Daniel Arndt

Check matching View extents in parallel algorithms

1

Copied from https://kokkosteam.slack.com/archives/G5CBLMFLP/p1707745107838919: > in the std-like parallel algorithms, when passing 2 views, we don't seem to actually check that they have the same size, we just take begin and...

SYCL: Improve team_reduce implementation

3

Related to https://github.com/kokkos/kokkos/pull/5334#issuecomment-1787497628. #5334 exposed that `team_reduce` in `SYCL` doesn't use arbitrarily large types in general. The current implementation relies on using some fixed-size local memory (that is also used...

Add test for TeamPolicy array reduction and fix Cuda and HIP

5

In reference to https://github.com/kokkos/kokkos/pull/6293#issue-1810833445. It seems that we don't test array reductions for TeamPolicy in the CI. This pull request copies a very similar test to use array reductions instead...

SYCL: Clean up TeamPolicy implementation

1

This pull request is based on #6562 and avoids using `m_team_reduce`. The advantage is that we don't need to explicitly allocate memory for `TeamPolicy` but only when needed. The problem...

Allow specifying scratch level for ScratchSpace at compile-time

7

The oneAPI compiler emits warnings whenever scratch memory is used since we don't expose what the correct address space is. For level 0, we essentially want `sycl::local_ptr` and for level...

Cuda: Avoid repeated template instantiations for UniqueToken and idForInstance

1

https://github.com/dealii/dealii/issues/16065 shows that in a `deal.II` build with `Kokkos+Cuda` some `Kokkos` templates are the most expensive templates as reported by https://github.com/aras-p/ClangBuildAnalyzer. Specifically, these are `idForInstance` and the `Kokkos::View` used in...

SYCL: Unroll shuffle loops for top-level parallel_reduce and parallel_scan

1

Similar to #6562, unrolling the loops for shuffles improves performance significantly. For a simple dot product benchmark we are seeing on Intel PVC | elements | old shuffle (GB/s) |...

Avoid gcc-13 warning about out-of-range access for subviews

12

Using gcc-13 I am seeing warnings such as ``` [...] [ 27%] Building CXX object core/unit_test/CMakeFiles/Kokkos_CoreUnitTest_Serial2.dir/serial/TestSerial_SubView_c07.cpp.o In file included from /tmp/kokkos/core/src/Kokkos_View.hpp:490, from /tmp/kokkos/core/src/Kokkos_Parallel.hpp:31, from /tmp/kokkos/core/src/Kokkos_MemoryPool.hpp:26, from /tmp/kokkos/core/src/Kokkos_TaskScheduler.hpp:34, from /tmp/kokkos/core/src/Serial/Kokkos_Serial.hpp:36, from...

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

13

Related to thread-safety questions in #6051 and #4385. This pull request ensures that kernels submitted to the same execution space instance from multiple threads don't run concurrently. Also, calling `fence`...

Consistency between Range/MDRange/TeamPolicy reduction interface and nested reduction interface

2

As noted in https://kokkosteam.slack.com/archives/C5BGU5NDQ/p1690480258950949?thread_ts=1690468200.131139&cid=C5BGU5NDQ, we currently don't support functors as reducers for nested reductions but for Range/MDRange/TeamPolicy reductiions. We should consider unifying the interfaces so that an example like ```C++...