + byteOffset), handle));
}
catch
{
Release();
throw;
}
}
}
string Ctor(char[] value, int startIndex, int length)
{
if (value == null)
{
throw new ArgumentNullException(nameof(value));
}
if (startIndex < 0)
{
throw new ArgumentOutOfRangeException(nameof(startIndex), SR.ArgumentOutOfRange_StartIndex);
}
if (length < 0)
{
throw new ArgumentOutOfRangeException(nameof(length), SR.ArgumentOutOfRange_NegativeLength);
}
if (startIndex > value.Length - length)
{
throw new ArgumentOutOfRangeException(nameof(startIndex), SR.ArgumentOutOfRange_Index);
}
if (length == 0)
{
return(Empty);
}
string result = FastAllocateString(length);
Buffer.Memmove(
elementCount: (uint)result.Length, // derefing Length now allows JIT to prove 'result' not null below
destination: ref result._firstChar,
source: ref Unsafe.Add(ref MemoryMarshal.GetArrayDataReference(value), startIndex));
return(result);
}
private static int CompareOrdinalHelper(string strA, int indexA, int countA, string strB, int indexB, int countB)
{
Debug.Assert(strA != null);
Debug.Assert(strB != null);
Debug.Assert(indexA >= 0 && indexB >= 0);
Debug.Assert(countA >= 0 && countB >= 0);
Debug.Assert(indexA + countA <= strA.Length && indexB + countB <= strB.Length);
return(SpanHelpers.SequenceCompareTo(ref Unsafe.Add(ref strA.GetRawStringData(), indexA), countA, ref Unsafe.Add(ref strB.GetRawStringData(), indexB), countB));
}
/// <summary>
/// Computes a 64-hash using the Marvin algorithm.
/// </summary>
public static long ComputeHash(ref byte data, int count, ulong seed)
{
uint ucount = (uint)count;
uint p0 = (uint)seed;
uint p1 = (uint)(seed >> 32);
int byteOffset = 0; // declared as signed int so we don't have to cast everywhere (it's passed to Unsafe.Add() and used for nothing else.)
while (ucount >= 8)
{
p0 += Unsafe.As <byte, uint>(ref Unsafe.Add(ref data, byteOffset));
Block(ref p0, ref p1);
p0 += Unsafe.As <byte, uint>(ref Unsafe.Add(ref data, byteOffset + 4));
Block(ref p0, ref p1);
byteOffset += 8;
ucount -= 8;
}
switch (ucount)
{
case 4:
p0 += Unsafe.As <byte, uint>(ref Unsafe.Add(ref data, byteOffset));
Block(ref p0, ref p1);
goto case 0;
case 0:
p0 += 0x80u;
break;
case 5:
p0 += Unsafe.As <byte, uint>(ref Unsafe.Add(ref data, byteOffset));
byteOffset += 4;
Block(ref p0, ref p1);
goto case 1;
case 1:
p0 += 0x8000u | Unsafe.Add(ref data, byteOffset);
break;
case 6:
p0 += Unsafe.As <byte, uint>(ref Unsafe.Add(ref data, byteOffset));
byteOffset += 4;
Block(ref p0, ref p1);
goto case 2;
case 2:
p0 += 0x800000u | Unsafe.As <byte, ushort>(ref Unsafe.Add(ref data, byteOffset));
break;
case 7:
p0 += Unsafe.As <byte, uint>(ref Unsafe.Add(ref data, byteOffset));
byteOffset += 4;
Block(ref p0, ref p1);
goto case 3;
case 3:
p0 += 0x80000000u | (((uint)(Unsafe.Add(ref data, byteOffset + 2))) << 16) | (uint)(Unsafe.As <byte, ushort>(ref Unsafe.Add(ref data, byteOffset)));
break;
default:
Debug.Fail("Should not get here.");
break;
}
Block(ref p0, ref p1);
Block(ref p0, ref p1);
return((((long)p1) << 32) | p0);
}
private static int GetIndexOfFirstInvalidUtf8Sequence(ref byte inputBuffer, int inputLength, out int scalarCount, out int surrogatePairCount)
{
// The fields below control where we read from the buffer.
IntPtr inputBufferCurrentOffset = IntPtr.Zero;
int tempScalarCount = inputLength;
int tempSurrogatePairCount = 0;
// If the sequence is long enough, try running vectorized "is this sequence ASCII?"
// logic. We perform a small test of the first few bytes to make sure they're all
// ASCII before we incur the cost of invoking the vectorized code path.
if (Vector.IsHardwareAccelerated)
{
if (IntPtr.Size >= 8)
{
// Test first 16 bytes and check for all-ASCII.
if ((inputLength >= 2 * sizeof(ulong) + 3 * Vector <byte> .Count) && QWordAllBytesAreAscii(ReadAndFoldTwoQWordsUnaligned(ref inputBuffer)))
{
inputBufferCurrentOffset = ConsumeAsciiBytesVectorized(ref Unsafe.Add(ref inputBuffer, 2 * sizeof(ulong)), inputLength - 2 * sizeof(ulong)) + 2 * sizeof(ulong);
}
}
else
{
// Test first 8 bytes and check for all-ASCII.
if ((inputLength >= 2 * sizeof(uint) + 3 * Vector <byte> .Count) && DWordAllBytesAreAscii(ReadAndFoldTwoDWordsUnaligned(ref inputBuffer)))
{
inputBufferCurrentOffset = ConsumeAsciiBytesVectorized(ref Unsafe.Add(ref inputBuffer, 2 * sizeof(uint)), inputLength - 2 * sizeof(uint)) + 2 * sizeof(uint);
}
}
}
int inputBufferRemainingBytes = inputLength - ConvertIntPtrToInt32WithoutOverflowCheck(inputBufferCurrentOffset);
// Begin the main loop.
#if DEBUG
long lastOffsetProcessed = -1; // used for invariant checking in debug builds
#endif
while (inputBufferRemainingBytes >= sizeof(uint))
{
// Read 32 bits at a time. This is enough to hold any possible UTF8-encoded scalar.
