CHANGELOG.md

CHANGELOG

4.4.01

Full Changelog

Features:

• Introduce new SequentialHostInit view allocation property #7229

Backend and Architecture Enhancements:

CUDA:

• Experimental support for unified memory mode (intended for Grace-Hopper etc.) #6823

Bug Fixes

• OpenMP: Fix issue related to the visibility of an internal symbol with shared libraries that affected ScatterView in particular #7284
• Fix implicit copy assignment operators in few AVX2 masks being deleted #7296

4.4.00

Full Changelog

Features:

• Add Kokkos::View conversions from and to std::mdspan #6830 #7069

Backend and Architecture Enhancements:

CUDA:

• nvcc_wrapper: Adding ability to process --disable-warnings flag #6936
• Use recommended/max team size functions in Cuda ParallelFor and Reduce constructors #6891
• Improve compile-times when building with Kokkos_ENABLE_DEBUG_BOUNDS_CHECK in Cuda #7013

HIP:

• Use HIP builtin atomics #6882 #7000
• Enable user-specified compiler and linker flags for AMD GPUs #7127

SYCL:

• Add support for Graphs #6912
• Fix multi-GPU support #6887
• Improve performance of reduction and scan operations #6562, #6750
• Fix lock for guarding scratch space in TeamPolicy parallel_reduce #6988
• Include submission command queue property information into SYCL::print_configuration() #7004

OpenACC:

• Make TeamPolicy parallel_for execute on the correct async queue #7012

OpenMPTarget:

• Honor user requested loop ordering in MDRange policy #6925
• Prevent data races by guarding the scratch space used in parallel_scan #6998

HPX:

• Workaround issue with template argument deduction to support compilation with NVCC #7015

General Enhancements

• Improve performance of view copies in host parallel regions #6730
• Harmonize convertibility rules of Kokkos::RandomAccessIterator with Views #6929
• Add a check precondition non-overlapping ranges for the adjacent_difference algorithm in debug mode #6922
• Add deduction guides for TeamPolicy #7030
• SIMD: Allow flexible vector width for 32 bit types #6802
• Updates for Kokkos::Array: add kokkos_swap(Array<T, N>) specialization #6943, add Kokkos::to_array #6375, make Kokkos::Array equality-comparable #7148
• Structured binding support for Kokkos::complex #7040

Build System Changes

• Do not require OpenMP support for languages other than CXX #6965
• Update Intel GPU architectures in Makefile #6895
• Fix use of OpenMP with Cuda or HIP as compile language #6972
• Define and enforce new minimum compiler versions for C++20 support #7128, #7123
• Add nvidia Grace CPU architecture: Kokkos_ARCH_ARMV9_GRACE #7158
• Fix Makefile.kokkos for Threads #6896
• Remove support for NVHPC as CUDA device compiler #6987
• Fix using CUDAToolkit for CMake 3.28.4 and higher #7062

Incompatibilities (i.e. breaking changes)

• Drop Kokkos::Array special treatment in Views #6906
• Drop Experimental::RawMemoryAllocationFailure #7145

Deprecations

• Remove Experimental::LayoutTiled class template and deprecate is_layouttiled trait #6907
• Deprecate Kokkos::layout_iterate_type_selector #7076
• Deprecate specialization of Kokkos::pair for a single element #6947
• Deprecate deep_copy of UnorderedMap of different size #6812
• Deprecate trailing Proxy template argument of Kokkos::Array #6934
• Deprecate implicit conversions of integers to ChunkSize #7151
• Deprecate implicit conversions to execution spaces #7156

Bug Fixes

• Do not return a copy of the input functor in Experimental::for_each #6910
• Fix realloc on views of non-default constructible element types #6993
• Fix undefined behavior in View initialization or fill with zeros #7014
• Fix sort_by_key on host execution spaces when building with NVCC #7059
• Fix using shared libraries and -fvisibility=hidden #7065
• Fix view reference counting when functor copy constructor throws in parallel dispatch #6289
• Fix initialize(InitializationSetting) for handling print_configuration setting #7098
• Thread safety fixes for the Serial and OpenMP backend #7080, #6151

4.3.01

Full Changelog

Backend and Architecture Enhancements:

HIP:

• MI300 support unified memory #6877

Bug Fixes

• Serial: Use the provided execution space instance in TeamPolicy #6951
• nvcc_wrapper: bring back support for --fmad option #6931
• Fix CUDA reduction overflow for RangePolicy #6578

4.3.00 (2024-03-19)

Full Changelog

Features:

• Add Experimental::sort_by_key(exec, keys, values) algorithm #6801

Backend and Architecture Enhancements:

CUDA:

• Experimental multi-GPU support (from the same process) #6782
• Link against CUDA libraries even with KOKKOS_ENABLE_COMPILE_AS_CMAKE_LANGUAGE #6701
• Don't use the compiler launcher script if the CMake compile language is CUDA. #6704
• nvcc(wrapper): adding "long" and "short" versions for all flags #6615

HIP:

• Fix compilation when using amdclang (with ROCm >= 5.7) and RDC #6857
• Use rocthrust for sorting, when available #6793

SYCL:

• We only support OneAPI SYCL implementation: add check during initialization
  • Error out on initialization if the backend is different from ext_oneapi_* #6784
  • Filter GPU devices for ext_onapi_* GPU devices #6758
• Performance Improvements
  • Avoid unnecessary zero-memset of the scratch flags in SYCL #6739
  • Use host-pinned memory to copy reduction/scan result #6500
• Address deprecations after oneAPI 2023.2.0 #6577
• Make sure to call find_dependency for oneDPL if necessary #6870

OpenMPTarget:

• Use LLVM extensions for dynamic shared memory #6380
• Guard scratch memory usage in ParallelReduce #6585
• Update linker flags for Intel GPUs update #6735
• Improve handling of printf on Intel GPUs #6652

OpenACC:

• Add atomics support #6446
• Make the OpenACC backend asynchronous #6772

Threads:

• Add missing broadcast to TeamThreadRange parallel_scan #6601

OpenMP:

• Improve performance of view initializations and filling with zeros #6573

General Enhancements

• Improve performance of random number generation when using a normal distribution on GPUs #6556
• Allocate temporary view with the user-provided execution space instance and do not initialize in unique algorithm #6598
• Add deduction guide for Kokkos::Array #6373
• Provide new public headers <Kokkos_Clamp.hpp> and <Kokkos_MinMax.hpp> #6687
• Fix/improvement to remove_if parallel algorithm: use the provided execution space instance for temporary allocations and drop unnecessaryinitialization + avoid evaluating twice the predicate during final pass #6747
• Add runtime function to query the number of devices and make device ID consistent with KOKKOS_VISIBLE_DEVICES #6713
• simd: support vector_aligned_tag #6243
• Avoid unnecessary allocation when default constructing Bitset #6524
• Fix constness for views in std algorithms #6813
• Improve error message on unsafe implicit conversion in MDRangePolicy #6855
• CTAD (deduction guides) for RangePolicy #6850
• CTAD (deduction guides) for MDRangePolicy #5516

Build System Changes

• Require Kokkos_ENABLE_ATOMICS_BYPASS option to bypass atomic operation for Serial backend only builds #6692
• Add support for RISCV and the Milk-V's Pioneer #6773
• Add C++26 standard to CMake setup #6733
• Fix Makefile when using gnu_generate_makefile.sh and make >= 4.3 #6606
• Cuda: Fix configuring with CMake >= 3.28.4 - temporary fallback to internal CudaToolkit.cmake #6898

Incompatibilities (i.e. breaking changes)

• Remove all DEPRECATED_CODE_3 option and all code that was guarded by it #6523
• Drop guards to accommodate external code defining KOKKOS_ASSERT #6665
• Profiling::ProfilingSection(std::string) constructor marked explicit and nodiscard #6690
• Add bound check preconditions for RangePolicy and MDRangePolicy #6617 #6726
• Add checks for unsafe implicit conversions in RangePolicy #6754
• Remove Kokkos::[b]half_t volatile overloads #6579
• Remove KOKKOS_IMPL_DO_NOT_USE_PRINTF #6593
• Check matching static extents in View constructor #5190
• Tools(profiling): fix typo Kokkos_Tools_Optim[i]zationGoal #6642
• Remove variadic range policy constructor (disallow passing multiple trailing chunk size arguments) #6845
• Improve message on view out of bounds access and always abort #6861
• Drop KOKKOS_ENABLE_INTEL_MM_ALLOC macro #6797
• Remove Kokkos::Experimental::LogicalMemorySpace (without going through deprecation) #6557
• Remove Experimental::HBWSpace and support for linking against memkind #6791
• Drop librt TPL and associated KOKKOS_ENABLE_LIBRT macro #6798
• Drop support for old CPU architectures (ARCH_BGQ, ARCH_POWER7, ARCH_WSM and associated ARCH_SSE4 macro) #6806
• Drop support for deprecated command-line arguments and environment variables #6744

Deprecations

• Provide kokkos_swap as part of Core and deprecate Experimental::swap in Algorithms #6697
• Deprecate {Cuda,HIP}::detect_device_count() and Cuda::[detect_]device_arch() #6710
• Deprecate ExecutionSpace::in_parallel() #6582

Bug Fixes

• Fix team-level MDRange reductions: #6511
• Fix CUDA and SYCL small value type (16-bit) team reductions #5334
• Enable {transform_}exclusive_scan in place #6667
• fill_random overload that do not take an execution space instance argument should fence #6658
• HIP,Cuda,OpenMPTarget: Fixup use provided execution space when copying host inaccessible reduction result #6777
• Fix typo in cuda_func_set_attribute[s]_wrapper preventing proper setting of desired occupancy #6786
• Avoid undefined behavior due to conversion between signed and unsigned integers in shift_{right, left}_team_impl #6821
• Fix a bug in Makefile.kokkos when using AMD GPU architectures as AMD_GFXYYY #6892

4.2.01 (2023-12-07)

Full Changelog

Backend and Architecture Enhancements:

CUDA:

• Add warp sync for parallel_reduce to avoid race condition #6630, #6746

HIP:

• Fix Graph "multiple definition of" linking error (missing inline specifier) #6624
• Add support for gfx940 (AMD Instinct MI300 GPU) #6671

Build System

• CMake: Don't let Kokkos set CMAKE_CXX_FLAGS for Trilinos builds #6742

Bug Fixes

• Remove deprecation warning for AllocationMechanism for GCC <11.0 #6653
• Fix bug early tools finalize with non-default host execution instances #6635
• Fix various issues for MSVC CUDA builds #6659
• Fix "extra ;" warning with -pedantic flag in <Kokkos_SIMD_Scalar.hpp> #6510

4.2.00 (2023-11-06)

Full Changelog

Features:

• SIMD: significant improvements to SIMD support and alignment with C++26 SIMD
  • add Kokkos::abs overload for SIMD types #6069
  • add generator constructors #6347
  • convert binary operators to hidden friends #6320
  • add shift operators #6109
  • add float support #6177
  • add remaining gather_from and scatter_to overloads #6220
  • define simd math function overloads in the Kokkos namespace #6465, #6487
  • Kokkos_ENABLE_NATIVE=ON autodetects SIMD types supported #6188
  • fix AVX2 SIMD support for ZEN2 AMD CPU #6238
• Kokkos::printf #6083
• Kokkos::sort: support custom comparator #6253
• half_t and bhalf_t numeric traits #5778
• half_t and bhalf_t mixed comparisons #6407
• half_t and bhalf_t mathematical functions #6124
• TeamThreadRange parallel_scan with return value #6090, #6301, #6302, #6303, #6307
• ThreadVectorRange parallel_scan with return value #6235, #6242, #6308, #6305, #6292
• Add team-level std algorithms #6200, #6205, #6207, #6208, #6209, #6210, #6211, #6212, #6213, #6256, #6258, #6350, #6351
• Serial: Allow for distinct execution space instances #6441

Backend and Architecture Enhancements:

CUDA:

• Fixed potential data race in Cuda parallel_reduce #6236
• Use cudaMallocAsync by default #6402
• Bugfix for using Kokkos from a thread of execution #6299

HIP:

• New naming convention for AMD GPU: VEGA906, VEGA908, VEGA90A, NAVI1030 to AMD_GFX906, AMD_GFX908, AMD_GFX90A, AMD_GFX1030 #6266
• Add initial support for gfx942: #6358
• Improve reduction performance #6229
• Deprecate HIP(hipStream_t,bool) constructor #6401
• Add support for Graph #6370
• Improve reduction performance when using Teams #6284
• Fix concurrency calculation #6479
• Fix potential data race in HIP parallel_reduce #6429

SYCL:

• Enforce external sycl::queues to be in-order #6246
• Improve reduction performance: #6272 #6271 #6270 #6264
• Allow using the SYCL execution space on AMD GPUs #6321
• Allow sorting via native oneDPL to support Views with stride=1 #6322
• Make in-order queues the default via macro #6189

OpenACC:

• Support Clacc compiler #6250

General Enhancements

• Add missing is_*_view traits and is_*_view_v helper variable templates for DynRankView, DynamicView, OffsetView, ScatterView containers #6195
• Make nvcc_wrapper and compiler_launcher scripts more portable by switching to a #!/usr/bin/env shebang #6357
• Add an improved Kokkos::malloc / Kokkos::free performance test #6377
• Ensure Views with size==0 can be used with deep_copy #6273
• Kokkos::abort is moved to header Kokkos_Abort.hpp #6445
• KOKKOS_ASSERT, KOKKOS_EXPECTS, KOKKOS_ENSURES are moved to header Kokkos_Assert.hpp #6445
• Add a permuted-index mode to the gups benchmark #6378
• Check for overflow during backend initialization #6159
• Make constraints on Kokkos::sort more visible #6234 and cleanup API #6239
• Add converting assignment to DualView: #6474

Build System Changes

• Export Kokkos_CXX_COMPILER_VERSION #6282
• Disable default oneDPL support in Trilinos #6342

Incompatibilities (i.e. breaking changes)

