Try new rand(Float32) in kernel

[simsurace]

So this rand is basically an alternative to the GPUArrays.gpu_rand?

This is rand from Random, made compatible by redefining default_rng on the device and using our own RNG. So yes, this is definitely the path forward, but we'll obviously have to flesh out the implementation by making sure the necessary APIs are GPU compatible and either overriding more calls or extending the RNG.

Originally posted by @maleadt in JuliaGPU/CUDA.jl#772 (comment)

[simsurace]

@maleadt Just to let you know: I just tried the implementation of rand from CUDA.jl/#772 to replace the GPUArrays.gpu_rand(Float32, CUDA.CuKernelContext(), randstates) by rand(Float32) everywhere in my kernel kernel_BTRS. The resulting speed is almost identical (if slightly faster) using the new function. Note that in Julia 1.5 and using the GPUArrays sampler, the speed is quite a bit faster issue #8.

[maleadt]

That's good, although the memory usage of the new implementation will be (much) larger for now. I plan to revisit it before releasing CUDA 3.0 though. The performance difference with 1.5 is a problem, would be good if you could reduce to something I can have a look at 🙂

[simsurace]

Yes, I will look into that and provide a minimal example!

[maleadt]

Awesome, thanks.

[simsurace]

@maleadt See comment on issue #8

[simsurace]

Still not fully functional, see JuliaGPU/CUDA.jl#788 (comment)