diff --git a/csrc/kernels.cu b/csrc/kernels.cu
index 1c3e723..4e744fb 100644
--- a/csrc/kernels.cu
+++ b/csrc/kernels.cu
@@ -1768,7 +1768,6 @@ template<typename T, int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_
 
   __shared__ float smem_row_absmax_values[ITEMS_PER_THREAD*THREADS];
   __shared__ int smem_row_nnz_values[TILE_ROWS];
-  //__shared__ float smem_col_absmax_values[ITEMS_PER_THREAD*THREADS];
 
   half local_data[ITEMS_PER_THREAD];
   float local_data_fp32[ITEMS_PER_THREAD];
@@ -1828,13 +1827,14 @@ template<typename T, int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_
       local_col_absmax_values[j] = fmaxf(local_col_absmax_values[j], __half2float(local_data[j]));
 
     // 3. compute row max (per block); store in smem to accumulate full global mem transation
-    __syncthreads();
 
     // this is slow as it uses extra registers, but we need this to be compatible with Kepler and Maxwell (no fp16 units)
     #pragma unroll ITEMS_PER_THREAD
     for(int j = 0; j < ITEMS_PER_THREAD; j++)
       local_data_fp32[j] = local_data[j];
 
+    __syncthreads();
+
     row_absmax = (float)BlockRowReduce(temp_storage.rowreduce).Reduce(local_data_fp32, cub::Max());
     if(SPARSE_DECOMP)
     {