Pour ceux que ca interresse, j'ai posé la question sur le forum officiel d'opencl https://www.khronos.org/message_boards/viewtopic.php?f=28&t=5566
La réponse:

It's interesting that there is a way to query device capabilities (using "CL_DEVICE_SINGLE_FP_CONFIG"), but not set them.

Pour CUDA, ca existe. Voir https://devtalk.nvidia.com/default/topic/483264/cuda-programming-and-performance/static-rounding-modes/