Support fastdg and h/l in benchmark generation

python/dune/perftool/sumfact/autotune.py

@@ -64,6 +64,7 @@ def generate_standalone_code(sf, filename, logname):
                       "#include<dune/perftool/common/tsc.hh>\n",
                       "#include<dune/perftool/common/vectorclass.hh>\n",
                       "#include<dune/perftool/sumfact/onedquadrature.hh>\n",
+                      "#include<dune/perftool/sumfact/horizontaladd.hh>\n",
                       "#include<random>\n",
                       "#include<fstream>\n",
                       "\n"
@@ -98,6 +99,16 @@ def generate_standalone_code(sf, filename, logname):
                       "  char buffer1[{}] __attribute__ ((aligned (32)));\n".format(size),
                       ])
 
+        # Setup fastdg inputs
+        for arg in sf.interface.signature_args:
+            if "jacobian" in arg:
+                f.write("{} = 0;\n".format(arg))
+            else:
+                basis_size = product(m.basis_size for m in sf.matrix_sequence)
+                if sf.within_inames:
+                    basis_size = basis_size * basis_size
+                f.write("RF {}[{}] __attribute__ ((aligned (32)));\n".format(arg.split()[-1], product(m.basis_size for m in sf.matrix_sequence)))
+
         # Write stuff into the input buffer
         f.writelines(["  {0} *input = ({0} *)buffer0;\n".format(real),
                       "  {0} *output = ({0} *)buffer{1};\n".format(real, sf.length % 2),
@@ -129,7 +140,7 @@ def generate_standalone_code(sf, filename, logname):
         # Add setup code for theta matrices. We add some lines not necessary,
         # but it would be more work to remove them than keeping them.
         for line in lp.generate_body(constructor_knl).split("\n")[1:-1]:
-            if "gfsu" not in line and "meshwidth" not in line:
+            if "gfsu" not in line and "meshwidth" not in line and "geometry" not in line:
                 f.write("  {}\n".format(line))
 
         # INtroduces a variable that makes sure that the kernel cannot be optimized away