Merge pull request lammps#4322 from stanmoore1/kk_update_4.4.0

Update Kokkos library in LAMMPS to v4.4.1

cmake/Modules/Packages/KOKKOS.cmake

@@ -8,8 +8,24 @@ endif()
 ########################################################################
 # consistency checks and Kokkos options/settings required by LAMMPS
 if(Kokkos_ENABLE_CUDA)
-  message(STATUS "KOKKOS: Enabling CUDA LAMBDA function support")
-  set(Kokkos_ENABLE_CUDA_LAMBDA ON CACHE BOOL "" FORCE)
+  option(Kokkos_ENABLE_IMPL_CUDA_MALLOC_ASYNC "CUDA asynchronous malloc support" OFF)
+  mark_as_advanced(Kokkos_ENABLE_IMPL_CUDA_MALLOC_ASYNC)
+  if(Kokkos_ENABLE_IMPL_CUDA_MALLOC_ASYNC)
+    message(STATUS "KOKKOS: CUDA malloc async support enabled")
+  else()
+    message(STATUS "KOKKOS: CUDA malloc async support disabled")
+  endif()
+endif()
+if(Kokkos_ENABLE_HIP)
+  option(Kokkos_ENABLE_HIP_MULTIPLE_KERNEL_INSTANTIATIONS "Enable multiple kernel instantiations with HIP" ON)
+  mark_as_advanced(Kokkos_ENABLE_HIP_MULTIPLE_KERNEL_INSTANTIATIONS)
+  option(Kokkos_ENABLE_ROCTHRUST "Use RoCThrust library" ON)
+  mark_as_advanced(Kokkos_ENABLE_ROCTHRUST)
+
+  if(Kokkos_ARCH_AMD_GFX942 OR Kokkos_ARCH_AMD_GFX940)
+    option(Kokkos_ENABLE_IMPL_HIP_UNIFIED_MEMORY "Enable unified memory with HIP" ON)
+    mark_as_advanced(Kokkos_ENABLE_IMPL_HIP_UNIFIED_MEMORY)
+  endif()
 endif()
 # Adding OpenMP compiler flags without the checks done for
 # BUILD_OMP can result in compile failures. Enforce consistency.
@@ -18,6 +34,15 @@ if(Kokkos_ENABLE_OPENMP)
     message(FATAL_ERROR "Must enable BUILD_OMP with Kokkos_ENABLE_OPENMP")
   endif()
 endif()
+
+if(Kokkos_ENABLE_SERIAL)
+  if(NOT (Kokkos_ENABLE_OPENMP OR Kokkos_ENABLE_THREADS OR
+    Kokkos_ENABLE_CUDA OR Kokkos_ENABLE_HIP OR Kokkos_ENABLE_SYCL
+    OR Kokkos_ENABLE_OPENMPTARGET))
+  option(Kokkos_ENABLE_ATOMICS_BYPASS "Disable atomics for Kokkos Serial Backend" ON)
+  mark_as_advanced(Kokkos_ENABLE_ATOMICS_BYPASS)
+  endif()
+endif()
 ########################################################################
 
 option(EXTERNAL_KOKKOS "Build against external kokkos library" OFF)
@@ -45,8 +70,8 @@ if(DOWNLOAD_KOKKOS)
   list(APPEND KOKKOS_LIB_BUILD_ARGS "-DCMAKE_CXX_EXTENSIONS=${CMAKE_CXX_EXTENSIONS}")
   list(APPEND KOKKOS_LIB_BUILD_ARGS "-DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE}")
   include(ExternalProject)
-  set(KOKKOS_URL "https://github.com/kokkos/kokkos/archive/4.3.01.tar.gz" CACHE STRING "URL for KOKKOS tarball")
-  set(KOKKOS_MD5 "243de871b3dc2cf3990c1c404032df83" CACHE STRING "MD5 checksum of KOKKOS tarball")
+  set(KOKKOS_URL "https://github.com/kokkos/kokkos/archive/4.4.01.tar.gz" CACHE STRING "URL for KOKKOS tarball")
+  set(KOKKOS_MD5 "de6ee80d00b6212b02bfb7f1e71a8392" CACHE STRING "MD5 checksum of KOKKOS tarball")
   mark_as_advanced(KOKKOS_URL)
   mark_as_advanced(KOKKOS_MD5)
   GetFallbackURL(KOKKOS_URL KOKKOS_FALLBACK)
@@ -71,7 +96,7 @@ if(DOWNLOAD_KOKKOS)
   add_dependencies(LAMMPS::KOKKOSCORE kokkos_build)
   add_dependencies(LAMMPS::KOKKOSCONTAINERS kokkos_build)
 elseif(EXTERNAL_KOKKOS)
-  find_package(Kokkos 4.3.01 REQUIRED CONFIG)
+  find_package(Kokkos 4.4.01 REQUIRED CONFIG)
   target_link_libraries(lammps PRIVATE Kokkos::kokkos)
 else()
   set(LAMMPS_LIB_KOKKOS_SRC_DIR ${LAMMPS_LIB_SOURCE_DIR}/kokkos)

