Fix VCLCast for padded kernels

python/dune/perftool/sumfact/tabulation.py

@@ -22,7 +22,7 @@ from dune.perftool.generation import (class_member,
                                       )
 from dune.perftool.loopy.buffer import get_buffer_temporary
 from dune.perftool.loopy.target import dtype_floatingpoint
-from dune.perftool.loopy.vcl import ExplicitVCLCast
+from dune.perftool.loopy.vcl import ExplicitVCLCast, get_vcl_type_size
 from dune.perftool.pdelab.localoperator import (name_domain_field,
                                                 lop_template_range_field,
                                                 )
@@ -196,7 +196,7 @@ class BasisTabulationMatrixArray(BasisTabulationMatrixBase):
         # Check whether we can realize this by broadcasting the values of a simple tabulation
         if len(set(self.tabs)) == 1:
             theta = self.tabs[0].pymbolic(indices[:-1])
-            return prim.Call(ExplicitVCLCast(dtype_floatingpoint(), vector_width=len(self.tabs)), (theta,))
+            return prim.Call(ExplicitVCLCast(dtype_floatingpoint(), vector_width=get_vcl_type_size(dtype_floatingpoint())), (theta,))
 
         abbrevs = tuple("{}x{}".format("d" if t.derivative else "",
                                        "s{}".format(t.slice_index) if t.slice_size is not None else "")