# Classic BNNS API ## Topics ### N-dimensional array descriptor essentials [`struct BNNSLayerData`](/documentation/Accelerate/BNNSLayerData) A structure containing common layer parameters. [`enum Shape`](/documentation/Accelerate/BNNS/Shape) Constants that describe the size and data layout of an n-dimensional array descriptor. [`struct BNNSDataLayout`](/documentation/Accelerate/BNNSDataLayout) Constants that describe the data type of an n-dimensional array. [`struct BNNSDataType`](/documentation/Accelerate/BNNSDataType) BNNS Data Types. [`struct BNNSNDArrayDescriptor`](/documentation/Accelerate/BNNSNDArrayDescriptor) A structure that describes the shape, stride, data type, and, optionally, the memory location of an n-dimensional array. [`func BNNSDataLayoutGetRank(BNNSDataLayout) -> Int`](/documentation/Accelerate/BNNSDataLayoutGetRank(_:)) ### General filters [`typealias BNNSFilter`](/documentation/Accelerate/BNNSFilter) An opaque type that represents a filter. [`class Layer`](/documentation/Accelerate/BNNS/Layer) The base class for layer objects that wrap filters and manage deinitialization. [`class UnaryLayer`](/documentation/Accelerate/BNNS/UnaryLayer) The base class for layers that accept a single input. [`class BinaryLayer`](/documentation/Accelerate/BNNS/BinaryLayer) The base class for layers that accept two inputs. [`struct BNNSFilterParameters`](/documentation/Accelerate/BNNSFilterParameters) A structure that contains common filter parameters. [`func BNNSFilterDestroy(BNNSFilter?)`](/documentation/Accelerate/BNNSFilterDestroy(_:)) Destroys the specified filter, releasing all resources allocated for it. [`typealias BNNSAlloc`](/documentation/Accelerate/BNNSAlloc) A type-alias for a user-provided memory allocation function. [`typealias BNNSFree`](/documentation/Accelerate/BNNSFree) A type-alias for a user-provided memory deallocation function. ### Activation layers [`func BNNSFilterCreateVectorActivationLayer(UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateVectorActivationLayer(_:_:_:_:)) [`class ActivationLayer`](/documentation/Accelerate/BNNS/ActivationLayer) A layer object that wraps an activation filter and manages its deinitialization. [`struct BNNSActivationFunction`](/documentation/Accelerate/BNNSActivationFunction) Constants that describe activation functions. [`struct BNNSActivation`](/documentation/Accelerate/BNNSActivation) A set of parameters that describe common activation functions. [`struct BNNSLayerParametersActivation`](/documentation/Accelerate/BNNSLayerParametersActivation) A set of parameters that define an activation layer. [`func BNNSFilterCreateLayerActivation(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerActivation(_:_:)) Returns a new activation layer. [`func BNNSDirectApplyActivationBatch(UnsafePointer, UnsafePointer?, Int, Int, Int) -> Int32`](/documentation/Accelerate/BNNSDirectApplyActivationBatch(_:_:_:_:_:)) Applies an activation filter to a set of input objects, writing out the result to a set of output objects. [`static func applyActivation(activation: BNNS.ActivationFunction, axes: [Int], input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, batchSize: Int, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/applyActivation(activation:axes:input:output:batchSize:filterParameters:)) Applies an activation function on the specified axes. [`static func applyActivation(activation: BNNS.ActivationFunction, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, batchSize: Int, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/applyActivation(activation:input:output:batchSize:filterParameters:)) Applies the specified activation function. ### Arithmetic layers [`class UnaryArithmeticLayer`](/documentation/Accelerate/BNNS/UnaryArithmeticLayer) A layer object that wraps a unary arithmetic filter and manages its deinitialization. [`class BinaryArithmeticLayer`](/documentation/Accelerate/BNNS/BinaryArithmeticLayer) A layer object that wraps a binary arithmetic filter and manages its deinitialization. [`class TernaryArithmeticLayer`](/documentation/Accelerate/BNNS/TernaryArithmeticLayer) A layer object that wraps a ternary arithmetic filter and manages its deinitialization. [`struct BNNSDescriptorType`](/documentation/Accelerate/BNNSDescriptorType) Constants that describe the input and output types of an arithmetic operation. [`struct BNNSArithmeticUnary`](/documentation/Accelerate/BNNSArithmeticUnary) A structure that contains the input and output of an arithmetic operation with a single input. [`struct BNNSArithmeticBinary`](/documentation/Accelerate/BNNSArithmeticBinary) A structure that contains the inputs and output of an arithmetic operation with two inputs. [`struct BNNSArithmeticTernary`](/documentation/Accelerate/BNNSArithmeticTernary) A structure that contains the inputs and output of an arithmetic operation with three inputs. [`struct BNNSArithmeticFunction`](/documentation/Accelerate/BNNSArithmeticFunction) Constants that define arithmetic operations. [`struct BNNSLayerParametersArithmetic`](/documentation/Accelerate/BNNSLayerParametersArithmetic) A structure that contains the parameters of an arithmetic layer. [`func BNNSFilterCreateLayerArithmetic(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerArithmetic(_:_:)) Returns a new arithmetic layer. [`func BNNSArithmeticFilterApplyBatch(BNNSFilter?, Int, Int, UnsafeMutablePointer, UnsafePointer, UnsafeMutableRawPointer, Int) -> Int32`](/documentation/Accelerate/BNNSArithmeticFilterApplyBatch(_:_:_:_:_:_:_:)) Applies an arithmetic filter to a set of input objects, writing the result to a set of output objects. [`func BNNSArithmeticFilterApplyBackwardBatch(BNNSFilter?, Int, Int, UnsafeMutablePointer?, UnsafePointer?, UnsafeMutablePointer>, UnsafePointer, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int) -> Int32`](/documentation/Accelerate/BNNSArithmeticFilterApplyBackwardBatch(_:_:_:_:_:_:_:_:_:_:_:)) Applies an arithmetic filter backward to generate input gradients. ### Compute norm functions [`static func computeNorm(input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axes: [Int]?) throws`](/documentation/Accelerate/BNNS/computeNorm(input:output:axes:)) Computes the Euclidean norm and writes the result to the output tensor. [`static func computeNormBackward(input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axes: [Int]?, outputGradient: BNNSNDArrayDescriptor, generatingInputGradient: BNNSNDArrayDescriptor) throws`](/documentation/Accelerate/BNNS/computeNormBackward(input:output:axes:outputGradient:generatingInputGradient:)) Backpropogates gradients for the compute norm function. [`func BNNSComputeNorm(UnsafeMutablePointer, UnsafePointer, BNNSNormType, UInt32) -> Int32`](/documentation/Accelerate/BNNSComputeNorm(_:_:_:_:)) Computes the specified norm over an entire tensor or the specified axes. [`func BNNSComputeNormBackward(UnsafeRawPointer, UnsafeMutablePointer, UnsafeRawPointer, UnsafePointer, BNNSNormType, UInt32) -> Int32`](/documentation/Accelerate/BNNSComputeNormBackward(_:_:_:_:_:_:)) Backpropogates gradients for the compute norm function. [`struct BNNSNormType`](/documentation/Accelerate/BNNSNormType) Constants that describe norm types. ### Convolution layers [`struct BNNSConvolutionLayerParameters`](/documentation/Accelerate/BNNSConvolutionLayerParameters) A structure containing convolution parameters. [`func BNNSFilterCreateConvolutionLayer(UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateConvolutionLayer(_:_:_:_:)) Returns a convolution filter, initialized with input, output, layer, and filter parameters. [`class ConvolutionLayer`](/documentation/Accelerate/BNNS/ConvolutionLayer) A layer object that wraps a convolution filter and manages its deinitialization. [`struct BNNSLayerParametersConvolution`](/documentation/Accelerate/BNNSLayerParametersConvolution) A structure that contains the parameters of a convolution layer. [`func BNNSFilterCreateLayerConvolution(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerConvolution(_:_:)) Returns a new convolution layer. [`func BNNSFilterCreateLayerTransposedConvolution(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerTransposedConvolution(_:_:)) Returns a new transposed convolution layer. ### Crop-resize layers [`class CropResizeLayer`](/documentation/Accelerate/BNNS/CropResizeLayer) A layer object that wraps a crop-resize filter and manages its deinitialization. [`func BNNSCropResize(UnsafePointer, UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSCropResize(_:_:_:_:_:)) Extracts and resizes regions of interest of an input tensor. [`func BNNSCropResizeBackward(UnsafePointer, UnsafeMutablePointer, UnsafePointer, UnsafePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSCropResizeBackward(_:_:_:_:_:)) Applies a crop-resize filter backward to generate gradients. [`struct BNNSLayerParametersCropResize`](/documentation/Accelerate/BNNSLayerParametersCropResize) A set of parameters that describe a crop-resize operation. [`struct BNNSBoxCoordinateMode`](/documentation/Accelerate/BNNSBoxCoordinateMode) Constants that define the convention to specify the four bounding box coordinates for crop-resize operations. [`struct BNNSLinearSamplingMode`](/documentation/Accelerate/BNNSLinearSamplingMode) Constants that specify how a crop-resize layer samples a grid. ### Dropout layers [`class DropoutLayer`](/documentation/Accelerate/BNNS/DropoutLayer) A layer object that wraps a dropout filter and manages its deinitialization. [`struct BNNSLayerParametersDropout`](/documentation/Accelerate/BNNSLayerParametersDropout) A structure that contains the parameters of a dropout layer. [`func BNNSFilterCreateLayerDropout(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerDropout(_:_:)) Returns a new dropout layer. ### Embedding layers [`class EmbeddingLayer`](/documentation/Accelerate/BNNS/EmbeddingLayer) A layer object that wraps an embedding filter and manages its deinitialization. [`struct BNNSLayerParametersEmbedding`](/documentation/Accelerate/BNNSLayerParametersEmbedding) A structure that contains the parameters of an embedding layer. [`func BNNSFilterCreateLayerEmbedding(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerEmbedding(_:_:)) Returns a new embedding layer. ### Fully connected layers [`struct BNNSFullyConnectedLayerParameters`](/documentation/Accelerate/BNNSFullyConnectedLayerParameters) A structure containing fully connected layer parameters. [`func BNNSFilterCreateFullyConnectedLayer(UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateFullyConnectedLayer(_:_:_:_:)) Returns a fully connected filter, initialized with input, output, layer, and filter parameters. [`class FullyConnectedLayer`](/documentation/Accelerate/BNNS/FullyConnectedLayer) A layer object that wraps a fully connected filter and manages its deinitialization. [`struct BNNSLayerParametersFullyConnected`](/documentation/Accelerate/BNNSLayerParametersFullyConnected) A structure that contains the parameters of a fully connected layer. [`func BNNSFilterCreateLayerFullyConnected(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerFullyConnected(_:_:)) Returns a new fully connected layer. ### Fused layers [`protocol FusableLayerParameters`](/documentation/Accelerate/FusableLayerParameters) [`class FusedParametersLayer`](/documentation/Accelerate/BNNS/FusedParametersLayer) A layer object that wraps a fused layer and manages its deinitialization. [`class FusedConvolutionNormalizationLayer`](/documentation/Accelerate/BNNS/FusedConvolutionNormalizationLayer) A layer object that wraps a fused, convolution normalization layer and manages its deinitialization. [`class FusedFullyConnectedNormalizationLayer`](/documentation/Accelerate/BNNS/FusedFullyConnectedNormalizationLayer) A layer object that wraps a fused, fully connected normalization layer and manages its deinitialization. [`struct BNNSFilterType`](/documentation/Accelerate/BNNSFilterType) Constants that define the component filters of a fused layer. [`func BNNSFilterCreateFusedLayer(Int, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateFusedLayer(_:_:_:_:)) Returns a new fused layer. [`func BNNSFusedFilterApplyBatch(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeMutableRawPointer, Int, Bool) -> Int32`](/documentation/Accelerate/BNNSFusedFilterApplyBatch(_:_:_:_:_:_:_:)) Applies a fused filter to a set of input objects, writing the result to a set of output objects. [`func BNNSFusedFilterApplyMultiInputBatch(BNNSFilter?, Int, Int, UnsafeMutablePointer, UnsafePointer, UnsafeMutableRawPointer, Int, Bool) -> Int32`](/documentation/Accelerate/BNNSFusedFilterApplyMultiInputBatch(_:_:_:_:_:_:_:_:)) Applies a multiple-input fused filter to a set of input objects, writing the result to a set of output objects. [`func BNNSFusedFilterApplyBackwardBatch(BNNSFilter?