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.

Files changed (911) hide show
  1. cuda/cccl/headers/include/cub/agent/agent_adjacent_difference.cuh +12 -38
  2. cuda/cccl/headers/include/cub/agent/agent_batch_memcpy.cuh +16 -40
  3. cuda/cccl/headers/include/cub/agent/agent_for.cuh +2 -28
  4. cuda/cccl/headers/include/cub/agent/agent_histogram.cuh +24 -56
  5. cuda/cccl/headers/include/cub/agent/agent_merge_sort.cuh +12 -38
  6. cuda/cccl/headers/include/cub/agent/agent_radix_sort_downsweep.cuh +31 -56
  7. cuda/cccl/headers/include/cub/agent/agent_radix_sort_histogram.cuh +31 -35
  8. cuda/cccl/headers/include/cub/agent/agent_radix_sort_onesweep.cuh +47 -48
  9. cuda/cccl/headers/include/cub/agent/agent_radix_sort_upsweep.cuh +39 -42
  10. cuda/cccl/headers/include/cub/agent/agent_reduce.cuh +33 -60
  11. cuda/cccl/headers/include/cub/agent/agent_reduce_by_key.cuh +18 -44
  12. cuda/cccl/headers/include/cub/agent/agent_rle.cuh +26 -55
  13. cuda/cccl/headers/include/cub/agent/agent_scan.cuh +22 -49
  14. cuda/cccl/headers/include/cub/agent/agent_scan_by_key.cuh +15 -41
  15. cuda/cccl/headers/include/cub/agent/agent_segmented_radix_sort.cuh +9 -35
  16. cuda/cccl/headers/include/cub/agent/agent_select_if.cuh +20 -49
  17. cuda/cccl/headers/include/cub/agent/agent_sub_warp_merge_sort.cuh +14 -40
  18. cuda/cccl/headers/include/cub/agent/agent_three_way_partition.cuh +18 -40
  19. cuda/cccl/headers/include/cub/agent/agent_topk.cuh +0 -2
  20. cuda/cccl/headers/include/cub/agent/agent_unique_by_key.cuh +20 -46
  21. cuda/cccl/headers/include/cub/agent/single_pass_scan_operators.cuh +3 -28
  22. cuda/cccl/headers/include/cub/block/block_adjacent_difference.cuh +7 -31
  23. cuda/cccl/headers/include/cub/block/block_discontinuity.cuh +10 -34
  24. cuda/cccl/headers/include/cub/block/block_exchange.cuh +120 -154
  25. cuda/cccl/headers/include/cub/block/block_histogram.cuh +28 -52
  26. cuda/cccl/headers/include/cub/block/block_load.cuh +124 -146
  27. cuda/cccl/headers/include/cub/block/block_load_to_shared.cuh +0 -16
  28. cuda/cccl/headers/include/cub/block/block_merge_sort.cuh +58 -87
  29. cuda/cccl/headers/include/cub/block/block_radix_rank.cuh +81 -100
  30. cuda/cccl/headers/include/cub/block/block_radix_sort.cuh +92 -156
  31. cuda/cccl/headers/include/cub/block/block_raking_layout.cuh +8 -32
  32. cuda/cccl/headers/include/cub/block/block_reduce.cuh +21 -46
  33. cuda/cccl/headers/include/cub/block/block_run_length_decode.cuh +51 -79
  34. cuda/cccl/headers/include/cub/block/block_scan.cuh +94 -401
  35. cuda/cccl/headers/include/cub/block/block_shuffle.cuh +10 -34
  36. cuda/cccl/headers/include/cub/block/block_store.cuh +73 -97
  37. cuda/cccl/headers/include/cub/block/radix_rank_sort_operations.cuh +2 -29
  38. cuda/cccl/headers/include/cub/block/specializations/block_histogram_atomic.cuh +5 -29
  39. cuda/cccl/headers/include/cub/block/specializations/block_histogram_sort.cuh +25 -49
  40. cuda/cccl/headers/include/cub/block/specializations/block_reduce_raking.cuh +12 -34
  41. cuda/cccl/headers/include/cub/block/specializations/block_reduce_raking_commutative_only.cuh +10 -34
  42. cuda/cccl/headers/include/cub/block/specializations/block_reduce_warp_reductions.cuh +3 -27
  43. cuda/cccl/headers/include/cub/block/specializations/block_scan_raking.cuh +12 -36
  44. cuda/cccl/headers/include/cub/block/specializations/block_scan_warp_scans.cuh +9 -33
  45. cuda/cccl/headers/include/cub/config.cuh +2 -26
  46. cuda/cccl/headers/include/cub/cub.cuh +3 -27
  47. cuda/cccl/headers/include/cub/detail/array_utils.cuh +2 -26
  48. cuda/cccl/headers/include/cub/detail/choose_offset.cuh +2 -28
  49. cuda/cccl/headers/include/cub/detail/detect_cuda_runtime.cuh +3 -27
  50. cuda/cccl/headers/include/cub/detail/device_double_buffer.cuh +0 -2
  51. cuda/cccl/headers/include/cub/detail/device_memory_resource.cuh +1 -3
  52. cuda/cccl/headers/include/cub/detail/fast_modulo_division.cuh +2 -28
  53. cuda/cccl/headers/include/cub/detail/integer_utils.cuh +0 -2
  54. cuda/cccl/headers/include/cub/detail/launcher/cuda_driver.cuh +0 -2
  55. cuda/cccl/headers/include/cub/detail/launcher/cuda_runtime.cuh +0 -2
  56. cuda/cccl/headers/include/cub/detail/mdspan_utils.cuh +0 -2
  57. cuda/cccl/headers/include/cub/detail/ptx-json/README.md +7 -12
  58. cuda/cccl/headers/include/cub/detail/ptx-json/array.h +6 -33
  59. cuda/cccl/headers/include/cub/detail/ptx-json/json.h +13 -36
  60. cuda/cccl/headers/include/cub/detail/ptx-json/object.h +9 -38
  61. cuda/cccl/headers/include/cub/detail/ptx-json/string.h +58 -32
  62. cuda/cccl/headers/include/cub/detail/ptx-json/value.h +51 -51
  63. cuda/cccl/headers/include/cub/detail/ptx-json-parser.h +7 -31
  64. cuda/cccl/headers/include/cub/detail/rfa.cuh +2 -27
  65. cuda/cccl/headers/include/cub/detail/strong_load.cuh +3 -29
  66. cuda/cccl/headers/include/cub/detail/strong_store.cuh +3 -29
  67. cuda/cccl/headers/include/cub/detail/temporary_storage.cuh +2 -9
  68. cuda/cccl/headers/include/cub/detail/type_traits.cuh +0 -2
  69. cuda/cccl/headers/include/cub/detail/uninitialized_copy.cuh +6 -31
  70. cuda/cccl/headers/include/cub/detail/unsafe_bitcast.cuh +2 -25
  71. cuda/cccl/headers/include/cub/device/device_adjacent_difference.cuh +2 -26
  72. cuda/cccl/headers/include/cub/device/device_for.cuh +3 -5
  73. cuda/cccl/headers/include/cub/device/device_histogram.cuh +3 -27
  74. cuda/cccl/headers/include/cub/device/device_memcpy.cuh +2 -26
  75. cuda/cccl/headers/include/cub/device/device_merge_sort.cuh +2 -26
  76. cuda/cccl/headers/include/cub/device/device_partition.cuh +3 -27
  77. cuda/cccl/headers/include/cub/device/device_radix_sort.cuh +3 -27
  78. cuda/cccl/headers/include/cub/device/device_reduce.cuh +10 -31
  79. cuda/cccl/headers/include/cub/device/device_run_length_encode.cuh +3 -27
  80. cuda/cccl/headers/include/cub/device/device_scan.cuh +16 -34
  81. cuda/cccl/headers/include/cub/device/device_segmented_radix_sort.cuh +3 -27
  82. cuda/cccl/headers/include/cub/device/device_segmented_reduce.cuh +3 -27
  83. cuda/cccl/headers/include/cub/device/device_segmented_sort.cuh +2 -26
  84. cuda/cccl/headers/include/cub/device/device_select.cuh +3 -27
  85. cuda/cccl/headers/include/cub/device/dispatch/dispatch_adjacent_difference.cuh +2 -28
  86. cuda/cccl/headers/include/cub/device/dispatch/dispatch_batch_memcpy.cuh +2 -27
  87. cuda/cccl/headers/include/cub/device/dispatch/dispatch_copy_mdspan.cuh +0 -2
  88. cuda/cccl/headers/include/cub/device/dispatch/dispatch_for.cuh +3 -29
  89. cuda/cccl/headers/include/cub/device/dispatch/dispatch_histogram.cuh +14 -34
  90. cuda/cccl/headers/include/cub/device/dispatch/dispatch_merge_sort.cuh +5 -30
  91. cuda/cccl/headers/include/cub/device/dispatch/dispatch_radix_sort.cuh +4 -29
  92. cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce.cuh +5 -32
  93. cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce_by_key.cuh +3 -29
  94. cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce_deterministic.cuh +2 -29
  95. cuda/cccl/headers/include/cub/device/dispatch/dispatch_reduce_nondeterministic.cuh +1 -2
  96. cuda/cccl/headers/include/cub/device/dispatch/dispatch_rle.cuh +47 -59
  97. cuda/cccl/headers/include/cub/device/dispatch/dispatch_scan.cuh +21 -30
  98. cuda/cccl/headers/include/cub/device/dispatch/dispatch_scan_by_key.cuh +2 -27
  99. cuda/cccl/headers/include/cub/device/dispatch/dispatch_segmented_sort.cuh +3 -27
  100. cuda/cccl/headers/include/cub/device/dispatch/dispatch_select_if.cuh +3 -27
  101. cuda/cccl/headers/include/cub/device/dispatch/dispatch_streaming_reduce.cuh +0 -2
  102. cuda/cccl/headers/include/cub/device/dispatch/dispatch_streaming_reduce_by_key.cuh +51 -36
  103. cuda/cccl/headers/include/cub/device/dispatch/dispatch_three_way_partition.cuh +3 -28
  104. cuda/cccl/headers/include/cub/device/dispatch/dispatch_topk.cuh +0 -1
  105. cuda/cccl/headers/include/cub/device/dispatch/dispatch_transform.cuh +27 -55
  106. cuda/cccl/headers/include/cub/device/dispatch/dispatch_unique_by_key.cuh +4 -28
  107. cuda/cccl/headers/include/cub/device/dispatch/kernels/{for_each.cuh → kernel_for_each.cuh} +0 -2
  108. cuda/cccl/headers/include/cub/device/dispatch/kernels/{histogram.cuh → kernel_histogram.cuh} +149 -157
  109. cuda/cccl/headers/include/cub/device/dispatch/kernels/{merge_sort.cuh → kernel_merge_sort.cuh} +0 -2
  110. cuda/cccl/headers/include/cub/device/dispatch/kernels/{radix_sort.cuh → kernel_radix_sort.cuh} +0 -2
  111. cuda/cccl/headers/include/cub/device/dispatch/kernels/{reduce.cuh → kernel_reduce.cuh} +2 -28
  112. cuda/cccl/headers/include/cub/device/dispatch/kernels/{scan.cuh → kernel_scan.cuh} +2 -28
  113. cuda/cccl/headers/include/cub/device/dispatch/kernels/{segmented_reduce.cuh → kernel_segmented_reduce.cuh} +3 -29
  114. cuda/cccl/headers/include/cub/device/dispatch/kernels/{segmented_sort.cuh → kernel_segmented_sort.cuh} +0 -1
  115. cuda/cccl/headers/include/cub/device/dispatch/kernels/{three_way_partition.cuh → kernel_three_way_partition.cuh} +0 -1
  116. cuda/cccl/headers/include/cub/device/dispatch/kernels/{transform.cuh → kernel_transform.cuh} +11 -11
  117. cuda/cccl/headers/include/cub/device/dispatch/kernels/{unique_by_key.cuh → kernel_unique_by_key.cuh} +0 -1
  118. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_adjacent_difference.cuh +2 -26
  119. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_batch_memcpy.cuh +2 -26
  120. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_for.cuh +2 -28
  121. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_histogram.cuh +6 -26
  122. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_merge.cuh +2 -26
  123. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_merge_sort.cuh +5 -31
  124. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_radix_sort.cuh +31 -33
  125. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_reduce.cuh +15 -40
  126. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh +2 -26
  127. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_run_length_encode.cuh +2 -28
  128. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan.cuh +20 -44
  129. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan_by_key.cuh +2 -26
  130. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_segmented_sort.cuh +20 -45
  131. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_select_if.cuh +2 -27
  132. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_three_way_partition.cuh +11 -36
  133. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_topk.cuh +0 -1
  134. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_transform.cuh +2 -27
  135. cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_unique_by_key.cuh +14 -40
  136. cuda/cccl/headers/include/cub/grid/grid_even_share.cuh +3 -27
  137. cuda/cccl/headers/include/cub/grid/grid_mapping.cuh +3 -27
  138. cuda/cccl/headers/include/cub/grid/grid_queue.cuh +3 -27
  139. cuda/cccl/headers/include/cub/iterator/arg_index_input_iterator.cuh +3 -27
  140. cuda/cccl/headers/include/cub/iterator/cache_modified_input_iterator.cuh +3 -27
  141. cuda/cccl/headers/include/cub/iterator/cache_modified_output_iterator.cuh +3 -27
  142. cuda/cccl/headers/include/cub/iterator/tex_obj_input_iterator.cuh +3 -27
  143. cuda/cccl/headers/include/cub/thread/thread_load.cuh +3 -28
  144. cuda/cccl/headers/include/cub/thread/thread_operators.cuh +3 -27
  145. cuda/cccl/headers/include/cub/thread/thread_reduce.cuh +3 -26
  146. cuda/cccl/headers/include/cub/thread/thread_scan.cuh +3 -29
