Changes

SYCL: Cleanup device selection

Adding SECURITY.md file

This is based on the one from OSSF Scorecard

Cleanup KOKKOS_CONFIGURE_CORE

Avoid detecting unwanted fences in the parallel_scan_no_fence test

Add test for current CTAD support with RangePolicy (#6803)

* Add test for current CTAD support with RangePolicy

Co-authored-by: Nevin Liber <nliber@anl.gov>

* Rework CTAD test to avoid "memeber <bla> was declared but never referenced" warnings with icpc 19

* Attempt to fix CI

* Attempt to eliminate maybe unused warning in icpc

* Disable CTAD tests for nvcc < 11.2
as compiler bugs prevent CTAD expressions inside decltype

---------

Co-authored-by: Nevin Liber <nliber@anl.gov>
Co-authored-by: Nevin ":-)" Liber <nliber+github@gmail.com>

multi-GPU support: Add test for all policies (#6782)

* Cuda multi-GPU support: Test with managed and unmanaged Views

* Move check for cuda_device

* Also test copying between devices

* Refactor using StreamsAndDevices

* Don't use shared_ptr

Implement sort_by_key (#6801)

* Implement sort_by_key

* Address review comments

* Make passed in view const ref

* Fix _via_sort for OpenMPTarget

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

* Rip out ROCThrust for now

* Few changes to address the comments

* Fix SYCL

* Fix SYCL implementation

---------

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

Fix scorecard workflow (#6831)

* Add Scorecrad

* another one

* Don't run on pull requests

* Add comments back in

Remove redundant RangePolicy constructor (#6841)

* Remove redundant RangePolicy constructor

* Disable test case for Cuda <11.2 completely

Bump Google Benchmark version v1.{6.2 -> 7.1} in CMake FetchContent

Remove variadic range policy constructor (#6845)

* Removed variadic ctor from RangePolicy, as well as extra
set(...) methods.

* In RangePolicy:
Deprecated set(ChunkSize)
In the ctor that takes a ChunkSize, set the chunk size via
set_chunk_size instead of via member initializers (to be more consistent
with other code)

* Removed superfluous "inline" from two of the RangePolicy constructors

Avoid unused variable warning in TestRangePolicyCTAD.cpp

bytes_and_flops with CMake

Fix some warnings as well

bytes_and_flops fix a small bug in command line argument

add gather benchmark to CMake

add atomic benchmark to CMake

add policy_performance benchmark to CMake

fix some warnings in policy_performance benchmark

policy_benchmark: fix indentation

Benchmarks: disable 2 benchmarks for OpenMPTarget

Apply Rahul suggestion to disable two benchmarks that are causing Internal Compiler Errors with OpenMPTarget.

bytes_and_flops: fix a counter name

Fix C-style cast

Fix amdclang++ compilation (#6857)

* Fix amdclang++ compilation

* Add guards for hipcc

Fix fallback implementation for sort_by_key (#6856)

* Fix fallback implementation for sort_by_key

* Guard with KOKKOS_ENABLE_ONEDPL

* Drop sort_on_device

* Improve wording

* Improve comment

Update the unsafe implicit conversion error message in MDRangePolicy (#6855)

* Updated the error message in MDRangePolicy about unsafe implicit conversions

* Addressed clang warnings

Split some classes from Kokkos_ViewMapping (#6859)

* move ViewOffset and ViewDataAnalysis to a separate header

* only include Kokkos_Macro in ViewDataAnalysis

update mdspan tpl

CTAD (deduction guides) for RangePolicy (#6850)

* Removed the check for NVCC compiler version when testing CTAD,
as decltype(RangePolicy(...)) isn't compiling under any version
of Cuda.

* Added in the explicit deduction guides for RangePolicy:
• Correctness when passing in an execution space
• Workaround for nvcc as RangePolicy<...> doesn't have any
template parameters that can be deduced, so gcc/clang assume that a
matching ctor in the primary template deduces to RangePolicy<> while
nvcc assumes it is a bug.

Rewrote the tests to be of the form:

[[maybe_unused]] static inline auto rpxy = RangePolicy(x, y);
static_assert(is_same_v<RangePolicy<WhatShouldBeHere>, decltype(rpxy)>);

This form avoids the most vexing parse.o, and seems to pass on all
compilers.

Added tests for SomeExecutionSpace which is guaranteed not to be
DefaultExecutionSpace.

Added calls to ImplicitlyConvertibleToDefaultExecutionSpace::operator
DefaultExecutionSpace() and SomeExecutionSpace::concurrency() to
avoid maybe_unused errors under some compilers.

Notes:

The default constructed CTAD RangePolicy uses list initialization syntax
(curlies) instead of parentheses to get around a gcc 8.2 compiler bug.

nestodes uses "Kokkos::DefaultExecutionSpace" instead of "auto" in order
to trigger the implicit conversion.