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int, UnsafeMutablePointer?>?) -> Int32`](/documentation/Accelerate/BNNSFusedFilterApplyBackwardBatch(_:_:_:_:_:_:_:_:_:_:_:)) Applies a fused filter backward to generate input gradients. [`func BNNSFusedFilterApplyBackwardMultiInputBatch(BNNSFilter?, Int, Int, UnsafeMutablePointer?, UnsafePointer?, UnsafeMutablePointer>, UnsafePointer, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int, UnsafeMutablePointer?>?) -> Int32`](/documentation/Accelerate/BNNSFusedFilterApplyBackwardMultiInputBatch(_:_:_:_:_:_:_:_:_:_:_:_:)) Applies a multiple-input fused filter backward to generate input gradients. ### Gather and scatter operations [`static func gather(input: BNNSNDArrayDescriptor, indices: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axis: Int, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/gather(input:indices:output:axis:filterParameters:)) Gathers the elements of a tensor along a single axis. [`static func gatherND(input: BNNSNDArrayDescriptor, indices: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/gatherND(input:indices:output:filterParameters:)) Gathers the slices of a tensor. [`static func scatter(input: BNNSNDArrayDescriptor, indices: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axis: Int, reductionFunction: BNNS.ReductionFunction, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/scatter(input:indices:output:axis:reductionFunction:filterParameters:)) Scatters the elements of a tensor along a single axis. [`static func scatterND(input: BNNSNDArrayDescriptor, indices: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, reductionFunction: BNNS.ReductionFunction, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/scatterND(input:indices:output:reductionFunction:filterParameters:)) Scatters the slices of a tensor. [`func BNNSGather(Int, UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSGather(_:_:_:_:_:)) Gathers the elements of a tensor along a single axis. [`func BNNSGatherND(UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSGatherND(_:_:_:_:)) Gathers the slices of a tensor. [`func BNNSScatter(Int, BNNSReduceFunction, UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSScatter(_:_:_:_:_:_:)) Scatters the elements of a tensor along a single axis. [`func BNNSScatterND(BNNSReduceFunction, UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSScatterND(_:_:_:_:_:)) Scatters the slices of a tensor. ### Loss layers [`class LossLayer`](/documentation/Accelerate/BNNS/LossLayer) A layer object that wraps a loss filter and manages its deinitialization. [`struct BNNSLossFunction`](/documentation/Accelerate/BNNSLossFunction) Constants that describe loss functions. [`struct BNNSLossReductionFunction`](/documentation/Accelerate/BNNSLossReductionFunction) Constants that describe reduction functions used by a loss layer. [`struct BNNSLayerParametersLossBase`](/documentation/Accelerate/BNNSLayerParametersLossBase) A structure that contains the parameters of a loss layer. [`struct BNNSLayerParametersLossHuber`](/documentation/Accelerate/BNNSLayerParametersLossHuber) A structure that contains the parameters of a Huber loss layer. [`struct BNNSLayerParametersLossSigmoidCrossEntropy`](/documentation/Accelerate/BNNSLayerParametersLossSigmoidCrossEntropy) A structure that contains the parameters of a sigmoid cross entropy loss layer. [`struct BNNSLayerParametersLossSoftmaxCrossEntropy`](/documentation/Accelerate/BNNSLayerParametersLossSoftmaxCrossEntropy) A structure that contains the parameters of a softmax cross entropy loss layer. [`struct BNNSLayerParametersLossYolo`](/documentation/Accelerate/BNNSLayerParametersLossYolo) A structure that contains the parameters of a You Only Look Once (YOLO) loss layer. [`func BNNSFilterCreateLayerLoss(UnsafeRawPointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerLoss(_:_:)) Returns a new loss layer. [`func BNNSLossFilterApplyBatch(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeRawPointer, Int, UnsafeRawPointer?, Int, UnsafeMutableRawPointer, UnsafeMutablePointer?, Int) -> Int32`](/documentation/Accelerate/BNNSLossFilterApplyBatch(_:_:_:_:_:_:_:_:_:_:_:)) Applies a loss filter to a set of input objects, writing the result to a set of output objects. [`func BNNSLossFilterApplyBackwardBatch(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeMutablePointer, Int, UnsafeRawPointer, Int, UnsafeRawPointer?, Int, UnsafePointer, Int) -> Int32`](/documentation/Accelerate/BNNSLossFilterApplyBackwardBatch(_:_:_:_:_:_:_:_:_:_:_:_:)) Applies a loss filter backward to generate gradients. ### K-nearest neighbors calculation [`struct NearestNeighbors`](/documentation/Accelerate/BNNS/NearestNeighbors) A structure that calculates k-nearest neighbors. [`typealias BNNSNearestNeighbors`](/documentation/Accelerate/BNNSNearestNeighbors) A k-nearest neighbors object. [`func BNNSCreateNearestNeighbors(UInt32, UInt32, UInt32, BNNSDataType, UnsafePointer?) -> BNNSNearestNeighbors?