cuda-cccl 0.3.2__cp313-cp313-manylinux_2_24_aarch64.whl → 0.3.4__cp313-cp313-manylinux_2_24_aarch64.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of cuda-cccl might be problematic. Click here for more details.
- cuda/cccl/headers/include/cub/agent/agent_adjacent_difference.cuh +12 -38
- cuda/cccl/headers/include/cub/agent/agent_batch_memcpy.cuh +16 -40
- cuda/cccl/headers/include/cub/agent/agent_for.cuh +2 -28
- cuda/cccl/headers/include/cub/agent/agent_histogram.cuh +24 -56
- cuda/cccl/headers/include/cub/agent/agent_merge_sort.cuh +12 -38
- cuda/cccl/headers/include/cub/agent/agent_radix_sort_downsweep.cuh +31 -56
- cuda/cccl/headers/include/cub/agent/agent_radix_sort_histogram.cuh +31 -35
- cuda/cccl/headers/include/cub/agent/agent_radix_sort_onesweep.cuh +47 -48
- cuda/cccl/headers/include/cub/agent/agent_radix_sort_upsweep.cuh +39 -42
- cuda/cccl/headers/include/cub/agent/agent_reduce.cuh +33 -60
- cuda/cccl/headers/include/cub/agent/agent_reduce_by_key.cuh +18 -44
- cuda/cccl/headers/include/cub/agent/agent_rle.cuh +26 -55
- cuda/cccl/headers/include/cub/agent/agent_scan.cuh +22 -49
- cuda/cccl/headers/include/cub/agent/agent_scan_by_key.cuh +15 -41
- cuda/cccl/headers/include/cub/agent/agent_segmented_radix_sort.cuh +9 -35
- cuda/cccl/headers/include/cub/agent/agent_select_if.cuh +20 -49
- cuda/cccl/headers/include/cub/agent/agent_sub_warp_merge_sort.cuh +14 -40
- cuda/cccl/headers/include/cub/agent/agent_three_way_partition.cuh +18 -40
- cuda/cccl/headers/include/cub/agent/agent_topk.cuh +0 -2
- cuda/cccl/headers/include/cub/agent/agent_unique_by_key.cuh +20 -46
- cuda/cccl/headers/include/cub/agent/single_pass_scan_operators.cuh +3 -28
- cuda/cccl/headers/include/cub/block/block_adjacent_difference.cuh +7 -31
- cuda/cccl/headers/include/cub/block/block_discontinuity.cuh +10 -34
- cuda/cccl/headers/include/cub/block/block_exchange.cuh +120 -154
- cuda/cccl/headers/include/cub/block/block_histogram.cuh +28 -52
- cuda/cccl/headers/include/cub/block/block_load.cuh +124 -146
- cuda/cccl/headers/include/cub/block/block_load_to_shared.cuh +0 -16
- cuda/cccl/headers/include/cub/block/block_merge_sort.cuh +58 -87
- cuda/cccl/headers/include/cub/block/block_radix_rank.cuh +81 -100
- cuda/cccl/headers/include/cub/block/block_radix_sort.cuh +92 -156
- cuda/cccl/headers/include/cub/block/block_raking_layout.cuh +8 -32
- cuda/cccl/headers/include/cub/block/block_reduce.cuh +21 -46
- cuda/cccl/headers/include/cub/block/block_run_length_decode.cuh +51 -79
- cuda/cccl/headers/include/cub/block/block_scan.cuh +94 -401
- cuda/cccl/headers/include/cub/block/block_shuffle.cuh +10 -34
- cuda/cccl/headers/include/cub/block/block_store.cuh +73 -97
- cuda/cccl/headers/include/cub/block/radix_rank_sort_operations.cuh +2 -29
- cuda/cccl/headers/include/cub/block/specializations/block_histogram_atomic.cuh +5 -29
- cuda/cccl/headers/include/cub/block/specializations/block_histogram_sort.cuh +25 -49
- cuda/cccl/headers/include/cub/block/specializations/block_reduce_raking.cuh +12 -34
- cuda/cccl/headers/include/cub/block/specializations/block_reduce_raking_commutative_only.cuh +10 -34
- cuda/cccl/headers/include/cub/block/specializations/block_reduce_warp_reductions.cuh +3 -27
- cuda/cccl/headers/include/cub/block/specializations/block_scan_raking.cuh +12 -36
- cuda/cccl/headers/include/cub/block/specializations/block_scan_warp_scans.cuh +9 -33
- cuda/cccl/headers/include/cub/config.cuh +2 -26
- cuda/cccl/headers/include/cub/cub.cuh +3 -27
- cuda/cccl/headers/include/cub/detail/array_utils.cuh +2 -26
- cuda/cccl/headers/include/cub/detail/choose_offset.cuh +2 -28
- cuda/cccl/headers/include/cub/detail/detect_cuda_runtime.cuh +3 -27
- cuda/cccl/headers/include/cub/detail/device_double_buffer.cuh +0 -2
- cuda/cccl/headers/include/cub/detail/device_memory_resource.cuh +1 -3
- cuda/cccl/headers/include/cub/detail/fast_modulo_division.cuh +2 -28
- cuda/cccl/headers/include/cub/detail/integer_utils.cuh +0 -2
- cuda/cccl/headers/include/cub/detail/launcher/cuda_driver.cuh +0 -2
- cuda/cccl/headers/include/cub/detail/launcher/cuda_runtime.cuh +0 -2
- cuda/cccl/headers/include/cub/detail/mdspan_utils.cuh +0 -2
- cuda/cccl/headers/include/cub/detail/ptx-json/README.md +7 -12
- cuda/cccl/headers/include/cub/detail/ptx-json/array.h +6 -33
- cuda/cccl/headers/include/cub/detail/ptx-json/json.h +13 -36
- cuda/cccl/headers/include/cub/detail/ptx-json/object.h +9 -38
- cuda/cccl/headers/include/cub/detail/ptx-json/string.h +58 -32
- cuda/cccl/headers/include/cub/detail/ptx-json/value.h +51 -51
- cuda/cccl/headers/include/cub/detail/ptx-json-parser.h +7 -31
- cuda/cccl/headers/include/cub/detail/rfa.cuh +2 -27
- cuda/cccl/headers/include/cub/detail/strong_load.cuh +3 -29
- cuda/cccl/headers/include/cub/detail/strong_store.cuh +3 -29
- cuda/cccl/headers/include/cub/detail/temporary_storage.cuh +2 -9
- cuda/cccl/headers/include/cub/detail/type_traits.cuh +0 -2
- cuda/cccl/headers/include/cub/detail/uninitialized_copy.cuh +6 -31
- cuda/cccl/headers/include/cub/detail/unsafe_bitcast.cuh +2 -25
- cuda/cccl/headers/include/cub/device/device_adjacent_difference.cuh +2 -26
- cuda/cccl/headers/include/cub/device/device_for.cuh +3 -5
- cuda/cccl/headers/include/cub/device/device_histogram.cuh +3 -27
- cuda/cccl/headers/include/cub/device/device_memcpy.cuh +2 -26
- cuda/cccl/headers/include/cub/device/device_merge_sort.cuh +2 -26
- cuda/cccl/headers/include/cub/device/device_partition.cuh +3 -27
- cuda/cccl/headers/include/cub/device/device_radix_sort.cuh +3 -27
- cuda/cccl/headers/include/cub/device/device_reduce.cuh +10 -31
- cuda/cccl/headers/include/cub/device/device_run_length_encode.cuh +3 -27
- cuda/cccl/headers/include/cub/device/device_scan.cuh +16 -34
- cuda/cccl/headers/include/cub/device/device_segmented_radix_sort.cuh +3 -27
- cuda/cccl/headers/include/cub/device/device_segmented_reduce.cuh +3 -27
- cuda/cccl/headers/include/cub/device/device_segmented_sort.cuh +2 -26
- cuda/cccl/headers/include/cub/device/device_select.cuh +3 -27
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_adjacent_difference.cuh +2 -28
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_batch_memcpy.cuh +2 -27
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_copy_mdspan.cuh +0 -2
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_for.cuh +3 -29
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_histogram.cuh +14 -34
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_merge_sort.cuh +5 -30
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_radix_sort.cuh +4 -29
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce.cuh +5 -32
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce_by_key.cuh +3 -29
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce_deterministic.cuh +2 -29
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce_nondeterministic.cuh +1 -2
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_rle.cuh +47 -59
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_scan.cuh +21 -30
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_scan_by_key.cuh +2 -27
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_segmented_sort.cuh +3 -27
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_select_if.cuh +3 -27
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_streaming_reduce.cuh +0 -2
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_streaming_reduce_by_key.cuh +51 -36
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_three_way_partition.cuh +3 -28
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_topk.cuh +0 -1
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_transform.cuh +27 -55
- cuda/cccl/headers/include/cub/device/dispatch/dispatch_unique_by_key.cuh +4 -28
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{for_each.cuh → kernel_for_each.cuh} +0 -2
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{histogram.cuh → kernel_histogram.cuh} +149 -157
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{merge_sort.cuh → kernel_merge_sort.cuh} +0 -2
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{radix_sort.cuh → kernel_radix_sort.cuh} +0 -2
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{reduce.cuh → kernel_reduce.cuh} +2 -28
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{scan.cuh → kernel_scan.cuh} +2 -28
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{segmented_reduce.cuh → kernel_segmented_reduce.cuh} +3 -29
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{segmented_sort.cuh → kernel_segmented_sort.cuh} +0 -1
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{three_way_partition.cuh → kernel_three_way_partition.cuh} +0 -1
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{transform.cuh → kernel_transform.cuh} +11 -11
- cuda/cccl/headers/include/cub/device/dispatch/kernels/{unique_by_key.cuh → kernel_unique_by_key.cuh} +0 -1
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_adjacent_difference.cuh +2 -26
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_batch_memcpy.cuh +2 -26
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_for.cuh +2 -28
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_histogram.cuh +6 -26
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_merge.cuh +2 -26
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_merge_sort.cuh +5 -31
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_radix_sort.cuh +31 -33
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_reduce.cuh +15 -40
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh +2 -26
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_run_length_encode.cuh +2 -28
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan.cuh +20 -44
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan_by_key.cuh +2 -26
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_segmented_sort.cuh +20 -45
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_select_if.cuh +2 -27
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_three_way_partition.cuh +11 -36
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_topk.cuh +0 -1
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_transform.cuh +2 -27
- cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_unique_by_key.cuh +14 -40
- cuda/cccl/headers/include/cub/grid/grid_even_share.cuh +3 -27
- cuda/cccl/headers/include/cub/grid/grid_mapping.cuh +3 -27
- cuda/cccl/headers/include/cub/grid/grid_queue.cuh +3 -27
- cuda/cccl/headers/include/cub/iterator/arg_index_input_iterator.cuh +3 -27
- cuda/cccl/headers/include/cub/iterator/cache_modified_input_iterator.cuh +3 -27
- cuda/cccl/headers/include/cub/iterator/cache_modified_output_iterator.cuh +3 -27
- cuda/cccl/headers/include/cub/iterator/tex_obj_input_iterator.cuh +3 -27
- cuda/cccl/headers/include/cub/thread/thread_load.cuh +3 -28
- cuda/cccl/headers/include/cub/thread/thread_operators.cuh +3 -27
- cuda/cccl/headers/include/cub/thread/thread_reduce.cuh +3 -26
- cuda/cccl/headers/include/cub/thread/thread_scan.cuh +3 -29
- cuda/cccl/headers/include/cub/thread/thread_search.cuh +3 -27
- cuda/cccl/headers/include/cub/thread/thread_simd.cuh +0 -2
- cuda/cccl/headers/include/cub/thread/thread_sort.cuh +2 -26
- cuda/cccl/headers/include/cub/thread/thread_store.cuh +3 -27
- cuda/cccl/headers/include/cub/util_allocator.cuh +3 -27
- cuda/cccl/headers/include/cub/util_arch.cuh +3 -29
- cuda/cccl/headers/include/cub/util_cpp_dialect.cuh +2 -26
- cuda/cccl/headers/include/cub/util_debug.cuh +3 -27
- cuda/cccl/headers/include/cub/util_device.cuh +18 -59
- cuda/cccl/headers/include/cub/util_macro.cuh +4 -28
- cuda/cccl/headers/include/cub/util_math.cuh +2 -28
- cuda/cccl/headers/include/cub/util_namespace.cuh +3 -28
- cuda/cccl/headers/include/cub/util_policy_wrapper_t.cuh +3 -27
- cuda/cccl/headers/include/cub/util_ptx.cuh +6 -30
- cuda/cccl/headers/include/cub/util_temporary_storage.cuh +3 -29
- cuda/cccl/headers/include/cub/util_type.cuh +5 -32
- cuda/cccl/headers/include/cub/util_vsmem.cuh +2 -28
- cuda/cccl/headers/include/cub/version.cuh +2 -26
- cuda/cccl/headers/include/cub/warp/specializations/warp_exchange_shfl.cuh +10 -35
- cuda/cccl/headers/include/cub/warp/specializations/warp_exchange_smem.cuh +5 -30
- cuda/cccl/headers/include/cub/warp/specializations/warp_reduce_shfl.cuh +15 -39
- cuda/cccl/headers/include/cub/warp/specializations/warp_reduce_smem.cuh +5 -35
- cuda/cccl/headers/include/cub/warp/specializations/warp_scan_shfl.cuh +22 -46
- cuda/cccl/headers/include/cub/warp/specializations/warp_scan_smem.cuh +3 -27
- cuda/cccl/headers/include/cub/warp/warp_exchange.cuh +2 -26
- cuda/cccl/headers/include/cub/warp/warp_load.cuh +4 -27
- cuda/cccl/headers/include/cub/warp/warp_merge_sort.cuh +2 -26
- cuda/cccl/headers/include/cub/warp/warp_reduce.cuh +3 -22
- cuda/cccl/headers/include/cub/warp/warp_scan.cuh +3 -27
- cuda/cccl/headers/include/cub/warp/warp_store.cuh +4 -27
- cuda/cccl/headers/include/cub/warp/warp_utils.cuh +0 -2
- cuda/cccl/headers/include/cuda/__barrier/barrier.h +1 -1