CTAD (deduction guides) for MDRangePolicy (#5516)

* CTAD for MDRangePolicy

* WIP: Moved CTAD tests from TestMDRangePolicyConstructors.hpp to
TestMDRangePolicyCTAD.cpp

* WIP Added [[maybe_unused]] to all static inline variables in
CTAD tests

* Removed TEST_EXECSPACE from the MDRangePolicy CTAD tests
(as that isn't needed for compile time only tests) and
associated cleanup

* Marked
struct ImplicitlyConvertibleToDefaultExecutionSpace::operator Kokkos::DefaultExecutionSpace() const;
as [[maybe_unused]] to make it consistent with the
other policy CTAD tests (in other PRs)

* Added a Workaround for nvc++ (CUDA-11.7-NVHPC) ignoring
[[maybe_unused]] on

ImplicitlyConvertibleToDefaultExecutionSpace::operator
Kokkos::DefaultExecutionSpace() const

by defining it and implicitly calling it in another [[maybe_unused]] static inline
variable.

* Workaround for HIP-ROCm-5.2 "declared but never referenced"

* Added MDRangePolicy CTAD tests for initializer_lists
Fixed a comment

Add support for rocThrust in sort when using HIP (#6793)

* Use rocthrust in sort when using HIP

* Fix reviewer's comments

* Make sure that we don't compile Kokkos for every supported architecture when enabling rocthrust

* Export ROCTHRUST as a Kokkos dependency

Add HIP specialization for sort-by-key

Fix unused variable warning in TestMDRangePolicyCTAD.cpp

Improve message on view out of bounds access and always abort (#6861)

* Add test for view out-of-bounds access

* [DO NOT MERGE] enable bound checking in view accesses unconditionally

* Abort on out-of-bounds access on the host side instead of throwing

* Improve out-of-bounds error message

* Add test for printing multidimensional index

* Try to get offsetview to compile

* Do not bother with the label on the device side

* Update test to reflect that device does not print the label

* Fixup OffsetView out-of-bounds on device

* Revert "Update test to reflect that device does not print the label"

This reverts commit 7dbecbf635386556d3557c78a495970b439a3e05.

* Revert "Do not bother with the label on the device side"

This reverts commit 40348e88d265f7dfaf42354db5be18db21d55808.

* Fixup UNMANAGED -> UNAVAILABLE

* Only enable the test when debug bound checking is enabled

* Revert "[DO NOT MERGE] enable bound checking in view accesses unconditionally"

This reverts commit ff9d411fa0e8898f3db60055a12f222e9836fc57.

* Fix typo formated -> formatted

* Drop unecesary cast to void and Kokkos::Impl:: qualification

* Add test with mixed integer types

* Improve function name check_bounds -> within_range

* Prefer right fold per review

* Per review drop immediately invoked lambda trick on the host side

* Silent warnings about tracker variable not being used on the device

Fix linking with rothrust in downstream applications

SYCL: Make sure to call find_dependency for oneDPL if necessary  (#6870)

* Make sure to call find_dependency for oneDPL if necessary

* Move linking with oneDPL to algorithms

* Don't guard KOKKOS_IMPORT_TPL by CMake language

Don't use Fedora development version in GitHub CI

kokkos_tpls.cmake: update default option to enable rocthrust

Avoid configuration issues with Trilinos of the form:
Make Error at kokkos/cmake/kokkos_tpls.cmake:29 (MESSAGE):
 Enabled TPL ROCTHRUST inside TriBITS build, but this can only be enabled in
 a standalone build

Scratch space fix for MultiGPU (#6866)

* Use CudaSpace::(de)allocate() instead of Kokkos_SharedAllocate.hpp for Cuda scratch memory

Fixes bug where all scratch mem was allocated on default device

* Add scratch test to MultiGPU testing

* fix typo in multi-gpu test

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

* Use reductions for error accumulator

* Remove exec_ arg from ScratchFunctor (not needed anymore)

* avoid implicit capture of this in test

---------

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

[ci skip] Bump version number to 4.3.99

Fix Makefile.kokkos for Threads

Fix a bug in Makefile when using AMD GPU architectures (#6892)

* Fix bug in Makefile when using AMD GPU architectures

* Fix indentation

* Fix documentation of the architecture to match the code

Cuda: Fix configuring with CMake 3.28.4 (#6898)

* Cuda: Fix configuring with CMake 3.29.0

* CMake 3.28.4 is also affected

Update Intel GPU architectures in Makefile (#6895)

* Update Intel GPU architectures in Makefile

* Add some comments

Move `Kokkos::Array` tests to a more suitable place (#6905)

* Move Kokkos::Array tests to a more suitable place

* Workaround bogous(?) compile error with Array::operator[] not being constexpr

Fixed and simplified the unit test

SYCL: Cleanup device selection

Adding SECURITY.md file

This is based on the one from OSSF Scorecard

Cleanup KOKKOS_CONFIGURE_CORE

Avoid detecting unwanted fences in the parallel_scan_no_fence test

Add test for current CTAD support with RangePolicy (#6803)

* Add test for current CTAD support with RangePolicy

Co-authored-by: Nevin Liber <nliber@anl.gov>

* Rework CTAD test to avoid "memeber <bla> was declared but never referenced" warnings with icpc 19