Debug.Assert(inputLength - (int)inputBufferCurrentOffset >= sizeof(uint));
uint thisDWord = Unsafe.ReadUnaligned <uint>(ref Unsafe.Add(ref inputBuffer, inputBufferCurrentOffset));
AfterReadDWord:
#if DEBUG
Debug.Assert(lastOffsetProcessed < (long)inputBufferCurrentOffset, "Algorithm should've made forward progress since last read.");
lastOffsetProcessed = (long)inputBufferCurrentOffset;
#endif
// First, check for the common case of all-ASCII bytes.
if (DWordAllBytesAreAscii(thisDWord))
{
// We read an all-ASCII sequence.
inputBufferCurrentOffset += 4;
inputBufferRemainingBytes -= 4;
// If we saw a sequence of all ASCII, there's a good chance a significant amount of following data is also ASCII.
// Below is basically unrolled loops with poor man's vectorization.
if (inputBufferRemainingBytes >= 5 * sizeof(uint))
{
// The JIT produces better codegen for aligned reads than it does for
// unaligned reads, and we want the processor to operate at maximum
// efficiency in the loop that follows, so we'll align the references
// now. It's OK to do this without pinning because the GC will never
// move a heap-allocated object in a manner that messes with its
// alignment.
{
ref byte refToCurrentDWord = ref Unsafe.Add(ref inputBuffer, inputBufferCurrentOffset);
thisDWord = Unsafe.ReadUnaligned <uint>(ref refToCurrentDWord);
if (!DWordAllBytesAreAscii(thisDWord))
{
goto AfterReadDWordSkipAllBytesAsciiCheck;
}
int adjustment = GetNumberOfBytesToNextDWordAlignment(ref refToCurrentDWord);
inputBufferCurrentOffset += adjustment;
// will adjust 'bytes remaining' value after below loop
}
// At this point, the input buffer offset points to an aligned DWORD.
// We also know that there's enough room to read at least four DWORDs from the stream.
IntPtr inputBufferFinalOffsetAtWhichCanSafelyLoop = (IntPtr)(inputLength - 4 * sizeof(uint));
do
{
ref uint currentReadPosition = ref Unsafe.As <byte, uint>(ref Unsafe.Add(ref inputBuffer, inputBufferCurrentOffset));
if (!DWordAllBytesAreAscii(currentReadPosition | Unsafe.Add(ref currentReadPosition, 1)))
{
goto LoopTerminatedEarlyDueToNonAsciiData;
}
if (!DWordAllBytesAreAscii(Unsafe.Add(ref currentReadPosition, 2) | Unsafe.Add(ref currentReadPosition, 3)))
{
inputBufferCurrentOffset += 2 * sizeof(uint);
goto LoopTerminatedEarlyDueToNonAsciiData;
}
inputBufferCurrentOffset += 4 * sizeof(uint);
} while (IntPtrIsLessThanOrEqualTo(inputBufferCurrentOffset, inputBufferFinalOffsetAtWhichCanSafelyLoop));
inputBufferRemainingBytes = inputLength - ConvertIntPtrToInt32WithoutOverflowCheck(inputBufferCurrentOffset);
continue; // need to perform a bounds check because we might be running out of data
LoopTerminatedEarlyDueToNonAsciiData:
// We know that there's *at least* two DWORDs of data remaining in the buffer.
// We also know that one of them (or both of them) contains non-ASCII data somewhere.
// Let's perform a quick check here to bypass the logic at the beginning of the main loop.
thisDWord = Unsafe.As <byte, uint>(ref Unsafe.Add(ref inputBuffer, inputBufferCurrentOffset));
if (DWordAllBytesAreAscii(thisDWord))
{
inputBufferCurrentOffset += 4;
thisDWord = Unsafe.As <byte, uint>(ref Unsafe.Add(ref inputBuffer, inputBufferCurrentOffset));
}
inputBufferRemainingBytes = inputLength - ConvertIntPtrToInt32WithoutOverflowCheck(inputBufferCurrentOffset);
goto AfterReadDWordSkipAllBytesAsciiCheck;
}
continue;
}
public static unsafe void ClearWithReferences(ref IntPtr ip, nuint pointerSizeLength)
{
Debug.Assert((int)Unsafe.AsPointer(ref ip) % sizeof(IntPtr) == 0, "Should've been aligned on natural word boundary.");
// First write backward 8 natural words at a time.
// Writing backward allows us to get away with only simple modifications to the
// mov instruction's base and index registers between loop iterations.
for (; pointerSizeLength >= 8; pointerSizeLength -= 8)
{
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -1) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -2) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -3) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -4) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -5) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -6) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -7) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -8) = default(IntPtr);
}
Debug.Assert(pointerSizeLength <= 7);
// The logic below works by trying to minimize the number of branches taken for any
// given range of lengths. For example, the lengths [ 4 .. 7 ] are handled by a single
// branch, [ 2 .. 3 ] are handled by a single branch, and [ 1 ] is handled by a single
// branch.
//
// We can write both forward and backward as a perf improvement. For example,
// the lengths [ 4 .. 7 ] can be handled by zeroing out the first four natural
// words and the last 3 natural words. In the best case (length = 7), there are
// no overlapping writes. In the worst case (length = 4), there are three
// overlapping writes near the middle of the buffer. In perf testing, the
// penalty for performing duplicate writes is less expensive than the penalty
// for complex branching.
if (pointerSizeLength >= 4)
{
goto Write4To7;
}
else if (pointerSizeLength >= 2)
{
goto Write2To3;
}
else if (pointerSizeLength > 0)
{
goto Write1;
}
else
{
return; // nothing to write
}
Write4To7:
Debug.Assert(pointerSizeLength >= 4);
// Write first four and last three.
Unsafe.Add(ref ip, 2) = default(IntPtr);
Unsafe.Add(ref ip, 3) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -3) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -2) = default(IntPtr);
Write2To3:
Debug.Assert(pointerSizeLength >= 2);
// Write first two and last one.
Unsafe.Add(ref ip, 1) = default(IntPtr);
Unsafe.Add(ref Unsafe.Add(ref ip, (IntPtr)pointerSizeLength), -1) = default(IntPtr);
Write1:
Debug.Assert(pointerSizeLength >= 1);
// Write only element.
ip = default(IntPtr);
}
public static int ComputeHash32(ref byte data, uint count, uint p0, uint p1)
{
// Control flow of this method generally flows top-to-bottom, trying to
// minimize the number of branches taken for large (>= 8 bytes, 4 chars) inputs.