- cuda/cccl/headers/include/cuda/__barrier/barrier_arrive_tx.h +0 -1
- cuda/cccl/headers/include/cuda/__barrier/barrier_block_scope.h +277 -235
- cuda/cccl/headers/include/cuda/__barrier/barrier_expect_tx.h +0 -1
- cuda/cccl/headers/include/cuda/__driver/driver_api.h +13 -0
- cuda/cccl/headers/include/cuda/__execution/determinism.h +0 -2
- cuda/cccl/headers/include/cuda/__execution/output_ordering.h +0 -2
- cuda/cccl/headers/include/cuda/__functional/maximum.h +25 -7
- cuda/cccl/headers/include/cuda/__functional/minimum.h +25 -7
- cuda/cccl/headers/include/cuda/__functional/minimum_maximum_common.h +52 -0
- cuda/cccl/headers/include/cuda/__functional/proclaim_return_type.h +0 -2
- cuda/cccl/headers/include/cuda/__iterator/counting_iterator.h +13 -4
- cuda/cccl/headers/include/cuda/__iterator/zip_function.h +4 -2
- cuda/cccl/headers/include/cuda/__mdspan/host_device_accessor.h +0 -1
- cuda/cccl/headers/include/cuda/__memcpy_async/cp_async_bulk_shared_global.h +28 -7
- cuda/cccl/headers/include/cuda/__memcpy_async/dispatch_memcpy_async.h +1 -1
- cuda/cccl/headers/include/cuda/__memcpy_async/elect_one.h +52 -0
- cuda/cccl/headers/include/cuda/__memcpy_async/memcpy_async_tx.h +2 -3
- cuda/cccl/headers/include/cuda/__memcpy_async/memcpy_completion.h +1 -7
- cuda/cccl/headers/include/cuda/__memcpy_async/try_get_barrier_handle.h +0 -1
- cuda/cccl/headers/include/cuda/__memory/get_device_address.h +1 -1
- cuda/cccl/headers/include/cuda/__memory/ranges_overlap.h +126 -0
- cuda/cccl/headers/include/cuda/__memory_resource/any_resource.h +898 -0
- cuda/cccl/headers/include/cuda/__memory_resource/device_memory_pool.h +149 -0
- cuda/cccl/headers/include/cuda/__memory_resource/get_property.h +3 -3
- cuda/cccl/headers/include/cuda/__memory_resource/legacy_managed_memory_resource.h +148 -0
- cuda/cccl/headers/include/cuda/__memory_resource/legacy_pinned_memory_resource.h +139 -0
- cuda/cccl/headers/include/cuda/__memory_resource/managed_memory_pool.h +146 -0
- cuda/cccl/headers/include/cuda/__memory_resource/memory_resource_base.h +578 -0
- cuda/cccl/headers/include/cuda/__memory_resource/pinned_memory_pool.h +188 -0
- cuda/cccl/headers/include/cuda/__memory_resource/properties.h +3 -3
- cuda/cccl/headers/include/cuda/__memory_resource/resource.h +37 -3
- cuda/cccl/headers/include/cuda/__numeric/add_overflow.h +13 -3
- cuda/cccl/headers/include/cuda/__numeric/div_overflow.h +150 -0
- cuda/cccl/headers/include/cuda/__numeric/overflow_cast.h +2 -2
- cuda/cccl/headers/include/cuda/__numeric/sub_overflow.h +344 -0
- cuda/cccl/headers/include/cuda/__nvtx/nvtx3.h +0 -6
- cuda/cccl/headers/include/cuda/__ptx/instructions/shfl_sync.h +1 -1
- cuda/cccl/headers/include/cuda/__ptx/pragmas/enable_smem_spilling.h +47 -0
- cuda/cccl/headers/include/cuda/{std/__cuda → __runtime}/api_wrapper.h +3 -3
- cuda/cccl/headers/include/cuda/__stream/get_stream.h +0 -1
- cuda/cccl/headers/include/cuda/{__fwd/barrier_native_handle.h → __stream/internal_streams.h} +17 -15
- cuda/cccl/headers/include/cuda/__utility/__basic_any/basic_any_ptr.h +2 -2
- cuda/cccl/headers/include/cuda/__utility/__basic_any/basic_any_value.h +1 -0
- cuda/cccl/headers/include/cuda/__utility/__basic_any/semiregular.h +1 -0
- cuda/cccl/headers/include/cuda/__utility/__basic_any/virtcall.h +2 -1
- cuda/cccl/headers/include/cuda/barrier +42 -16
- cuda/cccl/headers/include/cuda/memory +1 -0
- cuda/cccl/headers/include/cuda/memory_resource +6 -1
- cuda/cccl/headers/include/cuda/numeric +2 -0
- cuda/cccl/headers/include/cuda/pipeline +3 -2
- cuda/cccl/headers/include/cuda/ptx +1 -0
- cuda/cccl/headers/include/cuda/std/__algorithm/unique_copy.h +0 -2
- cuda/cccl/headers/include/cuda/std/__atomic/api/reference.h +1 -1
- cuda/cccl/headers/include/cuda/std/__atomic/functions/cuda_ptx_derived.h +115 -58
- cuda/cccl/headers/include/cuda/std/__atomic/functions/cuda_ptx_generated.h +844 -378
- cuda/cccl/headers/include/cuda/std/__atomic/functions/cuda_ptx_generated_helper.h +12 -5
- cuda/cccl/headers/include/cuda/std/__atomic/functions/host.h +31 -0
- cuda/cccl/headers/include/cuda/std/__atomic/types/small.h +10 -0
- cuda/cccl/headers/include/cuda/std/__atomic/types.h +2 -3
- cuda/cccl/headers/include/cuda/std/__bit/byteswap.h +37 -13
- cuda/cccl/headers/include/cuda/std/__cccl/builtin.h +0 -28
- cuda/cccl/headers/include/cuda/std/__cccl/dialect.h +7 -0
- cuda/cccl/headers/include/cuda/std/__cccl/epilogue.h +10 -0
- cuda/cccl/headers/include/cuda/std/__cccl/exceptions.h +2 -45
- cuda/cccl/headers/include/cuda/std/__cccl/is_non_narrowing_convertible.h +0 -2
- cuda/cccl/headers/include/cuda/std/__cccl/prologue.h +8 -0
- cuda/cccl/headers/include/cuda/std/__chrono/calendar.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/day.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/duration.h +13 -17
- cuda/cccl/headers/include/cuda/std/__chrono/file_clock.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/high_resolution_clock.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/month.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/steady_clock.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/system_clock.h +0 -2
- cuda/cccl/headers/include/cuda/std/__chrono/time_point.h +5 -8
- cuda/cccl/headers/include/cuda/std/__chrono/year.h +0 -2
- cuda/cccl/headers/include/cuda/std/__cmath/error_functions.h +4 -0
- cuda/cccl/headers/include/cuda/std/__cmath/exponential_functions.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/fdim.h +4 -0
- cuda/cccl/headers/include/cuda/std/__cmath/fma.h +4 -0
- cuda/cccl/headers/include/cuda/std/__cmath/fpclassify.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/gamma.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/hyperbolic_functions.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/hypot.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/inverse_hyperbolic_functions.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/inverse_trigonometric_functions.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/isfinite.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/isinf.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/isnan.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/logarithms.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/min_max.h +2 -2
- cuda/cccl/headers/include/cuda/std/__cmath/remainder.h +4 -0
- cuda/cccl/headers/include/cuda/std/__cmath/roots.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/rounding_functions.h +2 -3
- cuda/cccl/headers/include/cuda/std/__cmath/traits.h +4 -0
- cuda/cccl/headers/include/cuda/std/__cmath/trigonometric_functions.h +2 -3
- cuda/cccl/headers/include/cuda/std/__complex/complex.h +0 -6
- cuda/cccl/headers/include/cuda/std/__complex/exponential_functions.h +2 -2
- cuda/cccl/headers/include/cuda/std/__concepts/concept_macros.h +27 -1
- cuda/cccl/headers/include/cuda/std/__concepts/equality_comparable.h +2 -4
- cuda/cccl/headers/include/cuda/std/__exception/cuda_error.h +15 -36
- cuda/cccl/headers/include/cuda/std/__exception/exception_macros.h +93 -0
- cuda/cccl/headers/include/cuda/std/{detail/libcxx/include/stdexcept → __exception/throw_error.h} +3 -3
- cuda/cccl/headers/include/cuda/std/__expected/expected.h +28 -43
- cuda/cccl/headers/include/cuda/std/__expected/unexpected.h +2 -10
- cuda/cccl/headers/include/cuda/std/__format/format_arg_store.h +2 -2
- cuda/cccl/headers/include/cuda/std/__functional/bind.h +6 -6
- cuda/cccl/headers/include/cuda/std/__functional/function.h +2 -6
- cuda/cccl/headers/include/cuda/std/__functional/invoke.h +5 -5
- cuda/cccl/headers/include/cuda/std/__fwd/allocator.h +5 -0
- cuda/cccl/headers/include/cuda/std/__fwd/array.h +2 -2
- cuda/cccl/headers/include/cuda/std/__fwd/char_traits.h +12 -0
- cuda/cccl/headers/include/cuda/std/__fwd/expected.h +46 -0
- cuda/cccl/headers/include/cuda/std/__fwd/get.h +21 -22
- cuda/cccl/headers/include/cuda/std/{detail/libcxx/include/iosfwd → __fwd/ios.h} +5 -10
- cuda/cccl/headers/include/cuda/std/__fwd/mdspan.h +19 -10
- cuda/cccl/headers/include/cuda/std/__fwd/optional.h +2 -2
- cuda/cccl/headers/include/cuda/std/__fwd/reference_wrapper.h +5 -0
- cuda/cccl/headers/include/cuda/std/__fwd/span.h +2 -2
- cuda/cccl/headers/include/cuda/std/__fwd/string.h +7 -0
- cuda/cccl/headers/include/cuda/std/__fwd/string_view.h +18 -0
- cuda/cccl/headers/include/cuda/std/__fwd/tuple.h +3 -0
- cuda/cccl/headers/include/cuda/std/__fwd/unexpected.h +40 -0
- cuda/cccl/headers/include/cuda/std/{__type_traits/is_reference_wrapper.h → __fwd/variant.h} +16 -15
- cuda/cccl/headers/include/cuda/std/__internal/features.h +14 -0
- cuda/cccl/headers/include/cuda/std/__iterator/istream_iterator.h +1 -1
- cuda/cccl/headers/include/cuda/std/__iterator/istreambuf_iterator.h +1 -1
- cuda/cccl/headers/include/cuda/std/__iterator/iter_swap.h +58 -40
- cuda/cccl/headers/include/cuda/std/__iterator/ostream_iterator.h +1 -1
- cuda/cccl/headers/include/cuda/std/__iterator/ostreambuf_iterator.h +1 -1
- cuda/cccl/headers/include/cuda/std/__iterator/reverse_iterator.h +0 -5
- cuda/cccl/headers/include/cuda/std/__limits/numeric_limits.h +4 -18
- cuda/cccl/headers/include/cuda/std/__linalg/conj_if_needed.h +1 -2
- cuda/cccl/headers/include/cuda/std/__linalg/conjugate_transposed.h +0 -2
- cuda/cccl/headers/include/cuda/std/__linalg/conjugated.h +0 -2
- cuda/cccl/headers/include/cuda/std/__linalg/scaled.h +0 -4
- cuda/cccl/headers/include/cuda/std/__linalg/transposed.h +0 -5
- cuda/cccl/headers/include/cuda/std/__mdspan/concepts.h +3 -10
- cuda/cccl/headers/include/cuda/std/__mdspan/extents.h +4 -15
- cuda/cccl/headers/include/cuda/std/__mdspan/layout_left.h +4 -4
- cuda/cccl/headers/include/cuda/std/__mdspan/layout_right.h +4 -4
- cuda/cccl/headers/include/cuda/std/__mdspan/layout_stride.h +2 -4
- cuda/cccl/headers/include/cuda/std/__mdspan/mdspan.h +3 -3
- cuda/cccl/headers/include/cuda/std/__mdspan/submdspan_helper.h +1 -1
- cuda/cccl/headers/include/cuda/std/__memory/allocator_arg_t.h +1 -0
- cuda/cccl/headers/include/cuda/std/__memory/allocator_traits.h +6 -12
- cuda/cccl/headers/include/cuda/std/__memory/construct_at.h +1 -5
- cuda/cccl/headers/include/cuda/std/__memory/is_sufficiently_aligned.h +7 -2
- cuda/cccl/headers/include/cuda/std/__memory/uninitialized_algorithms.h +1 -0
- cuda/cccl/headers/include/cuda/std/__memory/uses_allocator.h +5 -0
- cuda/cccl/headers/include/cuda/std/__new/allocate.h +5 -0
- cuda/cccl/headers/include/cuda/{__barrier/barrier_native_handle.h → std/__new/device_new.h} +9 -24
- cuda/cccl/headers/include/cuda/std/__new_ +1 -0
- cuda/cccl/headers/include/cuda/std/__optional/optional.h +5 -4
- cuda/cccl/headers/include/cuda/std/__optional/optional_ref.h +4 -4
- cuda/cccl/headers/include/cuda/std/__random/linear_congruential_engine.h +1 -1
- cuda/cccl/headers/include/cuda/std/__random/philox_engine.h +562 -0
- cuda/cccl/headers/include/cuda/std/__random/seed_seq.h +204 -0
- cuda/cccl/headers/include/cuda/std/__random_ +2 -0
- cuda/cccl/headers/include/cuda/std/__ranges/concepts.h +7 -19
- cuda/cccl/headers/include/cuda/std/__ranges/movable_box.h +2 -4
- cuda/cccl/headers/include/cuda/std/__ranges/owning_view.h +5 -4
- cuda/cccl/headers/include/cuda/std/__ranges/repeat_view.h +1 -1
- cuda/cccl/headers/include/cuda/std/__string/string_view.h +5 -5
- cuda/cccl/headers/include/cuda/std/__tuple_dir/apply.h +82 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/get.h +122 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/sfinae_helpers.h +0 -160
- cuda/cccl/headers/include/cuda/std/__tuple_dir/structured_bindings.h +123 -129
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tie.h +55 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple.h +457 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_cat.h +158 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_constraints.h +286 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_element.h +7 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_leaf.h +452 -0
- cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_like.h +1 -2
- cuda/cccl/headers/include/cuda/std/__type_traits/is_comparable.h +78 -0
- cuda/cccl/headers/include/cuda/std/__type_traits/is_convertible.h +1 -1
- cuda/cccl/headers/include/cuda/std/__type_traits/is_fully_bounded_array.h +47 -0
- cuda/cccl/headers/include/cuda/std/__type_traits/is_swappable.h +0 -2
- cuda/cccl/headers/include/cuda/std/__utility/in_place.h +4 -24
- cuda/cccl/headers/include/cuda/std/__utility/integer_sequence.h +0 -2
- cuda/cccl/headers/include/cuda/std/__utility/pair.h +20 -20
- cuda/cccl/headers/include/cuda/std/__utility/rel_ops.h +0 -2
- cuda/cccl/headers/include/cuda/std/__variant/bad_variant_access.h +74 -0
- cuda/cccl/headers/include/cuda/std/__variant/comparison.h +207 -0
- cuda/cccl/headers/include/cuda/std/__variant/get.h +192 -0
- cuda/cccl/headers/include/cuda/std/__variant/hash.h +82 -0
