Open bwmeyers opened 11 months ago
We need to ensure that the HIP versions produce sensible results that match the CUDA-only version, both in terms of accuracy (i.e., can we do a byte-to-byte comparison?) and also from a performance perspective.
We need to ensure that the HIP versions produce sensible results that match the CUDA-only version, both in terms of accuracy (i.e., can we do a byte-to-byte comparison?) and also from a performance perspective.