Support for rand from Julia Base on device code
Closed this issue · 2 comments
williamfgc commented
CUDA.jl has support for using functionality like rand(Distributions.Uniform(-1, 1)) on kernel code, which AMDGPU.jl flags as invalid LLVM-IR.
What's the best way to add this kind of functionality? Happy to help.
vchuravy commented
Also needed by @utkarsh530 in SciML/DiffEqGPU.jl#227
jpsamaroo commented
Probably in the short term we can copy-pasta most of CUDA.jl's implementation.