[skip ci] Restructure conditionals

python/dune/codegen/sumfact/realization.py

@@ -247,23 +247,28 @@ def realize_sumfact_kernel_function(sf):
             if sf.vectorized:
                 insn_args["forced_iname_deps"] = frozenset({vec_iname[0].name})
             insn_dep = sf.interface.realize_direct(matprod, output_inames, out_shape, **insn_args)
-        else:
-            # Get a temporary that interprets the base storage of the output.
-            #
-            # Note: In this step the reordering of the fastest directions
-            # is happening. The new direction (out_inames[0]) and the
-            # corresponding shape (out_shape[0]) goes to the end (slowest
-            # direction) and everything stays column major (ftags->fortran
-            # style).
-            #
-            # If we are in the last step we reverse the permutation.
+        elif l == len(matrix_sequence) - 1:
             output_shape = tuple(out_shape[1:]) + (out_shape[0],)
-            if l == len(matrix_sequence) - 1:
-                # TODO: Move permutations to interface
-                output_shape = permute_backward(output_shape, sf.cost_permutation)
-                if sf.stage == 3:
-                    output_shape = permute_backward(output_shape, sf.interface.quadrature_permutation)
 
+            # TODO: Move permutations to interface
+            output_shape = permute_backward(output_shape, sf.cost_permutation)
+            if sf.stage == 3:
+                output_shape = permute_backward(output_shape, sf.interface.quadrature_permutation)
+
+            out = buffer.get_temporary("buff_step{}_out".format(l),
+                                       shape=output_shape + vec_shape,
+                                       dim_tags=ftags,
+                                       )
+
+            # Issue the reduction instruction that implements the multiplication
+            # at the same time store the instruction ID for the next instruction to depend on
+            insn_dep = frozenset({instruction(assignee=prim.Subscript(prim.Variable(out), output_inames + vec_iname),
+                                              expression=matprod,
+                                              **insn_args
+                                              )
+                                  })
+        else:
+            output_shape = tuple(out_shape[1:]) + (out_shape[0],)
             out = buffer.get_temporary("buff_step{}_out".format(l),
                                        shape=output_shape + vec_shape,
                                        dim_tags=ftags,