// Copyright Epic Games, Inc. All Rights Reserved.

#include "TextureCompressorModule.h"
#include "Math/RandomStream.h"
#include "ChildTextureFormat.h"
#include "Containers/IndirectArray.h"
#include "ImageCoreUtils.h"
#include "Stats/Stats.h"
#include "Async/AsyncWork.h"
#include "Async/ParallelFor.h"
#include "Modules/ModuleManager.h"
#include "Engine/TextureDefines.h"
#include "TextureFormatManager.h"
#include "TextureBuildUtilities.h"
#include "Interfaces/ITextureFormat.h"
#include "Misc/Paths.h"
#include "Tasks/Task.h"
#include "ImageCore.h"
#include <cmath>

#if PLATFORM_WINDOWS
#include "Windows/WindowsHWrapper.h"
#endif

DEFINE_LOG_CATEGORY_STATIC(LogTextureCompressor, Log, All);

/*------------------------------------------------------------------------------
	Mip-Map Generation
------------------------------------------------------------------------------*/

// NOTE: mip gen wrap/clamp does NOT correspond to Texture Address Wrap/Clamp setting !!
//		it comes from bPreserveBorder !
enum EMipGenAddressMode
{
	MGTAM_Wrap,
	MGTAM_Clamp,
	MGTAM_BorderBlack,
};

/**
 * 2D view into one slice of an image.
 */
struct FImageView2D
{
	/** Pointer to colors in the slice. */
	FLinearColor* SliceColors;
	/** Width of the slice. */
	int64 SizeX;
	/** Height of the slice. */
	int64 SizeY;

	FImageView2D() : SliceColors(nullptr), SizeX(0), SizeY(0) {}

	/** Initialization constructor. */
	FImageView2D(FImage& Image, int32 SliceIndex)
	{
		SizeX = Image.SizeX;
		SizeY = Image.SizeY;
		check( SliceIndex < Image.NumSlices );
		SliceColors = (&Image.AsRGBA32F()[0]) + SliceIndex * SizeY * SizeX;
	}

	/** Access a single texel. */
	FLinearColor& Access(int32 X, int32 Y)
	{
		return SliceColors[X + Y * SizeX];
	}

	/** Const access to a single texel. */
	const FLinearColor& Access(int32 X, int32 Y) const
	{
		return SliceColors[X + Y * SizeX];
	}

	bool IsValid() const { return SliceColors != nullptr; }

	static const FImageView2D ConstructConst(const FImage& Image, int32 SliceIndex)
	{
		return FImageView2D(const_cast<FImage&>(Image), SliceIndex);
	}

};

// 2D sample lookup with input conversion
// requires SourceImageData.SizeX and SourceImageData.SizeY to be power of two
template <EMipGenAddressMode AddressMode>
static const FLinearColor& LookupSourceMip(const FImageView2D& SourceImageData, int32 X, int32 Y)
{
	if(AddressMode == MGTAM_Wrap)
	{
		// wrap
		// ! requires pow2 sizes
		checkSlow( FMath::IsPowerOfTwo(SourceImageData.SizeX) );
		checkSlow( FMath::IsPowerOfTwo(SourceImageData.SizeY) );

		X = (int32)((uint32)X) & (SourceImageData.SizeX - 1);
		Y = (int32)((uint32)Y) & (SourceImageData.SizeY - 1);
	}
	else if(AddressMode == MGTAM_Clamp)
	{
		// clamp
		X = FMath::Clamp(X, 0, SourceImageData.SizeX - 1);
		Y = FMath::Clamp(Y, 0, SourceImageData.SizeY - 1);
	}
	else if(AddressMode == MGTAM_BorderBlack)
	{
		// border color 0
		if((uint32)X >= (uint32)SourceImageData.SizeX
			|| (uint32)Y >= (uint32)SourceImageData.SizeY)
		{
			static FLinearColor Black(0, 0, 0, 0);
			return Black;
		}
	}
	else
	{
		check(0);
	}
	return SourceImageData.Access(X, Y);
}

// Same functionality as above, but for 1D lookup with explicit size
template <EMipGenAddressMode AddressMode>
static const FLinearColor& LookupSourceMip(const FLinearColor* Data, int32 Size, int32 X)
{
	if (AddressMode == MGTAM_Wrap)
	{
		// wrap
		// ! requires pow2 sizes
		checkSlow( FMath::IsPowerOfTwo(Size) );

		X = (int32)((uint32)X) & (Size - 1);
	}
	else if (AddressMode == MGTAM_Clamp)
	{
		// clamp
		X = FMath::Clamp(X, 0, Size - 1);
	}
	else if (AddressMode == MGTAM_BorderBlack)
	{
		// border color 0
		if ((uint32)X >= (uint32)Size)
		{
			static FLinearColor Black(0, 0, 0, 0);
			return Black;
		}
	}
	else
	{
		check(0);
	}
	return Data[X];
}

// Kernel class for image filtering operations like image downsampling
// at max MaxKernelExtend x MaxKernelExtend
class FImageKernel2D
{
public:
	FImageKernel2D() :FilterTableSize(0)
	{
	}

	// @param TableSize1D 2 for 2x2, 4 for 4x4, 6 for 6x6, 8 for 8x8
	// @param SharpenFactor can be negative to blur
	// generate normalized 2D Kernel with sharpening
	void BuildSeparatableGaussWithSharpen(uint32 TableSize1D, float SharpenFactor = 0.0f)
	{
		if(TableSize1D > MaxKernelExtend)
		{
			TableSize1D = MaxKernelExtend;
		}

		float* Table1D = KernelWeights1D;
		float NegativeTable1D[MaxKernelExtend];

		FilterTableSize = TableSize1D;
		
		if(TableSize1D == 2)
		{
			// 2x2 kernel: simple average
			// SharpenFactor is ignored
			// this is TMGS_SimpleAverage
			Table1D[0] = Table1D[1] = 0.5f;
			KernelWeights[0] = KernelWeights[1] = KernelWeights[2] = KernelWeights[3] = 0.25f;
			return;
		}
		else if(SharpenFactor < 0.0f)
		{
			// blur only
			// this is TMGS_Blur

			// TMGS_Blur will always give us TableSize > 2
			check( TableSize1D > 2 );

			BuildGaussian1D(Table1D, TableSize1D, 1.0f, -SharpenFactor);
			BuildFilterTable2DFrom1D(KernelWeights, Table1D, TableSize1D);
			return;
		}
		else if(TableSize1D == 4)
		{
			// 4x4 kernel with sharpen or blur: can alias a bit
			// this is not used by standard TMGS_ mip options
			//  one thing that can get you in here is GenerateTopMip
			//	because it takes the standard 8 size and does /2
			BuildFilterTable1DBase(Table1D, TableSize1D, 1.0f + SharpenFactor);
			BuildFilterTable1DBase(NegativeTable1D, TableSize1D, -SharpenFactor);
			BlurFilterTable1D(NegativeTable1D, TableSize1D, 1);
		}
		else if(TableSize1D == 6)
		{
			// 6x6 kernel with sharpen or blur: still can alias
			// this is not used by standard TMGS_ mip options
			BuildFilterTable1DBase(Table1D, TableSize1D, 1.0f + SharpenFactor);
			BuildFilterTable1DBase(NegativeTable1D, TableSize1D, -SharpenFactor);
			BlurFilterTable1D(NegativeTable1D, TableSize1D, 2);
		}
		else if(TableSize1D == 8)
		{
			//8x8 kernel with sharpen
			// these are the TMGS_Sharpen filters

			// * 2 to get similar appearance as for TableSize 6
			SharpenFactor = SharpenFactor * 2.0f;

			BuildFilterTable1DBase(Table1D, TableSize1D, 1.0f + SharpenFactor);
			// positive lobe is blurred a bit for better quality
			BlurFilterTable1D(Table1D, TableSize1D, 1);
			BuildFilterTable1DBase(NegativeTable1D, TableSize1D, -SharpenFactor);
			BlurFilterTable1D(NegativeTable1D, TableSize1D, 3);
		}
		else 
		{
			// not yet supported
			check(0);
		}

		AddFilterTable1D(Table1D, NegativeTable1D, TableSize1D);
		BuildFilterTable2DFrom1D(KernelWeights, Table1D, TableSize1D);
	}

	inline uint32 GetFilterTableSize() const
	{
		return FilterTableSize;
	}

	inline float Get1D(uint32 X) const
	{
		checkSlow(X < FilterTableSize);
		return KernelWeights1D[X];
	}

	inline float GetAt(uint32 X, uint32 Y) const
	{
		checkSlow(X < FilterTableSize);
		checkSlow(Y < FilterTableSize);
		return KernelWeights[X + Y * FilterTableSize];
	}

private:

	inline static float NormalDistribution(float X, float Variance)
	{
		const float StandardDeviation = FMath::Sqrt(Variance);
		return FMath::Exp(-FMath::Square(X) / (2.0f * Variance)) / (StandardDeviation * FMath::Sqrt(2.0f * (float)PI));
	}

	// support even and non even sized filters
	static void BuildGaussian1D(float *InOutTable, uint32 TableSize, float Sum, float Variance)
	{
		float Center = TableSize * 0.5f - 0.5f;
		float CurrentSum = 0;
		for(uint32 i = 0; i < TableSize; ++i)
		{
			float Actual = NormalDistribution(i - Center, Variance);
			InOutTable[i] = Actual;
			CurrentSum += Actual;
		}
		// Normalize
		float InvSum = Sum / CurrentSum;
		for(uint32 i = 0; i < TableSize; ++i)
		{
			InOutTable[i] *= InvSum;
		}
	}

	//
	static void BuildFilterTable1DBase(float *InOutTable, uint32 TableSize, float Sum )
	{
		// we require a even sized filter
		check(TableSize % 2 == 0);

		float Inner = 0.5f * Sum;

		uint32 Center = TableSize / 2;
		for(uint32 x = 0; x < TableSize; ++x)
		{
			if(x == Center || x == Center - 1)
			{
				// center elements
				InOutTable[x] = Inner;
			}
			else
			{
				// outer elements
				InOutTable[x] = 0.0f;
			}
		}
	}

	// InOutTable += InTable
	static void AddFilterTable1D( float *InOutTable, const float *InTable, uint32 TableSize )
	{
		for(uint32 x = 0; x < TableSize; ++x)
		{
			InOutTable[x] += InTable[x];
		}
	}

	// @param Times 1:box, 2:triangle, 3:pow2, 4:pow3, ...
	// can be optimized with double buffering but doesn't need to be fast
	static void BlurFilterTable1D( float *InOutTable, uint32 TableSize, uint32 Times )
	{
		check(Times>0);
		check(TableSize<32);

		float Intermediate[32];

		for(uint32 Pass = 0; Pass < Times; ++Pass)
		{
			for(uint32 x = 0; x < TableSize; ++x)
			{
				Intermediate[x] = InOutTable[x];
			}

			for(uint32 x = 0; x < TableSize; ++x)
			{
				float sum = Intermediate[x];

				if(x)
				{
					sum += Intermediate[x-1];	
				}
				if(x < TableSize - 1)
				{
					sum += Intermediate[x+1];	
				}

				InOutTable[x] = sum / 3.0f;
			}
		}
	}

	static void BuildFilterTable2DFrom1D( float *OutTable2D, float *InTable1D, uint32 TableSize )
	{
		for(uint32 y = 0; y < TableSize; ++y)
		{
			for(uint32 x = 0; x < TableSize; ++x)
			{
				OutTable2D[x + y * TableSize] = InTable1D[y] * InTable1D[x];
			}
		}
	}

	// at max we support MaxKernelExtend x MaxKernelExtend kernels
	const static uint32 MaxKernelExtend = 12;
	// 0 if no kernel was setup yet
	uint32 FilterTableSize;
	// normalized, means the sum of it should be 1.0f
	float KernelWeights[MaxKernelExtend * MaxKernelExtend];
	float KernelWeights1D[MaxKernelExtend];
};

static float DetermineScaledThreshold(float Threshold, float Scale)
{
	check(Threshold > 0.f && Scale > 0.f);

	// Assuming Scale > 0 and Threshold > 0, find ScaledThreshold such that
	//	 x * Scale >= Threshold
	// is exactly equivalent to
	//	 x >= ScaledThreshold.
	//
	// This is for a test that was originally written in the first form that we want to
	// transform to the second form without changing results (which would in turn change
	// texture cooks).
	//
	// In exact arithmetic, this is just ScaledThreshold = Threshold / Scale.
	//
	// In floating point, we need to consider rounding. Computed in floating point
	// and assuming round-to-nearest (breaking ties towards even), we get 
	//
	//	 RN(x * Scale) >= Threshold
	//
	// The smallest conceivable x that passes RN(x * Scale) >= Threshold is
	// x = (Threshold - 0.5u) / Scale, landing exactly halfway with the rounding
	// going up; this is slightly less than an exact Threshold/Scale.
	//
	// For regular floating point division, we get
	//	 RN(Threshold / Scale)
	// = (Threshold / Scale) * (1 + e),  |e| < 0.5u (the inequality is strict for divisions)
	//
	// That gets us relatively close to the target value, but we have no guarantee that rounding
	// on the division was in the direction we wanted. Just check whether our target inequality
	// is satisfied and bump up or down to the next representable float as required.
	float ScaledThreshold = Threshold / Scale;
	float SteppedDown = std::nextafter(ScaledThreshold, 0.f);

	// We want ScaledThreshold to be the smallest float such that
	//	 ScaledThreshold * Scale >= Threshold
	// meaning the next-smaller float below ScaledThreshold (which is SteppedDown)
	// should not be >=Threshold. 

	if (SteppedDown * Scale >= Threshold)
	{
		// We were too large, go down by 1 ulp
		ScaledThreshold = SteppedDown;
	}
	else if (ScaledThreshold * Scale < Threshold)
	{
		// We were too small, go up by 1 ulp
		ScaledThreshold = std::nextafter(ScaledThreshold, 2.f * ScaledThreshold);
	}

	// We should now have the right threshold:
	check(ScaledThreshold * Scale >= Threshold); // ScaledThreshold is large enough
	check(std::nextafter(ScaledThreshold, 0.f) * Scale < Threshold); // next below is too small

	return ScaledThreshold;
}


static FVector4f ComputeAlphaCoverage(const FVector4f Thresholds, const FVector4f Scales, const FImageView2D& SourceImageData)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.ComputeAlphaCoverage);

	FVector4f Coverage(0, 0, 0, 0);

	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob,SourceImageData.SizeX,SourceImageData.SizeY);

	if ( Thresholds[0] == 0.f && Thresholds[1] == 0.f && Thresholds[2] == 0.f )
	{
		// common case that only channel 3 (A) is used for alpha coverage :
		
		check( Thresholds[3] != 0.f );

		const float ThresholdScaled = DetermineScaledThreshold(Thresholds[3] , Scales[3]);
		
		int64 CommonResult = 0;
		ParallelFor( TEXT("Texture.ComputeAlphaCoverage.PF"),NumJobs,1, [&](int32 Index)
		{
			int32 StartIndex = Index * NumRowsEachJob;
			int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, SourceImageData.SizeY);
			int32 LocalCoverage = 0;
			for (int32 y = StartIndex; y < EndIndex; ++y)
			{
				const FLinearColor * RowPixels = &SourceImageData.Access(0,y);

				for (int32 x = 0; x < SourceImageData.SizeX; ++x)
				{
					LocalCoverage += (RowPixels[x].A >= ThresholdScaled);
				}
			}

			FPlatformAtomics::InterlockedAdd(&CommonResult, LocalCoverage);
		}, EParallelForFlags::Unbalanced);

		Coverage[3] = float(CommonResult) / float( (int64) SourceImageData.SizeX * SourceImageData.SizeY);
		
		UE_LOG(LogTextureCompressor, VeryVerbose, TEXT("Thresholds = 000 %f Coverage = 000 %f"),  \
			Thresholds[3], Coverage[3] );
	}
	else
	{
		FVector4f ThresholdsScaled;

		for (int32 i = 0; i < 4; ++i)
		{
			// Skip channel if Threshold is 0
			if (Thresholds[i] == 0)
			{
				// stuff a value that we will always be less than
				ThresholdsScaled[i] = FLT_MAX;
			}
			else
			{
				check( Scales[i] != 0.f );
				ThresholdsScaled[i] = DetermineScaledThreshold( Thresholds[i] , Scales[i] );
			}
		}

		int64 CommonResults[4] = { 0, 0, 0, 0 };
		ParallelFor( TEXT("Texture.ComputeAlphaCoverage.PF"),NumJobs,1, [&](int32 Index)
		{
			int32 StartIndex = Index * NumRowsEachJob;
			int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, SourceImageData.SizeY);
			int32 LocalCoverage[4] = { 0, 0, 0, 0 };
			for (int32 y = StartIndex; y < EndIndex; ++y)
			{
				const FLinearColor * RowPixels = &SourceImageData.Access(0,y);

				for (int32 x = 0; x < SourceImageData.SizeX; ++x)
				{
					const FLinearColor & PixelValue = RowPixels[x];

					// Calculate coverage for each channel
					for (int32 i = 0; i < 4; ++i)
					{
						LocalCoverage[i] += ( PixelValue.Component(i) >= ThresholdsScaled[i] );
					}
				}
			}

			for (int32 i = 0; i < 4; ++i)
			{
				FPlatformAtomics::InterlockedAdd(&CommonResults[i], LocalCoverage[i]);
			}
		}, EParallelForFlags::Unbalanced);

		for (int32 i = 0; i < 4; ++i)
		{
			Coverage[i] = float(CommonResults[i]) / float((int64) SourceImageData.SizeX * SourceImageData.SizeY);
		}
		
		UE_LOG(LogTextureCompressor, VeryVerbose, TEXT("Thresholds = %f %f %f %f Coverage = %f %f %f %f"),  \
			Thresholds[0], Thresholds[1], Thresholds[2], Thresholds[3], \
			Coverage[0], Coverage[1], Coverage[2], Coverage[3] );
	}
	
	return Coverage;
}

static FVector4f ComputeAlphaScale(const FVector4f Coverages, const FVector4f AlphaThresholds, const FImageView2D& SourceImageData)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.ComputeAlphaScale);

	// This function is not a good way to do this
	// but we cannot change it without changing output pixels
	// A better method would be to histogram the channel and scale the histogram to meet the desired threshold
	// even if using this binary search method, you should remember which value gave the closest result
	//	 don't assume that each binary search step is an improvement
	// 

	FVector4f MinAlphaScales (0, 0, 0, 0);
	FVector4f MaxAlphaScales (4, 4, 4, 4);
	FVector4f AlphaScales (1, 1, 1, 1);

	//Binary Search to find Alpha Scale
	// limit binary search to 8 steps
	for (int32 i = 0; i < 8; ++i)
	{
		FVector4f ComputedCoverages = ComputeAlphaCoverage(AlphaThresholds, AlphaScales, SourceImageData);
		
		UE_LOG(LogTextureCompressor, VeryVerbose, TEXT("Tried AlphaScale = %f ComputedCoverage = %f Goal = %f"), AlphaScales[3], ComputedCoverages[3], Coverages[3] );

		for (int32 j = 0; j < 4; ++j)
		{
			if (AlphaThresholds[j] == 0 || fabsf(ComputedCoverages[j] - Coverages[j]) < KINDA_SMALL_NUMBER)
			{
				continue;
			}

			if (ComputedCoverages[j] < Coverages[j])
			{
				MinAlphaScales[j] = AlphaScales[j];
			}
			else if (ComputedCoverages[j] > Coverages[j])
			{
				MaxAlphaScales[j] = AlphaScales[j];
			}

			// guess alphascale is best at next midpoint :
			//  this means we wind up returning an alphascale value we have never tested
			AlphaScales[j] = (MinAlphaScales[j] + MaxAlphaScales[j]) * 0.5f;
		}

		// Equals default tolerance is KINDA_SMALL_NUMBER so it checks the same condition as above
		if (ComputedCoverages.Equals(Coverages))
		{
			break;
		}
	}

	UE_LOG(LogTextureCompressor, VeryVerbose, TEXT("Final AlphaScales = %f %f %f %f"), AlphaScales[0], AlphaScales[1], AlphaScales[2], AlphaScales[3] );

	return AlphaScales;
}


static void GenerateMip2x2Simple(
	const FImageView2D& SourceImageData, 
	FImageView2D& DestImageData)
{
	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob,DestImageData.SizeX,DestImageData.SizeY);

	ParallelFor( TEXT("Texture.GenerateMip2x2Simple.PF"),NumJobs,1, [&](int32 Index)
	{
		int32 StartIndex = Index * NumRowsEachJob;
		int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, DestImageData.SizeY);
		for (int32 DestY = StartIndex; DestY < EndIndex; ++DestY)
		{
			float* DestRow = &DestImageData.Access(0, DestY).Component(0);
			const float* SourceRow0 = &SourceImageData.Access(0, 2*DestY).Component(0);
			const float* SourceRow1 = &SourceImageData.Access(0, 2*DestY+1).Component(0);

			const VectorRegister4Float Mul = VectorSetFloat1(0.25f);
			for (int32 DestX = 0; DestX < DestImageData.SizeX; DestX++)
			{
				VectorRegister4Float A = VectorLoad(&SourceRow0[0]);
				VectorRegister4Float B = VectorLoad(&SourceRow0[4]);
				VectorRegister4Float C = VectorLoad(&SourceRow1[0]);
				VectorRegister4Float D = VectorLoad(&SourceRow1[4]);
				VectorRegister4Float Sum = VectorAdd(VectorAdd(VectorAdd(A, B), C), D);
				VectorRegister4Float Avg = VectorMultiply(Sum, Mul);
				VectorStore(Avg, &DestRow[0]);
				SourceRow0 += 8;
				SourceRow1 += 8;
				DestRow += 4;
			}
		}
	}, EParallelForFlags::Unbalanced);
}

template <EMipGenAddressMode AddressMode>
static void GenerateMipUnfiltered(const FImageView2D& SourceImageData, FImageView2D& DestImageData, FVector4f AlphaScale, uint32 ScaleFactor)
{
	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob, DestImageData.SizeX, DestImageData.SizeY);

