fix local_gpu_multinomial opt handling of broadcastable.

theano/sandbox/multinomial.py

@@ -45,7 +45,8 @@ class MultinomialFromUniform(Op):
             odtype = pvals.dtype
         else:
             odtype = self.odtype
-        return Apply(self, [pvals, unis], [T.matrix(dtype=odtype)])
+        out = T.tensor(dtype=odtype, broadcastable=pvals.type.broadcastable)
+        return Apply(self, [pvals, unis], [out])
 
     def grad(self, ins, outgrads):
         pvals, unis = ins
@@ -180,7 +181,9 @@ class GpuMultinomialFromUniform(MultinomialFromUniform, GpuOp):
             raise NotImplementedError(
                 'GpuMultinomialFromUniform works only if '
                 'self.odtype == pvals.dtype', odtype, pvals.dtype)
-        return Apply(self, [pvals, unis], [pvals.type()])
+        br = (pvals.broadcastable[1], pvals.broadcastable[0])
+        out = CudaNdarrayType(broadcastable=br)()
+        return Apply(self, [pvals, unis], [out])
 
     def perform(self, node, ins, outs):
         #The perform from parent don't work with CudaNdarray.  We

theano/sandbox/test_multinomial.py

@@ -138,3 +138,17 @@ def test_gpu_opt():
     pval = pval / pval.sum(axis=1)[:, None]
     uval = numpy.ones_like(pval[:, 0]) * 0.5
     mval = f(pval, uval)
+
+    # Test with a row, it was failing in the past.
+    r = tensor.frow()
+    m = multinomial.MultinomialFromUniform('auto')(r, u)
+    assert m.dtype == 'float32', m.dtype
+    m_gpu = cuda.gpu_from_host(m)
+
+    f = function([r, u], m_gpu, allow_input_downcast=True, mode=get_mode(True))
+    assert any([type(node.op) is multinomial.GpuMultinomialFromUniform
+                for node in f.maker.fgraph.toposort()])
+    pval = numpy.arange(1 * 4, dtype='float32').reshape((1, 4))+0.1
+    pval = pval / pval.sum(axis=1)[:, None]
+    uval = numpy.ones_like(pval[:, 0]) * 0.5
+    mval2 = f(pval, uval)