generateDenseLayerWithAccuracy

fun generateDenseLayerWithAccuracy(node: GraphNode, addNode: GraphNode? = null, denseIndex: Int? = null): LayerCode(source)

Generates C code for Dense layer operations with exact numerical consistency. Follows existing DefaultCpuOps implementation patterns for matrix-vector multiplication.

The generated code implements: output = input * weight^T + bias This matches the Linear layer forward pass: input.matmul(weight.t()) + bias

Enhanced for numerical accuracy by:

  • Using consistent floating-point operations with DefaultCpuOps

  • Implementing proper accumulation order to minimize floating-point errors

  • Adding bounds checking for array access safety

  • Ensuring direct output writing optimization when possible

Return

LayerCode containing generated C code fragment

Parameters

node

GraphNode representing a Dense/Linear layer