`](/documentation/Accelerate/BNNSCreateNearestNeighbors(_:_:_:_:_:)) Returns a new k-nearest neighbors object. [`func BNNSNearestNeighborsLoad(BNNSNearestNeighbors?, UInt32, UnsafeRawPointer) -> Int32`](/documentation/Accelerate/BNNSNearestNeighborsLoad(_:_:_:)) Adds new sample data to a k-nearest neighbors object. [`func BNNSNearestNeighborsGetInfo(BNNSNearestNeighbors?, Int32, UnsafeMutablePointer?, UnsafeMutableRawPointer?) -> Int32`](/documentation/Accelerate/BNNSNearestNeighborsGetInfo(_:_:_:_:)) Calculates the sorted indices and Euclidean distances of the k-nearest neighbors to a specified sample data point. [`func BNNSDestroyNearestNeighbors(BNNSNearestNeighbors?)`](/documentation/Accelerate/BNNSDestroyNearestNeighbors(_:)) Destroys a k-nearest neighbors object. ### Matrix multiplication [`func BNNSDirectApplyBroadcastMatMul(Bool, Bool, Float, UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer?)`](/documentation/Accelerate/BNNSDirectApplyBroadcastMatMul(_:_:_:_:_:_:_:)) Applies a broadcast matrix multiplication operation directly to two input matrices. [`class BroadcastMatrixMultiplyLayer`](/documentation/Accelerate/BNNS/BroadcastMatrixMultiplyLayer) A layer object that wraps a broadcast matrix multiply filter and manages its deinitialization. [`struct BNNSLayerParametersBroadcastMatMul`](/documentation/Accelerate/BNNSLayerParametersBroadcastMatMul) A set of parameters that define a broadcast matrix multiply layer. [`func BNNSFilterCreateLayerBroadcastMatMul(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerBroadcastMatMul(_:_:)) Returns a new broadcast matrix multiply layer. [`func BNNSMatMulWorkspaceSize(Bool, Bool, Float, UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer?) -> Int`](/documentation/Accelerate/BNNSMatMulWorkspaceSize(_:_:_:_:_:_:_:)) Returns the workspace size that a matrix multiply operation requires. [`func BNNSMatMul(Bool, Bool, Float, UnsafePointer, UnsafePointer, UnsafePointer, UnsafeMutableRawPointer?, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSMatMul(_:_:_:_:_:_:_:_:)) Applies a matrix multiplication operation directly to two input matrices. [`static func applyMatrixMultiplication(inputA: BNNSNDArrayDescriptor, transposed: Bool, inputB: BNNSNDArrayDescriptor, transposed: Bool, output: BNNSNDArrayDescriptor, alpha: Float, workspace: UnsafeMutableRawBufferPointer?, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/applyMatrixMultiplication(inputA:transposed:inputB:transposed:output:alpha:workspace:filterParameters:)) Performs a matrix multiplication operation directly on two input matrices. [`static func matrixMultiplicationWorkspaceSize(inputA: BNNSNDArrayDescriptor, transposed: Bool, inputB: BNNSNDArrayDescriptor, transposed: Bool, output: BNNSNDArrayDescriptor, alpha: Float, filterParameters: BNNSFilterParameters?) -> Int?`](/documentation/Accelerate/BNNS/matrixMultiplicationWorkspaceSize(inputA:transposed:inputB:transposed:output:alpha:filterParameters:)) Returns the workspace size that a matrix multiply operation requires. ### Multihead attention layers [`struct BNNSMHAProjectionParameters`](/documentation/Accelerate/BNNSMHAProjectionParameters) A structure that contains multihead attention projection parameters. [`struct BNNSLayerParametersMultiheadAttention`](/documentation/Accelerate/BNNSLayerParametersMultiheadAttention) A structure that contains the parameters of a multihead attention layer. [`func BNNSFilterCreateLayerMultiheadAttention(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerMultiheadAttention(_:_:)) Returns a new multihead attention layer. [`func BNNSApplyMultiheadAttention(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeRawPointer, Int, UnsafePointer?, Int, UnsafeRawPointer, Int, UnsafeMutableRawPointer, Int, UnsafePointer?, UnsafeMutablePointer?, UnsafeMutableRawPointer?, UnsafeMutablePointer?, UnsafeMutableRawPointer?) -> Int32`](/documentation/Accelerate/BNNSApplyMultiheadAttention(_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:)) Applies a mutihead attention filter to a set of input objects, writing the result to a set of output objects. [`func BNNSApplyMultiheadAttentionBackward(BNNSFilter?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer?, UnsafeRawPointer?, Int, UnsafePointer?, Int, UnsafeMutablePointer?, UnsafeRawPointer?, Int, UnsafeMutablePointer?, UnsafePointer?, UnsafeMutablePointer?, UnsafeMutablePointer?, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int, UnsafeMutableRawPointer?, UnsafeMutablePointer?, UnsafeMutableRawPointer?) -> Int32`](/documentation/Accelerate/BNNSApplyMultiheadAttentionBackward(_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:_:)) Applies a multihead attention filter backward to generate gradients. ### Normalization layers [`class NormalizationLayer`](/documentation/Accelerate/BNNS/NormalizationLayer) A layer object that wraps a normalization filter and manages its deinitialization. [`struct BNNSLayerParametersNormalization`](/documentation/Accelerate/BNNSLayerParametersNormalization) A structure that contains the parameters of a normalization layer. [`func BNNSFilterCreateLayerNormalization(BNNSFilterType, UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerNormalization(_:_:_:)) Returns a new normalization layer. [`func BNNSNormalizationFilterApplyBatch(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeMutableRawPointer, Int, Bool) -> Int32`](/documentation/Accelerate/BNNSNormalizationFilterApplyBatch(_:_:_:_:_:_:_:)) Applies a normalization filter to a set of input objects, writing the result to a set of output objects. [`func BNNSNormalizationFilterApplyBackwardBatch(BNNSFilter?, Int, UnsafeMutablePointer?