make Sparse structured_dot work when both matrix are sparse.

theano/sparse/basic.py

@@ -790,7 +790,11 @@ class StructuredDot(gof.Op):
         dtype_out = scalar.upcast(a.type.dtype, b.type.dtype)
         if b.type.ndim != 2:
             raise NotImplementedError('non-matrix b')
-        return gof.Apply(self, [a,b], [tensor.tensor(dtype_out, (False, b.type.broadcastable[1]))])
+
+        if _is_sparse_variable(b):
+            return gof.Apply(self, [a,b], [SparseType(a.type.format,dtype_out)()])
+        else:
+            return gof.Apply(self, [a,b], [tensor.tensor(dtype_out, (False, b.type.broadcastable[1]))])
 
     def perform(self, node, (a,b), (out,)):
         if a.shape[1] != b.shape[0]:
@@ -798,6 +802,11 @@ class StructuredDot(gof.Op):
 
         #variable = a.dot(b)  # deprecated
         variable = a * b
+        if isinstance(node.outputs[0].type,SparseType):
+            assert _is_sparse(variable)
+            out[0] = variable
+            return
+
         assert _is_dense(variable) # scipy 0.7 automatically converts to dense
 
         # dot of an NxM sparse matrix, with a Mx1 dense matrix, returns vector not matrix

theano/sparse/tests/test_basic.py

@@ -344,6 +344,28 @@ class test_structureddot(unittest.TestCase):
         outvals = f(kernvals,imvals)
         print outvals
 
+    def test_dot_sparse_sparse(self):
+        #test dot for 2 input sparse matrix
+        sparse_dtype = 'float64'
+        for sparse_format in ['csc','csr']:
+            a = SparseType(sparse_format, dtype=sparse_dtype)()
+            b = SparseType(sparse_format, dtype=sparse_dtype)()
+            d = theano.dot(a,b)
+            f = theano.function([a,b], theano.Out(d, borrow=True))
+            topo = f.maker.env.toposort()
+            for M,N,K,nnz in [(4,3,2,3),
+                              (40,30,20,3),
+                              (40,30,20,30),
+                              (400,3000,200,6000),
+                              ]:
+                if sparse_format == 'csc':
+                    spmat = sp.csc_matrix(random_lil((M,N), sparse_dtype, nnz))
+                    spmat2 = sp.csc_matrix(random_lil((N,K), sparse_dtype, nnz))
+                elif sparse_format == 'csr':
+                    spmat = sp.csr_matrix(random_lil((M,N), sparse_dtype, nnz))
+                    spmat2 = sp.csr_matrix(random_lil((N,K), sparse_dtype, nnz))
+                f(spmat,spmat2)
+
     def test_csc_correct_output_faster_than_scipy(self):
         sparse_dtype = 'float64'
         dense_dtype = 'float64'