	ParallelFor(TEXT("Texture.GenerateMipUnfiltered.PF"), NumJobs, 1, [&](int32 Index)
	{
		VectorRegister4Float AlphaScaleV = VectorLoad(&AlphaScale[0]);
		int32 StartIndex = Index * NumRowsEachJob;
		int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, DestImageData.SizeY);
		for (int32 DestY = StartIndex; DestY < EndIndex; ++DestY)
		{
			for (int32 DestX = 0; DestX < DestImageData.SizeX; DestX++)
			{
				const int32 SourceX = DestX * ScaleFactor;
				const int32 SourceY = DestY * ScaleFactor;
				const FLinearColor& Sample = LookupSourceMip<AddressMode>(SourceImageData, SourceX, SourceY);
				VectorRegister4Float FilteredColor = VectorLoad(&Sample.Component(0));

				// Apply computed alpha scales to each channel
				FilteredColor = VectorMultiply(FilteredColor, AlphaScaleV);

				// Set the destination pixel.
				VectorStore(FilteredColor, &DestImageData.Access(DestX, DestY).Component(0));
			}
		}
	}, EParallelForFlags::Unbalanced);
}

template <EMipGenAddressMode AddressMode>
static void GenerateMip2x2(const FImageView2D& SourceImageData, FImageView2D& DestImageData, FVector4f AlphaScale, uint32 ScaleFactor)
{
	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob, DestImageData.SizeX, DestImageData.SizeY);

	ParallelFor(TEXT("Texture.GenerateMip2x2.PF"), NumJobs, 1, [&](int32 Index)
	{
		VectorRegister4Float AlphaScaleV = VectorLoad(&AlphaScale[0]);
		int32 StartIndex = Index * NumRowsEachJob;
		int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, DestImageData.SizeY);
		for (int32 DestY = StartIndex; DestY < EndIndex; ++DestY)
		{
			for (int32 DestX = 0; DestX < DestImageData.SizeX; DestX++)
			{
				const int32 SourceX = DestX * ScaleFactor;
				const int32 SourceY = DestY * ScaleFactor;

				VectorRegister4Float A = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 0, SourceY + 0).Component(0));
				VectorRegister4Float B = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 1, SourceY + 0).Component(0));
				VectorRegister4Float C = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 0, SourceY + 1).Component(0));
				VectorRegister4Float D = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 1, SourceY + 1).Component(0));
				VectorRegister4Float FilteredColor = VectorAdd(VectorAdd(VectorAdd(A, B), C), D);
				FilteredColor = VectorMultiply(FilteredColor, VectorSetFloat1(0.25f));

				// Apply computed alpha scales to each channel
				FilteredColor = VectorMultiply(FilteredColor, AlphaScaleV);

				// Set the destination pixel.
				VectorStore(FilteredColor, &DestImageData.Access(DestX, DestY).Component(0));
			}
		}
	}, EParallelForFlags::Unbalanced);
}

static void AllocateTempForMips(
	TArray<FLinearColor>& TempData,
	int32 SourceSizeX,
	int32 SourceSizeY,
	int32 DestSizeX,
	int32 DestSizeY,
	bool bDoScaleMipsForAlphaCoverage,
	const FImageKernel2D& Kernel,
	uint32 ScaleFactor,
	bool bSharpenWithoutColorShift,
	bool bUnfiltered,
	bool bUseNewMipFilter)
{
	if (!bUseNewMipFilter)
	{
		// No need for extra memory if using old 2D filter
		return;
	}

	int32 KernelFilterTableSize = (int32)Kernel.GetFilterTableSize();

	if (KernelFilterTableSize == 2 &&
		ScaleFactor == 2 &&
		DestSizeX * 2 == SourceSizeX &&
		DestSizeY * 2 == SourceSizeY &&
		!bDoScaleMipsForAlphaCoverage &&
		!bUnfiltered)
	{
		// Will use GenerateMip2x2Simple
		return;
	}

	if (bUnfiltered)
	{
		// Will use GenerateMipUnfiltered
		return;
	}

	if (KernelFilterTableSize == 2)
	{
		// Will use GenerateMip2x2
		return;
	}

	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob, DestSizeX, DestSizeY);

	// Enough bytes to have one row per source width for each job thread
	// Make sure the rows do not overlap on same cache line
	int64 SizeInBytes = (sizeof(FLinearColor) * SourceSizeX + PLATFORM_CACHE_LINE_SIZE) * NumJobs;

	int64 ElementCount = FMath::DivideAndRoundUp<int64>(SizeInBytes, sizeof(FLinearColor));
	TempData.AddUninitialized(ElementCount);
}

template <EMipGenAddressMode AddressMode, bool bSharpenWithoutColorShift, int KernelSize = 0>
static void GenerateMipSharpened(
	const FImageView2D& SourceImageData,
	FImageView2D& DestImageData,
	const FImageKernel2D& Kernel,
	FVector4f AlphaScale,
	uint32 ScaleFactor)
{
	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob, DestImageData.SizeX, DestImageData.SizeY);

	ParallelFor(TEXT("Texture.GenerateMipSharpened.PF"), NumJobs, 1, [&](int32 Index)
	{
		// In case kernel size is passed as template argument use it as constant
		// This will allow compiler to unroll inner loops below
		const int32 KernelFilterTableSize = KernelSize ? KernelSize : (int32)Kernel.GetFilterTableSize();

		// if KernelFilterTableSize is odd, centered in-place filter can be applied
		// KernelFilterTableSize should be even for standard down-sampling
		const int32 KernelCenter = (KernelFilterTableSize - 1) / 2;

		VectorRegister4Float AlphaScaleV = VectorLoad(&AlphaScale[0]);

		int32 StartIndex = Index * NumRowsEachJob;
		int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, DestImageData.SizeY);

		for (int32 DestY = StartIndex; DestY < EndIndex; ++DestY)
		{
			for (int32 DestX = 0; DestX < DestImageData.SizeX; DestX++)
			{
				const int32 SourceX = DestX * ScaleFactor;
				const int32 SourceY = DestY * ScaleFactor;

				VectorRegister4Float Color = VectorZeroFloat();
				for (int32 KernelY = 0; KernelY < KernelFilterTableSize; ++KernelY)
				{
					for (int32 KernelX = 0; KernelX < KernelFilterTableSize; ++KernelX)
					{
						const FLinearColor& Sample = LookupSourceMip<AddressMode>(SourceImageData, SourceX + KernelX - KernelCenter, SourceY + KernelY - KernelCenter);
						VectorRegister4Float Weight = VectorSetFloat1(Kernel.GetAt(KernelX, KernelY));
						VectorRegister4Float WeightSample = VectorMultiply(Weight, VectorLoad(&Sample.Component(0)));
						Color = VectorAdd(Color, WeightSample);
					}
				}

				// This condition will be optimized away because it is constant template argument
				if (bSharpenWithoutColorShift)
				{
					VectorRegister4Float SharpenedColor = Color;

					// Luminace weights from FLinearColor::GetLuminance() function
					VectorRegister4Float LuminanceWeights = MakeVectorRegisterFloat(0.3f, 0.59f, 0.11f, 0.f);
					VectorRegister4Float NewLuminance = VectorDot3(SharpenedColor, LuminanceWeights);

					// simple 2x2 kernel to compute the color
					VectorRegister4Float A = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 0, SourceY + 0).Component(0));
					VectorRegister4Float B = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 1, SourceY + 0).Component(0));
					VectorRegister4Float C = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 0, SourceY + 1).Component(0));
					VectorRegister4Float D = VectorLoad(&LookupSourceMip<AddressMode>(SourceImageData, SourceX + 1, SourceY + 1).Component(0));
					VectorRegister4Float FilteredColor = VectorAdd(VectorAdd(VectorAdd(A, B), C), D);
					FilteredColor = VectorMultiply(FilteredColor, VectorSetFloat1(0.25f));

					VectorRegister4Float OldLuminance = VectorDot3(FilteredColor, LuminanceWeights);

					// if (OldLuminance > 0.001f) FilteredColor.RGB *= NewLuminance / OldLuminance;
					VectorRegister4Float CompareMask = VectorCompareGT(OldLuminance, VectorSetFloat1(0.001f));
					VectorRegister4Float Temp = VectorMultiply(FilteredColor, VectorDivide(NewLuminance, OldLuminance));
					FilteredColor = VectorSelect(CompareMask, Temp, FilteredColor);

					// FilteredColor.A = SharpenedColor.A
					VectorRegister4Float AlphaMask = MakeVectorRegisterFloatMask(0, 0, 0, 0xffffffff);
					FilteredColor = VectorSelect(AlphaMask, SharpenedColor, FilteredColor);

					// Apply computed alpha scales to each channel
					FilteredColor = VectorMultiply(FilteredColor, AlphaScaleV);

					// Set the destination pixel.
					VectorStore(FilteredColor, &DestImageData.Access(DestX, DestY).Component(0));
				}
				else
				{
					// Apply computed alpha scales to each channel
					Color = VectorMultiply(Color, AlphaScaleV);

					// Set the destination pixel.
					VectorStore(Color, &DestImageData.Access(DestX, DestY).Component(0));
				}
			}
		}
	}, EParallelForFlags::Unbalanced);
}

template <EMipGenAddressMode AddressMode, int KernelSize = 0>
static void GenerateMipSharpenedSeparable(
	const FImageView2D& SourceImageData,
	TArray<FLinearColor>& TempData,
	FImageView2D& DestImageData,
	const FImageKernel2D& Kernel,
	FVector4f AlphaScale,
	uint32 ScaleFactor)
{
	int32 NumRowsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForRows(NumRowsEachJob, DestImageData.SizeX, DestImageData.SizeY);

	// Verify that caller has allocated proper amount of bytes for temporary storage
	// This is allocated in AllocateTempForMips function above
	check(TempData.Num() >= SourceImageData.SizeX * NumJobs);

	ParallelFor(TEXT("Texture.GenerateMipSharpenedSeparable.PF"), NumJobs, 1, [&](int32 Index)
	{
		// In case kernel size is passed as template argument use it as constant
		// This will allow compiler to unroll inner loops below
		const int32 KernelFilterTableSize = KernelSize ? KernelSize : (int32)Kernel.GetFilterTableSize();

		// if KernelFilterTableSize is odd, centered in-place filter can be applied
		// KernelFilterTableSize should be even for standard down-sampling
		const int32 KernelCenter = (KernelFilterTableSize - 1) / 2;

		VectorRegister4Float AlphaScaleV = VectorLoad(&AlphaScale[0]);

		int32 StartIndex = Index * NumRowsEachJob;
		int32 EndIndex = FMath::Min(StartIndex + NumRowsEachJob, DestImageData.SizeY);

		FLinearColor* Temp = &TempData[Index * (TempData.Num() / NumJobs)];
		for (int32 DestY = StartIndex; DestY < EndIndex; ++DestY)
		{
			const int32 SourceY = DestY * ScaleFactor;

			for (int32 SourceX = 0; SourceX < SourceImageData.SizeX; SourceX++)
			{
				VectorRegister4Float Color = VectorZeroFloat();
				for (int32 KernelY = 0; KernelY < KernelFilterTableSize; ++KernelY)
				{
					const FLinearColor& Sample = LookupSourceMip<AddressMode>(SourceImageData, SourceX, SourceY + KernelY - KernelCenter);
					VectorRegister4Float Weight = VectorSetFloat1(Kernel.Get1D(KernelY));
					VectorRegister4Float WeightSample = VectorMultiply(Weight, VectorLoad(&Sample.Component(0)));
					Color = VectorAdd(Color, WeightSample);
				}
				VectorStore(Color, &Temp[SourceX].Component(0));
			}

			for (int32 DestX = 0; DestX < DestImageData.SizeX; DestX++)
			{
				const int32 SourceX = DestX * ScaleFactor;

				VectorRegister4Float Color = VectorZeroFloat();
				for (int32 KernelX = 0; KernelX < KernelFilterTableSize; ++KernelX)
				{
					const FLinearColor& Sample = LookupSourceMip<AddressMode>(Temp, SourceImageData.SizeX, SourceX + KernelX - KernelCenter);
					VectorRegister4Float Weight = VectorSetFloat1(Kernel.Get1D(KernelX));
					VectorRegister4Float WeightSample = VectorMultiply(Weight, VectorLoad(&Sample.Component(0)));
					Color = VectorAdd(Color, WeightSample);
				}

				// Apply computed alpha scales to each channel
				Color = VectorMultiply(Color, AlphaScaleV);

				// Set the destination pixel.
				VectorStore(Color, &DestImageData.Access(DestX, DestY).Component(0));
			}
		}
	}, EParallelForFlags::Unbalanced);
}

/**
* Generates a mip-map for an 2D B8G8R8A8 image using a filter with possible sharpening
* @param SourceImageData - The source image's data.
* @param TempData - Temporary storage required by separable mip filter, call AllocateTempForMips function to allocate it
* @param DestImageData - The destination image's data.
* @param ImageFormat - The format of both the source and destination images.
* @param FilterTable2D - [FilterTableSize * FilterTableSize]
* @param FilterTableSize - >= 2
* @param ScaleFactor 1 / 2:for downsampling
* @param bUseNewMipFilter - pass true to use new separatble mip filter
*/
template <EMipGenAddressMode AddressMode>
static void GenerateSharpenedMipB8G8R8A8Templ(
	const FImageView2D& SourceImageData, 
	TArray<FLinearColor>& TempData,
	FImageView2D& DestImageData, 
	bool bDoScaleMipsForAlphaCoverage,
	const FVector4f AlphaCoverages,
	const FVector4f AlphaThresholds,
	const FImageKernel2D& Kernel,
	uint32 ScaleFactor,
	bool bSharpenWithoutColorShift,
	bool bUnfiltered,
	bool bUseNewMipFilter)
{
	check( ScaleFactor == 1 || ScaleFactor == 2 );
	check( (SourceImageData.SizeX/ScaleFactor) == DestImageData.SizeX || DestImageData.SizeX == 1 );
	check( (SourceImageData.SizeY/ScaleFactor) == DestImageData.SizeY || DestImageData.SizeY == 1 );
	
	int32 KernelFilterTableSize = (int32) Kernel.GetFilterTableSize();

	checkf( KernelFilterTableSize >= 2, TEXT("Kernel table size %d, expected at least 2!"), KernelFilterTableSize);
	if ( KernelFilterTableSize == 2 )
	{
		// 2x2 is always box filter
		check( Kernel.GetAt(0,0) == 0.25f );
	}

	// Keep conditions here in sync with same conditions inside AllocateTempForMips function
	if ( KernelFilterTableSize == 2 &&
		ScaleFactor == 2 &&
		DestImageData.SizeX*2 == SourceImageData.SizeX &&
		DestImageData.SizeY*2 == SourceImageData.SizeY &&
		! bDoScaleMipsForAlphaCoverage &&
		! bUnfiltered )
	{
		// bSharpenWithoutColorShift is ignored for 2x2 filter
		GenerateMip2x2Simple(SourceImageData,DestImageData);
		return;
	}

	FVector4f AlphaScale(1, 1, 1, 1);
	if (bDoScaleMipsForAlphaCoverage)
	{
		AlphaScale = ComputeAlphaScale(AlphaCoverages, AlphaThresholds, SourceImageData);
	}

	if (bUnfiltered)
	{
		GenerateMipUnfiltered<AddressMode>(SourceImageData, DestImageData, AlphaScale, ScaleFactor);
		return;
	}

	if (KernelFilterTableSize == 2)
	{
		GenerateMip2x2<AddressMode>(SourceImageData, DestImageData, AlphaScale, ScaleFactor);
		return;
	}

	if (bUseNewMipFilter)
	{
		if (KernelFilterTableSize == 8)
		{
			GenerateMipSharpenedSeparable<AddressMode, 8>(SourceImageData, TempData, DestImageData, Kernel, AlphaScale, ScaleFactor);
		}
		else
		{
			GenerateMipSharpenedSeparable<AddressMode>(SourceImageData, TempData, DestImageData, Kernel, AlphaScale, ScaleFactor);
		}
	}
	else
	{
		if (bSharpenWithoutColorShift)
		{
			if (KernelFilterTableSize == 8)
			{
				GenerateMipSharpened<AddressMode, true, 8>(SourceImageData, DestImageData, Kernel, AlphaScale, ScaleFactor);
			}
			else
			{
				GenerateMipSharpened<AddressMode, true>(SourceImageData, DestImageData, Kernel, AlphaScale, ScaleFactor);
			}
		}
		else
		{
			if (KernelFilterTableSize == 8)
			{
				GenerateMipSharpened<AddressMode, false, 8>(SourceImageData, DestImageData, Kernel, AlphaScale, ScaleFactor);
			}
			else
			{
				GenerateMipSharpened<AddressMode, false>(SourceImageData, DestImageData, Kernel, AlphaScale, ScaleFactor);
			}
		}
	}
}

// to switch conveniently between different texture wrapping modes for the mip map generation
// the template can optimize the inner loop using a constant AddressMode
static void GenerateSharpenedMipB8G8R8A8(
	const FImageView2D& SourceImageData, 
	const FImageView2D& SourceImageData2, // Only used with volume texture.
	TArray<FLinearColor>& TempData,
	FImageView2D& DestImageData, 
	EMipGenAddressMode AddressMode, 
	bool bDoScaleMipsForAlphaCoverage,
	FVector4f AlphaCoverages,
	FVector4f AlphaThresholds,
	const FImageKernel2D &Kernel,
	uint32 ScaleFactor,
	bool bSharpenWithoutColorShift,
	bool bUnfiltered,
	bool bUseNewMipFilter
	)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateSharpenedMip);

	switch(AddressMode)
	{
	case MGTAM_Wrap:
		GenerateSharpenedMipB8G8R8A8Templ<MGTAM_Wrap>(SourceImageData, TempData, DestImageData, bDoScaleMipsForAlphaCoverage, AlphaCoverages, AlphaThresholds, Kernel, ScaleFactor, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
		break;
	case MGTAM_Clamp:
		GenerateSharpenedMipB8G8R8A8Templ<MGTAM_Clamp>(SourceImageData, TempData, DestImageData, bDoScaleMipsForAlphaCoverage, AlphaCoverages, AlphaThresholds, Kernel, ScaleFactor, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
		break;
	case MGTAM_BorderBlack:
		GenerateSharpenedMipB8G8R8A8Templ<MGTAM_BorderBlack>(SourceImageData, TempData, DestImageData, bDoScaleMipsForAlphaCoverage, AlphaCoverages, AlphaThresholds, Kernel, ScaleFactor, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
		break;
	default:
		check(0);
	}

	// For volume texture, do the average between the 2.
	if (SourceImageData2.IsValid() && !bUnfiltered)
	{
		FImage Temp(DestImageData.SizeX, DestImageData.SizeY, 1, ERawImageFormat::RGBA32F);
		FImageView2D TempImageData (Temp, 0);

		switch(AddressMode)
		{
		case MGTAM_Wrap:
			GenerateSharpenedMipB8G8R8A8Templ<MGTAM_Wrap>(SourceImageData2, TempData, TempImageData, bDoScaleMipsForAlphaCoverage, AlphaCoverages, AlphaThresholds, Kernel, ScaleFactor, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
			break;
		case MGTAM_Clamp:
			GenerateSharpenedMipB8G8R8A8Templ<MGTAM_Clamp>(SourceImageData2, TempData, TempImageData, bDoScaleMipsForAlphaCoverage, AlphaCoverages, AlphaThresholds, Kernel, ScaleFactor, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
			break;
		case MGTAM_BorderBlack:
			GenerateSharpenedMipB8G8R8A8Templ<MGTAM_BorderBlack>(SourceImageData2, TempData, TempImageData, bDoScaleMipsForAlphaCoverage, AlphaCoverages, AlphaThresholds, Kernel, ScaleFactor, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
			break;
		default:
			check(0);
		}

		const int32 NumColors = DestImageData.SizeX * DestImageData.SizeY;
		for (int32 ColorIndex = 0; ColorIndex < NumColors; ++ColorIndex)
		{
			DestImageData.SliceColors[ColorIndex] =
				(DestImageData.SliceColors[ColorIndex] + TempImageData.SliceColors[ColorIndex]) * 0.5f;
		}
	}
}

// Update border texels after normal mip map generation to preserve the colors there (useful for particles and decals).
static void GenerateMipBorder(
	const FImageView2D& SrcImageData, 
	FImageView2D& DestImageData
	)
{
	check( (SrcImageData.SizeX/2) == DestImageData.SizeX || DestImageData.SizeX == 1 );
	check( (SrcImageData.SizeY/2) == DestImageData.SizeY || DestImageData.SizeY == 1 );

	// this check is unnecessary if we always used MGTAM_Clamp here;
	bool bIsPow2 = FMath::IsPowerOfTwo(SrcImageData.SizeX) && FMath::IsPowerOfTwo(SrcImageData.SizeY);

	for ( int32 DestY = 0; DestY < DestImageData.SizeY; DestY++ )
	{
		for ( int32 DestX = 0; DestX < DestImageData.SizeX; )
		{
			FLinearColor FilteredColor(0, 0, 0, 0);
			{
				float WeightSum = 0.0f;
				for ( int32 KernelY = 0; KernelY < 2;  ++KernelY )
				{
					for ( int32 KernelX = 0; KernelX < 2;  ++KernelX )
					{
						const int32 SourceX = DestX * 2 + KernelX;
						const int32 SourceY = DestY * 2 + KernelY;

						// only average the source border
						if ( SourceX == 0 ||
							SourceX == SrcImageData.SizeX - 1 ||
							SourceY == 0 ||
							SourceY == SrcImageData.SizeY - 1 )
						{
							// I think this should have just always been MGTAM_Clamp
							//	but that changes existing content, so preserve old behavior of using _Wrap :(
							FLinearColor Sample;
							if ( bIsPow2 )
							{
								Sample = LookupSourceMip<MGTAM_Wrap>( SrcImageData, SourceX, SourceY );
							}
							else
							{
								Sample = LookupSourceMip<MGTAM_Clamp>( SrcImageData, SourceX, SourceY );
							}
							FilteredColor += Sample;
							WeightSum += 1.0f;
						}
					}
				}
				FilteredColor /= WeightSum;
			}

