Better I128 tuning

gevtushenko · Jun 6, 2023 · 0f57445 · 0f57445
1 parent ad56f13
commit 0f57445
Showing 1 changed file with 5 additions and 12 deletions.
diff --git a/cub/device/dispatch/dispatch_scan.cuh b/cub/device/dispatch/dispatch_scan.cuh
@@ -245,10 +245,10 @@ struct sm90_tuning<AccumT, 8>
 template <class AccumT>
 struct sm90_tuning<AccumT, 16>
 {
-  static constexpr int threads = 960;
-  static constexpr int items   = 20;
+  static constexpr int threads = 576;
+  static constexpr int items   = 21;
 
-  using delay_constructor = detail::fixed_delay_constructor_t<876, 635>;
+  using delay_constructor = detail::fixed_delay_constructor_t<860, 630>;
 };
 
 template <>
@@ -346,21 +346,14 @@ struct DeviceScanPolicy
   /// SM900
   struct Policy900 : ChainedPolicy<900, Policy900, Policy600>
   {
-    static constexpr BlockLoadAlgorithm load_algorithm   = sizeof(AccumT) > 8
-                                                             ? BLOCK_LOAD_DIRECT
-                                                             : BLOCK_LOAD_WARP_TRANSPOSE;
-    static constexpr BlockStoreAlgorithm store_algorithm = sizeof(AccumT) > 8
-                                                             ? BLOCK_STORE_DIRECT
-                                                             : BLOCK_STORE_WARP_TRANSPOSE;
-
     using tuning = detail::scan::sm90_tuning<AccumT>;
 
     using ScanPolicyT = policy_t<tuning::threads,
                                  tuning::items,
                                  AccumT,
-                                 load_algorithm,
+                                 ScanTransposedLoad,
                                  LOAD_DEFAULT,
-                                 store_algorithm,
+                                 ScanTransposedStore,
                                  BLOCK_SCAN_WARP_SCANS,
                                  typename tuning::delay_constructor>;
   };