- cuda/cccl/headers/include/cuda/std/__variant/sfinae_helpers.h +89 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant.h +250 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant_access.h +70 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant_base.h +683 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant_constraints.h +135 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant_match.h +126 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant_traits.h +184 -0
- cuda/cccl/headers/include/cuda/std/__variant/variant_visit.h +225 -0
- cuda/cccl/headers/include/cuda/std/__variant/visit.h +148 -0
- cuda/cccl/headers/include/cuda/std/array +1 -1
- cuda/cccl/headers/include/cuda/std/atomic +1 -1
- cuda/cccl/headers/include/cuda/std/bitset +2 -10
- cuda/cccl/headers/include/cuda/std/detail/libcxx/include/__config +6 -6
- cuda/cccl/headers/include/cuda/std/detail/libcxx/include/algorithm +1 -4
- cuda/cccl/headers/include/cuda/std/detail/libcxx/include/chrono +3 -6
- cuda/cccl/headers/include/cuda/std/functional +1 -1
- cuda/cccl/headers/include/cuda/std/initializer_list +8 -0
- cuda/cccl/headers/include/cuda/std/inplace_vector +6 -5
- cuda/cccl/headers/include/cuda/std/iterator +1 -1
- cuda/cccl/headers/include/cuda/std/numbers +0 -2
- cuda/cccl/headers/include/cuda/std/ratio +2 -2
- cuda/cccl/headers/include/cuda/std/span +2 -2
- cuda/cccl/headers/include/cuda/std/string_view +24 -42
- cuda/cccl/headers/include/cuda/std/tuple +18 -1
- cuda/cccl/headers/include/cuda/std/type_traits +0 -1
- cuda/cccl/headers/include/cuda/std/variant +8 -1
- cuda/cccl/headers/include/nv/target +2 -6
- cuda/cccl/headers/include/thrust/detail/adjacent_difference.inl +15 -2
- cuda/cccl/headers/include/thrust/detail/allocator/allocator_traits.h +0 -2
- cuda/cccl/headers/include/thrust/detail/allocator/copy_construct_range.h +0 -1
- cuda/cccl/headers/include/thrust/detail/allocator/destroy_range.h +0 -1
- cuda/cccl/headers/include/thrust/detail/allocator/fill_construct_range.h +0 -2
- cuda/cccl/headers/include/thrust/detail/allocator/malloc_allocator.h +0 -2
- cuda/cccl/headers/include/thrust/detail/allocator/no_throw_allocator.h +0 -2
- cuda/cccl/headers/include/thrust/detail/allocator/tagged_allocator.h +0 -2
- cuda/cccl/headers/include/thrust/detail/allocator/temporary_allocator.h +0 -2
- cuda/cccl/headers/include/thrust/detail/allocator_aware_execution_policy.h +0 -4
- cuda/cccl/headers/include/thrust/detail/binary_search.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/complex/arithmetic.h +2 -7
- cuda/cccl/headers/include/thrust/detail/complex/c99math.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/catrig.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/catrigf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/ccosh.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/ccoshf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/cexp.h +2 -7
- cuda/cccl/headers/include/thrust/detail/complex/cexpf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/clog.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/clogf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/cproj.h +2 -7
- cuda/cccl/headers/include/thrust/detail/complex/csinh.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/csinhf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/csqrt.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/csqrtf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/ctanh.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/ctanhf.h +2 -8
- cuda/cccl/headers/include/thrust/detail/complex/math_private.h +2 -8
- cuda/cccl/headers/include/thrust/detail/config/device_system.h +2 -0
- cuda/cccl/headers/include/thrust/detail/config/host_system.h +2 -0
- cuda/cccl/headers/include/thrust/detail/config/namespace.h +0 -1
- cuda/cccl/headers/include/thrust/detail/contiguous_storage.h +0 -2
- cuda/cccl/headers/include/thrust/detail/contiguous_storage.inl +0 -2
- cuda/cccl/headers/include/thrust/detail/copy.h +0 -2
- cuda/cccl/headers/include/thrust/detail/copy.inl +14 -4
- cuda/cccl/headers/include/thrust/detail/copy_if.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/count.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/equal.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/execute_with_allocator.h +4 -5
- cuda/cccl/headers/include/thrust/detail/extrema.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/fill.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/find.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/for_each.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/functional/actor.h +2 -5
- cuda/cccl/headers/include/thrust/detail/functional/operators.h +2 -5
- cuda/cccl/headers/include/thrust/detail/gather.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/generate.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/get_iterator_value.h +0 -2
- cuda/cccl/headers/include/thrust/detail/inner_product.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/internal_functional.h +1 -0
- cuda/cccl/headers/include/thrust/detail/logical.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/malloc_and_free.h +13 -1
- cuda/cccl/headers/include/thrust/detail/merge.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/mismatch.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/overlapped_copy.h +0 -4
- cuda/cccl/headers/include/thrust/detail/partition.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/random_bijection.h +0 -2
- cuda/cccl/headers/include/thrust/detail/range/head_flags.h +0 -2
- cuda/cccl/headers/include/thrust/detail/range/tail_flags.h +0 -2
- cuda/cccl/headers/include/thrust/detail/raw_reference_cast.h +0 -6
- cuda/cccl/headers/include/thrust/detail/reduce.inl +21 -3
- cuda/cccl/headers/include/thrust/detail/reference.h +27 -3
- cuda/cccl/headers/include/thrust/detail/remove.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/replace.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/reverse.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/scan.inl +21 -3
- cuda/cccl/headers/include/thrust/detail/scatter.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/sequence.inl +13 -1
- cuda/cccl/headers/include/thrust/detail/set_operations.inl +13 -1
- cuda/cccl/headers/include/thrust/detail/sort.inl +13 -1
- cuda/cccl/headers/include/thrust/detail/static_assert.h +0 -2
- cuda/cccl/headers/include/thrust/detail/static_map.h +0 -3
- cuda/cccl/headers/include/thrust/detail/swap_ranges.inl +13 -1
- cuda/cccl/headers/include/thrust/detail/tabulate.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/temporary_array.h +0 -4
- cuda/cccl/headers/include/thrust/detail/temporary_array.inl +0 -1
- cuda/cccl/headers/include/thrust/detail/temporary_buffer.h +14 -3
- cuda/cccl/headers/include/thrust/detail/transform_reduce.inl +13 -1
- cuda/cccl/headers/include/thrust/detail/transform_scan.inl +13 -1
- cuda/cccl/headers/include/thrust/detail/trivial_sequence.h +0 -2
- cuda/cccl/headers/include/thrust/detail/tuple_meta_transform.h +0 -2
- cuda/cccl/headers/include/thrust/detail/type_traits/is_call_possible.h +2 -7
- cuda/cccl/headers/include/thrust/detail/type_traits/is_commutative.h +0 -2
- cuda/cccl/headers/include/thrust/detail/type_traits/is_thrust_pointer.h +0 -4
- cuda/cccl/headers/include/thrust/detail/type_traits/pointer_traits.h +0 -4
- cuda/cccl/headers/include/thrust/detail/uninitialized_copy.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/uninitialized_fill.inl +14 -2
- cuda/cccl/headers/include/thrust/detail/unique.inl +21 -3
- cuda/cccl/headers/include/thrust/detail/vector_base.h +0 -2
- cuda/cccl/headers/include/thrust/detail/vector_base.inl +0 -2
- cuda/cccl/headers/include/thrust/execution_policy.h +10 -9
- cuda/cccl/headers/include/thrust/functional.h +0 -2
- cuda/cccl/headers/include/thrust/iterator/detail/device_system_tag.h +9 -4
- cuda/cccl/headers/include/thrust/iterator/detail/host_system_tag.h +8 -4
- cuda/cccl/headers/include/thrust/iterator/detail/iterator_adaptor_base.h +0 -1
- cuda/cccl/headers/include/thrust/iterator/detail/iterator_category_with_system_and_traversal.h +0 -1
- cuda/cccl/headers/include/thrust/iterator/detail/iterator_facade_category.h +0 -1
- cuda/cccl/headers/include/thrust/iterator/detail/minimum_system.h +0 -1
- cuda/cccl/headers/include/thrust/iterator/detail/tagged_iterator.h +0 -1
- cuda/cccl/headers/include/thrust/iterator/detail/tuple_of_iterator_references.h +2 -6
- cuda/cccl/headers/include/thrust/iterator/transform_input_output_iterator.h +0 -1
- cuda/cccl/headers/include/thrust/iterator/transform_iterator.h +0 -2
- cuda/cccl/headers/include/thrust/mr/allocator.h +0 -2
- cuda/cccl/headers/include/thrust/mr/device_memory_resource.h +9 -4
- cuda/cccl/headers/include/thrust/mr/disjoint_pool.h +10 -10
- cuda/cccl/headers/include/thrust/mr/disjoint_sync_pool.h +0 -2
- cuda/cccl/headers/include/thrust/mr/disjoint_tls_pool.h +0 -2
- cuda/cccl/headers/include/thrust/mr/fancy_pointer_resource.h +0 -2
- cuda/cccl/headers/include/thrust/mr/host_memory_resource.h +8 -4
- cuda/cccl/headers/include/thrust/mr/memory_resource.h +0 -2
- cuda/cccl/headers/include/thrust/mr/new.h +0 -2
- cuda/cccl/headers/include/thrust/mr/polymorphic_adaptor.h +0 -2
- cuda/cccl/headers/include/thrust/mr/pool.h +10 -10
- cuda/cccl/headers/include/thrust/mr/pool_options.h +4 -6
- cuda/cccl/headers/include/thrust/mr/sync_pool.h +0 -2
- cuda/cccl/headers/include/thrust/mr/tls_pool.h +0 -2
- cuda/cccl/headers/include/thrust/mr/validator.h +0 -2
- cuda/cccl/headers/include/thrust/per_device_resource.h +13 -1
- cuda/cccl/headers/include/thrust/random/detail/discard_block_engine.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/linear_congruential_engine.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/linear_congruential_engine_discard.h +2 -9
- cuda/cccl/headers/include/thrust/random/detail/linear_feedback_shift_engine.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/linear_feedback_shift_engine_wordmask.h +2 -9
- cuda/cccl/headers/include/thrust/random/detail/mod.h +2 -9
- cuda/cccl/headers/include/thrust/random/detail/normal_distribution.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/normal_distribution_base.h +2 -7
- cuda/cccl/headers/include/thrust/random/detail/random_core_access.h +2 -9
- cuda/cccl/headers/include/thrust/random/detail/subtract_with_carry_engine.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/uniform_int_distribution.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/uniform_real_distribution.inl +0 -2
- cuda/cccl/headers/include/thrust/random/detail/xor_combine_engine.inl +0 -2
- cuda/cccl/headers/include/thrust/random/discard_block_engine.h +0 -2
- cuda/cccl/headers/include/thrust/random/linear_congruential_engine.h +0 -2
- cuda/cccl/headers/include/thrust/random/linear_feedback_shift_engine.h +0 -2
- cuda/cccl/headers/include/thrust/random/normal_distribution.h +0 -2
- cuda/cccl/headers/include/thrust/random/subtract_with_carry_engine.h +0 -2
- cuda/cccl/headers/include/thrust/random/uniform_int_distribution.h +0 -2
- cuda/cccl/headers/include/thrust/random/uniform_real_distribution.h +0 -2
- cuda/cccl/headers/include/thrust/random/xor_combine_engine.h +0 -2
- cuda/cccl/headers/include/thrust/random.h +0 -2
- cuda/cccl/headers/include/thrust/system/cpp/detail/execution_policy.h +15 -11
- cuda/cccl/headers/include/thrust/system/cpp/detail/memory.inl +2 -7
- cuda/cccl/headers/include/thrust/system/cpp/memory.h +0 -1
- cuda/cccl/headers/include/thrust/system/cpp/memory_resource.h +0 -2
- cuda/cccl/headers/include/thrust/system/cpp/pointer.h +0 -2
- cuda/cccl/headers/include/thrust/system/cpp/vector.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/adjacent_difference.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/copy.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/copy_if.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/core/agent_launcher.h +2 -9
- cuda/cccl/headers/include/thrust/system/cuda/detail/core/triple_chevron_launch.h +4 -32
- cuda/cccl/headers/include/thrust/system/cuda/detail/core/util.h +2 -9
- cuda/cccl/headers/include/thrust/system/cuda/detail/count.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/cross_system.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/dispatch.h +23 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/equal.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/error.inl +2 -11
- cuda/cccl/headers/include/thrust/system/cuda/detail/execution_policy.h +2 -0
- cuda/cccl/headers/include/thrust/system/cuda/detail/extrema.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/fill.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/find.h +0 -5
- cuda/cccl/headers/include/thrust/system/cuda/detail/for_each.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/gather.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/generate.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/iter_swap.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/make_unsigned_special.h +2 -8
- cuda/cccl/headers/include/thrust/system/cuda/detail/malloc_and_free.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/memory.inl +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/merge.h +2 -26