• Ensure that Kokkos::complex only gets instantiated for cv-unqualified floating-point types #6251
• Removed (deprecated-3) support for volatile join operators in reductions #6385
• Enforce ViewCtorArgs restrictions for create_mirror_view #6304
• SIMD types for ARM NEON are not autodetected anymore but need Kokkos_ARCH_ARM_NEON or Kokkos_ARCH_NATIVE=ON #6394
• Remove #include <iostream> from headers where possible #6482

Deprecations

• Deprecated Kokkos::vector #6252
• All host allocation mechanisms except for STD_MALLOC have been deprecated #6341

Bug Fixes

• Missing memory fence in RandomPool::free_state functions #6290
• Fix for corner case in Kokkos::Experimental::is_partitioned algorithm #6257
• Fix initialization of scratch lock variables in the Cuda backend #6433
• Fixes for Kokkos::Array #6372
• Fixed symlink configure issue for Windows #6241
• OpenMPTarget init-join fix #6444
• Fix atomic operations bug for Min and Max #6435
• Fix implementation for cyl_bessel_i0 #6484
• Fix various NVCC warnings in BinSort, Array, and bit manipulation function templates #6483

4.1.00 (2023-06-16)

Full Changelog

Features:

• Add <Kokkos_BitManipulation.hpp> header #4577 #5907 #5967 #6101
• Add UnorderedMapInsertOpTypes #5877 and documentation #350
• Add multiple reducers support for team-level parallel reduce #5727

Backend and Architecture Enhancements:

CUDA:

• Allow NVCC 12 to compile using C++20 flag #5977
• Remove ability to disable CMake option Kokkos_ENABLE_CUDA_LAMBDA and unconditionally enable CUDA extended lambda support. #5964
• Drop unnecessary fences around the memory allocation when using CudaUVMSpace in views #6008

HIP:

• Improve performance for parallel_reduce. Use different parameters for LightWeight kernels #6029 and #6160

SYCL:

• Only pass one wrapper object in SYCL reductions #6047
• Improve and simplify parallel_scan implementation #6064
• Remove workaround for submit_barrier not being enqueued properly #5504
• Fix guards for using scratch space with SYCL #6003
• Fix compiling SYCL with KOKKOS_IMPL_DO_NOT_USE_PRINTF_USAGE #6219

OpenMPTarget:

• Improve hierarchical parallelism for Intel architectures #6043
• Enable Cray compiler for the OpenMPTarget backend. #5889

HPX:

• Update HPX backend to use HPX's sender/receiver functionality #5628
• Increase minimum required HPX version to 1.8.0 #6132
• Implement HPX::in_parallel #6143

General Enhancements

• Export CMake Kokkos_{CUDA,HIP}_ARCHITECTURES variables #5919 #5925
• Add Kokkos::Profiling::ScopedRegion #5959 #5972
• Add support for View::rank[_dynamic]()#5870
• Detect incompatible relocatable device code mode to prevent ODR violations #5991
• Add (experimental) support for 32-bit Darwin and PPC #5916
• Add missing half and bhalf specialization of the infinity numeric trait #6055
• Add is_dual_view trait and align further with regular view #6120
• Allow templated functors in parallel_for, parallel_reduce and parallel_scan #5976
• Define KOKKOS_COMPILER_INTEL_LLVM and only define at most one KOKKOS_COMPILER* macro #5906
• Allow linking against build tree #6078
• Allow passing a temporary std::vector to partition_space #6167
• Kokkos can be used as an external dependency in Trilinos #6142, #6157 #6163
• Left align demangled stacktrace output #6191
• Improve OpenMP affinity warning to include MPI concerns #6185

Build System Changes

• Drop Kokkos_ENABLE_LAUNCH_COMPILER option which had no effect #6148
• Export variables for relevant Kokkos options with cmake#6142

Incompatibilities (i.e. breaking changes)

• Desul atomics always enabled #5801
• Drop KOKKOS_ENABLE_CUDA_ASM* and KOKKOS_ENABLE_*_ATOMICS macros #5940
• Drop KOKKOS_ENABLE_RFO_PREFETCH macro #5944
• Deprecate Kokkos_ENABLE_CUDA_LAMBDA configuration option and force it to ON #5964
• Remove TriBITS Kokkos subpackages #6104
• Cuda: Remove unused attach_texture_object #6129
• Drop Kokkos_ENABLE_PROFILING_LOAD_PRINT configuration option #6150
• Drop pointless Kokkos{Algorithms,Containers}_config.h files #6108

Deprecations

• Deprecate BinSort, BinOp1D, and BinOp3D default constructors #6131

Bug Fixes

• Fix SYCLTeamMember to take arguments for scratch sizes as std::size_t #5981
• Fix Kokkos_SIMD with AVX2 on 64-bit architectures #6075
• Fix an incorrectly returning size for SIMD uint64_t in AVX2 #6004
• Fix missing avx512 header file with gcc versions before 10 #6183
• Fix incorrect results of parallel_reduce of types smaller than int on CUDA and HIP: #5745
• CMake: update package compatibility mode when building within Trilinos #6012
• Fix warnings generated from internal uses of ALL_t rather than Kokkos::ALL_t #6028
• Fix bug in hpcbind script: check for correct Slurm variable #6116
• KokkosTools: Don't call callbacks before backends are initialized #6114
• Fix global fence in Kokkos::resize(DynRankView) #6184
• Fix BinSort support for strided views #6081
• Fix missing is_*_view traits in containers #6195
• Fix broken OpenMP target on NVHPC #6171
• Sorting an empty view should exit early and not fail #6130

4.0.01 (2023-04-14)

Full Changelog

Backend and Architecture Enhancements:

CUDA:

• Allow NVCC 12 to compile using C++20 flag #6020
• Add CUDA Ada architecture support #6022

HIP:

• Add support for AMDGPU target NAVI31 / RX 7900 XT(X): gfx1100 #6021
• HIP: Fix warning from std::memcpy #6019

SYCL:

• Fix SYCLTeamMember to take arguments for scratch sizes as std::size_t #5986

General Enhancements

• Fixup 4.0 change log #6023

Build System Changes

• Cherry-pick TriBITS update from Trilinos #6037
• CMake: update package compatibility mode when building within Trilinos #6013

Bug Fixes

• Fix an incorrectly returning size for SIMD uint64_t in AVX2 #6011
• Desul atomics: wrong value for desul::Impl::numeric_limits_max<uint64_t> #6018
• Fix warning in some user code when using std::memcpy #6000
• Fix excessive build times using Makefile.kokkos #6068

4.0.0 (2023-02-21)

Full Changelog

Features:

• Allow value types without default constructor in Kokkos::View with Kokkos::WithoutInitializing #5307
• parallel_scan with View as result type. #5146
• Introduced SharedSpace, an alias for a MemorySpace that is accessible by every ExecutionSpace. The memory is moved and then accessed locally. #5289
• Introduced SharedHostPinnedSpace, an alias for a MemorySpace that is accessible by every ExecutionSpace. The memory is pinned to the host and accessed via zero-copy access. #5405
• Add team- and thread-level sort, sort_by_key algorithms. #5317
• Groundwork for MDSpan integration. #4973 and #5304
• Introduced MD version of hierarchical parallelism: TeamThreadMDRange, ThreadVectorMDRange and TeamVectorMDRange. #5238

Backend and Architecture Enhancements:

CUDA:

• Allow CUDA PTX forward compatibility #3612 #5536 #5527
• Add support for NVIDIA Hopper GPU architecture #5538
• Don't rely on synchronization behavior of default stream in CUDA and HIP #5391
• Improve CUDA cache config settings #5706

HIP:

• Move HIP, HIPSpace, HIPHostPinnedSpace, and HIPManagedSpace out of the Experimental namespace #5383
• Don't rely on synchronization behavior of default stream in CUDA and HIP #5391
• Export AMD architecture flag when using Trilinos #5528
• Fix linking error (see OLCF issue) when using amdclang: #5539
• Remove support for MI25 and added support for Navi 1030 #5522
• Fix race condition when using HSA_XNACK=1 #5755
• Add parameter to force using GlobalMemory launch mechanism. This can be used when encountering compiler bugs with ROCm 5.3 and 5.4 #5796

SYCL:

• Delegate choice of workgroup size for parallel_reduce with RangePolicy to the compiler. #5227
• SYCL RangePolicy: manually specify workgroup size through chunk size #4875

OpenMPTarget:

• Select the right device #5492

OpenMP:

• Add partition_space #5105

General Enhancements

• Implement OffsetView constructor taking pairs and ViewCtorProp #5303
• Promote math constants to Kokkos::numbers namespace #5434
• Add overloads of hypot math function that take 3 arguments #5341
• Add fma fused multiply-add math function #5428
• Views using MemoryTraits::Atomic don't need volatile overloads for the value type anymore. #5455
• Added is_team_handle trait #5375
• Refactor desul atomics to support compiling CUDA with NVC++ #5431 #5497 #5498
• Support finding libquadmath with native compiler support #5286
• Add architecture flags for MSVC #5673
• SIMD backend for ARM NEON #5829

Build System Changes

• Let CMake determine OpenMP flags. #4105
• Update minimum compiler versions. #5323
• Makefile and CMake support for C++23 #5283
• Do not add -cuda to the link line with NVHPC compiler when the CUDA backend is not actually enabled #5485
• Only add -latomic in generated GNU makefiles when OpenMPTarget backend is enabled #5501 #5537 (3.7 patch release candidate)
• Kokkos_ENABLE_CUDA_LAMBDA now ON by default with NVCC #5580
• Fix enabling of relocatable device code when using CUDA as CMake language #5564
• Fix cmake configuration with CUDA 12 #5691

Incompatibilities (i.e. breaking changes)

• Require C++17 #5277
• Turn setting Kokkos_CXX_STANDARD into an error #5293
• Remove all deprecations in Kokkos 3 #5297
• Remove KOKKOS_COMPILER_CUDA_VERSION #5430
• Drop reciprocal_overflow_threshold numeric trait #5326
• Move reduction_identity out of <Kokkos_NumericTraits.hpp> into a new <Kokkos_ReductionIdentity.hpp> header #5450
• Reduction and scan routines will report an error if the join() operator they would use takes volatile-qualified parameters #5409
• ENABLE_CUDA_UVM is dropped in favor of using SharedSpace as MemorySpace explicitly #5608
• Remove Kokkos_ENABLE_CUDA_LDG_INTRINSIC option #5623
• Don't rely on synchronization behavior of default stream in CUDA and HIP - this potentially will break unintended implicit synchronization with other libraries such as MPI #5391
• Make ExecutionSpace::concurrency() a non-static member function #5655 and related PRs
• Remove code guarded by KOKKOS_ENABLE_DEPRECATED_CODE_3

Deprecations

• Deprecate CudaUVMSpace::available() which always returned true #5614
• Deprecate volatile-qualified members from Kokkos::pair and Kokkos::complex #5412
• Deprecate KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_* macros #5824 (oversight in 3.6)

Bug Fixes

• Avoid allocating memory for UniqueToken #5300
• Fix pragma ivdep in Kokkos_OpenMP_Parallel.hpp #5356
• Fix configuring with Threads support when rerunning CMake #5486
• Fix View assignment between LayoutLeft and LayoutRight with static extents #5535 (3.7 patch release candidate)
• Add fence() calls to sorting routine overloads that don't take an execution space parameter #5389
• ClockTic changed to 64 bit to fix overflow on Power #5577 (incl. in 3.7.01 patch release)
• Fix incorrect offset in CUDA and HIP parallel_scan for < 4 byte types #5555 (3.7 patch release candidate)
• Fix incorrect alignment behavior of scratch allocations in some corner cases (e.g. very small allocations) #5687 (3.7 patch release candidate)
• Add missing ReductionIdentity<char> specialization #5798
• Don't install standard algorithms headers multiple times #5670
• Fix max scratch size calculation for level 0 scratch in CUDA and HIP #5718

3.7.02 (2023-05-17)

Full Changelog

Backends and Archs Enhancements:

CUDA

• Add Hopper support and update nvcc_wrapper to work with CUDA-12 #5693

General Enhancements:

• sprintf -> snprintf #5787

Build System:

• Add error message when not using hipcc and when CMAKE_CXX_STANDARD is not set #5945

Bug Fixes:

• Fix Scratch allocation alignment issues #5692
• Fix Intel Classic Compiler ICE #5710
• Don't install std algorithm headers multiple times #5711
• Fix static init order issue in InitalizationSettings #5721
• Fix src/dst Properties in deep_copy(DynamicView,View) #5732
• Fix build on Fedora Rawhide #5782
• Finalize HIP lock arrays #5694
• Fix CUDA lock arrays for current Desul #5812
• Set the correct device/context in InterOp tests #5701

3.7.01 (2022-12-01)

Full Changelog

Bug Fixes:

• Add fences to all sorting routines not taking an execution space instance argument #5547
• Fix repeated team_reduce without barrier #5552
• Fix memory spaces in create_mirror_view overloads using view_alloc #5521
• Allow as_view_of_rank_n() to be overloaded for "special" scalar types #5553
• Fix warning calling a __host__ function from a __host__ __device__ from View:: as_view_of_rank_n #5591
• OpenMPTarget: adding implementation to set device id. #5557
• Use Kokkos::atomic_load to Correct Race Condition Giving Rise to Seg Faulting Error in OpenMP tests #5559
• cmake: define KOKKOS_ARCH_A64FX #5561
• Only link against libatomic in gnu-make OpenMPTarget build #5565
• Fix static extents assignment for LayoutLeft/LayoutRight assignment #5566
• Do not add -cuda to the link line with NVHPC compiler when the CUDA backend is not actually enabled #5569
• Export the flags in KOKKOS_AMDGPU_OPTIONS when using Trilinos #5571
• Add support for detecting MPI local rank with MPICH and PMI #5570 #5582
• Remove listing of undefined TPL dependencies #5573
• ClockTic changed to 64 bit to fix overflow on Power #5592
• Fix incorrect offset in CUDA and HIP parallel scan for < 4 byte types #5607
• Fix initialization of Cuda lock arrays #5622

3.7.00 (2022-08-22)

Full Changelog

Features:

• Use non-volatile join() member functions and operator+= in parallel_reduce/scan #4931 #4954 #4951
• Add SIMD sub package (requires C++17) #5016
• Add is_finalized() #5247
• Promote mathematical functions from namespace Kokkos::Experimental to namespace Kokkos #4791
• Promote min, max, clamp, minmax functions from namespace Kokkos::Experimental to namespace Kokkos #5170
• Add round, logb, nextafter, copysign, and signbit math functions #4768
• Add HIPManagedSpace, similar to CudaUVMSpace #5112
• Accept view construction allocation properties in create_mirror[_view,_view_and_copy] and resize/realloc #5125 #5095 #5035 #4805 #4844
• Allow MemorySpace::allocate() to be called with execution space #4826
• Experimental: Compile time view subscriber #4197

Backends and Archs Enhancements:

• Add support for Sapphire Rapids Intel architecture #5015
• Add support for ICX, SKL and ICL Intel architectures #5013 #4929
• Add arch flags for Intel GPU Ponte Vecchio #4932
• SYCL: require GPU if GPU architecture was set at configuration time (i.e. do not allow fallback to CPU device) #5264 #5222
• SYCL: Add SYCL::sycl_queue() for interoperability #5241
• SYCL: Loosen restriction for using built-in sycl::group_broadcast #4552
• SYCL: preserve address space #4396
• OpenMPTarget: Adding a workaound for team scan #5219
• OpenMPTarget: Adding logic to skip the kernel launch if league_size=0 #5067
• OpenMPTarget: Make sure Kokkos::abort() causes abnormal program termination when called on the host-side #4808
• HIP: Make HIPHostPinnedSpace coarse-grained #5152
• Refactor OpenMP parallel_for implementation to use more native OpenMP constructs #4664
• Add option to optimize for local CPU architecture Kokkos_ARCH_NATIVE #4930

Implemented enhancements

• Add command line argument/environment variable to print the configuration #5233
• Improve error message in view memory access violations #4950
• Remove unnecessary fences in View initialization #4823
• Make View::shmem_size() device-callable #4936
• Update numerics support for __float128 #5081
• Add log10 overload for Kokkos::complex #5009
• Add [[nodiscard]] to ScopeGuard #5224
• Add structured binding support for Kokkos::Array #4962
• Enable accessing Kokkos::Array elements in constant expressions #4916
• Mark as_view_of_rank_n as KOKKOS_FUNCTION #5248
• Cleanup/rework fence overloads #5148
• Assert that Layout construction from extents is valid in functions taking integer extents #5209
• Add fill_random overload that takes an execution space as first argument #5181
• Avoid some unnecessary fences in parallel_reduce/scan #5154
• Include KOKKOS_ENABLE_LIBDL in options when printing configuration #5086
• DynRankView: make layout() return the same as a corresponding static View #5026
• Use _mm_malloc for icpx #5012
• Avoid forcing matching execution spaces in BinSort constructor and sort() #4919
• Check number of bins in BinSort #4890
• Improve performance in parallel STL-like algorithms #4887 #4886
• Disable memset on A64FX and launch parallel_for instead (performance) #4884
• Allow non-power-of-two team sizes for team reductions and scans #4809

Harmonization of Kokkos execution environment initialization:

• Warn when unable to detect local MPI rank and user explicitly asked for it #5263
• Refactor parsing of command line arguments and environment variables #5221
• Refactor device selection at initialization #5211
• Rename tools settings for consistency #5201
• Print help only once #5128
• Update precedence rule in initialization #5130
• Warn instead of just ignoring user settings when kokkos-tools is disabled #5088
• Drop numa args in threads backend initialization #5127
• Warn users when a flag prefixed with -[-]kokkos is not recognized and do not remove it #5256
• Give back to Core what belongs to Core (aka moving tune_internals option from Tools back to Core) #5202

Build system updates:

• nvcc_wrapper: filter out -pedantic-errors from nvcc options #5235
• nvcc_wrapper: add known nvcc option --source-in-ptx #5052
• Link libdl as interface library #5179
• Only show GPU architectures with enabled corresponding backend #5119
• Enable optional external desul build #5021 #5132
• Export Kokkos_CXX_STANDARD variable with CMake #5068
• Suppress warnings with nvc++ #5031
• Disallow multiple host architectures in CMake #4996
• Do not include compiler warning flags in the compile option of the cmake target #4989
• AOT flags for OpenMPTarget targeting Intel GPUs #4915
• Repurpose Kokkos_ARCH_INTEL_GEN for SYCL to mean JIT to be conforming with OMPT #4894
• Replace amdgpu-target with offload-arch #4874
• Do not enable kokkos_launch_compiler when CMAKE_CXX_COMPILER_LAUNCHER is set #4870
• Move CMake version check up #4797

Incompatibilities:

• Remove KOKKOS_THREAD_LOCAL #5064
• Remove KOKKOS_ENABLE_POSIX_MEMALIGN #5011
• Remove unused KOKKOS_ENABLE_TM #4995
• Remove unused cmakedefine KOKKOS_ENABLE_COMPILER_WARNINGS #4883
• Remove unused KOKKOS_ENABLE_DUALVIEW_MODIFY_CHECK #4882
• Drop Instruction Set Architecture (ISA) macros #4981
• Warn in ScopeGuard about illegal usage #5250

Deprecations:

• Guard against non-public header inclusion #5178
• Raise deprecation warnings if non empty WorkTag class is used #5230
• Deprecate parallel_* overloads taking the label as trailing argument #5141
• Deprecate nested types in functional #5185
• Deprecate InitArguments struct and replace it with InitializationSettings #5135
• Deprecate finalize_all() #5134
• Deprecate command line arguments (other than --help) that are not prefixed with kokkos-* #5120
• Deprecate --[kokkos-]numa cmdline arg and KOKKOS_NUMA env var #5117
• Deprecate --[kokkos-]threads command line argument in favor of --[kokkos-]num-threads #5111
• Deprecate Kokkos::is_reducer_type #4957
• Deprecate OffsetView constructors taking index_list_type #4810
• Deprecate overloads of Kokkos::sort taking a parameter bool always_use_kokkos_sort #5382
• Warn about parallel_reduce cases that call join() with volatile-qualified arguments #5215

Bug Fixes:

• CUDA Reductions: Fix data races reported by Nvidia compute-sanitizer #4855
• Work around Intel compiler bug #5301
• Avoid allocating memory for UniqueToken #5300
• DynamicView: Properly resize mirror instances after construction #5276
• Remove Kokkos::Rank limit of 6 ranks #5271
• Do not forget to set last element to nullptr when removing a flag in Kokkos::initialize #5272
• Fix CUDA+MSVC build issue #5261
• Fix DynamicView::resize_serial #5220
• Fix cmake default compiler flags for unknown compiler #5217
• Fix move_backward #5191
• Fixing issue 5196 - missing symbol with intel compiler #5207
• Preserve KOKKOS_INVALID_INDEX in ViewDimension and ArrayLayout construction #5188
• Finalize deep_copy_space early avoiding printing to std::cerr for Cuda #5151
• Use correct policy in Threads MDRange parallel_reduce #5123
• Fix building with NVCC as the CXX compiler while the CUDA backend is not enabled #5115
• OpenMPTarget Index range fix for MDRange. #5089
• Fix bug with CUDA's team reduction for empty ranges #5079
• Fix using ZeroMemset for Serial #5077
• Fix Kokkos::Vector::push_back for default execution space #5047
• ScatterView: Fix ScatterMin/ScatterMax to use proper atomics #5045
• Fix calling ZeroMemset in deep_copy #5040
• Make View self-assignment not produce double-free #5024
• Guard against unrecognized pragma with intel compilers #5019
• Fix racing condition in HIPParallelLaunch #5008
• KokkosP: Fix device_id in profiling #4997
• Fix for Kokkos::vector::insert into empty vector with begin and end iterators #4988
• Fix Core header files installation #4984
• Fix bounds errors with Kokkos::sort #4980
• Fixup let RangePolicy::set_chunk_size return a reference to self #4918
• Fix allocating large Views #4907
• Fix combined reductions with Kokkos::View #4896
• Fixed _CUDA_ARCH__ to __CUDA_ARCH__ for CUDA LDG #4893
• Fixup View::access() truncate parameter pack #4876
• Fix abort with HIP backend for ROCm 5.0.2 and beyond #4873
• Fix HIP version when printing the configuration #4872
• Fix scratch lock array when using scratch level 1 #4871
• Fix Makefile.kokkos to work with fujitsu compiler #4867
• cmake: Correct link THREADS link option #4854
• UniqueToken impl_acquire function should be device only #4819
• Fix example calls to non existing static print_configuration #4806
• Fix requests for large team scratch sizes #4728

3.6.01 (2022-05-23)

Full Changelog

Bug Fixes:

• Fix Threads: Fix serial resizing scratch space (3.6.01 cherry-pick) #5109
• Fix ScatterMin/ScatterMax to use proper atomics (3.6.01 cherry-pick) #5046
• Fix allocating large Views #4907
• Fix bounds errors with Kokkos::sort #4980
• Fix HIP version when printing the configuration #4872
• Fixed _CUDA_ARCH__ to __CUDA_ARCH__ for CUDA LDG #4893
• Fixed an incorrect struct initialization #5028
• Fix racing condition in HIPParallelLaunch #5008
• Avoid deprecation warnings with OpenMPExec::validate_partition #4982
• Make View self-assignment not produce double-free #5024

3.6.00 (2022-02-18)

Full Changelog

Features:

• Add C++ standard algorithms #4315
• Implement fill_random for DynRankView #4763
• Add bhalf_t #4543 #4653
• Add mathematical constants #4519
• Allow Kokkos::{create_mirror*,resize,realloc} to be used with WithoutInitializing #4486 #4337
• Implement KOKKOS_IF_ON_{HOST,DEVICE} macros #4660
• Allow setting the CMake language for Kokkos #4323

Perf bug fix

• Desul: Add ScopeCaller #4690
• Enable Desul atomics by default when using Makefiles #4606
• Unique token improvement #4741 #4748

Other improvements:

• Add math function long double overload on the host side #4712

Deprecations:

• Array reductions with pointer return types #4756
• Deprecate partition_master, validate_partition #4737
• Deprecate Kokkos_ENABLE_PTHREAD in favor of Kokkos_ENABLE_THREADS #4619 ** pair with use std::threads **
• Deprecate log2(unsigned) -> int (removing in next release) #4595
• Deprecate Kokkos::Impl::is_view #4592
• Deprecate KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_* macros and the ActiveExecutionMemorySpace alias #4668

Backends and Archs Enhancements:

SYCL:

• Update required SYCL compiler version #4749
• Cap vector size to kernel maximum for SYCL #4704
• Improve check for compatibility of vector size and subgroup size in SYCL #4579
• Provide chunk_size for SYCL #4635
• Use host-pinned memory for SYCL kernel memory #4627
• Use shuffle-based algorithm for scalar reduction #4608
• Implement pool of USM IndirectKernelMemory #4596
• Provide valid default team size for SYCL #4481

CUDA:

• Add checks for shmem usage in parallel_reduce #4548

HIP:

• Add support for fp16 in the HIP backend #4688
• Disable multiple kernel instantiations when using HIP (configure with -DKokkos_ENABLE_HIP_MULTIPLE_KERNEL_INSTANTIATIONS=ON to use) #4644
• Fix HIP scratch use per instance #4439
• Change allocation header to 256B alignment for AMD VEGA architecture #4753
• Add generic KOKKOS_ARCH_VEGA macro #4782
• Require ROCm 4.5 #4689

HPX:

• Adapt to HPX 1.7.0 which is now required #4241

OpenMP:

• Fix thread deduction for OpenMP for thread_count==0 #4541

OpenMPTarget:

• Update memory space size_type to improve performance (size_t -> unsigned) #4779

Other Improvements:

• Improve NVHPC support #4599
• Add Kokkos::Experimental::{min,max,minmax,clamp} #4629 #4506
• Use device type as template argument in Containers and Algorithms #4724 #4675
• Implement Kokkos::sort with execution space #4490
• Kokkos::resize always error out for mismatch in runtime rank #4681
• Print current call stack when calling Kokkos::abort() from the host #4672 #4671
• Detect mismatch of execution spaces in functors #4655
• Improve view label access on host #4647
• Error out for const scalar return type in reduction #4645
• Don't allow calling UnorderdMap::value_at for a set #4639
• Add KOKKOS_COMPILER_NVHPC macro, disable quiet_NaN and signaling_NaN #4586
• Improve performance of local_deep_copy #4511
• Improve performance when sorting integers #4464
• Add missing numeric traits (denorm_min, reciprocal_overflow_threshold, {quiet,silent}_NaN}) and make them work on cv-qualified types #4466 #4415 #4473 #4443

Implemented enhancements BuildSystem

• Manually compute IntelLLVM compiler version for older CMake versions #4760
• Add Xptxas without = to nvcc_wrapper #4646
• Use external GoogleTest optionally #4563
• Silent warnings about multiple optimization flags with nvcc_wrapper #4502
• Use the same flags in Makefile.kokkos for POWER7/8/9 as for CMake #4483
• Fix support for A64FX architecture #4745

Incompatibilities:

• Drop KOKKOS_ARCH_HIP macro when using generated GNU makefiles #4786
• Remove gcc-toolchain auto add for clang in Makefile.kokkos #4762

Bug Fixes:

• Lock constant memory in Cuda/HIP kernel launch with a mutex (thread safety) #4525
• Fix overflow for large requested scratch allocation #4551
• Fix Windows build in mingw #4564
• Fix kokkos_launch_compiler: escape $ character #4769 #4703
• Fix math functions with NVCC and GCC 5 as host compiler #4733
• Fix shared build with Intel19 #4725
• Do not install empty desul/src/ directory #4714
• Fix wrong device_id computation in identifier_from_devid (Profiling Interface) #4694
• Fix a bug in CUDA scratch memory pool (abnormally high memory consumption) #4673
• Remove eval of command args in hpcbind #4630
• SYCL fix to run when no GPU is detected #4623
• Fix layout_strides::span for rank-0 views #4605
• Fix SYCL atomics for local memory #4585
• Hotfix mdrange_large_deep_copy for SYCL #4581
• Fix bug when sorting integer using the HIP backend #4570
• Fix compilation error when using HIP with RDC #4553
• DynamicView: Fix deallocation extent #4533
• SYCL fix running parallel_reduce with TeamPolicy for large ranges #4532
• Fix bash syntax error in nvcc_wrapper #4524
• OpenMPTarget team_policy reduce fixes for init/join reductions #4521
• Avoid hangs in the Threads backend #4499
• OpenMPTarget fix reduction bug in parallel_reduce for TeamPolicy #4491
• HIP fix scratch space per instance #4439
• OpenMPTarget fix team scratch allocation #4431

