Fix bias_add size calculation for half.

PiperOrigin-RevId: 167051419

1 files changed, 3 insertions, 3 deletions

diff --git a/tensorflow/core/kernels/bias_op_gpu.cu.cc b/tensorflow/core/kernels/bias_op_gpu.cu.cc
index e07ca5e0c4..ddc2d457b0 100644
--- a/tensorflow/core/kernels/bias_op_gpu.cu.cc
+++ b/tensorflow/core/kernels/bias_op_gpu.cu.cc
@@ -142,9 +142,9 @@ __global__ void BiasGradNCHW_SharedAtomics(const T* output_backprop,
                                            int group_size) {
   // Initialize the shared memory.
   typedef typename AccumulatorType<T>::type AccT;
-  __shared__ AccT s_data[32];
-  int32 s_data_size = sizeof(s_data) / sizeof(T);
-  for (int32 index = threadIdx.x; index < s_data_size; index += blockDim.x) {
+  const int32 kSDataSize = 32;
+  __shared__ AccT s_data[kSDataSize];
+  for (int32 index = threadIdx.x; index < kSDataSize; index += blockDim.x) {
     s_data[index] = AccT(0);
   }
   __syncthreads();