Fabian Knorr
Fabian Knorr
Here's a minimal example of this causing a miscompilation on the sm_75 CUDA platform: ```c++ #include constexpr int global = 20; struct Foo { int a; int bar() const {...
I have added some documentation and re-introduced the old accessor API including tests. I think that Level Zero support and optimization beyond the original state of reductions should happen in...
@illuhad This is now complete AFAIC.
Rebased to resolve conflicts; I believe the CI failure is spurious (network outage).
@DieGoldeneEnte I'm reluctant to put any more work into this as it hasn't seen any action in more than a year. But you're free to extract a sub-PR of features...
I have started [working on this](https://github.com/fknorr/hipSYCL/tree/reduction-api), the API surface is easy to adapt but I will have to see how difficult handling of reductions without identities and `poperty::reduction::initialize_to_identity` will turn...
In this distributed setting, would be interesting to benchmark common communication patterns (gather / scatter / alltoall) as well as scalar reductions from host buffers against the optimized MPI collectives.
The CI failure on `dpcpp:HEAD` is due to a bug in Clang 15 and not related to this PR.
> As discussed in person, the `std::in_place` constructor of `host_object` currently forwards to the initializer-list constructor of `T`, if one exists, which is unexpected. I have replaced the "universal" initializer...