3.5.00 (2021-10-19)

Full Changelog

Features:

• Add support for quad-precision math functions/traits #4098
• Adding ExecutionSpace partitioning function #4096
• Improve Python Interop Capabilities #4065
• Add half_t Kokkos::rand specialization #3922
• Add math special functions: erf, erfcx, expint1, Bessel functions, Hankel functions #3920
• Add missing common mathematical functions #4043 #4036 #4034
• Let the numeric traits be SFINAE-friendly #4038
• Add Desul atomics - enabling memory-order and memory-scope parameters #3247
• Add detection idiom from the C++ standard library extension version 2 #3980
• Fence Profiling Support in all backends #3966 #4304 #4258 #4232
• Significant SYCL enhancements (see below)

Deprecations:

• Deprecate CUDA_SAFE_CALL and HIP_SAFE_CALL #4249
• Deprecate Kokkos::Impl::Timer (Kokkos::Timer has been available for a long time) #4201
• Deprecate Experimental::MasterLock #4094
• Deprecate Kokkos_TaskPolicy.hpp (headers got reorganized, doesn't remove functionality) #4011
• Deprecate backward compatibility features #3978
• Update and deprecate is_space::host_memory/execution/mirror_space #3973

Backends and Archs Enhancements:

• Enabling constbitset constructors in kernels #4296
• Use ZeroMemset in View constructor to improve performance #4226
• Use memset in deep_copy #3944
• Add missing fence() calls in resize(View) that effectively do deep_copy(resized, orig) #4212
• Avoid allocations in resize and realloc #4207
• StaticCsrGraph: use device type instead of execution space to construct views #3991
• Consider std::sort when view is accessible from host #3929
• Fix CPP20 warnings except for volatile #4312

SYCL:

• Introduce SYCLHostUSMSpace #4268
• Implement SYCL TeamPolicy for vector_size > 1 #4183
• Enable 64bit ranges for SYCL #4211
• Don't print SYCL device info in execution space intialization #4168
• Improve SYCL MDRangePolicy performance #4161
• Use sub_groups in SYCL parallel_scan #4147
• Implement subgroup reduction for SYCL RangePolicy parallel_reduce #3940
• Use DPC++ broadcast extension in SYCL team_broadcast #4103
• Only fence in SYCL parallel_reduce for non-device-accessible result_ptr #4089
• Improve fencing behavior in SYCL backend #4088
• Fence all registered SYCL queues before deallocating memory #4086
• Implement SYCL::print_configuration #3992
• Reuse scratch memory in parallel_scan and TeamPolicy (decreases memory footprint) #3899 #3889

CUDA:

• Cuda improve heuristic for blocksize #4271
• Don't use [[deprecated]] for nvcc #4229
• Improve error message for NVHPC as host compiler #4227
• Update support for cuda reductions to work with types < 4bytes #4156
• Fix incompatible team size deduction in rare cases parallel_reduce #4142
• Remove UVM usage in DynamicView #4129
• Remove dependency between core and containers #4114
• Adding opt-in CudaMallocSync support when using CUDA version >= 11.2 #4026 #4233
• Fix a potential race condition in the CUDA backend #3999

HIP:

• Implement new blocksize deduction method for HIP Backend #3953
• Add multiple LaunchMechanism #3820
• Make HIP backend thread-safe #4170

Serial:

• Refactor Serial backend and fix thread-safety issue #4053

OpenMPTarget:

• OpenMPTarget: support array reductions in RangePolicy #4040
• OpenMPTarget: add MDRange parallel_reduce #4032
• OpenMPTarget: Fix bug in for the case of a reducer. #4044
• OpenMPTarget: verify process fix #4041

Implemented enhancements BuildSystem

Important BuildSystem Updates:

• Use hipcc architecture autodetection when Kokkos_ARCH is not set #3941
• Introduce Kokkos_ENABLE_DEPRECATION_WARNINGS and remove deprecated code with Kokkos_ENABLE_DEPRECATED_CODE_3 #4106 #3855

Other Improvements:

• Add allow-unsupported-compiler flag to nvcc-wrapper #4298
• nvcc_wrapper: fix errors in argument handling #3993
• Adds support for -time= and -time in nvcc_wrapper #4015
• nvcc_wrapper: suppress duplicates of GPU architecture and RDC flags #3968
• Fix TMPDIR support in nvcc_wrapper #3792
• NVHPC: update PGI compiler arch flags #4133
• Replace PGI with NVHPC (works for both) #4196
• Make sure that KOKKOS_CXX_HOST_COMPILER_ID is defined #4235
• Add options to Makefile builds for deprecated code and warnings #4215
• Use KOKKOS_CXX_HOST_COMPILER_ID for identifying CPU arch flags #4199
• Added support for Cray Clang to Makefile.kokkos #4176
• Add XLClang as compiler #4120
• Keep quoted compiler flags when passing to Trilinos #3987
• Add support for AMD Zen3 CPU architecture #3972
• Rename IntelClang to IntelLLVM #3945
• Add cppcoreguidelines-pro-type-cstyle-cast to clang-tidy #3522
• Add sve bit size definition for A64FX #3947 #3946
• Remove KOKKOS_ENABLE_DEBUG_PRINT_KERNEL_NAMES #4150

Other Changes:

Tool Enhancements:

• Retrieve original value from a point in a MultidimensionalSparseTuningProblem #3977
• Allow extension of built-in tuners with additional tuning axes #3961
• Added a categorical tuner #3955

Miscellaneous:

• hpcbind: Use double quotes around $@ when invoking user command #4284
• Add file and line to error message #3985
• Fix compiler warnings when compiling with nvc++ #4198
• Add OpenMPTarget CI build on AMD GPUs #4055
• CI: icpx is now part of intel container #4002

Incompatibilities:

• Remove pre CUDA 9 KOKKOS_IMPL_CUDA_* macros #4138

Bug Fixes:

• UnorderedMap::clear() should zero the size() #4130
• Add memory fence for HostSharedPtr::cleanup() #4144
• SYCL: Fix race conditions in TeamPolicy::parallel_reduce #4418
• Adding missing memory fence to serial exec space fence. #4292
• Fix using external SYCL queues in tests #4291
• Fix digits10 bug #4281
• Fixes constexpr errors with frounding-math on gcc < 10. #4278
• Fix compiler flags for PGI/NVHPC #4264
• Fix Zen2/3 also implying Zen Arch with Makefiles #4260
• Kokkos_Cuda.hpp: Fix shadow warning with cuda/11.0 #4252
• Fix issue w/ static initialization of function attributes #4242
• Disable long double hypot test on Power systems #4221
• Fix false sharing in random pool #4218
• Fix a missing memory_fence for debug shared alloc code #4216
• Fix two xl issues #4179
• Makefile.kokkos: fix (standard_in) 1: syntax error #4173
• Fixes for query_device example #4172
• Fix a bug when using HIP atomic with Kokkos::Complex #4159
• Fix mistaken logic in pthread creation #4157
• Define KOKKOS_ENABLE_AGGRESSIVE_VECTORIZATION when requesting Kokkos_ENABLE_AGGRESSIVE_VECTORIZATION=ON #4107
• Fix compilation with latest MSVC version #4102
• Fix incorrect macro definitions when compiling with Intel compiler on Windows #4087
• Fixup global buffer overflow in hand rolled string manipulation #4070
• Fixup heap buffer overflow in cmd line args parsing unit tests #4069
• Only add quotes in compiler flags for Trilinos if necessary #4067
• Fixed invocation of tools init callbacks #4061
• Work around SYCL JIT compiler issues with static variables #4013
• Fix TestDetectionIdiom.cpp test inclusion for Trilinos/TriBITS #4010
• Fixup allocation headers with OpenMPTarget backend #4003
• Add missing specialization for OMPT to Kokkos Random #3967
• Disable hypot long double test on power arches #3962
• Use different EBO workaround for MSVC (rebased) #3924
• Fix SYCL Kokkos::Profiling::(de)allocateData calls #3928

3.4.01 (2021-05-19)

Full Changelog

Bug Fixes:

• Windows: Remove atomic_compare_exchange_strong overload conflicts with Windows #4024
• OpenMPTarget: Fixup allocation headers with OpenMPTarget backend #4020
• OpenMPTarget: Add missing specailization for OMPT to Kokkos Random #4022
• AMD: Add support for AMD Zen3 CPU architecture #4021
• SYCL: Implement SYCL::print_configuration #4012
• Containers: staticcsrgraph: use device type instead of execution space to construct views #3998
• nvcc_wrapper: fix errors in argument handling, suppress duplicates of GPU architecture and RDC flags #4006
• CI: Add icpx testing to intel container #4004
• CMake/TRIBITS: Keep quoted compiler flags when passing to Trilinos #4007
• CMake: Rename IntelClang to IntelLLVM #3945

3.4.00 (2021-04-25)

Full Changelog

Highlights:

• SYCL Backend Almost Feature Complete
• OpenMPTarget Backend Almost Feature Complete
• Performance Improvements for HIP backend
• Require CMake 3.16 or newer
• Tool Callback Interface Enhancements
• cmath wrapper functions available now in Kokkos::Experimental

Features:

• Implement parallel_scan with ThreadVectorRange and Reducer #3861
• Implement SYCL Random #3849
• OpenMPTarget: Adding Implementation for nested reducers #3845
• Implement UniqueToken for SYCL #3833
• OpenMPTarget: UniqueToken::Global implementation #3823
• DualView sync's on ExecutionSpaces #3822
• SYCL outer TeamPolicy parallel_reduce #3818
• SYCL TeamPolicy::team_scan #3815
• SYCL MDRangePolicy parallel_reduce #3801
• Enable use of execution space instances in ScatterView #3786
• SYCL TeamPolicy nested parallel_reduce #3783
• OpenMPTarget: MDRange with TagType for parallel_for #3781
• Adding OpenMPTarget parallel_scan #3655
• SYCL basic TeamPolicy #3654
• OpenMPTarget: scratch memory implementation #3611

Implemented enhancements Backends and Archs:

• SYCL choose a specific GPU #3918
• [HIP] Lock access to scratch memory when using Teams #3916
• [HIP] fix multithreaded access to get_next_driver #3908
• Forward declare HIPHostPinnedSpace and SYCLSharedUSMSpace #3902
• Let SYCL USMObjectMem use SharedAllocationRecord #3898
• Implement clock_tic for SYCL #3893
• Don't use a static variable in HIPInternal::scratch_space #3866(kokkos#3866)
• Reuse memory for SYCL parallel_reduce #3873
• Update SYCL compiler in CI #3826
• Introduce HostSharedPtr to manage m_space_instance for Cuda/HIP/SYCL #3824
• [HIP] Use shuffle for range reduction #3811
• OpenMPTarget: Changes to the hierarchical parallelism #3808
• Remove ExtendedReferenceWrapper for SYCL parallel_reduce #3802
• Eliminate sycl_indirect_launch #3777
• OpenMPTarget: scratch implementation for parallel_reduce #3776
• Allow initializing SYCL execution space from sycl::queue and SYCL::impl_static_fence #3767
• SYCL TeamPolicy scratch memory alternative #3763
• Alternative implementation for SYCL TeamPolicy #3759
• Unify handling of synchronous errors in SYCL #3754
• core/Cuda: Half_t updates for cgsolve #3746
• Unify HIPParallelLaunch structures #3733
• Improve performance for SYCL parallel_reduce #3732
• Use consistent types in Kokkos_OpenMPTarget_Parallel.hpp #3703
• Implement non-blocking kernel launches for HIP backend #3697
• Change SYCLInternal::m_queue std::unique_ptr -> std::optional #3677
• Use alternative SYCL parallel_reduce implementation #3671
• Use runtime values in KokkosExp_MDRangePolicy.hpp #3626
• Clean up AnalyzePolicy #3564
• Changes for indirect launch of SYCL parallel reduce #3511

Implemented enhancements BuildSystem:

• Also require C++14 when building gtest #3912
• Fix compiling SYCL with OpenMP #3874
• Require C++17 for SYCL (at configuration time) #3869
• Add COMPILE_DEFINITIONS argument to kokkos_create_imported_tpl #3862
• Do not pass arch flags to the linker with no rdc #3846
• Try compiling C++14 check with C++14 support and print error message #3843
• Enable HIP with Cray Clang #3842
• Add an option to disable header self containment tests #3834
• CMake check for C++14 #3809
• Prefer -std=* over --std=* #3779
• Kokkos launch compiler updates #3778
• Updated comments and enabled no-op for kokkos_launch_compiler #3774
• Apple's Clang not correctly recognised #3772
• kokkos_launch_compiler + CUDA auto-detect arch #3770
• Add Spack test support for Kokkos #3753
• Split SYCL tests for aot compilation #3741
• Use consistent OpenMP flag for IntelClang #3735
• Add support for -Wno-deprecated-gpu-targets #3722
• Add configuration to target CUDA compute capability 8.6 #3713
• Added VERSION and SOVERSION to KOKKOS_INTERNAL_ADD_LIBRARY #3706
• Add fast-math to known NVCC flags #3699
• Add MI-100 arch string #3698
• Require CMake >=3.16 #3679
• KokkosCI.cmake, KokkosCTest.cmake.in, CTestConfig.cmake.in + CI updates #2844

Implemented enhancements Tools:

• Improve readability of the callback invocation in profiling #3860
• V1.1 Tools Interface: incremental, action-based #3812
• Enable launch latency simulations #3721
• Added metadata callback to tools interface #3711
• MDRange Tile Size Tuning #3688
• Added support for command-line args for kokkos-tools #3627
• Query max tile sizes for an MDRangePolicy, and set tile sizes on an existing policy #3481

Implemented enhancements Other:

• Try detecting ndevices in get_gpu #3921
• Use strcmp to compare names() #3909
• Add execution space arguments for constructor overloads that might allocate a new underlying View #3904
• Prefix labels in internal use of kokkos_malloc #3891
• Prefix labels for internal uses of SharedAllocationRecord #3890
• Add missing hypot math function #3880
• Unify algorithm unit tests to avoid code duplication #3851
• DualView.template view() better matches for Devices in UVMSpace cases #3857
• More extensive disentangling of Policy Traits #3829
• Replaced nanosleep and sched_yield with STL routines #3825
• Constructing Atomic Subviews #3810
• Metadata Declaration in Core #3729
• Allow using tagged final functor in parallel_reduce #3714
• Major duplicate code removal in SharedAllocationRecord specializations #3658

Fixed bugs:

• Provide forward declarations in Kokkos_ViewLayoutTiled.hpp for XL #3911
• Fixup absolute value of floating points in Kokkos complex #3882
• Address intel 17 ICE #3881
• Add missing pow(Kokkos::complex) overloads #3868
• Fix bug {pow, log}(Kokkos::complex) #3866(kokkos#3866)
• Cleanup writing to output streams in Cuda #3859
• Fixup cache CUDA fallback execution space instance used by DualView::sync #3856
• Fix cmake warning with pthread #3854
• Fix typo FOUND_CUDA_{DRIVVER -> DRIVER} #3852
• Fix bug in SYCL team_reduce #3848
• Atrocious bug in MDRange tuning #3803
• Fix compiling SYCL with Kokkos_ENABLE_TUNING=ON #3800
• Fixed command line parsing bug #3797
• Workaround race condition in SYCL parallel_reduce #3782
• Fix Atomic{Min,Max} for Kepler30 #3780
• Fix SYCL typo #3755
• Fixed Kokkos_install_additional_files macro #3752
• Fix a typo for Kokkos_ARCH_A64FX #3751
• OpenMPTarget: fixes and workarounds to work with "Release" build type #3748
• Fix parsing bug for number of devices command line argument #3724
• Avoid more warnings with clang and C++20 #3719
• Fix gcc-10.1 C++20 warnings #3718
• Fix cuda cache config not being set correct #3712
• Fix dualview deepcopy perftools #3701
• use drand instead of frand in drand #3696

Incompatibilities:

• Remove unimplemented member functions of SYCLDevice #3919
• Replace cl::sycl #3896
• Get rid of SYCL workaround in Kokkos_Complex.hpp #3884
• Replace most uses of if_c #3883
• Remove Impl::enable_if_type #3863
• Remove HostBarrier test #3847
• Avoid (void) interface #3836
• Remove VerifyExecutionCanAccessMemorySpace #3813
• Avoid duplicated code in ScratchMemorySpace #3793
• Remove superfluous FunctorFinal specialization #3788
• Rename cl::sycl -> sycl in Kokkos_MathematicalFunctions.hpp #3678
• Remove integer_sequence backward compatibility implementation #3533

Enabled tests:

• Fixup re-enable core performance tests #3903
• Enable more SYCL tests #3900
• Restrict MDRange Policy tests for Intel GPUs #3853
• Disable death tests for rawhide #3844
• OpenMPTarget: Block unit tests that do not pass with the nvidia compiler #3839
• Enable Bitset container test for SYCL #3830
• Enable some more SYCL tests #3744
• Enable SYCL atomic tests #3742
• Enable more SYCL perf_tests #3692
• Enable examples for SYCL #3691

3.3.01 (2021-01-06)

Full Changelog

Bug Fixes:

• Fix severe performance bug in DualView which added memcpys for sync and modify #3693
• Fix performance bug in CUDA backend, where the cuda Cache config was not set correct.

3.3.00 (2020-12-16)

Full Changelog

Features:

• Require C++14 as minimum C++ standard. C++17 and C++20 are supported too.
• HIP backend is nearly feature complete. Kokkos Dynamic Task Graphs are missing.
• Major update for OpenMPTarget: many capabilities now work. For details contact us.
• Added DPC++/SYCL backend: primary capabilites are working.
• Added Kokkos Graph API analogous to CUDA Graphs.
• Added parallel_scan support with TeamThreadRange #3536
• Added Logical Memory Spaces #3546
• Added initial half precision support #3439
• Experimental feature: control cuda occupancy #3379

Implemented enhancements Backends and Archs:

• Add a64fx and fujitsu Compiler support #3614
• Adding support for AMD gfx908 archictecture #3375
• SYCL parallel_for MDRangePolicy #3583
• SYCL add parallel_scan #3577
• SYCL custom reductions #3544
• SYCL Enable container unit tests #3550
• SYCL feature level 5 #3480
• SYCL Feature level 4 (parallel_for) #3474
• SYCL feature level 3 #3451
• SYCL feature level 2 #3447
• OpenMPTarget: Hierarchial reduction for + operator on scalars #3504
• OpenMPTarget hierarchical #3411
• HIP Add Impl::atomic_[store,load] #3440
• HIP enable global lock arrays #3418
• HIP Implement multiple occupancy paths for various HIP kernel launchers #3366

Implemented enhancements Policies:

• MDRangePolicy: Let it be semiregular #3494
• MDRangePolicy: Check narrowing conversion in construction #3527
• MDRangePolicy: CombinedReducers support #3395
• Kokkos Graph: Interface and Default Implementation #3362
• Kokkos Graph: add Cuda Graph implementation #3369
• TeamPolicy: implemented autotuning of team sizes and vector lengths #3206
• RangePolicy: Initialize all data members in default constructor #3509

Implemented enhancements BuildSystem:

• Auto-generate core test files for all backends #3488
• Avoid rewriting test files when calling cmake #3548
• RULE_LAUNCH_COMPILE and RULE_LAUNCH_LINK system for nvcc_wrapper #3136
• Adding -include as a known argument to nvcc_wrapper #3434
• Install hpcbind script #3402
• cmake/kokkos_tribits.cmake: add parsing for args #3457

Implemented enhancements Tools:

• Changed namespacing of Kokkos::Tools::Impl::Impl::tune_policy #3455
• Delegate to an impl allocate/deallocate method to allow specifying a SpaceHandle for MemorySpaces #3530
• Use the Kokkos Profiling interface rather than the Impl interface #3518
• Runtime option for tuning #3459
• Dual View Tool Events #3326

Implemented enhancements Other:

• Abort on errors instead of just printing #3528
• Enable C++14 macros unconditionally #3449
• Make ViewMapping trivially copyable #3436
• Rename struct ViewMapping to class #3435
• Replace enums in Kokkos_ViewMapping.hpp (removes -Wextra) #3422
• Use bool for enums representing bools #3416
• Fence active instead of default execution space instances #3388
• Refactor parallel_reduce fence usage #3359
• Moved Space EBO helpers to Kokkos_EBO #3357
• Add remove_cvref type trait #3340
• Adding identity type traits and update definition of identity_t alias #3339
• Add is_specialization_of type trait #3338
• Make ScratchMemorySpace semi-regular #3309
• Optimize min/max atomics with early exit on no-op case #3265
• Refactor Backend Development #2941

Fixed bugs:

• Fixup MDRangePolicy construction from Kokkos arrays #3591
• Add atomic functions for unsigned long long using gcc built-in #3588
• Fixup silent pointless comparison with zero in checked_narrow_cast (compiler workaround) #3566
• Fixes for ROCm 3.9 #3565
• Fix windows build issues which crept in for the CUDA build #3532
• HIP Fix atomics of large data types and clean up lock arrays #3529
• Pthreads fix exception resulting from 0 grain size #3510
• Fixup do not require atomic operation to be default constructible #3503
• Fix race condition in HIP backend #3467
• Replace KOKKOS_DEBUG with KOKKOS_ENABLE_DEBUG #3458
• Fix multi-stream team scratch space definition for HIP #3398
• HIP fix template deduction #3393
• Fix compiling with HIP and C++17 #3390
• Fix sigFPE in HIP blocksize deduction #3378
• Type alias change: replace CS with CTS to avoid conflicts with NVSHMEM #3348
• Clang compilation of CUDA backend on Windows #3345
• Fix HBW support #3343
• Added missing fences to unique token #3260

Incompatibilities:

• Remove unused utilities (forward, move, and expand_variadic) from Kokkos::Impl #3535
• Remove unused traits #3534
• HIP: Remove old HCC code #3301
• Prepare for deprecation of ViewAllocateWithoutInitializing #3264
• Remove ROCm backend #3148

3.2.01 (2020-11-17)

Full Changelog

Fixed bugs:

• Disallow KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE in shared library builds #3332
• Do not install libprinter-tool when testing is enabled #3313
• Fix restrict/alignment following refactor #3373
  • Intel fix: workaround compiler issue with using statement #3383
• Fix zero-length reductions #\3364
  • Pthread zero-length reduction fix #3452
  • HPX zero-length reduction fix #3470
  • cuda/9.2 zero-length reduction fix #3580
• Fix multi-stream scratch #\3269
• Guard KOKKOS_ALL_COMPILE_OPTIONS if Cuda is not enabled #3387
• Do not include link flags for Fortran linkage #3384
• Fix NVIDIA GPU arch macro with autodetection #3473
• Fix libdl/test issues with Trilinos #3543
  • Register Pthread as Tribits option to be enabled with Trilinos #3558

Implemented enhancements:

• Separate Cuda timing-based tests into their own executable #3407

3.2.00 (2020-08-19)

Full Changelog

Implemented enhancements:

• HIP:Enable stream in HIP #3163
• HIP:Add support for shuffle reduction for the HIP backend #3154
• HIP:Add implementations of missing HIPHostPinnedSpace methods for LAMMPS #3137
• HIP:Require HIP 3.5.0 or higher #3099
• HIP:WorkGraphPolicy for HIP #3096
• OpenMPTarget: Significant update to the new experimental backend. Requires C++17, works on Intel GPUs, reference counting fixes. #3169
• Windows Cuda support #3018
• Pass -Wext-lambda-captures-this to NVCC when support for __host__ __device__ lambda is enabled from CUDA 11 #3241
• Use explicit staging buffer for constant memory kernel launches and cleanup host/device synchronization #3234
• Various fixup to policies including making TeamPolicy default constructible and making RangePolicy and TeamPolicy assignable: #3202 , #3203 , #3196
• Annotations for DefaultExectutionSpace and DefaultHostExectutionSpace to use in static analysis #3189
• Add documentation on using Spack to install Kokkos and developing packages that depend on Kokkos #3187
• Add OpenMPTarget backend flags for NVC++ compiler #3185
• Move deep_copy/create_mirror_view on Experimental::OffsetView into Kokkos:: namespace #3166
• Allow for larger block size in HIP #3165
• View: Added names of Views to the different View initialize/free kernels #3159
• Cuda: Caching cudaFunctorAttributes and whether L1/Shmem prefer was set #3151
• BuildSystem: Improved performance in default configuration by defaulting to Release build #3131
• Cuda: Update CUDA occupancy calculation #3124
• Vector: Adding data() to Vector #3123
• BuildSystem: Add CUDA Ampere configuration support #3122
• General: Apply [[noreturn]] to Kokkos::abort when applicable #3106
• TeamPolicy: Validate storage level argument passed to TeamPolicy::set_scratch_size() #3098
• BuildSystem: Make kokkos_has_string() function in Makefile.kokkos case insensitive #3091
• Modify KOKKOS_FUNCTION macro for clang-tidy analysis #3087
• Move allocation profiling to allocate/deallocate calls #3084
• BuildSystem: FATAL_ERROR when attempting in-source build #3082
• Change enums in ScatterView to types #3076
• HIP: Changes for new compiler/runtime #3067
• Extract and use get_gpu #3061 , #3048
• Add is_allocated to View-like containers #3059
• Combined reducers for scalar references #3052
• Add configurable capacity for UniqueToken #3051
• Add installation testing #3034
• HIP: Add UniqueToken #3020
• Autodetect number of devices #3013

Fixed bugs:

• Check error code from cudaStreamSynchronize in CUDA fences #3255
• Fix issue with C++ standard flags when using nvcc\_wrapper with PGI #3254
• Add missing threadfence in lock-based atomics #3208
• Fix dedup of linker flags for shared lib on CMake <=3.12 #3176
• Fix memory leak with CUDA streams #3170
• BuildSystem: Fix OpenMP Target flags for Cray #3161
• ScatterView: fix for OpenmpTarget remove inheritance from reducers #3162
• BuildSystem: Set OpenMP flags according to host compiler #3127
• OpenMP: Fix logic for nested omp in partition_master bug #3101
• nvcc_wrapper: send --cudart to nvcc instead of host compiler #3092
• BuildSystem: Fixes for Cuda/11 and c++17 #3085
• HIP: Fix print_configuration #3080
• Conditionally define get_gpu #3072
• Fix bounds for ranges in random number generator #3069
• Fix Cuda minor arch check #3035
• BuildSystem: Add -expt-relaxed-constexpr flag to nvcc_wrapper #3021

Incompatibilities:

• Remove ETI support #3157
• Remove KOKKOS_INTERNAL_ENABLE_NON_CUDA_BACKEND #3147
• Remove core/unit_test/config #3146
• Removed the preprocessor branch for KOKKOS_ENABLE_PROFILING #3115
• Disable profiling with MSVC #3066

Closed issues:

• Silent error (Validate storage level arg to set_scratch_size) #3097

• Remove KOKKKOS_ENABLE_PROFILING Option #3095

• Cuda 11 -> allow C++17 #3083

• In source build failure not explained #3081

• Allow naming of Views for initialization kernel #3070

• DefaultInit tests failing when using CTest resource allocation feature #3040

• Add installation testing. #3037

• nvcc_wrapper needs to handle -expt-relaxed-constexpr flag #3017

• CPU core oversubscription warning on macOS with OpenMP backend #2996

• Default behavior of KOKKOS_NUM_DEVICES to use all devices available #2975

• Assert blocksize > 0 #2974

• Add ability to assign kokkos profile function from executable #2973

• ScatterView Support for the pre/post increment operator #2967

• Compiler issue: Cuda build with clang 10 has errors with the atomic unit tests #3237

• Incompatibility of flags for C++ standard with PGI v20.4 on Power9/NVIDIA V100 system #3252

• Error configuring as subproject #3140

• CMake fails with Nvidia compilers when the GPU architecture option is not supplied (Fix configure with OMPT and Cuda) #3207