// If small inputs (< 8 bytes, 4 chars) are given, this jumps to a "small inputs"
// handler at the end of the method.
if (count < 8)
{
// We can't run the main loop, but we might still have 4 or more bytes available to us.
// If so, jump to the 4 .. 7 bytes logic immediately after the main loop.
if (count >= 4)
{
goto Between4And7BytesRemain;
}
else
{
goto InputTooSmallToEnterMainLoop;
}
}
// Main loop - read 8 bytes at a time.
// The block function is unrolled 2x in this loop.
uint loopCount = count / 8;
Debug.Assert(loopCount > 0, "Shouldn't reach this code path for small inputs.");
do
{
// Most x86 processors have two dispatch ports for reads, so we can read 2x 32-bit
// values in parallel. We opt for this instead of a single 64-bit read since the
// typical use case for Marvin32 is computing String hash codes, and the particular
// layout of String instances means the starting data is never 8-byte aligned when
// running in a 64-bit process.
p0 += Unsafe.ReadUnaligned <uint>(ref data);
uint nextUInt32 = Unsafe.ReadUnaligned <uint>(ref Unsafe.AddByteOffset(ref data, 4));
// One block round for each of the 32-bit integers we just read, 2x rounds total.
Block(ref p0, ref p1);
p0 += nextUInt32;
Block(ref p0, ref p1);
// Bump the data reference pointer and decrement the loop count.
// Decrementing by 1 every time and comparing against zero allows the JIT to produce
// better codegen compared to a standard 'for' loop with an incrementing counter.
// Requires https://github.com/dotnet/coreclr/issues/7566 to be addressed first
// before we can realize the full benefits of this.
data = ref Unsafe.AddByteOffset(ref data, 8);
} while (--loopCount > 0);
// n.b. We've not been updating the original 'count' parameter, so its actual value is
// still the original data length. However, we can still rely on its least significant
// 3 bits to tell us how much data remains (0 .. 7 bytes) after the loop above is
// completed.
if ((count & 0b_0100) == 0)
{
goto DoFinalPartialRead;
}
Between4And7BytesRemain:
// If after finishing the main loop we still have 4 or more leftover bytes, or if we had
// 4 .. 7 bytes to begin with and couldn't enter the loop in the first place, we need to
// consume 4 bytes immediately and send them through one round of the block function.
Debug.Assert(count >= 4, "Only should've gotten here if the original count was >= 4.");
p0 += Unsafe.ReadUnaligned <uint>(ref data);
Block(ref p0, ref p1);
DoFinalPartialRead:
// Finally, we have 0 .. 3 bytes leftover. Since we know the original data length was at
// least 4 bytes (smaller lengths are handled at the end of this routine), we can safely
// read the 4 bytes at the end of the buffer without reading past the beginning of the
// original buffer. This necessarily means the data we're about to read will overlap with
// some data we've already processed, but we can handle that below.
Debug.Assert(count >= 4, "Only should've gotten here if the original count was >= 4.");
// Read the last 4 bytes of the buffer.
uint partialResult = Unsafe.ReadUnaligned <uint>(ref Unsafe.Add(ref Unsafe.AddByteOffset(ref data, (nuint)count & 7), -4));
// The 'partialResult' local above contains any data we have yet to read, plus some number
// of bytes which we've already read from the buffer. An example of this is given below
// for little-endian architectures. In this table, AA BB CC are the bytes which we still
// need to consume, and ## are bytes which we want to throw away since we've already
// consumed them as part of a previous read.
//
// (partialResult contains) (we want it to contain)
// count mod 4 = 0 -> [ ## ## ## ## | ] -> 0x####_#### -> 0x0000_0080
// count mod 4 = 1 -> [ ## ## ## ## | AA ] -> 0xAA##_#### -> 0x0000_80AA
// count mod 4 = 2 -> [ ## ## ## ## | AA BB ] -> 0xBBAA_#### -> 0x0080_BBAA
// count mod 4 = 3 -> [ ## ## ## ## | AA BB CC ] -> 0xCCBB_AA## -> 0x80CC_BBAA
count = ~count << 3;
if (BitConverter.IsLittleEndian)
{
partialResult >>= 8; // make some room for the 0x80 byte
partialResult |= 0x8000_0000u; // put the 0x80 byte at the beginning
partialResult >>= (int)count & 0x1F; // shift out all previously consumed bytes
}
else
{
partialResult <<= 8; // make some room for the 0x80 byte
partialResult |= 0x80u; // put the 0x80 byte at the end
partialResult <<= (int)count & 0x1F; // shift out all previously consumed bytes
}
DoFinalRoundsAndReturn:
// Now that we've computed the final partial result, merge it in and run two rounds of
// the block function to finish out the Marvin algorithm.
p0 += partialResult;
Block(ref p0, ref p1);
Block(ref p0, ref p1);
return((int)(p1 ^ p0));
InputTooSmallToEnterMainLoop:
// We had only 0 .. 3 bytes to begin with, so we can't perform any 32-bit reads.
// This means that we're going to be building up the final result right away and
// will only ever run two rounds total of the block function. Let's initialize
// the partial result to "no data".
if (BitConverter.IsLittleEndian)
{
partialResult = 0x80u;
}
else
{
partialResult = 0x80000000u;
}
if ((count & 0b_0001) != 0)
{
// If the buffer is 1 or 3 bytes in length, let's read a single byte now
// and merge it into our partial result. This will result in partialResult
// having one of the two values below, where AA BB CC are the buffer bytes.
//
// (little-endian / big-endian)
// [ AA ] -> 0x0000_80AA / 0xAA80_0000
// [ AA BB CC ] -> 0x0000_80CC / 0xCC80_0000
partialResult = Unsafe.AddByteOffset(ref data, (nuint)count & 2);
if (BitConverter.IsLittleEndian)
{
partialResult |= 0x8000;
}
else
{
partialResult <<= 24;
partialResult |= 0x800000u;
}
}
if ((count & 0b_0010) != 0)
{
// If the buffer is 2 or 3 bytes in length, let's read a single ushort now
// and merge it into the partial result. This will result in partialResult
// having one of the two values below, where AA BB CC are the buffer bytes.
