diff --git a/lib/gpu/Nvidia.makefile b/lib/gpu/Nvidia.makefile
index 21663ad85cb161c5acf97ee0463f295beaf58c17..1b4cfc469f4f4f650c2410a0903e38d62ed6b807 100644
--- a/lib/gpu/Nvidia.makefile
+++ b/lib/gpu/Nvidia.makefile
@@ -79,7 +79,9 @@ OBJS = $(OBJ_DIR)/lal_atom.o $(OBJ_DIR)/lal_ans.o \
$(OBJ_DIR)/lal_lj_cubic.o $(OBJ_DIR)/lal_lj_cubic_ext.o \
$(OBJ_DIR)/lal_ufm.o $(OBJ_DIR)/lal_ufm_ext.o \
$(OBJ_DIR)/lal_dipole_long_lj.o $(OBJ_DIR)/lal_dipole_long_lj_ext.o \
- $(OBJ_DIR)/lal_lj_expand_coul_long.o $(OBJ_DIR)/lal_lj_expand_coul_long_ext.o
+ $(OBJ_DIR)/lal_lj_expand_coul_long.o $(OBJ_DIR)/lal_lj_expand_coul_long_ext.o \
+ $(OBJ_DIR)/lal_coul_long_cs.o $(OBJ_DIR)/lal_coul_long_cs_ext.o \
+ $(OBJ_DIR)/lal_born_coul_long_cs.o $(OBJ_DIR)/lal_born_coul_long_cs_ext.o
CBNS = $(OBJ_DIR)/device.cubin $(OBJ_DIR)/device_cubin.h \
$(OBJ_DIR)/atom.cubin $(OBJ_DIR)/atom_cubin.h \
@@ -137,7 +139,9 @@ CBNS = $(OBJ_DIR)/device.cubin $(OBJ_DIR)/device_cubin.h \
$(OBJ_DIR)/lj_cubic.cubin $(OBJ_DIR)/lj_cubic_cubin.h \
$(OBJ_DIR)/ufm.cubin $(OBJ_DIR)/ufm_cubin.h \
$(OBJ_DIR)/dipole_long_lj.cubin $(OBJ_DIR)/dipole_long_lj_cubin.h \
- $(OBJ_DIR)/lj_expand_coul_long.cubin $(OBJ_DIR)/lj_expand_coul_long_cubin.h
+ $(OBJ_DIR)/lj_expand_coul_long.cubin $(OBJ_DIR)/lj_expand_coul_long_cubin.h \
+ $(OBJ_DIR)/coul_long_cs.cubin $(OBJ_DIR)/coul_long_cs_cubin.h \
+ $(OBJ_DIR)/born_coul_long_cs.cubin $(OBJ_DIR)/born_coul_long_cs_cubin.h
all: $(OBJ_DIR) $(GPU_LIB) $(EXECS)
@@ -837,6 +841,30 @@ $(OBJ_DIR)/lal_lj_expand_coul_long.o: $(ALL_H) lal_lj_expand_coul_long.h lal_lj_
$(OBJ_DIR)/lal_lj_expand_coul_long_ext.o: $(ALL_H) lal_lj_expand_coul_long.h lal_lj_expand_coul_long_ext.cpp lal_base_charge.h
$(CUDR) -o $@ -c lal_lj_expand_coul_long_ext.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/coul_long_cs.cubin: lal_coul_long_cs.cu lal_precision.h lal_preprocessor.h
+ $(CUDA) --cubin -DNV_KERNEL -o $@ lal_coul_long_cs.cu
+
+$(OBJ_DIR)/coul_long_cs_cubin.h: $(OBJ_DIR)/coul_long_cs.cubin $(OBJ_DIR)/coul_long_cs.cubin
+ $(BIN2C) -c -n coul_long_cs $(OBJ_DIR)/coul_long_cs.cubin > $(OBJ_DIR)/coul_long_cs_cubin.h
+
+$(OBJ_DIR)/lal_coul_long_cs.o: $(ALL_H) lal_coul_long_cs.h lal_coul_long_cs.cpp $(OBJ_DIR)/coul_long_cs_cubin.h $(OBJ_DIR)/lal_base_charge.o
+ $(CUDR) -o $@ -c lal_coul_long_cs.cpp -I$(OBJ_DIR)
+
+$(OBJ_DIR)/lal_coul_long_cs_ext.o: $(ALL_H) lal_coul_long_cs.h lal_coul_long_cs_ext.cpp lal_base_charge.h
+ $(CUDR) -o $@ -c lal_coul_long_cs_ext.cpp -I$(OBJ_DIR)
+
+$(OBJ_DIR)/born_coul_long_cs.cubin: lal_born_coul_long_cs.cu lal_precision.h lal_preprocessor.h
+ $(CUDA) --cubin -DNV_KERNEL -o $@ lal_born_coul_long_cs.cu
+
+$(OBJ_DIR)/born_coul_long_cs_cubin.h: $(OBJ_DIR)/born_coul_long_cs.cubin $(OBJ_DIR)/born_coul_long_cs.cubin
+ $(BIN2C) -c -n born_coul_long_cs $(OBJ_DIR)/born_coul_long_cs.cubin > $(OBJ_DIR)/born_coul_long_cs_cubin.h
+
+$(OBJ_DIR)/lal_born_coul_long_cs.o: $(ALL_H) lal_born_coul_long_cs.h lal_born_coul_long_cs.cpp $(OBJ_DIR)/born_coul_long_cs_cubin.h $(OBJ_DIR)/lal_base_charge.o
+ $(CUDR) -o $@ -c lal_born_coul_long_cs.cpp -I$(OBJ_DIR)
+
+$(OBJ_DIR)/lal_born_coul_long_cs_ext.o: $(ALL_H) lal_born_coul_long_cs.h lal_born_coul_long_cs_ext.cpp lal_base_charge.h
+ $(CUDR) -o $@ -c lal_born_coul_long_cs_ext.cpp -I$(OBJ_DIR)