  147. cuda/cccl/headers/include/cub/thread/thread_search.cuh +3 -27
  148. cuda/cccl/headers/include/cub/thread/thread_simd.cuh +0 -2
  149. cuda/cccl/headers/include/cub/thread/thread_sort.cuh +2 -26
  150. cuda/cccl/headers/include/cub/thread/thread_store.cuh +3 -27
  151. cuda/cccl/headers/include/cub/util_allocator.cuh +3 -27
  152. cuda/cccl/headers/include/cub/util_arch.cuh +3 -29
  153. cuda/cccl/headers/include/cub/util_cpp_dialect.cuh +2 -26
  154. cuda/cccl/headers/include/cub/util_debug.cuh +3 -27
  155. cuda/cccl/headers/include/cub/util_device.cuh +18 -59
  156. cuda/cccl/headers/include/cub/util_macro.cuh +4 -28
  157. cuda/cccl/headers/include/cub/util_math.cuh +2 -28
  158. cuda/cccl/headers/include/cub/util_namespace.cuh +3 -28
  159. cuda/cccl/headers/include/cub/util_policy_wrapper_t.cuh +3 -27
  160. cuda/cccl/headers/include/cub/util_ptx.cuh +6 -30
  161. cuda/cccl/headers/include/cub/util_temporary_storage.cuh +3 -29
  162. cuda/cccl/headers/include/cub/util_type.cuh +5 -32
  163. cuda/cccl/headers/include/cub/util_vsmem.cuh +2 -28
  164. cuda/cccl/headers/include/cub/version.cuh +2 -26
  165. cuda/cccl/headers/include/cub/warp/specializations/warp_exchange_shfl.cuh +10 -35
  166. cuda/cccl/headers/include/cub/warp/specializations/warp_exchange_smem.cuh +5 -30
  167. cuda/cccl/headers/include/cub/warp/specializations/warp_reduce_shfl.cuh +15 -39
  168. cuda/cccl/headers/include/cub/warp/specializations/warp_reduce_smem.cuh +5 -35
  169. cuda/cccl/headers/include/cub/warp/specializations/warp_scan_shfl.cuh +22 -46
  170. cuda/cccl/headers/include/cub/warp/specializations/warp_scan_smem.cuh +3 -27
  171. cuda/cccl/headers/include/cub/warp/warp_exchange.cuh +2 -26
  172. cuda/cccl/headers/include/cub/warp/warp_load.cuh +4 -27
  173. cuda/cccl/headers/include/cub/warp/warp_merge_sort.cuh +2 -26
  174. cuda/cccl/headers/include/cub/warp/warp_reduce.cuh +3 -22
  175. cuda/cccl/headers/include/cub/warp/warp_scan.cuh +3 -27
  176. cuda/cccl/headers/include/cub/warp/warp_store.cuh +4 -27
  177. cuda/cccl/headers/include/cub/warp/warp_utils.cuh +0 -2
  178. cuda/cccl/headers/include/cuda/__barrier/barrier.h +1 -1
  179. cuda/cccl/headers/include/cuda/__barrier/barrier_arrive_tx.h +0 -1
  180. cuda/cccl/headers/include/cuda/__barrier/barrier_block_scope.h +277 -235
  181. cuda/cccl/headers/include/cuda/__barrier/barrier_expect_tx.h +0 -1
  182. cuda/cccl/headers/include/cuda/__driver/driver_api.h +13 -0
  183. cuda/cccl/headers/include/cuda/__execution/determinism.h +0 -2
  184. cuda/cccl/headers/include/cuda/__execution/output_ordering.h +0 -2
  185. cuda/cccl/headers/include/cuda/__functional/maximum.h +25 -7
  186. cuda/cccl/headers/include/cuda/__functional/minimum.h +25 -7
  187. cuda/cccl/headers/include/cuda/__functional/minimum_maximum_common.h +52 -0
  188. cuda/cccl/headers/include/cuda/__functional/proclaim_return_type.h +0 -2
  189. cuda/cccl/headers/include/cuda/__iterator/counting_iterator.h +13 -4
  190. cuda/cccl/headers/include/cuda/__iterator/zip_function.h +4 -2
  191. cuda/cccl/headers/include/cuda/__mdspan/host_device_accessor.h +0 -1
  192. cuda/cccl/headers/include/cuda/__memcpy_async/cp_async_bulk_shared_global.h +28 -7
  193. cuda/cccl/headers/include/cuda/__memcpy_async/dispatch_memcpy_async.h +1 -1
  194. cuda/cccl/headers/include/cuda/__memcpy_async/elect_one.h +52 -0
  195. cuda/cccl/headers/include/cuda/__memcpy_async/memcpy_async_tx.h +2 -3
  196. cuda/cccl/headers/include/cuda/__memcpy_async/memcpy_completion.h +1 -7
  197. cuda/cccl/headers/include/cuda/__memcpy_async/try_get_barrier_handle.h +0 -1
  198. cuda/cccl/headers/include/cuda/__memory/get_device_address.h +1 -1
  199. cuda/cccl/headers/include/cuda/__memory/ranges_overlap.h +126 -0
  200. cuda/cccl/headers/include/cuda/__memory_resource/any_resource.h +898 -0
  201. cuda/cccl/headers/include/cuda/__memory_resource/device_memory_pool.h +149 -0
  202. cuda/cccl/headers/include/cuda/__memory_resource/get_property.h +3 -3
  203. cuda/cccl/headers/include/cuda/__memory_resource/legacy_managed_memory_resource.h +148 -0
  204. cuda/cccl/headers/include/cuda/__memory_resource/legacy_pinned_memory_resource.h +139 -0
  205. cuda/cccl/headers/include/cuda/__memory_resource/managed_memory_pool.h +146 -0
  206. cuda/cccl/headers/include/cuda/__memory_resource/memory_resource_base.h +578 -0
  207. cuda/cccl/headers/include/cuda/__memory_resource/pinned_memory_pool.h +188 -0
  208. cuda/cccl/headers/include/cuda/__memory_resource/properties.h +3 -3
  209. cuda/cccl/headers/include/cuda/__memory_resource/resource.h +37 -3
  210. cuda/cccl/headers/include/cuda/__numeric/add_overflow.h +13 -3
  211. cuda/cccl/headers/include/cuda/__numeric/div_overflow.h +150 -0
  212. cuda/cccl/headers/include/cuda/__numeric/overflow_cast.h +2 -2
  213. cuda/cccl/headers/include/cuda/__numeric/sub_overflow.h +344 -0
  214. cuda/cccl/headers/include/cuda/__nvtx/nvtx3.h +0 -6
  215. cuda/cccl/headers/include/cuda/__ptx/instructions/shfl_sync.h +1 -1
  216. cuda/cccl/headers/include/cuda/__ptx/pragmas/enable_smem_spilling.h +47 -0
  217. cuda/cccl/headers/include/cuda/{std/__cuda → __runtime}/api_wrapper.h +3 -3
  218. cuda/cccl/headers/include/cuda/__stream/get_stream.h +0 -1
  219. cuda/cccl/headers/include/cuda/{__fwd/barrier_native_handle.h → __stream/internal_streams.h} +17 -15
  220. cuda/cccl/headers/include/cuda/__utility/__basic_any/basic_any_ptr.h +2 -2
  221. cuda/cccl/headers/include/cuda/__utility/__basic_any/basic_any_value.h +1 -0
  222. cuda/cccl/headers/include/cuda/__utility/__basic_any/semiregular.h +1 -0
  223. cuda/cccl/headers/include/cuda/__utility/__basic_any/virtcall.h +2 -1
  224. cuda/cccl/headers/include/cuda/barrier +42 -16
  225. cuda/cccl/headers/include/cuda/memory +1 -0
  226. cuda/cccl/headers/include/cuda/memory_resource +6 -1
  227. cuda/cccl/headers/include/cuda/numeric +2 -0
  228. cuda/cccl/headers/include/cuda/pipeline +3 -2
  229. cuda/cccl/headers/include/cuda/ptx +1 -0
  230. cuda/cccl/headers/include/cuda/std/__algorithm/unique_copy.h +0 -2
  231. cuda/cccl/headers/include/cuda/std/__atomic/api/reference.h +1 -1
  232. cuda/cccl/headers/include/cuda/std/__atomic/functions/cuda_ptx_derived.h +115 -58
  233. cuda/cccl/headers/include/cuda/std/__atomic/functions/cuda_ptx_generated.h +844 -378
  234. cuda/cccl/headers/include/cuda/std/__atomic/functions/cuda_ptx_generated_helper.h +12 -5
  235. cuda/cccl/headers/include/cuda/std/__atomic/functions/host.h +31 -0
  236. cuda/cccl/headers/include/cuda/std/__atomic/types/small.h +10 -0
  237. cuda/cccl/headers/include/cuda/std/__atomic/types.h +2 -3
  238. cuda/cccl/headers/include/cuda/std/__bit/byteswap.h +37 -13
  239. cuda/cccl/headers/include/cuda/std/__cccl/builtin.h +0 -28
  240. cuda/cccl/headers/include/cuda/std/__cccl/dialect.h +7 -0
  241. cuda/cccl/headers/include/cuda/std/__cccl/epilogue.h +10 -0
  242. cuda/cccl/headers/include/cuda/std/__cccl/exceptions.h +2 -45
  243. cuda/cccl/headers/include/cuda/std/__cccl/is_non_narrowing_convertible.h +0 -2
  244. cuda/cccl/headers/include/cuda/std/__cccl/prologue.h +8 -0
  245. cuda/cccl/headers/include/cuda/std/__chrono/calendar.h +0 -2
  246. cuda/cccl/headers/include/cuda/std/__chrono/day.h +0 -2
  247. cuda/cccl/headers/include/cuda/std/__chrono/duration.h +13 -17
  248. cuda/cccl/headers/include/cuda/std/__chrono/file_clock.h +0 -2
  249. cuda/cccl/headers/include/cuda/std/__chrono/high_resolution_clock.h +0 -2
  250. cuda/cccl/headers/include/cuda/std/__chrono/month.h +0 -2
  251. cuda/cccl/headers/include/cuda/std/__chrono/steady_clock.h +0 -2
  252. cuda/cccl/headers/include/cuda/std/__chrono/system_clock.h +0 -2
  253. cuda/cccl/headers/include/cuda/std/__chrono/time_point.h +5 -8
  254. cuda/cccl/headers/include/cuda/std/__chrono/year.h +0 -2
  255. cuda/cccl/headers/include/cuda/std/__cmath/error_functions.h +4 -0
  256. cuda/cccl/headers/include/cuda/std/__cmath/exponential_functions.h +2 -3
  257. cuda/cccl/headers/include/cuda/std/__cmath/fdim.h +4 -0
  258. cuda/cccl/headers/include/cuda/std/__cmath/fma.h +4 -0
  259. cuda/cccl/headers/include/cuda/std/__cmath/fpclassify.h +2 -3
  260. cuda/cccl/headers/include/cuda/std/__cmath/gamma.h +2 -3
  261. cuda/cccl/headers/include/cuda/std/__cmath/hyperbolic_functions.h +2 -3
  262. cuda/cccl/headers/include/cuda/std/__cmath/hypot.h +2 -3
  263. cuda/cccl/headers/include/cuda/std/__cmath/inverse_hyperbolic_functions.h +2 -3
  264. cuda/cccl/headers/include/cuda/std/__cmath/inverse_trigonometric_functions.h +2 -3
  265. cuda/cccl/headers/include/cuda/std/__cmath/isfinite.h +2 -3
  266. cuda/cccl/headers/include/cuda/std/__cmath/isinf.h +2 -3
  267. cuda/cccl/headers/include/cuda/std/__cmath/isnan.h +2 -3
  268. cuda/cccl/headers/include/cuda/std/__cmath/logarithms.h +2 -3
  269. cuda/cccl/headers/include/cuda/std/__cmath/min_max.h +2 -2
  270. cuda/cccl/headers/include/cuda/std/__cmath/remainder.h +4 -0
  271. cuda/cccl/headers/include/cuda/std/__cmath/roots.h +2 -3
  272. cuda/cccl/headers/include/cuda/std/__cmath/rounding_functions.h +2 -3
  273. cuda/cccl/headers/include/cuda/std/__cmath/traits.h +4 -0
  274. cuda/cccl/headers/include/cuda/std/__cmath/trigonometric_functions.h +2 -3
  275. cuda/cccl/headers/include/cuda/std/__complex/complex.h +0 -6
  276. cuda/cccl/headers/include/cuda/std/__complex/exponential_functions.h +2 -2
  277. cuda/cccl/headers/include/cuda/std/__concepts/concept_macros.h +27 -1
  278. cuda/cccl/headers/include/cuda/std/__concepts/equality_comparable.h +2 -4
  279. cuda/cccl/headers/include/cuda/std/__exception/cuda_error.h +15 -36
  280. cuda/cccl/headers/include/cuda/std/__exception/exception_macros.h +93 -0
  281. cuda/cccl/headers/include/cuda/std/{detail/libcxx/include/stdexcept → __exception/throw_error.h} +3 -3
  282. cuda/cccl/headers/include/cuda/std/__expected/expected.h +28 -43
  283. cuda/cccl/headers/include/cuda/std/__expected/unexpected.h +2 -10
  284. cuda/cccl/headers/include/cuda/std/__format/format_arg_store.h +2 -2
  285. cuda/cccl/headers/include/cuda/std/__functional/bind.h +6 -6
  286. cuda/cccl/headers/include/cuda/std/__functional/function.h +2 -6
  287. cuda/cccl/headers/include/cuda/std/__functional/invoke.h +5 -5
  288. cuda/cccl/headers/include/cuda/std/__fwd/allocator.h +5 -0
  289. cuda/cccl/headers/include/cuda/std/__fwd/array.h +2 -2
  290. cuda/cccl/headers/include/cuda/std/__fwd/char_traits.h +12 -0
  291. cuda/cccl/headers/include/cuda/std/__fwd/expected.h +46 -0
  292. cuda/cccl/headers/include/cuda/std/__fwd/get.h +21 -22
  293. cuda/cccl/headers/include/cuda/std/{detail/libcxx/include/iosfwd → __fwd/ios.h} +5 -10
  294. cuda/cccl/headers/include/cuda/std/__fwd/mdspan.h +19 -10
  295. cuda/cccl/headers/include/cuda/std/__fwd/optional.h +2 -2
  296. cuda/cccl/headers/include/cuda/std/__fwd/reference_wrapper.h +5 -0
  297. cuda/cccl/headers/include/cuda/std/__fwd/span.h +2 -2
  298. cuda/cccl/headers/include/cuda/std/__fwd/string.h +7 -0
  299. cuda/cccl/headers/include/cuda/std/__fwd/string_view.h +18 -0
  300. cuda/cccl/headers/include/cuda/std/__fwd/tuple.h +3 -0
  301. cuda/cccl/headers/include/cuda/std/__fwd/unexpected.h +40 -0
  302. cuda/cccl/headers/include/cuda/std/{__type_traits/is_reference_wrapper.h → __fwd/variant.h} +16 -15
  303. cuda/cccl/headers/include/cuda/std/__internal/features.h +14 -0
  304. cuda/cccl/headers/include/cuda/std/__iterator/istream_iterator.h +1 -1
  305. cuda/cccl/headers/include/cuda/std/__iterator/istreambuf_iterator.h +1 -1
  306. cuda/cccl/headers/include/cuda/std/__iterator/iter_swap.h +58 -40
