Merge pull request #1752 from ndtrung81/gpu-mem-leak-init

Fixed bugs with kernel (re)compiling when the global device got clear and reinitialized

lib/gpu/lal_base_atomic.cpp

@@ -26,15 +26,16 @@ BaseAtomicT::BaseAtomic() : _compiled(false), _max_bytes(0)  {
   ans=new Answer<numtyp,acctyp>();
   nbor=new Neighbor();
   pair_program=NULL;
+  ucl_device=NULL;
 }
 
 template <class numtyp, class acctyp>
 BaseAtomicT::~BaseAtomic() {
   delete ans;
   delete nbor;
-  if (pair_program) delete pair_program;
   k_pair_fast.clear();
   k_pair.clear();
+  if (pair_program) delete pair_program;
 }
 
 template <class numtyp, class acctyp>
@@ -78,6 +79,8 @@ int BaseAtomicT::init_atomic(const int nlocal, const int nall,
   if (success!=0)
     return success;
 
+  if (ucl_device!=device->gpu) _compiled=false;
+
   ucl_device=device->gpu;
   atom=&device->atom;
 

lib/gpu/lal_base_charge.cpp

@@ -26,15 +26,16 @@ BaseChargeT::BaseCharge() : _compiled(false), _max_bytes(0) {
   ans=new Answer<numtyp,acctyp>();
   nbor=new Neighbor();
   pair_program=NULL;
+  ucl_device=NULL;
 }
 
 template <class numtyp, class acctyp>
 BaseChargeT::~BaseCharge() {
   delete ans;
   delete nbor;
-  if (pair_program) delete pair_program;
   k_pair_fast.clear();
   k_pair.clear();
+  if (pair_program) delete pair_program;
 }
 
 template <class numtyp, class acctyp>
@@ -78,6 +79,8 @@ int BaseChargeT::init_atomic(const int nlocal, const int nall,
   if (success!=0)
     return success;
 
+  if (ucl_device!=device->gpu) _compiled=false;
+
   ucl_device=device->gpu;
   atom=&device->atom;
 

lib/gpu/lal_base_dipole.cpp

@@ -26,15 +26,16 @@ BaseDipoleT::BaseDipole() : _compiled(false), _max_bytes(0) {
   ans=new Answer<numtyp,acctyp>();
   nbor=new Neighbor();
   pair_program=NULL;
+  ucl_device=NULL;
 }
 
 template <class numtyp, class acctyp>
 BaseDipoleT::~BaseDipole() {
   delete ans;
   delete nbor;
-  if (pair_program) delete pair_program;
   k_pair_fast.clear();
   k_pair.clear();
+  if (pair_program) delete pair_program;
 }
 
 template <class numtyp, class acctyp>
@@ -79,6 +80,8 @@ int BaseDipoleT::init_atomic(const int nlocal, const int nall,
   if (success!=0)
     return success;
 
+  if (ucl_device!=device->gpu) _compiled=false;
+
   ucl_device=device->gpu;
   atom=&device->atom;
 

lib/gpu/lal_base_dpd.cpp

@@ -26,15 +26,16 @@ BaseDPDT::BaseDPD() : _compiled(false), _max_bytes(0) {
   ans=new Answer<numtyp,acctyp>();
   nbor=new Neighbor();
   pair_program=NULL;
+  ucl_device=NULL;
 }
 
 template <class numtyp, class acctyp>
 BaseDPDT::~BaseDPD() {
   delete ans;
   delete nbor;
-  if (pair_program) delete pair_program;
   k_pair_fast.clear();
   k_pair.clear();
+  if (pair_program) delete pair_program;
 }
 
 template <class numtyp, class acctyp>
@@ -79,6 +80,8 @@ int BaseDPDT::init_atomic(const int nlocal, const int nall,
   if (success!=0)
     return success;
 
+  if (ucl_device!=device->gpu) _compiled=false;
+
   ucl_device=device->gpu;
   atom=&device->atom;
 

lib/gpu/lal_base_ellipsoid.cpp

@@ -36,6 +36,7 @@ BaseEllipsoidT::BaseEllipsoid() : _compiled(false), _max_bytes(0) {
   nbor_program=NULL;
   ellipsoid_program=NULL;
   lj_program=NULL;
+  ucl_device=NULL;
 }
 
 template <class numtyp, class acctyp>
@@ -92,7 +93,9 @@ int BaseEllipsoidT::init_base(const int nlocal, const int nall,
                   max_nbors,cell_size,true,1);
   if (success!=0)
     return success;
-  
+
+  if (ucl_device!=device->gpu) _compiled=false;
+
   ucl_device=device->gpu;
   atom=&device->atom;
 

lib/gpu/lal_base_three.cpp

@@ -28,6 +28,7 @@ BaseThreeT::BaseThree() : _compiled(false), _max_bytes(0) {
   ans2=new Answer<numtyp,acctyp>();
   #endif
   pair_program=NULL;
+  ucl_device=NULL;
 }
 
 template <class numtyp, class acctyp>
@@ -37,12 +38,12 @@ BaseThreeT::~BaseThree() {
   #ifdef THREE_CONCURRENT
   delete ans2;
   #endif
-  if (pair_program) delete pair_program;
   k_three_center.clear();
   k_three_end.clear();
   k_three_end_vatom.clear();
   k_pair.clear();
   k_short_nbor.clear();
+  if (pair_program) delete pair_program;
 }
 
 template <class numtyp, class acctyp>
@@ -94,6 +95,8 @@ int BaseThreeT::init_three(const int nlocal, const int nall,
   if (success!=0)
     return success;
 
+  if (ucl_device!=device->gpu) _compiled=false;
+
   ucl_device=device->gpu;
   atom=&device->atom;