- cuda/cccl/headers/include/thrust/system/cuda/detail/mismatch.h +7 -142
- cuda/cccl/headers/include/thrust/system/cuda/detail/parallel_for.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/partition.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/per_device_resource.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/reduce.h +0 -5
- cuda/cccl/headers/include/thrust/system/cuda/detail/reduce_by_key.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/remove.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/replace.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/reverse.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/scan.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/scan_by_key.h +0 -3
- cuda/cccl/headers/include/thrust/system/cuda/detail/scatter.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/set_operations.h +3 -5
- cuda/cccl/headers/include/thrust/system/cuda/detail/sort.h +8 -10
- cuda/cccl/headers/include/thrust/system/cuda/detail/temporary_buffer.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/transform.h +0 -1
- cuda/cccl/headers/include/thrust/system/cuda/detail/transform_reduce.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/detail/transform_scan.h +0 -2
- cuda/cccl/headers/include/thrust/system/cuda/detail/uninitialized_copy.h +1 -7
- cuda/cccl/headers/include/thrust/system/cuda/detail/uninitialized_fill.h +2 -7
- cuda/cccl/headers/include/thrust/system/cuda/detail/unique.h +0 -3
- cuda/cccl/headers/include/thrust/system/cuda/detail/unique_by_key.h +0 -4
- cuda/cccl/headers/include/thrust/system/cuda/error.h +2 -11
- cuda/cccl/headers/include/thrust/system/cuda/memory.h +2 -6
- cuda/cccl/headers/include/thrust/system/cuda/memory_resource.h +2 -9
- cuda/cccl/headers/include/thrust/system/cuda/pointer.h +2 -7
- cuda/cccl/headers/include/thrust/system/cuda/vector.h +2 -6
- cuda/cccl/headers/include/thrust/system/detail/bad_alloc.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/errno.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/error_category.inl +0 -4
- cuda/cccl/headers/include/thrust/system/detail/error_code.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/error_condition.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/adjacent_difference.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/adjacent_difference.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/binary_search.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/binary_search.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/copy.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/copy.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/copy_if.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/copy_if.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/count.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/count.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/equal.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/equal.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/extrema.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/extrema.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/fill.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/find.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/find.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/for_each.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/gather.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/gather.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/generate.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/inner_product.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/inner_product.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/logical.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/memory.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/memory.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/merge.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/merge.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/mismatch.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/mismatch.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/partition.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/partition.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/per_device_resource.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/reduce.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/reduce.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/reduce_by_key.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/reduce_by_key.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/remove.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/remove.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/replace.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/replace.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/reverse.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/reverse.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/scalar/binary_search.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/scalar/binary_search.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/scan.h +26 -12
- cuda/cccl/headers/include/thrust/system/detail/generic/scan_by_key.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/scan_by_key.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/scatter.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/scatter.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/select_system.h +0 -1
- cuda/cccl/headers/include/thrust/system/detail/generic/set_operations.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/set_operations.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/shuffle.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/shuffle.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/sort.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/sort.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/swap_ranges.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/swap_ranges.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/tabulate.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/tabulate.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/tag.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/temporary_buffer.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/temporary_buffer.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/transform.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/transform_reduce.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/transform_reduce.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/transform_scan.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/transform_scan.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_copy.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_copy.inl +2 -4
- cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_fill.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_fill.inl +0 -3
- cuda/cccl/headers/include/thrust/system/detail/generic/unique.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/unique.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/unique_by_key.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/generic/unique_by_key.inl +0 -2
- cuda/cccl/headers/include/thrust/system/detail/internal/decompose.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/adjacent_difference.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/assign_value.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/binary_search.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/copy.h +76 -5
- cuda/cccl/headers/include/thrust/system/detail/sequential/copy_backward.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/copy_if.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/extrema.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/find.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/for_each.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/general_copy.h +0 -3
- cuda/cccl/headers/include/thrust/system/detail/sequential/get_value.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/insertion_sort.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/iter_swap.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/malloc_and_free.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/merge.h +78 -6
- cuda/cccl/headers/include/thrust/system/detail/sequential/partition.h +0 -4
- cuda/cccl/headers/include/thrust/system/detail/sequential/reduce.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/reduce_by_key.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/remove.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/scan.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/scan_by_key.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/set_operations.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/sort.h +67 -6
- cuda/cccl/headers/include/thrust/system/detail/sequential/stable_merge_sort.h +310 -11
- cuda/cccl/headers/include/thrust/system/detail/sequential/stable_primitive_sort.h +78 -5
- cuda/cccl/headers/include/thrust/system/detail/sequential/stable_radix_sort.h +543 -7
- cuda/cccl/headers/include/thrust/system/detail/sequential/trivial_copy.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/unique.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/sequential/unique_by_key.h +0 -2
- cuda/cccl/headers/include/thrust/system/detail/system_error.inl +0 -2
- cuda/cccl/headers/include/thrust/system/error_code.h +0 -4
- cuda/cccl/headers/include/thrust/system/omp/detail/adjacent_difference.h +5 -25
- cuda/cccl/headers/include/thrust/system/omp/detail/assign_value.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/binary_search.h +5 -25
- cuda/cccl/headers/include/thrust/system/omp/detail/copy.h +40 -29
- cuda/cccl/headers/include/thrust/system/omp/detail/copy_if.h +11 -28
- cuda/cccl/headers/include/thrust/system/omp/detail/count.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/default_decomposition.h +26 -28
- cuda/cccl/headers/include/thrust/system/omp/detail/equal.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/execution_policy.h +18 -13
- cuda/cccl/headers/include/thrust/system/omp/detail/extrema.h +5 -25
- cuda/cccl/headers/include/thrust/system/omp/detail/fill.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/find.h +5 -25
- cuda/cccl/headers/include/thrust/system/omp/detail/for_each.h +47 -30
- cuda/cccl/headers/include/thrust/system/omp/detail/gather.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/generate.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/get_value.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/inner_product.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/iter_swap.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/logical.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/malloc_and_free.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/merge.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/mismatch.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/partition.h +26 -31
- cuda/cccl/headers/include/thrust/system/omp/detail/per_device_resource.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/pragma_omp.h +2 -26
- cuda/cccl/headers/include/thrust/system/omp/detail/reduce.h +35 -27
- cuda/cccl/headers/include/thrust/system/omp/detail/reduce_by_key.h +13 -28
- cuda/cccl/headers/include/thrust/system/omp/detail/reduce_intervals.h +56 -28
- cuda/cccl/headers/include/thrust/system/omp/detail/remove.h +26 -31
- cuda/cccl/headers/include/thrust/system/omp/detail/replace.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/reverse.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/scan.h +176 -17
- cuda/cccl/headers/include/thrust/system/omp/detail/scan_by_key.h +8 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/scatter.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/sequence.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/set_operations.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/sort.h +213 -28
- cuda/cccl/headers/include/thrust/system/omp/detail/swap_ranges.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/tabulate.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/temporary_buffer.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/transform.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/transform_reduce.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/transform_scan.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/uninitialized_copy.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/uninitialized_fill.h +2 -15
- cuda/cccl/headers/include/thrust/system/omp/detail/unique.h +21 -30
- cuda/cccl/headers/include/thrust/system/omp/detail/unique_by_key.h +17 -29
- cuda/cccl/headers/include/thrust/system/omp/memory.h +51 -9
- cuda/cccl/headers/include/thrust/system/omp/memory_resource.h +3 -7
- cuda/cccl/headers/include/thrust/system/omp/pointer.h +3 -7
- cuda/cccl/headers/include/thrust/system/omp/vector.h +3 -6