+
$(BIN_DIR)/nvc_get_devices: ./geryon/ucl_get_devices.cpp $(NVD_H)
$(CUDR) -o $@ ./geryon/ucl_get_devices.cpp -DUCL_CUDADR $(CUDA_LIB) -lcuda
diff --git a/lib/gpu/Opencl.makefile b/lib/gpu/Opencl.makefile
index 3e73e6256c7b998c449a60e32e2a3974f636b227..e6e84a0cf4c8907c4f24e2f6519dda79404d39e5 100644
--- a/lib/gpu/Opencl.makefile
+++ b/lib/gpu/Opencl.makefile
@@ -68,7 +68,9 @@ OBJS = $(OBJ_DIR)/lal_atom.o $(OBJ_DIR)/lal_answer.o \
$(OBJ_DIR)/lal_lj_cubic.o $(OBJ_DIR)/lal_lj_cubic_ext.o \
$(OBJ_DIR)/lal_ufm.o $(OBJ_DIR)/lal_ufm_ext.o \
$(OBJ_DIR)/lal_dipole_long_lj.o $(OBJ_DIR)/lal_dipole_long_lj_ext.o \
- $(OBJ_DIR)/lal_lj_expand_coul_long.o $(OBJ_DIR)/lal_lj_expand_coul_long_ext.o
+ $(OBJ_DIR)/lal_lj_expand_coul_long.o $(OBJ_DIR)/lal_lj_expand_coul_long_ext.o \
+ $(OBJ_DIR)/lal_coul_long_cs.o $(OBJ_DIR)/lal_coul_long_cs_ext.o \
+ $(OBJ_DIR)/lal_born_coul_long_cs.o $(OBJ_DIR)/lal_born_coul_long_cs_ext.o
KERS = $(OBJ_DIR)/device_cl.h $(OBJ_DIR)/atom_cl.h \
$(OBJ_DIR)/neighbor_cpu_cl.h $(OBJ_DIR)/pppm_cl.h \
@@ -98,7 +100,8 @@ KERS = $(OBJ_DIR)/device_cl.h $(OBJ_DIR)/atom_cl.h \
$(OBJ_DIR)/coul_debye_cl.h $(OBJ_DIR)/zbl_cl.h \
$(OBJ_DIR)/lj_cubic_cl.h $(OBJ_DIR)/vashishta_cl.h \
$(OBJ_DIR)/ufm_cl.h $(OBJ_DIR)/dipole_long_lj_cl.h \
- $(OBJ_DIR)/lj_expand_coul_long_cl.h
+ $(OBJ_DIR)/lj_expand_coul_long_cl.h $(OBJ_DIR)/coul_long_cs_cl.h \
+ $(OBJ_DIR)/born_coul_long_cs_cl.h
OCL_EXECS = $(BIN_DIR)/ocl_get_devices
@@ -609,6 +612,24 @@ $(OBJ_DIR)/lal_lj_expand_coul_long.o: $(ALL_H) lal_lj_expand_coul_long.h lal_lj_
$(OBJ_DIR)/lal_lj_expand_coul_long_ext.o: $(ALL_H) lal_lj_expand_coul_long.h lal_lj_expand_coul_long_ext.cpp lal_base_charge.h
$(OCL) -o $@ -c lal_lj_expand_coul_long_ext.cpp -I$(OBJ_DIR)
+$(OBJ_DIR)/coul_long_cs_cl.h: lal_coul_long_cs.cu $(PRE1_H)
+ $(BSH) ./geryon/file_to_cstr.sh coul_long_cs $(PRE1_H) lal_coul_long_cs.cu $(OBJ_DIR)/coul_long_cs_cl.h;
+
+$(OBJ_DIR)/lal_coul_long_cs.o: $(ALL_H) lal_coul_long_cs.h lal_coul_long_cs.cpp $(OBJ_DIR)/coul_long_cs_cl.h $(OBJ_DIR)/coul_long_cs_cl.h $(OBJ_DIR)/lal_base_charge.o
+ $(OCL) -o $@ -c lal_coul_long_cs.cpp -I$(OBJ_DIR)
+
+$(OBJ_DIR)/lal_coul_long_cs_ext.o: $(ALL_H) lal_coul_long_cs.h lal_coul_long_cs_ext.cpp lal_base_charge.h
+ $(OCL) -o $@ -c lal_coul_long_cs_ext.cpp -I$(OBJ_DIR)
+
+$(OBJ_DIR)/born_coul_long_cs_cl.h: lal_born_coul_long_cs.cu $(PRE1_H)
+ $(BSH) ./geryon/file_to_cstr.sh born_coul_long_cs $(PRE1_H) lal_born_coul_long_cs.cu $(OBJ_DIR)/born_coul_long_cs_cl.h;
+
+$(OBJ_DIR)/lal_born_coul_long_cs.o: $(ALL_H) lal_born_coul_long_cs.h lal_born_coul_long_cs.cpp $(OBJ_DIR)/born_coul_long_cs_cl.h $(OBJ_DIR)/born_coul_long_cs_cl.h $(OBJ_DIR)/lal_base_charge.o
+ $(OCL) -o $@ -c lal_born_coul_long_cs.cpp -I$(OBJ_DIR)
+
+$(OBJ_DIR)/lal_born_coul_long_cs_ext.o: $(ALL_H) lal_born_coul_long_cs.h lal_born_coul_long_cs_ext.cpp lal_base_charge.h
+ $(OCL) -o $@ -c lal_born_coul_long_cs_ext.cpp -I$(OBJ_DIR)
+
$(BIN_DIR)/ocl_get_devices: ./geryon/ucl_get_devices.cpp $(OCL_H)
$(OCL) -o $@ ./geryon/ucl_get_devices.cpp -DUCL_OPENCL $(OCL_LINK)
diff --git a/lib/gpu/lal_born_coul_long.cpp b/lib/gpu/lal_born_coul_long.cpp
index 68695c49380ea9a0810a98ca0273686c6d3eb61c..116d44d58fd340e13b75439dab4d380dd13de89b 100644
