Namespace ManagedCuda.NPP

Classes

JPEGCompression

The JPEG standard defines a flow of level shift, DCT and quantization for forward JPEG transform and inverse level shift, IDCT and de-quantization for inverse JPEG transform. This group has the functions for both forward and inverse functions.

NPPException

Exception thrown in NPP library if a native NPP function returns a negative error code

NPPImage_16sC1

NPPImage_16sC2

NPPImage_16sC3

NPPImage_16sC4

NPPImage_16scC1

NPPImage_16scC2

NPPImage_16scC3

NPPImage_16scC4

NPPImage_16uC1

NPPImage_16uC2

NPPImage_16uC3

NPPImage_16uC4

NPPImage_32fC1

NPPImage_32fC3

NPPImage_32fC4

NPPImage_32fcC1

NPPImage_32fcC2

NPPImage_32fcC3

NPPImage_32fcC4

NPPImage_32sC1

NPPImage_32sC3

NPPImage_32sC4

NPPImage_32scC1

NPPImage_32scC2

NPPImage_32scC3

NPPImage_32scC4

NPPImage_32uC1

NPPImage_32uC4

NPPImage_8sC1

NPPImage_8sC2

NPPImage_8sC3

NPPImage_8sC4

NPPImage_8uC1

NPPImage_8uC2

NPPImage_8uC3

NPPImage_8uC4

NPPImageBase

Abstract base class for derived NPP typed images.

NPPNativeMethods

C# Wrapper-Methods for NPP functions defined in npp.h, nppversion.h, nppcore.h, nppi.h, npps.h, nppdefs.h

NPPNativeMethods.NPPCore

nppcore.h

NPPNativeMethods.NPPi

nppi.h

NPPNativeMethods.NPPi.Abs

Absolute value of each pixel value in an image.

NPPNativeMethods.NPPi.AbsDiff

Pixel by pixel absolute difference between two images.

NPPNativeMethods.NPPi.AbsDiffConst

Determines absolute difference between each pixel of an image and a constant value.

NPPNativeMethods.NPPi.Add

Pixel by pixel addition of two images.

NPPNativeMethods.NPPi.AddConst

Adds a constant value to each pixel of an image.

NPPNativeMethods.NPPi.AddProduct

Pixel by pixel addition of product of pixels from two source images to floating point pixel values of destination image.

NPPNativeMethods.NPPi.AddSquare

Pixel by pixel addition of squared pixels from source image to floating point pixel values of destination image.

NPPNativeMethods.NPPi.AddWeighted

Pixel by pixel addition of alpha weighted pixel values from a source image to floating point pixel values of destination image.

NPPNativeMethods.NPPi.AffinTransforms

Affine warping, affine transform calculation Affine warping of an image is the transform of image pixel positions, defined by the following formulas: \f[ X_{new} = C_{00} * x + C_{01} * y + C_{02} \qquad Y_{new} = C_{10} * x + C_{11} * y + C_{12} \qquad C = \left[ \matrix{C_{00} & C_{01} & C_{02} \cr C_{10} & C_{11} & C_{12} } \right] \f] That is, any pixel with coordinates \f$(X_{new},Y_{new})\f$ in the transformed image is sourced from coordinates \f$(x,y)\f$ in the original image. The mapping \f$C\f$ is completely specified by 6 values \f$C_{ij}, i=\overline{0,1}, j=\overline{0,2}\f$. The transform maps parallel lines to parallel lines and preserves ratios of distances of points to lines. Implementation specific properties are discussed in each function's documentation.

NPPNativeMethods.NPPi.AlphaComp

Composite two images using alpha opacity values contained in each image.

NPPNativeMethods.NPPi.AlphaCompConst

Composite two images using constant alpha values.

NPPNativeMethods.NPPi.AlphaPremul

Premultiplies image pixels by image alpha opacity values.

NPPNativeMethods.NPPi.AlphaPremulConst

Premultiplies pixels of an image using a constant alpha value.

NPPNativeMethods.NPPi.And

Pixel by pixel logical and of images.

NPPNativeMethods.NPPi.AndConst

Pixel by pixel logical and of an image with a constant.

NPPNativeMethods.NPPi.AverageError

Primitives for computing the average error between two images.

Given two images Src1 and Src2 both with width W and height H.

If the image is in complex format, the absolute value is used for computation.

NPPNativeMethods.NPPi.AverageRelativeError

Primitives for computing the average relative error between two images.

If the image is in complex format, the absolute value is used for computation.

NPPNativeMethods.NPPi.BGRToCbYCr

BGR To CbYCr Conversion

NPPNativeMethods.NPPi.BGRToHLS

BGR to HLS

NPPNativeMethods.NPPi.BGRToLab

BGR to LAB

NPPNativeMethods.NPPi.BGRToYCbCr

BGR to YCbCr Conversion

NPPNativeMethods.NPPi.BGRToYCrCb

BGR to YCrCb Conversion

NPPNativeMethods.NPPi.BGRToYUV

BGR to YUV color conversion.

Here is how NPP converts gamma corrected RGB or BGR to YUV.

For digital RGB values in the range [0..255], Y has the range [0..255], U varies in the range [-112..+112], and V in the range [-157..+157].

To fit in the range of [0..255], a constant value of 128 is added to computed U and V values, and V is then saturated.

Npp32f nY = 0.299F * R + 0.587F * G + 0.114F * B;

Npp32f nU = (0.492F * ((Npp32f)nB - nY)) + 128.0F;

Npp32f nV = (0.877F * ((Npp32f)nR - nY)) + 128.0F;

if (nV > 255.0F) nV = 255.0F;

NPPNativeMethods.NPPi.BGRToYUV420

BGR To YUV420 Conversion

NPPNativeMethods.NPPi.BitDepthConversion

Convert bit-depth up and down.

The integer conversion methods do not involve any scaling. Conversions that reduce bit-depth saturate values exceeding the reduced range to the range's maximum/minimum value. When converting from floating-point values to integer values, a rounding mode can be specified. After rounding to integer values the values get saturated to the destination data type's range.

NPPNativeMethods.NPPi.CbYCrToBGR

CbYCrToBGR Conversion

NPPNativeMethods.NPPi.CbYCrToRGB

CbYCr to RGB Conversion

NPPNativeMethods.NPPi.ColorDebayer

Grayscale Color Filter Array to RGB Color Debayer conversion. Generates one RGB color pixel for every grayscale source pixel.