			// Set the destination pixel.
			//FLinearColor& DestColor = *(DestImageData.AsRGBA32F() + DestX + DestY * DestImageData.SizeX);
			FLinearColor& DestColor = DestImageData.Access(DestX, DestY);
			DestColor = FilteredColor;

			++DestX;

			if ( DestY > 0 &&
				DestY < DestImageData.SizeY - 1 &&
				DestX > 0 &&
				DestX < DestImageData.SizeX - 1 )
			{
				// jump over the non border area
				DestX += FMath::Max( 1, DestImageData.SizeX - 2 );
			}
		}
	}
}

// how should be treat lookups outside of the image
static EMipGenAddressMode ComputeAddressMode(const FImage & Image,const FTextureBuildSettings& Settings)
{
	if ( Settings.bPreserveBorder )
	{
		return Settings.bBorderColorBlack ? MGTAM_BorderBlack : MGTAM_Clamp;
	}
	
	// conditional on bUseNewMipFilter so old textures are not changed
	//	really this should be done all the time
	if ( Settings.bCubemap && Settings.bUseNewMipFilter )
	{
		// Cubemaps should Clamp on their faces, never wrap :
		return MGTAM_Clamp;
	}

	// Wrap uses AND so requires pow2 sizes ; change to Clamp if nonpow2
	bool bIsPow2 = FMath::IsPowerOfTwo(Image.SizeX) && FMath::IsPowerOfTwo(Image.SizeY);
	// 2d address mode, no need to look at volume z :
	/*
	if ( Settings.bVolume && ! FMath::IsPowerOfTwo(Image.NumSlices) )
	{
		bIsPow2 = false;
	}
	*/

	if ( ! bIsPow2 )
	{
		return MGTAM_Clamp;
	}

	// note: all textures Wrap by default even if their address mode is set to Clamp !?
	EMipGenAddressMode AddressMode = MGTAM_Wrap;

	if ( Settings.bUseNewMipFilter )
	{
		// in new filters, we do use address mode to change mipgen
		//	(isolated to new filters solely to avoid changing old content)

		// texture address mode is Wrap,Clamp,Mirror, with Wrap by default
		// treat Clamp & Mirror the same
		bool bAddressXWrap = (Settings.TextureAddressModeX == TA_Wrap);
		bool bAddressYWrap = (Settings.TextureAddressModeY == TA_Wrap);

		if ( !bAddressXWrap && !bAddressYWrap )
		{
			AddressMode = MGTAM_Clamp;
		}
		else if ( bAddressXWrap && bAddressYWrap )
		{
			AddressMode = MGTAM_Wrap;
		}
		else
		{
			// use Wrap for now to match legacy behavior
			// ideally we'd wrap one dimension and clamp the other, but that is not yet supported
			// @todo Oodle: separate wrap/clamp for X&Y ; remove the heavy templating on AddressMode
			AddressMode = MGTAM_Wrap;

			UE_LOG(LogTextureCompressor, Verbose, TEXT("Not ideal: Heterogeneous XY Wrap/Clamp will generate mips with Wrap on both dimensions") );
		}
	}

	return AddressMode;
}

static void GenerateTopMip(const FImage& SrcImage, FImage& DestImage, const FTextureBuildSettings& Settings)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateTopMip);

	// GenerateTopMip is only used for ApplyCompositeTexture
	// bApplyKernelToTopMip is not exposed to Texture GUI

	EMipGenAddressMode AddressMode = ComputeAddressMode(SrcImage,Settings);

	FImageKernel2D KernelDownsample;

	if ( Settings.MipSharpening < 0.f )
	{
		// negative Sharpening is a Gaussian
		//  this can make centered ("odd") filters, so the image doesn't shift
		int32 OddMipKernelSize = Settings.SharpenMipKernelSize | 1;
		KernelDownsample.BuildSeparatableGaussWithSharpen( OddMipKernelSize, Settings.MipSharpening );
	}
	else
	{	
		// non-Gaussians only support "even" filters
		//	this causes a half-pixel shift of the top mip
		// warn but then go ahead and do as requested
		UE_LOG(LogTextureCompressor, Warning, TEXT("GenerateTopMip used with non-Gaussian blur filter will cause half pixel shift"));

		KernelDownsample.BuildSeparatableGaussWithSharpen( Settings.SharpenMipKernelSize, Settings.MipSharpening );
	}

	DestImage.Init(SrcImage.SizeX, SrcImage.SizeY, SrcImage.NumSlices, SrcImage.Format, SrcImage.GammaSpace);

	const bool bSharpenWithoutColorShift = Settings.bSharpenWithoutColorShift;
	const bool bUnfiltered = Settings.MipGenSettings == TMGS_Unfiltered;
	const bool bUseNewMipFilter = Settings.bUseNewMipFilter;

	TArray<FLinearColor> TempData;
	AllocateTempForMips(TempData, SrcImage.SizeX, SrcImage.SizeY, DestImage.SizeX, DestImage.SizeY, false, KernelDownsample, 1, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);

	for (int32 SliceIndex = 0; SliceIndex < SrcImage.NumSlices; ++SliceIndex)
	{
		FImageView2D SrcView((FImage&)SrcImage, SliceIndex);
		FImageView2D DestView(DestImage, SliceIndex);

		// generate DestImage: down sample with sharpening
		GenerateSharpenedMipB8G8R8A8(
			SrcView, 
			FImageView2D(),
			TempData,
			DestView,
			AddressMode,
			false,
			FVector4f(0, 0, 0, 0),
			FVector4f(0, 0, 0, 0),
			KernelDownsample,
			1,
			Settings.bSharpenWithoutColorShift,
			bUnfiltered,
			bUseNewMipFilter);
	}
}

static FLinearColor Bilerp(
	const FLinearColor & Sample00,
	const FLinearColor & Sample10,
	const FLinearColor & Sample01,
	const FLinearColor & Sample11,
	float FracX,float FracY)
{
	//FMath::Lerp is :
	// return (T)(A + Alpha * (B-A));
	// must match that exactly

	VectorRegister4Float V00 = VectorLoad(&Sample00.Component(0));
	VectorRegister4Float V10 = VectorLoad(&Sample10.Component(0));
	VectorRegister4Float V01 = VectorLoad(&Sample01.Component(0));
	VectorRegister4Float V11 = VectorLoad(&Sample11.Component(0));

	VectorRegister4Float S0 = VectorAdd(V00, VectorMultiply(VectorSetFloat1(FracX), VectorSubtract(V10,V00)));
	VectorRegister4Float S1 = VectorAdd(V01, VectorMultiply(VectorSetFloat1(FracX), VectorSubtract(V11,V01)));
	VectorRegister4Float SS = VectorAdd(S0 , VectorMultiply(VectorSetFloat1(FracY), VectorSubtract(S1 ,S0 )));
	FLinearColor Out;
	VectorStore(SS,&Out.Component(0));
	
	/*
	// FLinearColor has math operators but they are scalar, not vector
	FLinearColor Sample0 = FMath::Lerp(Sample00, Sample10, FracX);
	FLinearColor Sample1 = FMath::Lerp(Sample01, Sample11, FracX);
	FLinearColor Ret = FMath::Lerp(Sample0, Sample1, FracY);
	check( Out == Ret );
	*/

	return Out;
}

// pixel centers are at XY = integers
//  range of X is [-0.5,-0.5] to [Width-0.5,Height-0.5]
static FLinearColor LookupSourceMipBilinear(const FImageView2D& SourceImageData, float X, float Y)
{
	X = FMath::Clamp(X, 0.f, SourceImageData.SizeX - 1.f);
	Y = FMath::Clamp(Y, 0.f, SourceImageData.SizeY - 1.f);
	int32 IntX0 = FMath::TruncToInt32(X);
	int32 IntY0 = FMath::TruncToInt32(Y);
	float FractX = X - IntX0;
	float FractY = Y - IntY0;
	int32 IntX1 = FMath::Min(IntX0+1, SourceImageData.SizeX-1);
	int32 IntY1 = FMath::Min(IntY0+1, SourceImageData.SizeY-1);
	
	const FLinearColor & Sample00 = SourceImageData.Access(IntX0,IntY0);
	const FLinearColor & Sample10 = SourceImageData.Access(IntX1,IntY0);
	const FLinearColor & Sample01 = SourceImageData.Access(IntX0,IntY1);
	const FLinearColor & Sample11 = SourceImageData.Access(IntX1,IntY1);

	return Bilerp(Sample00,Sample10,Sample01,Sample11,FractX,FractY);
}

// UV range is [0,1] , first pixel center is at 0.5/W
static FLinearColor LookupSourceMipBilinearUV(const FImageView2D& SourceImageData, float U, float V)
{
	float X = U * SourceImageData.SizeX - 0.5f;
	float Y = V * SourceImageData.SizeY - 0.5f;
	return LookupSourceMipBilinear(SourceImageData,X,Y);
}

// pixel centers are at XY = integers
//  range of X is [-0.5,-0.5] to [Width-0.5,Height-0.5]
static float LookupFloatBilinear(const float * FloatPlane,int32 SizeX,int32 SizeY, float X, float Y)
{
	X = FMath::Clamp(X, 0.f, SizeX - 1.f);
	Y = FMath::Clamp(Y, 0.f, SizeY - 1.f);
	int32 IntX0 = FMath::TruncToInt32(X);
	int32 IntY0 = FMath::TruncToInt32(Y);
	float FractX = X - IntX0;
	float FractY = Y - IntY0;
	int32 IntX1 = FMath::Min(IntX0+1, SizeX-1);
	int32 IntY1 = FMath::Min(IntY0+1, SizeY-1);
	
	float Sample00 = FloatPlane[ IntX0 + IntY0 * (int64) SizeX ];
	float Sample10 = FloatPlane[ IntX1 + IntY0 * (int64) SizeX ];
	float Sample01 = FloatPlane[ IntX0 + IntY1 * (int64) SizeX ];
	float Sample11 = FloatPlane[ IntX1 + IntY1 * (int64) SizeX ];
	float Sample0 = FMath::Lerp(Sample00, Sample10, FractX);
	float Sample1 = FMath::Lerp(Sample01, Sample11, FractX);
		
	return FMath::Lerp(Sample0, Sample1, FractY);
}

// UV range is [0,1] , first pixel center is at 0.5/W
static float LookupFloatBilinearUV(const float * FloatPlane,int32 SizeX,int32 SizeY, float U, float V)
{
	float X = U * SizeX - 0.5f;
	float Y = V * SizeY - 0.5f;
	return LookupFloatBilinear(FloatPlane,SizeX,SizeY,X,Y);
}

struct FTextureDownscaleSettings
{
	int32 BlockSize;
	float Downscale;
	uint8 DownscaleOptions;
	bool UseNewMipFilter;

	FTextureDownscaleSettings(const FTextureBuildSettings& BuildSettings)
	{
		Downscale = BuildSettings.Downscale;
		DownscaleOptions = BuildSettings.DownscaleOptions;
		BlockSize = 4; // <- hard coded to 4 even if texture is not compressed
		UseNewMipFilter = BuildSettings.bUseNewMipFilter;
	}
};

static float GetDownscaleFinalSizeAndClampedDownscale(int32 SrcImageWidth, int32 SrcImageHeight,  const FTextureDownscaleSettings& Settings, int32& OutWidth, int32& OutHeight)
{
	check(Settings.Downscale > 1.0f); // must be already handled.
	
	float Downscale = FMath::Clamp(Settings.Downscale, 1.f, 8.f);

	// note: more accurate would be to use FMath::Max(1, FMath::RoundToInt(SrcImage.SizeX / Downscale))
	int32 FinalSizeX = FMath::CeilToInt(SrcImageWidth / Downscale);
	int32 FinalSizeY = FMath::CeilToInt(SrcImageHeight / Downscale);

	// compute final size respecting image block size
	if (Settings.BlockSize > 1
		&& (SrcImageWidth % Settings.BlockSize) == 0
		&& (SrcImageHeight % Settings.BlockSize) == 0)
	{
		// the following code finds non-zero dimensions of the scaled image which preserve both aspect ratio and block alignment, 
		// it favors preserving aspect ratio at the expense of not scaling the desired factor when both are not possible
		int32 GCD = FMath::GreatestCommonDivisor(SrcImageWidth, SrcImageHeight);
		int32 ScalingGridSizeX = (SrcImageWidth / GCD) * Settings.BlockSize;
		// note: more accurate would be to use (SrcImage.SizeX / Downscale) instead of FinalSizeX here
		// GridSnap rounds to nearest, and can return zero
		FinalSizeX = FMath::GridSnap(FinalSizeX, ScalingGridSizeX);
		FinalSizeX = FMath::Max(ScalingGridSizeX, FinalSizeX);
		FinalSizeY = (int32)( ((int64)FinalSizeX * SrcImageHeight) / SrcImageWidth );
		// Final Size X and Y are gauranteed to be block aligned

		#if 0
		
		// simpler alternative :
		// choose the block count in the smaller dimension first
		// then make the larger dimension maintain aspect ratio
		int32 FinalNumBlocksX,FinalNumBlocksY;
		if ( SrcImage.SizeX >= SrcImage.SizeY )
		{
			FinalNumBlocksY = FMath::RoundToInt( SrcImage.SizeY / (Downscale * Settings.BlockSize) );
			FinalNumBlocksX = FMath::RoundToInt( FinalNumBlocksY * SrcImage.SizeX / (float)SrcImage.SizeY );
		}
		else
		{
			FinalNumBlocksX = FMath::RoundToInt( SrcImage.SizeX / (Downscale * Settings.BlockSize) );
			FinalNumBlocksY = FMath::RoundToInt( FinalNumBlocksX * SrcImage.SizeY / (float)SrcImage.SizeX );
		}

		FinalSizeX = FMath::Max(FinalNumBlocksX,1)*Settings.BlockSize;
		FinalSizeY = FMath::Max(FinalNumBlocksY,1)*Settings.BlockSize;

		#endif
	}

	OutWidth = FinalSizeX;
	OutHeight = FinalSizeY;
	return Downscale;
}

static void DownscaleImage(const FImage& SrcImage, FImage& DstImage, const FTextureDownscaleSettings& Settings)
{
	if (Settings.Downscale <= 1.f)
	{
		return;
	}
	
	// Settings.BlockSize == 4 always, even if not compressed
	// Downscale can only be applied if NoMipMaps, otherwise it is silently ignored
	//	 also only applied if Texture2D , ignored by all other texture types

	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.DownscaleImage);

	int32 FinalSizeX = 0, FinalSizeY =0;
	float Downscale = GetDownscaleFinalSizeAndClampedDownscale(SrcImage.SizeX, SrcImage.SizeY, Settings, FinalSizeX, FinalSizeY);
	
	//@todo OodleImageResize : replace this whole function with better image resizer if NewFilters

	// recompute Downscale factor because it may have changed due to block alignment
	// note: if aspect ratio was not exactly preserved, this could differ in X and Y
	Downscale = (float)SrcImage.SizeX / FinalSizeX;
	
	// while desired final size is > 2X smaller, do 2X resizes using the standard mip gen filter code

	FImage Image0;
	FImage Image1;
	FImage* ImageChain[2] = {&const_cast<FImage&>(SrcImage), &Image1};
	const bool bUnfiltered = Settings.DownscaleOptions == (uint8)ETextureDownscaleOptions::Unfiltered;
	const bool bUseNewMipFilter = Settings.UseNewMipFilter;
	
	// Scaledown using 2x2 average, use user specified filtering only for last iteration
	FImageKernel2D AvgKernel;
	AvgKernel.BuildSeparatableGaussWithSharpen(2);

	TArray<FLinearColor> TempData;
	AllocateTempForMips(TempData, SrcImage.SizeX, SrcImage.SizeY, FMath::Max(1, SrcImage.SizeX / 2), FMath::Max(1, SrcImage.SizeY / 2), false, AvgKernel, 2, false, bUnfiltered, bUseNewMipFilter);

	int32 NumIterations = 0;
	while(Downscale > 2.0f)
	{
		int32 DstSizeX = FMath::Max(1, ImageChain[0]->SizeX / 2 );
		int32 DstSizeY = FMath::Max(1, ImageChain[0]->SizeY / 2 );
		ImageChain[1]->Init(DstSizeX, DstSizeY, ImageChain[0]->NumSlices, ImageChain[0]->Format, ImageChain[0]->GammaSpace);

		FImageView2D SrcImageData(*ImageChain[0], 0);
		FImageView2D DstImageData(*ImageChain[1], 0);
		GenerateSharpenedMipB8G8R8A8Templ<MGTAM_Clamp>(
			SrcImageData, 
			TempData,
			DstImageData, 
			false,
			FVector4f(0, 0, 0, 0),
			FVector4f(0, 0, 0, 0), 
			AvgKernel, 
			2, 
			false,
			bUnfiltered,
			bUseNewMipFilter);

		if (NumIterations == 0)
		{
			ImageChain[0] = &Image0;
		}
		Swap(ImageChain[0], ImageChain[1]);
		
		NumIterations++;
		Downscale/= 2.f;
	}

	if (ImageChain[0]->SizeX == FinalSizeX &&
		ImageChain[0]->SizeY == FinalSizeY)
	{
		ImageChain[0]->CopyTo(DstImage, ImageChain[0]->Format, ImageChain[0]->GammaSpace);
		return;
	}
	
	int32 KernelSize = 2;
	float Sharpening = 0.0f;
	if (Settings.DownscaleOptions >= (uint8)ETextureDownscaleOptions::Sharpen0 && Settings.DownscaleOptions <= (uint8)ETextureDownscaleOptions::Sharpen10)
	{
		// 0 .. 2.0f
		Sharpening = ((int32)Settings.DownscaleOptions - (int32)ETextureDownscaleOptions::Sharpen0) * 0.2f;
		KernelSize = 8;
	}
	
	bool bBilinear = Settings.DownscaleOptions == (uint8)ETextureDownscaleOptions::SimpleAverage;
	
	FImageKernel2D KernelSharpen;
	KernelSharpen.BuildSeparatableGaussWithSharpen(KernelSize, Sharpening);
	const int32 KernelCenter = (int32)KernelSharpen.GetFilterTableSize() / 2 - 1;
		
	ImageChain[1] = &DstImage;
	if (ImageChain[0] == ImageChain[1])
	{
		ImageChain[0]->CopyTo(Image0, ImageChain[0]->Format, ImageChain[0]->GammaSpace);
		ImageChain[0] = &Image0;
	}
	
	ImageChain[1]->Init(FinalSizeX, FinalSizeY, ImageChain[0]->NumSlices, ImageChain[0]->Format, ImageChain[0]->GammaSpace);

	// recompute Downscale factor again for final arbitrary-factor resizes :
	// note: if aspect ratio was not exactly preserved, this could differ in X and Y
	Downscale = (float)ImageChain[0]->SizeX / FinalSizeX;

	FImageView2D SrcImageData(*ImageChain[0], 0);
	FImageView2D DstImageData(*ImageChain[1], 0);
				
	// @todo OodleImageResize : not sure this is a correct image resize without shift; does it get pixel center offsets right?
	for (int32 Y = 0; Y < FinalSizeY; ++Y)
	{
		float SourceY = Y * Downscale;
		int32 IntSourceY = FMath::RoundToInt(SourceY);
		
		for (int32 X = 0; X < FinalSizeX; ++X)
		{
			float SourceX = X * Downscale;
			int32 IntSourceX = FMath::RoundToInt(SourceX);

			FLinearColor FilteredColor(0,0,0,0);

			if (bUnfiltered)
			{
				FilteredColor = LookupSourceMip<MGTAM_Clamp>(SrcImageData, IntSourceX, IntSourceY);
			}
			else if(bBilinear)
			{
				FilteredColor = LookupSourceMipBilinear(SrcImageData, SourceX, SourceY);
			}
			else
			{
				for (uint32 KernelY = 0; KernelY < KernelSharpen.GetFilterTableSize();  ++KernelY)
				{
					for (uint32 KernelX = 0; KernelX < KernelSharpen.GetFilterTableSize();  ++KernelX)
					{
						float Weight = KernelSharpen.GetAt(KernelX, KernelY);
						FLinearColor Sample = LookupSourceMipBilinear(SrcImageData, SourceX + KernelX - KernelCenter, SourceY + KernelY - KernelCenter);
						FilteredColor += Weight	* Sample;
					}
				}
			}

			// Set the destination pixel.
			FLinearColor& DestColor = DstImageData.Access(X, Y);
			DestColor = FilteredColor;
		}
	}
}

void ITextureCompressorModule::GenerateMipChain(
	const FTextureBuildSettings& Settings,
	const FImage& BaseImage,
	TArray<FImage> &OutMipChain,
	uint32 MipChainDepth 
	)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateMipChain);

	// MipChainDepth is the number more to make, OutMipChain has some already
	// typically BaseImage == OutMipChain.Last()
	if ( MipChainDepth == 0 )
	{
		return;
	}

	check(BaseImage.Format == ERawImageFormat::RGBA32F);

	const FImage& BaseMip = BaseImage;
	const int32 SrcWidth = BaseMip.SizeX;
	const int32 SrcHeight= BaseMip.SizeY;
	const int32 SrcNumSlices = BaseMip.NumSlices;
	const ERawImageFormat::Type ImageFormat = ERawImageFormat::RGBA32F;

	const FImage* IntermediateSrcPtr;
	FImage* IntermediateDstPtr;

	// This will be used as a buffer for the mip processing
	FImage FirstTempImage;

	if (BaseMip.GammaSpace != EGammaSpace::Linear)
	{
		// copy base mip
		BaseMip.CopyTo(FirstTempImage, ERawImageFormat::RGBA32F, EGammaSpace::Linear);

		IntermediateSrcPtr = &FirstTempImage;
	}
	else
	{
		// It looks like the BaseMip can be reused for the intermediate source of the second Mip (assuming that the format was check earlier to be RGBA32F)
		IntermediateSrcPtr = &BaseMip;

		// This temp image will be first used as an intermediate destination for the third mip in the chain
		FirstTempImage.Init( FMath::Max<uint32>( 1, SrcWidth >> 2 ), FMath::Max<uint32>( 1, SrcHeight >> 2 ), Settings.bVolume ? FMath::Max<uint32>( 1, SrcNumSlices >> 2 ) : SrcNumSlices, ImageFormat );
	}

	// The image for the first destination
	FImage SecondTempImage(FMath::Max<uint32>( 1, SrcWidth >> 1 ), FMath::Max<uint32>( 1, SrcHeight >> 1 ), Settings.bVolume ? FMath::Max<uint32>( 1, SrcNumSlices >> 1 ) : SrcNumSlices, ImageFormat);
	IntermediateDstPtr = &SecondTempImage;

	// Filtering kernels.
	FImageKernel2D KernelSimpleAverage;
	FImageKernel2D KernelDownsample;
	KernelSimpleAverage.BuildSeparatableGaussWithSharpen( 2 );
	KernelDownsample.BuildSeparatableGaussWithSharpen( Settings.SharpenMipKernelSize, Settings.MipSharpening );

	//@TODO : add a true 3D kernel.

