Started by timer00002982696114861463.855119841427ca8550dd3d18acaa29a25d96d059dd31e7bf1427ca8550dd3d18acaa29a25d96d059dd31e7bfrefs/remotes/origin/develop1427ca8550dd3d18acaa29a25d96d059dd31e7bf1427ca8550dd3d18acaa29a25d96d059dd31e7bfrefs/remotes/origin/develop1427ca8550dd3d18acaa29a25d96d059dd31e7bf1427ca8550dd3d18acaa29a25d96d059dd31e7bfrefs/remotes/origin/develophttps://github.com/kokkos/kokkos/10199testReportfalse#98429826963875102Kokkos_nightly #984984false98455434FAILURE1791347478309https://cloud1.cees.ornl.gov/jenkins-ci/job/Kokkos_nightly/984/- tpls/desul/include/desul/atomics/Operator_Function_Objects.hpp9a98f4f33032594ec6ceae8b2a41ab1f0e4c7c941791222844000https://cloud1.cees.ornl.gov/jenkins-ci/user/dalg24Damien L-Gdalg24@gmail.comDesul atomics: Relax semantics of atomic floating-point min/max
Treat negative and positive zero as equivalent, meaning no guarantee on
the sign of the result with +0 and -0 as arguments to
atomic_fetch_{min,max}
Signed-off-by: Damien L-G <dalg24@gmail.com>
2026-10-05 11:54:04 -06009a98f4f33032594ec6ceae8b2a41ab1f0e4c7c94Desul atomics: Relax semantics of atomic floating-point min/maxedittpls/desul/include/desul/atomics/Operator_Function_Objects.hpp
- tpls/desul-hash.txt832da541bc34e8b84f4afc815bf8f0e2375c2eda1791222844000https://cloud1.cees.ornl.gov/jenkins-ci/user/dalg24Damien L-Gdalg24@gmail.comUpdate desul commit hash value
Signed-off-by: Damien L-G <dalg24@gmail.com>
2026-10-05 11:54:04 -0600832da541bc34e8b84f4afc815bf8f0e2375c2edaUpdate desul commit hash valueedittpls/desul-hash.txt
- core/src/Cuda/Kokkos_Cuda_BlockSize_Deduction.hpp23b5d90394f802d3853b403354acfcc5bff0a3a11791234526000https://cloud1.cees.ornl.gov/jenkins-ci/user/masterleinadDaniel Arndtarndtd@ornl.govCuda: Prefer block sizes above 32 threads
Signed-off-by: Daniel Arndt <arndtd@ornl.gov>
Co-authored-by: Nathan Ellingwood <ndellin@sandia.gov>
Signed-off-by: Daniel Arndt <arndtd@ornl.gov>
2026-10-05 17:08:46 -040023b5d90394f802d3853b403354acfcc5bff0a3a1Cuda: Prefer block sizes above 32 threadseditcore/src/Cuda/Kokkos_Cuda_BlockSize_Deduction.hpp
- core/src/Cuda/Kokkos_Cuda_KernelLaunch.hppcore/src/Cuda/Kokkos_Cuda_Instance.cppcore/unit_test/cuda/TestCuda_CUPTI_LaunchRoute.cppcore/src/Cuda/Kokkos_Cuda_Instance.hppcore/unit_test/CMakeLists.txt1427ca8550dd3d18acaa29a25d96d059dd31e7bf1791337766000https://cloud1.cees.ornl.gov/jenkins-ci/user/noreplynoreplynoreply@github.comRemove CUDA constant memory dispatch when `__grid_constant__` is available at compile time
Unconditionally uses __grid_constant__ when available with NVCC.
Adds a test to verify launch mechanism.
This is effectively reverting prior work to opt out of grid constant launch.
While there was a real perf regression reported in one application, a (extremely) deep dive
into the issue shows it was not a systemic problem but just some corner case of compiler
optimization choices. We can't work around all of them, and is just as likely some other code
gets the opposite effect. Furthermore, a workaround on the user code side was available.
Signed-off-by: Namjae Choi <Namjae.Choi@inl.gov>
Signed-off-by: Evan Weinberg <eweinberg@nvidia.com>
Co-authored-by: Namjae Choi <Namjae.Choi@inl.gov>
2026-10-06 19:49:26 -06001427ca8550dd3d18acaa29a25d96d059dd31e7bfRemove CUDA constant memory dispatch when `__grid_constant__` is available at compile timeeditcore/src/Cuda/Kokkos_Cuda_Instance.cppeditcore/src/Cuda/Kokkos_Cuda_Instance.hppeditcore/src/Cuda/Kokkos_Cuda_KernelLaunch.hppaddcore/unit_test/cuda/TestCuda_CUPTI_LaunchRoute.cppeditcore/unit_test/CMakeLists.txt
githttps://cloud1.cees.ornl.gov/jenkins-ci/user/masterleinadDaniel Arndthttps://cloud1.cees.ornl.gov/jenkins-ci/user/noreplynoreplyhttps://cloud1.cees.ornl.gov/jenkins-ci/user/dalg24Damien L-Gfalse985https://cloud1.cees.ornl.gov/jenkins-ci/job/Kokkos_nightly/985/983https://cloud1.cees.ornl.gov/jenkins-ci/job/Kokkos_nightly/983/