preferredStrides

Declaration

var preferredStrides: [Int] { get }

Discussion

During the specialization of an AIModel, a preferred memory layout for a given ndArray value may be set depending on structure of the model and which compute units it is specialized for. In some cases, this can result in a non-contiguous layout being preferred/required by the backing compute. In such case, you are still able to provide InferenceFunction.run normal contiguous ndArray values, however it may incur a copy to the preferred layout. As such, this property provides an opportunity for you to optimize performance by creating your source ndArray value with the preferred striding and avoiding that copy.

guard case .ndArray(let inputDescriptor) = inferenceFunctionDescriptor.inputDescriptor(of: "input") else {
  throw UnexpectedInferenceValueType()
}
let preferredStrides = inputDescriptor.preferredStrides
var ndArray = NDArray(shape: [theShape], scalarType: .float32, strides: preferredStrides)
var view = ndArray.mutableView(as: Float.self)
if let contiguousElements = view.contiguousElements {
  // The preferred strides were a normal contiguous layout
} else {
  // The preferred strides were non-contiguous
  view.withUnsafeMutablePointer { data, shape, strides in
    ... logic which respects whatever strides were preferred ...
  }
}