* Attempt to fix CI

* Attempt to eliminate maybe unused warning in icpc

* Disable CTAD tests for nvcc < 11.2
as compiler bugs prevent CTAD expressions inside decltype

---------

Co-authored-by: Nevin Liber <nliber@anl.gov>
Co-authored-by: Nevin ":-)" Liber <nliber+github@gmail.com>

multi-GPU support: Add test for all policies (#6782)

* Cuda multi-GPU support: Test with managed and unmanaged Views

* Move check for cuda_device

* Also test copying between devices

* Refactor using StreamsAndDevices

* Don't use shared_ptr

Implement sort_by_key (#6801)

* Implement sort_by_key

* Address review comments

* Make passed in view const ref

* Fix _via_sort for OpenMPTarget

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

* Rip out ROCThrust for now

* Few changes to address the comments

* Fix SYCL

* Fix SYCL implementation

---------

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

Fix scorecard workflow (#6831)

* Add Scorecrad

* another one

* Don't run on pull requests

* Add comments back in

Remove redundant RangePolicy constructor (#6841)

* Remove redundant RangePolicy constructor

* Disable test case for Cuda <11.2 completely

Bump Google Benchmark version v1.{6.2 -> 7.1} in CMake FetchContent

Remove variadic range policy constructor (#6845)

* Removed variadic ctor from RangePolicy, as well as extra
set(...) methods.

* In RangePolicy:
Deprecated set(ChunkSize)
In the ctor that takes a ChunkSize, set the chunk size via
set_chunk_size instead of via member initializers (to be more consistent
with other code)

* Removed superfluous "inline" from two of the RangePolicy constructors

Avoid unused variable warning in TestRangePolicyCTAD.cpp

bytes_and_flops with CMake

Fix some warnings as well

bytes_and_flops fix a small bug in command line argument

add gather benchmark to CMake

add atomic benchmark to CMake

add policy_performance benchmark to CMake

fix some warnings in policy_performance benchmark

policy_benchmark: fix indentation

Benchmarks: disable 2 benchmarks for OpenMPTarget

Apply Rahul suggestion to disable two benchmarks that are causing Internal Compiler Errors with OpenMPTarget.

bytes_and_flops: fix a counter name

Fix C-style cast

Fix amdclang++ compilation (#6857)

* Fix amdclang++ compilation

* Add guards for hipcc

Fix fallback implementation for sort_by_key (#6856)

* Fix fallback implementation for sort_by_key

* Guard with KOKKOS_ENABLE_ONEDPL

* Drop sort_on_device

* Improve wording

* Improve comment

Update the unsafe implicit conversion error message in MDRangePolicy (#6855)

* Updated the error message in MDRangePolicy about unsafe implicit conversions

* Addressed clang warnings

Split some classes from Kokkos_ViewMapping (#6859)

* move ViewOffset and ViewDataAnalysis to a separate header

* only include Kokkos_Macro in ViewDataAnalysis

update mdspan tpl

CTAD (deduction guides) for RangePolicy (#6850)

* Removed the check for NVCC compiler version when testing CTAD,
as decltype(RangePolicy(...)) isn't compiling under any version
of Cuda.

* Added in the explicit deduction guides for RangePolicy:
• Correctness when passing in an execution space
• Workaround for nvcc as RangePolicy<...> doesn't have any
template parameters that can be deduced, so gcc/clang assume that a
matching ctor in the primary template deduces to RangePolicy<> while
nvcc assumes it is a bug.

Rewrote the tests to be of the form:

[[maybe_unused]] static inline auto rpxy = RangePolicy(x, y);
static_assert(is_same_v<RangePolicy<WhatShouldBeHere>, decltype(rpxy)>);

This form avoids the most vexing parse.o, and seems to pass on all
compilers.

Added tests for SomeExecutionSpace which is guaranteed not to be
DefaultExecutionSpace.

Added calls to ImplicitlyConvertibleToDefaultExecutionSpace::operator
DefaultExecutionSpace() and SomeExecutionSpace::concurrency() to
avoid maybe_unused errors under some compilers.

Notes:

The default constructed CTAD RangePolicy uses list initialization syntax
(curlies) instead of parentheses to get around a gcc 8.2 compiler bug.

nestodes uses "Kokkos::DefaultExecutionSpace" instead of "auto" in order
to trigger the implicit conversion.

CTAD (deduction guides) for MDRangePolicy (#5516)

* CTAD for MDRangePolicy

* WIP: Moved CTAD tests from TestMDRangePolicyConstructors.hpp to
TestMDRangePolicyCTAD.cpp

* WIP Added [[maybe_unused]] to all static inline variables in
CTAD tests

* Removed TEST_EXECSPACE from the MDRangePolicy CTAD tests
(as that isn't needed for compile time only tests) and
associated cleanup

* Marked
struct ImplicitlyConvertibleToDefaultExecutionSpace::operator Kokkos::DefaultExecutionSpace() const;
as [[maybe_unused]] to make it consistent with the
other policy CTAD tests (in other PRs)

* Added a Workaround for nvc++ (CUDA-11.7-NVHPC) ignoring
[[maybe_unused]] on

ImplicitlyConvertibleToDefaultExecutionSpace::operator
Kokkos::DefaultExecutionSpace() const

by defining it and implicitly calling it in another [[maybe_unused]] static inline
variable.