	EMipGenAddressMode AddressMode = ComputeAddressMode(*IntermediateSrcPtr,Settings);
	bool bReDrawBorder = false;
	if( Settings.bPreserveBorder )
	{
		bReDrawBorder = !Settings.bBorderColorBlack;
	}

	// Calculate alpha coverage value to preserve along mip chain
	FVector4f AlphaCoverages(0, 0, 0, 0);
	if ( Settings.bDoScaleMipsForAlphaCoverage )
	{
		check(Settings.AlphaCoverageThresholds != FVector4f(0,0,0,0));
		check(IntermediateSrcPtr);
		const FImageView2D IntermediateSrcView = FImageView2D::ConstructConst(*IntermediateSrcPtr, 0);

		const FVector4f AlphaScales(1, 1, 1, 1);		
		AlphaCoverages = ComputeAlphaCoverage(Settings.AlphaCoverageThresholds, AlphaScales, IntermediateSrcView);
	}

	TArray<FLinearColor> DownsampleTempData;
	TArray<FLinearColor> AverageTempData;
	const bool bDoScaleMipsForAlphaCoverage = Settings.bDoScaleMipsForAlphaCoverage;
	const bool bSharpenWithoutColorShift = Settings.bSharpenWithoutColorShift;
	const bool bUnfiltered = Settings.MipGenSettings == TMGS_Unfiltered;
	const bool bUseNewMipFilter = Settings.bUseNewMipFilter;
	AllocateTempForMips(DownsampleTempData, BaseMip.SizeX, BaseMip.SizeY, SecondTempImage.SizeX, SecondTempImage.SizeY, bDoScaleMipsForAlphaCoverage, KernelDownsample, 2, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);
	AllocateTempForMips(AverageTempData, BaseMip.SizeX, BaseMip.SizeY, SecondTempImage.SizeX, SecondTempImage.SizeY, bDoScaleMipsForAlphaCoverage, KernelSimpleAverage, 2, bSharpenWithoutColorShift, bUnfiltered, bUseNewMipFilter);

	// Generate mips
	//  default value of MipChainDepth is MAX_uint32, means generate all mips down to 1x1
	//	(break inside the loop)
	for (; MipChainDepth != 0 ; --MipChainDepth)
	{
		check(IntermediateSrcPtr && IntermediateDstPtr);
		const FImage& IntermediateSrc = *IntermediateSrcPtr;
		FImage& IntermediateDst = *IntermediateDstPtr;
		
		if ( IntermediateSrc.SizeX == 1 && IntermediateSrc.SizeY == 1 && (!Settings.bVolume || IntermediateSrc.NumSlices == 1))
		{
			// should not have been called, starting mip is already small enough
			check(0);
			break;
		}

		// add new mip to TArray<FImage> &OutMipChain :
		//	placement new on TArray does AddUninitialized then constructs in the last element
		FImage& DestImage = *new(OutMipChain) FImage(IntermediateDst.SizeX, IntermediateDst.SizeY, IntermediateDst.NumSlices, ImageFormat);
		
		for (int32 SliceIndex = 0; SliceIndex < IntermediateDst.NumSlices; ++SliceIndex)
		{
			const int32 SrcSliceIndex = Settings.bVolume ? (SliceIndex * 2) : SliceIndex;
			const FImageView2D IntermediateSrcView = FImageView2D::ConstructConst(IntermediateSrc, SrcSliceIndex);
			FImageView2D IntermediateSrcView2;
			if ( Settings.bVolume )
			{
				if ( SrcSliceIndex + 1 < IntermediateSrc.NumSlices )
				{
					IntermediateSrcView2 = FImageView2D::ConstructConst(IntermediateSrc, SrcSliceIndex + 1);
				}
				else
				{
					// nonpow2 volume sizeZ , clamp slice index
					IntermediateSrcView2 = FImageView2D::ConstructConst(IntermediateSrc, SrcSliceIndex);
				}
			}
			FImageView2D DestView(DestImage, SliceIndex);
			FImageView2D IntermediateDstView(IntermediateDst, SliceIndex);

			// @@CB these two mipgens can run in parallel I believe?
			//	and the one to the output mip is not blocking
			//	we only need to block on the downsample one

			// DestView is the output mip
			GenerateSharpenedMipB8G8R8A8(
				IntermediateSrcView, 
				IntermediateSrcView2,
				DownsampleTempData,
				DestView,
				AddressMode,
				bDoScaleMipsForAlphaCoverage,
				AlphaCoverages,
				Settings.AlphaCoverageThresholds,
				KernelDownsample,
				2,
				bSharpenWithoutColorShift,
				bUnfiltered,
				bUseNewMipFilter);

			// generate IntermediateDstImage:
			// IntermediateDstImage will be the source for the next mip
			if ( Settings.bDownsampleWithAverage )
			{
				// down sample without sharpening for the next iteration
				// bDownsampleWithAverage comes from GetMipGenSettings
				// it is on by default for all cases except Blur
				// it means every mip is generated *twice*
				// the output mip is made above using Sharpen from the IntermediateSrc
				// then the next source is made here using 2x2 ("SimpleAverage")
				// the next IntermediateSrc is not my outputmip, it's what is made here

				GenerateSharpenedMipB8G8R8A8(
					IntermediateSrcView,
					IntermediateSrcView2,
					AverageTempData,
					IntermediateDstView,
					AddressMode,
					bDoScaleMipsForAlphaCoverage,
					AlphaCoverages,
					Settings.AlphaCoverageThresholds,
					KernelSimpleAverage,
					2,
					bSharpenWithoutColorShift,
					bUnfiltered,
					bUseNewMipFilter);
			}
		}

		if ( Settings.bDownsampleWithAverage == false )
		{
			TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateMipChain.memcpy);

			FMemory::Memcpy( (&IntermediateDst.AsRGBA32F()[0]), (&DestImage.AsRGBA32F()[0]),
				IntermediateDst.SizeX * IntermediateDst.SizeY * IntermediateDst.NumSlices * sizeof(FLinearColor) );
		}

		if ( bReDrawBorder )
		{
			for (int32 SliceIndex = 0; SliceIndex < IntermediateDst.NumSlices; ++SliceIndex)
			{
				const FImageView2D IntermediateSrcView = FImageView2D::ConstructConst(IntermediateSrc, SliceIndex);
				FImageView2D DestView(DestImage, SliceIndex);
				FImageView2D IntermediateDstView(IntermediateDst, SliceIndex);
				GenerateMipBorder( IntermediateSrcView, DestView );
				GenerateMipBorder( IntermediateSrcView, IntermediateDstView );
			}
		}

		// Once we've created mip-maps down to 1x1, we're done.
		if ( IntermediateDst.SizeX == 1 && IntermediateDst.SizeY == 1 && (!Settings.bVolume || IntermediateDst.NumSlices == 1))
		{
			break;
		}

		// last destination becomes next source
		if (IntermediateDstPtr == &SecondTempImage)
		{
			IntermediateDstPtr = &FirstTempImage;
			IntermediateSrcPtr = &SecondTempImage;
		}
		else
		{
			IntermediateDstPtr = &SecondTempImage;
			IntermediateSrcPtr = &FirstTempImage;
		}

		// Update the destination size for the next iteration.
		IntermediateDstPtr->SizeX = FMath::Max<uint32>( 1, IntermediateSrcPtr->SizeX >> 1 );
		IntermediateDstPtr->SizeY = FMath::Max<uint32>( 1, IntermediateSrcPtr->SizeY >> 1 );
		IntermediateDstPtr->NumSlices = Settings.bVolume ? FMath::Max<uint32>( 1, IntermediateSrcPtr->NumSlices >> 1 ) : SrcNumSlices;
	}

	{
		TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateMipChain.Free);
		
		// @@CB this is significant; detach trick again?

		FirstTempImage = FImage();
		SecondTempImage = FImage();
	}
}

/*------------------------------------------------------------------------------
	Angular Filtering for HDR Cubemaps.
------------------------------------------------------------------------------*/

/**
 * View in to an image that allows access by converting a direction to longitude and latitude.
 */
struct FImageViewLongLat
{
	/** Image colors. */
	FLinearColor* ImageColors;
	/** Width of the image. */
	int64 SizeX;
	/** Height of the image. */
	int64 SizeY;

	/** Initialization constructor. */
	explicit FImageViewLongLat(FImage& Image, int32 SliceIndex)
	{
		SizeX = Image.SizeX;
		SizeY = Image.SizeY;
		ImageColors = (&Image.AsRGBA32F()[0]) + SliceIndex * SizeY * SizeX;
	}

	/** Const access to a texel. */
	const FLinearColor & Access(int32 X, int32 Y) const
	{
		return ImageColors[X + Y * SizeX];
	}

	/** Makes a filtered lookup. */
	FLinearColor LookupFiltered(float X, float Y) const
	{
		// X is in (0,SizeX) (exclusive)
		// Y is in (0,SizeY)
		checkSlow( X > 0.f && X < SizeX );
		checkSlow( Y > 0.f && Y < SizeY );

		int32 X0 = (int32)X;
		int32 Y0 = (int32)Y;

		float FracX = X - X0;
		float FracY = Y - Y0;

		int32 X1 = X0 + 1;
		int32 Y1 = Y0 + 1;

		// wrap X :
		checkSlow( X0 >= 0 && X0 < SizeX );
		if ( X1 >= SizeX )
		{
			X1 = 0;
		}

		// clamp Y :
		// clamp should only ever change SizeY to SizeY -1 ?
		checkSlow( Y0 >= 0 && Y0 < SizeY );
		if ( Y1 >= SizeY )
		{
			Y1 = SizeY-1;
		}

		const FLinearColor & CornerRGB00 = Access(X0, Y0);
		const FLinearColor & CornerRGB10 = Access(X1, Y0);
		const FLinearColor & CornerRGB01 = Access(X0, Y1);
		const FLinearColor & CornerRGB11 = Access(X1, Y1);

		return Bilerp(CornerRGB00,CornerRGB10,CornerRGB01,CornerRGB11,FracX,FracY);
	}

	/** Makes a filtered lookup using a direction. */
	// see http://gl.ict.usc.edu/Data/HighResProbes
	// latitude-longitude panoramic format = equirectangular mapping

	// using floats saves ~6% of the total time
	//	but would change output
	// cycles_float = 135 cycles_double = 143
	FLinearColor LookupLongLatFloat(const FVector & NormalizedDirection) const
	{
		// atan2 returns in [-PI,PI]
		// acos returns in [0,PI]

		const float invPI = 1.f/PI;
		float X = (1.f + atan2f(NormalizedDirection.X, - NormalizedDirection.Z) * invPI) * 0.5f * SizeX;
		float Y = acosf(NormalizedDirection.Y)*invPI * SizeY;

		return LookupFiltered(X, Y);
	}
	
	FLinearColor LookupLongLatDouble(const FVector & NormalizedDirection) const
	{
		// this does the math in doubles then stores to floats :
		//	that was probably a mistake, but leave it to avoid patches
		float X = (1 + atan2(NormalizedDirection.X, - NormalizedDirection.Z) / PI) / 2 * SizeX;
		float Y = acos(NormalizedDirection.Y) / PI * SizeY;

		return LookupFiltered(X, Y);
	}
};

// transform world space vector to a space relative to the face
static inline FVector TransformSideToWorldSpace(uint32 CubemapFace, const FVector & InDirection)
{
	float x = InDirection.X, y = InDirection.Y, z = InDirection.Z;

	FVector Ret;

	// see http://msdn.microsoft.com/en-us/library/bb204881(v=vs.85).aspx
	switch(CubemapFace)
	{
		default: checkSlow(0);
		case 0: Ret = FVector(+z, -y, -x); break;
		case 1: Ret = FVector(-z, -y, +x); break;
		case 2: Ret = FVector(+x, +z, +y); break;
		case 3: Ret = FVector(+x, -z, -y); break;
		case 4: Ret = FVector(+x, -y, +z); break;
		case 5: Ret = FVector(-x, -y, -z); break;
	}

	// this makes it with the Unreal way (z and y are flipped)
	return FVector(Ret.X, Ret.Z, Ret.Y);
}

// transform vector relative to the face to world space
static inline FVector TransformWorldToSideSpace(uint32 CubemapFace, const FVector & InDirection)
{
	// undo Unreal way (z and y are flipped)
	float x = InDirection.X, y = InDirection.Z, z = InDirection.Y;

	FVector Ret;

	// see http://msdn.microsoft.com/en-us/library/bb204881(v=vs.85).aspx
	switch(CubemapFace)
	{
		default: checkSlow(0);
		case 0: Ret = FVector(-z, -y, +x); break;
		case 1: Ret = FVector(+z, -y, -x); break;
		case 2: Ret = FVector(+x, +z, +y); break;
		case 3: Ret = FVector(+x, -z, -y); break;
		case 4: Ret = FVector(+x, -y, +z); break;
		case 5: Ret = FVector(-x, -y, -z); break;
	}

	return Ret;
}

static inline FVector ComputeSSCubeDirectionAtTexelCenter(uint32 x, uint32 y, float InvSideExtent)
{
	// center of the texels
	FVector DirectionSS((x + 0.5f) * InvSideExtent * 2 - 1, (y + 0.5f) * InvSideExtent * 2 - 1, 1);
	DirectionSS.Normalize();
	return DirectionSS;
}

static inline FVector ComputeWSCubeDirectionAtTexelCenter(uint32 CubemapFace, uint32 x, uint32 y, float InvSideExtent)
{
	FVector DirectionSS = ComputeSSCubeDirectionAtTexelCenter(x, y, InvSideExtent);
	FVector DirectionWS = TransformSideToWorldSpace(CubemapFace, DirectionSS);
	return DirectionWS;
}

static uint32 ComputeLongLatCubemapExtents(int32 SrcImageSizeX, const uint32 MaxCubemapTextureResolution)
{
	return FMath::Clamp(1U << FMath::FloorLog2(SrcImageSizeX / 2), 32U, MaxCubemapTextureResolution);
}

void ITextureCompressorModule::GenerateBaseCubeMipFromLongitudeLatitude2D(FImage* OutMip, const FImage& SrcImage, const uint32 MaxCubemapTextureResolution, uint8 SourceEncodingOverride)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.CubeMipFromLongLat);

	FImage LongLatImage;
	SrcImage.Linearize(SourceEncodingOverride, LongLatImage);

	// TODO_TEXTURE: Expose target size to user.
	uint32 Extent = ComputeLongLatCubemapExtents(LongLatImage.SizeX, MaxCubemapTextureResolution);
	float InvExtent = 1.0f / Extent;
	OutMip->Init(Extent, Extent, SrcImage.NumSlices * 6, ERawImageFormat::RGBA32F, EGammaSpace::Linear);

	for (int32 Slice = 0; Slice < SrcImage.NumSlices; ++Slice)
	{
		FImageViewLongLat LongLatView(LongLatImage, Slice);

		// Parallel on the 6 faces :
		ParallelFor( TEXT("Texture.CubeMipFromLongLat.PF"),6,1, [&](int32 Face)
		{
			FImageView2D MipView(*OutMip, Slice * 6 + Face);
			for (uint32 y = 0; y < Extent; ++y)
			{
				for (uint32 x = 0; x < Extent; ++x)
				{
					FVector DirectionWS = ComputeWSCubeDirectionAtTexelCenter(Face, x, y, InvExtent);
					MipView.Access(x, y) = LongLatView.LookupLongLatDouble(DirectionWS);
				}
			}
		}, EParallelForFlags::Unbalanced);
	}
}

class FTexelProcessor
{
public:
	// @param InConeAxisSS - normalized, in side space
	// @param TexelAreaArray - precomputed area of each texel for correct weighting
	FTexelProcessor(const FVector& InConeAxisSS, float ConeAngle, const FLinearColor* InSideData, const float* InTexelAreaArray, uint32 InFullExtent)
		: ConeAxisSS(InConeAxisSS)
		, AccumulatedColor(0, 0, 0, 0)
		, SideData(InSideData)
		, TexelAreaArray(InTexelAreaArray)
		, FullExtent(InFullExtent)
	{
		ConeAngleSin = sinf(ConeAngle);
		ConeAngleCos = cosf(ConeAngle);

		// *2 as the position is from -1 to 1
		// / InFullExtent as x and y is in the range 0..InFullExtent-1
		PositionToWorldScale = 2.0f / InFullExtent;
		InvFullExtent = 1.0f / FullExtent;

		// examples: 0 to diffuse convolution, 0.95f for glossy
		DirDot = FMath::Min(FMath::Cos(ConeAngle), 0.9999f);

		InvDirOneMinusDot = 1.0f / (1.0f - DirDot);

		// precomputed sqrt(2.0f * 2.0f + 2.0f * 2.0f)
		float Sqrt8 = 2.8284271f;
		RadiusToWorldScale = Sqrt8 / (float)InFullExtent;
	}

	// @return true: yes, traverse deeper, false: not relevant
	bool TestIfRelevant(uint32 x, uint32 y, uint32 LocalExtent) const
	{
		float HalfExtent = LocalExtent * 0.5f; 
		float U = (x + HalfExtent) * PositionToWorldScale - 1.0f;
		float V = (y + HalfExtent) * PositionToWorldScale - 1.0f;

		float SphereRadius = RadiusToWorldScale * LocalExtent;

		FVector SpherePos(U, V, 1);

		return FMath::SphereConeIntersection(SpherePos, SphereRadius, ConeAxisSS, ConeAngleSin, ConeAngleCos);
	}

	void Process(uint32 x, uint32 y)
	{
		const FLinearColor* In = &SideData[x + y * FullExtent];
		
		FVector DirectionSS = ComputeSSCubeDirectionAtTexelCenter(x, y, InvFullExtent);

		float DotValue = ConeAxisSS | DirectionSS;

		if(DotValue > DirDot)
		{
			// 0..1, 0=at kernel border..1=at kernel center
			float KernelWeight = 1.0f - (1.0f - DotValue) * InvDirOneMinusDot;

			// apply smoothstep function (softer, less linear result)
			KernelWeight = KernelWeight * KernelWeight * (3 - 2 * KernelWeight);

			float AreaCompensation = TexelAreaArray[x + y * FullExtent];
			// AreaCompensation would be need for correctness but seems it has a but
			// as it looks much better (no seam) without, the effect is minor so it's deactivated for now.
//			float Weight = KernelWeight * AreaCompensation;
			float Weight = KernelWeight;

			AccumulatedColor.R += Weight * In->R;
			AccumulatedColor.G += Weight * In->G;
			AccumulatedColor.B += Weight * In->B;
			AccumulatedColor.A += Weight;
		}
	}

	// normalized, in side space
	FVector ConeAxisSS;

	FLinearColor AccumulatedColor;

	// cached for better performance
	float ConeAngleSin;
	float ConeAngleCos;
	float PositionToWorldScale;
	float RadiusToWorldScale;
	float InvFullExtent;
	// 0 to diffuse convolution, 0.95f for glossy
	float DirDot;
	float InvDirOneMinusDot;

	/** [x + y * FullExtent] */
	const FLinearColor* SideData;
	const float* TexelAreaArray;
	uint32 FullExtent;
};

template <class TVisitor>
void TCubemapSideRasterizer(TVisitor &TexelProcessor, int32 x, uint32 y, uint32 Extent)
{
	if(Extent > 1)
	{
		if(!TexelProcessor.TestIfRelevant(x, y, Extent))
		{
			return;
		}
		Extent /= 2;

		TCubemapSideRasterizer(TexelProcessor, x, y, Extent);
		TCubemapSideRasterizer(TexelProcessor, x + Extent, y, Extent);
		TCubemapSideRasterizer(TexelProcessor, x, y + Extent, Extent);
		TCubemapSideRasterizer(TexelProcessor, x + Extent, y + Extent, Extent);
	}
	else
	{
		TexelProcessor.Process(x, y);
	}
}

static FLinearColor IntegrateAngularArea(FImage& Image, FVector FilterDirectionWS, float ConeAngle, const float* TexelAreaArray)
{
	// Alpha channel is used to renormalize later
	FLinearColor ret(0, 0, 0, 0);
	int32 Extent = Image.SizeX;

	for(uint32 Face = 0; Face < 6; ++Face)
	{
		FImageView2D ImageView(Image, Face);
		FVector FilterDirectionSS = TransformWorldToSideSpace(Face, FilterDirectionWS);
		FTexelProcessor Processor(FilterDirectionSS, ConeAngle, &ImageView.Access(0,0), TexelAreaArray, Extent);