//
// (little-endian / big-endian)
// [ AA BB ] -> 0x0080_BBAA / 0xAABB_8000
// [ AA BB CC ] -> 0x80CC_BBAA / 0xAABB_CC80 (carried over from above)
if (BitConverter.IsLittleEndian)
{
partialResult <<= 16;
partialResult |= (uint)Unsafe.ReadUnaligned <ushort>(ref data);
}
else
{
partialResult |= (uint)Unsafe.ReadUnaligned <ushort>(ref data);
partialResult = BitOperations.RotateLeft(partialResult, 16);
}
}
// Everything is consumed! Go perform the final rounds and return.
goto DoFinalRoundsAndReturn;
}
public static void WriteHexByte(byte value, ref char buffer, int index)
{
Unsafe.Add(ref buffer, index) = HexTable[value >> 4];
Unsafe.Add(ref buffer, index + 1) = HexTable[value & 0xF];
}
public void IndexWithUnsafeReferenceArithmeticsOnArray0Impl(int x, int y, Vector4 v)
{
int elementOffset = (y * this.width) + x;
Unsafe.Add(ref this.array[0], elementOffset) = v;
}
public unsafe char this[int index] {
[Intrinsic]
get {
return(Unsafe.Add(ref _firstChar, index));
}
}
private static void WriteThreeBytes(ref byte destBytes, int i0)
{
destBytes = (byte)(i0 >> 16);
Unsafe.Add(ref destBytes, 1) = (byte)(i0 >> 8);
Unsafe.Add(ref destBytes, 2) = (byte)i0;
}
/// <inheritdoc />
protected override void OnFrameApply(ImageFrame <TPixel> source, Rectangle sourceRectangle, Configuration configuration)
{
DenseMatrix <float>[] kernels = { this.North, this.NorthWest, this.West, this.SouthWest, this.South, this.SouthEast, this.East, this.NorthEast };
int startY = sourceRectangle.Y;
int endY = sourceRectangle.Bottom;
int startX = sourceRectangle.X;
int endX = sourceRectangle.Right;
// Align start/end positions.
int minX = Math.Max(0, startX);
int maxX = Math.Min(source.Width, endX);
int minY = Math.Max(0, startY);
int maxY = Math.Min(source.Height, endY);
// we need a clean copy for each pass to start from
using (ImageFrame <TPixel> cleanCopy = source.Clone())
{
new ConvolutionProcessor <TPixel>(kernels[0]).Apply(source, sourceRectangle, configuration);
if (kernels.Length == 1)
{
return;
}
int shiftY = startY;
int shiftX = startX;
// Reset offset if necessary.
if (minX > 0)
{
shiftX = 0;
}
if (minY > 0)
{
shiftY = 0;
}
// Additional runs.
// ReSharper disable once ForCanBeConvertedToForeach
for (int i = 1; i < kernels.Length; i++)
{
using (ImageFrame <TPixel> pass = cleanCopy.Clone())
{
new ConvolutionProcessor <TPixel>(kernels[i]).Apply(pass, sourceRectangle, configuration);
Buffer2D <TPixel> passPixels = pass.PixelBuffer;
Buffer2D <TPixel> targetPixels = source.PixelBuffer;
ParallelFor.WithConfiguration(
minY,
maxY,
configuration,
y =>
{
int offsetY = y - shiftY;
ref TPixel passPixelsBase = ref MemoryMarshal.GetReference(passPixels.GetRowSpan(offsetY));
ref TPixel targetPixelsBase = ref MemoryMarshal.GetReference(targetPixels.GetRowSpan(offsetY));
for (int x = minX; x < maxX; x++)
{
int offsetX = x - shiftX;
// Grab the max components of the two pixels
ref TPixel currentPassPixel = ref Unsafe.Add(ref passPixelsBase, offsetX);
ref TPixel currentTargetPixel = ref Unsafe.Add(ref targetPixelsBase, offsetX);
var pixelValue = Vector4.Max(
currentPassPixel.ToVector4(),
currentTargetPixel.ToVector4());
currentTargetPixel.PackFromVector4(pixelValue);
}
});
private static void Test <T>(ref T first, int i)
{
Consume(Unsafe.Add(ref first, i));
}
internal static bool EndsWithOrdinalIgnoreCase(this ReadOnlySpan <char> span, ReadOnlySpan <char> value)
=> value.Length <= span.Length &&
Ordinal.EqualsIgnoreCase(
ref Unsafe.Add(ref MemoryMarshal.GetReference(span), span.Length - value.Length),
ref MemoryMarshal.GetReference(value),
value.Length);
public override ref byte GetPinnableMemoryAddress() => ref Unsafe.Add(ref this.Unwrap().GetPinnableMemoryAddress(), this.adjustment);
public byte GetPart(int i) => Unsafe.Add(ref Unsafe.As <BitBlock4, byte>(ref this), i);
public void SetPart(int i, byte value) => Unsafe.Add(ref Unsafe.As <BitBlock4, byte>(ref this), i) = value;
public Vector4 GetReferencesImpl(int x, int y)
{
int elementOffset = (y * this.width) + x;
return(Unsafe.Add(ref this.pinnable.Data, elementOffset));
}
public static int Compare(string?strA, int indexA, string?strB, int indexB, int length, StringComparison comparisonType)
{
CheckStringComparison(comparisonType);
if (strA == null || strB == null)
{
if (object.ReferenceEquals(strA, strB))
{
// They're both null
return(0);
}
return(strA == null ? -1 : 1);
}
if (length < 0)
{
throw new ArgumentOutOfRangeException(nameof(length), SR.ArgumentOutOfRange_NegativeLength);
}
if (indexA < 0 || indexB < 0)
{
string paramName = indexA < 0 ? nameof(indexA) : nameof(indexB);
throw new ArgumentOutOfRangeException(paramName, SR.ArgumentOutOfRange_Index);
}
if (strA.Length - indexA < 0 || strB.Length - indexB < 0)
{
string paramName = strA.Length - indexA < 0 ? nameof(indexA) : nameof(indexB);
throw new ArgumentOutOfRangeException(paramName, SR.ArgumentOutOfRange_Index);
}
if (length == 0 || (object.ReferenceEquals(strA, strB) && indexA == indexB))
{
return(0);
}
int lengthA = Math.Min(length, strA.Length - indexA);
int lengthB = Math.Min(length, strB.Length - indexB);
switch (comparisonType)
{
case StringComparison.CurrentCulture:
case StringComparison.CurrentCultureIgnoreCase:
return(CultureInfo.CurrentCulture.CompareInfo.Compare(strA, indexA, lengthA, strB, indexB, lengthB, GetCaseCompareOfComparisonCulture(comparisonType)));
