Open
Description
The product_roundoff() test in src/api/reductions/float_arithmetic.rs reports pretty lousy accuracy: there's a 0~3ulp discrepancy between the two multiplication methods on the f32x8 vector test, at least on some platforms. I unblocked the test for now by widening the amount of acceptable difference from 1ulp, but whatever is going on here should be fixed. It may just be an issue with the way the test is written. It is definitely not a new problem, however, as I only discovered it after clearing out all the other backlogged breakage in the repo, and it only appears to arise on the f32x8 vector test, so this isn't a high priority, as hopefully anyone who actually asks for 8 singles multiplied together knows what they are going to get.