--- a/lib/gpu/lal_born_coul_long.cpp
+++ b/lib/gpu/lal_born_coul_long.cpp
@@ -57,7 +57,7 @@ int BornCoulLongT::init(const int ntypes, double **host_cutsq, double **host_rho
const double g_ewald) {
int success;
success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
- _screen,born_coul_long,"k_born_long");
+ _screen,born_coul_long,"k_born_coul_long");
if (success!=0)
return success;
diff --git a/lib/gpu/lal_born_coul_long.cu b/lib/gpu/lal_born_coul_long.cu
index 4cb4ea448fbf74374ce8b7ca7bd95b7cfda6505a..71e5e0ae502e8847bc006788deaf82da9477868d 100644
--- a/lib/gpu/lal_born_coul_long.cu
+++ b/lib/gpu/lal_born_coul_long.cu
@@ -29,7 +29,7 @@ texture<int2> q_tex;
#define q_tex q_
#endif
-__kernel void k_born_long(const __global numtyp4 *restrict x_,
+__kernel void k_born_coul_long(const __global numtyp4 *restrict x_,
const __global numtyp4 *restrict coeff1,
const __global numtyp4 *restrict coeff2,
const int lj_types,
@@ -110,7 +110,7 @@ __kernel void k_born_long(const __global numtyp4 *restrict x_,
forcecoul = prefactor * (_erfc + EWALD_F*grij*expm2-factor_coul);
} else forcecoul = (numtyp)0.0;
- if (rsq < cutsq_sigma[mtype].y) {
+ if (rsq < cutsq_sigma[mtype].y) { // cut_ljsq
numtyp r = ucl_sqrt(rsq);
rexp = ucl_exp((cutsq_sigma[mtype].z-r)*coeff1[mtype].x);
r6inv = r2inv*r2inv*r2inv;
@@ -127,7 +127,7 @@ __kernel void k_born_long(const __global numtyp4 *restrict x_,
if (eflag>0) {
if (rsq < cut_coulsq)
e_coul += prefactor*(_erfc-factor_coul);
- if (rsq < coeff1[mtype].w) {
+ if (rsq < cutsq_sigma[mtype].y) {
numtyp e=coeff2[mtype].x*rexp - coeff2[mtype].y*r6inv
+ coeff2[mtype].z*r2inv*r6inv;
energy+=factor_lj*(e-coeff2[mtype].w);
@@ -149,7 +149,7 @@ __kernel void k_born_long(const __global numtyp4 *restrict x_,
} // if ii
}
-__kernel void k_born_long_fast(const __global numtyp4 *restrict x_,
+__kernel void k_born_coul_long_fast(const __global numtyp4 *restrict x_,
const __global numtyp4 *restrict coeff1_in,
const __global numtyp4 *restrict coeff2_in,
const __global numtyp *restrict sp_lj_in,
@@ -232,7 +232,7 @@ __kernel void k_born_long_fast(const __global numtyp4 *restrict x_,
forcecoul = prefactor * (_erfc + EWALD_F*grij*expm2-factor_coul);
} else forcecoul = (numtyp)0.0;
- if (rsq < cutsq_sigma[mtype].y) {
+ if (rsq < cutsq_sigma[mtype].y) { // cut_ljsq
numtyp r = ucl_sqrt(rsq);
rexp = ucl_exp((cutsq_sigma[mtype].z-r)*coeff1[mtype].x);
r6inv = r2inv*r2inv*r2inv;
@@ -249,7 +249,7 @@ __kernel void k_born_long_fast(const __global numtyp4 *restrict x_,
if (eflag>0) {
if (rsq < cut_coulsq)
e_coul += prefactor*(_erfc-factor_coul);
- if (rsq < coeff1[mtype].w) {
+ if (rsq < cutsq_sigma[mtype].y) {
numtyp e=coeff2[mtype].x*rexp - coeff2[mtype].y*r6inv
+ coeff2[mtype].z*r2inv*r6inv;
energy+=factor_lj*(e-coeff2[mtype].w);
diff --git a/lib/gpu/lal_born_coul_long_cs.cpp b/lib/gpu/lal_born_coul_long_cs.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..99e71a75d44b65bf0fb851eb2e084d406ae0566d
--- /dev/null
+++ b/lib/gpu/lal_born_coul_long_cs.cpp
@@ -0,0 +1,175 @@
+/***************************************************************************
+ born_coul_long_cs.cpp
+ -------------------
+ Trung Dac Nguyen (Northwestern)
+
+ Class for acceleration of the born/coul/long/cs pair style.
