Performance regression using `^`

[luraess]

As reported in luraess/JuliaGPUPerf#2 and luraess/JuliaGPUPerf#3, there is an issue significantly affecting performance when doing ^ operation within GPU kernels.

The Int32 on Int32 case (luraess/JuliaGPUPerf#2) may have been fixed as upon suggestion from @vchuravy by using

my_pow(x, p) = ccall("llvm.powi.f32.i32", llvmcall, Float32, (Float32, Int32), x, p)
#[...]
A[ix,iy] = B[ix,iy] + s*my_pow(C[ix,iy], pow_int)

But the Float32 and Float64 cases are still lacking behind.