Tools

What to reach for when the result is not what you expected.

Four questions, four tools. The order matters — each one is cheaper than the next, and the compiler has usually answered the first before you ask it.

Question	Tool	Page
What went wrong, and where?	Error types, log levels, IR dumps	Debugging
Is the IR wrong, or the device?	pypto.debug.torch_codegen	Torch codegen
What is on chip, and for how long?	pypto.tools.memory_map	Memory map
Where did the time go?	The L2 swimlane	Performance

The cheap checks first

Two things cost nothing to read, and the compiler has already produced both:

• report/perf_hints.log, written on every compile — what the compiler noticed but did not refuse: transfers below the hardware granularity, a matmul it could not tile, a pipeline depth that did not fit. One summary line also goes to stderr.
• The error message, when there is one. PyPTO distinguishes a user error from an internal one, and the distinction tells you whether to fix your code or file a bug — see Debugging.

See Also

• Precision — the workflow these tools serve when numbers are wrong.
• Performance — the same, when the numbers are right but slow.
• Execution — the compile and dispatch surface being debugged.