[v2] CUDA version uses fast math, SYCL does not

UoB-HPC / miniBUDE

A BUDE virtual-screening benchmark, in many programming models

Apache License 2.0

25 stars 13 forks source link

[v2] CUDA version uses fast math, SYCL does not #31

Closed illuhad closed 3 months ago

illuhad commented 4 months ago

At least the v2 version of miniBUDE (have not looked at the other one) uses --use_fast_math with CUDA, while not using -ffast-math with SYCL. This causes results to not be comparable between the two. Fast math significantly impacts performance for this app.

Note that this does not impact icpx which uses -ffast-math by default.

tom91136 commented 3 months ago

v2 is now main, so this should be resolved via 40ccb57.