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int, UnsafeMutablePointer?, UnsafeMutablePointer?) -> Int32`](/documentation/Accelerate/BNNSNormalizationFilterApplyBackwardBatch(_:_:_:_:_:_:_:_:_:_:)) Applies a normalization filter backward to generate gradients. ### Optimizers [`struct AdamOptimizer`](/documentation/Accelerate/BNNS/AdamOptimizer) An optimizer that uses the Adam optimization algorithm. [`struct AdamWOptimizer`](/documentation/Accelerate/BNNS/AdamWOptimizer) An optimizer that uses the AdamW optimization algorithm. [`struct RMSPropOptimizer`](/documentation/Accelerate/BNNS/RMSPropOptimizer) An optimizer that uses the root mean square propagation (RMSProp) optimization method. [`struct SGDMomentumOptimizer`](/documentation/Accelerate/BNNS/SGDMomentumOptimizer) An optimizer that uses the stochastic gradient descent (SGD) with the momentum optimization method. [`protocol BNNSOptimizer`](/documentation/Accelerate/BNNSOptimizer) [`struct BNNSOptimizerRegularizationFunction`](/documentation/Accelerate/BNNSOptimizerRegularizationFunction) A structure that contains optimizer regularization functions. [`struct BNNSOptimizerAdamFields`](/documentation/Accelerate/BNNSOptimizerAdamFields) A structure that contains the fields of an Adam optimizer. [`struct BNNSOptimizerAdamWithClippingFields`](/documentation/Accelerate/BNNSOptimizerAdamWithClippingFields) A structure that contains the fields of an Adam or AdamW optimizer that optionally clips the gradient by value or by norm. [`struct BNNSOptimizerRMSPropFields`](/documentation/Accelerate/BNNSOptimizerRMSPropFields) A structure that contains the fields of a root mean square propagation (RMSProp) optimizer. [`struct BNNSOptimizerRMSPropWithClippingFields`](/documentation/Accelerate/BNNSOptimizerRMSPropWithClippingFields) A structure that contains the fields of a root mean square propagation (RMSProp) optimizer that optionally clips the gradient by value or by norm. [`struct BNNSOptimizerSGDMomentumFields`](/documentation/Accelerate/BNNSOptimizerSGDMomentumFields) A structure that contains the fields of a stochastic gradient descent (SGD) with momentum optimizer. [`struct BNNSOptimizerSGDMomentumWithClippingFields`](/documentation/Accelerate/BNNSOptimizerSGDMomentumWithClippingFields) A structure that contains the fields of a stochastic gradient descent (SGD) with momentum optimizer that optionally clips the gradient by value or by norm. [`struct BNNSOptimizerSGDMomentumVariant`](/documentation/Accelerate/BNNSOptimizerSGDMomentumVariant) Constants that define SGD momentum variants. [`func BNNSOptimizerStep(BNNSOptimizerFunction, UnsafeRawPointer, Int, UnsafeMutablePointer>, UnsafeMutablePointer>, UnsafeMutablePointer?>?, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSOptimizerStep(_:_:_:_:_:_:_:)) Applies a single optimization step to one or more parameters. [`struct BNNSOptimizerFunction`](/documentation/Accelerate/BNNSOptimizerFunction) A structure that contains optimizer functions. ### Padding layers [`class PaddingLayer`](/documentation/Accelerate/BNNS/PaddingLayer) A layer object that wraps a padding filter and manages its deinitialization. [`struct BNNSPaddingMode`](/documentation/Accelerate/BNNSPaddingMode) Constants that define padding modes. [`struct BNNSLayerParametersPadding`](/documentation/Accelerate/BNNSLayerParametersPadding) A structure that contains the parameters of a padding layer. [`func BNNSFilterCreateLayerPadding(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerPadding(_:_:)) Returns a new loss layer. ### Permute layers [`class PermuteLayer`](/documentation/Accelerate/BNNS/PermuteLayer) A layer object that wraps a permute filter and manages its deinitialization. [`struct BNNSLayerParametersPermute`](/documentation/Accelerate/BNNSLayerParametersPermute) A structure that contains the parameters of a permute layer. [`func BNNSFilterCreateLayerPermute(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerPermute(_:_:)) Returns a new permute layer. [`func BNNSPermuteFilterApplyBackwardBatch(BNNSFilter?, Int, UnsafeMutablePointer, Int, UnsafePointer, Int) -> Int32`](/documentation/Accelerate/BNNSPermuteFilterApplyBackwardBatch(_:_:_:_:_:_:)) Applies a permute filter backward to generate gradients. ### Pooling layers [`struct BNNSPoolingLayerParameters`](/documentation/Accelerate/BNNSPoolingLayerParameters) A structure containing pooling layer parameters. [`func BNNSFilterCreatePoolingLayer(UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreatePoolingLayer(_:_:_:_:)) Returns a pooling filter, initialized with input, output, layer, and filter parameters. [`class PoolingLayer`](/documentation/Accelerate/BNNS/PoolingLayer) A layer object that wraps a pooling filter and manages its deinitialization. [`struct BNNSPoolingFunction`](/documentation/Accelerate/BNNSPoolingFunction) Constants that describe pooling functions. [`var BNNSPoolingFunctionAverage: BNNSPoolingFunction`](/documentation/Accelerate/BNNSPoolingFunctionAverage) [`var BNNSPoolingFunctionMax: BNNSPoolingFunction`](/documentation/Accelerate/BNNSPoolingFunctionMax) [`struct BNNSLayerParametersPooling`](/documentation/Accelerate/BNNSLayerParametersPooling) A structure that contains the parameters of a pooling layer. [`func BNNSFilterCreateLayerPooling(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerPooling(_:_:)) Returns a new pooling layer. [`func BNNSPoolingFilterApplyBatch(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeMutableRawPointer, Int, UnsafeMutablePointer?, Int) -> Int32`](/documentation/Accelerate/BNNSPoolingFilterApplyBatch(_:_:_:_:_:_:_:_:)) Applies a pooling filter to a set of input objects, writing the result to a set of output objects. [`func BNNSPoolingFilterApplyBackwardBatch(BNNSFilter?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int, UnsafeMutablePointer?, UnsafePointer?, Int) -> Int32`](/documentation/Accelerate/BNNSPoolingFilterApplyBackwardBatch(_:_:_:_:_:_:_:_:_:_:_:_:_:)) Applies a pooling filter backward to generate gradients. [`func BNNSPoolingFilterApplyBatchEx(BNNSFilter?, Int, UnsafeRawPointer, Int, UnsafeMutableRawPointer, Int, BNNSDataType, UnsafeMutableRawPointer?, Int) -> Int32`](/documentation/Accelerate/BNNSPoolingFilterApplyBatchEx(_:_:_:_:_:_:_:_:_:)) Applies a pooling filter to a set of input objects with support for multiple data types for indices. [`func BNNSPoolingFilterApplyBackwardBatchEx(BNNSFilter?