lib/kokkos/CHANGELOG.md

@@ -1,12 +1,103 @@
 # CHANGELOG
 
+## [4.4.01](https://github.com/kokkos/kokkos/tree/4.4.01)
+[Full Changelog](https://github.com/kokkos/kokkos/compare/4.0.00...4.4.01)
+
+### Features:
+* Introduce new SequentialHostInit view allocation property [\#7229](https://github.com/kokkos/kokkos/pull/7229)
+
+### Backend and Architecture Enhancements:
+
+#### CUDA:
+* Experimental support for unified memory mode (intended for Grace-Hopper etc.) [\#6823](https://github.com/kokkos/kokkos/pull/6823)
+
+### Bug Fixes
+* OpenMP: Fix issue related to the visibility of an internal symbol with shared libraries that affected `ScatterView` in particular [\#7284](https://github.com/kokkos/kokkos/pull/7284)
+* Fix implicit copy assignment operators in few AVX2 masks being deleted [#7296](https://github.com/kokkos/kokkos/pull/7296)
+
+## [4.4.00](https://github.com/kokkos/kokkos/tree/4.4.00)
+[Full Changelog](https://github.com/kokkos/kokkos/compare/4.3.01...4.4.00)
+
+### Features:
+* Add `Kokkos::View` conversions from and to [`std::mdspan`](https://en.cppreference.com/w/cpp/container/mdspan) [\#6830](https://github.com/kokkos/kokkos/pull/6830) [\#7069](https://github.com/kokkos/kokkos/pull/7069)
+
+### Backend and Architecture Enhancements:
+
+#### CUDA:
+* `nvcc_wrapper`: Adding ability to process `--disable-warnings` flag [\#6936](https://github.com/kokkos/kokkos/issues/6936)
+* Use recommended/max team size functions in Cuda ParallelFor and Reduce constructors [\#6891](https://github.com/kokkos/kokkos/issues/6891)
+* Improve compile-times when building with `Kokkos_ENABLE_DEBUG_BOUNDS_CHECK` in Cuda [\#7013](https://github.com/kokkos/kokkos/pull/7013)
+
+#### HIP:
+* Use HIP builtin atomics [\#6882](https://github.com/kokkos/kokkos/pull/6882) [\#7000](https://github.com/kokkos/kokkos/pull/7000)
+* Enable user-specified compiler and linker flags for AMD GPUs [\#7127](https://github.com/kokkos/kokkos/pull/7127)
+
+#### SYCL:
+* Add support for Graphs [\#6912](https://github.com/kokkos/kokkos/pull/6912)
+* Fix multi-GPU support [\#6887](https://github.com/kokkos/kokkos/pull/6887)
+* Improve performance of reduction and scan operations [\#6562](https://github.com/kokkos/kokkos/pull/6562), [\#6750](https://github.com/kokkos/kokkos/pull/6750)
+* Fix lock for guarding scratch space in `TeamPolicy` `parallel_reduce` [\#6988](https://github.com/kokkos/kokkos/pull/6988)
+* Include submission command queue property information into `SYCL::print_configuration()` [\#7004](https://github.com/kokkos/kokkos/pull/7004)
+
+#### OpenACC:
+* Make `TeamPolicy` `parallel_for` execute on the correct async queue [\#7012](https://github.com/kokkos/kokkos/pull/7012)
+
+#### OpenMPTarget:
+* Honor user requested loop ordering in `MDRange` policy [\#6925](https://github.com/kokkos/kokkos/pull/6925)
+* Prevent data races by guarding the scratch space used in `parallel_scan` [\#6998](https://github.com/kokkos/kokkos/pull/6998)
+
+#### HPX:
+* Workaround issue with template argument deduction to support compilation with NVCC [\#7015](https://github.com/kokkos/kokkos/pull/7015)
+
+### General Enhancements
+* Improve performance of view copies in host parallel regions [\#6730](https://github.com/kokkos/kokkos/pull/6730)
+* Harmonize convertibility rules of `Kokkos::RandomAccessIterator` with `View`s [\#6929](https://github.com/kokkos/kokkos/pull/6929)
+* Add a check precondition non-overlapping ranges for the `adjacent_difference` algorithm in debug mode [\#6922](https://github.com/kokkos/kokkos/pull/6922)
+* Add deduction guides for `TeamPolicy` [\#7030](https://github.com/kokkos/kokkos/pull/7030)
+* SIMD: Allow flexible vector width for 32 bit types [\#6802](https://github.com/kokkos/kokkos/pull/6802)
