[Openmp-commits] [PATCH] D51107: [LIBOMPTARGET] Add support for mapping of lambda captures.

[Alexey Bataev via Phabricator via Openmp-commits]

ABataev added a comment.

In https://reviews.llvm.org/D51107#1211035, @Hahnfeld wrote:

> In https://reviews.llvm.org/D51107#1210813, @ABataev wrote:
>
> > I tried it, it won't work.
> >
> > 1. Lambdas are captured as firstprivates, thus we cannot translate the host address of the lambda to the device address. `MEMBER_OF` does not work here because of that.
>
>
> I agree that this doesn't work with the current code. IIRC the reason is that libomptarget has two loops dealing with `map` clauses: one in `target_data_begin` (which also handles `MEMBER_OF`) and the one in `target` which this patch is modifying. The problem is that `firstprivate` arguments are handled in the second, so the one in `target_data_begin` (which runs first) doesn't have a device address to resolve entries with `MEMBER_OF`. IMO the proper solution would be to have all allocations in one single loop in `target_data_begin`. The loop in `target` should only deal with constructing the arguments passed to the device.
>
> To implement this efficiently we need a temporary data structure for each mapping so that we can store the device address allocated in `target_data_begin`. That pointer could then be queried in `target` to populate `tgt_args`.
>  (Ideally the data structure would wrap each entry in the `args_base`, `args`, `arg_sizes`, and `arg_types` arrays into a nice object. This would have functions to hide all this bitmasking to find out the type.)
>
> Long story short, I think that lambdas are indeed `firstprivate`. What needs to be fixed are `firstprivate` "mappings" together with `MEMBER_OF` (@AlexEichenberger  this already works in lomp, right?).


This requires some significant redesign of the library itself + possibly the compiler. It would be better if somebody else, more familiar with libomptarget, could implement this.


Repository:
  rOMP OpenMP

https://reviews.llvm.org/D51107