		// recursively split the (0,0)-(Extent-1,Extent-1), tests for intersection and processes only colors inside
		TCubemapSideRasterizer(Processor, 0, 0, Extent);
		ret += Processor.AccumulatedColor;
	}
	
	if(ret.A != 0)
	{
		float Inv = 1.0f / ret.A;

		ret.R *= Inv;
		ret.G *= Inv;
		ret.B *= Inv;
	}
	else
	{
		// should not happen
//		checkSlow(0);
	}

	ret.A = 0;

	return ret;
}

// @return 2 * computed triangle area 
static inline float TriangleArea2_3D(FVector A, FVector B, FVector C)
{
	return ((A-B) ^ (C-B)).Size();
}

static inline float ComputeTexelArea(uint32 x, uint32 y, float InvSideExtentMul2)
{
	float fU = x * InvSideExtentMul2 - 1;
	float fV = y * InvSideExtentMul2 - 1;

	FVector CornerA = FVector(fU, fV, 1);
	FVector CornerB = FVector(fU + InvSideExtentMul2, fV, 1);
	FVector CornerC = FVector(fU, fV + InvSideExtentMul2, 1);
	FVector CornerD = FVector(fU + InvSideExtentMul2, fV + InvSideExtentMul2, 1);

	CornerA.Normalize();
	CornerB.Normalize();
	CornerC.Normalize();
	CornerD.Normalize();

	return TriangleArea2_3D(CornerA, CornerB, CornerC) + TriangleArea2_3D(CornerC, CornerB, CornerD) * 0.5f;
}

/**
 * Generate a mip using angular filtering.
 * @param DestMip - The filtered mip.
 * @param SrcMip - The source mip which will be filtered.
 * @param ConeAngle - The cone angle with which to filter.
 */
static void GenerateAngularFilteredMip(FImage* DestMip, FImage& SrcMip, float ConeAngle)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateAngularFilteredMip);

	int32 MipExtent = DestMip->SizeX;
	float MipInvSideExtent = 1.0f / MipExtent;

	TArray<float> TexelAreaArray;
	TexelAreaArray.AddUninitialized(SrcMip.SizeX * SrcMip.SizeY);

	// precompute the area size for one face (is the same for each face)
	for(int32 y = 0; y < SrcMip.SizeY; ++y)
	{
		for(int32 x = 0; x < SrcMip.SizeX; ++x)
		{
			TexelAreaArray[x + y * (int64) SrcMip.SizeX] = ComputeTexelArea(x, y, MipInvSideExtent * 2);
		}
	}

	// We start getting gains running threaded upwards of sizes >= 128
	if (SrcMip.SizeX >= 128)
	{
		// Quick workaround: Do a thread per mip
		struct FAsyncGenerateMipsPerFaceWorker : public FNonAbandonableTask
		{
			int32 Face;
			FImage* DestMip;
			int32 Extent;
			float ConeAngle;
			const float* TexelAreaArray;
			FImage* SrcMip;
			FAsyncGenerateMipsPerFaceWorker(int32 InFace, FImage* InDestMip, int32 InExtent, float InConeAngle, const float* InTexelAreaArray, FImage* InSrcMip) :
				Face(InFace),
				DestMip(InDestMip),
				Extent(InExtent),
				ConeAngle(InConeAngle),
				TexelAreaArray(InTexelAreaArray),
				SrcMip(InSrcMip)
			{
			}

			void DoWork()
			{
				const float InvSideExtent = 1.0f / Extent;
				FImageView2D DestMipView(*DestMip, Face);
				for (int32 y = 0; y < Extent; ++y)
				{
					for (int32 x = 0; x < Extent; ++x)
					{
						FVector DirectionWS = ComputeWSCubeDirectionAtTexelCenter(Face, x, y, InvSideExtent);
						DestMipView.Access(x, y) = IntegrateAngularArea(*SrcMip, DirectionWS, ConeAngle, TexelAreaArray);
					}
				}
			}

			FORCEINLINE TStatId GetStatId() const
			{
				RETURN_QUICK_DECLARE_CYCLE_STAT(FAsyncGenerateMipsPerFaceWorker, STATGROUP_ThreadPoolAsyncTasks);
			}
		};

		typedef FAsyncTask<FAsyncGenerateMipsPerFaceWorker> FAsyncGenerateMipsPerFaceTask;
		TIndirectArray<FAsyncGenerateMipsPerFaceTask> AsyncTasks;

		for (int32 Face = 0; Face < 6; ++Face)
		{
			auto* AsyncTask = new FAsyncGenerateMipsPerFaceTask(Face, DestMip, MipExtent, ConeAngle, TexelAreaArray.GetData(), &SrcMip);
			AsyncTasks.Add(AsyncTask);
			AsyncTask->StartBackgroundTask();
		}

		for (int32 TaskIndex = 0; TaskIndex < AsyncTasks.Num(); ++TaskIndex)
		{
			auto& AsyncTask = AsyncTasks[TaskIndex];
			AsyncTask.EnsureCompletion();
		}
	}
	else
	{
		for (int32 Face = 0; Face < 6; ++Face)
		{
			FImageView2D DestMipView(*DestMip, Face);
			for (int32 y = 0; y < MipExtent; ++y)
			{
				for (int32 x = 0; x < MipExtent; ++x)
				{
					FVector DirectionWS = ComputeWSCubeDirectionAtTexelCenter(Face, x, y, MipInvSideExtent);
					DestMipView.Access(x, y) = IntegrateAngularArea(SrcMip, DirectionWS, ConeAngle, TexelAreaArray.GetData());
				}
			}
		}
	}
}

void ITextureCompressorModule::GenerateAngularFilteredMips(TArray<FImage>& InOutMipChain, int32 NumMips, uint32 DiffuseConvolveMipLevel)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.GenerateAngularFilteredMips);

	TArray<FImage> SrcMipChain;
	Exchange(SrcMipChain, InOutMipChain);
	InOutMipChain.Empty(NumMips);

	// note: should work on cube arrays but currently does not
	// GetMipGenSettings forces Angular off for anything but pure Cube classes (no arrays)
	check( SrcMipChain[0].NumSlices == 6 );

	// Generate simple averaged mips to accelerate angular filtering.
	for (int32 MipIndex = SrcMipChain.Num(); MipIndex < NumMips; ++MipIndex)
	{
		FImage& BaseMip = SrcMipChain[MipIndex - 1];
		int32 BaseExtent = BaseMip.SizeX;
		int32 MipExtent = FMath::Max(BaseExtent >> 1, 1);
		FImage* Mip = new(SrcMipChain) FImage(MipExtent, MipExtent, BaseMip.NumSlices, BaseMip.Format);

		for(int32 Face = 0; Face < 6; ++Face)
		{
			FImageView2D BaseMipView(BaseMip, Face);
			FImageView2D MipView(*Mip, Face);

			for(int32 y = 0; y < MipExtent; ++y)
			{
				for(int32 x = 0; x < MipExtent; ++x)
				{		
					FLinearColor Sum = (
						BaseMipView.Access(x*2, y*2) +
						BaseMipView.Access(x*2+1, y*2) +
						BaseMipView.Access(x*2, y*2+1) +
						BaseMipView.Access(x*2+1, y*2+1)
						) * 0.25f;
					MipView.Access(x,y) = Sum;
				}
			}
		}
	}

	int32 Extent = 1 << (NumMips - 1);
	int32 BaseExtent = Extent;
	for (int32 i = 0; i < NumMips; ++i)
	{
		// 0:top mip 1:lowest mip = diffuse convolve
		float NormalizedMipLevel = i / (float)(NumMips - DiffuseConvolveMipLevel);
		float AdjustedMipLevel = NormalizedMipLevel * NumMips;
		float NormalizedWidth = BaseExtent * FMath::Pow(2.0f, -AdjustedMipLevel);
		float TexelSize = 1.0f / NormalizedWidth;

		// 0.001f:sharp  .. PI/2: diffuse convolve
		// all lower mips are used for diffuse convolve
		// above that the angle blends from sharp to diffuse convolved version
		float ConeAngle = PI / 2.0f * TexelSize;

		// restrict to reasonable range
		ConeAngle = FMath::Clamp(ConeAngle, 0.002f, (float)PI / 2.0f);

		UE_LOG(LogTextureCompressor, Verbose, TEXT("GenerateAngularFilteredMips  %f %f %f %f %f"), NormalizedMipLevel, AdjustedMipLevel, NormalizedWidth, TexelSize, ConeAngle * 180 / PI);

		// 0:normal, -1:4x faster, +1:4 times slower but more precise, -2, 2 ...
		float QualityBias = 3.0f;

		// defined to result in a area of 1.0f (NormalizedArea)
		// optimized = 0.5f * FMath::Sqrt(1.0f / PI);
		float SphereRadius = 0.28209478f;
		float SegmentHeight = SphereRadius * (1.0f - FMath::Cos(ConeAngle));
		// compute SphereSegmentArea
		float AreaCoveredInNormalizedArea = 2 * PI * SphereRadius * SegmentHeight;
		checkSlow(AreaCoveredInNormalizedArea <= 0.5f);

		// unoptimized
		//	float FloatInputMip = FMath::Log2(FMath::Sqrt(AreaCoveredInNormalizedArea)) + InputMipCount - QualityBias;
		// optimized
		float FloatInputMip = 0.5f * FMath::Log2(AreaCoveredInNormalizedArea) + NumMips - QualityBias;
		uint32 InputMip = FMath::Clamp(FMath::TruncToInt(FloatInputMip), 0, NumMips - 1);

		FImage* Mip = new(InOutMipChain) FImage(Extent, Extent, 6, ERawImageFormat::RGBA32F);
		GenerateAngularFilteredMip(Mip, SrcMipChain[InputMip], ConeAngle);
		Extent = FMath::Max(Extent >> 1, 1);
	}
}

static bool NeedAdjustImageColors(const FTextureBuildSettings& InBuildSettings)
{
	const FColorAdjustmentParameters& InParams = InBuildSettings.ColorAdjustment;

	return
		!FMath::IsNearlyEqual(InParams.AdjustBrightness, 1.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustBrightnessCurve, 1.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustSaturation, 1.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustVibrance, 0.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustRGBCurve, 1.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustHue, 0.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustMinAlpha, 0.0f, (float)KINDA_SMALL_NUMBER) ||
		!FMath::IsNearlyEqual(InParams.AdjustMaxAlpha, 1.0f, (float)KINDA_SMALL_NUMBER) ||
		InBuildSettings.bChromaKeyTexture;
}

static inline void AdjustColorsOld(FLinearColor * Colors,int64 Count, const FTextureBuildSettings& InBuildSettings)
{
	const FColorAdjustmentParameters& Params = InBuildSettings.ColorAdjustment;
	
	// very similar to AdjustColorsNew
	// issues in here are mostly fixed in AdjustColorsNew

	// note: : not the same checks as bAdjustNeeded outside
	//	 but preserves legacy behavior
	bool bAdjustBrightnessCurve = (!FMath::IsNearlyEqual(Params.AdjustBrightnessCurve, 1.0f, (float)KINDA_SMALL_NUMBER) && Params.AdjustBrightnessCurve != 0.0f);
	bool bAdjustVibrance = (!FMath::IsNearlyZero(Params.AdjustVibrance, (float)KINDA_SMALL_NUMBER));
	bool bAdjustRGBCurve = (!FMath::IsNearlyEqual(Params.AdjustRGBCurve, 1.0f, (float)KINDA_SMALL_NUMBER) && Params.AdjustRGBCurve != 0.0f);
	bool bAdjustSaturation = ( Params.AdjustSaturation != 1.f || bAdjustVibrance );
	bool bAdjustValue = ( Params.AdjustBrightness != 1.f ) || bAdjustBrightnessCurve;

	float AdjustHue = Params.AdjustHue;
	if ( AdjustHue != 0.f )
	{
		// Params.AdjustHue should be in [0,360] , make sure
		if ( AdjustHue < 0.f || AdjustHue > 360.f )
		{
			AdjustHue = fmodf(AdjustHue, 360.0f);
			if ( AdjustHue < 0.f )
			{
				AdjustHue += 360.f;
			}
		}
	}
	
	// BuildSettings.ChromaKeyColor is an FColor
	FLinearColor ChromaKeyColor(InBuildSettings.ChromaKeyColor);

	bool bHDRSource = InBuildSettings.bHDRSource;
	bool bChromaKeyTexture = InBuildSettings.bChromaKeyTexture;
	float ChromaKeyThreshold = InBuildSettings.ChromaKeyThreshold + SMALL_NUMBER;

	for(int64 i=0;i<Count;i++)
	{
		FLinearColor OriginalColor = Colors[i];
	
		// note: non-HDR source data in [0,1] can drift outside of [0,1]
		//   and then bad things will happen here
		//  left alone to preserve old behavior

		if (bChromaKeyTexture && (OriginalColor.Equals(ChromaKeyColor, ChromaKeyThreshold)))
		{
			OriginalColor = FLinearColor::Transparent;

			//note: strange: no return.  processing continues on the transparent color...
			//	  this was likely unintentional
			//Colors[i] = FLinearColor::Transparent;
			//continue;
		}

		// NOTE: if OriginalColor has HDR/floats in it, this function does not handle it well
		//	it implicitly discards negatives (and negatives cause color shifts)
		//	for values > 1 the clamp behavior is very strange

		// Convert to HSV
		FLinearColor HSVColor = OriginalColor.LinearRGBToHSV();
		float& PixelHue = HSVColor.R;
		float& PixelSaturation = HSVColor.G;
		float& PixelValue = HSVColor.B;

		float OriginalLuminance = PixelValue;

		if ( bAdjustValue )
		{
			// Apply brightness adjustment
			PixelValue *= Params.AdjustBrightness;

			// Apply brightness power adjustment
			if ( bAdjustBrightnessCurve )
			{
				// Raise HSV.V to the specified power
				PixelValue = FMath::Pow(PixelValue, Params.AdjustBrightnessCurve);
			}
		
			// Clamp brightness if non-HDR
			if (!bHDRSource)
			{
				PixelValue = FMath::Clamp(PixelValue, 0.0f, 1.0f);
			}
		}
		
		if ( bAdjustSaturation )
		{
			// PixelSaturation is >= 0 but not <= 1
			//  because negative RGB can come into this function which gives Saturation > 1

			// Apply "vibrance" adjustment
			if ( bAdjustVibrance )
			{
				// note: AdjustVibrance is disabled for HDR source in the Texture UPROPERTIES
				//    (unclear why, this is no worse than anything else here on HDR)
				const float SatRaisePow = 5.0f;
				const float InvSatRaised = FMath::Pow(1.0f - PixelSaturation, SatRaisePow);

				const float ClampedVibrance = FMath::Clamp(Params.AdjustVibrance, 0.0f, 1.0f);
				const float HalfVibrance = ClampedVibrance * 0.5f;

				const float SatProduct = HalfVibrance * InvSatRaised;

				PixelSaturation += SatProduct;
			}

			// Apply saturation adjustment
			PixelSaturation *= Params.AdjustSaturation;
			PixelSaturation = FMath::Clamp(PixelSaturation, 0.0f, 1.0f);
		}

		// Apply hue adjustment
		if ( AdjustHue != 0.f )
		{
			// PixelHue is [0,360) but AdjustHue is [0,360]
			PixelHue += AdjustHue;

			// Clamp HSV values
			if ( PixelHue >= 360.f )
			{
				PixelHue -= 360.f;
			}
		}

		// Convert back to a linear color
		FLinearColor LinearColor = HSVColor.HSVToLinearRGB();

		// Apply RGB curve adjustment (linear space)
		if ( bAdjustRGBCurve )
		{
			LinearColor.R = FMath::Pow(LinearColor.R, Params.AdjustRGBCurve);
			LinearColor.G = FMath::Pow(LinearColor.G, Params.AdjustRGBCurve);
			LinearColor.B = FMath::Pow(LinearColor.B, Params.AdjustRGBCurve);
		}

		// Clamp HDR RGB channels to 1 or the original luminance (max original RGB channel value), whichever is greater
		// note: this is a very odd thing to do
		//		clamping at OriginalLuminance if you do AdjustBrightness or AdjustBrightnessCurve ?
		//	    that would keep values brighter than 1.f unchanged, but bring up lower ones to 1.f
		if (bHDRSource)
		{
			LinearColor.R = FMath::Clamp(LinearColor.R, 0.0f, (OriginalLuminance > 1.0f ? OriginalLuminance : 1.0f));
			LinearColor.G = FMath::Clamp(LinearColor.G, 0.0f, (OriginalLuminance > 1.0f ? OriginalLuminance : 1.0f));
			LinearColor.B = FMath::Clamp(LinearColor.B, 0.0f, (OriginalLuminance > 1.0f ? OriginalLuminance : 1.0f));
		}

		// Remap the alpha channel
		LinearColor.A = FMath::Lerp(Params.AdjustMinAlpha, Params.AdjustMaxAlpha, OriginalColor.A);

		Colors[i] = LinearColor;
	}
}

// see also AdjustColorsOld
static inline void AdjustColorsNew(FLinearColor* Colors, int64 Count, const FTextureBuildSettings& InBuildSettings)
{
	const FColorAdjustmentParameters& Params = InBuildSettings.ColorAdjustment;

	// @todo Oodle : not the same checks as bAdjustNeeded outside
	//	 but preserves legacy behavior
	bool bAdjustBrightnessCurve = (!FMath::IsNearlyEqual(Params.AdjustBrightnessCurve, 1.0f, (float)KINDA_SMALL_NUMBER) && Params.AdjustBrightnessCurve != 0.0f);
	bool bAdjustVibrance = (!FMath::IsNearlyZero(Params.AdjustVibrance, (float)KINDA_SMALL_NUMBER));
	bool bAdjustRGBCurve = (!FMath::IsNearlyEqual(Params.AdjustRGBCurve, 1.0f, (float)KINDA_SMALL_NUMBER) && Params.AdjustRGBCurve != 0.0f);
	bool bAdjustSaturation = ( Params.AdjustSaturation != 1.f || bAdjustVibrance );
	bool bAdjustValue = ( Params.AdjustBrightness != 1.f ) || bAdjustBrightnessCurve;

	float AdjustHue = Params.AdjustHue;
	if ( AdjustHue != 0.f )
	{
		// Params.AdjustHue should be in [0,360] , make sure
		if ( AdjustHue < 0.f || AdjustHue > 360.f )
		{
			AdjustHue = fmodf(AdjustHue, 360.0f);
			if ( AdjustHue < 0.f )
			{
				AdjustHue += 360.f;
			}
		}
	}
	
	// BuildSettings.ChromaKeyColor is an FColor
	FLinearColor ChromaKeyColor(InBuildSettings.ChromaKeyColor);
	bool bChromaKeyTexture = InBuildSettings.bChromaKeyTexture;
	float ChromaKeyThreshold = InBuildSettings.ChromaKeyThreshold + SMALL_NUMBER;

	bool bHDRSource = InBuildSettings.bHDRSource;

	for (int64 i=0; i<Count; i++)
	{
		if (bChromaKeyTexture && (Colors[i].Equals(ChromaKeyColor, ChromaKeyThreshold)))
		{
			Colors[i] = FLinearColor::Transparent;
			continue;
		}

		FLinearColor OriginalColor = Colors[i];
			
		if (!bHDRSource)
		{
			// Ensure we are clamped as expected (can drift out of clamp due to previous processing)
			// if you wind up even very slightly out of [0,1] range this function does bad things
			OriginalColor.R = FMath::Clamp(OriginalColor.R, 0.0f, 1.f);
			OriginalColor.G = FMath::Clamp(OriginalColor.G, 0.0f, 1.f);
			OriginalColor.B = FMath::Clamp(OriginalColor.B, 0.0f, 1.f);
		}
		else
		{
			/*
			if ( OriginalColor.R < 0 || OriginalColor.G < 0 || OriginalColor.B < 0 )
			{
				// need to log texture name for this to be of any use :
				UE_CALL_ONCE( [&](){
					UE_LOG(LogTextureCompressor, Warning,
						TEXT("Negative pixel values (%f, %f, %f) are not expected"),
						OriginalColor.R, OriginalColor.G, OriginalColor.B);
				} );
			}
			*/
			
			// yes HDR source, but we don't support negatives in Adjust
			// clamp to >= 0 :
			OriginalColor.R = FMath::Max(OriginalColor.R, 0.0f);
			OriginalColor.G = FMath::Max(OriginalColor.G, 0.0f);
			OriginalColor.B = FMath::Max(OriginalColor.B, 0.0f);
		}

		// Convert to HSV
		FLinearColor HSVColor = OriginalColor.LinearRGBToHSV();
		float& PixelHue = HSVColor.R;
		float& PixelSaturation = HSVColor.G;
		float& PixelValue = HSVColor.B;

		if ( bAdjustValue )
		{
			// Apply brightness adjustment
			PixelValue *= Params.AdjustBrightness;

			// Apply brightness power adjustment
			if ( bAdjustBrightnessCurve )
			{
				// Raise HSV.V to the specified power
				PixelValue = FMath::Pow(PixelValue, Params.AdjustBrightnessCurve);
			}
		
			// Clamp brightness if non-HDR
			if (!bHDRSource)
			{
				PixelValue = FMath::Clamp(PixelValue, 0.0f, 1.0f);
			}
		}
		
		if ( bAdjustSaturation )
		{
			// PixelSaturation is >= 0 but not <= 1
			//  because negative RGB can come into this function which gives Saturation > 1

			// Apply "vibrance" adjustment
			if ( bAdjustVibrance )
			{
				const float InvSat = 1.0f - PixelSaturation;
				const float InvSatRaised = InvSat * InvSat * InvSat * InvSat * InvSat;

				const float ClampedVibrance = FMath::Clamp(Params.AdjustVibrance, 0.0f, 1.0f);
				const float HalfVibrance = ClampedVibrance * 0.5f;

				const float SatProduct = HalfVibrance * InvSatRaised;

				PixelSaturation += SatProduct;
			}

			// Apply saturation adjustment
			PixelSaturation *= Params.AdjustSaturation;
			PixelSaturation = FMath::Clamp(PixelSaturation, 0.0f, 1.0f);
		}

		// Apply hue adjustment
		if ( AdjustHue != 0.f )
		{
			// PixelHue is [0,360) but AdjustHue is [0,360]
			PixelHue += AdjustHue;

			// Clamp HSV values
			if ( PixelHue >= 360.f )
			{
				PixelHue -= 360.f;
			}
		}

