Merge branch 'main' into enh/deprecate_abi_v2_v3

NVIDIA · Jan 30, 2025 · d646310 · d646310
2 parents 827f531 + 77a6a45
commit d646310
Show file tree

Hide file tree

Showing 61 changed files with 394 additions and 2,766 deletions.
diff --git a/cub/cub/agent/agent_batch_memcpy.cuh b/cub/cub/agent/agent_batch_memcpy.cuh
@@ -642,14 +642,12 @@ private:
     TilePrefixCallbackOp<BufferOffsetT,
                          ::cuda::std::plus<>,
                          BLevBufferOffsetTileState,
-                         0,
                          typename AgentMemcpySmallBuffersPolicyT::buff_delay_constructor>;
 
   using BLevBlockScanPrefixCallbackOpT =
     TilePrefixCallbackOp<BlockOffsetT,
                          ::cuda::std::plus<>,
                          BLevBlockOffsetTileState,
-                         0,
                          typename AgentMemcpySmallBuffersPolicyT::block_delay_constructor>;
 
   //-----------------------------------------------------------------------------

diff --git a/cub/cub/agent/agent_histogram.cuh b/cub/cub/agent/agent_histogram.cuh
@@ -172,9 +172,6 @@ namespace histogram
  *
  * @tparam OffsetT
  *   Signed integer type for global offsets
- *
- * @tparam LEGACY_PTX_ARCH
- *   PTX compute capability (unused)
  */
 template <typename AgentHistogramPolicyT,
           int PRIVATIZED_SMEM_BINS,
@@ -184,8 +181,7 @@ template <typename AgentHistogramPolicyT,
           typename CounterT,
           typename PrivatizedDecodeOpT,
           typename OutputDecodeOpT,
-          typename OffsetT,
-          int LEGACY_PTX_ARCH = 0>
+          typename OffsetT>
 struct AgentHistogram
 {
   //---------------------------------------------------------------------
@@ -930,8 +926,7 @@ template <typename AgentHistogramPolicyT,
           typename CounterT,
           typename PrivatizedDecodeOpT,
           typename OutputDecodeOpT,
-          typename OffsetT,
-          int LEGACY_PTX_ARCH = 0>
+          typename OffsetT>
 using AgentHistogram CCCL_DEPRECATED_BECAUSE("This class is considered an implementation detail and the public "
                                              "interface will be removed.") =
   detail::histogram::AgentHistogram<
@@ -943,7 +938,6 @@ using AgentHistogram CCCL_DEPRECATED_BECAUSE("This class is considered an implem
     CounterT,
     PrivatizedDecodeOpT,
     OutputDecodeOpT,
-    OffsetT,
-    LEGACY_PTX_ARCH>;
+    OffsetT>;
 
 CUB_NAMESPACE_END
diff --git a/cub/cub/agent/agent_reduce_by_key.cuh b/cub/cub/agent/agent_reduce_by_key.cuh
@@ -276,7 +276,7 @@ struct AgentReduceByKey
   // Callback type for obtaining tile prefix during block scan
   using DelayConstructorT = typename AgentReduceByKeyPolicyT::detail::delay_constructor_t;
   using TilePrefixCallbackOpT =
-    TilePrefixCallbackOp<OffsetValuePairT, ReduceBySegmentOpT, ScanTileStateT, 0, DelayConstructorT>;
+    TilePrefixCallbackOp<OffsetValuePairT, ReduceBySegmentOpT, ScanTileStateT, DelayConstructorT>;
 
   // Key and value exchange types
   using KeyExchangeT   = KeyOutputT[TILE_ITEMS + 1];

diff --git a/cub/cub/agent/agent_rle.cuh b/cub/cub/agent/agent_rle.cuh
@@ -258,7 +258,7 @@ struct AgentRle
   // Callback type for obtaining tile prefix during block scan
   using DelayConstructorT = typename AgentRlePolicyT::detail::delay_constructor_t;
   using TilePrefixCallbackOpT =
-    TilePrefixCallbackOp<LengthOffsetPair, ReduceBySegmentOpT, ScanTileStateT, 0, DelayConstructorT>;
+    TilePrefixCallbackOp<LengthOffsetPair, ReduceBySegmentOpT, ScanTileStateT, DelayConstructorT>;
 
   // Warp exchange types
   using WarpExchangePairs = WarpExchange<LengthOffsetPair, ITEMS_PER_THREAD>;

diff --git a/cub/cub/agent/agent_scan.cuh b/cub/cub/agent/agent_scan.cuh
@@ -201,7 +201,7 @@ struct AgentScan
 
   // Callback type for obtaining tile prefix during block scan
   using DelayConstructorT     = typename AgentScanPolicyT::detail::delay_constructor_t;
-  using TilePrefixCallbackOpT = TilePrefixCallbackOp<AccumT, ScanOpT, ScanTileStateT, 0 /* PTX */, DelayConstructorT>;
+  using TilePrefixCallbackOpT = TilePrefixCallbackOp<AccumT, ScanOpT, ScanTileStateT, DelayConstructorT>;
 
   // Stateful BlockScan prefix callback type for managing a running total while
   // scanning consecutive tiles

diff --git a/cub/cub/agent/agent_scan_by_key.cuh b/cub/cub/agent/agent_scan_by_key.cuh
@@ -179,7 +179,7 @@ struct AgentScanByKey
 
   using DelayConstructorT = typename AgentScanByKeyPolicyT::detail::delay_constructor_t;
   using TilePrefixCallbackT =
-    TilePrefixCallbackOp<FlagValuePairT, ReduceBySegmentOpT, ScanTileStateT, 0, DelayConstructorT>;
+    TilePrefixCallbackOp<FlagValuePairT, ReduceBySegmentOpT, ScanTileStateT, DelayConstructorT>;
 
   using BlockScanT = BlockScan<FlagValuePairT, BLOCK_THREADS, AgentScanByKeyPolicyT::SCAN_ALGORITHM, 1, 1>;
 

diff --git a/cub/cub/agent/agent_segment_fixup.cuh b/cub/cub/agent/agent_segment_fixup.cuh
@@ -84,7 +84,8 @@ template <int _BLOCK_THREADS,
           BlockLoadAlgorithm _LOAD_ALGORITHM,
           CacheLoadModifier _LOAD_MODIFIER,
           BlockScanAlgorithm _SCAN_ALGORITHM>
-struct AgentSegmentFixupPolicy
+struct CCCL_DEPRECATED_BECAUSE("This class is considered an implementation detail and the public "
+                               "interface will be removed.") AgentSegmentFixupPolicy
 {
   enum
   {

diff --git a/cub/cub/agent/agent_select_if.cuh b/cub/cub/agent/agent_select_if.cuh
@@ -274,7 +274,7 @@ struct AgentSelectIf
   // Callback type for obtaining tile prefix during block scan
   using DelayConstructorT = typename AgentSelectIfPolicyT::detail::delay_constructor_t;
   using TilePrefixCallbackOpT =
-    TilePrefixCallbackOp<OffsetT, ::cuda::std::plus<>, MemoryOrderedTileStateT, 0, DelayConstructorT>;
+    TilePrefixCallbackOp<OffsetT, ::cuda::std::plus<>, MemoryOrderedTileStateT, DelayConstructorT>;
 
   // Item exchange type
   using ItemExchangeT = InputT[TILE_ITEMS];