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer?, Int, UnsafeRawPointer?, Int, UnsafeMutablePointer, Int, UnsafeMutablePointer?, BNNSDataType, UnsafeRawPointer?, Int) -> Int32`](/documentation/Accelerate/BNNSPoolingFilterApplyBackwardBatchEx(_:_:_:_:_:_:_:_:_:_:_:_:_:_:)) Applies a pooling filter backward to generate gradients with support for multiple data types for indices. ### Quantization functions [`static func quantize(batchSize: Int, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axis: Int?, scale: BNNSNDArrayDescriptor?, bias: BNNSNDArrayDescriptor?, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/quantize(batchSize:input:output:axis:scale:bias:filterParameters:)) Quantizes the input tensor and writes the result to the output tensor. [`static func dequantize(batchSize: Int, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axis: Int?, scale: BNNSNDArrayDescriptor?, bias: BNNSNDArrayDescriptor?, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/dequantize(batchSize:input:output:axis:scale:bias:filterParameters:)) Dequantizes the input tensor and writes the result to the output tensor. [`struct BNNSQuantizerFunction`](/documentation/Accelerate/BNNSQuantizerFunction) Constants that describe quantization functions. [`struct BNNSLayerParametersQuantization`](/documentation/Accelerate/BNNSLayerParametersQuantization) A structure that contains the parameters of a quantization layer. [`func BNNSDirectApplyQuantizer(UnsafePointer, UnsafePointer?, Int, Int, Int) -> Int32`](/documentation/Accelerate/BNNSDirectApplyQuantizer(_:_:_:_:_:)) Applies a quantization layer directly to two input matrices. ### Random number generation [`class RandomGenerator`](/documentation/Accelerate/BNNS/RandomGenerator) A random number generator. [`func BNNSCreateRandomGenerator(BNNSRandomGeneratorMethod, UnsafePointer?) -> BNNSRandomGenerator?`](/documentation/Accelerate/BNNSCreateRandomGenerator(_:_:)) Returns a new random number generator using an internally generated random seed. [`func BNNSCreateRandomGeneratorWithSeed(BNNSRandomGeneratorMethod, UInt64, UnsafePointer?) -> BNNSRandomGenerator?`](/documentation/Accelerate/BNNSCreateRandomGeneratorWithSeed(_:_:_:)) Returns a new random number generator using the specified seed. [`struct BNNSRandomGeneratorMethod`](/documentation/Accelerate/BNNSRandomGeneratorMethod) Constants that describe random number generation methods. [`typealias BNNSRandomGenerator`](/documentation/Accelerate/BNNSRandomGenerator) A pointer to a random number generator object. [`func BNNSRandomFillUniformInt(BNNSRandomGenerator?, UnsafeMutablePointer, Int64, Int64) -> Int32`](/documentation/Accelerate/BNNSRandomFillUniformInt(_:_:_:_:)) Fills the specified tensor with random integer values from the continuous uniform distribution within a range. [`func BNNSRandomFillUniformFloat(BNNSRandomGenerator?, UnsafeMutablePointer, Float, Float) -> Int32`](/documentation/Accelerate/BNNSRandomFillUniformFloat(_:_:_:_:)) Fills the specified tensor with random floating-point values from the continuous uniform distribution within a range. [`func BNNSRandomFillNormalFloat(BNNSRandomGenerator?, UnsafeMutablePointer, Float, Float) -> Int32`](/documentation/Accelerate/BNNSRandomFillNormalFloat(_:_:_:_:)) Fills the specified tensor with random floating-point values mapped to a normal distribution. [`func BNNSRandomFillCategoricalFloat(BNNSRandomGenerator?, UnsafePointer, UnsafePointer, Bool) -> Int32`](/documentation/Accelerate/BNNSRandomFillCategoricalFloat(_:_:_:_:)) Fills the specified tensor with random values from the categorical distributions with the given event probabilities. [`func BNNSRandomGeneratorStateSize(BNNSRandomGenerator?) -> Int`](/documentation/Accelerate/BNNSRandomGeneratorStateSize(_:)) Returns the state size, in bytes, of a random number generator. [`func BNNSRandomGeneratorGetState(BNNSRandomGenerator?, Int, UnsafeMutableRawPointer) -> Int32`](/documentation/Accelerate/BNNSRandomGeneratorGetState(_:_:_:)) Returns the state of a random number generator. [`func BNNSRandomGeneratorSetState(BNNSRandomGenerator?, Int, UnsafeMutableRawPointer) -> Int32`](/documentation/Accelerate/BNNSRandomGeneratorSetState(_:_:_:)) Sets the state of a random number generator. [`func BNNSDestroyRandomGenerator(BNNSRandomGenerator?)`](/documentation/Accelerate/BNNSDestroyRandomGenerator(_:)) Destroys a random number generator. ### Recurrent layers [`struct BNNSLSTMDataDescriptor`](/documentation/Accelerate/BNNSLSTMDataDescriptor) A structure that contains the input-output, hidden, and cell state n-dimensional array descriptors for a long short-term memory (LSTM) layer. [`struct BNNSLSTMGateDescriptor`](/documentation/Accelerate/BNNSLSTMGateDescriptor) A structure that describes a long short-term memory (LSTM) gate layer. [`struct BNNSLayerFlags`](/documentation/Accelerate/BNNSLayerFlags) Options that control the behavior of a long short-term memory (LSTM) layer. [`struct BNNSLayerParametersLSTM`](/documentation/Accelerate/BNNSLayerParametersLSTM) A structure that contains the parameters of a long short-term memory (LSTM) layer. [`func BNNSComputeLSTMTrainingCacheCapacity(UnsafePointer) -> Int`](/documentation/Accelerate/BNNSComputeLSTMTrainingCacheCapacity(_:)) Returns the minimum bytes capacity of the training cache buffer a long short-term memory (LSTM) layer uses when it’s applied. [`func BNNSDirectApplyLSTMBatchTrainingCaching(UnsafePointer, UnsafePointer?, UnsafeMutableRawPointer?, Int) -> Int32`](/documentation/Accelerate/BNNSDirectApplyLSTMBatchTrainingCaching(_:_:_:_:)) Applies a long short-term memory (LSTM) layer directly to an input. [`func BNNSDirectApplyLSTMBatchBackward(UnsafePointer, UnsafePointer, UnsafePointer?, UnsafeRawPointer?, Int) -> Int32`](/documentation/Accelerate/BNNSDirectApplyLSTMBatchBackward(_:_:_:_:_:)) Applies a long short-term memory (LSTM) filter backward to generate gradients. ### Reduction layers [`class ReductionLayer`](/documentation/Accelerate/BNNS/ReductionLayer) A layer object that wraps a reduction filter and manages its deinitialization. [`static func applyReduction(BNNS.ReductionFunction, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, weights: BNNSNDArrayDescriptor?, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/applyReduction(_:input:output:weights:filterParameters:)) Applies the specified reduction function. [`struct BNNSReduceFunction`](/documentation/Accelerate/BNNSReduceFunction) Constants that describe reduction functions. [`struct BNNSLayerParametersReduction`](/documentation/Accelerate/BNNSLayerParametersReduction) A set of parameters that define a reduction layer. [`func BNNSFilterCreateLayerReduction(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerReduction(_:_:)) Returns a new reduction layer. [`func BNNSDirectApplyReduction(UnsafePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSDirectApplyReduction(_:_:)) Applies a reduction operation directly to an input tensor. ### Resize layers [`class ResizeLayer`](/documentation/Accelerate/BNNS/ResizeLayer) A layer object that wraps a resize filter and manages its deinitialization. [`struct BNNSInterpolationMethod`](/documentation/Accelerate/BNNSInterpolationMethod) Constants that describe interpolation methods. [`struct BNNSLayerParametersResize`](/documentation/Accelerate/BNNSLayerParametersResize) A structure that contains the parameters of a resize layer. [`func BNNSFilterCreateLayerResize(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerResize(_:_:)) Returns a new resize layer. ### Sparse layers [`func BNNSNDArrayGetDataSize(UnsafePointer) -> Int`](/documentation/Accelerate/BNNSNDArrayGetDataSize(_:)) Returns the size, in bytes, that an array descriptor requires. [`func BNNSNDArrayFullyConnectedSparsifySparseCOO(UnsafePointer, UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?, Int, UnsafeMutableRawPointer?, Int, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSNDArrayFullyConnectedSparsifySparseCOO(_:_:_:_:_:_:_:_:_:)) Converts a sparse tensor from the standardized coordinate list (COO) layout to a device-specific sparse layout that BNNS fully connected layers use. [`func BNNSNDArrayFullyConnectedSparsifySparseCSR(UnsafePointer, UnsafePointer, UnsafePointer, UnsafePointer, UnsafeMutablePointer, UnsafePointer?, Int, UnsafeMutableRawPointer?, Int, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSNDArrayFullyConnectedSparsifySparseCSR(_:_:_:_:_:_:_:_:_:_:)) Converts a sparse tensor from the standardized compressed sparse row (CSR) layout to a device-specific sparse layout that BNNS fully connected layers use. [`static func sparsify(batchSize: Int, inputLayout: BNNS.SparseLayout, inputDenseShape: BNNSNDArrayDescriptor, inputValues: BNNSNDArrayDescriptor, output: inout BNNSNDArrayDescriptor, sparseParameters: BNNS.SparseParameters?, workspace: UnsafeMutableRawBufferPointer?, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/FullyConnectedLayer/sparsify(batchSize:inputLayout:inputDenseShape:inputValues:output:sparseParameters:workspace:filterParameters:)) Converts a sparse tensor from a standardized sparse layout to a device-specific sparse layout that Fully Connected uses. [`struct SparseParameters`](/documentation/Accelerate/BNNS/SparseParameters) A data structure that provides a hint to the sparsity function. [`enum SparseLayout`](/documentation/Accelerate/BNNS/SparseLayout) Constants that specify standardized sparse layouts that BNNS can convert to opaque. [`enum SparsityType`](/documentation/Accelerate/BNNS/SparsityType) Constants that specify patterns in the sparsity. [`var BNNSSparsityTypeUnstructured: BNNSSparsityType`](/documentation/Accelerate/BNNSSparsityTypeUnstructured) ### Tensor comparison layers [`static func compare(BNNSNDArrayDescriptor, BNNSNDArrayDescriptor, using: BNNS.RelationalOperator, output: BNNSNDArrayDescriptor) throws`](/documentation/Accelerate/BNNS/compare(_:_:using:output:)) Performs an elementwise comparison of two array descriptors using the specified relational operator. [`struct BNNSRelationalOperator`](/documentation/Accelerate/BNNSRelationalOperator) Constants that describe relational operations. [`func BNNSCompareTensor(UnsafePointer, UnsafePointer, BNNSRelationalOperator, UnsafeMutablePointer) -> Int32`](/documentation/Accelerate/BNNSCompareTensor(_:_:_:_:)) Returns a tensor of Boolean type by comparing or performing a logical operation between two inputs. ### Tensor contraction layers [`struct BNNSLayerParametersTensorContraction`](/documentation/Accelerate/BNNSLayerParametersTensorContraction) A structure that contains the parameters of a tensor-contraction layer. [`func BNNSFilterCreateLayerTensorContraction(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerTensorContraction(_:_:)) Returns a new tensor-contraction layer. ### Top-k layers [`static func applyTopK(k: Int, input: BNNSNDArrayDescriptor, bestValues: BNNSNDArrayDescriptor, bestIndices: BNNSNDArrayDescriptor, axis: Int, batchSize: Int, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/applyTopK(k:input:bestValues:bestIndices:axis:batchSize:filterParameters:)) Applies a top-k filter directly to an input. [`static func applyInTopK(k: Int, input: BNNSNDArrayDescriptor, testIndices: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axis: Int, batchSize: Int, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/applyInTopK(k:input:testIndices:output:axis:batchSize:filterParameters:)) Applies an in-top-k filter directly to an input. ### Top-k layers [`func BNNSDirectApplyTopK(Int, Int, Int, UnsafePointer, Int, UnsafeMutablePointer, Int, UnsafeMutablePointer?, Int, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSDirectApplyTopK(_:_:_:_:_:_:_:_:_:_:)) Applies a top-k filter directly to an input. [`func BNNSDirectApplyInTopK(Int, Int, Int, UnsafePointer, Int, UnsafePointer, Int, UnsafeMutablePointer, Int, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSDirectApplyInTopK(_:_:_:_:_:_:_:_:_:_:)) Applies an in-top-k filter directly to an input. ### Utility functions [`static func copy(BNNSNDArrayDescriptor, to: BNNSNDArrayDescriptor, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/copy(_:to:filterParameters:)) Copies the contents of an n-dimensional array descriptor to another