+
+ __________________________________________________________________________
+ This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+ __________________________________________________________________________
+
+ begin :
+ email : trung.nguyen@northwestern.edu
+ ***************************************************************************/
+
+#ifdef USE_OPENCL
+#include "born_coul_long_cs_cl.h"
+#elif defined(USE_CUDART)
+const char *born_coul_long_cs=0;
+#else
+#include "born_coul_long_cs_cubin.h"
+#endif
+
+#include "lal_born_coul_long_cs.h"
+#include <cassert>
+using namespace LAMMPS_AL;
+#define BornCoulLongCST BornCoulLongCS<numtyp, acctyp>
+
+extern Device<PRECISION,ACC_PRECISION> device;
+
+template <class numtyp, class acctyp>
+BornCoulLongCST::BornCoulLongCS() : BaseCharge<numtyp,acctyp>(),
+ _allocated(false) {
+}
+
+template <class numtyp, class acctyp>
+BornCoulLongCST::~BornCoulLongCS() {
+ clear();
+}
+
+template <class numtyp, class acctyp>
+int BornCoulLongCST::bytes_per_atom(const int max_nbors) const {
+ return this->bytes_per_atom_atomic(max_nbors);
+}
+
+template <class numtyp, class acctyp>
+int BornCoulLongCST::init(const int ntypes, double **host_cutsq, double **host_rhoinv,
+ double **host_born1, double **host_born2, double **host_born3,
+ double **host_a, double **host_c, double **host_d,
+ double **host_sigma, double **host_offset,
+ double *host_special_lj, const int nlocal,
+ const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size,
+ const double gpu_split, FILE *_screen,
+ double **host_cut_ljsq, const double host_cut_coulsq,
+ double *host_special_coul, const double qqrd2e,
+ const double g_ewald) {
+ int success;
+ success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
+ _screen,born_coul_long_cs,"k_born_coul_long_cs");
+ if (success!=0)
+ return success;
+
+ // If atom type constants fit in shared memory use fast kernel
+ int lj_types=ntypes;
+ shared_types=false;
+ int max_shared_types=this->device->max_shared_types();
+ if (lj_types<=max_shared_types && this->_block_size>=max_shared_types) {
+ lj_types=max_shared_types;
+ shared_types=true;
+ }
+ _lj_types=lj_types;
+
+ // Allocate a host write buffer for data initialization
+ UCL_H_Vec<numtyp> host_write(lj_types*lj_types*32,*(this->ucl_device),
+ UCL_WRITE_ONLY);
+
+ for (int i=0; i<lj_types*lj_types; i++)
+ host_write[i]=0.0;
+
+ coeff1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
+ this->atom->type_pack4(ntypes,lj_types,coeff1,host_write,host_rhoinv,
+ host_born1,host_born2,host_born3);
+
+ coeff2.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
+ this->atom->type_pack4(ntypes,lj_types,coeff2,host_write,host_a,host_c,
+ host_d,host_offset);
+
+ cutsq_sigma.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
+ this->atom->type_pack4(ntypes,lj_types,cutsq_sigma,host_write,host_cutsq,
+ host_cut_ljsq,host_sigma);
+
+ sp_lj.alloc(8,*(this->ucl_device),UCL_READ_ONLY);
+ for (int i=0; i<4; i++) {
+ host_write[i]=host_special_lj[i];
+ host_write[i+4]=host_special_coul[i];
+ }
+ ucl_copy(sp_lj,host_write,8,false);
+
+ _cut_coulsq=host_cut_coulsq;
+ _qqrd2e=qqrd2e;
+ _g_ewald=g_ewald;
+
+ _allocated=true;
+ this->_max_bytes=coeff1.row_bytes()+coeff2.row_bytes()
+ +cutsq_sigma.row_bytes()+sp_lj.row_bytes();
+ return 0;
+}
+
+template <class numtyp, class acctyp>
+void BornCoulLongCST::clear() {
+ if (!_allocated)
+ return;
+ _allocated=false;
+
+ coeff1.clear();
+ coeff2.clear();
+ cutsq_sigma.clear();
+ sp_lj.clear();
+ this->clear_atomic();
+}
+
+template <class numtyp, class acctyp>
+double BornCoulLongCST::host_memory_usage() const {
+ return this->host_memory_usage_atomic()+sizeof(BornCoulLongCS<numtyp,acctyp>);
+}
+
+// ---------------------------------------------------------------------------
+// Calculate energies, forces, and torques
+// ---------------------------------------------------------------------------
+template <class numtyp, class acctyp>
+void BornCoulLongCST::loop(const bool _eflag, const bool _vflag) {
+ // Compute the block size and grid size to keep all cores busy
+ const int BX=this->block_size();
+ int eflag, vflag;
+ if (_eflag)
+ eflag=1;
+ else
+ eflag=0;
+
+ if (_vflag)
+ vflag=1;
+ else
+ vflag=0;
+
+ int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
+ (BX/this->_threads_per_atom)));
+
+ int ainum=this->ans->inum();
+ int nbor_pitch=this->nbor->nbor_pitch();
+ this->time_pair.start();
+ if (shared_types) {
+ this->k_pair_fast.set_size(GX,BX);
+ this->k_pair_fast.run(&this->atom->x, &coeff1, &coeff2, &sp_lj,
+ &this->nbor->dev_nbor,
+ &this->_nbor_data->begin(),
+ &this->ans->force,
+ &this->ans->engv, &eflag, &vflag,
+ &ainum, &nbor_pitch, &this->atom->q,
+ &cutsq_sigma, &_cut_coulsq, &_qqrd2e,
+ &_g_ewald, &this->_threads_per_atom);
+ } else {
+ this->k_pair.set_size(GX,BX);
+ this->k_pair.run(&this->atom->x, &coeff1, &coeff2, &_lj_types, &sp_lj,
+ &this->nbor->dev_nbor, &this->_nbor_data->begin(),
+ &this->ans->force, &this->ans->engv,
+ &eflag, &vflag, &ainum,
+ &nbor_pitch, &this->atom->q,
+ &cutsq_sigma, &_cut_coulsq,
+ &_qqrd2e, &_g_ewald, &this->_threads_per_atom);
+ }
+ this->time_pair.stop();
+}
+
+template class BornCoulLongCS<PRECISION,ACC_PRECISION>;
diff --git a/lib/gpu/lal_born_coul_long_cs.cu b/lib/gpu/lal_born_coul_long_cs.cu
new file mode 100644
index 0000000000000000000000000000000000000000..d18f573c11332d2846f25f99802a35fc390627b6
--- /dev/null
+++ b/lib/gpu/lal_born_coul_long_cs.cu
@@ -0,0 +1,325 @@
+// **************************************************************************
+// born_coul_long_cs.cu
+// -------------------
+// Trung Dac Nguyen (Northwestern)
+//
+// Device code for acceleration of the born/coul/long/cs pair style
+//
+// __________________________________________________________________________
+// This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+// __________________________________________________________________________
+//
+// begin :
+// email : trung.nguyen@northwestern.edu
+// ***************************************************************************/
+
+#ifdef NV_KERNEL
+
+#include "lal_aux_fun1.h"
+#ifndef _DOUBLE_DOUBLE
+texture<float4> pos_tex;
+texture<float> q_tex;
+#else
+texture<int4,1> pos_tex;
+texture<int2> q_tex;
+#endif
+
+#else
+#define pos_tex x_
+#define q_tex q_
+#endif
+
+#define B0 (numtyp)-0.1335096380159268
+#define B1 (numtyp)-2.57839507e-1
+#define B2 (numtyp)-1.37203639e-1
+#define B3 (numtyp)-8.88822059e-3
+#define B4 (numtyp)-5.80844129e-3
+#define B5 (numtyp)1.14652755e-1
+#define EPSILON (numtyp)1.0e-20
+#define EPS_EWALD (numtyp)1.0e-6
+#define EPS_EWALD_SQR (numtyp)1.0e-12
+
+__kernel void k_born_coul_long_cs(const __global numtyp4 *restrict x_,
+ const __global numtyp4 *restrict coeff1,
+ const __global numtyp4 *restrict coeff2,
+ const int lj_types,
+ const __global numtyp *restrict sp_lj_in,
+ const __global int *dev_nbor,
+ const __global int *dev_packed,
+ __global acctyp4 *restrict ans,
+ __global acctyp *restrict engv,
+ const int eflag, const int vflag, const int inum,
+ const int nbor_pitch,
+ const __global numtyp *restrict q_,
+ const __global numtyp4 *restrict cutsq_sigma,
+ const numtyp cut_coulsq, const numtyp qqrd2e,
+ const numtyp g_ewald, const int t_per_atom) {
+ int tid, ii, offset;
+ atom_info(t_per_atom,ii,tid,offset);
+
+ __local numtyp sp_lj[8];
+ sp_lj[0]=sp_lj_in[0];
+ sp_lj[1]=sp_lj_in[1];
