Accelerate

Basic neural network subroutines (BNNS) is a collection of functions that you use to implement and run neural networks, using previously obtained training data.

Quadrature provides an approximation of the definite integral of a function over a finite or infinite interval.

This section describes the C and Swift APIs for performing common routines on a single vector in vDSP. These functions perform tasks such as finding the absolute value of a vector, compressing the values of a vector, or converting between single and double precision vectors.

Alpha compositing (also known as alpha blending) is the process of layering multiple images, with the alpha value for a pixel in a given layer indicating what fraction of the colors from lower layers are seen through the color at the given level. The functions described in this reference operate on the alpha values of pixels either by blending alpha values or clipping them.

Conversion functions change an image from one image format into another. These functions work with the formats supported by vImage (Planar8, PlanarF, ARGB8888, ARGBFFFF, RGBA8888, and RGBAFFFF) but they can also change between a supported format to one that’s not supported by vImage (such as RGB565). Conversion functions can also fill buffers with a color, overwrite channels, permute channels, flatten data, and clip data.

Convolution functions implement various techniques for smoothing or sharpening an image by replacing a pixel with a weighted sum of itself and nearby pixels. Image convolution does not alter the size of an image.

The vImage framework provides one function for filtering data prior to decompression.

Geometric functions rotate, resize, and distort the geometry of images. vImage provides both high-level (rotation, scaling, and warping) and low-level geometric functions (reflection, shearing, and low-level rotation).

Histogram functions calculate image histograms or manipulate a histogram to modify an image. There are a number of reasons to apply histogram operations to an image. An image may not make full use of the possible range of intensity values—for example, most of its pixels may be fairly dark, making details difficult to see. Changing the image so that it has a more uniform histogram can improve contrast. Also, it may be easier to compare two images (with respect to texture or other aspects) if you change each histogram to match some standard histogram.

Morphological functions change the shape of an object by performing dilatation, erosion, maximum, and minimum operations. Dilation expands objects. Erosion contracts them. Maximum is a special case of dilation, while minimum is a special case of erosion. The precise nature of the expanding or shrinking is determined by a kernel (also known as a structure element) provided by the caller. The number of rows and number of columns of the image does not change after applying a morphological operation.

Transformation functions alter the values of pixels in the image. Unlike convolutions, transformation functions do not depend on the values of nearby pixels. The vImage transformation functions fall into four broad categories:

The data types and constants defined in this document are used by vImage functions. The primary vImage data type is the vImage buffer, which contains a pointer to image data along with other image data information. The vImage framework also defines data types for planar and interleaved pixel types, a resampling callback filter, and an affine transform. It provides constants that specify errors that can be returned by vImage functions and flags that you can pass to a function to specify a variety of processing options.