		// Convert back to a linear color
		FLinearColor LinearColor = HSVColor.HSVToLinearRGB();

		// Apply RGB curve adjustment (linear space)
		if ( bAdjustRGBCurve )
		{
			LinearColor.R = FMath::Pow(LinearColor.R, Params.AdjustRGBCurve);
			LinearColor.G = FMath::Pow(LinearColor.G, Params.AdjustRGBCurve);
			LinearColor.B = FMath::Pow(LinearColor.B, Params.AdjustRGBCurve);
		}

		// Remap the alpha channel
		LinearColor.A = FMath::Lerp(Params.AdjustMinAlpha, Params.AdjustMaxAlpha, FMath::Clamp(OriginalColor.A, 0.f, 1.f));

		Colors[i] = LinearColor;
	}
}

void ITextureCompressorModule::AdjustImageColors(FImage& Image, const FTextureBuildSettings& InBuildSettings)
{
	const FColorAdjustmentParameters& InParams = InBuildSettings.ColorAdjustment;
	check( Image.SizeX > 0 && Image.SizeY > 0 );

	// @todo Oodle : this bAdjustNeeded is not checking the same conditions to enable these adjustments
	//		as is used inside the AdjustColors() routine
	//		this is how it was done in the past, so keep it the same to preserve legacy operation
	//	if possible in the future factor this Needed check out so it is shared code

	bool bAdjustNeeded = NeedAdjustImageColors(InBuildSettings);

	if ( bAdjustNeeded )
	{
		TRACE_CPUPROFILER_EVENT_SCOPE(Texture.AdjustImageColors);

		const int64 NumPixels = (int64)Image.SizeX * Image.SizeY * Image.NumSlices;
		TArrayView64<FLinearColor> ImageColors = Image.AsRGBA32F();

		int64 NumPixelsEachJob;
		int32 NumJobs = ImageParallelForComputeNumJobsForPixels(NumPixelsEachJob,NumPixels);

		ParallelFor( TEXT("Texture.AdjustImageColorsFunc.PF"),NumJobs,1, [&](int32 Index)
		{
			int64 StartIndex = Index * NumPixelsEachJob;
			int64 EndIndex = FMath::Min(StartIndex + NumPixelsEachJob, NumPixels);

			FLinearColor * First = &ImageColors[StartIndex];
			int64 Count = EndIndex-StartIndex;

			// Use new AdjustColors code only for newly added textures
			// So existing textures maintain exactly same output as before
			if (InBuildSettings.bUseNewMipFilter)
			{
				AdjustColorsNew(First, Count, InBuildSettings);
			}
			else
			{
				AdjustColorsOld(First, Count, InBuildSettings);
			}
		}, EParallelForFlags::Unbalanced);
	}
}

/**
 * Compute the alpha channel how BokehDOF needs it setup
 *
 * @param	Image	Image to adjust
 */
static void ComputeBokehAlpha(FImage& Image)
{
	check( Image.SizeX > 0 && Image.SizeY > 0 );

	const int64 NumPixels = Image.GetNumPixels();
	TArrayView64<FLinearColor> ImageColors = Image.AsRGBA32F();

	// compute LinearAverage
	FLinearColor LinearAverage;
	{
		FLinearColor LinearSum(0, 0, 0, 0);
		for( int64 CurPixelIndex = 0; CurPixelIndex < NumPixels; ++CurPixelIndex )
		{
			LinearSum += ImageColors[ CurPixelIndex ];
		}
		LinearAverage = LinearSum / (float)NumPixels;
	}

	FLinearColor Scale(1, 1, 1, 1);

	// we want to normalize the image to have 0.5 as average luminance, this is assuming clamping doesn't happen (can happen when using a very small Bokeh shape)
	{
		float RGBLum = (LinearAverage.R + LinearAverage.G + LinearAverage.B) / 3.0f;

		// ideally this would be 1 but then some pixels would need to be >1 which is not supported for the textureformat we want to use.
		// The value affects the occlusion computation of the BokehDOF
		const float LumGoal = 0.25f;

		// clamp to avoid division by 0
		Scale *= LumGoal / FMath::Max(RGBLum, 0.001f);
	}

	{
		for( int64 CurPixelIndex = 0; CurPixelIndex < NumPixels; ++CurPixelIndex )
		{
			const FLinearColor OriginalColor = ImageColors[ CurPixelIndex ];

			// Convert to a linear color
			FLinearColor LinearColor = OriginalColor * Scale;
			float RGBLum = (LinearColor.R + LinearColor.G + LinearColor.B) / 3.0f;
			LinearColor.A = FMath::Clamp(RGBLum, 0.0f, 1.0f);
			ImageColors[ CurPixelIndex ] = LinearColor;
		}
	}
}

/**
 * Replicates the contents of the red channel to the green, blue, and alpha channels.
 */
static void ReplicateRedChannel( TArray<FImage>& InOutMipChain )
{
	const uint32 MipCount = InOutMipChain.Num();
	for ( uint32 MipIndex = 0; MipIndex < MipCount; ++MipIndex )
	{
		FImage& SrcMip = InOutMipChain[MipIndex];
		FLinearColor* FirstColor = (&SrcMip.AsRGBA32F()[0]);
		FLinearColor* LastColor = FirstColor + SrcMip.GetNumPixels();
		for ( FLinearColor* Color = FirstColor; Color < LastColor; ++Color )
		{
			*Color = FLinearColor( Color->R, Color->R, Color->R, Color->R );
		}
	}
}

/**
 * Replicates the contents of the alpha channel to the red, green, and blue channels.
 */
static void ReplicateAlphaChannel( TArray<FImage>& InOutMipChain )
{
	const uint32 MipCount = InOutMipChain.Num();
	for ( uint32 MipIndex = 0; MipIndex < MipCount; ++MipIndex )
	{
		FImage& SrcMip = InOutMipChain[MipIndex];
		FLinearColor* FirstColor = (&SrcMip.AsRGBA32F()[0]);
		FLinearColor* LastColor = FirstColor + SrcMip.GetNumPixels();
		for ( FLinearColor* Color = FirstColor; Color < LastColor; ++Color )
		{
			*Color = FLinearColor( Color->A, Color->A, Color->A, Color->A );
		}
	}
}

/**
 * Flips the contents of the green channel.
 * @param InOutMipChain - The mip chain on which the green channel shall be flipped.
 */
static void FlipGreenChannel( FImage& Image )
{
	FLinearColor* FirstColor = (&Image.AsRGBA32F()[0]);
	FLinearColor* LastColor = FirstColor + (Image.SizeX * Image.SizeY * Image.NumSlices);
	for ( FLinearColor* Color = FirstColor; Color < LastColor; ++Color )
	{
		Color->G = 1.0f - FMath::Clamp(Color->G, 0.0f, 1.0f);
	}
}

/** Calculate a scale per 4x4 block of each image, and apply it to the red/green channels. Store scale in the blue channel. */
static void ApplyYCoCgBlockScale(TArray<FImage>& InOutMipChain)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.ApplyYCoCgBlockScale);

	const uint32 MipCount = InOutMipChain.Num();
	for (uint32 MipIndex = 0; MipIndex < MipCount; ++MipIndex)
	{
		FImage& SrcMip = InOutMipChain[MipIndex];
		FLinearColor* FirstColor = (&SrcMip.AsRGBA32F()[0]);

		int32 BlockWidthX = SrcMip.SizeX / 4;
		int32 BlockWidthY = SrcMip.SizeY / 4;

		for (int32 Slice = 0; Slice < SrcMip.NumSlices; ++Slice)
		{
			FLinearColor* SliceFirstColor = FirstColor + SrcMip.GetSliceNumPixels() * Slice;

			for (int32 Y = 0; Y < BlockWidthY; ++Y)
			{
				FLinearColor* RowFirstColor = SliceFirstColor + Y * 4 * (int64) SrcMip.SizeX;

				for (int32 X = 0; X < BlockWidthX; ++X)
				{
					FLinearColor* BlockFirstColor = RowFirstColor + (X * 4);

					// Iterate block to find MaxComponent
					float MaxComponent = 0.f;
					for (int32 BlockY = 0; BlockY < 4; ++BlockY)
					{
						FLinearColor* Color = BlockFirstColor + BlockY * SrcMip.SizeX;
						for (int32 BlockX = 0; BlockX < 4; ++BlockX, ++Color)
						{
							MaxComponent = FMath::Max(FMath::Abs(Color->R - 128.f / 255.f), MaxComponent);
							MaxComponent = FMath::Max(FMath::Abs(Color->G - 128.f / 255.f), MaxComponent);
						}
					}

					const float Scale = (MaxComponent < 32.f / 255.f) ? 4.f : (MaxComponent < 64.f / 255.f) ? 2.f : 1.f;
					const float OutB = (Scale - 1.f) * 8.f / 255.f;

					// Iterate block to modify for scale
					for (int32 BlockY = 0; BlockY < 4; ++BlockY)
					{
						FLinearColor* Color = BlockFirstColor + BlockY * SrcMip.SizeX;
						for (int32 BlockX = 0; BlockX < 4; ++BlockX, ++Color)
						{
							const float OutR = (Color->R - 128.f / 255.f) * Scale + 128.f / 255.f;
							const float OutG = (Color->G - 128.f / 255.f) * Scale + 128.f / 255.f;

							*Color = FLinearColor(OutR, OutG, OutB, Color->A);
						}
					}
				}
			}
		}
	}
}

#if 0
static float RoughnessToSpecularPower(float Roughness)
{
	float Div = FMath::Pow(Roughness, 4);

	// Roughness of 0 should result in a high specular power
	float MaxSpecPower = 10000000000.0f;
	Div = FMath::Max(Div, 2.0f / (MaxSpecPower + 2.0f));

	return 2.0f / Div - 2.0f;
}

static float SpecularPowerToRoughness(float SpecularPower)
{
	float Out = FMath::Pow( SpecularPower * 0.5f + 1.0f, -0.25f );

	return Out;
}
#endif

static float CompositeNormalLengthToRoughness(const float LengthN, float Roughness, float CompositePower)
{
	float Variance = ( 1.0f - LengthN ) / LengthN;
	Variance = Variance - 0.00004f;
	if ( Variance <= 0.f )
	{
		return Roughness;
	}

	Variance *= CompositePower;
		
#if 0
	float Power = RoughnessToSpecularPower( Roughness );
	Power = Power / ( 1.0f + Variance * Power );
	Roughness = SpecularPowerToRoughness( Power );
#else
	// Refactored above to avoid divide by zero
	float a = Roughness * Roughness;
	float a2 = a * a;
	float B = 2.0f * Variance * (a2 - 1.0f);
	a2 = ( B - a2 ) / ( B - 1.0f );
	Roughness = FMath::Pow( a2, 0.25f );
#endif

	return Roughness;
}

static float CompositeNormalToRoughness(const FLinearColor & NormalColor, float Roughness, float CompositePower)
{
	FVector3f Normal = FVector3f(NormalColor.R * 2.0f - 1.0f, NormalColor.G * 2.0f - 1.0f, NormalColor.B * 2.0f - 1.0f);
				
	// Toksvig estimation of variance
	float LengthN = FMath::Min( Normal.Size(), 1.0f );

	return CompositeNormalLengthToRoughness(LengthN,Roughness,CompositePower);
}

// @param CompositeTextureMode original type ECompositeTextureMode
//	write roughness to specified channel of DestRoughness, computed from SourceNormals
static bool ApplyCompositeTexture(FImage& DestRoughness, const FImage& SourceNormals, uint8 CompositeTextureMode, float CompositePower)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.ApplyCompositeTexture);

	FLinearColor* FirstColor = (&DestRoughness.AsRGBA32F()[0]);
		
	float* TargetValuePtr;

	switch((ECompositeTextureMode)CompositeTextureMode)
	{
		case CTM_NormalRoughnessToRed:
			TargetValuePtr = &FirstColor->R;
			break;
		case CTM_NormalRoughnessToGreen:
			TargetValuePtr = &FirstColor->G;
			break;
		case CTM_NormalRoughnessToBlue:
			TargetValuePtr = &FirstColor->B;
			break;
		case CTM_NormalRoughnessToAlpha:
			TargetValuePtr = &FirstColor->A;
			break;
		default:
			UE_LOG(LogTextureCompressor, Error, TEXT("Invalid CompositeTextureMode"));
			return false;
	}
	
	if ( DestRoughness.SizeX == SourceNormals.SizeX && DestRoughness.SizeY == SourceNormals.SizeY &&
		DestRoughness.NumSlices == SourceNormals.NumSlices )
	{
		int64 Count = (int64) DestRoughness.SizeX * DestRoughness.SizeY * DestRoughness.NumSlices;

		const FLinearColor* NormalColors = (&SourceNormals.AsRGBA32F()[0]);

		for ( int64 i=0; i<Count; i++ )
		{
			const FLinearColor & NormalColor = NormalColors[i];

			float Roughness = TargetValuePtr[i*4];
			TargetValuePtr[i*4] = CompositeNormalToRoughness(NormalColor,Roughness,CompositePower);
		}
	}
	else
	{
		// sizes don't match, stretch with bilinear filter (filter the normal lengths, not the normal colors)

		if ( SourceNormals.NumSlices != 1 ||
			 DestRoughness.NumSlices != 1 )
		{
			UE_LOG(LogTextureCompressor, Warning, TEXT("CompositeTexture XY sizes don't match, stretch not supported because slice count is not 1 : %d vs %d"),
				SourceNormals.NumSlices,DestRoughness.NumSlices);
			return false;
		}

		TArray<float> SourceNormalLengths;

		{
			int64 SourceNormalCount = (int64) SourceNormals.SizeX * SourceNormals.SizeY;
			SourceNormalLengths.SetNum(SourceNormalCount);

			const FLinearColor* NormalColors = (&SourceNormals.AsRGBA32F()[0]);

			for ( int64 i=0; i<SourceNormalCount; i++ )
			{
				const FLinearColor & NormalColor = NormalColors[i];
				
				FVector3f Normal = FVector3f(NormalColor.R * 2.0f - 1.0f, NormalColor.G * 2.0f - 1.0f, NormalColor.B * 2.0f - 1.0f);
				
				float LengthN = FMath::Min( Normal.Size(), 1.0f );

				SourceNormalLengths[i] = LengthN;
			}
		}

		//const FImageView2D NormalColors(const_cast<FImage &>(SourceNormals),0);
		const float * SourceNormalLengthsPlane = &SourceNormalLengths[0];

		float InvDestW = 1.f/DestRoughness.SizeX;
		float InvDestH = 1.f/DestRoughness.SizeY;

		for( int64 DestY=0; DestY< DestRoughness.SizeY; DestY++)
		{
			float V = (DestY + 0.5f) * InvDestH;
			for( int64 DestX=0; DestX< DestRoughness.SizeX; DestX++)
			{
				float U = (DestX + 0.5f) * InvDestW;
				
				//const FLinearColor NormalColor = LookupSourceMipBilinearUV(NormalColors,U,V);
				const float NormalLength = LookupFloatBilinearUV(SourceNormalLengthsPlane,SourceNormals.SizeX,SourceNormals.SizeY,U,V);

				float Roughness = *TargetValuePtr;
				*TargetValuePtr = CompositeNormalLengthToRoughness(NormalLength,Roughness,CompositePower);
				TargetValuePtr += 4;
			}
		}
	}

	return true;
}

/*------------------------------------------------------------------------------
	Image Compression.
------------------------------------------------------------------------------*/

void FTextureBuildSettings::GetEncodedTextureDescriptionWithPixelFormat(FEncodedTextureDescription* OutTextureDescription, EPixelFormat InEncodedPixelFormat, int32 InEncodedMip0SizeX, int32 InEncodedMip0SizeY, int32 InEncodedMip0NumSlices, int32 InMipCount) const
{
	FEncodedTextureDescription& TextureDescription = *OutTextureDescription;
	TextureDescription = FEncodedTextureDescription();
	TextureDescription.bCubeMap = bCubemap;
	TextureDescription.bTextureArray = bTextureArray;
	TextureDescription.bVolumeTexture = bVolume;
	TextureDescription.NumMips = InMipCount;
	TextureDescription.PixelFormat = InEncodedPixelFormat;

	TextureDescription.TopMipSizeX = InEncodedMip0SizeX;
	TextureDescription.TopMipSizeY = InEncodedMip0SizeY;
	TextureDescription.TopMipVolumeSizeZ = bVolume ? InEncodedMip0NumSlices : 1;
	if (bTextureArray)
	{
		if (bCubemap)
		{
			TextureDescription.ArraySlices = InEncodedMip0NumSlices / 6;
		}
		else
		{
			TextureDescription.ArraySlices = InEncodedMip0NumSlices;
		}
	}
	else
	{
		TextureDescription.ArraySlices = 1;
	}
}

void FTextureBuildSettings::GetEncodedTextureDescription(FEncodedTextureDescription* OutTextureDescription, const ITextureFormat* InTextureFormat, int32 InEncodedMip0SizeX, int32 InEncodedMip0SizeY, int32 InEncodedMip0NumSlices, int32 InMipCount, bool bInImageHasAlphaChannel) const
{
	EPixelFormat EncodedPixelFormat = InTextureFormat->GetEncodedPixelFormat(*this, bInImageHasAlphaChannel);
	GetEncodedTextureDescriptionWithPixelFormat(OutTextureDescription, EncodedPixelFormat, InEncodedMip0SizeX, InEncodedMip0SizeY, InEncodedMip0NumSlices, InMipCount);
}

// compress mip-maps in InMipChain and add mips to Texture, might alter the source content
static bool CompressMipChain(
	const ITextureFormat* TextureFormat,
	const TArray<FImage>& MipChain,
	const FTextureBuildSettings& Settings,
	const bool bImageHasAlphaChannel,
	FStringView DebugTexturePathName,
	TArray<FCompressedImage2D>& OutMips,
	uint32& OutNumMipsInTail,
	uint32& OutExtData)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.CompressMipChain)

	// Determine ExtData (platform specific data) and NumMipsInTail.
	// ExtData gets passed 
	FEncodedTextureExtendedData ExtendedData;
	FEncodedTextureDescription TextureDescription;
	Settings.GetEncodedTextureDescription(&TextureDescription, TextureFormat, MipChain[0].SizeX, MipChain[0].SizeY, MipChain[0].NumSlices, MipChain.Num(), bImageHasAlphaChannel);
	{
		ExtendedData = TextureFormat->GetExtendedDataForTexture(TextureDescription, Settings.LODBias);
		OutNumMipsInTail = ExtendedData.NumMipsInTail;
		OutExtData = ExtendedData.ExtData;
	}

	int32 MipCount = MipChain.Num();
	check(MipCount >= (int32)ExtendedData.NumMipsInTail);
	// This number was too small (128) for current hardware and caused too many
	// context switch for work taking < 1ms. Bump the value for 2020 CPUs.
	const int32 MinAsyncCompressionSize = 512;
	const bool bAllowParallelBuild = TextureFormat->AllowParallelBuild();
	bool bCompressionSucceeded = true;

	// Mip tail is when the last few mips get grouped together in the hardware layout.
	// Treat not having a mip tail as having a mip tail with 1 mip in it, which is
	// equivalent and lets us simplify the logic.
	int32 FirstMipTailIndex = MipCount - 1;
	int32 MipTailCount = 1;

	if (ExtendedData.NumMipsInTail > 1)
	{
		MipTailCount = ExtendedData.NumMipsInTail;
		FirstMipTailIndex = MipCount - MipTailCount;
	}

	uint32 StartCycles = FPlatformTime::Cycles();

	// Set up one task for the base mip, one task for everything after. Since each mip level
	// has 4x the pixels as the one below it (8x for volumes), work for mip levels is highly
	// unbalanced and there's not much use spawning extra tasks past that: for a 2D texture,
	// the entire tail after the base mip (all remaining mips combined) has 1/3 the number of
	// pixels the base mip does.
	OutMips.Empty(MipCount);
	OutMips.AddDefaulted(MipCount);

	FIntVector3 Mip0Dimensions = TextureDescription.GetMipDimensions(0);
	int32 Mip0NumSlicesNoDepth = TextureDescription.GetNumSlices_NoDepth();

	auto ProcessMips =
		[&TextureFormat, &MipChain, &OutMips, &Mip0Dimensions, Mip0NumSlicesNoDepth, FirstMipTailIndex, MipTailCount, ExtData = ExtendedData.ExtData, &Settings, &DebugTexturePathName, bImageHasAlphaChannel](int32 MipBegin, int32 MipEnd)
	{
		bool bSuccess = true;

		for (int32 MipIndex = MipBegin; MipIndex < MipEnd; ++MipIndex)
		{
			TRACE_CPUPROFILER_EVENT_SCOPE(Texture.CompressImage);

			// We always compress 1 mip at a time, unless the platform requests that the mip tail
			// gets packed in to a single mip level (NumMipsInTail).
			int32 MipsToCompress = 1;
			if (MipIndex == FirstMipTailIndex)
			{
				MipsToCompress = MipTailCount;
			}

			bSuccess = bSuccess && TextureFormat->CompressImageEx(
				&MipChain[MipIndex],
				MipsToCompress,
				Settings,
				Mip0Dimensions,
				Mip0NumSlicesNoDepth,
				MipIndex,
				OutMips.Num(),
				DebugTexturePathName,
				bImageHasAlphaChannel,
				ExtData,
				OutMips[MipIndex]
			);

			// no longer possible, because hashing is done on a thread
			// todo: restore this by making the images refcounted to track lifetime better
			/*
			// note: MipChain[MipIndex].RawData may be freed or mutated by CompressImage
			// do not use it after the call to CompressImage
			// go ahead and free it now if CompressImage didn't :
			for(int MipSubIndex=0;MipSubIndex<MipsToCompress;MipSubIndex++)
			{
				MipChain[MipIndex+MipSubIndex].RawData.Empty();
			}
			*/
		}

		return bSuccess;
	};

	if (bAllowParallelBuild &&
		FirstMipTailIndex > 0 &&
		FMath::Min(MipChain[0].SizeX, MipChain[0].SizeY) >= MinAsyncCompressionSize)
	{
		// Spawn async job to compress all mips below base
		auto AsyncTask = UE::Tasks::Launch(TEXT("Texture.CompressLowerMips"),
			[&ProcessMips, FirstMipTailIndex]()
			{
				TRACE_CPUPROFILER_EVENT_SCOPE(Texture.CompressLowerMips);
				return ProcessMips(1, FirstMipTailIndex + 1);
			},
			LowLevelTasks::ETaskPriority::BackgroundNormal
		);

