cuda-cccl 0.1.3.2.0.dev438__cp312-cp312-manylinux_2_24_aarch64.whl → 0.3.1__cp312-cp312-manylinux_2_24_aarch64.whl

cuda/cccl/headers/include/cub/device/dispatch/kernels/reduce.cuh

@@ -47,9 +47,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace reduce
+namespace detail::reduce
 {
 
 /**
@@ -172,6 +170,10 @@ __launch_bounds__(int(ChainedPolicyT::ActivePolicy::ReducePolicy::BLOCK_THREADS)
                 AccumT,
                 TransformOpT>;
 
+  static_assert(sizeof(typename AgentReduceT::TempStorage) <= max_smem_per_block,
+                "cub::DeviceReduce ran out of CUDA shared memory, which we judged to be extremely unlikely. Please "
+                "file an issue at: https://github.com/NVIDIA/cccl/issues");
+
   // Shared memory storage
   __shared__ typename AgentReduceT::TempStorage temp_storage;
 
@@ -253,6 +255,10 @@ CUB_DETAIL_KERNEL_ATTRIBUTES __launch_bounds__(
                 AccumT,
                 TransformOpT>;
 
+  static_assert(sizeof(typename AgentReduceT::TempStorage) <= max_smem_per_block,
+                "cub::DeviceReduce ran out of CUDA shared memory, which we judged to be extremely unlikely. Please "
+                "file an issue at: https://github.com/NVIDIA/cccl/issues");
+
   // Shared memory storage
   __shared__ typename AgentReduceT::TempStorage temp_storage;
 
@@ -572,7 +578,6 @@ CUB_DETAIL_KERNEL_ATTRIBUTES __launch_bounds__(int(
   }
 }
 
-} // namespace reduce
-} // namespace detail
+} // namespace detail::reduce
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/kernels/scan.cuh

@@ -42,9 +42,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace scan
+namespace detail::scan
 {
 
 /******************************************************************************
@@ -186,7 +184,6 @@ __launch_bounds__(int(ChainedPolicyT::ActivePolicy::ScanPolicyT::BLOCK_THREADS))
   AgentScanT(temp_storage, d_in, d_out, scan_op, real_init_value).ConsumeRange(num_items, tile_state, start_tile);
 }
 
-} // namespace scan
-} // namespace detail
+} // namespace detail::scan
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/kernels/segmented_reduce.cuh

@@ -43,9 +43,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace reduce
+namespace detail::reduce
 {
 
 /// Normalize input iterator to segment offset
@@ -318,7 +316,6 @@ __launch_bounds__(int(ChainedPolicyT::ActivePolicy::ReducePolicy::BLOCK_THREADS)
   }
 }
 
-} // namespace reduce
-} // namespace detail
+} // namespace detail::reduce
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/kernels/segmented_sort.cuh

@@ -29,6 +29,56 @@ using local_segment_index_t = ::cuda::std::uint32_t;
 // Type used for total number of segments and to index within segments globally
 using global_segment_offset_t = ::cuda::std::int64_t;
 
+template <typename OffsetT, typename BeginOffsetIteratorT, typename EndOffsetIteratorT>
+struct LargeSegmentsSelectorT
+{
+  OffsetT value{};
+  BeginOffsetIteratorT d_offset_begin{};
+  EndOffsetIteratorT d_offset_end{};
+  global_segment_offset_t base_segment_offset{};
+
+#if !_CCCL_COMPILER(NVRTC)
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE
+  LargeSegmentsSelectorT(OffsetT value, BeginOffsetIteratorT d_offset_begin, EndOffsetIteratorT d_offset_end)
+      : value(value)
+      , d_offset_begin(d_offset_begin)
+      , d_offset_end(d_offset_end)
+  {}
+#endif
+
+  _CCCL_DEVICE _CCCL_FORCEINLINE bool operator()(local_segment_index_t segment_id) const
+  {
+    const OffsetT segment_size =
+      d_offset_end[base_segment_offset + segment_id] - d_offset_begin[base_segment_offset + segment_id];
+    return segment_size > value;
+  }
+};
+
+template <typename OffsetT, typename BeginOffsetIteratorT, typename EndOffsetIteratorT>
+struct SmallSegmentsSelectorT
+{
+  OffsetT value{};
+  BeginOffsetIteratorT d_offset_begin{};
+  EndOffsetIteratorT d_offset_end{};
+  global_segment_offset_t base_segment_offset{};
+
+#if !_CCCL_COMPILER(NVRTC)
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE
+  SmallSegmentsSelectorT(OffsetT value, BeginOffsetIteratorT d_offset_begin, EndOffsetIteratorT d_offset_end)
+      : value(value)
+      , d_offset_begin(d_offset_begin)
+      , d_offset_end(d_offset_end)
+  {}
+#endif
+
+  _CCCL_DEVICE _CCCL_FORCEINLINE bool operator()(local_segment_index_t segment_id) const
+  {
+    const OffsetT segment_size =
+      d_offset_end[base_segment_offset + segment_id] - d_offset_begin[base_segment_offset + segment_id];
+    return segment_size < value;
+  }
+};
+
 /**
  * @brief Fallback kernel, in case there's not enough segments to
  *        take advantage of partitioning.
@@ -89,7 +139,7 @@ __launch_bounds__(ChainedPolicyT::ActivePolicy::LargeSegmentPolicy::BLOCK_THREAD
 {
   using ActivePolicyT       = typename ChainedPolicyT::ActivePolicy;
   using LargeSegmentPolicyT = typename ActivePolicyT::LargeSegmentPolicy;
-  using MediumPolicyT       = typename ActivePolicyT::SmallAndMediumSegmentedSortPolicyT::MediumPolicyT;
+  using MediumPolicyT       = typename ActivePolicyT::MediumSegmentPolicy;
 
