Small fixes to make compilation of single precision work

python/dune/perftool/loopy/vcl.py

@@ -95,4 +95,4 @@ def vcl_function_mangler(knl, func, arg_dtypes):
     if func == "horizontal_add":
         dtype = arg_dtypes[0]
         vcl = lp.types.NumpyType(get_vcl_type(dtype))
-        return lp.CallMangleInfo("horizontal_add", (lp.types.NumpyType(dtype),), (vcl,))
+        return lp.CallMangleInfo("horizontal_add", (lp.types.NumpyType(dtype.dtype),), (vcl,))

python/dune/perftool/pdelab/driver/__init__.py

@@ -15,6 +15,7 @@ from dune.perftool.generation import (generator_factory,
                                       global_context,
                                       include_file,
                                       cached,
+                                      pre_include,
                                       preamble,
                                       )
 from dune.perftool.options import get_option

python/dune/perftool/sumfact/tabulation.py

@@ -399,7 +399,7 @@ class PolynomialLookup(FunctionIdentifier):
 def polynomial_lookup_mangler(target, func, dtypes):
     if isinstance(func, PolynomialLookup):
         dtype = dtype_floatingpoint()
-        return CallMangleInfo(func.name, (dtype,), (NumpyType(np.int32), NumpyType(dtype)))
+        return CallMangleInfo(func.name, (NumpyType(dtype),), (NumpyType(np.int32), NumpyType(dtype)))
 
 
 def define_theta(name, tabmat, additional_indices=(), width=None):