+* Updates for `Kokkos::Array`: add `kokkos_swap(Array<T, N>)` specialization [\#6943](https://github.com/kokkos/kokkos/pull/6943), add `Kokkos::to_array` [\#6375](https://github.com/kokkos/kokkos/pull/6375),  make `Kokkos::Array` equality-comparable [\#7148](https://github.com/kokkos/kokkos/pull/7148)
+* Structured binding support for `Kokkos::complex` [\#7040](https://github.com/kokkos/kokkos/pull/7040)
+
+### Build System Changes
+* Do not require OpenMP support for languages other than CXX [\#6965](https://github.com/kokkos/kokkos/pull/6965)
+* Update Intel GPU architectures in Makefile [\#6895](https://github.com/kokkos/kokkos/pull/6895)
+* Fix use of OpenMP with Cuda or HIP as compile language [\#6972](https://github.com/kokkos/kokkos/pull/6972)
+* Define and enforce new minimum compiler versions for C++20 support [\#7128](https://github.com/kokkos/kokkos/pull/7128), [\#7123](https://github.com/kokkos/kokkos/pull/7123)
+* Add nvidia Grace CPU architecture: `Kokkos_ARCH_ARMV9_GRACE` [\#7158](https://github.com/kokkos/kokkos/pull/7158)
+* Fix Makefile.kokkos for Threads [\#6896](https://github.com/kokkos/kokkos/pull/6896)
+* Remove support for NVHPC as CUDA device compiler [\#6987](https://github.com/kokkos/kokkos/pull/6987)
+* Fix using CUDAToolkit for CMake 3.28.4 and higher [\#7062](https://github.com/kokkos/kokkos/pull/7062)
+
+### Incompatibilities (i.e. breaking changes)
+* Drop `Kokkos::Array` special treatment in `View`s [\#6906](https://github.com/kokkos/kokkos/pull/6906)
+* Drop `Experimental::RawMemoryAllocationFailure` [\#7145](https://github.com/kokkos/kokkos/pull/7145)
+
+### Deprecations
+* Remove `Experimental::LayoutTiled` class template and deprecate `is_layouttiled` trait [\#6907](https://github.com/kokkos/kokkos/pull/6907)
+* Deprecate `Kokkos::layout_iterate_type_selector` [\#7076](https://github.com/kokkos/kokkos/pull/7076)
+* Deprecate specialization of `Kokkos::pair` for a single element [\#6947](https://github.com/kokkos/kokkos/pull/6947)
+* Deprecate `deep_copy` of `UnorderedMap` of different size [\#6812](https://github.com/kokkos/kokkos/pull/6812)
+* Deprecate trailing `Proxy` template argument of `Kokkos::Array` [\#6934](https://github.com/kokkos/kokkos/pull/6934)
+* Deprecate implicit conversions of integers to `ChunkSize` [\#7151](https://github.com/kokkos/kokkos/pull/7151)
+* Deprecate implicit conversions to execution spaces [\#7156](https://github.com/kokkos/kokkos/pull/7156)
+
+### Bug Fixes
+* Do not return a copy of the input functor in `Experimental::for_each` [\#6910](https://github.com/kokkos/kokkos/pull/6910)
+* Fix `realloc` on views of non-default constructible element types [\#6993](https://github.com/kokkos/kokkos/pull/6993)
+* Fix undefined behavior in `View` initialization or fill with zeros [\#7014](https://github.com/kokkos/kokkos/pull/7014)
+* Fix `sort_by_key` on host execution spaces when building with NVCC [\#7059](https://github.com/kokkos/kokkos/pull/7059)
+* Fix using shared libraries and -fvisibility=hidden [\#7065](https://github.com/kokkos/kokkos/pull/7065)
+* Fix view reference counting when functor copy constructor throws in parallel dispatch [\#6289](https://github.com/kokkos/kokkos/pull/6289)
+* Fix `initialize(InitializationSetting)` for handling `print_configuration` setting [\#7098](https://github.com/kokkos/kokkos/pull/7098)
+* Thread safety fixes for the Serial and OpenMP backend [\#7080](https://github.com/kokkos/kokkos/pull/7080), [\#6151](https://github.com/kokkos/kokkos/pull/6151)
+
 ## [4.3.01](https://github.com/kokkos/kokkos/tree/4.3.01)
 [Full Changelog](https://github.com/kokkos/kokkos/compare/4.3.00...4.3.01)
 
