Out of memory during precompile

[juliagpu]

Out of memory during precompile

[maleadt]

Would it be acceptable to specify CUDANATIVE_INITIALIZE=false as environment variable to prevent CUDAnative.jl from initializing device 0, and require you to device!(whatever) explicitly?
Doing so automatically upon the first API call, ie. how CUDA does it, is tricky, as I don't want to prefix every CUDAnative & CUDAdrv function with a initialized || do_initialize() check, nor pay the performance penalty of more automatic schemes (eg. a hook mechanism for @apicall).

[maleadt]

Adding a simple, well-typed hook turned out not so expensive, so we have full lazy initialization as per CUDA with JuliaGPU/CUDAnative.jl@c36ce17

$ DEBUG=true julia examples/vadd.jl
┌ Debug: Initializing CUDA for default device 0 after API call
│   name = "GeForce GTX TITAN"
│   apicall = :cuMemAlloc
└ @ CUDAnative init.jl:28

Adding an explicit device! disables this mechanism.

[MikeInnes]

Can you reproduce this just loading CUDAnative? I expect that's where the fix needs to be.

[denizyuret]

Yes, I can confirm that using CUDAnative causes the same problem. Just fill up the memory of gpu0 and it does not see there is other gpus to use.