- cuda/cccl/headers/include/thrust/system/system_error.h +0 -2
- cuda/cccl/headers/include/thrust/system/tbb/detail/adjacent_difference.h +4 -25
- cuda/cccl/headers/include/thrust/system/tbb/detail/assign_value.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/binary_search.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/copy.h +38 -29
- cuda/cccl/headers/include/thrust/system/tbb/detail/copy_if.h +91 -24
- cuda/cccl/headers/include/thrust/system/tbb/detail/count.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/equal.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/execution_policy.h +17 -13
- cuda/cccl/headers/include/thrust/system/tbb/detail/extrema.h +4 -25
- cuda/cccl/headers/include/thrust/system/tbb/detail/fill.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/find.h +4 -25
- cuda/cccl/headers/include/thrust/system/tbb/detail/for_each.h +47 -28
- cuda/cccl/headers/include/thrust/system/tbb/detail/gather.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/generate.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/get_value.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/inner_product.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/iter_swap.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/logical.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/malloc_and_free.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/merge.h +254 -29
- cuda/cccl/headers/include/thrust/system/tbb/detail/mismatch.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/partition.h +25 -31
- cuda/cccl/headers/include/thrust/system/tbb/detail/per_device_resource.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/reduce.h +95 -29
- cuda/cccl/headers/include/thrust/system/tbb/detail/reduce_by_key.h +345 -28
- cuda/cccl/headers/include/thrust/system/tbb/detail/reduce_intervals.h +4 -26
- cuda/cccl/headers/include/thrust/system/tbb/detail/remove.h +32 -42
- cuda/cccl/headers/include/thrust/system/tbb/detail/replace.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/reverse.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/scan.h +265 -30
- cuda/cccl/headers/include/thrust/system/tbb/detail/scan_by_key.h +7 -17
- cuda/cccl/headers/include/thrust/system/tbb/detail/scatter.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/sequence.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/set_operations.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/sort.h +244 -32
- cuda/cccl/headers/include/thrust/system/tbb/detail/swap_ranges.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/tabulate.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/temporary_buffer.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/transform.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/transform_reduce.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/transform_scan.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/uninitialized_copy.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/uninitialized_fill.h +2 -15
- cuda/cccl/headers/include/thrust/system/tbb/detail/unique.h +23 -33
- cuda/cccl/headers/include/thrust/system/tbb/detail/unique_by_key.h +16 -29
- cuda/cccl/headers/include/thrust/system/tbb/memory.h +52 -24
- cuda/cccl/headers/include/thrust/system/tbb/memory_resource.h +4 -22
- cuda/cccl/headers/include/thrust/system/tbb/pointer.h +4 -22
- cuda/cccl/headers/include/thrust/system/tbb/vector.h +4 -21
- cuda/cccl/headers/include/thrust/transform.h +14 -3
- cuda/cccl/headers/include/thrust/type_traits/integer_sequence.h +0 -4
- cuda/cccl/headers/include/thrust/type_traits/is_contiguous_iterator.h +0 -1
- cuda/cccl/headers/include/thrust/type_traits/is_operator_less_or_greater_function_object.h +0 -4
- cuda/cccl/headers/include/thrust/type_traits/is_operator_plus_function_object.h +0 -4
- cuda/cccl/headers/include/thrust/type_traits/is_trivially_relocatable.h +0 -4
- cuda/cccl/headers/include/thrust/universal_allocator.h +8 -0
- cuda/cccl/headers/include/thrust/universal_vector.h +9 -0
- cuda/cccl/headers/include/thrust/zip_function.h +2 -28
- cuda/compute/__init__.py +4 -0
- cuda/compute/_bindings.pyi +26 -3
- cuda/compute/_bindings_impl.pyx +143 -1
- cuda/compute/algorithms/__init__.py +9 -5
- cuda/compute/algorithms/_sort/__init__.py +23 -0
- cuda/compute/algorithms/{_merge_sort.py → _sort/_merge_sort.py} +10 -10
- cuda/compute/algorithms/{_radix_sort.py → _sort/_radix_sort.py} +9 -58
- cuda/compute/algorithms/_sort/_segmented_sort.py +288 -0
- cuda/compute/algorithms/_sort/_sort_common.py +52 -0
- cuda/compute/cu12/_bindings_impl.cpython-313-aarch64-linux-gnu.so +0 -0
- cuda/compute/cu12/cccl/libcccl.c.parallel.so +0 -0
- cuda/compute/cu13/_bindings_impl.cpython-313-aarch64-linux-gnu.so +0 -0
- cuda/compute/cu13/cccl/libcccl.c.parallel.so +0 -0
- cuda_cccl-0.3.4.dist-info/METADATA +78 -0
- {cuda_cccl-0.3.2.dist-info → cuda_cccl-0.3.4.dist-info}/RECORD +830 -867
- cuda/cccl/headers/include/cuda/__memory_resource/resource_ref.h +0 -652
- cuda/cccl/headers/include/cuda/std/detail/libcxx/include/tuple +0 -1365
- cuda/cccl/headers/include/cuda/std/detail/libcxx/include/variant +0 -2144
- cuda/cccl/headers/include/thrust/detail/integer_math.h +0 -113
- cuda/cccl/headers/include/thrust/system/detail/adl/adjacent_difference.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/assign_value.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/binary_search.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/copy.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/copy_if.h +0 -52
- cuda/cccl/headers/include/thrust/system/detail/adl/count.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/equal.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/extrema.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/fill.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/find.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/for_each.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/gather.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/generate.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/get_value.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/inner_product.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/iter_swap.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/logical.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/malloc_and_free.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/merge.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/mismatch.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/partition.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/per_device_resource.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/reduce.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/reduce_by_key.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/remove.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/replace.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/reverse.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/scan.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/scan_by_key.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/scatter.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/sequence.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/set_operations.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/sort.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/swap_ranges.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/tabulate.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/temporary_buffer.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/transform.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/transform_reduce.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/transform_scan.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/uninitialized_copy.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/uninitialized_fill.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/unique.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/adl/unique_by_key.h +0 -51
- cuda/cccl/headers/include/thrust/system/detail/generic/scan.inl +0 -85
- cuda/cccl/headers/include/thrust/system/detail/sequential/copy.inl +0 -119
- cuda/cccl/headers/include/thrust/system/detail/sequential/merge.inl +0 -145
- cuda/cccl/headers/include/thrust/system/detail/sequential/sort.inl +0 -116
- cuda/cccl/headers/include/thrust/system/detail/sequential/stable_merge_sort.inl +0 -356
- cuda/cccl/headers/include/thrust/system/detail/sequential/stable_primitive_sort.inl +0 -124
- cuda/cccl/headers/include/thrust/system/detail/sequential/stable_radix_sort.inl +0 -586
- cuda/cccl/headers/include/thrust/system/omp/detail/copy.inl +0 -74
- cuda/cccl/headers/include/thrust/system/omp/detail/copy_if.inl +0 -59
- cuda/cccl/headers/include/thrust/system/omp/detail/default_decomposition.inl +0 -65
- cuda/cccl/headers/include/thrust/system/omp/detail/for_each.inl +0 -87
- cuda/cccl/headers/include/thrust/system/omp/detail/memory.inl +0 -93
- cuda/cccl/headers/include/thrust/system/omp/detail/partition.inl +0 -102
- cuda/cccl/headers/include/thrust/system/omp/detail/reduce.inl +0 -78
- cuda/cccl/headers/include/thrust/system/omp/detail/reduce_by_key.inl +0 -65
- cuda/cccl/headers/include/thrust/system/omp/detail/reduce_intervals.inl +0 -103
- cuda/cccl/headers/include/thrust/system/omp/detail/remove.inl +0 -87
- cuda/cccl/headers/include/thrust/system/omp/detail/sort.inl +0 -265
- cuda/cccl/headers/include/thrust/system/omp/detail/unique.inl +0 -71
- cuda/cccl/headers/include/thrust/system/omp/detail/unique_by_key.inl +0 -75
- cuda/cccl/headers/include/thrust/system/tbb/detail/copy.inl +0 -73
- cuda/cccl/headers/include/thrust/system/tbb/detail/copy_if.inl +0 -136
- cuda/cccl/headers/include/thrust/system/tbb/detail/for_each.inl +0 -91
- cuda/cccl/headers/include/thrust/system/tbb/detail/memory.inl +0 -94
- cuda/cccl/headers/include/thrust/system/tbb/detail/merge.inl +0 -327
- cuda/cccl/headers/include/thrust/system/tbb/detail/partition.inl +0 -98
- cuda/cccl/headers/include/thrust/system/tbb/detail/reduce.inl +0 -137
- cuda/cccl/headers/include/thrust/system/tbb/detail/reduce_by_key.inl +0 -400
- cuda/cccl/headers/include/thrust/system/tbb/detail/remove.inl +0 -87
- cuda/cccl/headers/include/thrust/system/tbb/detail/scan.inl +0 -312
- cuda/cccl/headers/include/thrust/system/tbb/detail/sort.inl +0 -295
- cuda/cccl/headers/include/thrust/system/tbb/detail/unique.inl +0 -71
- cuda/cccl/headers/include/thrust/system/tbb/detail/unique_by_key.inl +0 -75
- cuda_cccl-0.3.2.dist-info/METADATA +0 -42
- {cuda_cccl-0.3.2.dist-info → cuda_cccl-0.3.4.dist-info}/WHEEL +0 -0
- {cuda_cccl-0.3.2.dist-info → cuda_cccl-0.3.4.dist-info}/licenses/LICENSE +0 -0
|
@@ -83,7 +83,6 @@ namespace cuda_cub
|
|
|
83
83
|
{
|
|
84
84
|
namespace detail
|
|
85
85
|
{
|
|
86
|
-
|
|
87
86
|
template <cub::SelectImpl SelectionOpt,
|
|
88
87
|
typename Derived,
|
|
89
88
|
typename InputIt,
|
|
@@ -212,7 +211,6 @@ select_unique(execution_policy<Derived>& policy, InputIt first, InputIt last, Ou
|
|
|
212
211
|
|
|
213
212
|
return output;
|
|
214
213
|
}
|
|
215
|
-
|
|
216
214
|
} // namespace detail
|
|
217
215
|
|
|
218
216
|
//-------------------------
|
|
@@ -279,7 +277,6 @@ unique_count(execution_policy<Derived>& policy, ForwardIt first, ForwardIt last,
|
|
|
279
277
|
return 1
|
|
280
278
|
+ thrust::count_if(policy, it, ::cuda::std::next(it, size - 1), zip_adj_not_predicate<BinaryPred>{binary_pred});
|
|
281
279
|
}
|
|
282
|
-
|
|
283
280
|
} // namespace cuda_cub
|
|
284
281
|
THRUST_NAMESPACE_END
|
|
285
282
|
|
|
@@ -79,10 +79,8 @@ _CCCL_HOST_DEVICE thrust::pair<OutputIterator1, OutputIterator2> unique_by_key_c
|
|
|
79
79
|
|
|
80
80
|
namespace cuda_cub
|
|
81
81
|
{
|
|
82
|
-
|
|
83
82
|
namespace detail
|
|
84
83
|
{
|
|
85
|
-
|
|
86
84
|
template <typename Derived,
|
|
87
85
|
typename KeyInputIt,
|
|
88
86
|
typename ValInputIt,
|
|
@@ -238,7 +236,6 @@ THRUST_RUNTIME_FUNCTION pair<KeyOutputIt, ValOutputIt> unique_by_key(
|
|
|
238
236
|
|
|
239
237
|
return result_end;
|
|
240
238
|
}
|
|
241
|
-
|
|
242
239
|
} // namespace detail
|
|
243
240
|
|
|
244
241
|
//-------------------------
|
|
@@ -300,7 +297,6 @@ unique_by_key(execution_policy<Derived>& policy, KeyInputIt keys_first, KeyInput
|
|
|
300
297
|
using key_type = thrust::detail::it_value_t<KeyInputIt>;
|
|
301
298
|
return cuda_cub::unique_by_key(policy, keys_first, keys_last, values_first, ::cuda::std::equal_to<key_type>());
|
|
302
299
|
}
|
|
303
|
-
|
|
304
300
|
} // namespace cuda_cub
|
|
305
301
|
THRUST_NAMESPACE_END
|
|
306
302
|
|
|
@@ -36,10 +36,6 @@ THRUST_NAMESPACE_BEGIN
|
|
|
36
36
|
|
|
37
37
|
namespace system
|
|
38
38
|
{
|
|
39
|
-
|
|
40
|
-
namespace cuda
|
|
41
|
-
{
|
|
42
|
-
|
|
43
39
|
// To construct an error_code after a CUDA Runtime error:
|
|
44
40
|
//
|
|
45
41
|
// error_code(::cudaGetLastError(), cuda_category())
|
|
@@ -47,9 +43,8 @@ namespace cuda
|
|
|
47
43
|
// XXX N3000 prefers enum class errc { ... }
|
|
48
44
|
/*! Namespace for CUDA Runtime errors.