   const auto segment_id = static_cast<local_segment_index_t>(blockIdx.x);
   OffsetT segment_begin = d_begin_offsets[segment_id];
@@ -253,7 +303,7 @@ template <SortOrder Order,
           typename BeginOffsetIteratorT,
           typename EndOffsetIteratorT,
           typename OffsetT>
-__launch_bounds__(ChainedPolicyT::ActivePolicy::SmallAndMediumSegmentedSortPolicyT::BLOCK_THREADS)
+__launch_bounds__(ChainedPolicyT::ActivePolicy::SmallSegmentPolicy::BLOCK_THREADS)
   CUB_DETAIL_KERNEL_ATTRIBUTES void DeviceSegmentedSortKernelSmall(
     local_segment_index_t small_segments,
     local_segment_index_t medium_segments,
@@ -272,10 +322,9 @@ __launch_bounds__(ChainedPolicyT::ActivePolicy::SmallAndMediumSegmentedSortPolic
   const local_segment_index_t tid = threadIdx.x;
   const local_segment_index_t bid = blockIdx.x;
 
-  using ActivePolicyT         = typename ChainedPolicyT::ActivePolicy;
-  using SmallAndMediumPolicyT = typename ActivePolicyT::SmallAndMediumSegmentedSortPolicyT;
-  using MediumPolicyT         = typename SmallAndMediumPolicyT::MediumPolicyT;
-  using SmallPolicyT          = typename SmallAndMediumPolicyT::SmallPolicyT;
+  using ActivePolicyT = typename ChainedPolicyT::ActivePolicy;
+  using SmallPolicyT  = typename ActivePolicyT::SmallSegmentPolicy;
+  using MediumPolicyT = typename ActivePolicyT::MediumSegmentPolicy;
 
   constexpr auto threads_per_medium_segment = static_cast<local_segment_index_t>(MediumPolicyT::WARP_THREADS);
   constexpr auto threads_per_small_segment  = static_cast<local_segment_index_t>(SmallPolicyT::WARP_THREADS);
@@ -286,11 +335,9 @@ __launch_bounds__(ChainedPolicyT::ActivePolicy::SmallAndMediumSegmentedSortPolic
   using SmallAgentWarpMergeSortT =
     sub_warp_merge_sort::AgentSubWarpSort<Order == SortOrder::Descending, SmallPolicyT, KeyT, ValueT, OffsetT>;
 
-  constexpr auto segments_per_medium_block =
-    static_cast<local_segment_index_t>(SmallAndMediumPolicyT::SEGMENTS_PER_MEDIUM_BLOCK);
+  constexpr auto segments_per_medium_block = static_cast<local_segment_index_t>(MediumPolicyT::SEGMENTS_PER_BLOCK);
 
-  constexpr auto segments_per_small_block =
-    static_cast<local_segment_index_t>(SmallAndMediumPolicyT::SEGMENTS_PER_SMALL_BLOCK);
+  constexpr auto segments_per_small_block = static_cast<local_segment_index_t>(SmallPolicyT::SEGMENTS_PER_BLOCK);
 
   __shared__ union
   {

cuda/cccl/headers/include/cub/device/dispatch/kernels/transform.cuh

@@ -202,14 +202,18 @@ _CCCL_HOST_DEVICE _CCCL_CONSTEVAL auto load_store_type()
   }
 }
 
