From c0b967054e144e74e2365aa50790410006aa540e Mon Sep 17 00:00:00 2001
From: Trung Nguyen <ndactrung@gmail.com>
Date: Thu, 16 Sep 2021 17:27:44 -0500
Subject: [PATCH] Fixed bugs with zero local atoms (similar to what has been
 done to PPPM interp)

---
 lib/gpu/lal_amoeba.cpp | 28 +++++++++++++++++++---------
 lib/gpu/lal_pppm.cpp   |  6 ++++--
 2 files changed, 23 insertions(+), 11 deletions(-)

diff --git a/lib/gpu/lal_amoeba.cpp b/lib/gpu/lal_amoeba.cpp
index 8adabbe6d5..6bf93a3eb5 100644
--- a/lib/gpu/lal_amoeba.cpp
+++ b/lib/gpu/lal_amoeba.cpp
@@ -128,14 +128,18 @@ double AmoebaT::host_memory_usage() const {
 // ---------------------------------------------------------------------------
 template <class numtyp, class acctyp>
 int AmoebaT::udirect2b(const int eflag, const int vflag) {
+  int ainum=this->ans->inum(); 
+  if (ainum == 0)
+    return 0;
+
   int _nall=this->atom->nall();
   int nbor_pitch=this->nbor->nbor_pitch();
-  int ainum=this->ans->inum(); 
 
   // Compute the block size and grid size to keep all cores busy
   const int BX=this->block_size();
   int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
                                (BX/this->_threads_per_atom)));
+  this->time_pair.start();
 
   // Build the short neighbor list if not done yet
   if (!this->short_nbor_avail) {
@@ -164,14 +168,17 @@ int AmoebaT::udirect2b(const int eflag, const int vflag) {
 // ---------------------------------------------------------------------------
 template <class numtyp, class acctyp>
 int AmoebaT::umutual2b(const int eflag, const int vflag) {
+  int ainum=this->ans->inum();
+  if (ainum == 0)
+    return 0;
+
+  int _nall=this->atom->nall();
+  int nbor_pitch=this->nbor->nbor_pitch();
+
   // Compute the block size and grid size to keep all cores busy
   const int BX=this->block_size();
   int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
                                (BX/this->_threads_per_atom)));
-
-  int _nall=this->atom->nall();
-  int ainum=this->ans->inum();
-  int nbor_pitch=this->nbor->nbor_pitch();
   this->time_pair.start();
 
   // Build the short neighbor list if not done yet
@@ -200,14 +207,17 @@ int AmoebaT::umutual2b(const int eflag, const int vflag) {
 // ---------------------------------------------------------------------------
 template <class numtyp, class acctyp>
 int AmoebaT::polar_real(const int eflag, const int vflag) {
+  int ainum=this->ans->inum();
+  if (ainum == 0)
+    return 0;
+
+  int _nall=this->atom->nall();
+  int nbor_pitch=this->nbor->nbor_pitch();
+
   // Compute the block size and grid size to keep all cores busy
   const int BX=this->block_size();
   int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
                                (BX/this->_threads_per_atom)));
-
-  int _nall=this->atom->nall();
-  int ainum=this->ans->inum();
-  int nbor_pitch=this->nbor->nbor_pitch();
   this->time_pair.start();
 
   // Build the short neighbor list if not done yet
diff --git a/lib/gpu/lal_pppm.cpp b/lib/gpu/lal_pppm.cpp
index 6e8fe237a6..87ab6fe775 100644
--- a/lib/gpu/lal_pppm.cpp
+++ b/lib/gpu/lal_pppm.cpp
@@ -342,13 +342,15 @@ void PPPMT::interp(const grdtyp qqrd2e_scale) {
   vd_brick.update_device(true);
   time_in.stop();
 
+  int ainum=this->ans->inum();
+  if (ainum==0)
+    return;
+
   time_interp.start();
   // Compute the block size and grid size to keep all cores busy
   int BX=this->block_size();
   int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/BX));
 
-  int ainum=this->ans->inum();
-
   k_interp.set_size(GX,BX);
   k_interp.run(&atom->x, &atom->q, &ainum, &vd_brick, &d_rho_coeff,
                &_npts_x, &_npts_yx, &_brick_x, &_brick_y, &_brick_z, &_delxinv,