How REX Finished The LLVM 21 GPU Benchmark Suite

Thu, 30 Apr 2026 00:00:00 +0000

The previous post removed the last confusing process-lifetime tax from the LLVM 21 comparison. GPU-total profiling had already shown that the suspicious pathfinder and srad rows were not device-side native LLVM wins. The process-exit teardown investigation then explained why some short wall-clock runs still looked worse for REX even when kernel and transfer totals were clean.

At that point, the LLVM 21 performance work finally had a coherent scoreboard.

This post is the wrap-up for that state. It is not another chronological debugging log. The chronological work is covered by the previous posts. This one answers a different question:

Wrapup on ./Code

How REX Finished The LLVM 21 GPU Benchmark Suite