Multi-kernel compilation units (multiple entry points per module)

[tetsuo-cpp]

Summary

Allow a single Forth source (or module) to define multiple GPU kernel entry points and non-kernel device functions, instead of one \! kernel per compilation unit.

Motivation

Real GPU libraries ship several related kernels (e.g. attn_fwd, attn_fwd_kv, softmax_fix). Today the model is effectively one kernel name per file/compile. Multi-entry modules simplify demos, metadata, and host APIs.

Proposed model

\! module attention

\! kernel attn_fwd
\! param ...
...

\! kernel attn_fwd_cached
\! param ...
...

: helper { ... -- }   \ device function shared by both
  ...
;

Semantics

• Multiple gpu.func kernels inside one gpu.module
• Shared device words callable from any kernel in the unit
• Params/shared scopes: per-kernel (preferred) vs module-global — decide and document
• warpforthc emits one PTX module; metadata lists all kernels

Acceptance criteria

• Two kernels in one source both appear in PTX and are launchable by name
• Shared helper word usable from both (once params-in-words or explicit args exist)
• Metadata schema supports multiple kernels
• LIT tests for parsing and GPU-module structure
• Error on duplicate kernel names

Related

• Params/shared in word definitions
• Kernel metadata emission
• Python package launch-by-name