GPUFFT Namespace Reference

Classes
struct	ComplexFFTImage1D
struct	ConvolutionWithTextureImage1D
class	FFTDescription
struct	TwoForOneRealFFTImage1D

Typedefs
typedef FVector3f	FPreFilter

Enumerations
enum class	FFT_XFORM_TYPE : int { FORWARD_HORIZONTAL = 0 , FORWARD_VERTICAL = 1 , INVERSE_HORIZONTAL = 2 , INVERSE_VERTICAL = 3 }

Functions
uint32	BitSize (uint32 N)
TArray< uint32 >	GetFactors (const uint32 PowTwoLength, const uint32 PowTwoBase)
void	SwapContents (FRDGTextureRef &TmpBuffer, FRDGTextureRef &DstBuffer)
bool	IsHorizontal (const FFT_XFORM_TYPE &XFormType)
bool	IsForward (const FFT_XFORM_TYPE &XFormType)
uint32	MaxScanLineLength ()
void	FFTImage2D (FRDGBuilder &GraphBuilder, const FGlobalShaderMap *ShaderMap, const FIntPoint &FrequencySize, bool bHorizontalFirst, FRDGTextureSRVRef SrcTexture, const FIntRect &ROIRect, FRDGTextureRef ResultTexture)
void	ConvolutionWithTextureImage2D (FRDGBuilder &GraphBuilder, ERDGPassFlags ComputePassFlags, const FGlobalShaderMap *ShaderMap, const FIntPoint &FrequencySize, bool bHorizontalFirst, FRDGTextureRef TransformedKernel, FRDGTextureSRVRef SrcTexture, const FIntRect &ROIRect, FRDGTextureRef DstTexture, const FIntRect &DstRect, const FPreFilter &PreFilter, FRDGBufferRef PostFilterParameters, ETextureCreateFlags AdditionalTextureCreateFlags)
FIntPoint	Convolution2DBufferSize (const FIntPoint &FrequencySize, const bool bHorizontalFirst, const FIntPoint &SrcExtent)
void	CopyImage2D (FRDGBuilder &GraphBuilder, ERDGPassFlags ComputePassFlags, const FGlobalShaderMap *ShaderMap, FRDGTextureSRVRef SrcTexture, const FIntRect &SrcWindow, FRDGTextureRef DstTexture, const FIntRect &DstWindow, const FPreFilter &PreFilter=FPreFilter(TNumericLimits< float >::Max(), TNumericLimits< float >::Lowest(), 0.f))

Typedef Documentation

FPreFilter

typedef FVector3f GPUFFT::FPreFilter

Define a prefilter to be applied to the pixel Luma, used when forward transforming image data. Used to boost the incoming pixel luma of bright pixels - useful to accentuate the influence of bright pixels during a convolution.

PreFilter.x = MinLuma PreFilter.y = MaxLuma PreFilter.z = Multiplier

When used in TwoForOneRealFFTImage, pixels Luma is limited by MaxLuma and Luma greater than MinLuma is scaled as Multiplier * (Luma - MinLuma) + MinLuma;

Enumeration Type Documentation

FFT_XFORM_TYPE

enum class GPUFFT::FFT_XFORM_TYPE : int

strong

Enumerate the FFT direction and type for up to two dimensions.

Enumerator
FORWARD_HORIZONTAL
FORWARD_VERTICAL
INVERSE_HORIZONTAL
INVERSE_VERTICAL

Function Documentation

BitSize()

uint32 GPUFFT::BitSize

(

uint32

N

)

Computes the minimal number of bits required to represent the in number N

Convolution2DBufferSize()

FIntPoint GPUFFT::Convolution2DBufferSize	(	const FIntPoint &	FrequencySize,
		const bool	bHorizontalFirst,
		const FIntPoint &	SrcExtent
	)

ConvolutionWithTextureImage2D()

void GPUFFT::ConvolutionWithTextureImage2D	(	FRDGBuilder &	GraphBuilder,
		ERDGPassFlags	ComputePassFlags,
		const FGlobalShaderMap *	ShaderMap,
		const FIntPoint &	FrequencySize,
		bool	bHorizontalFirst,
		FRDGTextureRef	TransformedKernel,
		FRDGTextureSRVRef	SrcTexture,
		const FIntRect &	ROIRect,
		FRDGTextureRef	DstTexture,
		const FIntRect &	DstRect,
		const FPreFilter &	PreFilter,
		FRDGBufferRef	PostFilterParameters,
		ETextureCreateFlags	AdditionalTextureCreateFlags
	)

Two Dimensional convolution of an image with a pre-convolved kernel.