		// Compress base mip on this thread, join with async compress of other mips
		bCompressionSucceeded = ProcessMips(0, 1);
		bCompressionSucceeded &= AsyncTask.GetResult();
	}
	else
	{
		// Compress all mips at once on this thread
		bCompressionSucceeded = ProcessMips(0, FirstMipTailIndex + 1);
	}

	// Fill out the dimensions for the packed mip tail, should we have one
	for (int32 MipIndex = FirstMipTailIndex + 1; MipIndex < MipCount; ++MipIndex)
	{
		FCompressedImage2D& PrevMip = OutMips[MipIndex - 1];
		FCompressedImage2D& DestMip = OutMips[MipIndex];
		DestMip.SizeX = FMath::Max(1, PrevMip.SizeX >> 1);
		DestMip.SizeY = FMath::Max(1, PrevMip.SizeY >> 1);
		DestMip.SizeZ = Settings.bVolume ? FMath::Max(1, PrevMip.SizeZ >> 1) : PrevMip.SizeZ;
		DestMip.PixelFormat = PrevMip.PixelFormat;
	}

	if (!bCompressionSucceeded)
	{
		OutMips.Empty();
	}

	uint32 EndCycles = FPlatformTime::Cycles();
	UE_LOG(LogTextureCompressor,Verbose,TEXT("Compressed %dx%dx%d %s in %fms"),
		MipChain[0].SizeX,
		MipChain[0].SizeY,
		MipChain[0].NumSlices,
		*Settings.TextureFormatName.ToString(),
		FPlatformTime::ToMilliseconds( EndCycles-StartCycles )
		);

	return bCompressionSucceeded;
}

// only useful for normal maps, fixed bad input (denormalized normals) and improved quality (quantization artifacts)
static void NormalizeMip(FImage& InOutMip)
{
	TRACE_CPUPROFILER_EVENT_SCOPE(Texture.NormalizeMip);

	//@todo Oodle : change FVector to FVector3f all over Texture code!
	const FVector NormalIfZero(0.f,0.f,1.f);

	int64 NumPixels = InOutMip.GetNumPixels();
	int64 NumPixelsEachJob;
	int32 NumJobs = ImageParallelForComputeNumJobsForPixels(NumPixelsEachJob,NumPixels);
	
	FLinearColor * ImageColors = InOutMip.AsRGBA32F().GetData();

	ParallelFor( TEXT("Texture.NormalizeMip.PF"),NumJobs,1, [&](int32 Index)
	{
		int64 StartIndex = Index * NumPixelsEachJob;
		int64 EndIndex = FMath::Min(StartIndex + NumPixelsEachJob, NumPixels);
		for (int64 i = StartIndex; i < EndIndex; ++i)
		{
			// @todo Oodle: SIMD NormalizeColors

			FLinearColor& Color = ImageColors[i];

			// this uses doubles due to FVector :(

			FVector Normal(Color.R * 2.0f - 1.0f, Color.G * 2.0f - 1.0f, Color.B * 2.0f - 1.0f);

			// GetSafeNormal returns Vec(0,0,0) by default for tiny input, instead return flat/up
			Normal = Normal.GetSafeNormal(UE_SMALL_NUMBER,NormalIfZero);

			Color = FLinearColor(Normal.X * 0.5f + 0.5f, Normal.Y * 0.5f + 0.5f, Normal.Z * 0.5f + 0.5f, Color.A);
		}
	}, EParallelForFlags::Unbalanced);
}


// Returns true if the target texture size is different and padding/stretching is required.
static bool GetPowerOfTwoTargetTextureSize(int32 InMip0SizeX, int32 InMip0SizeY, int32 InMip0NumSlices, bool bInIsVolume, ETexturePowerOfTwoSetting::Type InPow2Setting, int32& OutTargetSizeX, int32& OutTargetSizeY, int32& OutTargetSizeZ)
{
	check(InPow2Setting != ETexturePowerOfTwoSetting::None);

	int32 TargetTextureSizeX = InMip0SizeX;
	int32 TargetTextureSizeY = InMip0SizeY;
	int32 TargetTextureSizeZ = bInIsVolume ? InMip0NumSlices : 1; // Only used for volume texture.

	const int32 PowerOfTwoTextureSizeX = FMath::RoundUpToPowerOfTwo(TargetTextureSizeX);
	const int32 PowerOfTwoTextureSizeY = FMath::RoundUpToPowerOfTwo(TargetTextureSizeY);
	const int32 PowerOfTwoTextureSizeZ = FMath::RoundUpToPowerOfTwo(TargetTextureSizeZ);

	switch (InPow2Setting)
	{
	// None should not get here

	case ETexturePowerOfTwoSetting::PadToPowerOfTwo:
		TargetTextureSizeX = PowerOfTwoTextureSizeX;
		TargetTextureSizeY = PowerOfTwoTextureSizeY;
		TargetTextureSizeZ = PowerOfTwoTextureSizeZ;
		break;

	case ETexturePowerOfTwoSetting::PadToSquarePowerOfTwo:
		TargetTextureSizeX = TargetTextureSizeY = TargetTextureSizeZ =
			FMath::Max3<int32>(PowerOfTwoTextureSizeX, PowerOfTwoTextureSizeY, PowerOfTwoTextureSizeZ);
		break;

	default:
		checkf(false, TEXT("Unknown entry in ETexturePowerOfTwoSetting::Type"));
		break;
	}

	// Z only matters as a sampling dimension if we are a volume texture.
	if (bInIsVolume == false)
	{
		TargetTextureSizeZ = InMip0NumSlices;
	}

	OutTargetSizeX = TargetTextureSizeX;
	OutTargetSizeY = TargetTextureSizeY;
	OutTargetSizeZ = TargetTextureSizeZ;

	return (TargetTextureSizeX != InMip0SizeX) ||
		(TargetTextureSizeY != InMip0SizeY) ||
		(bInIsVolume && TargetTextureSizeZ != InMip0NumSlices);
}


int32 ITextureCompressorModule::GetMipCountForBuildSettings(
	int32 InMip0SizeX, int32 InMip0SizeY, int32 InMip0NumSlices,
	int32 InExistingMipCount,
	const FTextureBuildSettings& BuildSettings,
	int32& OutMip0SizeX, int32& OutMip0SizeY, int32& OutMip0NumSlices)
{
	// AFAICT LatLongCubeMaps don't do any of this - pow2 is broken with them but it runs, and max texture stuff
	// is handled internally in the extents function.
	int32 BaseSizeX = InMip0SizeX;
	int32 BaseSizeY = InMip0SizeY;
	int32 BaseSizeZ = BuildSettings.bVolume ? InMip0NumSlices : 1; // Volume textures are the only type that mip their Z, arrays and cubes are fixed.

	// LatLong sources are clamped in ComputeLongLatCubemapExtents
	if (BuildSettings.bLongLatSource == false)
	{
		ETexturePowerOfTwoSetting::Type PowerOfTwoMode = (ETexturePowerOfTwoSetting::Type)BuildSettings.PowerOfTwoMode;
		if (BuildSettings.MipGenSettings != TMGS_LeaveExistingMips &&
			PowerOfTwoMode != ETexturePowerOfTwoSetting::None)
		{
			int32 TargetSizeX, TargetSizeY, TargetSizeZ;
			bool NeedsAdjustment = GetPowerOfTwoTargetTextureSize(BaseSizeX, BaseSizeY, BaseSizeY, BuildSettings.bVolume, PowerOfTwoMode, TargetSizeX, TargetSizeY, TargetSizeZ);
			if (NeedsAdjustment)
			{
				// In this case we are regenerating the entire mip chain.
				InExistingMipCount = 1;
				BaseSizeX = TargetSizeX;
				BaseSizeY = TargetSizeY;
				BaseSizeZ = TargetSizeZ; // volume textures already accounted for
			}
			// Otherwise we have valid pow2 so we can reuse any existing mips and regenerate
			// any missing tail mips.
		}

		// Max texture resolution strips off mips that are above the limit.
		int64 MaxTextureResolution = BuildSettings.MaxTextureResolution;
		int32 GeneratedMipCount = FImageCoreUtils::GetMipCountFromDimensions(BaseSizeX, BaseSizeY, BaseSizeZ, BuildSettings.bVolume);
		int32 i = 0;
		for (; i < GeneratedMipCount; i++)
		{
			// The code in BuildTextureMips doesn't worry about fitting Z in volume textures...
			// \todo volume texture MaxTextureSize. The old code ignored size Z, so we do too. I'm not sure
			// there's ever a case where volume textures have a Z that's bigger than X/Y.
			int32 MipSizeX = FMath::Max<uint32>(1, BaseSizeX >> i);
			int32 MipSizeY = FMath::Max<uint32>(1, BaseSizeY >> i);
			int32 MipSizeZ = BuildSettings.bVolume ? FMath::Max<uint32>(1, BaseSizeZ >> i) : BaseSizeZ;

			if (MipSizeX <= MaxTextureResolution &&
				MipSizeY <= MaxTextureResolution)
			{
				BaseSizeX = MipSizeX;
				BaseSizeY = MipSizeY;
				BaseSizeZ = MipSizeZ;
				break;
			}
		}

		if (BuildSettings.Downscale > 1.0f)
		{
			int32 DownscaledSizeX = 0, DownscaledSizeY = 0;
			GetDownscaleFinalSizeAndClampedDownscale(BaseSizeX, BaseSizeY, FTextureDownscaleSettings(BuildSettings), DownscaledSizeX, DownscaledSizeY);

			if (BuildSettings.bVolume)
			{
				UE_LOG(LogTextureCompressor, Error, TEXT("Downscaling volumes not yet supported - should have been handled in GetTextureBuildSettings!"));
			}
			check(BuildSettings.bVolume == false);

			BaseSizeX = DownscaledSizeX;
			BaseSizeY = DownscaledSizeY;
		}

		// Volumes are the only thing where num slices changes.
		if (BuildSettings.bVolume == false)
		{
			OutMip0NumSlices = InMip0NumSlices;
		}
		else
		{
			OutMip0NumSlices = BaseSizeZ;
		}
	}
	else
	{
		uint32 LongLatCubemapExtents = ComputeLongLatCubemapExtents(BaseSizeX, BuildSettings.MaxTextureResolution);
		BaseSizeX = LongLatCubemapExtents;
		BaseSizeY = LongLatCubemapExtents;
		OutMip0NumSlices = 6 * InMip0NumSlices;
	}

	// At this point we have a base mip size that is valid.
	OutMip0SizeX = BaseSizeX;
	OutMip0SizeY = BaseSizeY;

	if (BuildSettings.MipGenSettings == TMGS_NoMipmaps)
	{
		return 1;
	}

	// LeaveExisting is often unintentionally used as "NoMipmaps", so if they bring in 1 mip, leave it as 1 mip.
	if (BuildSettings.MipGenSettings == TMGS_LeaveExistingMips &&
		InExistingMipCount == 1)
	{
		return 1;
	}

	// NumOutputMips is the number of mips that would be made if you made a full mip chain
	//  eg. 256 makes 9 mips , 300 also makes 9 mips
	return FImageCoreUtils::GetMipCountFromDimensions(BaseSizeX, BaseSizeY, BaseSizeZ, BuildSettings.bVolume);
}


/**
 * Texture compression module
 */
class FTextureCompressorModule : public ITextureCompressorModule
{
public:
	FTextureCompressorModule()
	{
	}
	
	// compute a hash used to identify the contents of a mip chain
	// this is used by bulkdata diff tool, stored in asset registry
	// it is for debug tools only and skipping it is optional
	static FIoHash ComputeMipChainHash(const TArray<FImage> & IntermediateMipChain)
	{
		TRACE_CPUPROFILER_EVENT_SCOPE(Texture.ComputeMipChainHash);

		// @todo: ComputeMipChainHash is quite slow and often not necessary
		//	todo: use xxhash instead of blake3
		//  todo: don't do it at all when it's not needed
		//	  eg. in interactive Editor runs and in UEFN contentworker

		// Hash the mips before we compress them. This gets saved as part of the derived data and then added to the
		// diff tags during cook so we can catch determinism issues.
		FIoHashBuilder MipHashBuilder;
		for (const FImage& Mip : IntermediateMipChain)
		{
			const FImageInfo* Info = &Mip;
			// Can't just hash FImageInfo due to struct padding RNG making the hash non-deterministic.
			MipHashBuilder.Update(&Info->SizeX, sizeof(Info->SizeX));
			MipHashBuilder.Update(&Info->SizeY, sizeof(Info->SizeY));
			MipHashBuilder.Update(&Info->NumSlices, sizeof(Info->NumSlices));
			MipHashBuilder.Update(&Info->Format, sizeof(Info->Format));
			MipHashBuilder.Update(&Info->GammaSpace, sizeof(Info->GammaSpace));

			check( Mip.RawData.Num() != 0 );

			// show perf of hash function :
			//double t1 = FPlatformTime::Seconds();
			MipHashBuilder.Update(Mip.RawData.GetData(), Mip.RawData.Num());
			//double t2 = FPlatformTime::Seconds();
			//UE_LOG(LogTextureCompressor, Display, TEXT("Hashed %lld bytes in %.3f millis = %.3f GB/s"),Mip.RawData.Num(),(t2-t1)*1000,(Mip.RawData.Num()/(t2-t1))/(1000*1000*1000));
			//Blake3 is ~3.5 GB/s , around 4X slower than xxHash
		}

		return MipHashBuilder.Finalize();
	}

	virtual bool BuildTexture(
		const TArray<FImage>& SourceMips,
		const TArray<FImage>& AssociatedNormalSourceMips,
		const FTextureBuildSettings& BuildSettings,
		FStringView DebugTexturePathName,
		TArray<FCompressedImage2D>& OutTextureMips,
		uint32& OutNumMipsInTail,
		uint32& OutExtData,
		UE::TextureBuildUtilities::FTextureBuildMetadata* OutMetadata
	)
	{
		//TRACE_CPUPROFILER_EVENT_SCOPE(Texture.BuildTexture);

		const ITextureFormat* TextureFormat = nullptr;

		ITextureFormatManagerModule* TFM = GetTextureFormatManager();
		if (TFM)
		{
			TextureFormat = TFM->FindTextureFormat(BuildSettings.TextureFormatName);
		}
		if (TextureFormat == nullptr)
		{
			UE_LOG(LogTextureCompressor, Warning,
				TEXT("Failed to find compressor for texture format '%s'. [%.*s]"),
				*BuildSettings.TextureFormatName.ToString(),
				DebugTexturePathName.Len(),DebugTexturePathName.GetData()
			);

			return false;
		}
		
		// @todo Oodle: option to dump the Source image here
		//		we have dump in TextureFormatOodle for the after-processing (before encoding) image
		//		get a dump spot for before-processing as well

		TArray<FImage> IntermediateMipChain;

		bool bSourceMipsAlphaDetected = false;
		if (OutMetadata)
		{
			bSourceMipsAlphaDetected = FImageCore::DetectAlphaChannel(SourceMips[0]);
		}

		// allow to leave texture in sRGB in case compressor accepts other than non-F32 input source
		// otherwise linearizing will force format to be RGBA32F
		const bool bNeedLinearize = !TextureFormat->CanAcceptNonF32Source(BuildSettings.TextureFormatName) || AssociatedNormalSourceMips.Num() != 0;
		if (!BuildTextureMips(SourceMips, BuildSettings, bNeedLinearize, IntermediateMipChain, DebugTexturePathName))
		{
			return false;
		}
		
		// apply roughness adjustment depending on normal map variation
		if (AssociatedNormalSourceMips.Num())
		{
			TRACE_CPUPROFILER_EVENT_SCOPE(Texture.AssociatedNormalSourceMips);

			// ECompositeTextureMode is only NormalRoughness
			//  AssociatedNormalSourceMips should be a normal map

			TArray<FImage> IntermediateAssociatedNormalSourceMipChain;

			FTextureBuildSettings DefaultSettings;

			// apply a smooth Gaussian filter to the top level of the normal map
			// the original comment says :
			//  "helps to reduce aliasing further"
			// what's happening here is the blur on the top mip will reduce the length of normals in rough areas
			//	whereas without it the top mip would always have normals of length 1.0 , hence zero roughness per Toksvig
			DefaultSettings.MipSharpening = -3.5f;
			//DefaultSettings.MipSharpening = -2.5f; // CB: I think smaller blend looks better but don't change existing data
			DefaultSettings.SharpenMipKernelSize = 6;
			DefaultSettings.bApplyKernelToTopMip = true;

			// important to make accurate computation with normal length
			//  note this normalizes the top mip *before* the gaussian blur
			DefaultSettings.bRenormalizeTopMip = true;
			DefaultSettings.bNormalizeNormals = false; // do not normalize after mip gen, we want shortening

			// use new mip filter setting from build settings
			DefaultSettings.bUseNewMipFilter = BuildSettings.bUseNewMipFilter;

			if (!BuildTextureMips(AssociatedNormalSourceMips, DefaultSettings, true, IntermediateAssociatedNormalSourceMipChain, DebugTexturePathName))
			{
				UE_LOG(LogTextureCompressor, Warning, TEXT("Failed to generate texture mips for composite texture [%.*s]"),
					DebugTexturePathName.Len(),DebugTexturePathName.GetData());

				return false;
			}

			if (!ApplyCompositeTextureToMips(IntermediateMipChain, IntermediateAssociatedNormalSourceMipChain, BuildSettings.CompositeTextureMode, BuildSettings.CompositePower, BuildSettings.LODBias))
			{
				UE_LOG(LogTextureCompressor, Warning, TEXT("ApplyCompositeTextureToMips failed [%.*s]"),
					DebugTexturePathName.Len(),DebugTexturePathName.GetData());

				return false;
			}
		}

		
		// DetectAlphaChannel on the top mip of the generated mip chain. SoonTM this will use the source mip chain. Testing has
		// shown this to be 99.9% the same, and allows us to get the pixel format earlier.
		const bool bImageHasAlphaChannel = BuildSettings.GetTextureExpectsAlphaInPixelFormat(FImageCore::DetectAlphaChannel(IntermediateMipChain[0]));
		
		UE::Tasks::TTask<FIoHash> HashingTask;
		TArray<FImageInfo> SaveImageInfos;
		
		//FIoHash DebugOnly_HashBefore = ComputeMipChainHash(IntermediateMipChain);

		if (OutMetadata)
		{
			TRACE_CPUPROFILER_EVENT_SCOPE(Texture.MipHashBuilder);

			SaveImageInfos.SetNum(IntermediateMipChain.Num());
			for (int32 i=0;i<IntermediateMipChain.Num();++i)
			{
				SaveImageInfos[i] = IntermediateMipChain[i];
			}

			// The metadata is about trying to determine bImageHasAlphaChannel _before_ we launch a task, which 
			// means it must be on the actual source mips, not the post-processed mips. At some point enough textures
			// will have this saved as part of the creation process and we can rely on it - but for now we route it
			// back out so that when textures happen to get saved, it goes with it.
			OutMetadata->bSourceMipsAlphaDetected = bSourceMipsAlphaDetected;

			// hash value is not needed immediately, so don't block on it :
			HashingTask = UE::Tasks::Launch(TEXT("Texture.MipHashBuilder.Task"), [&](){ return ComputeMipChainHash(IntermediateMipChain); } );
		}
		
		bool bCompressSucceeded = CompressMipChain(TextureFormat, IntermediateMipChain, BuildSettings, bImageHasAlphaChannel, DebugTexturePathName,
					OutTextureMips, OutNumMipsInTail, OutExtData);
		
		// make sure IntermediateMipChain isn't changed:
		//FIoHash DebugOnly_HashAfter = ComputeMipChainHash(IntermediateMipChain);
		//check( DebugOnly_HashBefore == DebugOnly_HashAfter );

		if (OutMetadata)
		{
			// now block on the hash task
			OutMetadata->PreEncodeMipsHash = HashingTask.GetResult();
		
			//check( DebugOnly_HashBefore == OutMetadata->PreEncodeMipsHash );
		
			// (rough) check that IntermediateMipChain was not modified by the TextureFormats :
			check( SaveImageInfos.Num() == IntermediateMipChain.Num() );
			for (int32 i=0;i<IntermediateMipChain.Num();++i)
			{
				// check that FImageInfo is the same
				// does not check pixel values
				check( SaveImageInfos[i] == IntermediateMipChain[i] );
			}
		}

		{
			TRACE_CPUPROFILER_EVENT_SCOPE(Texture.Free_IntermediateMipChain);

			// this is quite slow; detach to an async task?
			//IntermediateMipChain.Empty();

			TArray<FImage>* PtrIntermediateMipChain = new TArray<FImage>( MoveTemp(IntermediateMipChain) );
			
			//TArray<FImage>* PtrIntermediateMipChain = new TArray<FImage>;
			//Swap(*PtrIntermediateMipChain,IntermediateMipChain);

			UE::Tasks::Launch(TEXT("Texture.Free_IntermediateMipChain.Task"),
				[PtrIntermediateMipChain]() // by value, not ref
				{
					TRACE_CPUPROFILER_EVENT_SCOPE(Texture.Free_IntermediateMipChain.Task);
					PtrIntermediateMipChain->Empty();
					delete PtrIntermediateMipChain;
					return true;
				},
				LowLevelTasks::ETaskPriority::BackgroundNormal);
		}

		return bCompressSucceeded;
	}

	// IModuleInterface implementation.
	void StartupModule()
	{
	}

	void ShutdownModule()
	{
	}

private:


	bool BuildTextureMips(
		const TArray<FImage>& InSourceMipChain,
		const FTextureBuildSettings& BuildSettings,
		const bool bNeedLinearize,
		TArray<FImage>& OutMipChain,
		FStringView DebugTexturePathName)
	{
		TRACE_CPUPROFILER_EVENT_SCOPE(Texture.BuildTextureMips);
		
		check(InSourceMipChain.Num() > 0);
		check(InSourceMipChain[0].SizeX > 0 && InSourceMipChain[0].SizeY > 0 && InSourceMipChain[0].NumSlices > 0);