* Workaround for HIP-ROCm-5.2 "declared but never referenced"

* Added MDRangePolicy CTAD tests for initializer_lists
Fixed a comment

Add support for rocThrust in sort when using HIP (#6793)

* Use rocthrust in sort when using HIP

* Fix reviewer's comments

* Make sure that we don't compile Kokkos for every supported architecture when enabling rocthrust

* Export ROCTHRUST as a Kokkos dependency

Add HIP specialization for sort-by-key

Fix unused variable warning in TestMDRangePolicyCTAD.cpp

Improve message on view out of bounds access and always abort (#6861)

* Add test for view out-of-bounds access

* [DO NOT MERGE] enable bound checking in view accesses unconditionally

* Abort on out-of-bounds access on the host side instead of throwing

* Improve out-of-bounds error message

* Add test for printing multidimensional index

* Try to get offsetview to compile

* Do not bother with the label on the device side

* Update test to reflect that device does not print the label

* Fixup OffsetView out-of-bounds on device

* Revert "Update test to reflect that device does not print the label"

This reverts commit 7dbecbf635386556d3557c78a495970b439a3e05.

* Revert "Do not bother with the label on the device side"

This reverts commit 40348e88d265f7dfaf42354db5be18db21d55808.

* Fixup UNMANAGED -> UNAVAILABLE

* Only enable the test when debug bound checking is enabled

* Revert "[DO NOT MERGE] enable bound checking in view accesses unconditionally"

This reverts commit ff9d411fa0e8898f3db60055a12f222e9836fc57.

* Fix typo formated -> formatted

* Drop unecesary cast to void and Kokkos::Impl:: qualification

* Add test with mixed integer types

* Improve function name check_bounds -> within_range

* Prefer right fold per review

* Per review drop immediately invoked lambda trick on the host side

* Silent warnings about tracker variable not being used on the device

Fix linking with rothrust in downstream applications

SYCL: Make sure to call find_dependency for oneDPL if necessary  (#6870)

* Make sure to call find_dependency for oneDPL if necessary

* Move linking with oneDPL to algorithms

* Don't guard KOKKOS_IMPORT_TPL by CMake language

Don't use Fedora development version in GitHub CI

kokkos_tpls.cmake: update default option to enable rocthrust

Avoid configuration issues with Trilinos of the form:
Make Error at kokkos/cmake/kokkos_tpls.cmake:29 (MESSAGE):
 Enabled TPL ROCTHRUST inside TriBITS build, but this can only be enabled in
 a standalone build

Scratch space fix for MultiGPU (#6866)

* Use CudaSpace::(de)allocate() instead of Kokkos_SharedAllocate.hpp for Cuda scratch memory

Fixes bug where all scratch mem was allocated on default device

* Add scratch test to MultiGPU testing

* fix typo in multi-gpu test

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

* Use reductions for error accumulator

* Remove exec_ arg from ScratchFunctor (not needed anymore)

* avoid implicit capture of this in test

---------

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

[ci skip] Bump version number to 4.3.99

Fix Makefile.kokkos for Threads

Fix a bug in Makefile when using AMD GPU architectures (#6892)

* Fix bug in Makefile when using AMD GPU architectures

* Fix indentation

* Fix documentation of the architecture to match the code

Cuda: Fix configuring with CMake 3.28.4 (#6898)

* Cuda: Fix configuring with CMake 3.29.0

* CMake 3.28.4 is also affected

Update Intel GPU architectures in Makefile (#6895)

* Update Intel GPU architectures in Makefile

* Add some comments

Move `Kokkos::Array` tests to a more suitable place (#6905)

* Move Kokkos::Array tests to a more suitable place

* Workaround bogous(?) compile error with Array::operator[] not being constexpr

Fixed and simplified the unit test

SYCL: Cleanup device selection

Adding SECURITY.md file

This is based on the one from OSSF Scorecard

Cleanup KOKKOS_CONFIGURE_CORE

Avoid detecting unwanted fences in the parallel_scan_no_fence test

Add test for current CTAD support with RangePolicy (#6803)

* Add test for current CTAD support with RangePolicy

Co-authored-by: Nevin Liber <nliber@anl.gov>

* Rework CTAD test to avoid "memeber <bla> was declared but never referenced" warnings with icpc 19

* Attempt to fix CI

* Attempt to eliminate maybe unused warning in icpc

* Disable CTAD tests for nvcc < 11.2
as compiler bugs prevent CTAD expressions inside decltype

---------

Co-authored-by: Nevin Liber <nliber@anl.gov>
Co-authored-by: Nevin ":-)" Liber <nliber+github@gmail.com>

multi-GPU support: Add test for all policies (#6782)

* Cuda multi-GPU support: Test with managed and unmanaged Views

* Move check for cuda_device

* Also test copying between devices

* Refactor using StreamsAndDevices

* Don't use shared_ptr

Implement sort_by_key (#6801)

* Implement sort_by_key

* Address review comments

* Make passed in view const ref

* Fix _via_sort for OpenMPTarget

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

* Rip out ROCThrust for now

* Few changes to address the comments

* Fix SYCL

* Fix SYCL implementation

---------

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

Fix scorecard workflow (#6831)

* Add Scorecrad

* another one

* Don't run on pull requests

* Add comments back in

Remove redundant RangePolicy constructor (#6841)

* Remove redundant RangePolicy constructor

* Disable test case for Cuda <11.2 completely

Bump Google Benchmark version v1.{6.2 -> 7.1} in CMake FetchContent

Remove variadic range policy constructor (#6845)

* Removed variadic ctor from RangePolicy, as well as extra
set(...) methods.

