cmake: Switchable valgrind/callgrind instrumentation for the main loop.

CMakeLists.txt

@@ -43,6 +43,7 @@ option(WITH_TCL    "Build tcl interface"     ON)
 option(WITH_CUDA   "Build with GPU support"  ON)
 option(WITH_TESTS  "Enable tests"            ON)
 option(WITH_SCAFACOS "Build with Scafacos support" ON)
+option(WITH_VALGRIND_INSTRUMENTATION "Build with valgrind instrumentation markers" OFF)
 
 # choose the name of the config file
 set(MYCONFIG_NAME "myconfig.hpp"
@@ -101,6 +102,16 @@ if(WITH_SCAFACOS)
   endif(SCAFACOS_FOUND)
 endif(WITH_SCAFACOS)
 
+if(WITH_VALGRIND_INSTRUMENTATION)
+  find_package(PkgConfig)
+  pkg_check_modules(VALGRIND valgrind)
+  if(VALGRIND_FOUND)
+    set(VALGRIND_INSTRUMENTATION 1)
+    message(STATUS ${VALGRIND_INCLUDE_DIRS})
+    include_directories(${VALGRIND_INCLUDE_DIRS})
+  endif(VALGRIND_FOUND)
+endif(WITH_VALGRIND_INSTRUMENTATION)
+
 include(RequireCXX11)
 
 #######################################################################

cmake/cmake_config.cmakein

@@ -17,6 +17,8 @@
 
 #cmakedefine SCAFACOS
 
+#cmakedefine VALGRIND_INSTRUMENTATION
+
 #define PACKAGE_NAME "${PROJECT_NAME}"
 
 #define __PRETTYFUNC__ "n/a"

src/core/integrate.cpp

@@ -64,6 +64,10 @@
 #include "immersed_boundary/ibm_volume_conservation.hpp"
 #include "minimize_energy.hpp"
 
+#ifdef VALGRIND_INSTRUMENTATION
+#include <callgrind.h>
+#endif
+
 /************************************************
  * DEFINES
  ************************************************/
@@ -325,6 +329,10 @@ void integrate_vv(int n_steps, int reuse_forces)
 
   n_verlet_updates = 0;
 
+#ifdef VALGRIND_INSTRUMENTATION
+  CALLGRIND_START_INSTRUMENTATION;
+#endif
+
   /* Integration loop */
   for (int step=0; step<n_steps; step++) {
     INTEG_TRACE(fprintf(stderr,"%d: STEP %d\n", this_node, step));
@@ -550,6 +558,10 @@ void integrate_vv(int n_steps, int reuse_forces)
     #endif
   }
 
+#ifdef VALGRIND_INSTRUMENTATION
+  CALLGRIND_STOP_INSTRUMENTATION;
+#endif
+
   /* verlet list statistics */
   if(n_verlet_updates>0) verlet_reuse = n_steps/(double) n_verlet_updates;
   else verlet_reuse = 0;