Skip to content

CHANGELOG for 5.2 #8944

Closed
Closed

Description

@dalg24

This is the changelog for the upcoming 5.2 release

Features:

  • Rank<1> MDRangePolicy support #9028
  • Introduce generic parallel_for with RangePolicy taking a team handle in hierarchical parallelism - enables writing generic functions with a nested parallel_for that can be called from host or device #8367
  • create_mirror_view_and_copy one argument convenience overload #9185
  • Add new Random Number generator type SFC64 #9017
  • Support using mdspan style template arguments for View thus enabling explicit specification of index_type #9185

Backend and Architecture Enhancements:

CUDA:

  • Support more than 48kB of scratch memory request for level 0 (shared memory) #9012
  • Enable C++23 support with NVCC (available since CUDA 13.3.0) #9264

HIP:

  • Improve performance of load/store atomic by leveraging HIP compiler built-ins instead of generating them via CAS #9040
  • Add AMD GFX1151 (Strix Halo / Radeon 8060S) architecture support #9179
  • Add AMD Radeon 860M / RDNA3.5 / gfx1152 architecture support #9242
  • Add AMD GFX1101 (Radeon RX 7800 XT, RX 7700 XT, RX 7700) architecture support #9230
  • Improve performance of deep_copy(v, 0) on MI300A #9125

SYCL:

  • Fix Windows build issues #9043

OpenACC:

  • Fix OpenACC parallel_scan chunk boundary race #9078

NextSilicon:

  • Add NextSilicon execution space and NextSiliconSharedSpace memory space #8998

HPX:

Threads:

OpenMP:

Serial:

General Enhancements

  • Add index_type (signed) to execution space instance classes and harmonize with size_type (unsigned) across different backends #7441
  • Add noexcept annotations to move constructors and assignment operators of execution spaces #8990
  • Add std::uint32_t support in Kokkos SIMD Neon backend #8942
  • Import math constants from the standard library #9016
  • Increase max level 1 team scratch size from 20MB to 80MB per team #9084
  • core(graph): rename interoperability graph access with <backend>_ prefix #9159
  • core(graph): node kind enum #9170
  • Add create_mirror_view_and_copy(a_view) overload #9185
  • Sepperated out (limited) iterator support from algorithms and moved to core #6684
  • Check extent preconditions on Views in std::algorithms #6811
  • Make the flag argument in simd constructors and load/store functions optional aligning with C++26 #9211 #9275
  • Use C++11 style attributes in definition of KOKKOS_FORCEINLINE_FUNCTION and add MSVC support #9229
  • Refactored nested for loops within host tiles to enhance readability and maintainability #9209
  • Introduce BadAlloc exception thrown by out of memory allocation requests #9260
  • MDRangePolicy: skip grid-stride loop when not needed (CUDA/HIP/SYCL) #9142 #9250
  • Introduce KOKKOS_FORCEINLINE_[CLASS_]LAMBDA macros #9239
  • Ensure the C++20 module kokkos.core can be used with latest MSVC and gcc-16. #9148, #9149.

Build System Changes

  • Build system prints output of compilation check #8986
  • nvcc_wrapper: Handle response files. #9175

Incompatibilities (i.e. breaking changes)

  • Forbid passing identical view arguments to deep_copy() #9033

Deprecations

  • Deprecate [static_]extent(r) calls with r >= rank() to align with mdspan preconditions #9076
  • Remove code guarded by KOKKOS_ENABLE_DEPRECATED_CODE_4 since Kokkos 4.2 and before #8957
  • Remove code guarded by KOKKOS_ENABLE_DEPRECATED_CODE_4 since Kokkos 4.4 and before #8966
  • Remove code guarded by KOKKOS_ENABLE_DEPRECATED_CODE_4 since Kokkos 4.6 and before #8980 #9009
  • (Partial) Remove code guarded by KOKKOS_ENABLE_DEPRECATED_CODE_4 in Kokkos 4.7 #9002
  • (Partial) Remove code guarded by KOKKOS_ENABLE_DEPRECATED_CODE_4 in Kokkos 5.0 #9013
  • Remove deprecated code path disabling enforcement of RangePolicy preconditions #9007
  • Remove miscellaneous deprecated code #9005
  • Remove untracked deprecated code 4 #9029
  • Remove all SIMD deprecated code #9008
  • Remove (deprecated) CUDA_UVM macro and option #8968
  • Remove deprecated CUDA_LAMBDA and CUDA_LDG_INTRINSIC macros and options #8960
  • Remove (deprecated) DEBUG_DUALVIEW_MODIFY_CHECK macro and option #9010
  • Remove SpaceAccessibility::deepcopy #9018
  • Deprecate Experimental HIP and SYCL symbols #9232
  • Deprecate extents for dimensions larger than the rank #9075

Bug Fixes

  • Fix View::extent_int(r) to behave as View::extent(r) with r>=rank #9072
  • Fix failures when using combined reducers with 8 or 16 bit datatypes on CUDA and HIP #8989
  • Only use __atomic_max_fetch with floating point types from LLVM 22.1 on #8991
  • Fix simd arithmetic operators calling host-only functions on device #9004
  • Fix compile failures on AppleClang 14 and Clang 14 and 15 when using simd #9065
  • Fix error message at compile time when index type is specified twice as template argument to the execution policy #9085
  • Fix an ambiguous call compile error when calling the masked version of simd reduce with default arguments #9079
  • Add bounds check to prevent OOB access in tuner dimension indexing #9090
  • Atomics: fix warning with lock free fetch op on pointers #9086 #9132
  • Fix RangePolicy CUDA/HIP performance regression with static batch size 1 introduced in Kokkos 5.0 #9123, #9124
  • Fix bug in compiler detection for RISC-V preventing use of clang++ #9155
  • Fix segfaults when doing unaligned simd stores on AVX512 #9184
  • Fix signed integer overflow UB in random generators’ range functions #9216
  • Fixing memory error in make_candidate_set tuning function #7875
  • Fix thread-safety for parallel_scan #9300
  • Fix ambiguous overload in Kokkos::resize with ExecutionSpace #9128
  • Fix subview construction between managed and unmanaged Views #8967

Activity

  1. pinned this issue on Mar 20, 2026
  2. daxmawal commented on Apr 30, 2026

    @daxmawal
  3. unpinned this issue on Jul 8, 2026
  4. rppawlo commented on Jul 9, 2026

    @rppawlo
    Contributor

    Did #9165 not make it into this release?

  5. dalg24 commented on Jul 10, 2026

    @dalg24
    MemberAuthor

    Did #9165 not make it into this release?

    It is in the RC branch. I could find a thread on the developer slack channel asking why it was not on the change log and Christian answered it did not need to appear in it.

  6. rppawlo commented on Jul 10, 2026

    @rppawlo
    Contributor

    Thanks! I figured you just put every PR into the changelog. Didn't realize the list was selective.

  7. dalg24 commented on Jul 10, 2026

    @dalg24
    MemberAuthor

    Thanks! I figured you just put every PR into the changelog. Didn't realize the list was selective.

    We curate the changelog. If we listed them all, there would be several hundred entries, with many being irrelevant for the end user.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions