[skip ci] Simulate cache misses in autotune benchmarks

Run the function call over multiple data sets.

python/dune/codegen/sumfact/autotune.py

@@ -381,7 +381,6 @@ def generate_standalone_kernel_code(kernel, signature, filename, transformations
     codegen_benchmark_function = textwrap.indent(codegen_benchmark_function, '  ')
 
     # Declare function arguments
-    datasets = 512
     codegen_declare_arguments = []
     codegen_declare_input = []
     function_arguments = [a for a in kernel.args if a.name in arguments]
@@ -403,10 +402,10 @@ def generate_standalone_kernel_code(kernel, signature, filename, transformations
             size *= min_stride
             alignment = arg.dtype.itemsize
             real = type_floatingpoint()
-            codegen_declare_input.append(('  {} {}[{}] __attribute__ ((aligned ({})));\n'.format(real,
-                                                                                                 arg.name,
-                                                                                                 size,
-                                                                                                 alignment)))
+            codegen_declare_input.append(('{} {}[datasets][{}] __attribute__ ((aligned ({})));\n'.format(real,
+                                                                                                         arg.name,
+                                                                                                         size,
+                                                                                                         alignment)))
 
     codegen_declare_arguments = ''.join(codegen_declare_arguments)
     codegen_declare_arguments = textwrap.indent(codegen_declare_arguments, '  ')
@@ -419,6 +418,10 @@ def generate_standalone_kernel_code(kernel, signature, filename, transformations
     for arg in function_arguments:
         if 'fastdg' in arg.name:
             lines = _initialize_arg(arg)
+            lines = ['  ' + a for a in lines]
+            lines = [a.replace(arg.name + ';', arg.name + '[i];') for a in lines]
+            lines.insert(0, 'for(std::size_t i=0; i<datasets; ++i){')
+            lines.append('}')
             codegen_initialize_input += '\n'.join(lines) + '\n'
         else:
             lines = _initialize_arg(arg)
@@ -427,8 +430,14 @@ def generate_standalone_kernel_code(kernel, signature, filename, transformations
     codegen_initialize_input = textwrap.indent(codegen_initialize_input, '  ')
 
     # Call the benchmark function
-    codegen_call_benchmark_function = kernel.name + '({})'.format(','.join(arguments)) + ';'
-    codegen_call_benchmark_function = textwrap.indent(codegen_call_benchmark_function, '      ')
+    arguments_with_datasets = arguments.copy()
+    arguments_with_datasets = [a if 'fastdg' not in a else a + '[i]' for a in arguments]
+    codegen_call_benchmark_function = 'for (std::size_t i=0; i<datasets; ++i){\n'
+    codegen_call_benchmark_function += '  ' + kernel.name + '({})'.format(','.join(arguments_with_datasets)) + ';\n'
+    for arg in input_arguments:
+        codegen_call_benchmark_function += 'benchmark::DoNotOptimize({}[i][0]);\n'.format(arg)
+    codegen_call_benchmark_function += '}'
+    codegen_call_benchmark_function = textwrap.indent(codegen_call_benchmark_function, '    ')
 
     # Replace placeholders in benchmark template
     benchmark = benchmark.replace('CODEGEN_TRANSFORMATIONS', codegen_transformations)