GpuMlpWeights

1 min read

Import: import system.gpu;

This reference lists the public declarations for this part of the library. See Reading API signatures for parameter and result notation. Shader-only declarations can be used only inside @compute, @vertex, @fragment, inferred helpers and ray tracing shader functions; using them elsewhere is a compile error.

GpuMlpWeights

One uploaded neural network layer: a weight matrix plus biases, stored in the layout the device's cooperative vector math reads fastest. Create one with GpuDevice.upload_mlp and pass it to a shader parameter, giving its sizes with @coop_shape(inputs, outputs). Dispatching a layer whose sizes differ from the parameter's fails with GpuError.InvalidResource. Inside the shader, apply evaluates the layer. Shaders that use it compile for GpuTarget.Vulkan, Dx12 and Cuda, and run on devices with GpuFeature.CooperativeVectors; elsewhere dispatching fails with GpuError.UnsupportedCapability.

public class GpuMlpWeights<T: unmanaged>

inputs

The number of values the layer reads.

public get inputs(): int

outputs

The number of values the layer produces.

public get outputs(): int

apply

Computes weights * input + biases inside a shader. input has inputs elements and the result outputs elements.

public extern fun apply(input: GpuCoopVec<T>): GpuCoopVec<T>