+ sp_lj[2]=sp_lj_in[2];
+ sp_lj[3]=sp_lj_in[3];
+ sp_lj[4]=sp_lj_in[4];
+ sp_lj[5]=sp_lj_in[5];
+ sp_lj[6]=sp_lj_in[6];
+ sp_lj[7]=sp_lj_in[7];
+
+ acctyp energy=(acctyp)0;
+ acctyp e_coul=(acctyp)0;
+ acctyp4 f;
+ f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
+ acctyp virial[6];
+ for (int i=0; i<6; i++)
+ virial[i]=(acctyp)0;
+
+ if (ii<inum) {
+ int nbor, nbor_end;
+ int i, numj;
+ __local int n_stride;
+ nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
+ n_stride,nbor_end,nbor);
+
+ numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
+ numtyp qtmp; fetch(qtmp,i,q_tex);
+ int itype=ix.w;
+
+ for ( ; nbor<nbor_end; nbor+=n_stride) {
+ int j=dev_packed[nbor];
+
+ numtyp factor_lj, factor_coul;
+ factor_lj = sp_lj[sbmask(j)];
+ factor_coul = sp_lj[sbmask(j)+4];
+ j &= NEIGHMASK;
+
+ numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
+ int jtype=jx.w;
+
+ // Compute r12
+ numtyp delx = ix.x-jx.x;
+ numtyp dely = ix.y-jx.y;
+ numtyp delz = ix.z-jx.z;
+ numtyp rsq = delx*delx+dely*dely+delz*delz;
+
+ int mtype=itype*lj_types+jtype;
+ if (rsq<cutsq_sigma[mtype].x) { // cutsq
+ numtyp forcecoul, forceborn, force, r6inv, prefactor, _erfc;
+ numtyp rexp = (numtyp)0.0;
+
+ if (rsq < cut_coulsq) {
+ rsq += EPSILON;
+
+ numtyp r2inv = ucl_recip(rsq);
+ numtyp r = ucl_rsqrt(r2inv);
+ fetch(prefactor,j,q_tex);
+ prefactor *= qqrd2e * qtmp;
+ if (factor_coul<(numtyp)1.0) {
+ numtyp grij = g_ewald * (r+EPS_EWALD);
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= (r+EPS_EWALD);
+ forcecoul = prefactor * (_erfc + EWALD_F*grij*expm2 - ((numtyp)1.0-factor_coul));
+ r2inv = ucl_recip(rsq + EPS_EWALD_SQR);
+ forcecoul *= r2inv;
+ } else {
+ numtyp grij = g_ewald * r;
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= r;
+ forcecoul = prefactor*(_erfc + EWALD_F*grij*expm2);
+ forcecoul *= r2inv;
+ }
+ } else forcecoul = (numtyp)0.0;
+
+ numtyp r2inv = ucl_recip(rsq);
+ if (rsq < cutsq_sigma[mtype].y) { // cut_ljsq
+ numtyp r = ucl_sqrt(rsq);
+ rexp = ucl_exp((cutsq_sigma[mtype].z-r)*coeff1[mtype].x);
+ r6inv = r2inv*r2inv*r2inv;
+ forceborn = (coeff1[mtype].y*r*rexp - coeff1[mtype].z*r6inv
+ + coeff1[mtype].w*r2inv*r6inv)*factor_lj;
+ } else forceborn = (numtyp)0.0;
+
+ force = forceborn*r2inv + forcecoul;
+
+ f.x+=delx*force;
+ f.y+=dely*force;
+ f.z+=delz*force;
+
+ if (eflag>0) {
+ if (rsq < cut_coulsq) {
+ e_coul += prefactor*_erfc;
+ if (factor_coul<(numtyp)1.0)
+ e_coul -= ((numtyp)1.0-factor_coul)*prefactor;
+ }
+ if (rsq < cutsq_sigma[mtype].y) {
+ numtyp e=coeff2[mtype].x*rexp - coeff2[mtype].y*r6inv
+ + coeff2[mtype].z*r2inv*r6inv;
+ energy+=factor_lj*(e-coeff2[mtype].w);
+ }
+ }
+ if (vflag>0) {
+ virial[0] += delx*delx*force;
+ virial[1] += dely*dely*force;
+ virial[2] += delz*delz*force;
+ virial[3] += delx*dely*force;
+ virial[4] += delx*delz*force;
+ virial[5] += dely*delz*force;
+ }
+ }
+
+ } // for nbor
+ store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
+ vflag,ans,engv);
+ } // if ii
+}
+
+__kernel void k_born_coul_long_cs_fast(const __global numtyp4 *restrict x_,
+ const __global numtyp4 *restrict coeff1_in,
+ const __global numtyp4 *restrict coeff2_in,
+ const __global numtyp *restrict sp_lj_in,
+ const __global int *dev_nbor,
+ const __global int *dev_packed,
+ __global acctyp4 *restrict ans,
+ __global acctyp *restrict engv,
+ const int eflag, const int vflag, const int inum,
+ const int nbor_pitch,
+ const __global numtyp *restrict q_,
+ const __global numtyp4 *restrict cutsq_sigma,
+ const numtyp cut_coulsq, const numtyp qqrd2e,
+ const numtyp g_ewald, const int t_per_atom) {
+ int tid, ii, offset;
+ atom_info(t_per_atom,ii,tid,offset);
+
