diff --git a/src/CUDA/ddot.jl b/src/CUDA/ddot.jl
index 0a7740b..638ca24 100644
--- a/src/CUDA/ddot.jl
+++ b/src/CUDA/ddot.jl
@@ -9,5 +9,6 @@
     @inbounds for i=1:n
         v += dx[i]*dy[i]
     end
+    CUDA.sync_threads()
     return v
 end
diff --git a/src/CUDA/dspcg.jl b/src/CUDA/dspcg.jl
index 4895d72..605d48b 100644
--- a/src/CUDA/dspcg.jl
+++ b/src/CUDA/dspcg.jl
@@ -51,6 +51,7 @@
                 end
             end
         end
+        CUDA.sync_threads()
         nfree = CUDA.shfl_sync(0xffffffff, nfree, 1)
 
         # Exit if there are no free constraints.