* In RangePolicy:
Deprecated set(ChunkSize)
In the ctor that takes a ChunkSize, set the chunk size via
set_chunk_size instead of via member initializers (to be more consistent
with other code)

* Removed superfluous "inline" from two of the RangePolicy constructors

Avoid unused variable warning in TestRangePolicyCTAD.cpp

bytes_and_flops with CMake

Fix some warnings as well

bytes_and_flops fix a small bug in command line argument

add gather benchmark to CMake

add atomic benchmark to CMake

add policy_performance benchmark to CMake

fix some warnings in policy_performance benchmark

policy_benchmark: fix indentation

Benchmarks: disable 2 benchmarks for OpenMPTarget

Apply Rahul suggestion to disable two benchmarks that are causing Internal Compiler Errors with OpenMPTarget.

bytes_and_flops: fix a counter name

Fix C-style cast

Fix amdclang++ compilation (#6857)

* Fix amdclang++ compilation

* Add guards for hipcc

Fix fallback implementation for sort_by_key (#6856)

* Fix fallback implementation for sort_by_key

* Guard with KOKKOS_ENABLE_ONEDPL

* Drop sort_on_device

* Improve wording

* Improve comment

Update the unsafe implicit conversion error message in MDRangePolicy (#6855)

* Updated the error message in MDRangePolicy about unsafe implicit conversions

* Addressed clang warnings

Split some classes from Kokkos_ViewMapping (#6859)

* move ViewOffset and ViewDataAnalysis to a separate header

* only include Kokkos_Macro in ViewDataAnalysis

update mdspan tpl

CTAD (deduction guides) for RangePolicy (#6850)

* Removed the check for NVCC compiler version when testing CTAD,
as decltype(RangePolicy(...)) isn't compiling under any version
of Cuda.

* Added in the explicit deduction guides for RangePolicy:
• Correctness when passing in an execution space
• Workaround for nvcc as RangePolicy<...> doesn't have any
template parameters that can be deduced, so gcc/clang assume that a
matching ctor in the primary template deduces to RangePolicy<> while
nvcc assumes it is a bug.

Rewrote the tests to be of the form:

[[maybe_unused]] static inline auto rpxy = RangePolicy(x, y);
static_assert(is_same_v<RangePolicy<WhatShouldBeHere>, decltype(rpxy)>);

This form avoids the most vexing parse.o, and seems to pass on all
compilers.

Added tests for SomeExecutionSpace which is guaranteed not to be
DefaultExecutionSpace.

Added calls to ImplicitlyConvertibleToDefaultExecutionSpace::operator
DefaultExecutionSpace() and SomeExecutionSpace::concurrency() to
avoid maybe_unused errors under some compilers.

Notes:

The default constructed CTAD RangePolicy uses list initialization syntax
(curlies) instead of parentheses to get around a gcc 8.2 compiler bug.

nestodes uses "Kokkos::DefaultExecutionSpace" instead of "auto" in order
to trigger the implicit conversion.

CTAD (deduction guides) for MDRangePolicy (#5516)

* CTAD for MDRangePolicy

* WIP: Moved CTAD tests from TestMDRangePolicyConstructors.hpp to
TestMDRangePolicyCTAD.cpp

* WIP Added [[maybe_unused]] to all static inline variables in
CTAD tests

* Removed TEST_EXECSPACE from the MDRangePolicy CTAD tests
(as that isn't needed for compile time only tests) and
associated cleanup

* Marked
struct ImplicitlyConvertibleToDefaultExecutionSpace::operator Kokkos::DefaultExecutionSpace() const;
as [[maybe_unused]] to make it consistent with the
other policy CTAD tests (in other PRs)

* Added a Workaround for nvc++ (CUDA-11.7-NVHPC) ignoring
[[maybe_unused]] on

ImplicitlyConvertibleToDefaultExecutionSpace::operator
Kokkos::DefaultExecutionSpace() const

by defining it and implicitly calling it in another [[maybe_unused]] static inline
variable.