Source and destination images must have even width and height. Missing pixel colors are generated using bilinear interpolation with chroma correlation of generated green values (eInterpolation MUST be set to 0). eGrid allows the user to specify the Bayer grid registration position at source image location oSrcROI.x, oSrcROI.y relative to pSrc. Possible registration positions are:

BGGR | RGGB | GBRG | GRBG

B G || R G || G B || G R

G R || G B || R G || B G

If it becomes necessary to access source pixels outside source image then the source image borders are mirrored.

Here is how the algorithm works. R, G, and B base pixels from the source image are used unmodified. To generate R values for those G pixels, the average of R(x - 1, y) and R(x + 1, y) or R(x, y - 1) and R(x, y + 1) is used depending on whether the left and right or top and bottom pixels are R base pixels. To generate B values for those G pixels, the same algorithm is used using nearest B values. For an R base pixel, if there are no B values in the upper, lower, left, or right adjacent pixels then B is the average of B values in the 4 diagonal (G base) pixels. The same algorithm is used using R values to generate the R value of a B base pixel. Chroma correlation is applied to generated G values only, for a B base pixel G(x - 1, y) and G(x + 1, y) are averaged or G(x, y - 1) and G(x, y + 1) are averaged depending on whether the absolute difference between B(x, y) and the average of B(x - 2, y) and B(x + 2, y) is smaller than the absolute difference between B(x, y) and the average of B(x, y - 2) and B(x, y + 2). For an R base pixel the same algorithm is used testing against the surrounding R values at those offsets. If the horizontal and vertical differences are the same at one of those pixels then the average of the four left, right, upper and lower G values is used instead.

NPPNativeMethods.NPPi.ColorLUT

Perform image color processing using members of various types of color look up tables.

NPPNativeMethods.NPPi.ColorLUTCubic

Perform image color processing using linear interpolation between members of various types of color look up tables.

NPPNativeMethods.NPPi.ColorLUTLinear

Perform image color processing using linear interpolation between members of various types of color look up tables.

NPPNativeMethods.NPPi.ColorLUTPalette

Perform image color processing using various types of bit range restricted palette color look up tables.

NPPNativeMethods.NPPi.ColorLUTTrilinear

Perform image color processing using 3D trilinear interpolation between members of various types of color look up tables.

NPPNativeMethods.NPPi.ColorProcessing

Color manipuliation functions.

NPPNativeMethods.NPPi.ColorToGray

RGB Color to Gray conversion using user supplied conversion coefficients.

Here is how NPP converts gamma corrected RGB Color to Gray using user supplied conversion coefficients.

nGray =  aCoeffs[0] * R + aCoeffs[1] * G + aCoeffs[2] * B;

NPPNativeMethods.NPPi.ColorTwist

Perform color twist pixel processing. Color twist consists of applying the following formula to each image pixel using coefficients from the user supplied color twist host matrix array as follows where dst[x] and src[x] represent destination pixel and source pixel channel or plane x.

dst[0] = aTwist[0][0] * src[0] + aTwist[0][1] * src[1] + aTwist[0][2] * src[2] + aTwist[0][3]

dst[1] = aTwist[1][0] * src[0] + aTwist[1][1] * src[1] + aTwist[1][2] * src[2] + aTwist[1][3]

dst[2] = aTwist[2][0] * src[0] + aTwist[2][1] * src[1] + aTwist[2][2] * src[2] + aTwist[2][3]

NPPNativeMethods.NPPi.ColorTwistBatch

Perform color twist pixel batch processing. Color twist consists of applying the following formula to each image pixel using coefficients from one or more user supplied color twist device memory matrix arrays as follows where dst[x] and src[x] represent destination pixel and source pixel channel or plane x.The full sized coefficient matrix should be sent for all pixel channel sizes, the function will process the appropriate coefficients and channels for the corresponding pixel size. ColorTwistBatch generally takes the same parameter list as ColorTwist except that there is a list of N instances of those parameters (N > 1) and that list is passed in device memory; The matrix pointers referenced for each image in the batch also need to point to device memory matrix values.A convenient data structure is provided that allows for easy initialization of the parameter lists.The only restriction on these functions is that there is one single ROI which is applied respectively to each image in the batch. The primary purpose of this function is to provide improved performance for batches of smaller images as long as GPU resources are available. Therefore it is recommended that the function not be used for very large images as there may not be resources available for processing several large images simultaneously.

NPPNativeMethods.NPPi.Compare

Compare the pixels of two images and create a binary result image. In case of multi-channel image types, the condition must be fulfilled for all channels, otherwise the comparison is considered false.

The "binary" result image is of type 8u_C1. False is represented by 0, true by NPP_MAX_8U.

NPPNativeMethods.NPPi.CompColorKey

Composition color key

NPPNativeMethods.NPPi.ComplexImageMorphology

Complex image morphological operations.

NPPNativeMethods.NPPi.CompressionDCT

Image compression primitives.

NPPNativeMethods.NPPi.Convolution

General purpose 2D convolution filters.

NPPNativeMethods.NPPi.CopyConstBorder

Methods for copying images and padding borders with a constant, user-specifiable color.

NPPNativeMethods.NPPi.CopyReplicateBorder

Methods for copying images and padding borders with a replicates of the nearest source image pixel color.

NPPNativeMethods.NPPi.CopySubpix

Functions for copying linearly interpolated images using source image subpixel coordinates

NPPNativeMethods.NPPi.CopyWrapBorder

Methods for copying images and padding borders with wrapped replications of the source image pixel colors.

NPPNativeMethods.NPPi.CountInRange

Primitives for computing the amount of pixels that fall into the specified intensity range. The lower bound and the upper bound are inclusive.

NPPNativeMethods.NPPi.Dilate3x3Border

Dilation using a 3x3 mask with the anchor at its center pixel with border control.

If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

NPPNativeMethods.NPPi.DilationWithBorderControl

Dilation with border control. Dilation computes the output pixel as the maximum pixel value of the pixels under the mask. Pixels who's corresponding mask values are zero do not participate in the maximum search.

If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

NPPNativeMethods.NPPi.Div

Pixel by pixel division of two images.

NPPNativeMethods.NPPi.DivConst

Divides each pixel of an image by a constant value.

NPPNativeMethods.NPPi.DivRound

Pixel by pixel division of two images using result rounding modes.

NPPNativeMethods.NPPi.DotProd

Primitives for computing the dot product of two images.

