I’m working on code that optionally uses double precision via the cl_khr_fp64 extension. So far, so good. It works nicely (and is surprisingly fast) on a GTX 285. However, the code als compiles and runs on GPUs that are not capable of dual precision, specifically a GT 240. It does “something”, but obviously the calculated values are completely wrong.
Isn’t that a pretty harsh bug? In this case either compilation or kernel execution should fail.
I’m on Linux x86_64 with 256.35 drivers.