-template <typename VectorizedPolicy, typename Offset, typename F, typename RandomAccessIteratorOut, typename... InputT>
+template <typename VectorizedPolicy,
+          typename Offset,
+          typename F,
+          typename RandomAccessIteratorOut,
+          typename... RandomAccessIteratorsIn>
 _CCCL_DEVICE void transform_kernel_vectorized(
   Offset num_items,
   int num_elem_per_thread_prefetch,
   bool can_vectorize,
   F f,
   RandomAccessIteratorOut out,
-  const InputT*... ins)
+  RandomAccessIteratorsIn... ins)
 {
   constexpr int block_dim        = VectorizedPolicy::block_threads;
   constexpr int items_per_thread = VectorizedPolicy::items_per_thread_vectorized;
@@ -240,9 +244,12 @@ _CCCL_DEVICE void transform_kernel_vectorized(
   constexpr int load_store_size = VectorizedPolicy::load_store_word_size;
   using load_store_t            = decltype(load_store_type<load_store_size>());
   using output_t                = it_value_t<RandomAccessIteratorOut>;
-  using result_t                = ::cuda::std::decay_t<::cuda::std::invoke_result_t<F, const InputT&...>>;
+  using result_t = ::cuda::std::decay_t<::cuda::std::invoke_result_t<F, const it_value_t<RandomAccessIteratorsIn>&...>>;
   // picks output type size if there are no inputs
-  constexpr int element_size     = int{first_item(sizeof(InputT)..., size_of<output_t>)};
+  constexpr int element_size     = int{first_nonzero_value(
+    (sizeof(it_value_t<RandomAccessIteratorsIn>)
+     * THRUST_NS_QUALIFIER::is_contiguous_iterator_v<RandomAccessIteratorsIn>) ...,
+    size_of<output_t>)};
   constexpr int load_store_count = (items_per_thread * element_size) / load_store_size;
 
   static_assert((items_per_thread * element_size) % load_store_size == 0);
@@ -258,18 +265,35 @@ _CCCL_DEVICE void transform_kernel_vectorized(
 
   auto provide_array = [&](auto... inputs) {
     // load inputs
-    // TODO(bgruber): we could support fancy iterators for loading here as well (and only vectorize some inputs)
-    [[maybe_unused]] auto load_tile_vectorized = [&](auto* in, auto& input) {
-      auto in_vec    = reinterpret_cast<const load_store_t*>(in);
-      auto input_vec = reinterpret_cast<load_store_t*>(input.data());
-      _CCCL_PRAGMA_UNROLL_FULL()
-      for (int i = 0; i < load_store_count; ++i)
+    [[maybe_unused]] auto load_tile = [](auto in, auto& input) {
+      if constexpr (THRUST_NS_QUALIFIER::is_contiguous_iterator_v<decltype(in)>)
       {
-        input_vec[i] = in_vec[i * VectorizedPolicy::block_threads + threadIdx.x];
+        auto in_vec    = reinterpret_cast<const load_store_t*>(in) + threadIdx.x;
+        auto input_vec = reinterpret_cast<load_store_t*>(input.data());
+        _CCCL_PRAGMA_UNROLL_FULL()
+        for (int i = 0; i < load_store_count; ++i)
+        {
+          input_vec[i] = in_vec[i * VectorizedPolicy::block_threads];
+        }
+      }
+      else
+      {
+        constexpr int elems = load_store_size / element_size;
+        in += threadIdx.x * elems;
+        _CCCL_PRAGMA_UNROLL_FULL()
+        for (int i = 0; i < load_store_count; ++i)
+        {
+          _CCCL_PRAGMA_UNROLL_FULL()
+          for (int j = 0; j < elems; ++j)
+          {
+            input[i * elems + j] = in[i * elems * VectorizedPolicy::block_threads + j];
+          }
+        }
       }
     };
     _CCCL_PDL_GRID_DEPENDENCY_SYNC();
-    (load_tile_vectorized(ins, inputs), ...);
+    (load_tile(ins, inputs), ...);
+
     // Benchmarks showed up to 38% slowdown on H200 (some improvements as well), so omitted. See #5249 for details.
     // _CCCL_PDL_TRIGGER_NEXT_LAUNCH();
 
@@ -280,7 +304,7 @@ _CCCL_DEVICE void transform_kernel_vectorized(
       output[i] = f(inputs[i]...);
     }
   };
-  provide_array(uninitialized_array<InputT, items_per_thread>{}...);
+  provide_array(uninitialized_array<it_value_t<RandomAccessIteratorsIn>, items_per_thread>{}...);
 
