Refactored precompute_induce to overlap data transfers with kernel launches

lib/gpu/lal_amoeba_ext.cpp

@@ -162,6 +162,17 @@ void amoeba_gpu_compute_polar_real(int *host_amtype, int *host_amgroup, double *
                               eflag_in, vflag_in, eatom, vatom, aewald, felec, off2, tep_ptr);
 }
 
+void amoeba_gpu_precompute_induce(const int inum_full, const int bsorder,
+                          double ***host_thetai1, double ***host_thetai2,
+                          double ***host_thetai3, int** igrid,
+                          const int nzlo_out, const int nzhi_out,
+                          const int nylo_out, const int nyhi_out,
+                          const int nxlo_out, const int nxhi_out) {
+   AMOEBAMF.precompute_induce(inum_full, bsorder, host_thetai1, host_thetai2,
+                              host_thetai3, igrid, nzlo_out, nzhi_out,
+                              nylo_out, nyhi_out, nxlo_out, nxhi_out);
+}
+
 void amoeba_gpu_fphi_uind(const int inum_full, const int bsorder,
                           double ***host_thetai1, double ***host_thetai2,
                           double ***host_thetai3, int** igrid, double ****host_grid_brick,

lib/gpu/lal_base_amoeba.cpp

@@ -580,7 +580,7 @@ void BaseAmoebaT::precompute_induce(const int inum_full, const int bsorder,
     _thetai1.alloc(_max_thetai_size*bsorder,*(this->ucl_device),UCL_READ_ONLY);
     _thetai2.alloc(_max_thetai_size*bsorder,*(this->ucl_device),UCL_READ_ONLY);
     _thetai3.alloc(_max_thetai_size*bsorder,*(this->ucl_device),UCL_READ_ONLY);
-    _igrid.alloc(_max_thetai_size*4,*(this->ucl_device),UCL_READ_ONLY);
+    _igrid.alloc(_max_thetai_size*4,*(this->ucl_device),UCL_WRITE_ONLY,UCL_READ_ONLY);
 
     _fdip_phi1.alloc(_max_thetai_size*10,*(this->ucl_device),UCL_READ_WRITE);
     _fdip_phi2.alloc(_max_thetai_size*10,*(this->ucl_device),UCL_READ_WRITE);
@@ -674,7 +674,7 @@ void BaseAmoebaT::compute_fphi_uind(const int inum_full, const int bsorder,
                                     double ***host_thetai1, double ***host_thetai2,
                                     double ***host_thetai3, int** host_igrid,
                                     double ****host_grid_brick,
-                                    void** host_fdip_phi1,
+                                    void **host_fdip_phi1,
                                     void **host_fdip_phi2,
                                     void **host_fdip_sum_phi,
                                     const int nzlo_out, const int nzhi_out,
@@ -682,16 +682,7 @@ void BaseAmoebaT::compute_fphi_uind(const int inum_full, const int bsorder,
                                     const int nxlo_out, const int nxhi_out,
                                     bool& first_iteration)
 {
-  // allocation/resize and transfers before the first iteration
+  // TODO: find out why this (dummy) host alloc helps the cgrid_brick update_device() work correcly
-  
-  if (first_iteration) {
-    precompute_induce(inum_full, bsorder, host_thetai1, host_thetai2,
-                      host_thetai3, host_igrid, nzlo_out, nzhi_out,
-                      nylo_out,  nyhi_out, nxlo_out, nxhi_out);
-    first_iteration = false;
-  }
-
-  // TODO: find out why this host alloc helps the cgrid_brick update_device() work correcly
   UCL_H_Vec<numtyp> hdummy;
   hdummy.alloc(1, *(this->ucl_device), UCL_READ_ONLY);
 

lib/gpu/lal_hippo_ext.cpp

@@ -193,6 +193,17 @@ void hippo_gpu_compute_polar_real(int *host_amtype, int *host_amgroup, double **
                              eflag_in, vflag_in, eatom, vatom, aewald, felec, off2, tep_ptr);
 }
 
+void hippo_gpu_precompute_induce(const int inum_full, const int bsorder,
+                          double ***host_thetai1, double ***host_thetai2,
+                          double ***host_thetai3, int** igrid,
+                          const int nzlo_out, const int nzhi_out,
+                          const int nylo_out, const int nyhi_out,
+                          const int nxlo_out, const int nxhi_out) {
+   HIPPOMF.precompute_induce(inum_full, bsorder, host_thetai1, host_thetai2,
+                              host_thetai3, igrid, nzlo_out, nzhi_out,
+                              nylo_out, nyhi_out, nxlo_out, nxhi_out);
+}
+
 void hippo_gpu_fphi_uind(const int inum_full, const int bsorder,
                           double ***host_thetai1, double ***host_thetai2,
                           double ***host_thetai3, int** igrid, double ****host_grid_brick,

src/GPU/pair_amoeba_gpu.cpp

@@ -88,6 +88,13 @@ void amoeba_gpu_compute_umutual2b(int *host_amtype, int *host_amgroup,
 
 void amoeba_gpu_update_fieldp(void **fieldp_ptr);
 
+void amoeba_gpu_precompute_induce(const int inum_full, const int bsorder,
+                          double ***host_thetai1, double ***host_thetai2,
+                          double ***host_thetai3, int** igrid,
+                          const int nzlo_out, const int nzhi_out,
+                          const int nylo_out, const int nyhi_out,
+                          const int nxlo_out, const int nxhi_out);
+
 void amoeba_gpu_fphi_uind(const int inum_full, const int bsorder,
                           double ***host_thetai1, double ***host_thetai2,
                           double ***host_thetai3, int** igrid,
@@ -294,6 +301,12 @@ void PairAmoebaGPU::induce()
 
   first_induce_iteration  = true;
 
+  amoeba_gpu_precompute_induce(atom->nlocal, bsorder, thetai1,
+                       thetai2, thetai3, igrid,
+                       ic_kspace->nzlo_out, ic_kspace->nzhi_out,
+                       ic_kspace->nylo_out, ic_kspace->nyhi_out,
+                       ic_kspace->nxlo_out, ic_kspace->nxhi_out);
+
   // set cutoffs, taper coeffs, and PME params
   // create qfac here, free at end of polar()
 

src/GPU/pair_hippo_gpu.cpp

@@ -105,6 +105,13 @@ void hippo_gpu_compute_umutual2b(int *host_amtype, int *host_amgroup,
 
 void hippo_gpu_update_fieldp(void **fieldp_ptr);
 
+void hippo_gpu_precompute_induce(const int inum_full, const int bsorder,
+                          double ***host_thetai1, double ***host_thetai2,
+                          double ***host_thetai3, int** igrid,
+                          const int nzlo_out, const int nzhi_out,
+                          const int nylo_out, const int nyhi_out,
+                          const int nxlo_out, const int nxhi_out);
+
 void hippo_gpu_fphi_uind(const int inum_full, const int bsorder,
                           double ***host_thetai1, double ***host_thetai2,
                           double ***host_thetai3, int** igrid,
@@ -419,6 +426,12 @@ void PairHippoGPU::induce()
 
   first_induce_iteration  = true;
 
+  hippo_gpu_precompute_induce(atom->nlocal, bsorder, thetai1,
+                       thetai2, thetai3, igrid,
+                       ic_kspace->nzlo_out, ic_kspace->nzhi_out,
+                       ic_kspace->nylo_out, ic_kspace->nyhi_out,
+                       ic_kspace->nxlo_out, ic_kspace->nxhi_out);
+
   // set cutoffs, taper coeffs, and PME params
   // create qfac here, free at end of polar()