GPU Kernel Baseline

When to Use

Use this skill when the user provides PyTorch logic or a kernel demo and asks to:

• Write a GPU kernel for the target platform.
• Build a baseline from scratch.
• Prepare kernel.py, reference.py, test_kernel.py, and baseline_report.md for later profile-driven optimization.

Workflow

This stage first understands the PyTorch semantics, then learns the framework APIs (CuteDSL or FlyDSL) through enabled knowledge tools, implements kernel.py and test_kernel.py, validates correctness, records performance, writes baseline_report.md, and writes memory/v0.json.

The orchestrator installs discovered plugin instructions at .atrex_plugins/instructions.md.

Phase 1: Understand PyTorch Semantics