NPPNativeMethods.NPPi.Dup

Functions for duplicating a single channel image in a multiple channel image.

NPPNativeMethods.NPPi.Erode3x3Border

Erosion using a 3x3 mask with the anchor at its center pixel with border control.

If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

NPPNativeMethods.NPPi.ErosionWithBorderControl

Erosion computes the output pixel as the minimum pixel value of the pixels under the mask. Pixels who's corresponding mask values are zero do not participate in the minimum search.

If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

NPPNativeMethods.NPPi.Exp

Exponential value of each pixel in an image.

NPPNativeMethods.NPPi.FilterBilateralGaussBorder

Filters the image using a bilateral Gaussian filter kernel with border control:

If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

For this filter the anchor point is always the central element of the kernel.

Coefficients of the bilateral filter kernel depend on their position in the kernel and on the value of some source image pixels overlayed by the filter kernel.

Only source image pixels with both coordinates divisible by nDistanceBetweenSrcPixels are used in calculations. The value of an output pixel \f$d\f$ is

\f[d = \frac{\sum_{h=-nRadius}^{nRadius}\sum_{w=-nRadius}^{nRadius}W1(h,w)\cdot W2(h,w)\cdot S(h,w)}{\sum_{h=-nRadius}^{nRadius}\sum_{w=-nRadius}^{nRadius}W1(h,w)\cdot W2(h,w)}\f] where h and w are the corresponding kernel width and height indexes, S(h,w) is the value of the source image pixel overlayed by filter kernel position (h,w), W1(h,w) is func(nValSquareSigma, (S(h,w) - S(0,0))) where S(0,0) is the value of the source image pixel at the center of the kernel, W2(h,w) is func(nPosSquareSigma, sqrt(hh+ww)), and func is the following formula \f[func(S,I) = exp(-\frac{I^2}{2.0F\cdot S^2})\f]

Currently only the NPP_BORDER_REPLICATE border type operations are supported.

NPPNativeMethods.NPPi.FilterBorder

General purpose 2D convolution filter with border control.

Pixels under the mask are multiplied by the respective weights in the mask and the results are summed. Before writing the result pixel the sum is scaled back via division by nDivisor. If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

Currently only the Replicate and Mirror border type operation are supported.

NPPNativeMethods.NPPi.FilterBorder32f

General purpose 2D convolution filter using floating-point weights with border control.

Currently only the Replicate and Mirror border type operation are supported.

NPPNativeMethods.NPPi.FilterCannyBorder

Performs Canny edge detection on a single channel 8-bit grayscale image and outputs a single channel 8-bit image consisting of 0x00 and 0xFF values with 0xFF representing edge pixels. The algorithm consists of three phases. The first phase generates two output images consisting of a single channel 16-bit signed image containing magnitude values and a single channel 32-bit floating point image containing the angular direction of those magnitude values. This phase is accomplished by calling the appropriate GradientVectorBorder filter function based on the filter type, filter mask size, and norm type requested. The next phase uses those magnitude and direction images to suppress non-maximum magnitude values which are lower than the values of either of its two nearest neighbors in the same direction as the test magnitude pixel in the 3x3 surrounding magnitude pixel neighborhood. This phase outputs a new magnitude image with non-maximum pixel values suppressed. Finally, in the third phase, the new magnitude image is passed through a hysteresis threshold filter that filters out any magnitude values that are not connected to another edge magnitude value. In this phase, any magnitude value above the high threshold value is automatically accepted, any magnitude value below the low threshold value is automatically rejected. For magnitude values that lie between the low and high threshold, values are only accepted if one of their two neighbors in the same direction in the 3x3 neighborhood around them lies above the low threshold value. In other words, if they are connected to an active edge. J. Canny recommends that the ratio of high to low threshold limit be in the range two or three to one, based on predicted signal-to-noise ratios. The final output of the third phase consists of a single channel 8-bit unsigned image of 0x00 and 0xFF values based on whether they are accepted or rejected during threshold testing.

Currently only the NPP_BORDER_REPLICATE border type operation is supported. Borderless output can be accomplished by using a larger source image than the destination and adjusting oSrcSize and oSrcOffset parameters accordingly.

NPPNativeMethods.NPPi.FilterGaussBorder

If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image.

Currently only the NPP_BORDER_REPLICATE and NPP_BORDER_MIRROR border type operation is supported.

Filters the image using a Gaussian filter kernel:

1/16 2/16 1/16

2/16 4/16 2/16

1/16 2/16 1/16

2/571 7/571 12/571 7/571 2/571

7/571 31/571 52/571 31/571 7/571

12/571 52/571 127/571 52/571 12/571

7/571 31/571 52/571 31/571 7/571

2/571 7/571 12/571 7/571 2/571

NPPNativeMethods.NPPi.FilterGaussPyramid

Filters the image using a separable Gaussian filter kernel with user supplied floating point coefficients with downsampling and border control. If the downsampling rate is equivalent to an integer value then unnecessary source pixels are just skipped. If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image. Filters the image using a separable Gaussian filter kernel with user supplied floating point coefficients with upsampling and border control. If the upsampling rate is equivalent to an integer value then unnecessary source pixels are just skipped. If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image. Currently only the Replicate and Mirror border type operation are supported.

NPPNativeMethods.NPPi.FilterHarrisCornersBorder

Performs Harris Corner detection on a single channel 8-bit grayscale image and outputs a single channel 32-bit floating point image consisting the corner response at each pixel of the image. The algorithm consists of two phases. The first phase generates the floating point product of XX, YY, and XY gradients at each pixel in the image. The type of gradient used is controlled by the eFilterType and eMaskSize parameters. The second phase averages those products over a window of either 3x3 or 5x5 pixels around the center pixel then generates the Harris corner response at that pixel which is output in the destination image. The Harris response value is determined as H = ((XX///YY - XY///XY) - (nK///((XX + YY)///(XX + YY))))///nScale.