 ### Backend and Architecture Enhancements:
 
 #### HIP:
-* MI300 support unified memory support [\#6877](https://github.com/kokkos/kokkos/pull/6877)
+* MI300 support unified memory [\#6877](https://github.com/kokkos/kokkos/pull/6877)
 
 ### Bug Fixes
 * Serial: Use the provided execution space instance in TeamPolicy [\#6951](https://github.com/kokkos/kokkos/pull/6951)

lib/kokkos/CITATION.cff

@@ -0,0 +1,65 @@
+cff-version: 1.2.0
+title: Kokkos
+message: >-
+  If you use this software, please cite the overview paper
+type: software
+authors:
+  - name: The Kokkos authors
+    website: https://kokkos.org/community/team/
+identifiers:
+  - type: url
+    website: https://kokkos.org/kokkos-core-wiki/citation.html
+repository-code: 'https://github.com/kokkos/kokkos'
+url: 'https://kokkos.org/'
+license: Apache-2.0
+preferred-citation:
+  type: article
+  authors:
+  - given-names: Christian R.
+    family-names: Trott
+  - given-names: Damien
+    family-names: Lebrun-Grandié
+  - given-names: Daniel
+    family-names: Arndt
+  - family-names: Ciesko
+    given-names: Jan
+  - given-names: Vinh
+    family-names: Dang
+  - family-names: Ellingwood
+    given-names: Nathan
+  - given-names: Rahulkumar
+    family-names: Gayatri
+  - given-names: Evan
+    family-names: Harvey
+  - given-names: Daisy S.
+    family-names: Hollman
+  - given-names: Dan
+    family-names: Ibanez
+  - given-names: Nevin
+    family-names: Liber
+  - given-names: Jonathan
+    family-names: Madsen
+  - given-names: Jeff
+    family-names: Miles
+  - given-names: David
+    family-names: Poliakoff
+  - given-names: Amy
+    family-names: Powell
+  - given-names: Sivasankaran
+    family-names: Rajamanickam
+  - given-names: Mikael
+    family-names: Simberg
+  - given-names: Dan
+    family-names: Sunderland
+  - given-names: Bruno
+    family-names: Turcksin
+  - given-names: Jeremiah
+    family-names: Wilke
+  doi: 10.1109/TPDS.2021.3097283
+  journal: IEEE Transactions on Parallel and Distributed Systems
+  start: 805
+  end: 817
+  title: "Kokkos 3: Programming Model Extensions for the Exascale Era"
+  volume: 33
+  issue: 4
+  year: 2022

lib/kokkos/CMakeLists.txt

@@ -150,7 +150,7 @@ ENDIF()
 
 
 set(Kokkos_VERSION_MAJOR 4)
-set(Kokkos_VERSION_MINOR 3)
+set(Kokkos_VERSION_MINOR 4)
 set(Kokkos_VERSION_PATCH 1)
 set(Kokkos_VERSION "${Kokkos_VERSION_MAJOR}.${Kokkos_VERSION_MINOR}.${Kokkos_VERSION_PATCH}")
 message(STATUS "Kokkos version: ${Kokkos_VERSION}")

lib/kokkos/Makefile.kokkos

@@ -11,7 +11,7 @@ CXXFLAGS += $(SHFLAGS)
 endif
 
 KOKKOS_VERSION_MAJOR = 4
-KOKKOS_VERSION_MINOR = 3
+KOKKOS_VERSION_MINOR = 4
 KOKKOS_VERSION_PATCH = 1
 KOKKOS_VERSION = $(shell echo $(KOKKOS_VERSION_MAJOR)*10000+$(KOKKOS_VERSION_MINOR)*100+$(KOKKOS_VERSION_PATCH) | bc)
 
@@ -21,11 +21,11 @@ KOKKOS_DEVICES ?= "OpenMP"
 # Options:
 # Intel:    KNC,KNL,SNB,HSW,BDW,SKL,SKX,ICL,ICX,SPR
 # NVIDIA:   Kepler,Kepler30,Kepler32,Kepler35,Kepler37,Maxwell,Maxwell50,Maxwell52,Maxwell53,Pascal60,Pascal61,Volta70,Volta72,Turing75,Ampere80,Ampere86,Ada89,Hopper90
-# ARM:      ARMv80,ARMv81,ARMv8-ThunderX,ARMv8-TX2,A64FX
+# ARM:      ARMv80,ARMv81,ARMv8-ThunderX,ARMv8-TX2,A64FX,ARMv9-Grace
 # IBM:      Power8,Power9
-# AMD-GPUS: AMD_GFX906,AMD_GFX908,AMD_GFX90A,AMD_GFX940,AMD_GFX942,AMD_GFX1030,AMD_GFX1100,AMD_GFX1103
+# AMD-GPUS: AMD_GFX906,AMD_GFX908,AMD_GFX90A,AMD_GFX940,AMD_GFX942,AMD_GFX1030,AMD_GFX1100
 # AMD-CPUS: AMDAVX,Zen,Zen2,Zen3
-# Intel-GPUs: Gen9,Gen11,Gen12LP,DG1,XeHP,PVC
+# Intel-GPUs: Intel_Gen,Intel_Gen9,Intel_Gen11,Intel_Gen12LP,Intel_DG1,Intel_XeHP,Intel_PVC
 KOKKOS_ARCH ?= ""
 # Options: yes,no
 KOKKOS_DEBUG ?= "no"
@@ -41,7 +41,7 @@ KOKKOS_STANDALONE_CMAKE ?= "no"
 
 # Default settings specific options.
 # Options: force_uvm,use_ldg,rdc,enable_lambda,enable_constexpr,disable_malloc_async
-KOKKOS_CUDA_OPTIONS ?= "enable_lambda"
+KOKKOS_CUDA_OPTIONS ?= "disable_malloc_async"
 
 # Options: rdc
 KOKKOS_HIP_OPTIONS ?= ""
@@ -328,12 +328,43 @@ KOKKOS_INTERNAL_USE_ARCH_ICL := $(call kokkos_has_string,$(KOKKOS_ARCH),ICL)
 KOKKOS_INTERNAL_USE_ARCH_ICX := $(call kokkos_has_string,$(KOKKOS_ARCH),ICX)
 KOKKOS_INTERNAL_USE_ARCH_SPR := $(call kokkos_has_string,$(KOKKOS_ARCH),SPR)
 
-KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelGen)
+# Traditionally, we supported, e.g., IntelGen9 instead of Intel_Gen9. The latter
+# matches the CMake option but we also accept the former for backward-compatibility.
 KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN9 := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelGen9)
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN9), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN9 := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_Gen9)
+endif
 KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN11 := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelGen11)
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN11), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN11 := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_Gen11)
+endif
 KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN12LP := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelGen12LP)
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN12LP), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN12LP := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_Gen12LP)
+endif
+KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelGen9)
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN9), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN9 := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_Gen9)
+endif
+KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN_SET := $(shell expr $(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN9)  \
+                                                     + $(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN11) \
+                                                     + $(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN12LP))
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN_SET), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelGen)
+  ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN), 0)
+    KOKKOS_INTERNAL_USE_ARCH_INTEL_GEN := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_Gen)
+  endif
+endif
 KOKKOS_INTERNAL_USE_ARCH_INTEL_DG1 := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelDG1)
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_DG1), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_DG1 := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_DG1)
+endif
 KOKKOS_INTERNAL_USE_ARCH_INTEL_XEHP := $(call kokkos_has_string,$(KOKKOS_ARCH),IntelXeHP)
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_INTEL_XEHP), 0)
+  KOKKOS_INTERNAL_USE_ARCH_INTEL_XEHP := $(call kokkos_has_string,$(KOKKOS_ARCH),Intel_XeHP)
+endif
+# Traditionally the architecture was called PVC instead of Intel_PVC. This
+# version makes us accept IntelPVC and Intel_PVC as well.
 KOKKOS_INTERNAL_USE_ARCH_INTEL_PVC := $(call kokkos_has_string,$(KOKKOS_ARCH),PVC)
 