descriptor of the same shape. [`static func transpose(input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, firstTransposeAxis: Int, secondTransposeAxis: Int, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/transpose(input:output:firstTransposeAxis:secondTransposeAxis:filterParameters:)) Transposes a tensor by swapping two of its dimensions. [`func BNNSCopy(UnsafeMutablePointer, UnsafePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSCopy(_:_:_:)) Copies the contents of an n-dimensional array descriptor to another of the same shape. [`func BNNSTranspose(UnsafeMutablePointer, UnsafePointer, Int, Int, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSTranspose(_:_:_:_:_:)) Transposes a tensor by swapping two of its dimensions. [`func BNNSGetPointer(BNNSFilter?, BNNSPointerSpecifier) -> BNNSNDArrayDescriptor`](/documentation/Accelerate/BNNSGetPointer(_:_:)) Returns an n-dimensional array descriptor that contains a reference to a filter-data member. [`struct BNNSPointerSpecifier`](/documentation/Accelerate/BNNSPointerSpecifier) Constants that specify which pointer the BNNS get filter function returns. [`class GramLayer`](/documentation/Accelerate/BNNS/GramLayer) A layer object that wraps a Gram matrix filter and manages its deinitialization. [`struct BNNSLayerParametersGram`](/documentation/Accelerate/BNNSLayerParametersGram) A set of parameters that define a Gram matrix layer. [`func BNNSFilterCreateLayerGram(UnsafePointer, UnsafePointer?) -> BNNSFilter?`](/documentation/Accelerate/BNNSFilterCreateLayerGram(_:_:)) Returns a new Gram matrix layer. [`static func clip(to: ClosedRange, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor) throws`](/documentation/Accelerate/BNNS/clip(to:input:output:)) Clips the input tensor to a closed range and writes the result to the output tensor. [`static func clipByNorm(threshold: Float, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, axes: [Int]?) throws`](/documentation/Accelerate/BNNS/clipByNorm(threshold:input:output:axes:)) Clips the input tensor to a Euclidean norm and writes the result to the output tensor. [`static func clipByGlobalNorm(threshold: Float, inputs: [BNNSNDArrayDescriptor], outputs: [BNNSNDArrayDescriptor], globalNorm: Float) throws`](/documentation/Accelerate/BNNS/clipByGlobalNorm(threshold:inputs:outputs:globalNorm:)) Clips the input tensors to a global Euclidean norm and writes the result to the output tensors. [`func BNNSClipByValue(UnsafeMutablePointer, UnsafePointer, Float, Float) -> Int32`](/documentation/Accelerate/BNNSClipByValue(_:_:_:_:)) Clips a tensor’s values to the specified minimum and maximum values. [`func BNNSClipByNorm(UnsafeMutablePointer, UnsafePointer, Float, UInt32) -> Int32`](/documentation/Accelerate/BNNSClipByNorm(_:_:_:_:)) Clips a tensor’s values to a maximum Euclidean norm. [`func BNNSClipByGlobalNorm(UnsafeMutablePointer>, UnsafeMutablePointer>, Int, Float, Float) -> Int32`](/documentation/Accelerate/BNNSClipByGlobalNorm(_:_:_:_:_:)) Clips a tensor’s values to a maximum global Euclidean norm. [`static func copyBandPart(BNNSNDArrayDescriptor, to: BNNSNDArrayDescriptor, lowerBandCount: Int?, upperBandCount: Int?, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/copyBandPart(_:to:lowerBandCount:upperBandCount:filterParameters:)) Copies the specified subdiagonals and superdiagonals of a matrix, and sets other elements to zero. [`static func shuffle(BNNS.ShuffleType, input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/shuffle(_:input:output:filterParameters:)) Rearranges elements in a tensor according to shuffle type. [`enum ShuffleType`](/documentation/Accelerate/BNNS/ShuffleType) Constants that specify a shuffle type. [`static func tile(input: BNNSNDArrayDescriptor, output: BNNSNDArrayDescriptor, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/tile(input:output:filterParameters:)) Generates an output tensor by tiling an input tensor multiple times. [`static func tileBackward(outputGradient: BNNSNDArrayDescriptor, generatingInputGradient: BNNSNDArrayDescriptor, filterParameters: BNNSFilterParameters?) throws`](/documentation/Accelerate/BNNS/tileBackward(outputGradient:generatingInputGradient:filterParameters:)) Applies a tile filter backward to generate an input gradient. ### Errors [`enum Error`](/documentation/Accelerate/BNNS/Error) [`func BNNSBandPart(Int32, Int32, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSBandPart(_:_:_:_:_:)) Copies the specified subdiagonals and superdiagonals of a matrix, and sets other elements to zero. [`func BNNSShuffle(BNNSShuffleType, UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSShuffle(_:_:_:_:)) Rearranges elements in a tensor according to shuffle type. [`struct BNNSShuffleType`](/documentation/Accelerate/BNNSShuffleType) Constants that specify a shuffle type. [`func BNNSTile(UnsafePointer, UnsafeMutablePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSTile(_:_:_:)) Generates an output tensor by tiling an input tensor multiple times. [`func BNNSTileBackward(UnsafeMutablePointer, UnsafePointer, UnsafePointer?) -> Int32`](/documentation/Accelerate/BNNSTileBackward(_:_:_:)) Applies a tile filter backward to generate an input gradient. ### Macros [`var BNNSDataTypeFloat16: BNNSDataType`](/documentation/Accelerate/BNNSDataTypeFloat16) [`var BNNSDataTypeFloat32: BNNSDataType`](/documentation/Accelerate/BNNSDataTypeFloat32) [`var BNNSDataTypeIndexed8: BNNSDataType`](/documentation/Accelerate/BNNSDataTypeIndexed8) [`var BNNSDataTypeInt16: BNNSDataType`](/documentation/Accelerate/BNNSDataTypeInt16) [`var BNNSDataTypeInt32: BNNSDataType`](/documentation/Accelerate/BNNSDataTypeInt32) [`var BNNSDataTypeInt8: BNNSDataType`](/documentation/Accelerate/BNNSDataTypeInt8) [`var BNNSFlagsUseClientPtr: BNNSFlags`](/documentation/Accelerate/BNNSFlagsUseClientPtr) --- Copyright © 2026 Apple Inc. All rights reserved. | [Terms of Use](https://www.apple.com/legal/internet-services/terms/site.html) | [Privacy Policy](https://www.apple.com/privacy/privacy-policy)