  307. cuda/cccl/headers/include/cuda/std/__iterator/ostream_iterator.h +1 -1
  308. cuda/cccl/headers/include/cuda/std/__iterator/ostreambuf_iterator.h +1 -1
  309. cuda/cccl/headers/include/cuda/std/__iterator/reverse_iterator.h +0 -5
  310. cuda/cccl/headers/include/cuda/std/__limits/numeric_limits.h +4 -18
  311. cuda/cccl/headers/include/cuda/std/__linalg/conj_if_needed.h +1 -2
  312. cuda/cccl/headers/include/cuda/std/__linalg/conjugate_transposed.h +0 -2
  313. cuda/cccl/headers/include/cuda/std/__linalg/conjugated.h +0 -2
  314. cuda/cccl/headers/include/cuda/std/__linalg/scaled.h +0 -4
  315. cuda/cccl/headers/include/cuda/std/__linalg/transposed.h +0 -5
  316. cuda/cccl/headers/include/cuda/std/__mdspan/concepts.h +3 -10
  317. cuda/cccl/headers/include/cuda/std/__mdspan/extents.h +4 -15
  318. cuda/cccl/headers/include/cuda/std/__mdspan/layout_left.h +4 -4
  319. cuda/cccl/headers/include/cuda/std/__mdspan/layout_right.h +4 -4
  320. cuda/cccl/headers/include/cuda/std/__mdspan/layout_stride.h +2 -4
  321. cuda/cccl/headers/include/cuda/std/__mdspan/mdspan.h +3 -3
  322. cuda/cccl/headers/include/cuda/std/__mdspan/submdspan_helper.h +1 -1
  323. cuda/cccl/headers/include/cuda/std/__memory/allocator_arg_t.h +1 -0
  324. cuda/cccl/headers/include/cuda/std/__memory/allocator_traits.h +6 -12
  325. cuda/cccl/headers/include/cuda/std/__memory/construct_at.h +1 -5
  326. cuda/cccl/headers/include/cuda/std/__memory/is_sufficiently_aligned.h +7 -2
  327. cuda/cccl/headers/include/cuda/std/__memory/uninitialized_algorithms.h +1 -0
  328. cuda/cccl/headers/include/cuda/std/__memory/uses_allocator.h +5 -0
  329. cuda/cccl/headers/include/cuda/std/__new/allocate.h +5 -0
  330. cuda/cccl/headers/include/cuda/{__barrier/barrier_native_handle.h → std/__new/device_new.h} +9 -24
  331. cuda/cccl/headers/include/cuda/std/__new_ +1 -0
  332. cuda/cccl/headers/include/cuda/std/__optional/optional.h +5 -4
  333. cuda/cccl/headers/include/cuda/std/__optional/optional_ref.h +4 -4
  334. cuda/cccl/headers/include/cuda/std/__random/linear_congruential_engine.h +1 -1
  335. cuda/cccl/headers/include/cuda/std/__random/philox_engine.h +562 -0
  336. cuda/cccl/headers/include/cuda/std/__random/seed_seq.h +204 -0
  337. cuda/cccl/headers/include/cuda/std/__random_ +2 -0
  338. cuda/cccl/headers/include/cuda/std/__ranges/concepts.h +7 -19
  339. cuda/cccl/headers/include/cuda/std/__ranges/movable_box.h +2 -4
  340. cuda/cccl/headers/include/cuda/std/__ranges/owning_view.h +5 -4
  341. cuda/cccl/headers/include/cuda/std/__ranges/repeat_view.h +1 -1
  342. cuda/cccl/headers/include/cuda/std/__string/string_view.h +5 -5
  343. cuda/cccl/headers/include/cuda/std/__tuple_dir/apply.h +82 -0
  344. cuda/cccl/headers/include/cuda/std/__tuple_dir/get.h +122 -0
  345. cuda/cccl/headers/include/cuda/std/__tuple_dir/sfinae_helpers.h +0 -160
  346. cuda/cccl/headers/include/cuda/std/__tuple_dir/structured_bindings.h +123 -129
  347. cuda/cccl/headers/include/cuda/std/__tuple_dir/tie.h +55 -0
  348. cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple.h +457 -0
  349. cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_cat.h +158 -0
  350. cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_constraints.h +286 -0
  351. cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_element.h +7 -0
  352. cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_leaf.h +452 -0
  353. cuda/cccl/headers/include/cuda/std/__tuple_dir/tuple_like.h +1 -2
  354. cuda/cccl/headers/include/cuda/std/__type_traits/is_comparable.h +78 -0
  355. cuda/cccl/headers/include/cuda/std/__type_traits/is_convertible.h +1 -1
  356. cuda/cccl/headers/include/cuda/std/__type_traits/is_fully_bounded_array.h +47 -0
  357. cuda/cccl/headers/include/cuda/std/__type_traits/is_swappable.h +0 -2
  358. cuda/cccl/headers/include/cuda/std/__utility/in_place.h +4 -24
  359. cuda/cccl/headers/include/cuda/std/__utility/integer_sequence.h +0 -2
  360. cuda/cccl/headers/include/cuda/std/__utility/pair.h +20 -20
  361. cuda/cccl/headers/include/cuda/std/__utility/rel_ops.h +0 -2
  362. cuda/cccl/headers/include/cuda/std/__variant/bad_variant_access.h +74 -0
  363. cuda/cccl/headers/include/cuda/std/__variant/comparison.h +207 -0
  364. cuda/cccl/headers/include/cuda/std/__variant/get.h +192 -0
  365. cuda/cccl/headers/include/cuda/std/__variant/hash.h +82 -0
  366. cuda/cccl/headers/include/cuda/std/__variant/sfinae_helpers.h +89 -0
  367. cuda/cccl/headers/include/cuda/std/__variant/variant.h +250 -0
  368. cuda/cccl/headers/include/cuda/std/__variant/variant_access.h +70 -0
  369. cuda/cccl/headers/include/cuda/std/__variant/variant_base.h +683 -0
  370. cuda/cccl/headers/include/cuda/std/__variant/variant_constraints.h +135 -0
  371. cuda/cccl/headers/include/cuda/std/__variant/variant_match.h +126 -0
  372. cuda/cccl/headers/include/cuda/std/__variant/variant_traits.h +184 -0
  373. cuda/cccl/headers/include/cuda/std/__variant/variant_visit.h +225 -0
  374. cuda/cccl/headers/include/cuda/std/__variant/visit.h +148 -0
  375. cuda/cccl/headers/include/cuda/std/array +1 -1
  376. cuda/cccl/headers/include/cuda/std/atomic +1 -1
  377. cuda/cccl/headers/include/cuda/std/bitset +2 -10
  378. cuda/cccl/headers/include/cuda/std/detail/libcxx/include/__config +6 -6
  379. cuda/cccl/headers/include/cuda/std/detail/libcxx/include/algorithm +1 -4
  380. cuda/cccl/headers/include/cuda/std/detail/libcxx/include/chrono +3 -6
  381. cuda/cccl/headers/include/cuda/std/functional +1 -1
  382. cuda/cccl/headers/include/cuda/std/initializer_list +8 -0
  383. cuda/cccl/headers/include/cuda/std/inplace_vector +6 -5
  384. cuda/cccl/headers/include/cuda/std/iterator +1 -1
  385. cuda/cccl/headers/include/cuda/std/numbers +0 -2
  386. cuda/cccl/headers/include/cuda/std/ratio +2 -2
  387. cuda/cccl/headers/include/cuda/std/span +2 -2
  388. cuda/cccl/headers/include/cuda/std/string_view +24 -42
  389. cuda/cccl/headers/include/cuda/std/tuple +18 -1
  390. cuda/cccl/headers/include/cuda/std/type_traits +0 -1
  391. cuda/cccl/headers/include/cuda/std/variant +8 -1
  392. cuda/cccl/headers/include/nv/target +2 -6
  393. cuda/cccl/headers/include/thrust/detail/adjacent_difference.inl +15 -2
  394. cuda/cccl/headers/include/thrust/detail/allocator/allocator_traits.h +0 -2
  395. cuda/cccl/headers/include/thrust/detail/allocator/copy_construct_range.h +0 -1
  396. cuda/cccl/headers/include/thrust/detail/allocator/destroy_range.h +0 -1
  397. cuda/cccl/headers/include/thrust/detail/allocator/fill_construct_range.h +0 -2
  398. cuda/cccl/headers/include/thrust/detail/allocator/malloc_allocator.h +0 -2
  399. cuda/cccl/headers/include/thrust/detail/allocator/no_throw_allocator.h +0 -2
  400. cuda/cccl/headers/include/thrust/detail/allocator/tagged_allocator.h +0 -2
  401. cuda/cccl/headers/include/thrust/detail/allocator/temporary_allocator.h +0 -2
  402. cuda/cccl/headers/include/thrust/detail/allocator_aware_execution_policy.h +0 -4
  403. cuda/cccl/headers/include/thrust/detail/binary_search.inl +14 -2
  404. cuda/cccl/headers/include/thrust/detail/complex/arithmetic.h +2 -7
  405. cuda/cccl/headers/include/thrust/detail/complex/c99math.h +2 -8
  406. cuda/cccl/headers/include/thrust/detail/complex/catrig.h +2 -8
  407. cuda/cccl/headers/include/thrust/detail/complex/catrigf.h +2 -8
  408. cuda/cccl/headers/include/thrust/detail/complex/ccosh.h +2 -8
  409. cuda/cccl/headers/include/thrust/detail/complex/ccoshf.h +2 -8
  410. cuda/cccl/headers/include/thrust/detail/complex/cexp.h +2 -7
  411. cuda/cccl/headers/include/thrust/detail/complex/cexpf.h +2 -8
  412. cuda/cccl/headers/include/thrust/detail/complex/clog.h +2 -8
  413. cuda/cccl/headers/include/thrust/detail/complex/clogf.h +2 -8
  414. cuda/cccl/headers/include/thrust/detail/complex/cproj.h +2 -7
  415. cuda/cccl/headers/include/thrust/detail/complex/csinh.h +2 -8
  416. cuda/cccl/headers/include/thrust/detail/complex/csinhf.h +2 -8
  417. cuda/cccl/headers/include/thrust/detail/complex/csqrt.h +2 -8
  418. cuda/cccl/headers/include/thrust/detail/complex/csqrtf.h +2 -8
  419. cuda/cccl/headers/include/thrust/detail/complex/ctanh.h +2 -8
  420. cuda/cccl/headers/include/thrust/detail/complex/ctanhf.h +2 -8
  421. cuda/cccl/headers/include/thrust/detail/complex/math_private.h +2 -8
  422. cuda/cccl/headers/include/thrust/detail/config/device_system.h +2 -0
  423. cuda/cccl/headers/include/thrust/detail/config/host_system.h +2 -0
  424. cuda/cccl/headers/include/thrust/detail/config/namespace.h +0 -1
  425. cuda/cccl/headers/include/thrust/detail/contiguous_storage.h +0 -2
  426. cuda/cccl/headers/include/thrust/detail/contiguous_storage.inl +0 -2
  427. cuda/cccl/headers/include/thrust/detail/copy.h +0 -2
  428. cuda/cccl/headers/include/thrust/detail/copy.inl +14 -4
  429. cuda/cccl/headers/include/thrust/detail/copy_if.inl +14 -2
  430. cuda/cccl/headers/include/thrust/detail/count.inl +14 -2
  431. cuda/cccl/headers/include/thrust/detail/equal.inl +14 -2
  432. cuda/cccl/headers/include/thrust/detail/execute_with_allocator.h +4 -5
  433. cuda/cccl/headers/include/thrust/detail/extrema.inl +14 -2
  434. cuda/cccl/headers/include/thrust/detail/fill.inl +14 -2
  435. cuda/cccl/headers/include/thrust/detail/find.inl +14 -2
  436. cuda/cccl/headers/include/thrust/detail/for_each.inl +14 -2
  437. cuda/cccl/headers/include/thrust/detail/functional/actor.h +2 -5
  438. cuda/cccl/headers/include/thrust/detail/functional/operators.h +2 -5
  439. cuda/cccl/headers/include/thrust/detail/gather.inl +14 -2
  440. cuda/cccl/headers/include/thrust/detail/generate.inl +14 -2
  441. cuda/cccl/headers/include/thrust/detail/get_iterator_value.h +0 -2
  442. cuda/cccl/headers/include/thrust/detail/inner_product.inl +14 -2
  443. cuda/cccl/headers/include/thrust/detail/internal_functional.h +1 -0
  444. cuda/cccl/headers/include/thrust/detail/logical.inl +14 -2
  445. cuda/cccl/headers/include/thrust/detail/malloc_and_free.h +13 -1
  446. cuda/cccl/headers/include/thrust/detail/merge.inl +14 -2
  447. cuda/cccl/headers/include/thrust/detail/mismatch.inl +14 -2
  448. cuda/cccl/headers/include/thrust/detail/overlapped_copy.h +0 -4
  449. cuda/cccl/headers/include/thrust/detail/partition.inl +14 -2
  450. cuda/cccl/headers/include/thrust/detail/random_bijection.h +0 -2
  451. cuda/cccl/headers/include/thrust/detail/range/head_flags.h +0 -2
  452. cuda/cccl/headers/include/thrust/detail/range/tail_flags.h +0 -2
  453. cuda/cccl/headers/include/thrust/detail/raw_reference_cast.h +0 -6
  454. cuda/cccl/headers/include/thrust/detail/reduce.inl +21 -3
  455. cuda/cccl/headers/include/thrust/detail/reference.h +27 -3
  456. cuda/cccl/headers/include/thrust/detail/remove.inl +14 -2
  457. cuda/cccl/headers/include/thrust/detail/replace.inl +14 -2
  458. cuda/cccl/headers/include/thrust/detail/reverse.inl +14 -2
  459. cuda/cccl/headers/include/thrust/detail/scan.inl +21 -3
  460. cuda/cccl/headers/include/thrust/detail/scatter.inl +14 -2
  461. cuda/cccl/headers/include/thrust/detail/sequence.inl +13 -1
  462. cuda/cccl/headers/include/thrust/detail/set_operations.inl +13 -1
  463. cuda/cccl/headers/include/thrust/detail/sort.inl +13 -1
  464. cuda/cccl/headers/include/thrust/detail/static_assert.h +0 -2
  465. cuda/cccl/headers/include/thrust/detail/static_map.h +0 -3
  466. cuda/cccl/headers/include/thrust/detail/swap_ranges.inl +13 -1
  467. cuda/cccl/headers/include/thrust/detail/tabulate.inl +14 -2
  468. cuda/cccl/headers/include/thrust/detail/temporary_array.h +0 -4
  469. cuda/cccl/headers/include/thrust/detail/temporary_array.inl +0 -1