• PGI compiler being passed the gcc -fopenmp flag #3125

• Cuda: Memory leak when using CUDA stream #3167

• RangePolicy has an implicitly deleted assignment operator #3192

• MemorySpace::allocate needs to have memory pool counting. #3064

• Missing write fence for lock based atomics on CUDA #3038

• CUDA compute capability version check problem #3026

• Make DynRankView fencing consistent #3014

• nvcc_wrapper cant handle -Xcompiler -o out.o #2993

• Reductions of non-trivial types of size 4 fail in CUDA shfl operations #2990

• complex_double misalignment in reduce, clang+CUDA #2989

• Span of degenerated (zero-length) subviews is not zero in some special cases #2979

• Rank 1 custom layouts dont work as expected. #2840

3.1.01 (2020-04-14)

Full Changelog

Fixed bugs:

• Fix complex_double misalignment in reduce, clang+CUDA #2989
• Fix compilation fails when profiling disabled and CUDA enabled #3001
• Fix cuda reduction of non-trivial scalars of size 4 #2990
• Configure and install version file when building in Trilinos #2957
• Fix OpenMPTarget build missing include and namespace #3000
• fix typo in KOKKOS_SET_EXE_PROPERTY() #2959
• Fix non-zero span subviews of zero sized subviews #2979

3.1.00 (2020-04-14)

Full Changelog

Features:

• HIP Support for AMD
• OpenMPTarget Support with clang
• Windows VS19 (Serial) Support #1533

Implemented enhancements:

• generate_makefile.bash should allow tests to be disabled #2886
• clang/7+cuda/9 build -Werror-unused parameter error in nightly test #2884
• ScatterView memory space is not user settable #2826
• clang/8+cuda/10.0 build error with c++17 #2809
• warnings.... #2805
• Kokkos version in cpp define #2787
• Remove Defunct QThreads Backend #2751
• Improve Kokkos::fence behavior with multiple execution spaces #2659
• polylithic(?) initialization of Kokkos #2658
• Unnecessary(?) check for host execution space initialization from Cuda initialization #2652
• Kokkos error reporting failures with CUDA GPUs in exclusive mode #2471
• atomicMax equivalent (and other atomics) #2401
• Fix alignment for Kokkos::complex #2255
• Warnings with Cuda 10.1 #2206
• dual view with Kokkos::ViewAllocateWithoutInitializing #2188
• Check error code from cudaOccupancyMaxActiveBlocksPerMultiprocessor #2172
• Add non-member Kokkos::resize/realloc for DualView #2170
• Construct DualView without initialization #2046
• Expose is_assignable to determine if one view can be assigned to another #1936
• profiling label #1935
• team_broadcast of bool failed on CUDA backend #1908
• View static_extent #660
• Misleading Kokkos::Cuda::initialize ERROR message when compiled for wrong GPU architecture #1944
• Cryptic Error When Malloc Fails #2164
• Drop support for intermediate standards in CMake #2336

Fixed bugs:

• DualView sync_device with length zero creates cuda errors #2946
• building with nvcc and clang (or clang based XL) as host compiler: "Kokkos::atomic_fetch_min(volatile int *, int)" has already been defined #2903
• Cuda 9.1,10.1 debug builds failing due to -Werror=unused-parameter #2880
• clang -Werror: Kokkos_FixedBufferMemoryPool.hpp:140:28: error: unused parameter 'alloc_size' #2869
• intel/16.0.1, intel/17.0.1 nightly build failures with debugging enabled #2867
• intel/16.0.1 debug build errors #2863
• xl/16.1.1 with cpp14, openmp build, nightly test failures #2856
• Intel nightly test failures: team_vector #2852
• Kokkos Views with intmax/2<N<intmax can hang during construction #2850
• workgraph_fib test seg-faults with threads backend and hwloc #2797
• cuda.view_64bit test hangs on Power8+Kepler37 system - develop and 2.9.00 branches #2771
• device_type for Kokkos_Random ? #2693
• "More than one tag given" error in Experimental::require() #2608
• Segfault on Marvell from our finalization stack #2542

3.0.00 (2020-01-27)

Full Changelog

Implemented enhancements:

• BuildSystem: Standalone Modern CMake Support #2104
• StyleFormat: ClangFormat Style #2157
• Documentation: Document build system and CMake philosophy #2263
• BuildSystem: Add Alias with Namespace Kokkos:: to Interal Libraries #2530
• BuildSystem: Universal Kokkos find_package #2099
• BuildSystem: Dropping support for Kokkos_{DEVICES,OPTIONS,ARCH} in CMake #2329
• BuildSystem: Set Kokkos_DEVICES and Kokkos_ARCH variables in exported CMake configuration #2193
• BuildSystem: Drop support for CUDA 7 and CUDA 8 #2489
• BuildSystem: Drop CMake option SEPARATE_TESTS #2266
• BuildSystem: Support expt-relaxed-constexpr same as expt-extended-lambda #2411
• BuildSystem: Add Xnvlink to command line options allowed in nvcc_wrapper #2197
• BuildSystem: Install Kokkos config files and target files to lib/cmake/Kokkos #2162
• BuildSystem: nvcc_wrappers and c++ 14 #2035
• BuildSystem: Kokkos version major/version minor (Feature request) #1930
• BuildSystem: CMake namespaces (and other modern cmake cleanup) #1924
• BuildSystem: Remove capability to install Kokkos via GNU Makefiles #2332
• Documentation: Remove PDF ProgrammingGuide in Kokkos replace with link #2244
• View: Add Method to Resize View without Initialization #2048
• Vector: implement “insert” method for Kokkos_Vector (as a serial function on host) #2437

Fixed bugs:

• ParallelScan: Kokkos::parallel\scan fix race condition seen in inter-block fence #2681
• OffsetView: Kokkos::OffsetView missing constructor which takes pointer #2247
• OffsetView: Kokkos::OffsetView: allow offset=0 #2246
• DeepCopy: Missing DeepCopy instrumentation in Kokkos #2522
• nvcc_wrapper: --host-only fails with multiple -W* flags #2484
• nvcc_wrapper: taking first -std option is counterintuitive #2553
• Subview: Error taking subviews of views with static_extents of min rank #2448
• TeamPolicy: reducers with valuetypes without += broken on CUDA #2410
• Libs: Fix inconsistency of Kokkos library names in Kokkos and Trilinos #1902
• Complex: operator>> for complex<T> uses std::ostream, not std::istream #2313
• Macros: Restrict not honored for non-intel compilers #1922

2.9.00 (2019-06-24)

Full Changelog

Implemented enhancements:

• Capability: CUDA Streams #1723
• Capability: CUDA Stream support for parallel_reduce #2061
• Capability: Feature Request: TeamVectorRange #713
• Capability: Adding HPX backend #2080
• Capability: TaskScheduler to have multiple queues #565
• Capability: Support for additional reductions in ScatterView #1674
• Capability: Request: deep_copy within parallel regions #689
• Capability: Feature Request: create\_mirror\_view\_without\_initializing #1765
• View: Use SFINAE to restrict possible View type conversions #2127
• Deprecation: Deprecate ExecutionSpace::fence() as static function and make it non-static #2140
• Deprecation: Deprecate LayoutTileLeft #2122
• Macros: KOKKOS_RESTRICT defined for non-Intel compilers #2038

Fixed bugs:

• Cuda: TeamThreadRange loop count on device is passed by reference to host static constexpr #1733
• Cuda: Build error with relocatable device code with CUDA 10.1 GCC 7.3 #2134
• Cuda: cudaFuncSetCacheConfig is setting CachePreferShared too often #2066
• Cuda: TeamPolicy doesn't throw then created with non-viable vector length and also doesn't backscale to viable one #2020
• Cuda: cudaMemcpy error for large league sizes on V100 #1991
• Cuda: illegal warp sync in parallel_reduce by functor on Turing 75 #1958
• TeamThreadRange: Inconsistent results from TeamThreadRange reduction #1905
• Atomics: atomic_fetch_oper & atomic_oper_fetch don't build for complex<float> #1964
• Views: Kokkos randomread Views leak memory #2155
• ScatterView: LayoutLeft overload currently non-functional #2165
• KNL: With intel 17.2.174 illegal instruction in random number test #2078
• Bitset: Enable copy constructor on device #2094
• Examples: do not compile due to template deduction error (multi_fem) #1928

2.8.00 (2019-02-05)

Full Changelog

Implemented enhancements:

• Capability, Tests: C++14 support and testing #1914
• Capability: Add environment variables for all command line arguments #1798
• Capability: --kokkos-ndevices not working for Slurm #1920
• View: Undefined behavior when deep copying from and to an empty unmanaged view #1967
• BuildSystem: nvcc_wrapper should stop immediately if nvcc is not in PATH #1861

Fixed bugs:

• Cuda: Fix Volta Issues 1 Non-deterministic behavior on Volta, runs fine on Pascal #1949
• Cuda: Fix Volta Issues 2 CUDA Team Scan gives wrong values on Volta with -G compile flag #1942
• Cuda: illegal warp sync in parallel_reduce by functor on Turing 75 #1958
• Threads: Pthreads backend does not handle RangePolicy with offset correctly #1976
• Atomics: atomic_fetch_oper has no case for Kokkos::complex<double> or other 16-byte types #1951
• MDRangePolicy: Fix zero-length range #1948
• TeamThreadRange: TeamThreadRange MaxLoc reduce doesnt compile #1909

2.7.24 (2018-11-04)

Full Changelog

Implemented enhancements:

• DualView: Add non-templated functions for sync, need_sync, view, modify #1858
• DualView: Avoid needlessly allocates and initializes modify_host and modify_device flag views #1831
• DualView: Incorrect deduction of "not device type" #1659
• BuildSystem: Add KOKKOS_ENABLE_CXX14 and KOKKOS_ENABLE_CXX17 #1602
• BuildSystem: Installed kokkos_generated_settings.cmake contains build directories instead of install directories #1838
• BuildSystem: KOKKOS_ARCH: add ticks to printout of improper arch setting #1649
• BuildSystem: Make core/src/Makefile for Cuda use needed nvcc_wrapper #1296
• Build: Support PGI as host compiler for NVCC #1828
• Build: Many Warnings Fixed e.g.#1786
• Capability: OffsetView with non-zero begin index #567
• Capability: Reductions into device side view #1788
• Capability: Add max_size to Kokkos::Array #1760
• Capability: View Assignment: LayoutStride -> LayoutLeft and LayoutStride -> LayoutRight #1594
• Capability: Atomic function allow implicit conversion of update argument #1571
• Capability: Add team_size_max with tagged functors #663
• Capability: Fix allignment of views from Kokkos_ScratchSpace should use different alignment #1700
• Capabilitiy: create_mirror_view_and_copy for DynRankView #1651
• Capability: DeepCopy HBWSpace / HostSpace #548
• ROCm: support team vector scan #1645
• ROCm: Merge from rocm-hackathon2 #1636
• ROCm: Add ParallelScanWithTotal #1611
• ROCm: Implement MDRange in ROCm #1314
• ROCm: Implement Reducers for Nested Parallelism Levels #963
• ROCm: Add asynchronous deep copy #959
• Tests: Memory pool test seems to allocate 8GB #1830
• Tests: Add unit_test for team_broadcast #734

Fixed bugs:

• BuildSystem: Makefile.kokkos gets gcc-toolchain wrong if gcc is cached #1841
• BuildSystem: kokkos_generated_settings.cmake placement is inconsistent #1771
• BuildSystem: Invalid escape sequence . in kokkos_functions.cmake #1661
• BuildSystem: Problem in Kokkos generated cmake file #1770
• BuildSystem: invalid file names on windows #1671
• Tests: reducers min/max_loc test fails randomly due to multiple min values and thus multiple valid locations #1681
• Tests: cuda.scatterview unit test causes "Bus error" when force_uvm and enable_lambda are enabled #1852
• Tests: cuda.cxx11 unit test fails when force_uvm and enable_lambda are enabled #1850
• Tests: threads.reduce_device_view_range_policy failing with Cuda/8.0.44 and RDC #1836
• Build: compile error when compiling Kokkos with hwloc 2.0.1 (on OSX 10.12.6, with g++ 7.2.0) #1506
• Build: dual_view.view broken with UVM #1834
• Build: White cuda/9.2 + gcc/7.2 warnings triggering errors #1833
• Build: warning: enum constant in boolean context #1813
• Capability: Fix overly conservative max_team_size thingy #1808
• DynRankView: Ctors taking ViewAllocateWithoutInitializing broken #1783
• Cuda: Apollo cuda.team_broadcast test fail with clang-6.0 #1762
• Cuda: Clang spurious test failure in impl_view_accessible #1753
• Cuda: Kokkos::complex<double> atomic deadlocks with Clang 6 Cuda build with -O0 #1752
• Cuda: LayoutStride Test fails for UVM as default memory space #1688
• Cuda: Scan wrong values on Volta #1676
• Cuda: Kokkos::deep_copy error with CudaUVM and Kokkos::Serial spaces #1652
• Cuda: cudaErrorInvalidConfiguration with debug build #1647
• Cuda: parallel_for with TeamPolicy::team_size_recommended with launch bounds not working -- reported by Daniel Holladay #1283
• Cuda: Using KOKKOS_CLASS_LAMBDA in a class with Kokkos::Random_XorShift64_Pool member data #1696
• Long Build Times on Darwin #1721
• Capability: Typo in Kokkos_Sort.hpp - BinOp3D - wrong comparison #1720
• Buffer overflow in SharedAllocationRecord in Kokkos_HostSpace.cpp #1673
• Serial unit test failure #1632

2.7.00 (2018-05-24)

Full Changelog

Part of the Kokkos C++ Performance Portability Programming EcoSystem 2.7

Implemented enhancements:

• Deprecate team_size auto adjusting to maximal value possible #1618
• DynamicView - remove restrictions to std::is_trivial types and value_type is power of two #1586
• Kokkos::StaticCrsGraph does not propagate memory traits (e.g., Unmanaged) #1581
• Adding ETI for DeepCopy / ViewFill etc. #1578
• Deprecate all the left over KOKKOS_HAVE_ Macros and Kokkos_OldMacros.hpp #1572
• Error if Kokkos_ARCH set in CMake #1555
• Deprecate ExecSpace::initialize / ExecSpace::finalize #1532
• New API for TeamPolicy property setting #1531
• clang 6.0 + cuda debug out-of-memory test failure #1521
• Cuda UniqueToken interface not consistent with other backends #1505
• Move Reducers out of Experimental namespace #1494
• Provide scope guard for initialize/finalize #1479
• Check Kokkos::is_initialized in SharedAllocationRecord dtor #1465
• Remove static list of allocations #1464
• Makefiles: Support single compile/link line use case #1402
• ThreadVectorRange with a range #1400
• Exclusive scan + last value API #1358
• Install kokkos_generated_settings.cmake #1348
• Kokkos arrays (not views!) don't do bounds checking in debug mode #1342
• Expose round-robin GPU assignment outside of initialize(int, char**) #1318
• DynamicView misses use_count and label function #1298
• View constructor should check arguments #1286
• False Positive on Oversubscription Warning #1207
• Allow (require) execution space for 1st arg of VerifyExecutionCanAccessMemorySpace #1192
• ROCm: Add ROCmHostPinnedSpace #958
• power of two functions #656
• CUDA 8 has 64bit __shfl #361
• Add TriBITS/CMake configure information about node types #243

Fixed bugs:

• CUDA atomic_fetch_sub for doubles is hitting CAS instead of intrinsic #1624
• Bug: use of ballot on Volta #1612
• Kokkos::deep_copy memory access failures #1583
• g++ -std option doubly set for cmake project #1548
• ViewFill for 1D Views of larger 32bit entries fails #1541
• CUDA Volta another warpsync bug #1520
• triple_nested_parallelism fails with KOKKOS_DEBUG and CUDA #1513
• Jenkins errors in Kokkos_SharedAlloc.cpp with debug build #1511
• Kokkos::Sort out-of-bounds with empty bins #1504
• Get rid of deprecated functions inside Kokkos #1484
• get_work_partition casts int64_t to int, causing a seg fault #1481
• NVCC bug with __device__ on defaulted function #1470
• CMake example broken with CUDA backend #1468

2.6.00 (2018-03-07)

Full Changelog

Part of the Kokkos C++ Performance Portability Programming EcoSystem 2.6

Implemented enhancements:

• Support NVIDIA Volta microarchitecture #1466
• Kokkos - Define empty functions when profiling disabled #1424
• Don't use __constant__ cache for lock arrays, enable once per run update instead of once per call #1385
• task dag enhancement. #1354
• Cuda task team collectives and stack size #1353
• Replace View operator acceptance of more than rank integers with 'access' function #1333
• Interoperability: Do not shut down backend execution space runtimes upon calling finalize. #1305
• shmem_size for LayoutStride #1291
• Kokkos::resize performs poorly on 1D Views #1270
• stride() is inconsistent with dimension(), extent(), etc. #1214
• Kokkos::sort defaults to std::sort on host #1208
• DynamicView with host size grow #1206
• Unmanaged View with Anonymous Memory Space #1175
• Sort subset of Kokkos::DynamicView #1160
• MDRange policy doesn't support lambda reductions #1054
• Add ability to set hook on Kokkos::finalize #714
• Atomics with Serial Backend - Default should be Disable? #549
• KOKKOS_ENABLE_DEPRECATED_CODE #1359

Fixed bugs:

• cuda_internal_maximum_warp_count returns 8, but I believe it should return 16 for P100 #1269
• Cuda: level 1 scratch memory bug (reported by Stan Moore) #1434
• MDRangePolicy Reduction requires value_type typedef in Functor #1379
• Kokkos DeepCopy between empty views fails #1369
• Several issues with new CMake build infrastructure (reported by Eric Phipps) #1365
• deep_copy between rank-1 host/device views of differing layouts without UVM no longer works (reported by Eric Phipps) #1363
• Profiling can't be disabled in CMake, and a parallel_for is missing for tasks (reported by Kyungjoo Kim) #1349
• get_work_partition int overflow (reported by berryj5) #1327
• Kokkos::deep_copy must fence even if the two views are the same #1303
• CudaUVMSpace::allocate/deallocate must fence #1302
• ViewResize on CUDA fails in Debug because of too many resources requested #1299
• Cuda 9 and intrepid2 calls from Panzer. #1183
• Slowdown due to tracking_enabled() in 2.04.00 (found by Albany app) #1016
• Bounds checking fails with zero-span Views (reported by Stan Moore) #1411

2.5.00 (2017-12-15)

Full Changelog

Part of the Kokkos C++ Performance Portability Programming EcoSystem 2.5

Implemented enhancements:

• Provide Makefile.kokkos logic for CMake and TriBITS #878
• Add Scatter View #825
• Drop gcc 4.7 and intel 14 from supported compiler list #603
• Enable construction of unmanaged view using common_view_alloc_prop #1170
• Unused Function Warning with XL #1267
• Add memory pool parameter check #1218
• CUDA9: Fix warning for unsupported long double #1189
• CUDA9: fix warning on defaulted function marking #1188
• CUDA9: fix warnings for deprecated warp level functions #1187
• Add CUDA 9.0 nightly testing #1174
• {OMPI,MPICH}_CXX hack breaks nvcc_wrapper use case #1166
• KOKKOS_HAVE_CUDA_LAMBDA became KOKKOS_CUDA_USE_LAMBDA #1274

Fixed bugs:

• MinMax Reducer with tagged operator doesn't compile #1251
• Reducers for Tagged operators give wrong answer #1250
• Kokkos not Compatible with Big Endian Machines? #1235
• Parallel Scan hangs forever on BG/Q #1234
• Threads backend doesn't compile with Clang on OS X #1232
• $(shell date) needs quote #1264
• Unqualified parallel_for call conflicts with user-defined parallel_for #1219
• KokkosAlgorithms: CMake issue in unit tests #1212
• Intel 18 Error: "simd pragma has been deprecated" #1210
• Memory leak in Kokkos::initialize #1194
• CUDA9: compiler error with static assert template arguments #1190
• Kokkos::Serial::is_initialized returns always true #1184
• Triple nested parallelism still fails on bowman #1093
• OpenMP openmp.range on Develop Runs Forever on POWER7+ with RHEL7 and GCC4.8.5 #995
• Rendezvous performance at global scope #985

2.04.11 (2017-10-28)

Full Changelog

Implemented enhancements:

• Add Subview pattern. #648
• Add Kokkos "global" is_initialized #1060
• Add create_mirror_view_and_copy #1161
• Add KokkosConcepts SpaceAccessibility function #1092
• Option to Disable Initialize Warnings #1142
• Mature task-DAG capability #320
• Promote Work DAG from experimental #1126
• Implement new WorkGraph push/pop #1108
• Kokkos_ENABLE_Cuda_Lambda should default ON #1101
• Add multidimensional parallel for example and improve unit test #1064
• Fix ROCm: Performance tests not building #1038
• Make KOKKOS_ALIGN_SIZE a configure-time option #1004
• Make alignment consistent #809
• Improve subview construction on Cuda backend #615

Fixed bugs:

• Kokkos::vector fixes for application #1134
• DynamicView non-power of two value_type #1177
• Memory pool bug #1154
• Cuda launch bounds performance regression bug #1140
• Significant performance regression in LAMMPS after updating Kokkos #1139
• CUDA compile error #1128
• MDRangePolicy neg idx test failure in debug mode #1113
• subview construction on Cuda backend #615

2.04.04 (2017-09-11)

Full Changelog

Implemented enhancements:

• OpenMP partition: set number of threads on nested level #1082
• Add StaticCrsGraph row() method #1071
• Enhance Kokkos complex operator overloading #1052
• Tell Trilinos packages about host+device lambda #1019
• Function markup for defaulted class members #952
• Add deterministic random number generator #857

Fixed bugs:

• Fix reduction_identity<T>::max for floating point numbers #1048
• Fix MD iteration policy ignores lower bound on GPUs #1041
• (Experimental) HBWSpace Linking issues in KokkosKernels #1094
• (Experimental) ROCm: algorithms/unit_tests test_sort failing with segfault #1070

2.04.00 (2017-08-16)

Full Changelog

Implemented enhancements:

• Added ROCm backend to support AMD GPUs
• Kokkos::complex<T> behaves slightly differently from std::complex<T> #1011
• Kokkos::Experimental::Crs constructor arguments were in the wrong order #992
• Work graph construction ease-of-use (one lambda for count and fill) #991
• when_all returns pointer of futures (improved interface) #990
• Allow assignment of LayoutLeft to LayoutRight or vice versa for rank-0 Views #594
• Changed the meaning of Kokkos_ENABLE_CXX11_DISPATCH_LAMBDA #1035

Fixed bugs:

• memory pool default constructor does not properly set member variables. #1007

2.03.13 (2017-07-27)

Full Changelog

Implemented enhancements:

• Disallow enabling both OpenMP and Threads in the same executable #406
• Make Kokkos::OpenMP respect OMP environment even if hwloc is available #630
• Improve Atomics Performance on KNL/Broadwell where PREFETCHW/RFO is Available #898
• Kokkos::resize should test whether dimensions have changed before resizing #904
• Develop performance-regression/acceptance tests #737
• Make the deep_copy Profiling hook a start/end system #890
• Add deep_copy Profiling hook #843
• Append tag name to parallel construct name for Profiling #842
• Add view label to View bounds error message for CUDA backend #870
• Disable printing the loaded profiling library #824
• "Declared but never referenced" warnings #853
• Warnings about lock_address_cuda_space #852
• WorkGraph execution policy #771
• Simplify makefiles by guarding compilation with appropriate KOKKOS_ENABLE_### macros #716
• Cmake build: wrong include install directory #668
• Derived View type and allocation #566
• Fix Compiler warnings when compiling core unit tests for Cuda #214

Fixed bugs:

• Out-of-bounds read in Kokkos_Layout.hpp #975
• CudaClang: Fix failing test with Clang 4.0 #941
• Respawn when memory pool allocation fails (not available memory) #940
• Memory pool aborts on zero allocation request, returns NULL for < minimum #939
• Error with TaskScheduler query of underlying memory pool #917
• Profiling::*Callee static variables declared in header #863
• calling *Space::name() causes compile error #862
• bug in Profiling::deallocateData #860
• task_depend test failing, CUDA 8.0 + Pascal + RDC #829
• [develop branch] Standalone cmake issues #826
• Kokkos CUDA failes to compile with OMPI_CXX and MPICH_CXX wrappers #776
• Task Team reduction on Pascal #767
• CUDA stack overflow with TaskDAG test #758
• TeamVector test on Cuda #670
• Clang 4.0 Cuda Build broken again #560

2.03.05 (2017-05-27)

Full Changelog

Implemented enhancements:

• Harmonize Custom Reductions over nesting levels #802
• Prevent users directly including KokkosCore_config.h #815
• DualView aborts on concurrent host/device modify (in debug mode) #814
• Abort when running on a NVIDIA CC5.0 or higher architecture with code compiled for CC < 5.0 #813
• Add "name" function to ExecSpaces #806
• Allow null Future in task spawn dependences #795
• Add Unit Tests for Kokkos::complex #785
• Add pow function for Kokkos::complex #784
• Square root of a complex #729
• Command line processing of --threads argument prevents users from having any commandline arguments starting with --threads #760
• Protected deprecated API with appropriate macro #756
• Allow task scheduler memory pool to be used by tasks #747
• View bounds checking on host-side performance: constructing a std::string #723
• Add check for AppleClang as compiler distinct from check for Clang. #705
• Uninclude source files for specific configurations to prevent link warning. #701
• Add --small option to snapshot script #697
• CMake Standalone Support #674
• CMake build unit test and install #808
• CMake: Fix having kokkos as a subdirectory in a pure cmake project #629
• Tribits macro assumes build directory is in top level source directory #654
• Use bin/nvcc_wrapper, not config/nvcc_wrapper #562
• Allow MemoryPool::allocate() to be called from multiple threads per warp. #487
• Allow MemoryPool::allocate\(\) to be called from multiple threads per warp. #487
• Move OpenMP 4.5 OpenMPTarget backend into Develop #456
• Testing on ARM testbed #288

Fixed bugs:

• Fix label in OpenMP parallel_reduce verify_initialized #834
• TeamScratch Level 1 on Cuda hangs #820
• [bug] memory pool. #786
• Some Reduction Tests fail on Intel 18 with aggressive vectorization on #774
• Error copying dynamic view on copy of memory pool #773
• CUDA stack overflow with TaskDAG test #758
• ThreadVectorRange Customized Reduction Bug #739
• set_scratch_size overflows #726
• Get wrong results for compiler checks in Makefile on OS X. #706
• Fix check if multiple host architectures enabled. #702
• Threads Backend Does not Pass on Cray Compilers #609
• Rare bug in memory pool where allocation can finish on superblock in empty state #452
• LDFLAGS in core/unit_test/Makefile: potential "undefined reference" to pthread lib #148

2.03.00 (2017-04-25)

Full Changelog

Implemented enhancements:

• UnorderedMap: make it accept Devices or MemorySpaces #711
• sort to accept DynamicView and [begin,end) indices #691
• ENABLE Macros should only be used via #ifdef or #if defined #675
• Remove impl/Kokkos_Synchronic_* #666
• Turning off IVDEP for Intel 14. #638
• Using an installed Kokkos in a target application using CMake #633
• Create Kokkos Bill of Materials #632
• MDRangePolicy and tagged evaluators #547
• Add PGI support #289

Fixed bugs:

• Output from PerTeam fails #733
• Cuda: architecture flag not added to link line #688
• Getting large chunks of memory for a thread team in a universal way #664
• Kokkos RNG normal() function hangs for small seed value #655
• Kokkos Tests Errors on Shepard/HSW Builds #644

2.02.15 (2017-02-10)

Full Changelog

Implemented enhancements:

