cos and sin huge compilation time in double precision

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi Everyone,

When trying to use cos and sin math functions in double-precision inside and OpenCL kernel I observed that the JIT compilation time was huge, typically several seconds. I tried to use the faster less-accurate native_sin() and native_cos() functions. The kernel doesn’t compile because the native_cos and native_sin don’t seem to be implemented in double precision despite the OpenCL specs state that they are.

  1. Could you please explain why compilation time is so long?
  2. Are the double-precision native_cos and sin functions implemented on Nvidia OpenCL?

Many Thanks

2

No one viewed my post or no one has an idea on this issue.

Thx??

3

Seems like native is only single-precision. (Native section only mentions single-precision)