  470. cuda/cccl/headers/include/thrust/detail/temporary_buffer.h +14 -3
  471. cuda/cccl/headers/include/thrust/detail/transform_reduce.inl +13 -1
  472. cuda/cccl/headers/include/thrust/detail/transform_scan.inl +13 -1
  473. cuda/cccl/headers/include/thrust/detail/trivial_sequence.h +0 -2
  474. cuda/cccl/headers/include/thrust/detail/tuple_meta_transform.h +0 -2
  475. cuda/cccl/headers/include/thrust/detail/type_traits/is_call_possible.h +2 -7
  476. cuda/cccl/headers/include/thrust/detail/type_traits/is_commutative.h +0 -2
  477. cuda/cccl/headers/include/thrust/detail/type_traits/is_thrust_pointer.h +0 -4
  478. cuda/cccl/headers/include/thrust/detail/type_traits/pointer_traits.h +0 -4
  479. cuda/cccl/headers/include/thrust/detail/uninitialized_copy.inl +14 -2
  480. cuda/cccl/headers/include/thrust/detail/uninitialized_fill.inl +14 -2
  481. cuda/cccl/headers/include/thrust/detail/unique.inl +21 -3
  482. cuda/cccl/headers/include/thrust/detail/vector_base.h +0 -2
  483. cuda/cccl/headers/include/thrust/detail/vector_base.inl +0 -2
  484. cuda/cccl/headers/include/thrust/execution_policy.h +10 -9
  485. cuda/cccl/headers/include/thrust/functional.h +0 -2
  486. cuda/cccl/headers/include/thrust/iterator/detail/device_system_tag.h +9 -4
  487. cuda/cccl/headers/include/thrust/iterator/detail/host_system_tag.h +8 -4
  488. cuda/cccl/headers/include/thrust/iterator/detail/iterator_adaptor_base.h +0 -1
  489. cuda/cccl/headers/include/thrust/iterator/detail/iterator_category_with_system_and_traversal.h +0 -1
  490. cuda/cccl/headers/include/thrust/iterator/detail/iterator_facade_category.h +0 -1
  491. cuda/cccl/headers/include/thrust/iterator/detail/minimum_system.h +0 -1
  492. cuda/cccl/headers/include/thrust/iterator/detail/tagged_iterator.h +0 -1
  493. cuda/cccl/headers/include/thrust/iterator/detail/tuple_of_iterator_references.h +2 -6
  494. cuda/cccl/headers/include/thrust/iterator/transform_input_output_iterator.h +0 -1
  495. cuda/cccl/headers/include/thrust/iterator/transform_iterator.h +0 -2
  496. cuda/cccl/headers/include/thrust/mr/allocator.h +0 -2
  497. cuda/cccl/headers/include/thrust/mr/device_memory_resource.h +9 -4
  498. cuda/cccl/headers/include/thrust/mr/disjoint_pool.h +10 -10
  499. cuda/cccl/headers/include/thrust/mr/disjoint_sync_pool.h +0 -2
  500. cuda/cccl/headers/include/thrust/mr/disjoint_tls_pool.h +0 -2
  501. cuda/cccl/headers/include/thrust/mr/fancy_pointer_resource.h +0 -2
  502. cuda/cccl/headers/include/thrust/mr/host_memory_resource.h +8 -4
  503. cuda/cccl/headers/include/thrust/mr/memory_resource.h +0 -2
  504. cuda/cccl/headers/include/thrust/mr/new.h +0 -2
  505. cuda/cccl/headers/include/thrust/mr/polymorphic_adaptor.h +0 -2
  506. cuda/cccl/headers/include/thrust/mr/pool.h +10 -10
  507. cuda/cccl/headers/include/thrust/mr/pool_options.h +4 -6
  508. cuda/cccl/headers/include/thrust/mr/sync_pool.h +0 -2
  509. cuda/cccl/headers/include/thrust/mr/tls_pool.h +0 -2
  510. cuda/cccl/headers/include/thrust/mr/validator.h +0 -2
  511. cuda/cccl/headers/include/thrust/per_device_resource.h +13 -1
  512. cuda/cccl/headers/include/thrust/random/detail/discard_block_engine.inl +0 -2
  513. cuda/cccl/headers/include/thrust/random/detail/linear_congruential_engine.inl +0 -2
  514. cuda/cccl/headers/include/thrust/random/detail/linear_congruential_engine_discard.h +2 -9
  515. cuda/cccl/headers/include/thrust/random/detail/linear_feedback_shift_engine.inl +0 -2
  516. cuda/cccl/headers/include/thrust/random/detail/linear_feedback_shift_engine_wordmask.h +2 -9
  517. cuda/cccl/headers/include/thrust/random/detail/mod.h +2 -9
  518. cuda/cccl/headers/include/thrust/random/detail/normal_distribution.inl +0 -2
  519. cuda/cccl/headers/include/thrust/random/detail/normal_distribution_base.h +2 -7
  520. cuda/cccl/headers/include/thrust/random/detail/random_core_access.h +2 -9
  521. cuda/cccl/headers/include/thrust/random/detail/subtract_with_carry_engine.inl +0 -2
  522. cuda/cccl/headers/include/thrust/random/detail/uniform_int_distribution.inl +0 -2
  523. cuda/cccl/headers/include/thrust/random/detail/uniform_real_distribution.inl +0 -2
  524. cuda/cccl/headers/include/thrust/random/detail/xor_combine_engine.inl +0 -2
  525. cuda/cccl/headers/include/thrust/random/discard_block_engine.h +0 -2
  526. cuda/cccl/headers/include/thrust/random/linear_congruential_engine.h +0 -2
  527. cuda/cccl/headers/include/thrust/random/linear_feedback_shift_engine.h +0 -2
  528. cuda/cccl/headers/include/thrust/random/normal_distribution.h +0 -2
  529. cuda/cccl/headers/include/thrust/random/subtract_with_carry_engine.h +0 -2
  530. cuda/cccl/headers/include/thrust/random/uniform_int_distribution.h +0 -2
  531. cuda/cccl/headers/include/thrust/random/uniform_real_distribution.h +0 -2
  532. cuda/cccl/headers/include/thrust/random/xor_combine_engine.h +0 -2
  533. cuda/cccl/headers/include/thrust/random.h +0 -2
  534. cuda/cccl/headers/include/thrust/system/cpp/detail/execution_policy.h +15 -11
  535. cuda/cccl/headers/include/thrust/system/cpp/detail/memory.inl +2 -7
  536. cuda/cccl/headers/include/thrust/system/cpp/memory.h +0 -1
  537. cuda/cccl/headers/include/thrust/system/cpp/memory_resource.h +0 -2
  538. cuda/cccl/headers/include/thrust/system/cpp/pointer.h +0 -2
  539. cuda/cccl/headers/include/thrust/system/cpp/vector.h +0 -1
  540. cuda/cccl/headers/include/thrust/system/cuda/detail/adjacent_difference.h +0 -4
  541. cuda/cccl/headers/include/thrust/system/cuda/detail/copy.h +0 -1
  542. cuda/cccl/headers/include/thrust/system/cuda/detail/copy_if.h +0 -4
  543. cuda/cccl/headers/include/thrust/system/cuda/detail/core/agent_launcher.h +2 -9
  544. cuda/cccl/headers/include/thrust/system/cuda/detail/core/triple_chevron_launch.h +4 -32
  545. cuda/cccl/headers/include/thrust/system/cuda/detail/core/util.h +2 -9
  546. cuda/cccl/headers/include/thrust/system/cuda/detail/count.h +0 -2
  547. cuda/cccl/headers/include/thrust/system/cuda/detail/cross_system.h +0 -2
  548. cuda/cccl/headers/include/thrust/system/cuda/detail/dispatch.h +23 -2
  549. cuda/cccl/headers/include/thrust/system/cuda/detail/equal.h +0 -2
  550. cuda/cccl/headers/include/thrust/system/cuda/detail/error.inl +2 -11
  551. cuda/cccl/headers/include/thrust/system/cuda/detail/execution_policy.h +2 -0
  552. cuda/cccl/headers/include/thrust/system/cuda/detail/extrema.h +0 -4
  553. cuda/cccl/headers/include/thrust/system/cuda/detail/fill.h +0 -1
  554. cuda/cccl/headers/include/thrust/system/cuda/detail/find.h +0 -5
  555. cuda/cccl/headers/include/thrust/system/cuda/detail/for_each.h +0 -1
  556. cuda/cccl/headers/include/thrust/system/cuda/detail/gather.h +0 -2
  557. cuda/cccl/headers/include/thrust/system/cuda/detail/generate.h +0 -2
  558. cuda/cccl/headers/include/thrust/system/cuda/detail/iter_swap.h +0 -1
  559. cuda/cccl/headers/include/thrust/system/cuda/detail/make_unsigned_special.h +2 -8
  560. cuda/cccl/headers/include/thrust/system/cuda/detail/malloc_and_free.h +0 -2
  561. cuda/cccl/headers/include/thrust/system/cuda/detail/memory.inl +0 -2
  562. cuda/cccl/headers/include/thrust/system/cuda/detail/merge.h +2 -26
  563. cuda/cccl/headers/include/thrust/system/cuda/detail/mismatch.h +7 -142
  564. cuda/cccl/headers/include/thrust/system/cuda/detail/parallel_for.h +0 -2
  565. cuda/cccl/headers/include/thrust/system/cuda/detail/partition.h +0 -4
  566. cuda/cccl/headers/include/thrust/system/cuda/detail/per_device_resource.h +0 -2
  567. cuda/cccl/headers/include/thrust/system/cuda/detail/reduce.h +0 -5
  568. cuda/cccl/headers/include/thrust/system/cuda/detail/reduce_by_key.h +0 -4
  569. cuda/cccl/headers/include/thrust/system/cuda/detail/remove.h +0 -2
  570. cuda/cccl/headers/include/thrust/system/cuda/detail/replace.h +0 -1
  571. cuda/cccl/headers/include/thrust/system/cuda/detail/reverse.h +0 -4
  572. cuda/cccl/headers/include/thrust/system/cuda/detail/scan.h +0 -4
  573. cuda/cccl/headers/include/thrust/system/cuda/detail/scan_by_key.h +0 -3
  574. cuda/cccl/headers/include/thrust/system/cuda/detail/scatter.h +0 -2
  575. cuda/cccl/headers/include/thrust/system/cuda/detail/set_operations.h +3 -5
  576. cuda/cccl/headers/include/thrust/system/cuda/detail/sort.h +8 -10
  577. cuda/cccl/headers/include/thrust/system/cuda/detail/temporary_buffer.h +0 -2
  578. cuda/cccl/headers/include/thrust/system/cuda/detail/transform.h +0 -1
  579. cuda/cccl/headers/include/thrust/system/cuda/detail/transform_reduce.h +0 -4
  580. cuda/cccl/headers/include/thrust/system/cuda/detail/transform_scan.h +0 -2
  581. cuda/cccl/headers/include/thrust/system/cuda/detail/uninitialized_copy.h +1 -7
  582. cuda/cccl/headers/include/thrust/system/cuda/detail/uninitialized_fill.h +2 -7
  583. cuda/cccl/headers/include/thrust/system/cuda/detail/unique.h +0 -3
  584. cuda/cccl/headers/include/thrust/system/cuda/detail/unique_by_key.h +0 -4
  585. cuda/cccl/headers/include/thrust/system/cuda/error.h +2 -11
  586. cuda/cccl/headers/include/thrust/system/cuda/memory.h +2 -6
  587. cuda/cccl/headers/include/thrust/system/cuda/memory_resource.h +2 -9
  588. cuda/cccl/headers/include/thrust/system/cuda/pointer.h +2 -7
  589. cuda/cccl/headers/include/thrust/system/cuda/vector.h +2 -6
  590. cuda/cccl/headers/include/thrust/system/detail/bad_alloc.h +0 -2
  591. cuda/cccl/headers/include/thrust/system/detail/errno.h +0 -2
  592. cuda/cccl/headers/include/thrust/system/detail/error_category.inl +0 -4
  593. cuda/cccl/headers/include/thrust/system/detail/error_code.inl +0 -2
  594. cuda/cccl/headers/include/thrust/system/detail/error_condition.inl +0 -2
  595. cuda/cccl/headers/include/thrust/system/detail/generic/adjacent_difference.h +0 -2
  596. cuda/cccl/headers/include/thrust/system/detail/generic/adjacent_difference.inl +0 -2
  597. cuda/cccl/headers/include/thrust/system/detail/generic/binary_search.h +0 -2
  598. cuda/cccl/headers/include/thrust/system/detail/generic/binary_search.inl +0 -3
  599. cuda/cccl/headers/include/thrust/system/detail/generic/copy.h +0 -2
  600. cuda/cccl/headers/include/thrust/system/detail/generic/copy.inl +0 -2
  601. cuda/cccl/headers/include/thrust/system/detail/generic/copy_if.h +0 -2
  602. cuda/cccl/headers/include/thrust/system/detail/generic/copy_if.inl +0 -3
  603. cuda/cccl/headers/include/thrust/system/detail/generic/count.h +0 -2
  604. cuda/cccl/headers/include/thrust/system/detail/generic/count.inl +0 -2
  605. cuda/cccl/headers/include/thrust/system/detail/generic/equal.h +0 -2
  606. cuda/cccl/headers/include/thrust/system/detail/generic/equal.inl +0 -2
  607. cuda/cccl/headers/include/thrust/system/detail/generic/extrema.h +0 -2
  608. cuda/cccl/headers/include/thrust/system/detail/generic/extrema.inl +0 -3
  609. cuda/cccl/headers/include/thrust/system/detail/generic/fill.h +0 -2
  610. cuda/cccl/headers/include/thrust/system/detail/generic/find.h +0 -2
  611. cuda/cccl/headers/include/thrust/system/detail/generic/find.inl +0 -2
  612. cuda/cccl/headers/include/thrust/system/detail/generic/for_each.h +0 -2
  613. cuda/cccl/headers/include/thrust/system/detail/generic/gather.h +0 -2
  614. cuda/cccl/headers/include/thrust/system/detail/generic/gather.inl +0 -2
  615. cuda/cccl/headers/include/thrust/system/detail/generic/generate.h +0 -2
  616. cuda/cccl/headers/include/thrust/system/detail/generic/inner_product.h +0 -2
  617. cuda/cccl/headers/include/thrust/system/detail/generic/inner_product.inl +0 -2
  618. cuda/cccl/headers/include/thrust/system/detail/generic/logical.h +0 -2
  619. cuda/cccl/headers/include/thrust/system/detail/generic/memory.h +0 -2