|
|
49
45
|
*/
|
|
50
|
-
namespace errc
|
|
46
|
+
namespace cuda::errc
|
|
51
47
|
{
|
|
52
|
-
|
|
53
48
|
/*! \p errc_t enumerates the kinds of CUDA Runtime errors.
|
|
54
49
|
*/
|
|
55
50
|
enum errc_t
|
|
@@ -113,10 +108,7 @@ enum errc_t
|
|
|
113
108
|
attempted_operation_not_supported = cudaErrorNotSupported,
|
|
114
109
|
startup_failure = cudaErrorStartupFailure
|
|
115
110
|
}; // end errc_t
|
|
116
|
-
|
|
117
|
-
} // end namespace errc
|
|
118
|
-
|
|
119
|
-
} // namespace cuda
|
|
111
|
+
} // namespace cuda::errc
|
|
120
112
|
|
|
121
113
|
/*! \return A reference to an object of a type derived from class \p thrust::error_category.
|
|
122
114
|
* \note The object's \p equivalent virtual functions shall behave as specified
|
|
@@ -147,7 +139,6 @@ inline error_code make_error_code(cuda::errc::errc_t e);
|
|
|
147
139
|
/*! \return <tt>error_condition(static_cast<int>(e), cuda::error_category())</tt>.
|
|
148
140
|
*/
|
|
149
141
|
inline error_condition make_error_condition(cuda::errc::errc_t e);
|
|
150
|
-
|
|
151
142
|
} // namespace system
|
|
152
143
|
|
|
153
144
|
namespace cuda_cub
|
|
@@ -39,7 +39,6 @@
|
|
|
39
39
|
THRUST_NAMESPACE_BEGIN
|
|
40
40
|
namespace cuda_cub
|
|
41
41
|
{
|
|
42
|
-
|
|
43
42
|
/*! Allocates an area of memory available to Thrust's <tt>cuda</tt> system.
|
|
44
43
|
* \param n Number of bytes to allocate.
|
|
45
44
|
* \return A <tt>cuda::pointer<void></tt> pointing to the beginning of the newly
|
|
@@ -93,17 +92,14 @@ using universal_host_pinned_allocator =
|
|
|
93
92
|
thrust::mr::stateless_resource_allocator<T, thrust::system::cuda::universal_host_pinned_memory_resource>;
|
|
94
93
|
} // namespace cuda_cub
|
|
95
94
|
|
|
96
|
-
namespace system
|
|
97
|
-
{
|
|
98
|
-
namespace cuda
|
|
95
|
+
namespace system::cuda
|
|
99
96
|
{
|
|
100
97
|
using thrust::cuda_cub::allocator;
|
|
101
98
|
using thrust::cuda_cub::free;
|
|
102
99
|
using thrust::cuda_cub::malloc;
|
|
103
100
|
using thrust::cuda_cub::universal_allocator;
|
|
104
101
|
using thrust::cuda_cub::universal_host_pinned_allocator;
|
|
105
|
-
} // namespace cuda
|
|
106
|
-
} // namespace system
|
|
102
|
+
} // namespace system::cuda
|
|
107
103
|
|
|
108
104
|
/*! \namespace thrust::cuda
|
|
109
105
|
* \brief \p thrust::cuda is a top-level alias for \p thrust::system::cuda.
|
|
@@ -39,15 +39,11 @@
|
|
|
39
39
|
|
|
40
40
|
THRUST_NAMESPACE_BEGIN
|
|
41
41
|
|
|
42
|
-
namespace system
|
|
42
|
+
namespace system::cuda
|
|
43
43
|
{
|
|
44
|
-
namespace cuda
|
|
45
|
-
{
|
|
46
|
-
|
|
47
44
|
//! \cond
|
|
48
45
|
namespace detail
|
|
49
46
|
{
|
|
50
|
-
|
|
51
47
|
using allocation_fn = cudaError_t (*)(void**, std::size_t);
|
|
52
48
|
using deallocation_fn = cudaError_t (*)(void*);
|
|
53
49
|
|
|
@@ -90,7 +86,6 @@ using managed_memory_resource =
|
|
|
90
86
|
detail::cuda_memory_resource<detail::cudaMallocManaged, cudaFree, thrust::cuda::universal_pointer<void>>;
|
|
91
87
|
using pinned_memory_resource =
|
|
92
88
|
detail::cuda_memory_resource<cudaMallocHost, cudaFreeHost, thrust::cuda::universal_host_pinned_pointer<void>>;
|
|
93
|
-
|
|
94
89
|
} // namespace detail
|
|
95
90
|
//! \endcond
|
|
96
91
|
|
|
@@ -108,9 +103,7 @@ using universal_memory_resource = detail::managed_memory_resource;
|
|
|
108
103
|
* cuda::universal_pointer.
|
|
109
104
|
*/
|
|
110
105
|
using universal_host_pinned_memory_resource = detail::pinned_memory_resource;
|
|
111
|
-
|
|
112
|
-
} // namespace cuda
|
|
113
|
-
} // namespace system
|
|
106
|
+
} // namespace system::cuda
|
|
114
107
|
|
|
115
108
|
namespace cuda
|
|
116
109
|
{
|
|
@@ -38,7 +38,6 @@
|
|
|
38
38
|
THRUST_NAMESPACE_BEGIN
|
|
39
39
|
namespace cuda_cub
|
|
40
40
|
{
|
|
41
|
-
|
|
42
41
|
/*! \p cuda::pointer stores a pointer to an object allocated in memory
|
|
43
42
|
* accessible by the \p cuda system. This type provides type safety when
|
|
44
43
|
* dispatching algorithms on ranges resident in \p cuda memory.
|
|
@@ -117,7 +116,6 @@ using universal_host_pinned_pointer = thrust::pointer<T, cpp::tag, ::cuda::std::
|
|
|
117
116
|
*/
|
|
118
117
|
template <typename T>
|
|
119
118
|
using reference = thrust::tagged_reference<T, thrust::cuda_cub::tag>;
|
|
120
|
-
|
|
121
119
|
} // namespace cuda_cub
|
|
122
120
|
|
|
123
121
|
/*! \addtogroup system_backends Systems
|
|
@@ -133,16 +131,13 @@ using reference = thrust::tagged_reference<T, thrust::cuda_cub::tag>;
|
|
|
133
131
|
* aliased in the top-level <tt>thrust::cuda</tt> namespace for easy access.
