encode(commandBuffer:inputs:outputs:)

Encodes GPU commands to evaluate the custom layer.

Declaration

optional func encode(commandBuffer: any MTLCommandBuffer, inputs: [any MTLTexture], outputs: [any MTLTexture]) throws

Parameters

commandBuffer:
A command buffer that defines the work the layer performs on the GPU.
inputs:
A texture array that represents the layer’s inputs.
outputs:
A texture array that represents the layer’s outputs.

Mentioned in

Creating and Integrating a Model with Custom Layers

Discussion

Implement this method to use the GPU to evaluate your layer. Fill commandBuffer with the GPU commands that evaluate the layer. Don’t commit the command buffer in this method; Core ML executes the command buffer after this method returns.

Improve your layer’s performance by caching the MTLComputePipelineState instances you create and intend to reuse in subsequent calls.

Implementing this method doesn’t guarantee that Core ML evaluates this layer on the GPU. For example, Core ML may evaluate the layer on the CPU if the system doesn’t have enough GPU’s resources to run the custom layer.

If you don’t implement this method, Core ML instead uses evaluate(inputs:outputs:).

For more information about using the GPU for general purpose programming, see Compute Processing.