True. And the resulting PTX is worse when it's generated from OpenCL C instead of CUDA C. I tested that recently with a toy FFT kernel and the CUDA pipeline pr

True. And the resulting PTX is worse when it's generated from OpenCL C instead of CUDA C. I tested that recently with a toy FFT kernel and the CUDA pipeline produced a lot more efficient FMA instructions.