[EPIC] Breaking changes for CCCL 3.0

[jrhemstad]

This issue is to keep track of breaking changes we'd like to make in the next major version release.

Removals

• Remove cub::ArrayWrapper #3665
• thrust global_workarounds.h
• thrust THRUST_HOST_BACKEND
• Drop CUB iterators existing in Thrust
• Drop thrust future, event and async
• Remove pre-C++17 code paths #3595
• Remove Agent* alias templates for the respective detail::<algorithm name>::Agent* #3467
• Drop support for VS 2017 in code #3375
• Drop CTK 11.x support in code #3284
• Remove any LEGACY_PTX_ARCH parameters #3552
• Dropcub::DeviceSpmv in favor of cuSPARSE #3070
• Internalize cuda/detail/core/* #3505
• Remove all code paths and policies below SM50 #3431
• Drop thrust::null_type #3508
• Move CUB kernel entry points to a detail namespace #3407
• Drop thrust universal iterator categories #3474
• Require C++17 for compiling libcu++
• Drop deprecated features from CUB util_ptx #3430
• Remove CDPv1 Support #3341
• Drop deprecated ::result_type etc. from Thrust function objects (and alias them to the libcu++ ones) #3271
• Drop cub::DivideAndRoundUp #3347
• Drop thrust::void_t #3362
• Drop cub::ValueCache #3346
• Drop all deprecated CUB APIs accepting a debug_synchronous parameter #3329
• Drop CI runs for CTK 11.x #3249
• Remove clang <14 from CI #3184
• Require C++17 for compiling Thrust and CUB #3252
• Drop Thrust's deprecated compiler macros #3301
• Drop CUB's util_compiler.cuh #3302
• Drop CUB's util_compiler.cuh for real #3340
• Drop CUB_RUNTIME_ENABLED and __THRUST_HAS_CUDART__ #3305
• Drop Thrust's compiler_fence.h #3300
• Drop Thrust legacy arch macros #3298
• Drop support for Intel ICC in code #3278
• Drop VS 2017 from CI #3286
• Drop CUB legacy macros in util_arch.cuh #3256
• Drop cub::Mutex #3250
• Drop Intel ICC from CI #3276
• Drop thrust::[unary|binary]_function #3273
• Drop thrust unary_traits and binary_traits #3259
• Drop thrust not1 and not2 #3263
• Drop support for GCC6 #3267
• Remove deprecated cub::min and thrust::remove_cvref #3450
• Remove cub::max

Deprecations

• Deprecate cub::RegBoundScaling and cub::MemBoundScaling #3663
• Probably cub::AliasTemporaries
• Internalize cub::PolicyWrapper
• Internalize cub::KernelConfig
• thrust::[make_][integer|index]_sequence in favor of libcu++
• Maybe the entire error code framework thrust/thrust/system/detail/error_*
• Maybe thrust::system::cuda::errc
• thrust::mr::validator (move to detail)
• thrust simple_defines.h
• thrust/detail/preprocessor.h and THRUST_PP* macros
• thrust::tuple_for_each
• parts of thrust memory_algorithms.h
• thrust event_error.h
• thrust::has_difference_type in advance.inl (move to detail)
• Replace cub::Int2Type by cuda::std::integral_constant and deprecate the former #3640
• Deprecate cub::FpLimits in favor of cuda::std::numeric_limits #3635
• Deprecate thrust iterator traversal #3562
• Deprecate DeviceSegmentedReduce::Arg{Min,Max} interface that returns result as cub::KeyValuePair #3482
• Replace parts of cub::Traits by numeric_limits and deprecate those #3381
• Replace use of thrust::identity by cuda::std::identity #2762
• Internalize triple_chevron #3648
• Internalize cuda/detail/core/* #3505
• AgentSegmentFixup, which was only used by cub::DeviceSpmv #3590
• Replace uses of CUB_IS_INT128_ENABLED by a CCCL macro and deprecate it #3403
• Deprecate thrust logical meta functions #3538
• Extend cuda::std::numeric_limits for __half, __nv_bfloat16, __nv_fp8_e4m3, __nv_fp8_e5m2 #3349
• THRUST_TUNING_ARCH Deprecate macros from cuda/detail/core/util.h #3504
• Deprecate CUB iterators which exist in Thrust or libcu++ #3261
• Deprecate thrust macros from type_deduction.h #3501
• Deprecate thrust event, future and more #3457
• Deprecate and replace thrust::cuda_cub iterators #3422
• Move CUB tunings and policy hubs to a detail namespace #3405
• Move all entities in cub/[block|warp]/specializations into a detail namespace #3409
• Deprecate a few CUB macros #3456
• Move CUB agents to internal namespace #3303
• Replace and deprecate thrust::cuda_cub::terminate #3421
• Deprecate and replace cub::min/max by cuda::std::min/max #3404
• Redefine and deprecate thrust::remove_cvref #3394
• Deprecate thrust::optional #3306
• Deprecate thrust::numeric_limits #3366
• Review and deprecate features from CUB util_ptx.cuh #3289
• Deprecate thrust::async #100
• Deprecate thrust::null_type #3367
• Deprecate cub::DeviceSpmv #896
• Deprecate cub::IterateThreadStore #3337
• Deprecate cub::Swap #3333
• Deprecate Thrust's cpp_compatibility.h macros #3299
• [FEA]: Deprecate GridBarrier and GridBarrierLifetime #1389

We should backport those to the 2.8 branch.

API breaking changes

• Make thrust::discard_iterator's value_type void #3399
• [FEA]: libcu++ iterator_traits should pick up std iterator_traits specializations #1930
• Swap order of template parameters of DispatchBatchMemcpy and DispatchReduce so policy hub is last
• Replace boolean parameters of CUB dispatchers by enumerations #3352
• Assert offset type in DispatchScan to be unsigned
• Move all CUB dispatchers into internal namespaces #3094
• Thrust universal_host_pinned_memory_resource's pointer type should probably be a universal_ptr #2659
• Fixes value type of thrust::tabulate_output_iterator #3573
• [FEA]: ceil_div should return the resulting type of its operation #2845
• [BUG]: thrust::transform_iterator<F, ...> is broken if F is not copy_assignable #2393

ABI breaking changes

• Deprecate and drop old libcu++ ABI namespaces #3424

Behavioral changes

• Unify accumulator type handling across CUB/Thrust
• Unify initial value handling across CUB/Thrust
• Remove multiplication and division nan handling <complex> Enable users to disable normalization for division too. libcudacxx#474 (review)

Cleanup

• Replace implementations in thrust/thrust/system/detail/sequential/* by libcu++ <algorithm>
• Replace THRUST_UNUSED_VAR by some CCCL macro
• Refactor and simplify Thrust iterators
• Review and deprecate macros from thrust/thrust/system/cuda/config.h
• Fix transform_iterator<identity> and drop result_of_adaptable_function #3652
• [FEA]: Refactor user-facing Thrust types to use libcu++  #34
• Specialize relevant cuda::(std::) types for __half/bfloat16/fp8 #525
• Replace mdspan implementation with one derived from the newer LLVM implementation #3353