* Workaround for HIP-ROCm-5.2 "declared but never referenced"

* Added MDRangePolicy CTAD tests for initializer_lists
Fixed a comment

Add support for rocThrust in sort when using HIP (#6793)

* Use rocthrust in sort when using HIP

* Fix reviewer's comments

* Make sure that we don't compile Kokkos for every supported architecture when enabling rocthrust

* Export ROCTHRUST as a Kokkos dependency

Add HIP specialization for sort-by-key

Fix unused variable warning in TestMDRangePolicyCTAD.cpp

Improve message on view out of bounds access and always abort (#6861)

* Add test for view out-of-bounds access

* [DO NOT MERGE] enable bound checking in view accesses unconditionally

* Abort on out-of-bounds access on the host side instead of throwing

* Improve out-of-bounds error message

* Add test for printing multidimensional index

* Try to get offsetview to compile

* Do not bother with the label on the device side

* Update test to reflect that device does not print the label

* Fixup OffsetView out-of-bounds on device

* Revert "Update test to reflect that device does not print the label"

This reverts commit 7dbecbf635386556d3557c78a495970b439a3e05.

* Revert "Do not bother with the label on the device side"

This reverts commit 40348e88d265f7dfaf42354db5be18db21d55808.

* Fixup UNMANAGED -> UNAVAILABLE

* Only enable the test when debug bound checking is enabled

* Revert "[DO NOT MERGE] enable bound checking in view accesses unconditionally"

This reverts commit ff9d411fa0e8898f3db60055a12f222e9836fc57.

* Fix typo formated -> formatted

* Drop unecesary cast to void and Kokkos::Impl:: qualification

* Add test with mixed integer types

* Improve function name check_bounds -> within_range

* Prefer right fold per review

* Per review drop immediately invoked lambda trick on the host side

* Silent warnings about tracker variable not being used on the device

Fix linking with rothrust in downstream applications

SYCL: Make sure to call find_dependency for oneDPL if necessary  (#6870)

* Make sure to call find_dependency for oneDPL if necessary

* Move linking with oneDPL to algorithms

* Don't guard KOKKOS_IMPORT_TPL by CMake language

Don't use Fedora development version in GitHub CI

kokkos_tpls.cmake: update default option to enable rocthrust

Avoid configuration issues with Trilinos of the form:
Make Error at kokkos/cmake/kokkos_tpls.cmake:29 (MESSAGE):
 Enabled TPL ROCTHRUST inside TriBITS build, but this can only be enabled in
 a standalone build

Scratch space fix for MultiGPU (#6866)

* Use CudaSpace::(de)allocate() instead of Kokkos_SharedAllocate.hpp for Cuda scratch memory

Fixes bug where all scratch mem was allocated on default device

* Add scratch test to MultiGPU testing

* fix typo in multi-gpu test

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

* Use reductions for error accumulator

* Remove exec_ arg from ScratchFunctor (not needed anymore)

* avoid implicit capture of this in test

---------

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

[ci skip] Bump version number to 4.3.99

Fix Makefile.kokkos for Threads

Fix a bug in Makefile when using AMD GPU architectures (#6892)

* Fix bug in Makefile when using AMD GPU architectures

* Fix indentation

* Fix documentation of the architecture to match the code

Cuda: Fix configuring with CMake 3.28.4 (#6898)

* Cuda: Fix configuring with CMake 3.29.0

* CMake 3.28.4 is also affected

Update Intel GPU architectures in Makefile (#6895)

* Update Intel GPU architectures in Makefile

* Add some comments

Move `Kokkos::Array` tests to a more suitable place (#6905)

* Move Kokkos::Array tests to a more suitable place

* Workaround bogous(?) compile error with Array::operator[] not being constexpr

Fixed and simplified the unit test

SYCL: Cleanup device selection

Adding SECURITY.md file

This is based on the one from OSSF Scorecard

Cleanup KOKKOS_CONFIGURE_CORE

Avoid detecting unwanted fences in the parallel_scan_no_fence test

Add test for current CTAD support with RangePolicy (#6803)

* Add test for current CTAD support with RangePolicy

Co-authored-by: Nevin Liber <nliber@anl.gov>

* Rework CTAD test to avoid "memeber <bla> was declared but never referenced" warnings with icpc 19

* Attempt to fix CI

* Attempt to eliminate maybe unused warning in icpc

* Disable CTAD tests for nvcc < 11.2
as compiler bugs prevent CTAD expressions inside decltype

---------

Co-authored-by: Nevin Liber <nliber@anl.gov>
Co-authored-by: Nevin ":-)" Liber <nliber+github@gmail.com>

multi-GPU support: Add test for all policies (#6782)

* Cuda multi-GPU support: Test with managed and unmanaged Views

* Move check for cuda_device

* Also test copying between devices

* Refactor using StreamsAndDevices

* Don't use shared_ptr

Implement sort_by_key (#6801)

* Implement sort_by_key

* Address review comments

* Make passed in view const ref

* Fix _via_sort for OpenMPTarget

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

* Rip out ROCThrust for now

* Few changes to address the comments

* Fix SYCL

* Fix SYCL implementation

---------

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

Fix scorecard workflow (#6831)

* Add Scorecrad

* another one

* Don't run on pull requests

* Add comments back in

Remove redundant RangePolicy constructor (#6841)

* Remove redundant RangePolicy constructor

* Disable test case for Cuda <11.2 completely

Bump Google Benchmark version v1.{6.2 -> 7.1} in CMake FetchContent

Remove variadic range policy constructor (#6845)

* Removed variadic ctor from RangePolicy, as well as extra
set(...) methods.

