In graphics (Shader) environments, these instructions only operate on scalar integers. Ban the floating point case. OpenCL supports atomic_float