+ __local numtyp4 coeff1[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
+ __local numtyp4 coeff2[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
+ __local numtyp sp_lj[8];
+ if (tid<8)
+ sp_lj[tid]=sp_lj_in[tid];
+ if (tid<MAX_SHARED_TYPES*MAX_SHARED_TYPES) {
+ coeff1[tid]=coeff1_in[tid];
+ if (eflag>0)
+ coeff2[tid]=coeff2_in[tid];
+ }
+
+ acctyp energy=(acctyp)0;
+ acctyp e_coul=(acctyp)0;
+ acctyp4 f;
+ f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
+ acctyp virial[6];
+ for (int i=0; i<6; i++)
+ virial[i]=(acctyp)0;
+
+ __syncthreads();
+
+ if (ii<inum) {
+ int nbor, nbor_end;
+ int i, numj;
+ __local int n_stride;
+ nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
+ n_stride,nbor_end,nbor);
+
+ numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
+ numtyp qtmp; fetch(qtmp,i,q_tex);
+ int iw=ix.w;
+ int itype=fast_mul((int)MAX_SHARED_TYPES,iw);
+
+ for ( ; nbor<nbor_end; nbor+=n_stride) {
+ int j=dev_packed[nbor];
+
+ numtyp factor_lj, factor_coul;
+ factor_lj = sp_lj[sbmask(j)];
+ factor_coul = sp_lj[sbmask(j)+4];
+ j &= NEIGHMASK;
+
+ numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
+ int mtype=itype+jx.w;
+
+ // Compute r12
+ numtyp delx = ix.x-jx.x;
+ numtyp dely = ix.y-jx.y;
+ numtyp delz = ix.z-jx.z;
+ numtyp rsq = delx*delx+dely*dely+delz*delz;
+
+ if (rsq<cutsq_sigma[mtype].x) { // cutsq
+ numtyp forcecoul, forceborn, force, r6inv, prefactor, _erfc;
+ numtyp rexp = (numtyp)0.0;
+
+ if (rsq < cut_coulsq) {
+ rsq += EPSILON;
+
+ numtyp r2inv = ucl_recip(rsq);
+ numtyp r = ucl_rsqrt(r2inv);
+ fetch(prefactor,j,q_tex);
+ prefactor *= qqrd2e * qtmp;
+ if (factor_coul<(numtyp)1.0) {
+ numtyp grij = g_ewald * (r+EPS_EWALD);
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= (r+EPS_EWALD);
+ forcecoul = prefactor * (_erfc + EWALD_F*grij*expm2 - ((numtyp)1.0-factor_coul));
+ r2inv = ucl_recip(rsq + EPS_EWALD_SQR);
+ forcecoul *= r2inv;
+ } else {
+ numtyp grij = g_ewald * r;
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= r;
+ forcecoul = prefactor*(_erfc + EWALD_F*grij*expm2);
+ forcecoul *= r2inv;
+ }
+ } else forcecoul = (numtyp)0.0;
+
+ numtyp r2inv = ucl_recip(rsq);
+ if (rsq < cutsq_sigma[mtype].y) { // cut_ljsq
+ numtyp r = ucl_sqrt(rsq);
+ rexp = ucl_exp((cutsq_sigma[mtype].z-r)*coeff1[mtype].x);
+ r6inv = r2inv*r2inv*r2inv;
+ forceborn = (coeff1[mtype].y*r*rexp - coeff1[mtype].z*r6inv
+ + coeff1[mtype].w*r2inv*r6inv)*factor_lj;
+ } else forceborn = (numtyp)0.0;
+
+ force = forceborn*r2inv + forcecoul;
+
+ f.x+=delx*force;
+ f.y+=dely*force;
+ f.z+=delz*force;
+
+ if (eflag>0) {
+ if (rsq < cut_coulsq) {
+ e_coul += prefactor*_erfc;
+ if (factor_coul<(numtyp)1.0)
+ e_coul -= ((numtyp)1.0-factor_coul)*prefactor;
+ }
+ if (rsq < cutsq_sigma[mtype].y) {
+ numtyp e=coeff2[mtype].x*rexp - coeff2[mtype].y*r6inv
+ + coeff2[mtype].z*r2inv*r6inv;
+ energy+=factor_lj*(e-coeff2[mtype].w);
+ }
+ }
+ if (vflag>0) {
+ virial[0] += delx*delx*force;
+ virial[1] += dely*dely*force;
+ virial[2] += delz*delz*force;
+ virial[3] += delx*dely*force;
+ virial[4] += delx*delz*force;
+ virial[5] += dely*delz*force;
+ }
+ }
+
+ } // for nbor
+ store_answers_q(f,energy,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
+ vflag,ans,engv);
+ } // if ii
+}
+
diff --git a/lib/gpu/lal_born_coul_long_cs.h b/lib/gpu/lal_born_coul_long_cs.h
new file mode 100644
index 0000000000000000000000000000000000000000..2b383dbea952a7a33d4ef4bda4aeebcaf25ef8a2
--- /dev/null
+++ b/lib/gpu/lal_born_coul_long_cs.h
@@ -0,0 +1,88 @@
+/***************************************************************************
+ born_coul_long.h
+ -------------------
+ Trung Dac Nguyen (ORNL)
+
+ Class for acceleration of the born/coul/long/cs pair style.