* In RangePolicy:
Deprecated set(ChunkSize)
In the ctor that takes a ChunkSize, set the chunk size via
set_chunk_size instead of via member initializers (to be more consistent
with other code)

* Removed superfluous "inline" from two of the RangePolicy constructors

Avoid unused variable warning in TestRangePolicyCTAD.cpp

bytes_and_flops with CMake

Fix some warnings as well

bytes_and_flops fix a small bug in command line argument

add gather benchmark to CMake

add atomic benchmark to CMake

add policy_performance benchmark to CMake

fix some warnings in policy_performance benchmark

policy_benchmark: fix indentation

Benchmarks: disable 2 benchmarks for OpenMPTarget

Apply Rahul suggestion to disable two benchmarks that are causing Internal Compiler Errors with OpenMPTarget.

bytes_and_flops: fix a counter name

Fix C-style cast

Fix amdclang++ compilation (#6857)

* Fix amdclang++ compilation

* Add guards for hipcc

Fix fallback implementation for sort_by_key (#6856)

* Fix fallback implementation for sort_by_key

* Guard with KOKKOS_ENABLE_ONEDPL

* Drop sort_on_device

* Improve wording

* Improve comment

Update the unsafe implicit conversion error message in MDRangePolicy (#6855)

* Updated the error message in MDRangePolicy about unsafe implicit conversions

* Addressed clang warnings

Split some classes from Kokkos_ViewMapping (#6859)

* move ViewOffset and ViewDataAnalysis to a separate header

* only include Kokkos_Macro in ViewDataAnalysis

update mdspan tpl

CTAD (deduction guides) for RangePolicy (#6850)

* Removed the check for NVCC compiler version when testing CTAD,
as decltype(RangePolicy(...)) isn't compiling under any version
of Cuda.

* Added in the explicit deduction guides for RangePolicy:
• Correctness when passing in an execution space
• Workaround for nvcc as RangePolicy<...> doesn't have any
template parameters that can be deduced, so gcc/clang assume that a
matching ctor in the primary template deduces to RangePolicy<> while
nvcc assumes it is a bug.

Rewrote the tests to be of the form:

[[maybe_unused]] static inline auto rpxy = RangePolicy(x, y);
static_assert(is_same_v<RangePolicy<WhatShouldBeHere>, decltype(rpxy)>);

This form avoids the most vexing parse.o, and seems to pass on all
compilers.

Added tests for SomeExecutionSpace which is guaranteed not to be
DefaultExecutionSpace.

Added calls to ImplicitlyConvertibleToDefaultExecutionSpace::operator
DefaultExecutionSpace() and SomeExecutionSpace::concurrency() to
avoid maybe_unused errors under some compilers.

Notes:

The default constructed CTAD RangePolicy uses list initialization syntax
(curlies) instead of parentheses to get around a gcc 8.2 compiler bug.

nestodes uses "Kokkos::DefaultExecutionSpace" instead of "auto" in order
to trigger the implicit conversion.

CTAD (deduction guides) for MDRangePolicy (#5516)

* CTAD for MDRangePolicy

* WIP: Moved CTAD tests from TestMDRangePolicyConstructors.hpp to
TestMDRangePolicyCTAD.cpp

* WIP Added [[maybe_unused]] to all static inline variables in
CTAD tests

* Removed TEST_EXECSPACE from the MDRangePolicy CTAD tests
(as that isn't needed for compile time only tests) and
associated cleanup

* Marked
struct ImplicitlyConvertibleToDefaultExecutionSpace::operator Kokkos::DefaultExecutionSpace() const;
as [[maybe_unused]] to make it consistent with the
other policy CTAD tests (in other PRs)

* Added a Workaround for nvc++ (CUDA-11.7-NVHPC) ignoring
[[maybe_unused]] on

ImplicitlyConvertibleToDefaultExecutionSpace::operator
Kokkos::DefaultExecutionSpace() const

by defining it and implicitly calling it in another [[maybe_unused]] static inline
variable.

* Workaround for HIP-ROCm-5.2 "declared but never referenced"

* Added MDRangePolicy CTAD tests for initializer_lists
Fixed a comment

Add support for rocThrust in sort when using HIP (#6793)

* Use rocthrust in sort when using HIP

* Fix reviewer's comments

* Make sure that we don't compile Kokkos for every supported architecture when enabling rocthrust

* Export ROCTHRUST as a Kokkos dependency

Add HIP specialization for sort-by-key

Fix unused variable warning in TestMDRangePolicyCTAD.cpp

Improve message on view out of bounds access and always abort (#6861)

* Add test for view out-of-bounds access

* [DO NOT MERGE] enable bound checking in view accesses unconditionally

* Abort on out-of-bounds access on the host side instead of throwing

* Improve out-of-bounds error message

* Add test for printing multidimensional index

* Try to get offsetview to compile

* Do not bother with the label on the device side

* Update test to reflect that device does not print the label

* Fixup OffsetView out-of-bounds on device

* Revert "Update test to reflect that device does not print the label"

This reverts commit 7dbecbf635386556d3557c78a495970b439a3e05.

