metal-kernel

This skill guides you through implementing Metal kernels for PyTorch operators on Apple Silicon. Important: The goal of this skill is to use native Metal capabilities via the c10/metal/ infrastructure, NOT MPSGraph. Native Metal kernels provide better control, performance, and maintainability. There are two workflows covered by this skill: Both workflows involve: Location: aten/src/ATen/native/native_functions.yaml Find the operator entry and add MPS dispatch: When migrating an existing operator from MPSGraph to native Metal, consolidate the dispatch entry : Key change: Replace MPS: my_op_out_mps with adding MPS to the shared dispatch line (e.g., CPU, CUDA, MPS: my_op_out ). Dispatch naming conventions: Location: aten/src/ATen/native/mps/kernels/ For binary operations, use the convenience macros defined in BinaryKernel.metal : Common patterns: Example for atan2 (supports both float and int inputs): Note on complex types: Complex numbers in Metal are represented as vector types: Use is_complex_v to specialize for complex types in functors. utils.h: special_math.h: indexing.h: Location: aten/src/ATen/native/mps/operations/ Choose or create an appropriate file based on operation type: For structured kernels that use the TensorIterator pattern: For unary operations: When migrating from MPSGraph, also remove the old implementation: Remove from BinaryOps.mm (or UnaryOps.mm): Add to...