Tiny fix in using data type template parameters in blockwise and dire…

…ct_threadwise kernel
ROCm · Sep 15, 2021 · 2f4fe70 · 2f4fe70
1 parent 7177b7c
commit 2f4fe70
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 3 deletions.
diff --git a/...el/composable_kernel/include/tensor_operation/gridwise_generic_2d_reduction_blockwise.hpp b/...el/composable_kernel/include/tensor_operation/gridwise_generic_2d_reduction_blockwise.hpp
@@ -281,7 +281,7 @@ struct GridwiseReduction_xy_to_x_blockwise
                                             ThreadClusterLengths,
                                             Sequence<0, 1>,
                                             srcDataType,
-                                            dstDataType,
+                                            compType,
                                             src2dDescType,
                                             decltype(in_block_desc),
                                             Sequence<0, 1>,

diff --git a/...sable_kernel/include/tensor_operation/gridwise_generic_2d_reduction_direct_threadwise.hpp b/...sable_kernel/include/tensor_operation/gridwise_generic_2d_reduction_direct_threadwise.hpp
@@ -232,7 +232,7 @@ struct GridwiseReduction_xy_to_x_direct_threadwise
         index_t thread_global_1d_id = get_block_1d_id() * BlockSize + get_thread_local_1d_id();
 
         auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<srcDataType,
-                                                                    dstDataType,
+                                                                    compType,
                                                                     src2dDescType,
                                                                     decltype(ThreadBufferDesc),
                                                                     ThreadBufferLengths,
@@ -377,7 +377,7 @@ struct GridwiseReduction_xy_to_x_direct_threadwise
         index_t thread_global_1d_id = get_block_1d_id() * BlockSize + get_thread_local_1d_id();
 
         auto threadwise_src_val_load = ThreadwiseTensorSliceTransfer_v2<srcDataType,
-                                                                        dstDataType,
+                                                                        compType,
                                                                         src2dDescType,
                                                                         decltype(ThreadBufferDesc),
                                                                         ThreadBufferLengths,