[Openmp-dev] Libomptarget fatal error 1: '#pragma omp requires unified_shared_memory' not used consistently!

[Narayanaswamy, Ravi via Openmp-dev]

We can also discuss this in tomorrows meeting 
Ravi

-----Original Message-----
From: Openmp-dev <openmp-dev-bounces at lists.llvm.org> On Behalf Of Joachim Protze via Openmp-dev
Sent: Tuesday, March 02, 2021 8:46 AM
To: Alexey.Bataev <a.bataev at outlook.com>; Ye Luo <xw111luoye at gmail.com>
Cc: openmp-dev <openmp-dev at lists.llvm.org>
Subject: Re: [Openmp-dev] Libomptarget fatal error 1: '#pragma omp requires unified_shared_memory' not used consistently!

Am 02.03.21 um 15:58 schrieb Alexey.Bataev:
> Ye, generally speaking stack allocations can be replaced by managed 
> mem allocation. We have very similar functionality for #pragma omp 
> allocate, where static alloca can be replaced by __kmpc_alloc calls. 
> But, most probably, it would be good to extend #pragma omp allocate to 
> add managed memory support. Otherwise it would hard to implement it. 
> Or it can be implemented in the backend by adding 
> analysis/transformation of static allocas into managed memory alloc 
> calls if it is proved that the variable is supposed to be used in USM mode.

I'm not completely up-to-date with what the OpenMP allocator traits can de/prescribe (can they have an effect to stack allocations without explicit allocate directive?). But from my perspective, compiler and runtime should support two modes:

- completely rely on hardware support for USM, but error out at runtime, if the runtime determines that hardware support is not available during execution. Would it be possible to encode this into the sm-specific runtime library?

- rely on the use of an OpenMP allocator which uses managed memory. For compilation units with require USM, the compiler must use __kmpc_alloc for all stack variables, where it is not possible prove that they will not need managed memory, e.g., if they can escape to a function potentially spawing a target region.

- Joachim

FYI: I also posted a related question on this topic at the omp-lang list, asking for clarification on the expected behavior, if USM cannot be guaranteed.