		// Identify long-lat cubemaps.
		const bool bLongLatCubemap = BuildSettings.bLongLatSource;
		if (BuildSettings.bCubemap && !bLongLatCubemap)
		{
			if (BuildSettings.bTextureArray && (InSourceMipChain[0].NumSlices % 6) != 0)
			{
				// Cube array must have multiple of 6 slices
				return false;
			}
			if (!BuildSettings.bTextureArray && InSourceMipChain[0].NumSlices != 6)
			{
				// Non-array cube must have exactly 6 slices
				return false;
			}
		}

		// handling of bLongLatCubemap seems overly complicated
		//	what it should do is convert it right at the start here
		//	then treat it as a standard cubemap below, no special cases
		//	but that will change output :(

		// pSourceMips will track the current FImages we consider to be "source"
		const TArray<FImage> * pSourceMips = &InSourceMipChain;

		// first pad up to pow2 if requested
		ETexturePowerOfTwoSetting::Type PowerOfTwoMode = (ETexturePowerOfTwoSetting::Type) BuildSettings.PowerOfTwoMode;

		// if bPadOrStretchTextureis done, PaddedSourceMips is filled with a new image
		TArray<FImage> PaddedSourceMips;

		if ( PowerOfTwoMode != ETexturePowerOfTwoSetting::None )
		{
			const FImage& FirstSourceMipImage = (*pSourceMips)[0];

			int32 TargetTextureSizeX = 0;
			int32 TargetTextureSizeY = 0;
			int32 TargetTextureSizeZ = 0;			
			bool bPadOrStretchTexture = GetPowerOfTwoTargetTextureSize(
				FirstSourceMipImage.SizeX, FirstSourceMipImage.SizeY, FirstSourceMipImage.NumSlices,
				BuildSettings.bVolume, PowerOfTwoMode,
				TargetTextureSizeX, TargetTextureSizeY, TargetTextureSizeZ);

			if (bPadOrStretchTexture)
			{
				if (BuildSettings.MipGenSettings == TMGS_LeaveExistingMips)
				{
					// pad+leave existing is broken
					UE_LOG(LogTextureCompressor, Error,	TEXT("Texture padded to pow2 + LeaveExistingMips forbidden"));
					return false;
				}
				if ( bLongLatCubemap )
				{
					UE_LOG(LogTextureCompressor, Warning, TEXT("PadPow2 + LongLat cubemap doesn't work, continuing.."));
				}

				// Want to stretch or pad the texture
				bool bSuitableFormat = FirstSourceMipImage.Format == ERawImageFormat::RGBA32F;

				FImage Temp;
				if (!bSuitableFormat)
				{
					// convert to RGBA32F
					FirstSourceMipImage.CopyTo(Temp, ERawImageFormat::RGBA32F, EGammaSpace::Linear);
				}

				// space for one source mip and one destination mip
				const FImage& SourceImage = bSuitableFormat ? FirstSourceMipImage : Temp;
				FImage& TargetImage = *new (PaddedSourceMips) FImage(TargetTextureSizeX, TargetTextureSizeY, BuildSettings.bVolume ? TargetTextureSizeZ : SourceImage.NumSlices, SourceImage.Format);
				FLinearColor FillColor = BuildSettings.PaddingColor;

				FLinearColor* TargetPtr = (FLinearColor*)TargetImage.RawData.GetData();
				FLinearColor* SourcePtr = (FLinearColor*)SourceImage.RawData.GetData();
				check(SourceImage.GetBytesPerPixel() == sizeof(FLinearColor));
				check(TargetImage.GetBytesPerPixel() == sizeof(FLinearColor));

				for (int32 SliceIndex = 0; SliceIndex < SourceImage.NumSlices; ++SliceIndex)
				{
					for (int32 Y = 0; Y < TargetTextureSizeY; ++Y)
					{
						int32 XStart = 0;
						if (Y < SourceImage.SizeY)
						{
							XStart = SourceImage.SizeX;
							FMemory::Memcpy(TargetPtr, SourcePtr, SourceImage.SizeX * sizeof(FLinearColor));
							SourcePtr += SourceImage.SizeX;
							TargetPtr += SourceImage.SizeX;
						}

						for (int32 XPad = XStart; XPad < TargetImage.SizeX; ++XPad)
						{
							*TargetPtr++ = FillColor;
						}
					}
				}
				// Pad new slices for volume texture
				for (int32 SliceIndex = SourceImage.NumSlices; SliceIndex < TargetImage.NumSlices; ++SliceIndex)
				{
					for (int32 Y = 0; Y < TargetImage.SizeY; ++Y)
					{
						for (int32 X = 0; X< TargetImage.SizeX; ++X)
						{
							*TargetPtr++ = FillColor;
						}
					}
				}
				
				// change pSourceMips to point at the one padded image we made
				pSourceMips = &PaddedSourceMips;
			}
		}		

		// now pow2 pad is done
		// find a starting source that meets MaxTextureResolution limit

		int32 StartMip = 0;

		TArray<FImage> BuildSourceImageMips;

		if ( ! bLongLatCubemap )
		{
			int32 NumSourceMips = (BuildSettings.MipGenSettings == TMGS_LeaveExistingMips) ? pSourceMips->Num() : 1;
			
			int64 MaxTextureResolution = BuildSettings.MaxTextureResolution;

			// note that "LODBias" is very similar to MaxTextureResolution
			//	but for LODBias we go ahead and make all the mips here
			//	and then just don't serialize the top ones in TextureDerivedData
			// (LODBias is not actually LOD Bias, it means discard top N mips)

			// step through source mips to find one that meets MaxTextureResolution
			while( StartMip < NumSourceMips && (
				(*pSourceMips)[StartMip].SizeX > MaxTextureResolution ||
				(*pSourceMips)[StartMip].SizeY > MaxTextureResolution ) )
			{
				StartMip++;
			}

			if ( StartMip == NumSourceMips )
			{

				// bLongLatCubemap should not get here because cube size is made from MaxTextureSize
				check( ! bLongLatCubemap );

				// the source is larger than the compressor allows and no mip image exists to act as a smaller source.
				// We must generate a suitable source image:
				const FImage& BaseImage = pSourceMips->Last();
				bool bSuitableFormat = BaseImage.Format == ERawImageFormat::RGBA32F;
			
				check( MaxTextureResolution > 0 );
				check( BaseImage.SizeX > MaxTextureResolution || 
					   BaseImage.SizeY > MaxTextureResolution );

				FImage Temp;
				if (!bSuitableFormat)
				{
					// convert to RGBA32F
					BaseImage.CopyTo(Temp, ERawImageFormat::RGBA32F, EGammaSpace::Linear);
				}

				UE_LOG(LogTextureCompressor, Verbose,
					TEXT("Source image %dx%d too large for compressors max dimension (%d). Resizing."),
					BaseImage.SizeX,
					BaseImage.SizeY,
					BuildSettings.MaxTextureResolution
					);

				// make sure BuildSourceImageMips doesn't reallocate :
				constexpr int BuildSourceImageMipsMaxCount = 20; // plenty
				BuildSourceImageMips.Empty(BuildSourceImageMipsMaxCount);

				// Max Texture Size resizing happens here :
				// note we do not check for TMGS_Angular here
				GenerateMipChain(BuildSettings, bSuitableFormat ? BaseImage : Temp, BuildSourceImageMips, 1);

				while( BuildSourceImageMips.Last().SizeX > MaxTextureResolution || 
					   BuildSourceImageMips.Last().SizeY > MaxTextureResolution )
				{
					// note: now making mips one by one, rather than N in one call
					//	this is not exactly the same if AlphaCoverage processing is on
					check( BuildSourceImageMips.Num() < BuildSourceImageMipsMaxCount );
					GenerateMipChain(BuildSettings, BuildSourceImageMips.Last(), BuildSourceImageMips, 1);
				}
			
				check( BuildSourceImageMips.Last().SizeX <= MaxTextureResolution &&
					   BuildSourceImageMips.Last().SizeY <= MaxTextureResolution );

				// change pSourceMips to point at the mip chain we made
				pSourceMips = &BuildSourceImageMips;
				StartMip = BuildSourceImageMips.Num() - 1;
				// [StartMip] will now references BuildSourceImageMips.Last()
			}
		}

		// now shrinking to MaxTextureResolution is done, figure out which mips to use or make
		
		// Copy over base mip and any LeaveExisting, from SourceMips , starting at StartMip
		int32 CopyCount = pSourceMips->Num() - StartMip;
		check( CopyCount > 0 );
		
		int32 NumOutputMips;

		if ( BuildSettings.MipGenSettings == TMGS_NoMipmaps )
		{
			NumOutputMips = 1;
			CopyCount = 1;
		}
		else
		{
			const FImage & TopMip = (*pSourceMips)[StartMip];

			// NumOutputMips is the number of mips that would be made if you made a full mip chain
			//  eg. 256 makes 9 mips , 300 also makes 9 mips
			if (bLongLatCubemap)
			{
				NumOutputMips = FImageCoreUtils::GetMipCountFromDimensions(ComputeLongLatCubemapExtents(TopMip.SizeX, BuildSettings.MaxTextureResolution), 1, 1, false);
			}
			else
			{
				NumOutputMips = FImageCoreUtils::GetMipCountFromDimensions(TopMip.SizeX, TopMip.SizeY, TopMip.NumSlices, BuildSettings.bVolume);
			}

			// LeaveExistingMips is often inadvertently used as "NoMips", so if they only brought in 1 mip, leave it as
			// 1 mip.
			if (BuildSettings.MipGenSettings == TMGS_LeaveExistingMips &&
				InSourceMipChain.Num() == 1)
			{
				NumOutputMips = 1;
				check(CopyCount == 1);
			}

			// unless LeaveExistingMips, we only copy 1
			if (BuildSettings.MipGenSettings != TMGS_LeaveExistingMips)
			{
				CopyCount = 1;
			}
		}

		// we will output NumOutputMips
		OutMipChain.Empty(NumOutputMips);

		int32 GenerateCount = NumOutputMips - CopyCount;
		check( GenerateCount >= 0 );

		// avoid converting to RGBA32F linear format if there's no need for any extra processing of pixels
		// image will be left in BGRA8 format if possible
		const bool bNeedAdjustImageColors = NeedAdjustImageColors(BuildSettings);
		const bool bLinearize = bNeedLinearize || (GenerateCount > 0) || BuildSettings.bRenormalizeTopMip || (BuildSettings.Downscale > 1.f)
			|| BuildSettings.bHasColorSpaceDefinition || BuildSettings.bComputeBokehAlpha || BuildSettings.bFlipGreenChannel
			|| BuildSettings.bReplicateRed || BuildSettings.bReplicateAlpha || BuildSettings.bApplyYCoCgBlockScale
			|| BuildSettings.SourceEncodingOverride != 0 || bNeedAdjustImageColors || BuildSettings.bNormalizeNormals;

		for (int32 MipIndex = StartMip; MipIndex < StartMip + CopyCount; ++MipIndex)
		{
			const FImage& Image = (*pSourceMips)[MipIndex];

			// copy mips over + processing
			// this is a code dupe of the processing done in GenerateMipChain

			// create base for the mip chain
			FImage* Mip = new(OutMipChain) FImage();

			if (bLongLatCubemap)
			{
				// Generate the base mip from the long-lat source image.
				GenerateBaseCubeMipFromLongitudeLatitude2D(Mip, Image, BuildSettings.MaxTextureResolution, BuildSettings.SourceEncodingOverride);
	
				check( CopyCount == 1 );
			}
			else
			{
				// copy base source content to the base of the mip chain
				if(BuildSettings.bApplyKernelToTopMip)
				{
					FImage Temp;
					Image.Linearize(BuildSettings.SourceEncodingOverride, Temp);
					if(BuildSettings.bRenormalizeTopMip)
					{
						NormalizeMip(Temp);
					}

					GenerateTopMip(Temp, *Mip, BuildSettings);
				}
				else
				{
					if (bLinearize)
					{
						Image.Linearize(BuildSettings.SourceEncodingOverride, *Mip);
						if (BuildSettings.bRenormalizeTopMip)
						{
							NormalizeMip(*Mip);
						}
					}
					else
					{
						// if image is in BGRA8 format leave it, otherwise convert to RGBA32F
						//  we only support leaving images in source format if they are BGRA8 and require no processing (eg VT tiles)
						ERawImageFormat::Type DestFormat = Image.Format;
						EGammaSpace DestGammaSpace = Image.GammaSpace;
						if ( Image.Format != ERawImageFormat::BGRA8 )
						{
							DestFormat = ERawImageFormat::RGBA32F;
							DestGammaSpace = EGammaSpace::Linear;
						}
						Image.CopyTo(*Mip, DestFormat, DestGammaSpace);

						//@@CB todo : when Mip format == Image format, we can Move instead of Copy
						//	have to make sure that's okay with SourceMips/TextureData
					}
				}
			}

			//@todo Oodle: free SourceMips as we consume them?

			if (BuildSettings.Downscale > 1.f)
			{		
				DownscaleImage(*Mip, *Mip, FTextureDownscaleSettings(BuildSettings));
			}

			if (BuildSettings.bHasColorSpaceDefinition)
			{
				FImageCore::TransformToWorkingColorSpace(
					*Mip,
					FVector2d(BuildSettings.RedChromaticityCoordinate),
					FVector2d(BuildSettings.GreenChromaticityCoordinate),
					FVector2d(BuildSettings.BlueChromaticityCoordinate),
					FVector2d(BuildSettings.WhiteChromaticityCoordinate),
					static_cast<UE::Color::EChromaticAdaptationMethod>(BuildSettings.ChromaticAdaptationMethod));
			}

			// Apply color adjustments
			AdjustImageColors(*Mip, BuildSettings);

			if (BuildSettings.bComputeBokehAlpha)
			{
				// To get the occlusion in the BokehDOF shader working for all Bokeh textures.
				ComputeBokehAlpha(*Mip);
			}
			if (BuildSettings.bFlipGreenChannel)
			{
				FlipGreenChannel(*Mip);
			}
		}

		check( OutMipChain.Num() == CopyCount );
		check( GenerateCount == NumOutputMips - OutMipChain.Num() );

		// Generate any missing mips in the chain.
		if ( GenerateCount > 0 )
		{
			// Do angular filtering of cubemaps if requested.
			if (BuildSettings.MipGenSettings == TMGS_Angular)
			{
				check( BuildSettings.bCubemap );
				// note TMGS_Angular forces dim to next lower power of 2

				// note GenerateAngularFilteredMips reprocesses ALL the mips, not just GenerateCount
				// this should probably be outside the GenerateCount check (eg. always done, even if GenerateCount == 0 )
				// but putting it inside matches existing behavior
				// I guess it's moot because you can't set NoMipMips or LeaveExisting if you chose Angular

				GenerateAngularFilteredMips(OutMipChain, NumOutputMips, BuildSettings.DiffuseConvolveMipLevel);
			}
			else
			{
				// GenerateMipChain should bring us up to NumOutputMips
				//	but it doesn't take NumOutputMips as a param, makes its own decision
				//  we will check that it chose the same mip count after

				// you could pass GenerateCount as the large arg here
				//  and it should make the same result

				GenerateMipChain(BuildSettings, OutMipChain.Last(), OutMipChain, MAX_uint32);
			}
		}
		check(OutMipChain.Num() == NumOutputMips);

		int32 CalculatedMip0SizeX, CalculatedMip0SizeY, CalculatedMip0NumSlices;
		int32 CalculatedMipCount = GetMipCountForBuildSettings(
			InSourceMipChain[0].SizeX, InSourceMipChain[0].SizeY, InSourceMipChain[0].NumSlices, InSourceMipChain.Num(), BuildSettings,
			CalculatedMip0SizeX, CalculatedMip0SizeY, CalculatedMip0NumSlices);
		if (CalculatedMipCount != NumOutputMips ||
			CalculatedMip0SizeX != OutMipChain[0].SizeX ||
			CalculatedMip0SizeY != OutMipChain[0].SizeY ||
			CalculatedMip0NumSlices != OutMipChain[0].NumSlices)
		{
			UE_LOG(LogTextureCompressor, Error, TEXT("Texture %.*s generated unexpected mip chain: GetMipCountForBuildSettings expected %d mips, %dx%dx%d, got %d mips, %dx%dx%d!"), 
				DebugTexturePathName.Len(), DebugTexturePathName.GetData(), 
				CalculatedMipCount, CalculatedMip0SizeX, CalculatedMip0SizeY, CalculatedMip0NumSlices,
				NumOutputMips, OutMipChain[0].SizeX, OutMipChain[0].SizeY, OutMipChain[0].NumSlices);
		}

		// Apply post-mip generation adjustments.
		if ( BuildSettings.bNormalizeNormals )
		{
			TRACE_CPUPROFILER_EVENT_SCOPE(Texture.NormalizeMipChain);
			
			// Parallel on the mips, but not on tiny mips :
			int32 NumParallelMips = OutMipChain.Num();
			while ( NumParallelMips > 0 && OutMipChain[NumParallelMips-1].GetNumPixels() < 16384 )
			{
				NumParallelMips--;
			}

			ParallelFor( TEXT("Texture.NormalizeMipChain.PF"),NumParallelMips,1, [&](int32 Mip)
			{
				if ( Mip == NumParallelMips-1 )
				{
					// tiny mips in the tail done on one job:
					do
					{
						NormalizeMip(OutMipChain[Mip]);
						++Mip;
					} while( Mip < OutMipChain.Num() );
				}
				else
				{
					NormalizeMip(OutMipChain[Mip]);
				}
			}, EParallelForFlags::Unbalanced);

		}
	
		if (BuildSettings.bReplicateRed)
		{
			check( !BuildSettings.bReplicateAlpha ); // cannot both be set 
			ReplicateRedChannel(OutMipChain);
		}
		else if (BuildSettings.bReplicateAlpha)
		{
			ReplicateAlphaChannel(OutMipChain);
		}

		if (BuildSettings.bApplyYCoCgBlockScale)
		{
			ApplyYCoCgBlockScale(OutMipChain);
		}

		return true;
	}

	// @param CompositeTextureMode original type ECompositeTextureMode
	// @return true on success, false on failure. Can fail due to bad mismatched dimensions of incomplete mip chains.
	bool ApplyCompositeTextureToMips(TArray<FImage>& DestRoughnessMips, const TArray<FImage>& NormalSourceMips, uint8 CompositeTextureMode, float CompositePower, int32 DestLODBias)
	{
		check( DestRoughnessMips.Num() > 0 );
		check( NormalSourceMips.Num() > 0 );

		// NormalSourceMips is always a full mip chain, because we just made it, ignoring MipGen settings on the normal map texture
		// DestRoughnessMips are the mips being made in the current texture build, they could be an incomplete set if NoMipMips or LeaveExisting

		// must write to every mip of output :
		for(int32 DestLevel = 0; DestLevel < DestRoughnessMips.Num(); ++DestLevel)
		{
			if ( DestRoughnessMips[DestLevel].SizeX > NormalSourceMips[0].SizeX &&
				 DestRoughnessMips[DestLevel].SizeY > NormalSourceMips[0].SizeY )
			{
				// Normal map size is smaller than dest, no Roughness needed
				// at equal size, DO compute roughness as a Gaussian has been applied to filter normals
				//  to find a roughness at the top mip level
				continue;
			}

			// find a mip of source normals that is the same size as current mip, if possible
			int32 SourceNormalMipLevel = FMath::Min(DestLevel,NormalSourceMips.Num()-1);
			while( SourceNormalMipLevel > 0 && (
				NormalSourceMips[SourceNormalMipLevel].SizeX < DestRoughnessMips[DestLevel].SizeX ||
				NormalSourceMips[SourceNormalMipLevel].SizeY < DestRoughnessMips[DestLevel].SizeY) )
			{
				SourceNormalMipLevel--;
			}
			while( SourceNormalMipLevel < NormalSourceMips.Num()-1 && (
				NormalSourceMips[SourceNormalMipLevel].SizeX > DestRoughnessMips[DestLevel].SizeX ||
				NormalSourceMips[SourceNormalMipLevel].SizeY > DestRoughnessMips[DestLevel].SizeY) )
			{
				SourceNormalMipLevel++;
			}

			if ( 
				DestRoughnessMips[DestLevel].SizeX != NormalSourceMips[SourceNormalMipLevel].SizeX ||
				DestRoughnessMips[DestLevel].SizeY != NormalSourceMips[SourceNormalMipLevel].SizeY )
			{
				UE_LOG(LogTextureCompressor, Display, 
					TEXT( "ApplyCompositeTexture: Couldn't find matching mip size, will stretch.  (dest: %dx%d with %d mips, source: %dx%d with %d mips).  current: (dest: %dx%d at %d, source: %dx%d at %d)" ),
					DestRoughnessMips[0].SizeX,
					DestRoughnessMips[0].SizeY,
					DestRoughnessMips.Num(),
					NormalSourceMips[0].SizeX,
					NormalSourceMips[0].SizeY,
					NormalSourceMips.Num(),
					DestRoughnessMips[DestLevel].SizeX,
					DestRoughnessMips[DestLevel].SizeY,
					DestLevel,
					NormalSourceMips[SourceNormalMipLevel].SizeX,
					NormalSourceMips[SourceNormalMipLevel].SizeY,
					SourceNormalMipLevel
					);
			}

			if ( ! ApplyCompositeTexture(DestRoughnessMips[DestLevel], NormalSourceMips[SourceNormalMipLevel], CompositeTextureMode, CompositePower) )
			{
				return false;
			}
		}

		return true;
	}
};

IMPLEMENT_MODULE(FTextureCompressorModule, TextureCompressor)