NPPNativeMethods.NPPi.FilterHoughLine

Extracts Hough lines from a single channel 8-bit binarized (0, 255) source feature (canny edges, etc.) image and outputs a list of lines in point polar format representing the length(rho) and angle(theta) of each line from the origin of the normal to the line using the formula rho = x cos(theta) + y sin(theta). The level of discretization, nDelta, is specified as an input parameter.The performance and effectiveness of this function highly depends on this parameter with higher performance for larger numbers and more detailed results for lower numbers. Also, lines are not guaranteed to be added to the pDeviceLines list in the same order from one call to the next. However, all of the same lines will still be generated as long as nMaxLineCount is set large enough so that they all can fit in the list. To convert lines in point polar format back to cartesian lines use the following formula:

NPPNativeMethods.NPPi.FilterScharrHorizBorder

Filters the image using a horizontal Scharr filter kernel with border control:

3 10 3

0 0 0

-3 -10 -3

NPPNativeMethods.NPPi.FilterScharrVertBorder

Filters the image using a vertical Scharr filter kernel with border control:

3 10 3

0 0 0

-3 -10 -3

NPPNativeMethods.NPPi.FilterSobelCrossBorder

Filters the image using a second cross derivative Sobel filter kernel with border control.

NPPNativeMethods.NPPi.FilterSobelHorizBorder

Filters the image using a horizontal Sobel filter kernel with border control.

NPPNativeMethods.NPPi.FilterSobelHorizSecondBorder

Filters the image using a second derivative, horizontal Sobel filter kernel with border control.

NPPNativeMethods.NPPi.FilterSobelVertBorder

Filters the image using a vertical Sobel filter kernel with border control.

NPPNativeMethods.NPPi.FilterSobelVertSecondBorder

Filters the image using a second derivative, vertical Sobel filter kernel with border control.

NPPNativeMethods.NPPi.FilterThresholdAdaptiveBoxBorder

Computes the average pixel values of the pixels under a square mask with border control. If any portion of the mask overlaps the source image boundary the requested border type operation is applied to all mask pixels which fall outside of the source image. Once the neighborhood average around a source pixel is determined the souce pixel is compared to the average - nDelta and if the source pixel is greater than that average the corresponding destination pixel is set to nValGT, otherwise nValLE. Currently only the NPP_BORDER_REPLICATE border type operation is supported.

NPPNativeMethods.NPPi.FilterWienerBorder

Noise removal filtering of an image using an adaptive Wiener filter with border control.

Pixels under the source mask are used to generate statistics about the local neighborhood which are then used to control the amount of adaptive noise filtering locally applied.

Currently only the NPP_BORDER_REPLICATE border type operation is supported.

NPPNativeMethods.NPPi.FixedFilters

Fixed filters perform linear filtering operations (i.e. convolutions) with predefined kernels of fixed sizes.

NPPNativeMethods.NPPi.Gamma

Gamma correction

NPPNativeMethods.NPPi.GeometricTransforms

Routines manipulating an image's geometry.

NPPNativeMethods.NPPi.GetResizeRect

Returns NppiRect which represents the offset and size of the destination rectangle that would be generated by resizing the source NppiRect by the requested scale factors and shifts.

NPPNativeMethods.NPPi.GradientColorToGray

RGB Color to Gray Gradient conversion using user selected gradient distance method.

NPPNativeMethods.NPPi.GradientVectorPrewittBorder

RGB Color to Prewitt Gradient Vector conversion using user selected fixed mask size and gradient distance method. Functions support up to 4 optional single channel output gradient vectors, X (vertical), Y (horizontal), magnitude, and angle with user selectable distance methods. Output for a particular vector is disabled by supplying a NULL pointer for that vector. X and Y gradient vectors are in cartesian form in the destination data type.
Magnitude vectors are polar gradient form in the destination data type, angle is always in floating point polar gradient format. Only fixed mask sizes of 3x3 are supported. Only nppiNormL1 (sum) and nppiNormL2 (sqrt of sum of squares) distance methods are currently supported.

For the C1R versions of the function the pDstMag output image value for L1 normalization consists of the absolute value of the pDstX value plus the absolute value of the pDstY value at that particular image pixel location. For the C1R versions of the function the pDstMag output image value for L2 normalization consists of the square root of the pDstX value squared plus the pDstY value squared at that particular image pixel location. For the C1R versions of the function the pDstAngle output image value consists of the arctangent (atan2) of the pDstY value and the pDstX value at that particular image pixel location.

For the C3C1R versions of the function, regardless of the selected normalization method, the L2 normalization value is first determined for each or the pDstX and pDstY values for each source channel then the largest L2 normalization value (largest gradient) is used to select which of the 3 pDstX channel values are output to the pDstX image or pDstY channel values are output to the pDstY image. For the C3C1R versions of the function the pDstMag output image value for L1 normalizaton consists of the same technique used for the C1R version for each source image channel. Then the largest L2 normalization value is again used to select which of the 3 pDstMag channel values to output to the pDstMag image. For the C3C1R versions of the function the pDstMag output image value for L2 normalizaton consists of just outputting the largest per source channel L2 normalization value to the pDstMag image. For the C3C1R versions of the function the pDstAngle output image value consists of the same technique used for the C1R version calculated for each source image channel. Then the largest L2 normalization value is again used to select which of the 3 angle values to output to the pDstAngle image.

NPPNativeMethods.NPPi.GradientVectorScharrBorder

RGB Color to Scharr Gradient Vector conversion using user selected fixed mask size and gradient distance method. Functions support up to 4 optional single channel output gradient vectors, X (vertical), Y (horizontal), magnitude, and angle with user selectable distance methods. Output for a particular vector is disabled by supplying a NULL pointer for that vector. X and Y gradient vectors are in cartesian form in the destination data type.
Magnitude vectors are polar gradient form in the destination data type, angle is always in floating point polar gradient format. Only fixed mask sizes of 3x3 are supported. Only nppiNormL1 (sum) and nppiNormL2 (sqrt of sum of squares) distance methods are currently supported.

NPPNativeMethods.NPPi.GradientVectorSobelBorder

RGB Color to Sobel Gradient Vector conversion using user selected fixed mask size and gradient distance method. Functions support up to 4 optional single channel output gradient vectors, X (vertical), Y (horizontal), magnitude, and angle with user selectable distance methods. Output for a particular vector is disabled by supplying a NULL pointer for that vector. X and Y gradient vectors are in cartesian form in the destination data type.
Magnitude vectors are polar gradient form in the destination data type, angle is always in floating point polar gradient format. Only fixed mask sizes of 3x3 and 5x5 are supported. Only nppiNormL1 (sum) and nppiNormL2 (sqrt of sum of squares) distance methods are currently supported.

NPPNativeMethods.NPPi.Histogram

Histogram.