   // write output
   if constexpr (can_vectorize_store)

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_adjacent_difference.cuh

@@ -43,9 +43,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace adjacent_difference
+namespace detail::adjacent_difference
 {
 template <typename InputIteratorT, bool MayAlias>
 struct policy_hub
@@ -64,7 +62,6 @@ struct policy_hub
 
   using MaxPolicy = Policy500;
 };
-} // namespace adjacent_difference
-} // namespace detail
+} // namespace detail::adjacent_difference
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_batch_memcpy.cuh

@@ -43,9 +43,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace batch_memcpy
+namespace detail::batch_memcpy
 {
 /**
  * Parameterizable tuning policy type for AgentBatchMemcpy
@@ -115,7 +113,6 @@ struct policy_hub
 
   using MaxPolicy = Policy700;
 };
-} // namespace batch_memcpy
-} // namespace detail
+} // namespace detail::batch_memcpy
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_for.cuh

@@ -42,9 +42,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace for_each
+namespace detail::for_each
 {
 
 struct policy_hub_t
@@ -57,7 +55,6 @@ struct policy_hub_t
   using MaxPolicy = policy_500_t;
 };
 
-} // namespace for_each
-} // namespace detail
+} // namespace detail::for_each
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_histogram.cuh

@@ -46,9 +46,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace histogram
+namespace detail::histogram
 {
 enum class primitive_sample
 {
@@ -272,7 +270,6 @@ struct policy_hub
 
   using MaxPolicy = Policy1000;
 };
-} // namespace histogram
-} // namespace detail
+} // namespace detail::histogram
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_merge.cuh

@@ -42,9 +42,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace merge
+namespace detail::merge
 {
 template <typename KeyT, typename ValueT>
 struct policy_hub
@@ -73,7 +71,6 @@ struct policy_hub
 
   using max_policy = policy600;
 };
-} // namespace merge
-} // namespace detail
+} // namespace detail::merge
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_merge_sort.cuh

@@ -62,6 +62,14 @@ struct MergeSortPolicyWrapper<StaticPolicyT, ::cuda::std::void_t<decltype(Static
   {}
 
   CUB_DEFINE_SUB_POLICY_GETTER(MergeSort);
+
+#if defined(CUB_ENABLE_POLICY_PTX_JSON)
+  _CCCL_DEVICE static constexpr auto EncodedPolicy()
+  {
+    using namespace ptx_json;
+    return object<key<"MergeSortPolicy">() = MergeSort().EncodedPolicy()>();
+  }
+#endif
 };
 
 template <typename PolicyT>

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_radix_sort.cuh

@@ -46,9 +46,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace radix
+namespace detail::radix
 {
 // sm90 default
 template <size_t KeySize, size_t ValueSize, size_t OffsetSize>
@@ -1062,7 +1060,6 @@ struct policy_hub
   using MaxPolicy = Policy1000;
 };
 
-} // namespace radix
-} // namespace detail
+} // namespace detail::radix
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh

@@ -50,9 +50,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace reduce_by_key
+namespace detail::reduce_by_key
 {
 enum class primitive_key
 {
@@ -939,7 +937,6 @@ struct policy_hub
   };
   using MaxPolicy = Policy1000;
 };
-} // namespace reduce_by_key
-} // namespace detail
+} // namespace detail::reduce_by_key
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_run_length_encode.cuh

@@ -52,9 +52,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace rle
+namespace detail::rle
 {
 enum class primitive_key
 {
@@ -670,7 +668,6 @@ struct policy_hub
   using MaxPolicy = Policy1000;
 };
 } // namespace non_trivial_runs
-} // namespace rle
-} // namespace detail
+} // namespace detail::rle
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan.cuh

@@ -53,9 +53,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace scan
+namespace detail::scan
 {
 enum class keep_rejects
 {
@@ -615,7 +613,6 @@ struct policy_hub
 
   using MaxPolicy = Policy1000;
 };
-} // namespace scan
-} // namespace detail
+} // namespace detail::scan
 
 CUB_NAMESPACE_END

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan_by_key.cuh

@@ -49,9 +49,7 @@
 
 CUB_NAMESPACE_BEGIN
 
-namespace detail
-{
-namespace scan_by_key
+namespace detail::scan_by_key
 {
 enum class primitive_accum
 {
@@ -1007,7 +1005,6 @@ struct policy_hub
 
   using MaxPolicy = Policy1000;
 };
-} // namespace scan_by_key
-} // namespace detail
+} // namespace detail::scan_by_key
 
 CUB_NAMESPACE_END