• Containers: Adding block partitioning to StaticCrsGraph #625
• Kokkos Make System can induce Errors on Cray Volta System #610
• OpenMP: error out if KOKKOS_HAVE_OPENMP is defined but not _OPENMP #605
• CMake: fix standalone build with tests #604
• Change README (that GitHub shows when opening Kokkos project page) to tell users how to submit PRs #597
• Add correctness testing for all operators of Atomic View #420
• Allow assignment of Views with compatible memory spaces #290
• Build only one version of Kokkos library for tests #213
• Clean out old KOKKOS_HAVE_CXX11 macros clauses #156
• Harmonize Macro names #150

Fixed bugs:

• Cray and PGI: Kokkos_Parallel_Reduce #634
• Kokkos Make System can induce Errors on Cray Volta System #610
• Normal() function random number generator doesn't give the expected distribution #592

2.02.07 (2016-12-16)

Full Changelog

Implemented enhancements:

• Add CMake option to enable Cuda Lambda support #589
• Add CMake option to enable Cuda RDC support #588
• Add Initial Intel Sky Lake Xeon-HPC Compiler Support to Kokkos Make System #584
• Building Tutorial Examples #582
• Internal way for using ThreadVectorRange without TeamHandle #574
• Testing: Add testing for uvm and rdc #571
• Profiling: Add Memory Tracing and Region Markers #557
• nvcc_wrapper not installed with Kokkos built with CUDA through CMake #543
• Improve DynRankView debug check #541
• Benchmarks: Add Gather benchmark #536
• Testing: add spot_check option to test_all_sandia #535
• Deprecate Kokkos::Impl::VerifyExecutionCanAccessMemorySpace #527
• Add AtomicAdd support for 64bit float for Pascal #522
• Add Restrict and Aligned memory trait #517
• Kokkos Tests are Not Run using Compiler Optimization #501
• Add support for clang 3.7 w/ openmp backend #393
• Provide an error throw class #79

Fixed bugs:

• Cuda UVM Allocation test broken with UVM as default space #586
• Bug (develop branch only): multiple tests are now failing when forcing uvm usage. #570
• Error in generate_makefile.sh for Kokkos when Compiler is Empty String/Fails #568
• XL 13.1.4 incorrect C++11 flag #553
• Improve DynRankView debug check #541
• Installing Library on MAC broken due to cp -u #539
• Intel Nightly Testing with Debug enabled fails #534

2.02.01 (2016-11-01)

Full Changelog

Implemented enhancements:

• Add Changelog generation to our process. #506

Fixed bugs:

• Test scratch_request fails in Serial with Debug enabled #520
• Bug In BoundsCheck for DynRankView #516

2.02.00 (2016-10-30)

Full Changelog

Implemented enhancements:

• Add PowerPC assembly for grabbing clock register in memory pool #511
• Add GCC 6.x support #508
• Test install and build against installed library #498
• Makefile.kokkos adds expt-extended-lambda to cuda build with clang #490
• Add top-level makefile option to just test kokkos-core unit-test #485
• Split and harmonize Object Files of Core UnitTests to increase build parallelism #484
• LayoutLeft to LayoutLeft subview for 3D and 4D views #473
• Add official Cuda 8.0 support #468
• Allow C++1Z Flag for Class Lambda capture #465
• Add Clang 4.0+ compilation of Cuda code #455
• Possible Issue with Intel 17.0.098 and GCC 6.1.0 in Develop Branch #445
• Add name of view to "View bounds error" #432
• Move Sort Binning Operators into Kokkos namespace #421
• TaskPolicy - generate error when attempt to use uninitialized #396
• Import WithoutInitializing and AllowPadding into Kokkos namespace #325
• TeamThreadRange requires begin, end to be the same type #305
• CudaUVMSpace should track # allocations, due to CUDA limit on # UVM allocations #300
• Remove old View and its infrastructure #259

Fixed bugs:

• Bug in TestCuda_Other.cpp: most likely assembly inserted into Device code #515
• Cuda Compute Capability check of GPU is outdated #509
• multi_scratch test with hwloc and pthreads seg-faults. #504
• generate_makefile.bash: "make install" is broken #503
• make clean in Out of Source Build/Tests Does Not Work Correctly #502
• Makefiles for test and examples have issues in Cuda when CXX is not explicitly specified #497
• Dispatch lambda test directly inside GTEST macro doesn't work with nvcc #491
• UnitTests with HWLOC enabled fail if run with mpirun bound to a single core #489
• Failing Reducer Test on Mac with Pthreads #479
• make test Dumps Error with Clang Not Found #471
• OpenMP TeamPolicy member broadcast not using correct volatile shared variable #424
• TaskPolicy - generate error when attempt to use uninitialized #396
• New task policy implementation is pulling in old experimental code. #372
• MemoryPool unit test hangs on Power8 with GCC 6.1.0 #298

2.01.10 (2016-09-27)

Full Changelog

Implemented enhancements:

• Enable Profiling by default in Tribits build #438
• parallel_reduce(0), parallel_scan(0) unit tests #436
• data()==NULL after realloc with LayoutStride #351
• Fix tutorials to track new Kokkos::View #323
• Rename team policy set_scratch_size. #195

Fixed bugs:

• Possible Issue with Intel 17.0.098 and GCC 6.1.0 in Develop Branch #445
• Makefile spits syntax error #435
• Kokkos::sort fails for view with all the same values #422
• Generic Reducers: can't accept inline constructed reducer #404
• data\(\)==NULL after realloc with LayoutStride #351
• const subview of const view with compile time dimensions on Cuda backend #310
• Kokkos (in Trilinos) Causes Internal Compiler Error on CUDA 8.0.21-EA on POWER8 #307
• Core Oversubscription Detection Broken? #159

2.01.06 (2016-09-02)

Full Changelog

Implemented enhancements:

• Add "standard" reducers for lambda-supportable customized reduce #411
• TaskPolicy - single thread back-end execution #390
• Kokkos master clone tag #387
• Query memory requirements from task policy #378
• Output order of test_atomic.cpp is confusing #373
• Missing testing for atomics #341
• Feature request for Kokkos to provide Kokkos::atomic_fetch_max and atomic_fetch_min #336
• TaskPolicy<Cuda> performance requires teams mapped to warps #218

Fixed bugs:

• Reduce with Teams broken for custom initialize #407
• Failing Kokkos build on Debian #402
• Failing Tests on NVIDIA Pascal GPUs #398
• Algorithms: fill_random assumes dimensions fit in unsigned int #389
• Kokkos::subview with RandomAccess Memory Trait #385
• Build warning (signed / unsigned comparison) in Cuda implementation #365
• wrong results for a parallel_reduce with CUDA8 / Maxwell50 #352
• Hierarchical parallelism - 3 level unit test #344
• Can I allocate a View w/ both WithoutInitializing & AllowPadding? #324
• subview View layout determination #309
• Unit tests with Cuda - Maxwell #196

2.01.00 (2016-07-21)

Full Changelog

Implemented enhancements:

• Edit ViewMapping so assigning Views with the same custom layout compiles when const casting #327
• DynRankView: Performance improvement for operator() #321
• Interoperability between static and dynamic rank views #295
• subview member function ? #280
• Inter-operatibility between View and DynRankView. #245
• (Trilinos) build warning in atomic_assign, with Kokkos::complex #177
• View<>::shmem_size should runtime check for number of arguments equal to rank #176
• Custom reduction join via lambda argument #99
• DynRankView with 0 dimensions passed in at construction #293
• Inject view_alloc and friends into Kokkos namespace #292
• Less restrictive TeamPolicy reduction on Cuda #286
• deep_copy using remap with source execution space #267
• Suggestion: Enable opt-in L1 caching via nvcc-wrapper #261
• More flexible create_mirror functions #260
• Rename View::memory_span to View::required_allocation_size #256
• Use of subviews and views with compile-time dimensions #237
• Use of subviews and views with compile-time dimensions #237
• Kokkos::Timer #234
• Fence CudaUVMSpace allocations #230
• View::operator() accept std::is_integral and std::is_enum #227
• Allocating zero size View #216
• Thread scalable memory pool #212
• Add a way to disable memory leak output #194
• Kokkos exec space init should init Kokkos profiling #192
• Runtime rank wrapper for View #189
• Profiling Interface #158
• Fix View assignment (of managed to unmanaged) #153
• Add unit test for assignment of managed View to unmanaged View #152
• Check for oversubscription of threads with MPI in Kokkos::initialize #149
• Dynamic resizeable 1dimensional view #143
• Develop TaskPolicy for CUDA #142
• New View : Test Compilation Downstream #138
• New View Implementation #135
• Add variant of subview that lets users add traits #134
• NVCC-WRAPPER: Add --host-only flag #121
• Address gtest issue with TriBITS Kokkos build outside of Trilinos #117
• Make tests pass with -expt-extended-lambda on CUDA #108
• Dynamic scheduling for parallel_for and parallel_reduce #106
• Runtime or compile time error when reduce functor's join is not properly specified as const member function or with volatile arguments #105
• Error out when the number of threads is modified after kokkos is initialized #104
• Porting to POWER and remove assumption of X86 default #103
• Dynamic scheduling option for RangePolicy #100
• SharedMemory Support for Lambdas #81
• Recommended TeamSize for Lambdas #80
• Add Aggressive Vectorization Compilation mode #72
• Dynamic scheduling team execution policy #53
• UVM allocations in multi-GPU systems #50
• Synchronic in Kokkos::Impl #44
• index and dimension types in for loops #28
• Subview assign of 1D Strided with stride 1 to LayoutLeft/Right #1

Fixed bugs:

• misspelled variable name in Kokkos_Atomic_Fetch + missing unit tests #340
• seg fault Kokkos::Impl::CudaInternal::print_configuration #338
• Clang compiler error with named parallel_reduce, tags, and TeamPolicy. #335
• Shared Memory Allocation Error at parallel_reduce #311
• DynRankView: Fix resize and realloc #303
• Scratch memory and dynamic scheduling #279
• MemoryPool infinite loop when out of memory #312
• Kokkos DynRankView changes break Sacado and Panzer #299
• MemoryPool fails to compile on non-cuda non-x86 #297
• Random Number Generator Fix #296
• View template parameter ordering Bug #282
• Serial task policy broken. #281
• deep_copy with LayoutStride should not memcpy #262
• DualView::need_sync should be a const method #248
• Arbitrary-sized atomics on GPUs broken; loop forever #238
• boolean reduction value_type changes answer #225
• Custom init() function for parallel_reduce with array value_type #210
• unit_test Makefile is Broken - Recursively Calls itself until Machine Apocalypse. #202
• nvcc_wrapper Does Not Support -Xcompiler <compiler option> #198
• Kokkos exec space init should init Kokkos profiling #192
• Kokkos Threads Backend impl_shared_alloc Broken on Intel 16.1 (Shepard Haswell) #186
• pthread back end hangs if used uninitialized #182
• parallel_reduce of size 0, not calling init/join #175
• Bug in Threads with OpenMP enabled #173
• KokkosExp_SharedAlloc, m_team_work_index inaccessible #166
• 128-bit CAS without Assembly Broken? #161
• fatal error: Cuda/Kokkos_Cuda_abort.hpp: No such file or directory #157
• Power8: Fix OpenMP backend #139
• Data race in Kokkos OpenMP initialization #131
• parallel_launch_local_memory and cuda 7.5 #125
• Resize can fail with Cuda due to asynchronous dispatch #119
• Qthread taskpolicy initialization bug. #92
• Windows: sys/mman.h #89
• Windows: atomic_fetch_sub() #88
• Windows: snprintf #87
• Parallel_Reduce with TeamPolicy and league size of 0 returns garbage #85
• Throw with Cuda when using (2D) team_policy parallel_reduce with less than a warp size #76
• Scalar views don't work with Kokkos::Atomic memory trait #69
• Reduce the number of threads per team for Cuda #63
• Named Kernels fail for reductions with CUDA #60
• Kokkos View dimension_() for long returning unsigned int #20
• atomic test hangs with LLVM #6
• OpenMP Test should set omp_set_num_threads to 1 #4

Closed issues:

• develop branch broken with CUDA 8 and --expt-extended-lambda #354
• --arch=KNL with Intel 2016 build failure #349
• Error building with Cuda when passing -DKOKKOS_CUDA_USE_LAMBDA to generate_makefile.bash #343
• Can I safely use int indices in a 2-D View with capacity > 2B? #318
• Kokkos::ViewAllocateWithoutInitializing is not working #317
• Intel build on Mac OS X #277
• deleted #271
• Broken Mira build #268
• 32-bit build #246
• parallel_reduce with RDC crashes linker #232
• build of Kokkos_Sparse_MV_impl_spmv_Serial.cpp.o fails if you use nvcc and have cuda disabled #209
• Kokkos Serial execution space is not tested with TeamPolicy. #207
• Unit test failure on Hansen KokkosCore_UnitTest_Cuda_MPI_1 #200
• nvcc compiler warning: calling a __host__ function from a __host__ __device__ function is not allowed #180
• Intel 15 build error with defaulted "move" operators #171
• missing libkokkos.a during Trilinos 12.4.2 build, yet other libkokkos*.a libs are there #165
• Tie atomic updates to execution space or even to thread team? (speculation) #144
• New View: Compiletime/size Test #137
• New View : Performance Test #136
• Signed/unsigned comparison warning in CUDA parallel #130
• Kokkos::complex: Need op* w/ std::complex & real #126
• Use uintptr_t for casting pointers #110
• Default thread mapping behavior between P and Q threads. #91
• Windows: Atomic_Fetch_Exchange() return type #90
• Synchronic unit test is way too long #84
• nvcc_wrapper -> $(NVCC_WRAPPER) #42
• Check compiler version and print helpful message #39
• Kokkos shared memory on Cuda uses a lot of registers #31
• Can not pass unit test cuda.space without a GT 720 #25
• Makefile.kokkos lacks bounds checking option that CMake has #24
• Kokkos can not complete unit tests with CUDA UVM enabled #23
• Simplify teams + shared memory histogram example to remove vectorization #21
• Kokkos needs to rever to ${PROJECT_NAME}_ENABLE_CXX11 not Trilinos_ENABLE_CXX11 #17
• Kokkos Base Makefile adds AVX to KNC Build #16
• MS Visual Studio 2013 Build Errors #9
• subview(X, ALL(), j) for 2-D LayoutRight View X: should it view a column? #5

End_C++98 (2015-04-15)

* This Change Log was automatically generated by github_changelog_generator