PTX runQ on Nvidia-free hardware

[michael-swan]

Is your feature request related to a problem? Please describe.
runQ variant of PTX runners cannot build on non-target machines, i.e. a machine w/o Nvidia GPU or drivers installed.

Describe the solution you'd like
For the runQ path, we must not compile PTX through the module compilation API exposed by libcuda and instead use the CLI utilities such as ptxas and LLVM utils.

It would also be necessary to provide a way to specify the target information to runQ style TH functions, since that information is typically gleaned from libcuda.

[michael-swan]

@tomsmeding You mentioned removing llvm-hs as a dependency and instead shelling out, perhaps that makes the above more possible. I would be curious to gauge how challenging that would be.

[tomsmeding]

We indeed dropped llvm-hs as a dependency, but cuda is still there. There are long-term plans to remove that dependency as well, but that would be in favour of linking to libcuda directly and binding the functions we need ourselves, to avoid having to keep the Haskell bindings package up to date.

The runQ TH pipeline is not receiving much testing or love, at the moment. Many things are possible, but unfortunately we do not have much time to devote to Accelerate UX issues, so few things actually happen. Thank you for opening this feature request, because it is always useful to track interest and discuss possibilities.

While we would welcome contributions for this feature, please be aware that we have a big rewrite ongoing of the accelerate-llvm backends. The D.A.A.LLVM.PTX.Compile module is mostly unchanged, but many other modules have been. If a PR to address this issue is very difficult to port to the rewritten backend, we may be hesitant to merge it.

[michael-swan]

Good to know, thanks. I'll base all future contributions on this rewrite until it is merged. I'm looking forward to those improvements.

[tmcdonell]

IIRC the problem is that it initialises the device in order to probe various capabilities, but that’s it. Removing that part (e.g. add an option to runQ to tell it to compile for compute capability X.Y, rather than looking it up itself) should avoid the problem.