Fixed bug caused by missing synchronization in findInteractingBlocks()

platforms/cuda/src/kernels/findInteractingBlocks.cu

@@ -257,6 +257,7 @@ __device__ void storeInteractionData(unsigned short x, unsigned short* buffer, s
 
     for (int i = threadIdx.x; i < BUFFER_SIZE; i += blockDim.x)
         buffer[i] = INVALID;
+    __syncthreads();
 }
 
 /**

platforms/opencl/src/kernels/findInteractingBlocks.cl

@@ -226,6 +226,7 @@ void storeInteractionData(unsigned short x, __local unsigned short* buffer, __lo
 
     for (int i = get_local_id(0); i < BUFFER_SIZE; i += get_local_size(0))
         buffer[i] = INVALID;
+    barrier(CLK_LOCAL_MEM_FENCE);
 }
 
 /**