lite

build/Linux/Release/amdgpu/onnxruntime/core/providers/rocm/reduction/reduction_functions.cu

@@ -440,7 +440,7 @@ Status call_reduce_matrix_rows(hipStream_t stream, const TIn* input, TOut* outpu
     HIP_RETURN_IF_ERROR(hipMemsetAsync(output, 0, n * sizeof(TOut), stream));
   }
 
-  constexpr int max_num_threads_in_block = 512;
+  constexpr int max_num_threads_in_block = 256;
   constexpr int max_num_blocks_in_grid = 512;
   constexpr int load_count_per_thread = 4;
 

build/Linux/Release/amdgpu/onnxruntime/core/providers/rocm/reduction/reduction_ops.h

@@ -61,6 +61,8 @@ class ReduceKernel : public RocmKernel, public ReduceKernelBase<allow_multi_axes
   template <typename T, miopenReduceTensorIndices_t ReduceTensorIndices = MIOPEN_REDUCE_TENSOR_NO_INDICES>
   Status ComputeImplEx(OpKernelContext* ctx, miopenReduceTensorOp_t miopen_reduce_op) const;
 
+
+
   template <typename T, typename OutT, miopenReduceTensorIndices_t ReduceTensorIndices>
   Status ReduceKernelShared(
       const T* X,