|
|
134
132
|
*
|
|
135
133
|
*/
|
|
136
|
-
namespace system
|
|
137
|
-
{
|
|
138
|
-
namespace cuda
|
|
134
|
+
namespace system::cuda
|
|
139
135
|
{
|
|
140
136
|
using thrust::cuda_cub::pointer;
|
|
141
137
|
using thrust::cuda_cub::reference;
|
|
142
138
|
using thrust::cuda_cub::universal_host_pinned_pointer;
|
|
143
139
|
using thrust::cuda_cub::universal_pointer;
|
|
144
|
-
} // namespace cuda
|
|
145
|
-
} // namespace system
|
|
140
|
+
} // namespace system::cuda
|
|
146
141
|
/*! \}
|
|
147
142
|
*/
|
|
148
143
|
|
|
@@ -38,7 +38,6 @@
|
|
|
38
38
|
THRUST_NAMESPACE_BEGIN
|
|
39
39
|
namespace cuda_cub
|
|
40
40
|
{
|
|
41
|
-
|
|
42
41
|
/*! \p cuda::vector is a container that supports random access to elements,
|
|
43
42
|
* constant time removal of elements at the end, and linear time insertion
|
|
44
43
|
* and removal of elements at the beginning or in the middle. The number of
|
|
@@ -88,15 +87,12 @@ template <typename T>
|
|
|
88
87
|
using universal_host_pinned_vector = thrust::detail::vector_base<T, universal_host_pinned_allocator<T>>;
|
|
89
88
|
} // namespace cuda_cub
|
|
90
89
|
|
|
91
|
-
namespace system
|
|
92
|
-
{
|
|
93
|
-
namespace cuda
|
|
90
|
+
namespace system::cuda
|
|
94
91
|
{
|
|
95
92
|
using thrust::cuda_cub::universal_host_pinned_vector;
|
|
96
93
|
using thrust::cuda_cub::universal_vector;
|
|
97
94
|
using thrust::cuda_cub::vector;
|
|
98
|
-
} // namespace cuda
|
|
99
|
-
} // namespace system
|
|
95
|
+
} // namespace system::cuda
|
|
100
96
|
|
|
101
97
|
namespace cuda
|
|
102
98
|
{
|
|
@@ -32,7 +32,6 @@
|
|
|
32
32
|
THRUST_NAMESPACE_BEGIN
|
|
33
33
|
namespace system::detail
|
|
34
34
|
{
|
|
35
|
-
|
|
36
35
|
// define our own bad_alloc so we can set its .what()
|
|
37
36
|
class bad_alloc : public std::bad_alloc
|
|
38
37
|
{
|
|
@@ -56,6 +55,5 @@ public:
|
|
|
56
55
|
private:
|
|
57
56
|
std::string m_what;
|
|
58
57
|
}; // end bad_alloc
|
|
59
|
-
|
|
60
58
|
} // namespace system::detail
|
|
61
59
|
THRUST_NAMESPACE_END
|
|
@@ -35,7 +35,6 @@ THRUST_NAMESPACE_BEGIN
|
|
|
35
35
|
|
|
36
36
|
namespace system::detail
|
|
37
37
|
{
|
|
38
|
-
|
|
39
38
|
static const int eafnosupport = 9901;
|
|
40
39
|
static const int eaddrinuse = 9902;
|
|
41
40
|
static const int eaddrnotavail = 9903;
|
|
@@ -114,7 +113,6 @@ static const int eagain = 9976;
|
|
|
114
113
|
static const int enfile = 9977;
|
|
115
114
|
static const int emfile = 9978;
|
|
116
115
|
static const int emlink = 9979;
|
|
117
|
-
|
|
118
116
|
} // namespace system::detail
|
|
119
117
|
|
|
120
118
|
THRUST_NAMESPACE_END
|
|
@@ -36,7 +36,6 @@ THRUST_NAMESPACE_BEGIN
|
|
|
36
36
|
|
|
37
37
|
namespace system
|
|
38
38
|
{
|
|
39
|
-
|
|
40
39
|
error_category ::~error_category()
|
|
41
40
|
{
|
|
42
41
|
;
|
|
@@ -75,7 +74,6 @@ bool error_category ::operator<(const error_category& rhs) const
|
|
|
75
74
|
|
|
76
75
|
namespace detail
|
|
77
76
|
{
|
|
78
|
-
|
|
79
77
|
class generic_error_category : public error_category
|
|
80
78
|
{
|
|
81
79
|
public:
|
|
@@ -282,7 +280,6 @@ public:
|
|
|
282
280
|
}
|
|
283
281
|
}
|
|
284
282
|
}; // end system_category_result
|
|
285
|
-
|
|
286
283
|
} // namespace detail
|
|
287
284
|
|
|
288
285
|
inline const error_category& generic_category()
|
|
@@ -296,7 +293,6 @@ inline const error_category& system_category()
|
|
|
296
293
|
static const detail::system_error_category result;
|
|
297
294
|
return result;
|
|
298
295
|
}
|
|
299
|
-
|
|
300
296
|
} // namespace system
|
|
301
297
|
|
|
302
298
|
THRUST_NAMESPACE_END
|
|
@@ -32,7 +32,6 @@ THRUST_NAMESPACE_BEGIN
|
|
|
32
32
|
|
|
33
33
|
namespace system
|
|
34
34
|
{
|
|
35
|
-
|
|
36
35
|
error_code ::error_code()
|
|
37
36
|
: m_val(0)
|
|
38
37
|
, m_cat(&system_category())
|
|
@@ -167,7 +166,6 @@ bool operator!=(const error_condition& lhs, const error_condition& rhs)
|
|
|
167
166
|
{
|
|
168
167
|
return !(lhs == rhs);
|
|
169
168
|
} // end operator!=()
|
|
170
|
-
|
|
171
169
|
} // namespace system
|
|
172
170
|
|
|
173
171
|
THRUST_NAMESPACE_END
|
|
@@ -33,7 +33,6 @@ THRUST_NAMESPACE_BEGIN
|
|
|
33
33
|
|
|
34
34
|
namespace system
|
|
35
35
|
{
|
|
36
|
-
|
|
37
36
|
error_condition ::error_condition()
|
|
38
37
|
: m_val(0)
|
|
39
38
|
, m_cat(&generic_category())
|
|
@@ -115,7 +114,6 @@ bool operator<(const error_condition& lhs, const error_condition& rhs)
|
|
|
115
114
|
return lhs.category().operator<(rhs.category())
|
|
116
115
|
|| (lhs.category().operator==(rhs.category()) && (lhs.value() < rhs.value()));
|
|
117
116
|
} // end operator<()
|
|
118
|
-
|
|
119
117
|
} // namespace system
|
|
120
118
|
|
|
121
119
|
THRUST_NAMESPACE_END
|
|
@@ -34,7 +34,6 @@
|
|
|
34
34
|
THRUST_NAMESPACE_BEGIN
|
|
35
35
|
namespace system::detail::generic
|
|
36
36
|
{
|
|
37
|
-
|
|
38
37
|
template <typename DerivedPolicy, typename InputIterator, typename OutputIterator>
|
|
39
38
|
_CCCL_HOST_DEVICE OutputIterator adjacent_difference(
|
|
40
39
|
thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, OutputIterator result);
|
|
@@ -46,7 +45,6 @@ _CCCL_HOST_DEVICE OutputIterator adjacent_difference(
|
|
|
46
45
|
InputIterator last,
|
|
47
46
|
OutputIterator result,
|
|
48
47
|
BinaryFunction binary_op);
|
|
49
|
-
|
|
50
48
|
} // namespace system::detail::generic
|
|
51
49
|
THRUST_NAMESPACE_END
|
|
52
50
|
|
|
@@ -35,7 +35,6 @@
|
|
|
35
35
|
THRUST_NAMESPACE_BEGIN
|
|
36
36
|
namespace system::detail::generic
|
|
37
37
|
{
|
|
38
|
-
|
|
39
38
|
template <typename DerivedPolicy, typename InputIterator, typename OutputIterator>
|
|
40
39
|
_CCCL_HOST_DEVICE OutputIterator adjacent_difference(
|
|
41
40
|
thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, OutputIterator result)
|
|
@@ -74,6 +73,5 @@ _CCCL_HOST_DEVICE OutputIterator adjacent_difference(
|
|
|
74
73
|
|
|
75
74
|
return result + (last - first);
|
|
76
75
|
}
|
|
77
|
-
|
|
78
76
|
} // namespace system::detail::generic
|
|
79
77
|
THRUST_NAMESPACE_END
|
|
@@ -34,7 +34,6 @@
|
|
|
34
34
|
THRUST_NAMESPACE_BEGIN
|
|
35
35
|
namespace system::detail::generic
|
|
36
36
|
{
|
|
37
|
-
|
|
38
37
|
template <typename DerivedPolicy, typename ForwardIterator, typename T>
|
|
39
38
|
_CCCL_HOST_DEVICE ForwardIterator
|
|
40
39
|
lower_bound(thrust::execution_policy<DerivedPolicy>& exec, ForwardIterator begin, ForwardIterator end, const T& value);
|
|
@@ -154,7 +153,6 @@ _CCCL_HOST_DEVICE thrust::pair<ForwardIterator, ForwardIterator> equal_range(
|
|
|
154
153
|
ForwardIterator last,
|
|
155
154
|
const LessThanComparable& value,
|
|
156
155
|
StrictWeakOrdering comp);
|
|
157
|
-
|
|
158
156
|
} // namespace system::detail::generic
|
|
159
157
|
THRUST_NAMESPACE_END
|
|
160
158
|
|
|
@@ -48,7 +48,6 @@ namespace system::detail::generic
|
|
|
48
48
|
{
|
|
49
49
|
namespace detail
|
|
50
50
|
{
|
|
51
|
-
|
|
52
51
|
// short names to avoid nvcc bug
|
|
53
52
|
struct lbf
|
|
54
53
|
{
|
|
@@ -196,7 +195,6 @@ struct binary_search_less
|
|
|
196
195
|
return lhs < rhs;
|
|
197
196
|
}
|
|
198
197
|
};
|
|
199
|
-
|
|
200
198
|
} // end namespace detail
|
|
201
199
|
|
|
202
200
|
//////////////////////
|
|
@@ -379,6 +377,5 @@ _CCCL_HOST_DEVICE thrust::pair<ForwardIterator, ForwardIterator> equal_range(
|
|
|
379
377
|
ForwardIterator ub = thrust::upper_bound(exec, first, last, value, comp);
|
|
380
378
|
return thrust::make_pair(lb, ub);
|
|
381
379
|
}
|
|
382
|
-
|
|
383
380
|
} // namespace system::detail::generic
|
|
384
381
|
THRUST_NAMESPACE_END
|
|
@@ -30,7 +30,6 @@
|
|
|
30
30
|
THRUST_NAMESPACE_BEGIN
|
|
31
31
|
namespace system::detail::generic
|
|
32
32
|
{
|
|
33
|
-
|
|
34
33
|
template <typename DerivedPolicy, typename InputIterator, typename OutputIterator>
|
|
35
34
|
_CCCL_HOST_DEVICE OutputIterator
|
|
36
35
|
copy(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, OutputIterator result);
|
|
@@ -38,7 +37,6 @@ copy(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIt
|
|
|
38
37
|
template <typename DerivedPolicy, typename InputIterator, typename Size, typename OutputIterator>
|
|
39
38
|
_CCCL_HOST_DEVICE OutputIterator
|
|
40
39
|
copy_n(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, Size n, OutputIterator result);
|
|
41
|
-
|
|
42
40
|
} // namespace system::detail::generic
|
|
43
41
|
THRUST_NAMESPACE_END
|
|
44
42
|
|
|
@@ -36,7 +36,6 @@
|
|
|
36
36
|
THRUST_NAMESPACE_BEGIN
|
|
37
37
|
namespace system::detail::generic
|
|
38
38
|
{
|
|
39
|
-
|
|
40
39
|
template <typename DerivedPolicy, typename InputIterator, typename OutputIterator>
|
|
41
40
|
_CCCL_HOST_DEVICE OutputIterator
|
|
42
41
|
copy(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, OutputIterator result)
|
|
@@ -59,6 +58,5 @@ copy_n(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, Size
|
|
|
59
58
|
|
|
60
59
|
return thrust::get<1>(thrust::for_each_n(exec, zipped, n, functor_type{xfrm_type()}).get_iterator_tuple());
|
|
61
60
|
} // end copy_n()
|
|
62
|
-
|
|
63
61
|
} // namespace system::detail::generic
|
|
64
62
|
THRUST_NAMESPACE_END
|
|
@@ -30,7 +30,6 @@
|
|
|
30
30
|
THRUST_NAMESPACE_BEGIN
|
|
31
31
|
namespace system::detail::generic
|
|
32
32
|
{
|
|
33
|
-
|
|
34
33
|
template <typename DerivedPolicy, typename InputIterator, typename OutputIterator, typename Predicate>
|
|
35
34
|
_CCCL_HOST_DEVICE OutputIterator copy_if(
|
|
36
35
|
thrust::execution_policy<DerivedPolicy>& exec,
|
|
@@ -51,7 +50,6 @@ _CCCL_HOST_DEVICE OutputIterator copy_if(
|
|
|
51
50
|
InputIterator2 stencil,
|
|
52
51
|
OutputIterator result,
|
|
53
52
|
Predicate pred);
|
|
54
|
-
|
|
55
53
|
} // namespace system::detail::generic
|
|
56
54
|
THRUST_NAMESPACE_END
|
|
57
55
|
|
|