  620. cuda/cccl/headers/include/thrust/system/detail/generic/memory.inl +0 -3
  621. cuda/cccl/headers/include/thrust/system/detail/generic/merge.h +0 -2
  622. cuda/cccl/headers/include/thrust/system/detail/generic/merge.inl +0 -2
  623. cuda/cccl/headers/include/thrust/system/detail/generic/mismatch.h +0 -2
  624. cuda/cccl/headers/include/thrust/system/detail/generic/mismatch.inl +0 -2
  625. cuda/cccl/headers/include/thrust/system/detail/generic/partition.h +0 -2
  626. cuda/cccl/headers/include/thrust/system/detail/generic/partition.inl +0 -2
  627. cuda/cccl/headers/include/thrust/system/detail/generic/per_device_resource.h +0 -2
  628. cuda/cccl/headers/include/thrust/system/detail/generic/reduce.h +0 -2
  629. cuda/cccl/headers/include/thrust/system/detail/generic/reduce.inl +0 -2
  630. cuda/cccl/headers/include/thrust/system/detail/generic/reduce_by_key.h +0 -2
  631. cuda/cccl/headers/include/thrust/system/detail/generic/reduce_by_key.inl +0 -3
  632. cuda/cccl/headers/include/thrust/system/detail/generic/remove.h +0 -2
  633. cuda/cccl/headers/include/thrust/system/detail/generic/remove.inl +0 -2
  634. cuda/cccl/headers/include/thrust/system/detail/generic/replace.h +0 -2
  635. cuda/cccl/headers/include/thrust/system/detail/generic/replace.inl +0 -3
  636. cuda/cccl/headers/include/thrust/system/detail/generic/reverse.h +0 -2
  637. cuda/cccl/headers/include/thrust/system/detail/generic/reverse.inl +0 -2
  638. cuda/cccl/headers/include/thrust/system/detail/generic/scalar/binary_search.h +0 -2
  639. cuda/cccl/headers/include/thrust/system/detail/generic/scalar/binary_search.inl +0 -2
  640. cuda/cccl/headers/include/thrust/system/detail/generic/scan.h +26 -12
  641. cuda/cccl/headers/include/thrust/system/detail/generic/scan_by_key.h +0 -2
  642. cuda/cccl/headers/include/thrust/system/detail/generic/scan_by_key.inl +0 -3
  643. cuda/cccl/headers/include/thrust/system/detail/generic/scatter.h +0 -2
  644. cuda/cccl/headers/include/thrust/system/detail/generic/scatter.inl +0 -2
  645. cuda/cccl/headers/include/thrust/system/detail/generic/select_system.h +0 -1
  646. cuda/cccl/headers/include/thrust/system/detail/generic/set_operations.h +0 -2
  647. cuda/cccl/headers/include/thrust/system/detail/generic/set_operations.inl +0 -2
  648. cuda/cccl/headers/include/thrust/system/detail/generic/shuffle.h +0 -2
  649. cuda/cccl/headers/include/thrust/system/detail/generic/shuffle.inl +0 -2
  650. cuda/cccl/headers/include/thrust/system/detail/generic/sort.h +0 -2
  651. cuda/cccl/headers/include/thrust/system/detail/generic/sort.inl +0 -2
  652. cuda/cccl/headers/include/thrust/system/detail/generic/swap_ranges.h +0 -2
  653. cuda/cccl/headers/include/thrust/system/detail/generic/swap_ranges.inl +0 -3
  654. cuda/cccl/headers/include/thrust/system/detail/generic/tabulate.h +0 -2
  655. cuda/cccl/headers/include/thrust/system/detail/generic/tabulate.inl +0 -2
  656. cuda/cccl/headers/include/thrust/system/detail/generic/tag.h +0 -2
  657. cuda/cccl/headers/include/thrust/system/detail/generic/temporary_buffer.h +0 -2
  658. cuda/cccl/headers/include/thrust/system/detail/generic/temporary_buffer.inl +0 -2
  659. cuda/cccl/headers/include/thrust/system/detail/generic/transform.h +0 -2
  660. cuda/cccl/headers/include/thrust/system/detail/generic/transform_reduce.h +0 -2
  661. cuda/cccl/headers/include/thrust/system/detail/generic/transform_reduce.inl +0 -2
  662. cuda/cccl/headers/include/thrust/system/detail/generic/transform_scan.h +0 -2
  663. cuda/cccl/headers/include/thrust/system/detail/generic/transform_scan.inl +0 -2
  664. cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_copy.h +0 -2
  665. cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_copy.inl +2 -4
  666. cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_fill.h +0 -2
  667. cuda/cccl/headers/include/thrust/system/detail/generic/uninitialized_fill.inl +0 -3
  668. cuda/cccl/headers/include/thrust/system/detail/generic/unique.h +0 -2
  669. cuda/cccl/headers/include/thrust/system/detail/generic/unique.inl +0 -2
  670. cuda/cccl/headers/include/thrust/system/detail/generic/unique_by_key.h +0 -2
  671. cuda/cccl/headers/include/thrust/system/detail/generic/unique_by_key.inl +0 -2
  672. cuda/cccl/headers/include/thrust/system/detail/internal/decompose.h +0 -2
  673. cuda/cccl/headers/include/thrust/system/detail/sequential/adjacent_difference.h +0 -2
  674. cuda/cccl/headers/include/thrust/system/detail/sequential/assign_value.h +0 -2
  675. cuda/cccl/headers/include/thrust/system/detail/sequential/binary_search.h +0 -2
  676. cuda/cccl/headers/include/thrust/system/detail/sequential/copy.h +76 -5
  677. cuda/cccl/headers/include/thrust/system/detail/sequential/copy_backward.h +0 -2
  678. cuda/cccl/headers/include/thrust/system/detail/sequential/copy_if.h +0 -2
  679. cuda/cccl/headers/include/thrust/system/detail/sequential/extrema.h +0 -2
  680. cuda/cccl/headers/include/thrust/system/detail/sequential/find.h +0 -2
  681. cuda/cccl/headers/include/thrust/system/detail/sequential/for_each.h +0 -2
  682. cuda/cccl/headers/include/thrust/system/detail/sequential/general_copy.h +0 -3
  683. cuda/cccl/headers/include/thrust/system/detail/sequential/get_value.h +0 -2
  684. cuda/cccl/headers/include/thrust/system/detail/sequential/insertion_sort.h +0 -2
  685. cuda/cccl/headers/include/thrust/system/detail/sequential/iter_swap.h +0 -2
  686. cuda/cccl/headers/include/thrust/system/detail/sequential/malloc_and_free.h +0 -2
  687. cuda/cccl/headers/include/thrust/system/detail/sequential/merge.h +78 -6
  688. cuda/cccl/headers/include/thrust/system/detail/sequential/partition.h +0 -4
  689. cuda/cccl/headers/include/thrust/system/detail/sequential/reduce.h +0 -2
  690. cuda/cccl/headers/include/thrust/system/detail/sequential/reduce_by_key.h +0 -2
  691. cuda/cccl/headers/include/thrust/system/detail/sequential/remove.h +0 -2
  692. cuda/cccl/headers/include/thrust/system/detail/sequential/scan.h +0 -2
  693. cuda/cccl/headers/include/thrust/system/detail/sequential/scan_by_key.h +0 -2
  694. cuda/cccl/headers/include/thrust/system/detail/sequential/set_operations.h +0 -2
  695. cuda/cccl/headers/include/thrust/system/detail/sequential/sort.h +67 -6
  696. cuda/cccl/headers/include/thrust/system/detail/sequential/stable_merge_sort.h +310 -11
  697. cuda/cccl/headers/include/thrust/system/detail/sequential/stable_primitive_sort.h +78 -5
  698. cuda/cccl/headers/include/thrust/system/detail/sequential/stable_radix_sort.h +543 -7
  699. cuda/cccl/headers/include/thrust/system/detail/sequential/trivial_copy.h +0 -2
  700. cuda/cccl/headers/include/thrust/system/detail/sequential/unique.h +0 -2
  701. cuda/cccl/headers/include/thrust/system/detail/sequential/unique_by_key.h +0 -2
  702. cuda/cccl/headers/include/thrust/system/detail/system_error.inl +0 -2
  703. cuda/cccl/headers/include/thrust/system/error_code.h +0 -4
  704. cuda/cccl/headers/include/thrust/system/omp/detail/adjacent_difference.h +5 -25
  705. cuda/cccl/headers/include/thrust/system/omp/detail/assign_value.h +2 -15
  706. cuda/cccl/headers/include/thrust/system/omp/detail/binary_search.h +5 -25
  707. cuda/cccl/headers/include/thrust/system/omp/detail/copy.h +40 -29
  708. cuda/cccl/headers/include/thrust/system/omp/detail/copy_if.h +11 -28
  709. cuda/cccl/headers/include/thrust/system/omp/detail/count.h +2 -15
  710. cuda/cccl/headers/include/thrust/system/omp/detail/default_decomposition.h +26 -28
  711. cuda/cccl/headers/include/thrust/system/omp/detail/equal.h +2 -15
  712. cuda/cccl/headers/include/thrust/system/omp/detail/execution_policy.h +18 -13
  713. cuda/cccl/headers/include/thrust/system/omp/detail/extrema.h +5 -25
  714. cuda/cccl/headers/include/thrust/system/omp/detail/fill.h +2 -15
  715. cuda/cccl/headers/include/thrust/system/omp/detail/find.h +5 -25
  716. cuda/cccl/headers/include/thrust/system/omp/detail/for_each.h +47 -30
  717. cuda/cccl/headers/include/thrust/system/omp/detail/gather.h +2 -15
  718. cuda/cccl/headers/include/thrust/system/omp/detail/generate.h +2 -15
  719. cuda/cccl/headers/include/thrust/system/omp/detail/get_value.h +2 -15
  720. cuda/cccl/headers/include/thrust/system/omp/detail/inner_product.h +2 -15
  721. cuda/cccl/headers/include/thrust/system/omp/detail/iter_swap.h +2 -15
  722. cuda/cccl/headers/include/thrust/system/omp/detail/logical.h +2 -15
  723. cuda/cccl/headers/include/thrust/system/omp/detail/malloc_and_free.h +2 -15
  724. cuda/cccl/headers/include/thrust/system/omp/detail/merge.h +2 -15
  725. cuda/cccl/headers/include/thrust/system/omp/detail/mismatch.h +2 -15
  726. cuda/cccl/headers/include/thrust/system/omp/detail/partition.h +26 -31
  727. cuda/cccl/headers/include/thrust/system/omp/detail/per_device_resource.h +2 -15
  728. cuda/cccl/headers/include/thrust/system/omp/detail/pragma_omp.h +2 -26
  729. cuda/cccl/headers/include/thrust/system/omp/detail/reduce.h +35 -27
  730. cuda/cccl/headers/include/thrust/system/omp/detail/reduce_by_key.h +13 -28
  731. cuda/cccl/headers/include/thrust/system/omp/detail/reduce_intervals.h +56 -28
  732. cuda/cccl/headers/include/thrust/system/omp/detail/remove.h +26 -31
  733. cuda/cccl/headers/include/thrust/system/omp/detail/replace.h +2 -15
  734. cuda/cccl/headers/include/thrust/system/omp/detail/reverse.h +2 -15
  735. cuda/cccl/headers/include/thrust/system/omp/detail/scan.h +176 -17
  736. cuda/cccl/headers/include/thrust/system/omp/detail/scan_by_key.h +8 -15
  737. cuda/cccl/headers/include/thrust/system/omp/detail/scatter.h +2 -15
  738. cuda/cccl/headers/include/thrust/system/omp/detail/sequence.h +2 -15
  739. cuda/cccl/headers/include/thrust/system/omp/detail/set_operations.h +2 -15
  740. cuda/cccl/headers/include/thrust/system/omp/detail/sort.h +213 -28
  741. cuda/cccl/headers/include/thrust/system/omp/detail/swap_ranges.h +2 -15
  742. cuda/cccl/headers/include/thrust/system/omp/detail/tabulate.h +2 -15
  743. cuda/cccl/headers/include/thrust/system/omp/detail/temporary_buffer.h +2 -15
  744. cuda/cccl/headers/include/thrust/system/omp/detail/transform.h +2 -15
  745. cuda/cccl/headers/include/thrust/system/omp/detail/transform_reduce.h +2 -15
  746. cuda/cccl/headers/include/thrust/system/omp/detail/transform_scan.h +2 -15
  747. cuda/cccl/headers/include/thrust/system/omp/detail/uninitialized_copy.h +2 -15
  748. cuda/cccl/headers/include/thrust/system/omp/detail/uninitialized_fill.h +2 -15
  749. cuda/cccl/headers/include/thrust/system/omp/detail/unique.h +21 -30
  750. cuda/cccl/headers/include/thrust/system/omp/detail/unique_by_key.h +17 -29
  751. cuda/cccl/headers/include/thrust/system/omp/memory.h +51 -9
  752. cuda/cccl/headers/include/thrust/system/omp/memory_resource.h +3 -7
  753. cuda/cccl/headers/include/thrust/system/omp/pointer.h +3 -7
  754. cuda/cccl/headers/include/thrust/system/omp/vector.h +3 -6
  755. cuda/cccl/headers/include/thrust/system/system_error.h +0 -2
  756. cuda/cccl/headers/include/thrust/system/tbb/detail/adjacent_difference.h +4 -25
  757. cuda/cccl/headers/include/thrust/system/tbb/detail/assign_value.h +2 -15
  758. cuda/cccl/headers/include/thrust/system/tbb/detail/binary_search.h +2 -15
  759. cuda/cccl/headers/include/thrust/system/tbb/detail/copy.h +38 -29
  760. cuda/cccl/headers/include/thrust/system/tbb/detail/copy_if.h +91 -24
  761. cuda/cccl/headers/include/thrust/system/tbb/detail/count.h +2 -15
  762. cuda/cccl/headers/include/thrust/system/tbb/detail/equal.h +2 -15
  763. cuda/cccl/headers/include/thrust/system/tbb/detail/execution_policy.h +17 -13
