Notes

Longer pieces on instruction sets, and on how this reference is built and checked. New notes also appear in What's new and its RSS feed.

  1. What a zero is worth: checking a GPU decoder

    The AMD GPU decoder was checked against 162,054 lines LLVM assembled and named none of them wrongly. How far that zero reaches, why most lines get no answer, and the first version of the check, which would have reported zero while checking nothing.

  2. Measuring the reference instead of the data

    How we measured the accuracy of an instruction reference built by a language model, the 18 wrong RISC-V encodings it found, and the four times the audit itself was the problem.