case StringComparison.InvariantCulture:
case StringComparison.InvariantCultureIgnoreCase:
return(CompareInfo.Invariant.Compare(strA, indexA, lengthA, strB, indexB, lengthB, GetCaseCompareOfComparisonCulture(comparisonType)));
case StringComparison.Ordinal:
return(CompareOrdinalHelper(strA, indexA, lengthA, strB, indexB, lengthB));
default:
Debug.Assert(comparisonType == StringComparison.OrdinalIgnoreCase); // CheckStringComparison validated these earlier
return(Ordinal.CompareStringIgnoreCase(ref Unsafe.Add(ref strA.GetRawStringData(), indexA), lengthA, ref Unsafe.Add(ref strB.GetRawStringData(), indexB), lengthB));
}
}
public static unsafe RetainedVec Create <T>(RetainableMemory <T> memorySource, int start, int length, bool externallyOwned = false)
{
if (!memorySource.IsBlittableOffheap)
{
ThrowHelper.ThrowInvalidOperationException("Memory source must have IsBlittableOffheap = true to be used in RetainedVec.");
}
RetainedVec vs;
if (TypeHelper <T> .IsReferenceOrContainsReferences)
{
ThrowHelper.DebugAssert(memorySource.Pointer == default && memorySource._array != default);
// RM's offset goes to _pointerOrOffset
vs = new RetainedVec(externallyOwned ? null : memorySource,
memorySource._array,
(IntPtr)memorySource._offset,
memorySource.Length,
VecTypeHelper <T> .RuntimeTypeId);
}
else
{
ThrowHelper.DebugAssert(memorySource.Pointer != default && memorySource._array == default);
// RM's offset added to _pointerOrOffset
vs = new RetainedVec(externallyOwned ? null : memorySource, array: null, (IntPtr)Unsafe.Add <T>(memorySource.Pointer, memorySource._offset), memorySource.Length,
VecTypeHelper <T> .RuntimeTypeId);
}
return(vs.Clone(start, length, externallyOwned));
}
public static unsafe void ClearWithoutReferences(ref byte b, nuint byteLength)
{
if (byteLength == 0)
{
return;
}
#if CORECLR && (AMD64 || ARM64)
if (byteLength > 4096)
{
goto PInvoke;
}
Unsafe.InitBlockUnaligned(ref b, 0, (uint)byteLength);
return;
#else
// TODO: Optimize other platforms to be on par with AMD64 CoreCLR
// Note: It's important that this switch handles lengths at least up to 22.
// See notes below near the main loop for why.
// The switch will be very fast since it can be implemented using a jump
// table in assembly. See http://stackoverflow.com/a/449297/4077294 for more info.
switch (byteLength)
{
case 1:
b = 0;
return;
case 2:
Unsafe.As <byte, short>(ref b) = 0;
return;
case 3:
Unsafe.As <byte, short>(ref b) = 0;
Unsafe.Add <byte>(ref b, 2) = 0;
return;
case 4:
Unsafe.As <byte, int>(ref b) = 0;
return;
case 5:
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.Add <byte>(ref b, 4) = 0;
return;
case 6:
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
return;
case 7:
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.Add <byte>(ref b, 6) = 0;
return;
case 8:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
return;
case 9:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.Add <byte>(ref b, 8) = 0;
return;
case 10:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
return;
case 11:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.Add <byte>(ref b, 10) = 0;
return;
case 12:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
return;
case 13:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.Add <byte>(ref b, 12) = 0;
return;
case 14:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
return;
case 15:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
Unsafe.Add <byte>(ref b, 14) = 0;
return;
case 16:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
return;
case 17:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
Unsafe.Add <byte>(ref b, 16) = 0;
return;
case 18:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 16)) = 0;
return;
case 19:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 16)) = 0;
Unsafe.Add <byte>(ref b, 18) = 0;
return;
case 20:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 16)) = 0;
return;
case 21:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 16)) = 0;
Unsafe.Add <byte>(ref b, 20) = 0;
return;
case 22:
#if BIT64
Unsafe.As <byte, long>(ref b) = 0;
Unsafe.As <byte, long>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
#else
Unsafe.As <byte, int>(ref b) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 12)) = 0;
#endif
Unsafe.As <byte, int>(ref Unsafe.Add <byte>(ref b, 16)) = 0;
Unsafe.As <byte, short>(ref Unsafe.Add <byte>(ref b, 20)) = 0;
return;
}
// P/Invoke into the native version for large lengths
if (byteLength >= 512)
{
goto PInvoke;
}
nuint i = 0; // byte offset at which we're copying
if ((Unsafe.As <byte, int>(ref b) & 3) != 0)
{
if ((Unsafe.As <byte, int>(ref b) & 1) != 0)
{
Unsafe.AddByteOffset <byte>(ref b, i) = 0;
i += 1;
if ((Unsafe.As <byte, int>(ref b) & 2) != 0)
{
goto IntAligned;
}
}
Unsafe.As <byte, short>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
i += 2;
}
IntAligned:
// On 64-bit IntPtr.Size == 8, so we want to advance to the next 8-aligned address. If
// (int)b % 8 is 0, 5, 6, or 7, we will already have advanced by 0, 3, 2, or 1
// bytes to the next aligned address (respectively), so do nothing. On the other hand,
// if it is 1, 2, 3, or 4 we will want to copy-and-advance another 4 bytes until
// we're aligned.
