fix gh-4036. We need the shape and we don't compute it. It wasn't timmed, so…

theano/sandbox/cuda/opt.py

@@ -2770,7 +2770,8 @@ def local_abstractconv_gemm(node):
             (node.op.imshp is not None) and
             (None not in node.op.imshp[-2:]) and
             (node.op.kshp is not None) and
-                (None not in node.op.kshp)):
+            (None not in node.op.kshp) and
+             border_mode != "half"):
             # we know the kernel and output size
             prod1 = node.op.kshp[0] * node.op.kshp[1]
             prod2 = ((node.op.imshp[-2] - node.op.kshp[0] + 1) *