lite

build/Linux/Release/amdgpu/onnxruntime/core/providers/rocm/cu_inc/binary_elementwise_impl.cuh

@@ -192,7 +192,7 @@ void BinaryElementWiseNoBroadcastImpl(
 
   #ifdef USE_ROCM
   const int num_elements_per_thread = 2;
-  const int num_threads_per_block = 512;
+  const int num_threads_per_block = 256;
   #else
   const int num_elements_per_thread = GridDim::maxElementsPerThread;
   const int num_threads_per_block = GridDim::maxThreadsPerBlock;
@@ -228,7 +228,7 @@ void BinaryElementWiseImpl(
 
   #ifdef USE_ROCM
   const int num_elements_per_thread = 2;
-  const int num_threads_per_block = 512;
+  const int num_threads_per_block = 256;
   #else
   const int num_elements_per_thread = GridDim::maxElementsPerThread;
   const int num_threads_per_block = GridDim::maxThreadsPerBlock;