This week's theme: passing tests is wonderful. Knowing what they didn't catch is even better.
GPU kernels: symbolic checking looks beneath the tolerances
Experimental tool Racer checks PTX kernels against reference implementations using symbolic expressions. Its authors report finding 82 bugs across 23 libraries, including 76 they say were previously unknown.
That matters because floating-point tests often need tolerances. A kernel can be "close enough" numerically while still being logically wrong. Racer aims to catch some of those errors. For example, a reduction kernel accidentally adds a tiny input twice.…