NPPNativeMethods.NPPi.HistogramOfOrientedGradientsBorder

Performs Histogram Of Oriented Gradients operation on source image generating separate windows of Histogram Descriptors for each requested location.

This function implements the simplest form of functionality described by N.Dalal and B.Triggs.Histograms of Oriented Gradients for Human Detection.INRIA, 2005.

It supports overlapped contrast normalized block histogram output with L2 normalization only, no threshold clipping, and no pre or post gaussian smoothing of input images or histogram output values. It supports both single channel grayscale source images and three channel color images.For color images, the color channel with the highest magnitude value is used as that pixel's magnitude. Output is row order only. Descriptors are output consecutively with no separation padding if multiple descriptor output is requested (one desriptor per source image location). For example, common HOG parameters are 9 histogram bins per 8 by 8 pixel cell, 2 by 2 cells per block, with a descriptor window size of 64 horizontal by 128 vertical pixels yielding 7 by 15 overlapping blocks (1 cell overlap in both horizontal and vertical directions). This results in 9 bins* 4 cells* 7 horizontal overlapping blocks* 15 vertical overlapping blocks or 3780 32-bit floating point output values(bins) per descriptor window.

The number of horizontal overlapping block histogram bins per descriptor window width is determined by (((oHOGConfig.detectionWindowSize.width / oHOGConfig.histogramBlockSize) * 2) - 1) * oHOGConfig.nHistogramBins. The number of vertical overlapping block histograms per descriptor window height is determined by (((oHOGConfig.detectionWindowSize.height / oHOGConfig.histogramBlockSize) * 2) - 1) The offset of each descriptor window in the descriptors output buffer is therefore horizontal histogram bins per descriptor window width * vertical histograms per descriptor window height 32-bit floating point values relative to the previous descriptor window output.

The algorithm uses a 1D centered derivative mask of[-1, 0, +1] when generating input magnitude and angle gradients. Magnitudes are added to the two nearest histogram bins of oriented gradients between 0 and 180 degrees using a weighted linear interpolation of each magnitude value across the 2 nearest angular bin orientations. 2D overlapping blocks of histogram bins consisting of the bins from 2D arrangements of cells are then contrast normalized using L2 normalization and output to the corresponding histogram descriptor window for that particular window location in the window locations list.

NPPNativeMethods.NPPi.HLSToBGR

HLS to BGR

NPPNativeMethods.NPPi.HLSToRGB

HLS to RGB

NPPNativeMethods.NPPi.HSVToRGB

HSV to RGB

NPPNativeMethods.NPPi.ImageCompression

Image compression primitives.

NPPNativeMethods.NPPi.ImageMedianFilter

Result pixel value is the median of pixel values under the rectangular mask region.

NPPNativeMethods.NPPi.ImageProximity

Primitives for computing the proximity measure between a source image and a template image.

NPPNativeMethods.NPPi.Integral

Primitives for computing the integral image of a given image.

NPPNativeMethods.NPPi.IQA

Primitives for computing the image quality between two images, such as MSE, PSNR, SSIM, and MS-SSIM.

NPPNativeMethods.NPPi.LabelMarkers

Generate image connected region label markers to be used for later image segmentation.

NPPNativeMethods.NPPi.LabToBGR

LAB to BGR

NPPNativeMethods.NPPi.LeftShiftConst

Pixel by pixel left shift of an image by a constant value.

NPPNativeMethods.NPPi.LinearFilter1D

1D mask Linear Convolution Filter, with rescaling, for 8 bit images.

NPPNativeMethods.NPPi.LinearFixedFilters2D

2D linear fixed filters for 8 bit images.

NPPNativeMethods.NPPi.LinearTransforms

Linear image transforms, like Fourier and DCT transformations.

NPPNativeMethods.NPPi.Ln

Pixel by pixel natural logarithm of each pixel in an image.

NPPNativeMethods.NPPi.LUVToRGB

LUV to RGB

NPPNativeMethods.NPPi.Max

Maximum

NPPNativeMethods.NPPi.MaxIdx

Maximum Index

NPPNativeMethods.NPPi.MaximumError

Primitives for computing the maximum error between two images.

Given two images Src1 and Src2 both with width W and height H, the maximum error is defined as the largest absolute difference between pixels of two images.

If the image is in complex format, the absolute value of the complex number is provided.

NPPNativeMethods.NPPi.MaximumRelativeError

Primitives for computing the maximum relative error between two images.

If the image is in complex format, the absolute value is used for computation.

For multiple channles, the maximum relative error of all the channles is returned.

NPPNativeMethods.NPPi.MeanNew

Mean (new in CUDA 5)

NPPNativeMethods.NPPi.MeanStdDevNew

Mean + Std deviation (new in CUDA 5)

NPPNativeMethods.NPPi.MemAlloc

Image-Memory Allocation

ImageAllocator methods for 2D arrays of data. The allocators have width and height parameters to specify the size of the image data being allocated. They return a pointer to the newly created memory and return the numbers of bytes between successive lines.

If the memory allocation failed due to lack of free device memory or device memory fragmentation the routine returns 0.

All allocators return memory with line strides that are beneficial for performance. It is not mandatory to use these allocators. Any valid CUDA device-memory pointers can be used by the NPP primitives and there are no restrictions on line strides.

NPPNativeMethods.NPPi.MemCopy

Copy methods for images of various types. Images are passed to these primitives via a pointer to the image data (first pixel in the ROI) and a step-width, i.e. the number of bytes between successive lines.

The size of the area to be copied (region-of-interest, ROI) is specified via a Size struct.

NPPNativeMethods.NPPi.MemSet

Set methods for images of various types. Images are passed to these primitives via a pointer to the image data (first pixel in the ROI) and a step-width, i.e. the number of bytes between successive lines. The size of the area to be set (region-of-interest, ROI) is specified via a Size struct.

In addition to the image data and ROI, all methods have a parameter to specify the value being set. In case of single channel images this is a single value, in case of multi-channel, an array of values is passed.

NPPNativeMethods.NPPi.Min

Minimum

NPPNativeMethods.NPPi.MinIdx

Minimum index

NPPNativeMethods.NPPi.MinMaxEvery

Primitives for computing the minimal/maximal value of the pixel pair from two images.

NPPNativeMethods.NPPi.MinMaxIndxNew

Min / Max Index (new in CUDA 5)

NPPNativeMethods.NPPi.MinMaxNew

Min / Max (new in CUDA 5)

NPPNativeMethods.NPPi.MorphologyFilter2D

Image dilate and erod operations.