// The thing 1, 2, 3, and 4 have in common that the others don't is that if you
// subtract one from them, their 3rd lsb will not be set. Hence, the below check.
if (((Unsafe.As <byte, int>(ref b) - 1) & 4) == 0)
{
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
i += 4;
}
nuint end = byteLength - 16;
byteLength -= i; // lower 4 bits of byteLength represent how many bytes are left *after* the unrolled loop
// We know due to the above switch-case that this loop will always run 1 iteration; max
// bytes we clear before checking is 23 (7 to align the pointers, 16 for 1 iteration) so
// the switch handles lengths 0-22.
Debug.Assert(end >= 7 && i <= end);
// This is separated out into a different variable, so the i + 16 addition can be
// performed at the start of the pipeline and the loop condition does not have
// a dependency on the writes.
nuint counter;
do
{
counter = i + 16;
// This loop looks very costly since there appear to be a bunch of temporary values
// being created with the adds, but the jit (for x86 anyways) will convert each of
// these to use memory addressing operands.
// So the only cost is a bit of code size, which is made up for by the fact that
// we save on writes to b.
#if BIT64
Unsafe.As <byte, long>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
Unsafe.As <byte, long>(ref Unsafe.AddByteOffset <byte>(ref b, i + 8)) = 0;
#else
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i + 4)) = 0;
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i + 8)) = 0;
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i + 12)) = 0;
#endif
i = counter;
// See notes above for why this wasn't used instead
// i += 16;
}while (counter <= end);
if ((byteLength & 8) != 0)
{
#if BIT64
Unsafe.As <byte, long>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
#else
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i + 4)) = 0;
#endif
i += 8;
}
if ((byteLength & 4) != 0)
{
Unsafe.As <byte, int>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
i += 4;
}
if ((byteLength & 2) != 0)
{
Unsafe.As <byte, short>(ref Unsafe.AddByteOffset <byte>(ref b, i)) = 0;
i += 2;
}
if ((byteLength & 1) != 0)
{
Unsafe.AddByteOffset <byte>(ref b, i) = 0;
// We're not using i after this, so not needed
// i += 1;
}
return;
#endif
PInvoke:
RuntimeImports.RhZeroMemory(ref b, byteLength);
}
/// <inheritdoc />
protected override void OnFrameApply(ImageFrame <TPixel> source)
{
DenseMatrix <float>[] kernels = this.Kernels.Flatten();
int startY = this.SourceRectangle.Y;
int endY = this.SourceRectangle.Bottom;
int startX = this.SourceRectangle.X;
int endX = this.SourceRectangle.Right;
// Align start/end positions.
int minX = Math.Max(0, startX);
int maxX = Math.Min(source.Width, endX);
int minY = Math.Max(0, startY);
int maxY = Math.Min(source.Height, endY);
// we need a clean copy for each pass to start from
using (ImageFrame <TPixel> cleanCopy = source.Clone())
{
using (var processor = new ConvolutionProcessor <TPixel>(this.Configuration, kernels[0], true, this.Source, this.SourceRectangle))
{
processor.Apply(source);
}
if (kernels.Length == 1)
{
return;
}
int shiftY = startY;
int shiftX = startX;
// Reset offset if necessary.
if (minX > 0)
{
shiftX = 0;
}
if (minY > 0)
{
shiftY = 0;
}
var workingRect = Rectangle.FromLTRB(minX, minY, maxX, maxY);
// Additional runs.
// ReSharper disable once ForCanBeConvertedToForeach
for (int i = 1; i < kernels.Length; i++)
{
using (ImageFrame <TPixel> pass = cleanCopy.Clone())
{
using (var processor = new ConvolutionProcessor <TPixel>(this.Configuration, kernels[i], true, this.Source, this.SourceRectangle))
{
processor.Apply(pass);
}
Buffer2D <TPixel> passPixels = pass.PixelBuffer;
Buffer2D <TPixel> targetPixels = source.PixelBuffer;
ParallelHelper.IterateRows(
workingRect,
this.Configuration,
rows =>
{
for (int y = rows.Min; y < rows.Max; y++)
{
int offsetY = y - shiftY;
ref TPixel passPixelsBase = ref MemoryMarshal.GetReference(passPixels.GetRowSpan(offsetY));
ref TPixel targetPixelsBase = ref MemoryMarshal.GetReference(targetPixels.GetRowSpan(offsetY));
for (int x = minX; x < maxX; x++)
{
int offsetX = x - shiftX;
// Grab the max components of the two pixels
ref TPixel currentPassPixel = ref Unsafe.Add(ref passPixelsBase, offsetX);
ref TPixel currentTargetPixel = ref Unsafe.Add(ref targetPixelsBase, offsetX);
var pixelValue = Vector4.Max(
currentPassPixel.ToVector4(),
currentTargetPixel.ToVector4());
currentTargetPixel.FromVector4(pixelValue);
}
}
});
public PointerBasedEnumerator(void *origin, int offset, int length)
{
_ptr = Unsafe.Add <T>(origin, offset);
_remaining = length;
}
private static void CopyContactData(int contactCount, ref NonconvexConstraintContactData sourceContacts, ref NonconvexPrestepData targetContacts)
{
for (int i = 0; i < contactCount; ++i)
{
ref var sourceContact = ref Unsafe.Add(ref sourceContacts, i);
ref var targetContact = ref Unsafe.Add(ref targetContacts, i);
/// <inheritdoc/>
protected override void OnFrameApply(
ImageFrame <TPixel> source,
ImageFrame <TPixel> destination,
Rectangle sourceRectangle,
Configuration configuration)
{
int height = this.TargetDimensions.Height;
int width = this.TargetDimensions.Width;
Rectangle sourceBounds = source.Bounds();
var targetBounds = new Rectangle(0, 0, width, height);
// Since could potentially be resizing the canvas we might need to re-calculate the matrix
Matrix3x2 matrix = this.GetProcessingMatrix(sourceBounds, targetBounds);
// Convert from screen to world space.
