GpuBatchedDot: remove CNDA_THREAD_SYNC from streams implementation as it always…

theano/sandbox/cuda/blas.py

@@ -177,9 +177,9 @@ class GpuBatchedDot(GpuOp):
                                      (const float **) gpu_y, y_dim2,
                                      (const float **) gpu_x, x_dim2, &beta,
                                      gpu_z, y_dim2, x_dim0);
+            CNDA_THREAD_SYNC;
 
             CLEANUP();
-
             if (CUBLAS_STATUS_SUCCESS != err)
             {
                 PyErr_Format(PyExc_RuntimeError,
@@ -303,7 +303,6 @@ class GpuBatchedDot(GpuOp):
 
                 x += Sx[0]; y += Sy[0]; z += Sz[0];
             };
-            CNDA_THREAD_SYNC;
 
             for(int i = 0; i < N_STREAMS; i++) {
                 cudaStreamSynchronize(streams[i]);