  764. cuda/cccl/headers/include/thrust/system/tbb/detail/extrema.h +4 -25
  765. cuda/cccl/headers/include/thrust/system/tbb/detail/fill.h +2 -15
  766. cuda/cccl/headers/include/thrust/system/tbb/detail/find.h +4 -25
  767. cuda/cccl/headers/include/thrust/system/tbb/detail/for_each.h +47 -28
  768. cuda/cccl/headers/include/thrust/system/tbb/detail/gather.h +2 -15
  769. cuda/cccl/headers/include/thrust/system/tbb/detail/generate.h +2 -15
  770. cuda/cccl/headers/include/thrust/system/tbb/detail/get_value.h +2 -15
  771. cuda/cccl/headers/include/thrust/system/tbb/detail/inner_product.h +2 -15
  772. cuda/cccl/headers/include/thrust/system/tbb/detail/iter_swap.h +2 -15
  773. cuda/cccl/headers/include/thrust/system/tbb/detail/logical.h +2 -15
  774. cuda/cccl/headers/include/thrust/system/tbb/detail/malloc_and_free.h +2 -15
  775. cuda/cccl/headers/include/thrust/system/tbb/detail/merge.h +254 -29
  776. cuda/cccl/headers/include/thrust/system/tbb/detail/mismatch.h +2 -15
  777. cuda/cccl/headers/include/thrust/system/tbb/detail/partition.h +25 -31
  778. cuda/cccl/headers/include/thrust/system/tbb/detail/per_device_resource.h +2 -15
  779. cuda/cccl/headers/include/thrust/system/tbb/detail/reduce.h +95 -29
  780. cuda/cccl/headers/include/thrust/system/tbb/detail/reduce_by_key.h +345 -28
  781. cuda/cccl/headers/include/thrust/system/tbb/detail/reduce_intervals.h +4 -26
  782. cuda/cccl/headers/include/thrust/system/tbb/detail/remove.h +32 -42
  783. cuda/cccl/headers/include/thrust/system/tbb/detail/replace.h +2 -15
  784. cuda/cccl/headers/include/thrust/system/tbb/detail/reverse.h +2 -15
  785. cuda/cccl/headers/include/thrust/system/tbb/detail/scan.h +265 -30
  786. cuda/cccl/headers/include/thrust/system/tbb/detail/scan_by_key.h +7 -17
  787. cuda/cccl/headers/include/thrust/system/tbb/detail/scatter.h +2 -15
  788. cuda/cccl/headers/include/thrust/system/tbb/detail/sequence.h +2 -15
  789. cuda/cccl/headers/include/thrust/system/tbb/detail/set_operations.h +2 -15
  790. cuda/cccl/headers/include/thrust/system/tbb/detail/sort.h +244 -32
  791. cuda/cccl/headers/include/thrust/system/tbb/detail/swap_ranges.h +2 -15
  792. cuda/cccl/headers/include/thrust/system/tbb/detail/tabulate.h +2 -15
  793. cuda/cccl/headers/include/thrust/system/tbb/detail/temporary_buffer.h +2 -15
  794. cuda/cccl/headers/include/thrust/system/tbb/detail/transform.h +2 -15
  795. cuda/cccl/headers/include/thrust/system/tbb/detail/transform_reduce.h +2 -15
  796. cuda/cccl/headers/include/thrust/system/tbb/detail/transform_scan.h +2 -15
  797. cuda/cccl/headers/include/thrust/system/tbb/detail/uninitialized_copy.h +2 -15
  798. cuda/cccl/headers/include/thrust/system/tbb/detail/uninitialized_fill.h +2 -15
  799. cuda/cccl/headers/include/thrust/system/tbb/detail/unique.h +23 -33
  800. cuda/cccl/headers/include/thrust/system/tbb/detail/unique_by_key.h +16 -29
  801. cuda/cccl/headers/include/thrust/system/tbb/memory.h +52 -24
  802. cuda/cccl/headers/include/thrust/system/tbb/memory_resource.h +4 -22
  803. cuda/cccl/headers/include/thrust/system/tbb/pointer.h +4 -22
  804. cuda/cccl/headers/include/thrust/system/tbb/vector.h +4 -21
  805. cuda/cccl/headers/include/thrust/transform.h +14 -3
  806. cuda/cccl/headers/include/thrust/type_traits/integer_sequence.h +0 -4
  807. cuda/cccl/headers/include/thrust/type_traits/is_contiguous_iterator.h +0 -1
  808. cuda/cccl/headers/include/thrust/type_traits/is_operator_less_or_greater_function_object.h +0 -4
  809. cuda/cccl/headers/include/thrust/type_traits/is_operator_plus_function_object.h +0 -4
  810. cuda/cccl/headers/include/thrust/type_traits/is_trivially_relocatable.h +0 -4
  811. cuda/cccl/headers/include/thrust/universal_allocator.h +8 -0
  812. cuda/cccl/headers/include/thrust/universal_vector.h +9 -0
  813. cuda/cccl/headers/include/thrust/zip_function.h +2 -28
  814. cuda/compute/__init__.py +4 -0
  815. cuda/compute/_bindings.pyi +26 -3
  816. cuda/compute/_bindings_impl.pyx +143 -1
  817. cuda/compute/algorithms/__init__.py +9 -5
  818. cuda/compute/algorithms/_sort/__init__.py +23 -0
  819. cuda/compute/algorithms/{_merge_sort.py → _sort/_merge_sort.py} +10 -10
  820. cuda/compute/algorithms/{_radix_sort.py → _sort/_radix_sort.py} +9 -58
  821. cuda/compute/algorithms/_sort/_segmented_sort.py +288 -0
  822. cuda/compute/algorithms/_sort/_sort_common.py +52 -0
  823. cuda/compute/cu12/_bindings_impl.cpython-313-aarch64-linux-gnu.so +0 -0
  824. cuda/compute/cu12/cccl/libcccl.c.parallel.so +0 -0
  825. cuda/compute/cu13/_bindings_impl.cpython-313-aarch64-linux-gnu.so +0 -0
  826. cuda/compute/cu13/cccl/libcccl.c.parallel.so +0 -0
  827. cuda_cccl-0.3.4.dist-info/METADATA +78 -0
  828. {cuda_cccl-0.3.2.dist-info → cuda_cccl-0.3.4.dist-info}/RECORD +830 -867
  829. cuda/cccl/headers/include/cuda/__memory_resource/resource_ref.h +0 -652
  830. cuda/cccl/headers/include/cuda/std/detail/libcxx/include/tuple +0 -1365
  831. cuda/cccl/headers/include/cuda/std/detail/libcxx/include/variant +0 -2144
  832. cuda/cccl/headers/include/thrust/detail/integer_math.h +0 -113
  833. cuda/cccl/headers/include/thrust/system/detail/adl/adjacent_difference.h +0 -51
  834. cuda/cccl/headers/include/thrust/system/detail/adl/assign_value.h +0 -51
  835. cuda/cccl/headers/include/thrust/system/detail/adl/binary_search.h +0 -51
  836. cuda/cccl/headers/include/thrust/system/detail/adl/copy.h +0 -51
  837. cuda/cccl/headers/include/thrust/system/detail/adl/copy_if.h +0 -52
  838. cuda/cccl/headers/include/thrust/system/detail/adl/count.h +0 -51
  839. cuda/cccl/headers/include/thrust/system/detail/adl/equal.h +0 -51
  840. cuda/cccl/headers/include/thrust/system/detail/adl/extrema.h +0 -51
  841. cuda/cccl/headers/include/thrust/system/detail/adl/fill.h +0 -51
  842. cuda/cccl/headers/include/thrust/system/detail/adl/find.h +0 -51
  843. cuda/cccl/headers/include/thrust/system/detail/adl/for_each.h +0 -51
  844. cuda/cccl/headers/include/thrust/system/detail/adl/gather.h +0 -51
  845. cuda/cccl/headers/include/thrust/system/detail/adl/generate.h +0 -51
  846. cuda/cccl/headers/include/thrust/system/detail/adl/get_value.h +0 -51
  847. cuda/cccl/headers/include/thrust/system/detail/adl/inner_product.h +0 -51
  848. cuda/cccl/headers/include/thrust/system/detail/adl/iter_swap.h +0 -51
  849. cuda/cccl/headers/include/thrust/system/detail/adl/logical.h +0 -51
  850. cuda/cccl/headers/include/thrust/system/detail/adl/malloc_and_free.h +0 -51
  851. cuda/cccl/headers/include/thrust/system/detail/adl/merge.h +0 -51
  852. cuda/cccl/headers/include/thrust/system/detail/adl/mismatch.h +0 -51
  853. cuda/cccl/headers/include/thrust/system/detail/adl/partition.h +0 -51
  854. cuda/cccl/headers/include/thrust/system/detail/adl/per_device_resource.h +0 -51
  855. cuda/cccl/headers/include/thrust/system/detail/adl/reduce.h +0 -51
  856. cuda/cccl/headers/include/thrust/system/detail/adl/reduce_by_key.h +0 -51
  857. cuda/cccl/headers/include/thrust/system/detail/adl/remove.h +0 -51
  858. cuda/cccl/headers/include/thrust/system/detail/adl/replace.h +0 -51
  859. cuda/cccl/headers/include/thrust/system/detail/adl/reverse.h +0 -51
  860. cuda/cccl/headers/include/thrust/system/detail/adl/scan.h +0 -51
  861. cuda/cccl/headers/include/thrust/system/detail/adl/scan_by_key.h +0 -51
  862. cuda/cccl/headers/include/thrust/system/detail/adl/scatter.h +0 -51
  863. cuda/cccl/headers/include/thrust/system/detail/adl/sequence.h +0 -51
  864. cuda/cccl/headers/include/thrust/system/detail/adl/set_operations.h +0 -51
  865. cuda/cccl/headers/include/thrust/system/detail/adl/sort.h +0 -51
  866. cuda/cccl/headers/include/thrust/system/detail/adl/swap_ranges.h +0 -51
  867. cuda/cccl/headers/include/thrust/system/detail/adl/tabulate.h +0 -51
  868. cuda/cccl/headers/include/thrust/system/detail/adl/temporary_buffer.h +0 -51
  869. cuda/cccl/headers/include/thrust/system/detail/adl/transform.h +0 -51
  870. cuda/cccl/headers/include/thrust/system/detail/adl/transform_reduce.h +0 -51
  871. cuda/cccl/headers/include/thrust/system/detail/adl/transform_scan.h +0 -51
  872. cuda/cccl/headers/include/thrust/system/detail/adl/uninitialized_copy.h +0 -51
  873. cuda/cccl/headers/include/thrust/system/detail/adl/uninitialized_fill.h +0 -51
  874. cuda/cccl/headers/include/thrust/system/detail/adl/unique.h +0 -51
  875. cuda/cccl/headers/include/thrust/system/detail/adl/unique_by_key.h +0 -51
  876. cuda/cccl/headers/include/thrust/system/detail/generic/scan.inl +0 -85
  877. cuda/cccl/headers/include/thrust/system/detail/sequential/copy.inl +0 -119
  878. cuda/cccl/headers/include/thrust/system/detail/sequential/merge.inl +0 -145
  879. cuda/cccl/headers/include/thrust/system/detail/sequential/sort.inl +0 -116
  880. cuda/cccl/headers/include/thrust/system/detail/sequential/stable_merge_sort.inl +0 -356
  881. cuda/cccl/headers/include/thrust/system/detail/sequential/stable_primitive_sort.inl +0 -124
  882. cuda/cccl/headers/include/thrust/system/detail/sequential/stable_radix_sort.inl +0 -586
  883. cuda/cccl/headers/include/thrust/system/omp/detail/copy.inl +0 -74
  884. cuda/cccl/headers/include/thrust/system/omp/detail/copy_if.inl +0 -59
  885. cuda/cccl/headers/include/thrust/system/omp/detail/default_decomposition.inl +0 -65
  886. cuda/cccl/headers/include/thrust/system/omp/detail/for_each.inl +0 -87
  887. cuda/cccl/headers/include/thrust/system/omp/detail/memory.inl +0 -93
  888. cuda/cccl/headers/include/thrust/system/omp/detail/partition.inl +0 -102
  889. cuda/cccl/headers/include/thrust/system/omp/detail/reduce.inl +0 -78
  890. cuda/cccl/headers/include/thrust/system/omp/detail/reduce_by_key.inl +0 -65
  891. cuda/cccl/headers/include/thrust/system/omp/detail/reduce_intervals.inl +0 -103
  892. cuda/cccl/headers/include/thrust/system/omp/detail/remove.inl +0 -87
  893. cuda/cccl/headers/include/thrust/system/omp/detail/sort.inl +0 -265
  894. cuda/cccl/headers/include/thrust/system/omp/detail/unique.inl +0 -71
  895. cuda/cccl/headers/include/thrust/system/omp/detail/unique_by_key.inl +0 -75
  896. cuda/cccl/headers/include/thrust/system/tbb/detail/copy.inl +0 -73
  897. cuda/cccl/headers/include/thrust/system/tbb/detail/copy_if.inl +0 -136
  898. cuda/cccl/headers/include/thrust/system/tbb/detail/for_each.inl +0 -91
  899. cuda/cccl/headers/include/thrust/system/tbb/detail/memory.inl +0 -94
  900. cuda/cccl/headers/include/thrust/system/tbb/detail/merge.inl +0 -327
  901. cuda/cccl/headers/include/thrust/system/tbb/detail/partition.inl +0 -98
  902. cuda/cccl/headers/include/thrust/system/tbb/detail/reduce.inl +0 -137
  903. cuda/cccl/headers/include/thrust/system/tbb/detail/reduce_by_key.inl +0 -400
  904. cuda/cccl/headers/include/thrust/system/tbb/detail/remove.inl +0 -87
  905. cuda/cccl/headers/include/thrust/system/tbb/detail/scan.inl +0 -312
  906. cuda/cccl/headers/include/thrust/system/tbb/detail/sort.inl +0 -295
  907. cuda/cccl/headers/include/thrust/system/tbb/detail/unique.inl +0 -71
  908. cuda/cccl/headers/include/thrust/system/tbb/detail/unique_by_key.inl +0 -75
  909. cuda_cccl-0.3.2.dist-info/METADATA +0 -42
  910. {cuda_cccl-0.3.2.dist-info → cuda_cccl-0.3.4.dist-info}/WHEEL +0 -0
  911. {cuda_cccl-0.3.2.dist-info → cuda_cccl-0.3.4.dist-info}/licenses/LICENSE +0 -0
@@ -1,652 +0,0 @@
1
- //===----------------------------------------------------------------------===//
2
- //
3
- // Part of libcu++, the C++ Standard Library for your entire system,
4
- // under the Apache License v2.0 with LLVM Exceptions.
5
- // See https://llvm.org/LICENSE.txt for license information.