Matrix3x2.Invert(matrix, out matrix);
if (this.Sampler is NearestNeighborResampler)
{
ParallelHelper.IterateRows(
targetBounds,
configuration,
rows =>
{
for (int y = rows.Min; y < rows.Max; y++)
{
Span <TPixel> destRow = destination.GetPixelRowSpan(y);
for (int x = 0; x < width; x++)
{
var point = Point.Transform(new Point(x, y), matrix);
if (sourceBounds.Contains(point.X, point.Y))
{
destRow[x] = source[point.X, point.Y];
}
}
}
});
return;
}
int maxSourceX = source.Width - 1;
int maxSourceY = source.Height - 1;
(float radius, float scale, float ratio)xRadiusScale = this.GetSamplingRadius(source.Width, destination.Width);
(float radius, float scale, float ratio)yRadiusScale = this.GetSamplingRadius(source.Height, destination.Height);
float xScale = xRadiusScale.scale;
float yScale = yRadiusScale.scale;
var radius = new Vector2(xRadiusScale.radius, yRadiusScale.radius);
IResampler sampler = this.Sampler;
var maxSource = new Vector4(maxSourceX, maxSourceY, maxSourceX, maxSourceY);
int xLength = (int)MathF.Ceiling((radius.X * 2) + 2);
int yLength = (int)MathF.Ceiling((radius.Y * 2) + 2);
MemoryAllocator memoryAllocator = configuration.MemoryAllocator;
using (Buffer2D <float> yBuffer = memoryAllocator.Allocate2D <float>(yLength, height))
using (Buffer2D <float> xBuffer = memoryAllocator.Allocate2D <float>(xLength, height))
{
ParallelHelper.IterateRows(
targetBounds,
configuration,
rows =>
{
for (int y = rows.Min; y < rows.Max; y++)
{
ref TPixel destRowRef = ref MemoryMarshal.GetReference(destination.GetPixelRowSpan(y));
ref float ySpanRef = ref MemoryMarshal.GetReference(yBuffer.GetRowSpan(y));
ref float xSpanRef = ref MemoryMarshal.GetReference(xBuffer.GetRowSpan(y));
for (int x = 0; x < width; x++)
{
// Use the single precision position to calculate correct bounding pixels
// otherwise we get rogue pixels outside of the bounds.
var point = Vector2.Transform(new Vector2(x, y), matrix);
// Clamp sampling pixel radial extents to the source image edges
Vector2 maxXY = point + radius;
Vector2 minXY = point - radius;
// max, maxY, minX, minY
var extents = new Vector4(
MathF.Floor(maxXY.X + .5F),
MathF.Floor(maxXY.Y + .5F),
MathF.Ceiling(minXY.X - .5F),
MathF.Ceiling(minXY.Y - .5F));
int right = (int)extents.X;
int bottom = (int)extents.Y;
int left = (int)extents.Z;
int top = (int)extents.W;
extents = Vector4.Clamp(extents, Vector4.Zero, maxSource);
int maxX = (int)extents.X;
int maxY = (int)extents.Y;
int minX = (int)extents.Z;
int minY = (int)extents.W;
if (minX == maxX || minY == maxY)
{
continue;
}
// It appears these have to be calculated on-the-fly.
// Precalculating transformed weights would require prior knowledge of every transformed pixel location
// since they can be at sub-pixel positions on both axis.
// I've optimized where I can but am always open to suggestions.
if (yScale > 1 && xScale > 1)
{
CalculateWeightsDown(
top,
bottom,
minY,
maxY,
point.Y,
sampler,
yScale,
ref ySpanRef,
yLength);
CalculateWeightsDown(
left,
right,
minX,
maxX,
point.X,
sampler,
xScale,
ref xSpanRef,
xLength);
}
else
{
CalculateWeightsScaleUp(minY, maxY, point.Y, sampler, ref ySpanRef);
CalculateWeightsScaleUp(minX, maxX, point.X, sampler, ref xSpanRef);
}
// Now multiply the results against the offsets
Vector4 sum = Vector4.Zero;
for (int yy = 0, j = minY; j <= maxY; j++, yy++)
{
float yWeight = Unsafe.Add(ref ySpanRef, yy);
for (int xx = 0, i = minX; i <= maxX; i++, xx++)
{
float xWeight = Unsafe.Add(ref xSpanRef, xx);
// Values are first premultiplied to prevent darkening of edge pixels
var current = source[i, j].ToVector4();
Vector4Utils.Premultiply(ref current);
sum += current * xWeight * yWeight;
}
}
ref TPixel dest = ref Unsafe.Add(ref destRowRef, x);
// Reverse the premultiplication
Vector4Utils.UnPremultiply(ref sum);
dest.FromVector4(sum);
}
public static void StoreValueTypeFieldValueIntoValueType(TypedReference typedReference, int fieldOffset, object fieldValue, RuntimeTypeHandle fieldTypeHandle)
{
Debug.Assert(TypedReference.TargetTypeToken(typedReference).ToEETypePtr().IsValueType);
RuntimeImports.RhUnbox(fieldValue, ref Unsafe.Add <byte>(ref typedReference.Value, fieldOffset), fieldTypeHandle.ToEETypePtr());
}
/// <inheritdoc/>
protected override void OnFrameApply(ImageFrame <TPixel> source)
{
int sourceWidth = source.Width;
int sourceHeight = source.Height;
int tileWidth = (int)MathF.Ceiling(sourceWidth / (float)this.Tiles);
int tileHeight = (int)MathF.Ceiling(sourceHeight / (float)this.Tiles);
int tileCount = this.Tiles;
int halfTileWidth = tileWidth / 2;
int halfTileHeight = tileHeight / 2;
int luminanceLevels = this.LuminanceLevels;
// The image is split up into tiles. For each tile the cumulative distribution function will be calculated.
using (var cdfData = new CdfTileData(this.Configuration, sourceWidth, sourceHeight, this.Tiles, this.Tiles, tileWidth, tileHeight, luminanceLevels))
{
cdfData.CalculateLookupTables(source, this);
var tileYStartPositions = new List <(int y, int cdfY)>();
int cdfY = 0;
int yStart = halfTileHeight;
for (int tile = 0; tile < tileCount - 1; tile++)
{
tileYStartPositions.Add((yStart, cdfY));
cdfY++;
yStart += tileHeight;
}
Parallel.For(
0,
tileYStartPositions.Count,
new ParallelOptions {
MaxDegreeOfParallelism = this.Configuration.MaxDegreeOfParallelism
},
index =>
{
int y = tileYStartPositions[index].y;
int cdfYY = tileYStartPositions[index].cdfY;
// It's unfortunate that we have to do this per iteration.