Parameters

FrequencySize	- The nominal count of frequencies in both directions.
bHornizontalFirst	- Defines the data layout of the 2d transform in spectral space.
TransformedKernel	- The result of a two dimensional transform of the kernel. Note, the transform order in constructing the TransformedKernel (e.g. bHorizontal first) must match the call here to have compatible data layout.
ROIRect	- The Region Of Interest: Defines the src data in the SrcTexture and the location of the result in the ResultUAV
SrcTexture	- Input float4 Image. Each channel is a seperate signal to be convolved with the corresponding channel in the Kernel
ResultUAV	- UAV buffer that will hold the result.
TmpBuffer0	- Temp float4 buffer with pixelformat PF_A32B32G32R32F
TmpBuffer1	- Temp float4 buffer with pixelformat PF_A32B32G32R32F

NB: The TransformedKernel should have the same frequencies and layout as transforming the SrcTexture in two dimensions ie. FrequencySize + (bHorizontalFirst) ? FIntPoint(2, 0) : (0, 2);

NB: both TmpBuffers must be (bHorizontalFirst) ? FIntPoint(FrequencySize .X + 2, ROIRect.Size().Y) : FIntPoint(ROIRect.Size().X, FrequencySize .Y + 2);

NB: This function does not transition resources on the Src or Target. That must be done by the calling scope.

The Src and Dst textures can be re-used as the tmp buffers (Dst->TmpBuffer0, Src->TmpBuffer1) provided they are large enough.

CopyImage2D()

void GPUFFT::CopyImage2D	(	FRDGBuilder &	GraphBuilder,
		ERDGPassFlags	ComputePassFlags,
		const FGlobalShaderMap *	ShaderMap,
		FRDGTextureSRVRef	SrcTexture,
		const FIntRect &	SrcWindow,
		FRDGTextureRef	DstTexture,
		const FIntRect &	DstWindow,
		const FPreFilter &	PreFilter = FPreFilter(TNumericLimits<float>::Max(), TNumericLimits<float>::Lowest(), 0.f)
	)

Copy a float4 image and possibly amplify the intensity of select pixels.

Parameters

Context	- RHI and ShadersMap
SrcWindow	- Location of the data in the SrcTexture to be copied.
SrcTexture	- The two dimensional float4 input buffer to transform
DstWindow	- Window that images will be copied.
DstUVA	- The destination buffer
PreFilter	- Values use to boost the intensity of any pixels. Default value deactivates prefilter

Data in the SrcWindow is copied into the DstWindow with no-rescaling. If the SrcWindow is smaller then the DstWindow, zeros will pad the result. If the SrcWindow is larger then the DstWindow, data will be lost.

NB: This function does not transition resources. That must be done by the calling scope.

FFTImage2D()

void GPUFFT::FFTImage2D	(	FRDGBuilder &	GraphBuilder,
		const FGlobalShaderMap *	ShaderMap,
		const FIntPoint &	FrequencySize,
		bool	bHorizontalFirst,
		FRDGTextureSRVRef	SrcTexture,
		const FIntRect &	ROIRect,
		FRDGTextureRef	ResultTexture
	)

Two Dimensional transform of an image

Parameters

FrequencySize	- The nominal count of frequencies in both directions.
bHornizontalFirst	- Defines the data layout of the 2d transform in spectral space.
ROIRect	- The Region Of Interest: Defines the src data in the SrcTexture and the location of the result in the ResultUAV
SrcTexture	- Input float4 Image. Each channel is a seperate signal to be convolved with the corresponding channel in the Kernel
ResultBuffer	- UAV buffer that will hold the result.
TmpBuffer0	- Temp float4 buffer with pixelformat PF_A32B32G32R32F

NB: TmpBuffers must be (bHorizontalFirst) ? FIntPoint(FrequencySize .X + 2, ROIRect.Size().Y) : FIntPoint(ROIRect.Size().X, FrequencySize .Y + 2);

NB: This function does not transition resources on the Src or Target. That must be done by the calling scope.

GetFactors()

TArray< uint32 > GPUFFT::GetFactors	(	const uint32	PowTwoLength,
		const uint32	PowTwoBase
	)

Decompose the input PowTwoLenght, as PowTwoLength = N X PowTwoBase X PowTwoBase X .. X PowTwoBase NB: This assumes the PowTwoLength and PowTwoBase are powers of two.

IsForward()

bool GPUFFT::IsForward

(

const FFT_XFORM_TYPE &

XFormType

)

To determine if the transform is inverse or forward

IsHorizontal()

bool GPUFFT::IsHorizontal

(

const FFT_XFORM_TYPE &

XFormType

)

The direction of the associated transform type

MaxScanLineLength()

uint32 GPUFFT::MaxScanLineLength

(

)

The largest power of two length scan line that can be FFT'd with group shared memory. Larger sizes require a much slower (

SwapContents()

void GPUFFT::SwapContents	(	FRDGTextureRef &	TmpBuffer,
		FRDGTextureRef &	DstBuffer
	)