NPPNativeMethods.NPPi.Mul

Pixel by pixel multiply of two images.

NPPNativeMethods.NPPi.MulConst

Multiplies each pixel of an image by a constant value.

NPPNativeMethods.NPPi.MulConstScale

Multiplies each pixel of an image by a constant value then scales the result by the maximum value for the data bit width.

NPPNativeMethods.NPPi.MulScale

Pixel by pixel multiplies each pixel of two images then scales the result by the maximum value for the data bit width.

NPPNativeMethods.NPPi.NormDiff

Norm of pixel differences between two images.

NPPNativeMethods.NPPi.NormInf

Infinite Norm

NPPNativeMethods.NPPi.NormL1

L1 Norm

NPPNativeMethods.NPPi.NormL2

L2 Norm

NPPNativeMethods.NPPi.NormRel

Primitives for computing the relative error between two images.

NPPNativeMethods.NPPi.Not

Pixel by pixel logical not of image.

NPPNativeMethods.NPPi.NV12ToYUV420

NV12 to YUV420 color conversion.

NPPNativeMethods.NPPi.NV21ToBGR

NV21 to BGR color conversion

NPPNativeMethods.NPPi.NV21ToRGB

NV21 to RGB color conversion

NPPNativeMethods.NPPi.Or

Pixel by pixel logical or of images.

NPPNativeMethods.NPPi.OrConst

Pixel by pixel logical or of an image with a constant.

NPPNativeMethods.NPPi.PerspectiveTransforms

Perspective warping, perspective transform calculation Perspective warping of an image is the transform of image pixel positions, defined by the following formulas: \f[ X_{new} = \frac{C_{00} * x + C_{01} * y + C_{02}}{C_{20} * x + C_{21} * y + C_{22}} \qquad Y_{new} = \frac{C_{10} * x + C_{11} * y + C_{12}}{C_{20} * x + C_{21} * y + C_{22}} \qquad C = \left[ \matrix{C_{00} & C_{01} & C_{02} \cr C_{10} & C_{11} & C_{12} \cr C_{20} & C_{21} & C_{22} } \right] \f] That is, any pixel of the transformed image with coordinates \f$(X_{new},Y_{new})\f$ has a preimage with coordinates \f$(x,y)\f$. The mapping \f$C\f$ is fully defined by 8 values \f$C_{ij}, (i,j)=\overline{0,2}\f$, except of \f$C_{22}\f$, which is a normalizer. The transform has a property of mapping any convex quadrangle to a convex quadrangle, which is used in a group of functions nppiWarpPerspectiveQuad. The NPPI implementation of perspective transform has some issues which are discussed in each function's documentation.

NPPNativeMethods.NPPi.QualityIndex

Primitives for computing the image quality index of two images.

NPPNativeMethods.NPPi.RankFilters

Min, Median, and Max image filters.

NPPNativeMethods.NPPi.Remap

Remap chooses source pixels using pixel coordinates explicitely supplied in two 2D device memory image arrays pointed to by the pXMap and pYMap pointers.

The pXMap array contains the X coordinated and the pYMap array contains the Y coordinate of the corresponding source image pixel to use as input. These coordinates are in floating point format so fraction pixel positions can be used. The coordinates of the source pixel to sample are determined as follows:

nSrcX = pxMap[nDstX, nDstY]

nSrcY = pyMap[nDstX, nDstY]

In the Remap functions below source image clip checking is handled as follows:

If the source pixel fractional x and y coordinates are greater than or equal to oSizeROI.x and less than oSizeROI.x + oSizeROI.width and greater than or equal to oSizeROI.y and less than oSizeROI.y + oSizeROI.height then the source pixel is considered to be within the source image clip rectangle and the source image is sampled. Otherwise the source image is not sampled and a destination pixel is not written to the destination image.

NPPNativeMethods.NPPi.ResizeSqrPixel

Resizes images.

NPPNativeMethods.NPPi.RGBToCbYCr

RGB To CbYCr Conversion

NPPNativeMethods.NPPi.RGBToGray

RGB to CCIR601 Gray conversion.

Here is how NPP converts gamma corrected RGB to CCIR601 Gray.

nGray =  0.299F * R + 0.587F * G + 0.114F * B;

NPPNativeMethods.NPPi.RGBToHLS

RGB to HLS

NPPNativeMethods.NPPi.RGBToHSV

RGB to HSV

NPPNativeMethods.NPPi.RGBToLUV

RGB to LUV

NPPNativeMethods.NPPi.RGBToXYZ

RGB to XYZ

NPPNativeMethods.NPPi.RGBToYCbCr

RGB to YCbCr color conversion.

NPPNativeMethods.NPPi.RGBToYCbCr_JPEG

JPEG RGB to YCbCr color conversion.

NPPNativeMethods.NPPi.RGBToYCC

RGB to YCC

NPPNativeMethods.NPPi.RGBToYCrCb

RGB To YCrCB Conversion

NPPNativeMethods.NPPi.RGBToYUV

RGB to YUV Conversion

NPPNativeMethods.NPPi.RGBToYUV420

RGB to YUV420 Conversion

NPPNativeMethods.NPPi.RGBToYUV422

RGB To YUV422 Conversion

NPPNativeMethods.NPPi.RightShiftConst

Pixel by pixel right shift of an image by a constant value.

NPPNativeMethods.NPPi.SamplePatternConversion

Sample Pattern Conversion.

NPPNativeMethods.NPPi.Scale

Scale bit-depth up and down.

To map source pixel srcPixelValue to destination pixel dstPixelValue the following equation is used:

dstPixelValue = dstMinRangeValue + scaleFactor * (srcPixelValue - srcMinRangeValue)

where scaleFactor = (dstMaxRangeValue - dstMinRangeValue) / (srcMaxRangeValue - srcMinRangeValue).

For conversions between integer data types, the entire integer numeric range of the input data type is mapped onto the entire integer numeric range of the output data type.

For conversions to floating point data types the floating point data range is defined by the user supplied floating point values of nMax and nMin which are used as the dstMaxRangeValue and dstMinRangeValue respectively in the scaleFactor and dstPixelValue calculations and also as the saturation values to which output data is clamped.

When converting from floating-point values to integer values, nMax and nMin are used as the srcMaxRangeValue and srcMinRangeValue respectively in the scaleFactor and dstPixelValue calculations. Output values are saturated and clamped to the full output integer pixel value range.