6
- // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7
- // SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
8
- //
9
- //===----------------------------------------------------------------------===//
10
-
11
- #ifndef _CUDA__MEMORY_RESOURCE_RESOURCE_REF_H
12
- #define _CUDA__MEMORY_RESOURCE_RESOURCE_REF_H
13
-
14
- #include <cuda/std/detail/__config>
15
-
16
- #if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
17
- # pragma GCC system_header
18
- #elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
19
- # pragma clang system_header
20
- #elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
21
- # pragma system_header
22
- #endif // no system header
23
-
24
- #if defined(LIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE)
25
-
26
- # include <cuda/__memory_resource/get_property.h>
27
- # include <cuda/__memory_resource/properties.h>
28
- # include <cuda/__memory_resource/resource.h>
29
- # include <cuda/__stream/stream_ref.h>
30
- # include <cuda/std/__concepts/concept_macros.h>
31
- # include <cuda/std/__memory/addressof.h>
32
- # include <cuda/std/__type_traits/is_base_of.h>
33
- # include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
34
- # include <cuda/std/__type_traits/type_set.h>
35
- # include <cuda/std/__utility/exchange.h>
36
- # include <cuda/std/__utility/move.h>
37
- # include <cuda/std/cstddef>
38
-
39
- # include <cuda/std/__cccl/prologue.h>
40
-
41
- _CCCL_BEGIN_NAMESPACE_CUDA_MR
42
-
43
- union _AnyResourceStorage
44
- {
45
- _CCCL_API constexpr _AnyResourceStorage(void* __ptr = nullptr) noexcept
46
- : __ptr_(__ptr)
47
- {}
48
-
49
- void* __ptr_;
50
- char __buf_[3 * sizeof(void*)];
51
- };
52
-
53
- template <class _Resource>
54
- constexpr bool _IsSmall() noexcept
55
- {
56
- return (sizeof(_Resource) <= sizeof(_AnyResourceStorage)) //
57
- && (alignof(_AnyResourceStorage) % alignof(_Resource) == 0)
58
- && ::cuda::std::is_nothrow_move_constructible_v<_Resource>;
59
- }
60
-
61
- template <class _Resource>
62
- constexpr _Resource* _Any_resource_cast(_AnyResourceStorage* __object) noexcept
63
- {
64
- return static_cast<_Resource*>(_IsSmall<_Resource>() ? __object->__buf_ : __object->__ptr_);
65
- }
66
-
67
- template <class _Resource>
68
- constexpr const _Resource* _Any_resource_cast(const _AnyResourceStorage* __object) noexcept
69
- {
70
- return static_cast<const _Resource*>(_IsSmall<_Resource>() ? __object->__buf_ : __object->__ptr_);
71
- }
72
-
73
- enum class _WrapperType
74
- {
75
- _Reference,
76
- _Owning
77
- };
78
-
79
- enum class _AllocType
80
- {
81
- _Synchronous,
82
- _Default,
83
- };
84
-
85
- struct _Alloc_vtable
86
- {
87
- using _AllocFn = void* (*) (void*, size_t, size_t);
88
- using _DeallocFn = void (*)(void*, void*, size_t, size_t) noexcept;
89
- using _EqualFn = bool (*)(void*, void*);
90
- using _DestroyFn = void (*)(_AnyResourceStorage*) noexcept;
91
- using _MoveFn = void (*)(_AnyResourceStorage*, _AnyResourceStorage*) noexcept;
92
- using _CopyFn = void (*)(_AnyResourceStorage*, const _AnyResourceStorage*);
93
-
94
- bool __is_small;
95
- _AllocFn __alloc_fn;
96
- _DeallocFn __dealloc_fn;
97
- _EqualFn __equal_fn;
98
- _DestroyFn __destroy_fn;
99
- _MoveFn __move_fn;
100
- _CopyFn __copy_fn;
101
-
102
- constexpr _Alloc_vtable(
103
- bool __is_small_,
104
- _AllocFn __alloc_fn_,
105
- _DeallocFn __dealloc_fn_,
106
- _EqualFn __equal_fn_,
107
- _DestroyFn __destroy_fn_,
108
- _MoveFn __move_fn_,
109
- _CopyFn __copy_fn_) noexcept
110
- : __is_small(__is_small_)
111
- , __alloc_fn(__alloc_fn_)
112
- , __dealloc_fn(__dealloc_fn_)
113
- , __equal_fn(__equal_fn_)
114
- , __destroy_fn(__destroy_fn_)
115
- , __move_fn(__move_fn_)
116
- , __copy_fn(__copy_fn_)
117
- {}
118
- };
119
-
120
- struct _Async_alloc_vtable : public _Alloc_vtable
121
- {
122
- using _AsyncAllocFn = void* (*) (void*, size_t, size_t, ::cuda::stream_ref);
123
- using _AsyncDeallocFn = void (*)(void*, void*, size_t, size_t, ::cuda::stream_ref);
124
-
125
- _AsyncAllocFn __async_alloc_fn;
126
- _AsyncDeallocFn __async_dealloc_fn;
127
-
128
- constexpr _Async_alloc_vtable(
129
- bool __is_small_,
130
- _Alloc_vtable::_AllocFn __alloc_fn_,
131
- _Alloc_vtable::_DeallocFn __dealloc_fn_,
132
- _Alloc_vtable::_EqualFn __equal_fn_,
133
- _Alloc_vtable::_DestroyFn __destroy_fn_,
134
- _Alloc_vtable::_MoveFn __move_fn_,
135
- _Alloc_vtable::_CopyFn __copy_fn_,
136
- _AsyncAllocFn __async_alloc_fn_,
137
- _AsyncDeallocFn __async_dealloc_fn_) noexcept
138
- : _Alloc_vtable(__is_small_, __alloc_fn_, __dealloc_fn_, __equal_fn_, __destroy_fn_, __move_fn_, __copy_fn_)
139
- , __async_alloc_fn(__async_alloc_fn_)
140
- , __async_dealloc_fn(__async_dealloc_fn_)
141
- {}
142
- };
143
-
144
- struct _Resource_vtable_builder
145
- {
146
- template <_WrapperType _Wrapper_type>
147
- using __wrapper_type = ::cuda::std::integral_constant<_WrapperType, _Wrapper_type>;
148
-
149
- template <class _Resource, class _Property>
150
- static __property_value_t<_Property> _Get_property(const void* __res) noexcept
151
- {
152
- return get_property(*static_cast<const _Resource*>(__res), _Property{});
153
- }
154
-
155
- template <class _Resource>
156
- static void* _Alloc(void* __object, size_t __bytes, size_t __alignment)
157
- {
158
- return static_cast<_Resource*>(__object)->allocate_sync(__bytes, __alignment);
159
- }
160
-
161
- template <class _Resource>
162
- static void _Dealloc(void* __object, void* __ptr, size_t __bytes, size_t __alignment) noexcept
163
- {
164
- static_assert(noexcept(static_cast<_Resource*>(__object)->deallocate_sync(__ptr, __bytes, __alignment)));
165
- return static_cast<_Resource*>(__object)->deallocate_sync(__ptr, __bytes, __alignment);
166
- }
167
-
168
- template <class _Resource>
169
- static void* _Alloc_async(void* __object, size_t __bytes, size_t __alignment, ::cuda::stream_ref __stream)
170
- {
171
- return static_cast<_Resource*>(__object)->allocate(__stream, __bytes, __alignment);
172
- }
173
-
174
- template <class _Resource>
175
- static void
176
- _Dealloc_async(void* __object, void* __ptr, size_t __bytes, size_t __alignment, ::cuda::stream_ref __stream) noexcept
177
- {
178
- static_assert(noexcept(static_cast<_Resource*>(__object)->deallocate(__stream, __ptr, __bytes, __alignment)));
179
- return static_cast<_Resource*>(__object)->deallocate(__stream, __ptr, __bytes, __alignment);
180
- }
181
-
182
- template <class _Resource>
183
- static bool _Equal(void* __left, void* __rhs)
184
- {
185
- return *static_cast<_Resource*>(__left) == *static_cast<_Resource*>(__rhs);
186
- }
187
-
188
- template <class _Resource>
189
- static void _Destroy_impl(_AnyResourceStorage* __object_, __wrapper_type<_WrapperType::_Owning>) noexcept
190
- {
191
- _Resource* __object = _Any_resource_cast<_Resource>(__object_);
192
- if constexpr (_IsSmall<_Resource>())
193
- {
194
- __object->~_Resource();
195
- }
196
- else
197
- {
198
- delete __object;
199
- }
200
- }
201
-
202
- template <class _Resource>
203
- static void _Destroy_impl(_AnyResourceStorage*, __wrapper_type<_WrapperType::_Reference>) noexcept
204
- {}
205
-
206
- template <class _Resource, _WrapperType _Wrapper_type>
207
- static void _Destroy(_AnyResourceStorage* __object) noexcept
208
- {
209
- _Destroy_impl<_Resource>(__object, __wrapper_type<_Wrapper_type>{});
210
- }
211
-
212
- template <class _Resource>
213
- static void _Move_impl(
214
- _AnyResourceStorage* __object, _AnyResourceStorage* __other_, __wrapper_type<_WrapperType::_Owning>) noexcept
215
- {
216
- if constexpr (_IsSmall<_Resource>())
217
- {
218
- _Resource* __other = _Any_resource_cast<_Resource>(__other_);
219
- ::new (static_cast<void*>(__object->__buf_)) _Resource(::cuda::std::move(*__other));
220
- __other->~_Resource();
221
- }
222
- else
223
- {
224
- __object->__ptr_ = ::cuda::std::exchange(__other_->__ptr_, nullptr);
225
- }
226
- }
227
-
228
- template <class _Resource>
229
- static void _Move_impl(_AnyResourceStorage*, _AnyResourceStorage*, __wrapper_type<_WrapperType::_Reference>) noexcept
230
- {}
231
-
232
- template <class _Resource, _WrapperType _Wrapper_type>
233
- static void _Move(_AnyResourceStorage* __object, _AnyResourceStorage* __other) noexcept
234
- {
235
- _Move_impl<_Resource>(__object, __other, __wrapper_type<_Wrapper_type>{});
236
- }
237
-
238
- template <class _Resource>
239
- static void _Copy_impl(
240
- _AnyResourceStorage* __object, const _AnyResourceStorage* __other, __wrapper_type<_WrapperType::_Owning>) noexcept
241
- {
242
- if constexpr (_IsSmall<_Resource>())
243
- {
244
- ::new (static_cast<void*>(__object->__buf_)) _Resource(*_Any_resource_cast<_Resource>(__other));
245
- }
246
- else
247
- {
248
- __object->__ptr_ = new _Resource(*_Any_resource_cast<_Resource>(__other));
249
- }
250
- }
251
-
252
- template <class _Resource>
253
- static void
254
- _Copy_impl(_AnyResourceStorage*, const _AnyResourceStorage*, __wrapper_type<_WrapperType::_Reference>) noexcept
255
- {}
256
-
257
- template <class _Resource, _WrapperType _Wrapper_type>
258
- static void _Copy(_AnyResourceStorage* __object, const _AnyResourceStorage* __other)
259
- {
260
- _Copy_impl<_Resource>(__object, __other, __wrapper_type<_Wrapper_type>{});
261
- }
262
-
263
- _CCCL_TEMPLATE(class _Resource, _AllocType _Alloc_type, _WrapperType _Wrapper_type)
264
- _CCCL_REQUIRES((_Alloc_type == _AllocType::_Synchronous))
265
- static constexpr _Alloc_vtable _Create() noexcept
266
- {
267
- return {_IsSmall<_Resource>(),
268
- &_Resource_vtable_builder::_Alloc<_Resource>,
269
- &_Resource_vtable_builder::_Dealloc<_Resource>,
270
- &_Resource_vtable_builder::_Equal<_Resource>,
271
- &_Resource_vtable_builder::_Destroy<_Resource, _Wrapper_type>,
272
- &_Resource_vtable_builder::_Move<_Resource, _Wrapper_type>,
273
- &_Resource_vtable_builder::_Copy<_Resource, _Wrapper_type>};
274
- }
275
-
276
- _CCCL_TEMPLATE(class _Resource, _AllocType _Alloc_type, _WrapperType _Wrapper_type)
277
- _CCCL_REQUIRES((_Alloc_type == _AllocType::_Default))
278
- static constexpr _Async_alloc_vtable _Create() noexcept
279
- {
280
- return {_IsSmall<_Resource>(),
281
- &_Resource_vtable_builder::_Alloc<_Resource>,
282
- &_Resource_vtable_builder::_Dealloc<_Resource>,
283
- &_Resource_vtable_builder::_Equal<_Resource>,
284
- &_Resource_vtable_builder::_Destroy<_Resource, _Wrapper_type>,
285
- &_Resource_vtable_builder::_Move<_Resource, _Wrapper_type>,
286
- &_Resource_vtable_builder::_Copy<_Resource, _Wrapper_type>,
287
- &_Resource_vtable_builder::_Alloc_async<_Resource>,
288
- &_Resource_vtable_builder::_Dealloc_async<_Resource>};
289
- }
290
- };
291
-
292
- template <class _Property>
293
- using __property_fn_t = __property_value_t<_Property> (*)(const void*);
294
-
295
- template <class _Property>
296
- struct _Property_vtable
297
- {
298
- __property_fn_t<_Property> __property_fn = nullptr;
299
-
300
- constexpr _Property_vtable(__property_fn_t<_Property> __property_fn_) noexcept
301
- : __property_fn(__property_fn_)
302
- {}
303
- };
304
-
305
- template <_AllocType _Alloc_type, class... _Properties>
306
- class basic_resource_ref;
307
-
308
- template <class... _Properties>
309
- struct _Resource_vtable : public _Property_vtable<_Properties>...
310
- {
311
- constexpr _Resource_vtable(__property_fn_t<_Properties>... __property_fn_) noexcept
312
- : _Property_vtable<_Properties>(__property_fn_)...
313
- {}
314
-
315
- template <class... _OtherProperties>
316
- constexpr _Resource_vtable(const _Resource_vtable<_OtherProperties...>& __other) noexcept
317
- : _Property_vtable<_Properties>(__other._Property_vtable<_Properties>::__property_fn)...