 # NVIDIA based.
@@ -394,7 +425,8 @@ KOKKOS_INTERNAL_USE_ARCH_ARMV81 := $(call kokkos_has_string,$(KOKKOS_ARCH),ARMv8
 KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX := $(call kokkos_has_string,$(KOKKOS_ARCH),ARMv8-ThunderX)
 KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX2 := $(call kokkos_has_string,$(KOKKOS_ARCH),ARMv8-TX2)
 KOKKOS_INTERNAL_USE_ARCH_A64FX := $(call kokkos_has_string,$(KOKKOS_ARCH),A64FX)
-KOKKOS_INTERNAL_USE_ARCH_ARM := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_ARMV80)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV81)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX2)+$(KOKKOS_INTERNAL_USE_ARCH_A64FX) | bc))
+KOKKOS_INTERNAL_USE_ARCH_ARMV9_GRACE := $(call kokkos_has_string,$(KOKKOS_ARCH),ARMv9-Grace)
+KOKKOS_INTERNAL_USE_ARCH_ARM := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_ARMV80)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV81)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX2)+$(KOKKOS_INTERNAL_USE_ARCH_A64FX)+$(KOKKOS_INTERNAL_USE_ARCH_ARMV9_GRACE) | bc))
 
 # IBM based.
 KOKKOS_INTERNAL_USE_ARCH_POWER8 := $(call kokkos_has_string,$(KOKKOS_ARCH),Power8)
@@ -433,7 +465,6 @@ KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100 := $(call kokkos_has_string,$(KOKKOS_ARCH),
 ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100), 0)
   KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100 := $(call kokkos_has_string,$(KOKKOS_ARCH),NAVI1100)
 endif
-KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1103 := $(call kokkos_has_string,$(KOKKOS_ARCH),AMD_GFX1103)
 
 # Any AVX?
 KOKKOS_INTERNAL_USE_ARCH_AVX        := $(shell expr $(KOKKOS_INTERNAL_USE_ARCH_SNB) + $(KOKKOS_INTERNAL_USE_ARCH_AMDAVX))
@@ -758,6 +789,14 @@ ifeq ($(KOKKOS_INTERNAL_USE_ARCH_A64FX), 1)
   endif
 endif
 
+ifeq ($(KOKKOS_INTERNAL_USE_ARCH_ARMV9_GRACE), 1)
+  tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_ARMV9_GRACE")
+  tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_ARM_NEON")
+
+  KOKKOS_CXXFLAGS += -mcpu=neoverse-v2  -msve-vector-bits=128
+  KOKKOS_LDFLAGS += -mcpu=neoverse-v2  -msve-vector-bits=128
+endif
+
 ifeq ($(KOKKOS_INTERNAL_USE_ARCH_ZEN), 1)
   tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_ZEN")
   tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AVX2")
@@ -1119,11 +1158,6 @@ ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1100), 1)
   tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_GPU")
   KOKKOS_INTERNAL_HIP_ARCH_FLAG := --offload-arch=gfx1100
 endif
-ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AMD_GFX1103), 1)
-  tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_GFX1103")
-  tmp := $(call kokkos_append_header,"$H""define KOKKOS_ARCH_AMD_GPU")
-  KOKKOS_INTERNAL_HIP_ARCH_FLAG := --offload-arch=gfx1103
-endif
 
 
 ifeq ($(KOKKOS_INTERNAL_USE_HIP), 1)
@@ -1216,6 +1250,8 @@ ifeq ($(KOKKOS_INTERNAL_DISABLE_BUNDLED_MDSPAN), 0)
 endif
 tmp := $(call kokkos_append_header,"$H""define KOKKOS_ENABLE_IMPL_MDSPAN")
 
+tmp := $(call kokkos_append_header,"$H""define KOKKOS_ENABLE_IMPL_REF_COUNT_BRANCH_UNLIKELY")
+
 KOKKOS_INTERNAL_LS_CONFIG := $(shell ls KokkosCore_config.h 2>&1)
 
 ifeq ($(KOKKOS_INTERNAL_LS_CONFIG), KokkosCore_config.h)