PyPTO Coding

Use this chapter when writing or reviewing PyPTO-Lib kernels. The first six pages are the canonical coding style, split by what you are writing:

Page	Covers
L2 Programming	The two authoring forms, pl.Out / pl.InOut directions, pl.at regions and their optimizations, mixed cube + vector regions, and dynamic B / S shapes
Operations	Vector, cube, MTE and scalar ops — the four families a kernel body is written from
Loops	pl.range, pl.unroll, pl.parallel, pl.pipeline, pl.spmd, and where each one is legal
L3 Programming	Multi-card kernels: the host driver, HCCL window buffers, cross-rank data movement, and the notify / wait protocols that order it
Golden and Run	Writing a kernel's validation: specs, the Torch reference, the run call, and the conventional CLI flags
Naming and Comments	Constant naming, what a comment may say, and where allocations go

One further page covers work that leaves PyPTO entirely:

• CCE Extern Kernel covers hand-written mixed cube and vector kernels called through pl.jit.extern, including runtime, ABI, synchronization, and validation constraints.

Start with L2 Programming for every kernel change. Add L3 Programming when the kernel spans more than one card. Use the extern-kernel page only when the implementation crosses from PyPTO into hand-written CCE code.