318
- {}
319
-
320
- template <class _Resource>
321
- static constexpr _Resource_vtable _Create() noexcept
322
- {
323
- return {&_Resource_vtable_builder::_Get_property<_Resource, _Properties>...};
324
- }
325
- };
326
-
327
- template <class... _Properties>
328
- struct _Filtered;
329
-
330
- template <bool _IsUniqueProperty>
331
- struct _Property_filter
332
- {
333
- template <class _Property, class... _Properties>
334
- using _Filtered_properties =
335
- typename _Filtered<_Properties...>::_Filtered_vtable::template _Append_property<_Property>;
336
- };
337
-
338
- template <>
339
- struct _Property_filter<false>
340
- {
341
- template <class _Property, class... _Properties>
342
- using _Filtered_properties = typename _Filtered<_Properties...>::_Filtered_vtable;
343
- };
344
-
345
- template <class _Property, class... _Properties>
346
- struct _Filtered<_Property, _Properties...>
347
- {
348
- using _Filtered_vtable = typename _Property_filter<
349
- property_with_value<_Property> && !::cuda::std::__is_included_in_v<_Property, _Properties...>>::
350
- template _Filtered_properties<_Property, _Properties...>;
351
-
352
- template <class _OtherPropery>
353
- using _Append_property = _Filtered<_OtherPropery, _Property, _Properties...>;
354
-
355
- using _Vtable = _Resource_vtable<_Property, _Properties...>;
356
- };
357
-
358
- template <>
359
- struct _Filtered<>
360
- {
361
- using _Filtered_vtable = _Filtered<>;
362
-
363
- template <class _OtherPropery>
364
- using _Append_property = _Filtered<_OtherPropery>;
365
-
366
- using _Vtable = _Resource_vtable<>;
367
- };
368
-
369
- template <class... _Properties>
370
- using _Filtered_vtable = typename _Filtered<_Properties...>::_Filtered_vtable::_Vtable;
371
-
372
- template <_WrapperType _Wrapper_type>
373
- using __alloc_object_storage_t =
374
- ::cuda::std::_If<_Wrapper_type == _WrapperType::_Reference, void*, _AnyResourceStorage>;
375
-
376
- struct _Resource_ref_base
377
- {};
378
-
379
- template <class _Vtable, _WrapperType _Wrapper_type>
380
- struct _CCCL_DECLSPEC_EMPTY_BASES _Alloc_base : _Resource_ref_base
381
- {
382
- static_assert(::cuda::std::is_base_of_v<_Alloc_vtable, _Vtable>, "");
383
-
384
- _Alloc_base(void* __object_, const _Vtable* __static_vtabl_) noexcept
385
- : __object(__object_)
386
- , __static_vtable(__static_vtabl_)
387
- {}
388
-
389
- [[nodiscard]] void* allocate_sync(size_t __bytes, size_t __alignment = alignof(::cuda::std::max_align_t))
390
- {
391
- return __static_vtable->__alloc_fn(_Get_object(), __bytes, __alignment);
392
- }
393
-
394
- void deallocate_sync(void* _Ptr, size_t __bytes, size_t __alignment = alignof(::cuda::std::max_align_t)) noexcept
395
- {
396
- __static_vtable->__dealloc_fn(_Get_object(), _Ptr, __bytes, __alignment);
397
- }
398
-
399
- protected:
400
- static _CCCL_FORCEINLINE void* _Get_object_(bool, void* __object) noexcept
401
- {
402
- return __object;
403
- }
404
-
405
- static _CCCL_FORCEINLINE void* _Get_object_(const bool __is_small, const _AnyResourceStorage& __object) noexcept
406
- {
407
- const void* __pv = __is_small ? __object.__buf_ : __object.__ptr_;
408
- return const_cast<void*>(__pv);
409
- }
410
-
411
- void* _Get_object() const noexcept
412
- {
413
- return _Get_object_(__static_vtable->__is_small, this->__object);
414
- }
415
-
416
- __alloc_object_storage_t<_Wrapper_type> __object{};
417
- const _Vtable* __static_vtable = nullptr;
418
- };
419
-
420
- template <class _Vtable, _WrapperType _Wrapper_type>
421
- struct _Async_alloc_base : public _Alloc_base<_Vtable, _Wrapper_type>
422
- {
423
- static_assert(::cuda::std::is_base_of_v<_Async_alloc_vtable, _Vtable>, "");
424
-
425
- _Async_alloc_base(void* __object_, const _Vtable* __static_vtabl_) noexcept
426
- : _Alloc_base<_Vtable, _Wrapper_type>(__object_, __static_vtabl_)
427
- {}
428
-
429
- [[nodiscard]] void* allocate(::cuda::stream_ref __stream, size_t __bytes, size_t __alignment)
430
- {
431
- return this->__static_vtable->__async_alloc_fn(this->_Get_object(), __bytes, __alignment, __stream);
432
- }
433
-
434
- [[nodiscard]] void* allocate(::cuda::stream_ref __stream, size_t __bytes)
435
- {
436
- return this->__static_vtable->__async_alloc_fn(this->_Get_object(), __bytes, alignof(max_align_t), __stream);
437
- }
438
-
439
- void deallocate(::cuda::stream_ref __stream, void* _Ptr, size_t __bytes)
440
- {
441
- this->__static_vtable->__async_dealloc_fn(this->_Get_object(), _Ptr, __bytes, alignof(max_align_t), __stream);
442
- }
443
-
444
- void deallocate(::cuda::stream_ref __stream, void* _Ptr, size_t __bytes, size_t __alignment)
445
- {
446
- this->__static_vtable->__async_dealloc_fn(this->_Get_object(), _Ptr, __bytes, __alignment, __stream);
447
- }
448
- };
449
-
450
- template <class _Resource>
451
- _CCCL_CONCEPT _Is_resource_ref = ::cuda::std::convertible_to<_Resource&, _Resource_ref_base>;
452
-
453
- template <_AllocType _Alloc_type, _WrapperType _Wrapper_type>
454
- using _Resource_base =
455
- ::cuda::std::_If<_Alloc_type == _AllocType::_Synchronous,
456
- _Alloc_base<_Alloc_vtable, _Wrapper_type>,
457
- _Async_alloc_base<_Async_alloc_vtable, _Wrapper_type>>;
458
-
459
- template <_AllocType _Alloc_type>
460
- using _Vtable_store = ::cuda::std::_If<_Alloc_type == _AllocType::_Synchronous, _Alloc_vtable, _Async_alloc_vtable>;
461
-
462
- template <_AllocType _Alloc_type, _WrapperType _Wrapper_type, class _Resource>
463
- inline constexpr _Vtable_store<_Alloc_type> __alloc_vtable =
464
- _Resource_vtable_builder::template _Create<_Resource, _Alloc_type, _Wrapper_type>();
465
-
466
- struct _Resource_ref_helper
467
- {
468
- //! This is used from \c basic_any_resource to make it convertible to a \c basic_resource_ref
469
- template <_AllocType _Alloc_type, class... _Properties>
470
- static basic_resource_ref<_Alloc_type, _Properties...>
471
- _Construct(void* __object,
472
- const _Vtable_store<_Alloc_type>* __static_vtable,
473
- _Filtered_vtable<_Properties...> __properties) noexcept
474
- {
475
- return basic_resource_ref<_Alloc_type, _Properties...>(__object, __static_vtable, __properties);
476
- }
477
- };
478
-
479
- template <_AllocType _Alloc_type, class... _Properties>
480
- class basic_resource_ref
481
- : public _Resource_base<_Alloc_type, _WrapperType::_Reference>
482
- , private _Filtered_vtable<_Properties...>
483
- {
484
- private:
485
- static_assert(__contains_execution_space_property<_Properties...>,
486
- "The properties of cuda::mr::basic_resource_ref must contain at least one execution space property!");
487
-
488
- template <_AllocType, class...>
489
- friend class basic_resource_ref;
490
-
491
- template <class...>
492
- friend struct _Resource_vtable;
493
-
494
- friend struct _Resource_ref_helper;
495
-
496
- using __vtable = _Filtered_vtable<_Properties...>;
497
-
498
- //! @brief Checks whether \c _OtherProperties is a true superset of \c _Properties, accounting for host_accessible
499
- template <class... _OtherProperties>
500
- static constexpr bool __properties_match =
501
- ::cuda::std::__type_set_contains_v<::cuda::std::__make_type_set<_OtherProperties...>, _Properties...>;
502
-
503
- //! @brief Constructs a \c basic_resource_ref from a void*, a resource vtable ptr, and a vtable
504
- //! for the properties. This is used to create a \c basic_resource_ref from a \c basic_any_resource.
505
- explicit basic_resource_ref(
506
- void* __object_, const _Vtable_store<_Alloc_type>* __static_vtable, __vtable __properties) noexcept
507
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(__object_, __static_vtable)
508
- , __vtable(__properties)
509
- {}
510
-
511
- public:
512
- //! @brief Constructs a \c basic_resource_ref from a type that satisfies the \c synchronous_resource concept
513
- //! as well as all properties
514
- //! @param __res The resource to be wrapped within the \c basic_resource_ref
515
- _CCCL_TEMPLATE(class _Resource, _AllocType _Alloc_type2 = _Alloc_type)
516
- _CCCL_REQUIRES((!_Is_resource_ref<_Resource>) _CCCL_AND(_Alloc_type2 == _AllocType::_Synchronous)
517
- _CCCL_AND synchronous_resource_with<_Resource, _Properties...>)
518
- basic_resource_ref(_Resource& __res) noexcept
519
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(
520
- ::cuda::std::addressof(__res), &__alloc_vtable<_Alloc_type, _WrapperType::_Reference, _Resource>)
521
- , __vtable(__vtable::template _Create<_Resource>())
522
- {}
523
-
524
- //! @brief Constructs a \c synchronous_resource_ref from a type that satisfies the \c resource concept as well as all
525
- //! properties. This ignores the async interface of the passed in resource
526
- //! @param __res The resource to be wrapped within the \c synchronous_resource_ref
527
- _CCCL_TEMPLATE(class _Resource, _AllocType _Alloc_type2 = _Alloc_type)
528
- _CCCL_REQUIRES((!_Is_resource_ref<_Resource>) _CCCL_AND(_Alloc_type2 == _AllocType::_Default)
529
- _CCCL_AND resource_with<_Resource, _Properties...>)
530
- basic_resource_ref(_Resource& __res) noexcept
531
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(
532
- ::cuda::std::addressof(__res), &__alloc_vtable<_Alloc_type, _WrapperType::_Reference, _Resource>)
533
- , __vtable(__vtable::template _Create<_Resource>())
534
- {}
535
-
536
- //! @brief Constructs a \c basic_resource_ref from a type that satisfies the \c synchronous_resource concept
537
- //! as well as all properties
538
- //! @param __res Pointer to a resource to be wrapped within the \c basic_resource_ref
539
- _CCCL_TEMPLATE(class _Resource, _AllocType _Alloc_type2 = _Alloc_type)
540
- _CCCL_REQUIRES((!_Is_resource_ref<_Resource>) _CCCL_AND(_Alloc_type2 == _AllocType::_Synchronous)
541
- _CCCL_AND synchronous_resource_with<_Resource, _Properties...>)
542
- basic_resource_ref(_Resource* __res) noexcept
543
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(
544
- __res, &__alloc_vtable<_Alloc_type, _WrapperType::_Reference, _Resource>)
545
- , __vtable(__vtable::template _Create<_Resource>())
546
- {}
547
-
548
- //! @brief Constructs a \c basic_resource_ref from a type that satisfies the \c resource concept as well as all
549
- //! properties. This ignores the async interface of the passed in resource
550
- //! @param __res Pointer to a resource to be wrapped within the \c basic_resource_ref
551
- _CCCL_TEMPLATE(class _Resource, _AllocType _Alloc_type2 = _Alloc_type)
552
- _CCCL_REQUIRES((!_Is_resource_ref<_Resource>) _CCCL_AND(_Alloc_type2 == _AllocType::_Default)
553
- _CCCL_AND resource_with<_Resource, _Properties...>)
554
- basic_resource_ref(_Resource* __res) noexcept
555
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(
556
- __res, &__alloc_vtable<_Alloc_type, _WrapperType::_Reference, _Resource>)
557
- , __vtable(__vtable::template _Create<_Resource>())
558
- {}
559
-
560
- //! @brief Conversion from a \c basic_resource_ref with the same set of properties but in a different order
561
- //! @param __ref The other \c basic_resource_ref
562
- _CCCL_TEMPLATE(class... _OtherProperties)
563
- _CCCL_REQUIRES(__properties_match<_OtherProperties...>)
564
- basic_resource_ref(basic_resource_ref<_Alloc_type, _OtherProperties...> __ref) noexcept
565
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(__ref.__object, __ref.__static_vtable)
566
- , __vtable(static_cast<const _Filtered_vtable<_OtherProperties...>&>(__ref))
567
- {}
568
-
569
- //! @brief Conversion from a \c resource_ref with the same set of properties but in a different order to a
570
- //! \c synchronous_resource_ref
571
- //! @param __ref The other \c resource_ref
572
- _CCCL_TEMPLATE(_AllocType _OtherAllocType, class... _OtherProperties)
573
- _CCCL_REQUIRES((_OtherAllocType == _AllocType::_Default) _CCCL_AND(_OtherAllocType != _Alloc_type)
574
- _CCCL_AND __properties_match<_OtherProperties...>)
575
- basic_resource_ref(basic_resource_ref<_OtherAllocType, _OtherProperties...> __ref) noexcept
576
- : _Resource_base<_Alloc_type, _WrapperType::_Reference>(__ref.__object, __ref.__static_vtable)
577
- , __vtable(static_cast<const _Filtered_vtable<_OtherProperties...>&>(__ref))
578
- {}
579
-
580
- //! @brief Equality comparison between two \c basic_resource_ref
581
- //! @param __lhs The first \c basic_resource_ref
582
- //! @param __rhs The second \c basic_resource_ref
583
- //! @return Checks whether both resources have the same equality function stored in their vtable and if so returns
584
- //! the result of that equality comparison. Otherwise returns false.
585
- [[nodiscard]] friend bool operator==(const basic_resource_ref& __lhs, const basic_resource_ref& __rhs)
586
- {
587
- // BUGBUG: comparing function pointers like this can lead to false negatives:
588
- return (__lhs.__static_vtable->__equal_fn == __rhs.__static_vtable->__equal_fn)
589
- && __lhs.__static_vtable->__equal_fn(__lhs.__object, __rhs.__object);
590
- }
591
-
592
- //! @overload
593
- _CCCL_TEMPLATE(class... _OtherProperties)
594
- _CCCL_REQUIRES((sizeof...(_Properties) == sizeof...(_OtherProperties))
595
- _CCCL_AND __properties_match<_OtherProperties...>)
596
- [[nodiscard]] bool operator==(const basic_resource_ref<_Alloc_type, _OtherProperties...>& __rhs) const
597
- {
598
- // BUGBUG: comparing function pointers like this can lead to false negatives:
599
- return (this->__static_vtable->__equal_fn == __rhs.__static_vtable->__equal_fn)
600
- && this->__static_vtable->__equal_fn(this->__object, __rhs.__object);
601
- }
602
-
603
- //! @brief Inequality comparison between two \c basic_resource_ref
604
- //! @param __lhs The first \c basic_resource_ref
605
- //! @param __rhs The second \c basic_resource_ref
606
- //! @return Checks whether both resources have the same equality function stored in their vtable and if so returns
607
- //! the inverse result of that equality comparison. Otherwise returns true.
608
- [[nodiscard]] friend bool operator!=(const basic_resource_ref& __lhs, const basic_resource_ref& __rhs)
609
- {
610
- return !(__lhs == __rhs);
611
- }
612
-
613
- //! @overload
614
- _CCCL_TEMPLATE(class... _OtherProperties)
615
- _CCCL_REQUIRES((sizeof...(_Properties) == sizeof...(_OtherProperties))
616
- _CCCL_AND __properties_match<_OtherProperties...>)
617
- [[nodiscard]] bool operator!=(const basic_resource_ref<_Alloc_type, _OtherProperties...>& __rhs) const
618
- {
619
- return !(*this == __rhs);
620
- }
621
-
622
- //! @brief Forwards the stateless properties
623
- _CCCL_TEMPLATE(class _Property)
624
- _CCCL_REQUIRES((!property_with_value<_Property>) _CCCL_AND ::cuda::std::__is_included_in_v<_Property, _Properties...>)
625
- friend void get_property(const basic_resource_ref&, _Property) noexcept {}
626
-
627
- //! @brief Forwards the stateful properties
628
- _CCCL_TEMPLATE(class _Property)
629
- _CCCL_REQUIRES(property_with_value<_Property> _CCCL_AND ::cuda::std::__is_included_in_v<_Property, _Properties...>)
630
- [[nodiscard]] friend __property_value_t<_Property> get_property(const basic_resource_ref& __res, _Property) noexcept
631
- {
632
- return __res._Property_vtable<_Property>::__property_fn(__res.__object);
633
- }
634
- };
635
-
636
- //! @brief Type erased wrapper around a `resource` that satisfies \tparam _Properties
637
- //! @tparam _Properties The properties that any resource wrapped within the `synchronous_resource_ref` needs to satisfy
638
- template <class... _Properties>
639
- using synchronous_resource_ref = basic_resource_ref<_AllocType::_Synchronous, _Properties...>;
640
-
641
- //! @brief Type erased wrapper around a `resource` that satisfies \tparam _Properties
642
- //! @tparam _Properties The properties that any async resource wrapped within the `resource_ref` needs to satisfy
643
- template <class... _Properties>
644
- using resource_ref = basic_resource_ref<_AllocType::_Default, _Properties...>;
645
-
646
- _CCCL_END_NAMESPACE_CUDA_MR
647
-
648
- # include <cuda/std/__cccl/epilogue.h>
649
-
650
- #endif // LIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE
651
-
652
- #endif //_CUDA__MEMORY_RESOURCE_RESOURCE_REF_H