NPPNativeMethods.NPPi.Sqr

Square each pixel in an image.

NPPNativeMethods.NPPi.Sqrt

Pixel by pixel square root of each pixel in an image.

NPPNativeMethods.NPPi.Sub

Pixel by pixel subtraction of two images.

NPPNativeMethods.NPPi.SubConst

Subtracts a constant value from each pixel of an image.

NPPNativeMethods.NPPi.Sum

Sum of 8 bit images.

NPPNativeMethods.NPPi.SwapChannel

Methods for exchanging the color channels of an image.

The methods support arbitrary permutations of the original channels, including replication.

NPPNativeMethods.NPPi.Threshold

Threshold pixels.

NPPNativeMethods.NPPi.Transpose

Methods for transposing images of various types. Like matrix transpose, image transpose is a mirror along the image's diagonal (upper-left to lower-right corner).

NPPNativeMethods.NPPi.WindowSum1D

1D mask Window Sum for 8 bit images.

NPPNativeMethods.NPPi.Xor

Pixel by pixel logical exclusive or of images.

NPPNativeMethods.NPPi.XorConst

Pixel by pixel logical exclusive or of an image with a constant.

NPPNativeMethods.NPPi.XYZToRGB

XYZ to RGB

NPPNativeMethods.NPPi.YCbCrAndACrCbAndOther

YCbCr ACrCb ...

NPPNativeMethods.NPPi.YCbCrToBGR

YCbCr To BGR Conversion

NPPNativeMethods.NPPi.YCbCrToRGB

YCbCr to RGB color conversion.

NPPNativeMethods.NPPi.YCCToRGB

YCC to RGB

NPPNativeMethods.NPPi.YCrCbToRGB

YCrCB to RGB Conversion

NPPNativeMethods.NPPi.YUV420ToBGR

YUV420 to BGR Conversion

NPPNativeMethods.NPPi.YUV420ToRGB

YUV420 to RGB Conversion

NPPNativeMethods.NPPi.YUV422ToRGB

YUV422 To RGB Conversion

NPPNativeMethods.NPPi.YUVToBGR

YUV to BGR color conversion.

Here is how NPP converts YUV to gamma corrected RGB or BGR.

Npp32f nY = (Npp32f)Y;

Npp32f nU = (Npp32f)U - 128.0F;

Npp32f nV = (Npp32f)V - 128.0F;

Npp32f nR = nY + 1.140F * nV;

if (nR < 0.0F) nR = 0.0F;

if (nR > 255.0F) nR = 255.0F;

Npp32f nG = nY - 0.394F * nU - 0.581F * nV;

if (nG < 0.0F) nG = 0.0F;

if (nG > 255.0F) nG = 255.0F;

Npp32f nB = nY + 2.032F * nU;

if (nB < 0.0F) nB = 0.0F;

if (nB > 255.0F) nB = 255.0F;

NPPNativeMethods.NPPi.YUVToRGB

YUV to RGB Conversion

NPPNativeMethods.NPPs

npps.h

NPPNativeMethods.NPPs.AbsoluteValueSignal

Absolute value of each sample of a signal.

NPPNativeMethods.NPPs.AddC

Adds a constant value to each sample of a signal.

NPPNativeMethods.NPPs.AddProductC

Adds product of a constant and each sample of a source signal to the each sample of destination signal.

NPPNativeMethods.NPPs.AddProductSignal

Adds sample by sample product of two signals to the destination signal.

NPPNativeMethods.NPPs.AddSignal

Sample by sample addition of two signals.

NPPNativeMethods.NPPs.And

Sample by sample bitwise AND of samples from two signals.

NPPNativeMethods.NPPs.AndC

Bitwise AND of a constant and each sample of a signal.

NPPNativeMethods.NPPs.AverageError

Primitives for computing the Average error between two signals.

NPPNativeMethods.NPPs.AverageRelativeError

Primitives for computing the AverageRelative error between two signals.

NPPNativeMethods.NPPs.Cauchy

Determine Cauchy robust error function and its first and second derivatives for each sample of a signal.

NPPNativeMethods.NPPs.Convert

Routines for converting the sample-data type of signals.

NPPNativeMethods.NPPs.Copy

Copy methods for various type signals. Copy methods operate on signal data given as a pointer to the underlying data-type (e.g. 8-bit vectors would be passed as pointers to Npp8u type) and length of the vectors, i.e. the number of items.

NPPNativeMethods.NPPs.CountInRange

Count In Range

NPPNativeMethods.NPPs.CubeRootSignal

Cube root of each sample of a signal.

NPPNativeMethods.NPPs.DivC

Divides each sample of a signal by a constant.

NPPNativeMethods.NPPs.DivCRev

Divides a constant by each sample of a signal.

NPPNativeMethods.NPPs.DivRoundSignal

Sample by sample division of the samples of two signals with rounding.

NPPNativeMethods.NPPs.DivSignal

Sample by sample division of the samples of two signals.

NPPNativeMethods.NPPs.DotProduct

Dot Product

NPPNativeMethods.NPPs.ExponentSignal

e raised to the power of each sample of a signal.

NPPNativeMethods.NPPs.FilteringFunctions

Functions that provide functionality of generating output signal based on the input signal like signal integral, etc.

NPPNativeMethods.NPPs.InverseTangentSignal

Inverse tangent of each sample of a signal.

NPPNativeMethods.NPPs.LShiftC

Left shifts the bits of each sample of a signal by a constant amount.

NPPNativeMethods.NPPs.Max

Functions that provide global signal statistics like: average, standard deviation, minimum, etc.

NPPNativeMethods.NPPs.MaximumError

Primitives for computing the maximum error between two signals.

NPPNativeMethods.NPPs.MaximumRelativeError

Primitives for computing the MaximumRelative error between two signals.

NPPNativeMethods.NPPs.MeanStdDev

Mean and StdDev

NPPNativeMethods.NPPs.MemAlloc

Signal-allocator methods for allocating 1D arrays of data in device memory. All allocators have size parameters to specify the size of the signal (1D array) being allocated.

The allocator methods return a pointer to the newly allocated memory of appropriate type. If device-memory allocation is not possible due to resource constaints the allocators return 0 (i.e. NULL pointer).

All signal allocators allocate memory aligned such that it is beneficial to the performance of the majority of the signal-processing primitives. It is no mandatory however to use these allocators. Any valid CUDA device-memory pointers can be passed to NPP primitives.