ref TPixel sourceBase = ref source.GetPixelReference(0, 0);
int cdfX = 0;
int x = halfTileWidth;
for (int tile = 0; tile < tileCount - 1; tile++)
{
int tileY = 0;
int yEnd = Math.Min(y + tileHeight, sourceHeight);
int xEnd = Math.Min(x + tileWidth, sourceWidth);
for (int dy = y; dy < yEnd; dy++)
{
int dyOffSet = dy * sourceWidth;
int tileX = 0;
for (int dx = x; dx < xEnd; dx++)
{
ref TPixel pixel = ref Unsafe.Add(ref sourceBase, dyOffSet + dx);
float luminanceEqualized = InterpolateBetweenFourTiles(
pixel,
cdfData,
tileCount,
tileCount,
tileX,
tileY,
cdfX,
cdfYY,
tileWidth,
tileHeight,
luminanceLevels);
pixel.FromVector4(new Vector4(luminanceEqualized, luminanceEqualized, luminanceEqualized, pixel.ToVector4().W));
tileX++;
}
tileY++;
}
cdfX++;
x += tileWidth;
}
});
public static void StoreReferenceTypeFieldValueIntoValueType(TypedReference typedReference, int fieldOffset, object fieldValue)
{
Debug.Assert(TypedReference.TargetTypeToken(typedReference).ToEETypePtr().IsValueType);
Unsafe.As <byte, object>(ref Unsafe.Add <byte>(ref typedReference.Value, fieldOffset)) = fieldValue;
}
/// <summary>
/// Finds the assignments with the lowest global assignment cost.
/// See https://en.wikipedia.org/wiki/Hungarian_algorithm for an explanation of the algorithm used.
/// </summary>
/// <param name="assignmentCostsDoubles">The assignment costs.</param>
/// <returns>System.Int32[].</returns>
/// <exception cref="ArgumentNullException">assignmentCosts</exception>
/// <exception cref="ArgumentException">This algorithm implementation does not support cost matrices with fewer columns than rows - assignmentCosts</exception>
public static int[] FindAssignments([NotNull] Matrix <double> assignmentCostsDoubles)
{
if (assignmentCostsDoubles == null)
{
throw new ArgumentNullException(nameof(assignmentCostsDoubles));
}
var rows = assignmentCostsDoubles.RowCount;
var columns = assignmentCostsDoubles.ColumnCount;
if (rows > columns)
{
throw new ArgumentException("This algorithm implementation does not support cost matrices with fewer columns than rows", nameof(assignmentCostsDoubles));
}
var costs = new Storage <float>(rows, columns);
for (var row = 0; row < rows; row++)
{
var min = float.MaxValue;
for (var column = 0; column < columns; column++)
{
var cost = (float)assignmentCostsDoubles[row, column];
if (float.IsNegativeInfinity(cost))
{
costs[row, column] = cost = float.MinValue;
}
else if (float.IsPositiveInfinity(cost) || float.IsNaN(cost))
{
costs[row, column] = cost = float.MaxValue;
}
else
{
costs[row, column] = cost;
}
min = Math.Min(min, cost);
}
if (float.IsInfinity(min))
{
min = float.MinValue;
}
for (var column = 0; column < columns; column++)
{
costs[row, column] -= min;
}
}
var masks = new Storage <byte>(rows, columns);
var rowsCovered = new bool[rows];
var colsCovered = new bool[columns];
for (var row = 0; row < rows; row++)
{
for (var column = 0; column < columns; column++)
{
if (Math.Abs(costs[row, column]) <= 0 && !rowsCovered[row] && !colsCovered[column])
{
masks[row, column] = 1;
rowsCovered[row] = colsCovered[column] = true;
}
}
}
ClearCoveredFlags();
var path = new Location[columns * rows];
var pathStart = default(Location);
var step = 1;
while (step != -1)
{
switch (step)
{
case 1:
step = RunStep1();
break;
case 2:
step = RunStep2();
break;
case 3:
step = RunStep3();
break;
case 4:
step = RunStep4();
break;
default:
throw new Exception($"Unknown step number {step}");
}
}
var agentsTasks = new int[rows];
for (var row = 0; row < rows; row++)
{
for (var column = 0; column < columns; column++)
{
if (masks[row, column] == 1)
{
agentsTasks[row] = column;
break;
}
}
}
return(agentsTasks);
int RunStep1()
{
// The covered flags have are reset before this step
var coveredColsCount = 0;
if (Avx2.IsSupported && rows >= Vector256 <byte> .Count)
{
var maxVectorOffset = rows - rows % Vector256 <byte> .Count;
if (maxVectorOffset > 0)
{
var onesVector = Vector256.Create((byte)1);
for (var column = 0; column < columns; column++)
{
var currentColCovered = false;
ref var rowRef = ref masks.ColumnMajorBackingStore[column * rows];
for (var row = 0; row < maxVectorOffset; row += Vector256 <byte> .Count)
{
var masksRowVector = Unsafe.ReadUnaligned <Vector256 <byte> >(ref Unsafe.Add(ref rowRef, row));
var comparison = Avx2.CompareEqual(masksRowVector, onesVector);
var comparisonMask = Avx2.MoveMask(comparison);
if (comparisonMask != 0)
{
colsCovered[column] = true;
currentColCovered = true;
coveredColsCount++;
break;
}
}
if (!currentColCovered && maxVectorOffset < rows)
{
for (var row = maxVectorOffset; row < rows; row++)
{
if (Unsafe.Add(ref rowRef, row) == 1)
{
colsCovered[column] = true;
coveredColsCount++;
break;
}
}
}
}
}
}
public static object LoadReferenceTypeFieldValueFromValueType(TypedReference typedReference, int fieldOffset)
{
Debug.Assert(TypedReference.TargetTypeToken(typedReference).ToEETypePtr().IsValueType);
return(Unsafe.As <byte, object>(ref Unsafe.Add <byte>(ref typedReference.Value, fieldOffset)));
}