* Revert "Do not bother with the label on the device side"

This reverts commit 40348e88d265f7dfaf42354db5be18db21d55808.

* Fixup UNMANAGED -> UNAVAILABLE

* Only enable the test when debug bound checking is enabled

* Revert "[DO NOT MERGE] enable bound checking in view accesses unconditionally"

This reverts commit ff9d411fa0e8898f3db60055a12f222e9836fc57.

* Fix typo formated -> formatted

* Drop unecesary cast to void and Kokkos::Impl:: qualification

* Add test with mixed integer types

* Improve function name check_bounds -> within_range

* Prefer right fold per review

* Per review drop immediately invoked lambda trick on the host side

* Silent warnings about tracker variable not being used on the device

Fix linking with rothrust in downstream applications

SYCL: Make sure to call find_dependency for oneDPL if necessary  (#6870)

* Make sure to call find_dependency for oneDPL if necessary

* Move linking with oneDPL to algorithms

* Don't guard KOKKOS_IMPORT_TPL by CMake language

Don't use Fedora development version in GitHub CI

kokkos_tpls.cmake: update default option to enable rocthrust

Avoid configuration issues with Trilinos of the form:
Make Error at kokkos/cmake/kokkos_tpls.cmake:29 (MESSAGE):
 Enabled TPL ROCTHRUST inside TriBITS build, but this can only be enabled in
 a standalone build

Scratch space fix for MultiGPU (#6866)

* Use CudaSpace::(de)allocate() instead of Kokkos_SharedAllocate.hpp for Cuda scratch memory

Fixes bug where all scratch mem was allocated on default device

* Add scratch test to MultiGPU testing

* fix typo in multi-gpu test

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

* Use reductions for error accumulator

* Remove exec_ arg from ScratchFunctor (not needed anymore)

* avoid implicit capture of this in test

---------

Co-authored-by: Dong Hun Lee <59181952+ldh4@users.noreply.github.com>

[ci skip] Bump version number to 4.3.99

Fix Makefile.kokkos for Threads

Fix a bug in Makefile when using AMD GPU architectures (#6892)

* Fix bug in Makefile when using AMD GPU architectures

* Fix indentation

* Fix documentation of the architecture to match the code

Cuda: Fix configuring with CMake 3.28.4 (#6898)

* Cuda: Fix configuring with CMake 3.29.0

* CMake 3.28.4 is also affected

Update Intel GPU architectures in Makefile (#6895)

* Update Intel GPU architectures in Makefile

* Add some comments

Move `Kokkos::Array` tests to a more suitable place (#6905)

* Move Kokkos::Array tests to a more suitable place

* Workaround bogous(?) compile error with Array::operator[] not being constexpr

Fixed and simplified the unit test

SYCL: Cleanup device selection

Adding SECURITY.md file

This is based on the one from OSSF Scorecard

Cleanup KOKKOS_CONFIGURE_CORE

Avoid detecting unwanted fences in the parallel_scan_no_fence test

Add test for current CTAD support with RangePolicy (#6803)

* Add test for current CTAD support with RangePolicy

Co-authored-by: Nevin Liber <nliber@anl.gov>

* Rework CTAD test to avoid "memeber <bla> was declared but never referenced" warnings with icpc 19

* Attempt to fix CI

* Attempt to eliminate maybe unused warning in icpc

* Disable CTAD tests for nvcc < 11.2
as compiler bugs prevent CTAD expressions inside decltype

---------

Co-authored-by: Nevin Liber <nliber@anl.gov>
Co-authored-by: Nevin ":-)" Liber <nliber+github@gmail.com>

multi-GPU support: Add test for all policies (#6782)

* Cuda multi-GPU support: Test with managed and unmanaged Views

* Move check for cuda_device

* Also test copying between devices

* Refactor using StreamsAndDevices

* Don't use shared_ptr

Implement sort_by_key (#6801)

* Implement sort_by_key

* Address review comments

* Make passed in view const ref

* Fix _via_sort for OpenMPTarget

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

* Rip out ROCThrust for now

* Few changes to address the comments

* Fix SYCL

* Fix SYCL implementation

---------

Co-authored-by: Daniel Arndt <arndtd@ornl.gov>

Fix scorecard workflow (#6831)

* Add Scorecrad

* another one

* Don't run on pull requests

* Add comments back in

Remove redundant RangePolicy constructor (#6841)

* Remove redundant RangePolicy constructor

* Disable test case for Cuda <11.2 completely

Bump Google Benchmark version v1.{6.2 -> 7.1} in CMake FetchContent

Remove variadic range policy constructor (#6845)

* Removed variadic ctor from RangePolicy, as well as extra
set(...) methods.

* In RangePolicy:
Deprecated set(ChunkSize)
In the ctor that takes a ChunkSize, set the chunk size via
set_chunk_size instead of via member initializers (to be more consistent
with other code)

* Removed superfluous "inline" from two of the RangePolicy constructors

Avoid unused variable warning in TestRangePolicyCTAD.cpp