I’m currently writing an sph solver. I have both a version running solely on the CPU and one running on the GPU and it is very important that they are able to produce the exact same result. The problem is that I get a small rounding error on the GPU when executing a sqrtf() call. Even tho the rounding error is very small it causes noticeable visual differences between the two versions after a few iterations.
Is there a way to reduce the accuracy on the host side to match the one on the GPU side? This seams like the best solution since it is not an accuracy problem.
Thankful for all responses and thoughts upon this matter.