@@ -44,7 +44,6 @@ namespace system::detail::generic
|
|
|
44
44
|
{
|
|
45
45
|
namespace detail
|
|
46
46
|
{
|
|
47
|
-
|
|
48
47
|
template <typename IndexType,
|
|
49
48
|
typename DerivedPolicy,
|
|
50
49
|
typename InputIterator1,
|
|
@@ -84,7 +83,6 @@ _CCCL_HOST_DEVICE OutputIterator copy_if(
|
|
|
84
83
|
|
|
85
84
|
return result + output_size;
|
|
86
85
|
}
|
|
87
|
-
|
|
88
86
|
} // end namespace detail
|
|
89
87
|
|
|
90
88
|
template <typename DerivedPolicy, typename InputIterator, typename OutputIterator, typename Predicate>
|
|
@@ -141,6 +139,5 @@ _CCCL_HOST_DEVICE OutputIterator copy_if(
|
|
|
141
139
|
|
|
142
140
|
return result;
|
|
143
141
|
} // end copy_if()
|
|
144
|
-
|
|
145
142
|
} // namespace system::detail::generic
|
|
146
143
|
THRUST_NAMESPACE_END
|
|
@@ -30,7 +30,6 @@
|
|
|
30
30
|
THRUST_NAMESPACE_BEGIN
|
|
31
31
|
namespace system::detail::generic
|
|
32
32
|
{
|
|
33
|
-
|
|
34
33
|
template <typename DerivedPolicy, typename InputIterator, typename EqualityComparable>
|
|
35
34
|
_CCCL_HOST_DEVICE thrust::detail::it_difference_t<InputIterator>
|
|
36
35
|
count(thrust::execution_policy<DerivedPolicy>& exec,
|
|
@@ -41,7 +40,6 @@ count(thrust::execution_policy<DerivedPolicy>& exec,
|
|
|
41
40
|
template <typename DerivedPolicy, typename InputIterator, typename Predicate>
|
|
42
41
|
_CCCL_HOST_DEVICE thrust::detail::it_difference_t<InputIterator>
|
|
43
42
|
count_if(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, Predicate pred);
|
|
44
|
-
|
|
45
43
|
} // namespace system::detail::generic
|
|
46
44
|
THRUST_NAMESPACE_END
|
|
47
45
|
|
|
@@ -32,7 +32,6 @@
|
|
|
32
32
|
THRUST_NAMESPACE_BEGIN
|
|
33
33
|
namespace system::detail::generic
|
|
34
34
|
{
|
|
35
|
-
|
|
36
35
|
template <typename InputType, typename Predicate, typename CountType>
|
|
37
36
|
struct count_if_transform
|
|
38
37
|
{
|
|
@@ -79,6 +78,5 @@ count_if(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, Inp
|
|
|
79
78
|
::cuda::std::plus<CountType> binary_op;
|
|
80
79
|
return thrust::transform_reduce(exec, first, last, unary_op, CountType(0), binary_op);
|
|
81
80
|
} // end count_if()
|
|
82
|
-
|
|
83
81
|
} // namespace system::detail::generic
|
|
84
82
|
THRUST_NAMESPACE_END
|
|
@@ -30,7 +30,6 @@
|
|
|
30
30
|
THRUST_NAMESPACE_BEGIN
|
|
31
31
|
namespace system::detail::generic
|
|
32
32
|
{
|
|
33
|
-
|
|
34
33
|
template <typename DerivedPolicy, typename InputIterator1, typename InputIterator2>
|
|
35
34
|
_CCCL_HOST_DEVICE bool equal(
|
|
36
35
|
thrust::execution_policy<DerivedPolicy>& exec, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2);
|
|
@@ -42,7 +41,6 @@ equal(thrust::execution_policy<DerivedPolicy>& exec,
|
|
|
42
41
|
InputIterator1 last1,
|
|
43
42
|
InputIterator2 first2,
|
|
44
43
|
BinaryPredicate binary_pred);
|
|
45
|
-
|
|
46
44
|
} // namespace system::detail::generic
|
|
47
45
|
THRUST_NAMESPACE_END
|
|
48
46
|
|
|
@@ -34,7 +34,6 @@
|
|
|
34
34
|
THRUST_NAMESPACE_BEGIN
|
|
35
35
|
namespace system::detail::generic
|
|
36
36
|
{
|
|
37
|
-
|
|
38
37
|
template <typename DerivedPolicy, typename InputIterator1, typename InputIterator2>
|
|
39
38
|
_CCCL_HOST_DEVICE bool
|
|
40
39
|
equal(thrust::execution_policy<DerivedPolicy>& exec, InputIterator1 first1, InputIterator1 last1, InputIterator2 first2)
|
|
@@ -55,6 +54,5 @@ equal(thrust::execution_policy<DerivedPolicy>& exec,
|
|
|
55
54
|
{
|
|
56
55
|
return thrust::mismatch(exec, first1, last1, first2, binary_pred).first == last1;
|
|
57
56
|
}
|
|
58
|
-
|
|
59
57
|
} // namespace system::detail::generic
|
|
60
58
|
THRUST_NAMESPACE_END
|
|
@@ -35,7 +35,6 @@
|
|
|
35
35
|
THRUST_NAMESPACE_BEGIN
|
|
36
36
|
namespace system::detail::generic
|
|
37
37
|
{
|
|
38
|
-
|
|
39
38
|
template <typename DerivedPolicy, typename ForwardIterator>
|
|
40
39
|
_CCCL_HOST_DEVICE ForwardIterator
|
|
41
40
|
max_element(thrust::execution_policy<DerivedPolicy>& exec, ForwardIterator first, ForwardIterator last);
|
|
@@ -59,7 +58,6 @@ minmax_element(thrust::execution_policy<DerivedPolicy>& exec, ForwardIterator fi
|
|
|
59
58
|
template <typename DerivedPolicy, typename ForwardIterator, typename BinaryPredicate>
|
|
60
59
|
_CCCL_HOST_DEVICE thrust::pair<ForwardIterator, ForwardIterator> minmax_element(
|
|
61
60
|
thrust::execution_policy<DerivedPolicy>& exec, ForwardIterator first, ForwardIterator last, BinaryPredicate comp);
|
|
62
|
-
|
|
63
61
|
} // namespace system::detail::generic
|
|
64
62
|
THRUST_NAMESPACE_END
|
|
65
63
|
|
|
@@ -44,7 +44,6 @@ namespace system::detail::generic
|
|
|
44
44
|
{
|
|
45
45
|
namespace detail
|
|
46
46
|
{
|
|
47
|
-
|
|
48
47
|
//////////////
|
|
49
48
|
// Functors //
|
|
50
49
|
//////////////
|
|
@@ -148,7 +147,6 @@ struct duplicate_tuple
|
|
|
148
147
|
return thrust::make_tuple(t, t);
|
|
149
148
|
}
|
|
150
149
|
}; // end duplicate_tuple
|
|
151
|
-
|
|
152
150
|
} // end namespace detail
|
|
153
151
|
|
|
154
152
|
template <typename DerivedPolicy, typename ForwardIterator>
|
|
@@ -247,6 +245,5 @@ _CCCL_HOST_DEVICE thrust::pair<ForwardIterator, ForwardIterator> minmax_element(
|
|
|
247
245
|
return thrust::make_pair(first + thrust::get<1>(thrust::get<0>(result)),
|
|
248
246
|
first + thrust::get<1>(thrust::get<1>(result)));
|
|
249
247
|
} // end minmax_element()
|
|
250
|
-
|
|
251
248
|
} // namespace system::detail::generic
|
|
252
249
|
THRUST_NAMESPACE_END
|
|
@@ -33,7 +33,6 @@
|
|
|
33
33
|
THRUST_NAMESPACE_BEGIN
|
|
34
34
|
namespace system::detail::generic
|
|
35
35
|
{
|
|
36
|
-
|
|
37
36
|
template <typename DerivedPolicy, typename OutputIterator, typename Size, typename T>
|
|
38
37
|
_CCCL_HOST_DEVICE OutputIterator
|
|
39
38
|
fill_n(thrust::execution_policy<DerivedPolicy>& exec, OutputIterator first, Size n, const T& value)
|
|
@@ -49,6 +48,5 @@ fill(thrust::execution_policy<DerivedPolicy>& exec, ForwardIterator first, Forwa
|
|
|
49
48
|
// XXX consider using the placeholder expression _1 = value
|
|
50
49
|
thrust::generate(exec, first, last, thrust::detail::fill_functor<T>{value});
|
|
51
50
|
}
|
|
52
|
-
|
|
53
51
|
} // namespace system::detail::generic
|
|
54
52
|
THRUST_NAMESPACE_END
|
|
@@ -30,7 +30,6 @@
|
|
|
30
30
|
THRUST_NAMESPACE_BEGIN
|
|
31
31
|
namespace system::detail::generic
|
|
32
32
|
{
|
|
33
|
-
|
|
34
33
|
template <typename DerivedPolicy, typename InputIterator, typename T>
|
|
35
34
|
_CCCL_HOST_DEVICE InputIterator
|
|
36
35
|
find(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, const T& value);
|
|
@@ -42,7 +41,6 @@ find_if(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, Inpu
|
|
|
42
41
|
template <typename DerivedPolicy, typename InputIterator, typename Predicate>
|
|
43
42
|
_CCCL_HOST_DEVICE InputIterator
|
|
44
43
|
find_if_not(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, Predicate pred);
|
|
45
|
-
|
|
46
44
|
} // namespace system::detail::generic
|
|
47
45
|
THRUST_NAMESPACE_END
|
|
48
46
|
|
|
@@ -40,7 +40,6 @@
|
|
|
40
40
|
THRUST_NAMESPACE_BEGIN
|
|
41
41
|
namespace system::detail::generic
|
|
42
42
|
{
|
|
43
|
-
|
|
44
43
|
template <typename DerivedPolicy, typename InputIterator, typename T>
|
|
45
44
|
_CCCL_HOST_DEVICE InputIterator
|
|
46
45
|
find(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first, InputIterator last, const T& value)
|
|
@@ -132,6 +131,5 @@ find_if_not(thrust::execution_policy<DerivedPolicy>& exec, InputIterator first,
|
|
|
132
131
|
{
|
|
133
132
|
return thrust::find_if(exec, first, last, ::cuda::std::not_fn(pred));
|
|
134
133
|
} // end find()
|
|
135
|
-
|
|
136
134
|
} // namespace system::detail::generic
|
|
137
135
|
THRUST_NAMESPACE_END
|
|
@@ -37,7 +37,6 @@
|
|
|
37
37
|
THRUST_NAMESPACE_BEGIN
|
|
38
38
|
namespace system::detail::generic
|
|
39
39
|
{
|
|
40
|
-
|
|
41
40
|
template <typename DerivedPolicy, typename InputIterator, typename UnaryFunction>
|
|
42
41
|
_CCCL_HOST_DEVICE InputIterator
|
|
43
42
|
for_each(thrust::execution_policy<DerivedPolicy>&, InputIterator first, InputIterator, UnaryFunction)
|
|
@@ -53,6 +52,5 @@ for_each_n(thrust::execution_policy<DerivedPolicy>&, InputIterator first, Size,
|
|
|
53
52
|
static_assert(thrust::detail::depend_on_instantiation<InputIterator, false>::value, "unimplemented for this system");
|
|
54
53
|
return first;
|
|
55
54
|
} // end for_each_n()
|
|
56
|
-
|
|
57
55
|
} // namespace system::detail::generic
|
|
58
56
|
THRUST_NAMESPACE_END
|
|
@@ -30,7 +30,6 @@
|
|
|
30
30
|
THRUST_NAMESPACE_BEGIN
|
|
31
31
|
namespace system::detail::generic
|
|
32
32
|
{
|
|
33
|
-
|
|
34
33
|
template <typename DerivedPolicy, typename InputIterator, typename RandomAccessIterator, typename OutputIterator>
|
|
35
34
|
_CCCL_HOST_DEVICE OutputIterator gather(
|
|
36
35
|
thrust::execution_policy<DerivedPolicy>& exec,
|
|
@@ -66,7 +65,6 @@ _CCCL_HOST_DEVICE OutputIterator gather_if(
|
|
|
66
65
|
RandomAccessIterator input_first,
|
|
67
66
|
OutputIterator result,
|
|
68
67
|
Predicate pred);
|
|
69
|
-
|
|
70
68
|
} // namespace system::detail::generic
|
|
71
69
|
THRUST_NAMESPACE_END
|
|
72
70
|
|
|
@@ -34,7 +34,6 @@
|
|
|
34
34
|
THRUST_NAMESPACE_BEGIN
|
|
35
35
|
namespace system::detail::generic
|
|
36
36
|
{
|
|
37
|
-
|
|
38
37
|
template <typename DerivedPolicy, typename InputIterator, typename RandomAccessIterator, typename OutputIterator>
|
|
39
38
|
_CCCL_HOST_DEVICE OutputIterator gather(
|
|
40
39
|
thrust::execution_policy<DerivedPolicy>& exec,
|
|
@@ -91,6 +90,5 @@ _CCCL_HOST_DEVICE OutputIterator gather_if(
|
|
|
91
90
|
::cuda::std::identity{},
|
|
92
91
|
pred);
|
|
93
92
|
} // end gather_if()
|
|
94
|
-
|
|
95
93
|
} // namespace system::detail::generic
|
|
96
94
|
THRUST_NAMESPACE_END
|