https://blog.ouankou.com/tags/cpu/Recent content in Cpu on ./CodeHugoen-US© Anjia WangWed, 15 Apr 2026 15:24:20 -0700https://blog.ouankou.com/2026/04/01/how-rex-proves-lowered-openmp-semantics-with-cpu-equivalence-tests/Wed, 01 Apr 2026 00:00:00 +0000https://blog.ouankou.com/2026/04/01/how-rex-proves-lowered-openmp-semantics-with-cpu-equivalence-tests/<p>The previous post in this series focused on <code>lowering_rodinia</code>: reduced Rodinia-like inputs, stable invariants, and why GPU lowering should be tested on structural facts instead of brittle full-file snapshots.</p> <p>That leaves one obvious gap.</p> <p>Structural correctness is necessary, but it is not enough.</p> <p>A lowerer can still emit the right number of kernels, the right runtime include, the right helper ordering, and the right launch API shape while changing the program’s meaning. A broken reduction, a misplaced barrier, or an incorrect <code>if</code> plus <code>num_threads</code> interaction will not necessarily show up in a structural verifier. It will show up when the transformed program actually runs.</p>