division in CUDA Fortran

[Tuan]

I experience slow performance when using division operator in the kernel. May I ask how CUDA Fortran map a division to the native code, as in CUDA C there are different ways to do division (which work differently for integer, FP32, FP64…).

Thanks,
Tuan

[MatColgrove]

Hi Tuan,

By default, we use the more precise but slower “fdiv” function to perform divides. For the faster but less precise divide operator (“/”) use “-Mcuda=fastmath”.

Hope this helps,
Mat

[Tuan]

Thanks, Mat