diff --git a/scipy/sparse/compressed.py b/scipy/sparse/compressed.py
index 971e765c1153..5fa0785afdbd 100644
--- a/scipy/sparse/compressed.py
+++ b/scipy/sparse/compressed.py
@@ -10,8 +10,8 @@
 from base import spmatrix, isspmatrix, SparseEfficiencyWarning
 from data import _data_matrix
 import sparsetools
-from sputils import upcast, to_native, isdense, isshape, getdtype, \
-        isscalarlike, isintlike
+from sputils import upcast, upcast_char, to_native, isdense, isshape, \
+     getdtype, isscalarlike, isintlike
 
 
 class _cs_matrix(_data_matrix):
@@ -251,7 +251,8 @@ def _mul_vector(self, other):
         M,N = self.shape
 
         # output array
-        result = np.zeros(M, dtype=upcast(self.dtype, other.dtype))
+        result = np.zeros(M, dtype=upcast_char(self.dtype.char,
+                                               other.dtype.char))
 
         # csr_matvec or csc_matvec
         fn = getattr(sparsetools,self.format + '_matvec')
@@ -264,7 +265,8 @@ def _mul_multivector(self, other):
         M,N = self.shape
         n_vecs = other.shape[1] #number of column vectors
 
-        result = np.zeros((M,n_vecs), dtype=upcast(self.dtype,other.dtype) )
+        result = np.zeros((M,n_vecs), dtype=upcast_char(self.dtype.char,
+                                                        other.dtype.char))
 
         # csr_matvecs or csc_matvecs
         fn = getattr(sparsetools,self.format + '_matvecs')
diff --git a/scipy/sparse/coo.py b/scipy/sparse/coo.py
index 060eafc17d05..b88477d03710 100644
--- a/scipy/sparse/coo.py
+++ b/scipy/sparse/coo.py
@@ -11,7 +11,7 @@
 from sparsetools import coo_tocsr, coo_todense, coo_matvec
 from base import isspmatrix
 from data import _data_matrix
-from sputils import upcast, to_native, isshape, getdtype, isintlike
+from sputils import upcast, upcast_char, to_native, isshape, getdtype, isintlike
 
 class coo_matrix(_data_matrix):
     """
@@ -369,7 +369,8 @@ def _with_data(self,data,copy=True):
 
     def _mul_vector(self, other):
         #output array
-        result = np.zeros( self.shape[0], dtype=upcast(self.dtype,other.dtype) )
+        result = np.zeros( self.shape[0], dtype=upcast_char(self.dtype.char,
+                                                            other.dtype.char) )
         coo_matvec(self.nnz, self.row, self.col, self.data, other, result)
         return result
 
diff --git a/scipy/sparse/dia.py b/scipy/sparse/dia.py
index 0389ebb7f008..ed6a193545e7 100644
--- a/scipy/sparse/dia.py
+++ b/scipy/sparse/dia.py
@@ -8,7 +8,7 @@
 
 from base import isspmatrix, _formats
 from data import _data_matrix
-from sputils import isshape, upcast, getdtype
+from sputils import isshape, upcast, upcast_char, getdtype
 from sparsetools import dia_matvec
 
 class dia_matrix(_data_matrix):
@@ -165,7 +165,8 @@ def getnnz(self):
     def _mul_vector(self, other):
         x = other
 
-        y = np.zeros( self.shape[0], dtype=upcast(self.dtype,x.dtype))
+        y = np.zeros( self.shape[0], dtype=upcast_char(self.dtype.char,
+                                                       x.dtype.char))
 
         L = self.data.shape[1]
 
diff --git a/scipy/sparse/sputils.py b/scipy/sparse/sputils.py
index 982b385d4a46..e6cc39f626df 100644
--- a/scipy/sparse/sputils.py
+++ b/scipy/sparse/sputils.py
@@ -50,6 +50,14 @@ def upcast(*args):
 
     raise TypeError('no supported conversion for types: %s' % args)
 
+def upcast_char(*args):
+    """Same as `upcast` but taking dtype.char as input (faster)."""
+    t = _upcast_memo.get(args)
+    if t is not None:
+        return t
+    t = upcast(*map(np.dtype, args))
+    _upcast_memo[args] = t
+    return t
 
 def to_native(A):
     return np.asarray(A,dtype=A.dtype.newbyteorder('native'))