adjust permute to general vector size

python/dune/perftool/blockstructured/vectorization.py

@@ -33,6 +33,7 @@ def add_vcl_accum_insns(knl, iname_inner, iname_outer):
     vng = knl.get_var_name_generator()
     idg = knl.get_instruction_id_generator()
     new_vec_temporaries = dict()
+    vcl_size = get_vcl_type_size(np.float64)
     for insn in knl.instructions:
         # somehow CInstructions are not hashable....
         if isinstance(insn, lp.MultiAssignmentBase) and insn in accum_insns:
@@ -61,11 +62,11 @@ def add_vcl_accum_insns(knl, iname_inner, iname_outer):
             identifier_b = vng('b')
             new_vec_temporaries[identifier_a] = DuneTemporaryVariable(identifier_a, dtype=np.float64,
                                                                       shape=(2,)*(world_dimension()-1)
-                                                                            +(get_vcl_type_size(np.float64),),
+                                                                            +(vcl_size,),
                                                                       managed=True, scope=lp.temp_var_scope.PRIVATE,
                                                                       dim_tags=('f',)*(world_dimension()-1)+('vec',))
             new_vec_temporaries[identifier_b] = DuneTemporaryVariable(identifier_b, dtype=np.float64,
-                                                                      shape=(get_vcl_type_size(np.float64),), managed=True,
+                                                                      shape=(vcl_size,), managed=True,
                                                                       scope=lp.temp_var_scope.PRIVATE, dim_tags=('vec',))
 
             var_a = prim.Subscript(prim.Variable(identifier_a),
@@ -108,7 +109,7 @@ def add_vcl_accum_insns(knl, iname_inner, iname_outer):
             # r+=a[iy]
             id_accum = idg('insn_mod_accum')
             expr_accum = prim.Sum((var_a,
-                                   prim.Call(prim.Variable('permute4d<-1,{}>'.format(','.join(map(str,range(get_vcl_type_size(np.float64)-1))))),
+                                   prim.Call(prim.Variable('permute{}d<-1,{}>'.format(vcl_size,','.join(map(str,range(vcl_size-1))))),
                                              (var_b,)),
                                    substitute(insn.assignee, {iname_ix:0})))
             new_insns.append(lp.Assignment(assignee=substitute(insn.assignee,{iname_ix:0}),
@@ -121,7 +122,7 @@ def add_vcl_accum_insns(knl, iname_inner, iname_outer):
                              )
             # a[iy] = permute
             id_permute = idg('insn_permute')
-            expr_permute = prim.Call(prim.Variable('permute4d<3,{}>'.format(','.join(['-1']*(get_vcl_type_size(np.float64)-1)))),
+            expr_permute = prim.Call(prim.Variable('permute{}d<3,{}>'.format(vcl_size,','.join(['-1']*(vcl_size-1)))),
                                      (var_b,))
             new_insns.append(lp.Assignment(assignee=var_a,
                                            expression=expr_permute,
@@ -134,7 +135,7 @@ def add_vcl_accum_insns(knl, iname_inner, iname_outer):
 
             # tail handling, uses tail alias
             id_accum_tail = idg('insn_accum_tail')
-            subst_map = {iname_inner: 0, iname_outer: get_option("number_of_blocks")//get_vcl_type_size(np.float64),
+            subst_map = {iname_inner: 0, iname_outer: get_option("number_of_blocks")//vcl_size,
                          iname_ix: 0, insn.assignee_name: prim.Variable(insn.assignee_name+'_tail'),
                          **replace_tail_inames}
             assignee_tail = substitute(insn.assignee, subst_map)