I would like to test the accuracy of the optimizations added by the “-cl-fast-relaxed-math” or “-cl-unsafe-math-optimization” options. However, I have already observed a negative impact on accuracy due to “-cl-mad-enable”. Since both of the previous options enable fused multiply-add I need to find a way to “trick” the compiler (or better yet inform the compiler) when I don’t want a fused multiply-add to occur. In cuda this is achieved with the intrinsic function __fmul_rn, but I can’t seem to find an equivalent in OpenCL (there is no function like native_multiply). Does anyone know of a way of doing this?
Thanks in advance,