Also use the other arg of sub as output.

theano/sandbox/cuda/blocksparse.py

@@ -321,7 +321,6 @@ CudaNdarray_HOST_STRIDES(%(out)s)[0], CudaNdarray_HOST_STRIDES(%(out)s)[1],
         o, W, h, inputIdx, outputIdx = inputs
         go = grads[0]
 
-        # might revise that interface to not have a huge output
         Wgrad = sparse_block_outer_ss(W.zeros_like(),
                                       h, go, inputIdx, outputIdx)
         hgrad = sparse_block_gemv_ss(h.zeros_like(),
@@ -682,9 +681,9 @@ GpuElemwise{mul}(lr, SparseBlockOuterSS) -> SparseBlockOuterSS(..., alpha=lr)
 
     @opt.register_opt()
     @opt.local_optimizer([GpuElemwise])
-    def local_merge_blocksparse_beta(node):
+    def local_merge_blocksparse_output(node):
         if (isinstance(node.op, GpuElemwise) and
-            node.op.scalar_op == scalar.sub and
+            (node.op.scalar_op == scalar.sub or node.scalar_op == scalar.add) and
             node.nin == 2):
             ger = grab_ger(node.inputs[0])
             W = node.inputs[1]
@@ -693,8 +692,11 @@ GpuElemwise{mul}(lr, SparseBlockOuterSS) -> SparseBlockOuterSS(..., alpha=lr)
                 W = node.inputs[0]
             if ger is None:
                 return None
+            alpha = get.inputs[5]
+            if node.op.scalar_op == scalar.sub:
+                alpha = -alpha
             return [sparse_block_outer_ss(*([W] + ger.inputs[1:5] +
-                                            [-ger.inputs[5]]))]
+                                            [alpha]))]
 
 
 def sparse_block_dot_SS(W, h, inputIdx, b, outputIdx):