NPPNativeMethods.NPPs.Min

Functions that provide global signal statistics like: average, standard deviation, minimum, etc.

NPPNativeMethods.NPPs.MinMaxEvery

Performs the min or max operation on the samples of a signal.

NPPNativeMethods.NPPs.MinMaxIndex

Minimum / Maximum values / indices

NPPNativeMethods.NPPs.MulC

Multiplies each sample of a signal by a constant value.

NPPNativeMethods.NPPs.MulSignal

Sample by sample multiplication the samples of two signals.

NPPNativeMethods.NPPs.NaturalLogarithmSignal

Natural logarithm of each sample of a signal.

NPPNativeMethods.NPPs.Norm

Infinity Norm, L1 Norm, L2 Norm

NPPNativeMethods.NPPs.NormalizeSignal

Normalize each sample of a real or complex signal using offset and division operations.

NPPNativeMethods.NPPs.NormDiff

Infinity Norm Diff, L1 Norm Diff, L2 Norm Diff

NPPNativeMethods.NPPs.Not

Bitwise NOT of each sample of a signal.

NPPNativeMethods.NPPs.Or

Sample by sample bitwise OR of samples from two signals.

NPPNativeMethods.NPPs.OrC

Bitwise OR of a constant and each sample of a signal.

NPPNativeMethods.NPPs.RShiftC

Right shifts the bits of each sample of a signal by a constant amount.

NPPNativeMethods.NPPs.Set

Set methods for 1D vectors of various types. The copy methods operate on vector data given as a pointer to the underlying data-type (e.g. 8-bit vectors would be passed as pointers to Npp8u type) and length of the vectors, i.e. the number of items.

NPPNativeMethods.NPPs.SquareRootSignal

Square root of each sample of a signal.

NPPNativeMethods.NPPs.SquareSignal

Squares each sample of a signal.

NPPNativeMethods.NPPs.SubC

Subtracts a constant from each sample of a signal.

NPPNativeMethods.NPPs.SubCRev

Subtracts each sample of a signal from a constant.

NPPNativeMethods.NPPs.SubSignal

Sample by sample subtraction of the samples of two signals.

NPPNativeMethods.NPPs.Sum

Functions that provide global signal statistics like: average, standard deviation, minimum, etc.

NPPNativeMethods.NPPs.SumLn

Sums up the natural logarithm of each sample of a signal.

NPPNativeMethods.NPPs.TenTimesBaseTenLogarithmSignal

Ten times the decimal logarithm of each sample of a signal.

NPPNativeMethods.NPPs.Threshold

Performs the threshold operation on the samples of a signal by limiting the sample values by a specified constant value.

NPPNativeMethods.NPPs.Xor

Sample by sample bitwise XOR of samples from two signals.

NPPNativeMethods.NPPs.XorC

Bitwise XOR of a constant and each sample of a signal.

NPPNativeMethods.NPPs.Zero

Set signals to zero.

NPPNativeMethods.NPPs.ZeroCrossing

Count Zero Crossings

NPPWarning

WarningException thrown if configured and a native NPP function returns a positive error code

NPPWarningHandler

Singleton NPPWarning handler. Use the OnNPPWarning event to get notified when a NPP functions returns a NPP warning status code.

NPPWarningHandler.NPPWarningEventArgs

NPP warning event args

Structs

HaarBuffer

HaarClassifier

Npp16sc

Complex Number.

This struct represents a short complex number.

Npp32fc

Complex Number.

This struct represents a single floating-point complex number.

Npp32sc

Complex Number.

This struct represents a signed int complex number.

Npp64fc

Complex Number.

This struct represents a double floating-point complex number.

Npp64sc

Complex Number.

This struct represents a long long complex number.

NppiColorTwistBatchCXR

NppiDCTState

DCT state structure

NppiDecodeHuffmanSpec

NppiEncodeHuffmanSpec

NppiGraphcutState

graph-cut state structure

NppiHOGConfig

The NppiHOGConfig structure defines the configuration parameters for the HOG descriptor

NppiJpegDecodeJob

JPEG decode job used by \ref nppiJpegDecodeJob (see that for more documentation) The job describes piece of computation to be done.

NppiJpegDecodeJobMemory

NppiJpegFrameDescr

JPEG frame descriptor. Can hold from 1 to 4 components.

NppiJpegScanDescr

JPEG scan descriptor

NppiMirrorBatchCXR

NppiPoint

2D Point.

NppiRect

2D Rectangle

This struct contains position and size information of a rectangle in two space.

The rectangle's position is usually signified by the coordinate of its upper-left corner.

NppiResizeBatchCXR

NppiSize

2D Size

This struct typically represents the size of a a rectangular region in two space.

NppiWarpAffineBatchCXR

NppLibraryVersion

Npp Library Version.

NppPointPolar

2D Polar Point.

Enums

DifferentialKernel

Differential Filter types

GpuComputeCapability

Gpu Compute Capabilities

InterpolationMode

Filtering methods

MaskSize

Fixed filter-kernel sizes.

NppCmpOp

Compare Operator

NppHintAlgorithm

HintAlgorithm

NppiAlphaOp

NppiAxis

Axis

NppiBayerGridPosition

Bayer Grid Position Registration.

NppiBorderType

BorderType

NppiHuffmanTableType

NppiJpegDecodeJobKind

Type of job to execute. Usually you will need just SIMPLE for each scan, one MEMZERO at the beginning and FINALIZE at the end. See the example in \ref nppiJpegDecodeJob SIMPLE can be split into multiple jobs: PRE, CPU & GPU. Please note that if you don't use SIMPLE, you man need to add some memcopies and synchronizes as described in \ref nppiJpegDecodeJob.

NppiNorm

NppRoundMode

Rounding Modes

The enumerated rounding modes are used by a large number of NPP primitives to allow the user to specify the method by which fractional values are converted to integer values. Also see \ref rounding_modes.

For NPP release 5.5 new names for the three rounding modes are introduced that are based on the naming conventions for rounding modes set forth in the IEEE-754 floating-point standard. Developers are encouraged to use the new, longer names to be future proof as the legacy names will be deprecated in subsequent NPP releases.

NppStatus

Error Status Codes

Almost all NPP function return error-status information using these return codes. Negative return codes indicate errors, positive return codes indicate warnings, a return code of 0 indicates success.