Fix benchmark generation

Fixes the generation of small benchmarking code with kerncraft and likwid for vectorized and aligned kernels.

pystencils/kerncraft_coupling/templates/benchmark.c

-
 #include "kerncraft.h"
 #include <stdlib.h>
 #include <stdint.h>
 #include <stdbool.h>
 #include <math.h>
 #include <stdio.h>
+#include <assert.h>
 
 {{ includes }}
 
@@ -18,6 +18,43 @@ void dummy(void *);
 void timing(double* wcTime, double* cpuTime);
 extern int var_false;
 
+/* see waLBerla src/field/allocation/AlignedMalloc */
+void *aligned_malloc_with_offset( uint64_t size, uint64_t alignment, uint64_t offset )
+{
+    // With 0 alignment this function makes no sense
+    // use normal malloc instead
+    assert( alignment > 0 );
+    // Tests if alignment is power of two (assuming alignment>0)
+    assert( !(alignment & (alignment - 1)) );
+    assert( offset < alignment );
+
+    void *pa;  // pointer to allocated memory
+    void *ptr; // pointer to usable aligned memory
+
+    pa=std::malloc( (size+2*alignment-1 )+sizeof(void *));
+    if(!pa)
+        return nullptr;
+
+    // Find next aligned position, starting at pa+sizeof(void*)-1
+    ptr=(void*)( ((size_t)pa+sizeof(void *)+alignment-1) & ~(alignment-1));
+    ptr=(void*) ( (char*)(ptr) + alignment - offset);
+
+    // Store pointer to real allocated chunk just before usable chunk
+    *((void **)ptr-1)=pa;
+
+    assert( ((size_t)ptr+offset) % alignment == 0 );
+
+    return ptr;
+}
+
+void aligned_free( void *ptr )
+{
+    // assume that pointer to real allocated chunk is stored just before
+    // chunk that was given to user
+    if(ptr)
+        std::free(*((void **)ptr-1));
+}
+
 
 {{kernel_code}}
 
@@ -28,11 +65,14 @@ int main(int argc, char **argv)
   likwid_markerInit();
   {%- endif %}
 
-  {%- for field_name, dataType, size in fields %}
-
+  {%- for field_name, dataType, elements, size, offset, alignment in fields %}
   // Initialization {{field_name}}
-  double * {{field_name}} = (double *) aligned_malloc(sizeof({{dataType}}) * {{size}}, 64);
-  for (unsigned long long i = 0; i < {{size}}; ++i)
+  {%- if alignment > 0 %}
+  {{dataType}} * {{field_name}} = ({{dataType}} *) aligned_malloc_with_offset({{size}}, {{alignment}}, {{offset}});
+  {%- else %}
+  {{dataType}} * {{field_name}} = new {{dataType}}[{{elements}}];
+  {%- endif %}
+  for (unsigned long long i = 0; i < {{elements}}; ++i)
     {{field_name}}[i] = 0.23;
 
   if(var_false)
@@ -69,18 +109,18 @@ int main(int argc, char **argv)
       likwid_markerStartRegion("loop");
       {%- endif %}
     }
-    
+
     {%- if timing %}
     double wcStartTime, cpuStartTime, wcEndTime, cpuEndTime;
     timing(&wcStartTime, &cpuStartTime);
     {%- endif %}
-    
+
     for (; repeat > 0; --repeat)
     {
       {{kernelName}}({{call_argument_list}});
 
       // Dummy calls
-      {%- for field_name, dataType, size in fields %}
+      {%- for field_name, dataType, elements, size, offset, alignment in fields %}
       if(var_false) dummy((void*){{field_name}});
       {%- endfor %}
       {%- for constantName, dataType in constants %}
@@ -105,4 +145,13 @@ int main(int argc, char **argv)
   {%- if likwid %}
   likwid_markerClose();
   {%- endif %}
+
+  {%- for field_name, dataType, elements, size, offset, alignment in fields %}
+  {%- if alignment > 0 %}
+  aligned_free({{field_name}});
+  {%- else %}
+  delete[] {{field_name}};
+  {%- endif %}
+
+  {%- endfor %}
 }