+
+ __________________________________________________________________________
+ This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+ __________________________________________________________________________
+
+ begin :
+ email : trung.nguyen@northwestern.edu
+ ***************************************************************************/
+
+#ifndef LAL_BORN_COUL_LONG_CS_H
+#define LAL_BORN_COUL_LONG_CS_H
+
+#include "lal_base_charge.h"
+
+namespace LAMMPS_AL {
+
+template <class numtyp, class acctyp>
+class BornCoulLongCS : public BaseCharge<numtyp, acctyp> {
+ public:
+ BornCoulLongCS();
+ ~BornCoulLongCS();
+
+ /// Clear any previous data and set up for a new LAMMPS run
+ /** \param max_nbors initial number of rows in the neighbor matrix
+ * \param cell_size cutoff + skin
+ * \param gpu_split fraction of particles handled by device
+ *
+ * Returns:
+ * - 0 if successfull
+ * - -1 if fix gpu not found
+ * - -3 if there is an out of memory error
+ * - -4 if the GPU library was not compiled for GPU
+ * - -5 Double precision is not supported on card **/
+ int init(const int ntypes, double **host_cutsq, double **host_rhoinv,
+ double **host_born1, double **host_born2, double **host_born3,
+ double **host_a, double **host_c, double **host_d,
+ double **host_sigma, double **host_offset, double *host_special_lj,
+ const int nlocal, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size,
+ const double gpu_split, FILE *screen, double **host_cut_ljsq,
+ const double host_cut_coulsq, double *host_special_coul,
+ const double qqrd2e, const double g_ewald);
+
+ /// Clear all host and device data
+ /** \note This is called at the beginning of the init() routine **/
+ void clear();
+
+ /// Returns memory usage on device per atom
+ int bytes_per_atom(const int max_nbors) const;
+
+ /// Total host memory used by library for pair style
+ double host_memory_usage() const;
+
+ // --------------------------- TYPE DATA --------------------------
+
+ /// coeff1.x = rhoinv, coeff1.y = born1, coeff1.z = born2,
+ /// coeff1.w = born3
+ UCL_D_Vec<numtyp4> coeff1;
+ /// coeff2.x = a, coeff2.y = c, coeff2.z = d, coeff2.w = offset
+ UCL_D_Vec<numtyp4> coeff2;
+ /// cutsq_sigma.x = cutsq, cutsq_sigma.y = cutsq_lj,
+ /// cutsq_sigma.z = sigma
+ UCL_D_Vec<numtyp4> cutsq_sigma;
+ /// Special LJ values [0-3] and Special Coul values [4-7]
+ UCL_D_Vec<numtyp> sp_lj;
+
+ /// If atom type constants fit in shared memory, use fast kernels
+ bool shared_types;
+
+ /// Number of atom types
+ int _lj_types;
+
+ numtyp _cut_coulsq, _qqrd2e, _g_ewald;
+
+ private:
+ bool _allocated;
+ void loop(const bool _eflag, const bool _vflag);
+};
+
+}
+
+#endif
diff --git a/lib/gpu/lal_born_coul_long_cs_ext.cpp b/lib/gpu/lal_born_coul_long_cs_ext.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..43363e217eec3e3d8d29802b2d224d1857ebeb79
--- /dev/null
+++ b/lib/gpu/lal_born_coul_long_cs_ext.cpp
@@ -0,0 +1,132 @@
+/***************************************************************************
+ born_coul_long_cs_ext.cpp
+ -------------------
+ Trung Dac Nguyen (ORNL)
+
+ Functions for LAMMPS access to born/coul/long/cs acceleration routines.
+
+ __________________________________________________________________________
+ This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+ __________________________________________________________________________
+
+ begin :
+ email : trung.nguyen@northwestern.edu
+ ***************************************************************************/
+
+#include <iostream>
+#include <cassert>
+#include <cmath>
+
+#include "lal_born_coul_long_cs.h"
+
+using namespace std;
+using namespace LAMMPS_AL;
+
+static BornCoulLongCS<PRECISION,ACC_PRECISION> BCLCSMF;
+
+// ---------------------------------------------------------------------------
+// Allocate memory on host and device and copy constants to device
+// ---------------------------------------------------------------------------
+int bornclcs_gpu_init(const int ntypes, double **cutsq, double **host_rhoinv,
+ double **host_born1, double **host_born2, double **host_born3,
+ double **host_a, double **host_c, double **host_d,
+ double **sigma, double **offset, double *special_lj,
+ const int inum, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size, int &gpu_mode,
+ FILE *screen, double **host_cut_ljsq, double host_cut_coulsq,
+ double *host_special_coul, const double qqrd2e,
+ const double g_ewald) {
+ BCLCSMF.clear();
+ gpu_mode=BCLCSMF.device->gpu_mode();
+ double gpu_split=BCLCSMF.device->particle_split();
+ int first_gpu=BCLCSMF.device->first_device();
+ int last_gpu=BCLCSMF.device->last_device();
+ int world_me=BCLCSMF.device->world_me();
+ int gpu_rank=BCLCSMF.device->gpu_rank();
+ int procs_per_gpu=BCLCSMF.device->procs_per_gpu();
+
+ BCLCSMF.device->init_message(screen,"born/coul/long/cs",first_gpu,last_gpu);
+
+ bool message=false;
+ if (BCLCSMF.device->replica_me()==0 && screen)
+ message=true;
+
+ if (message) {
+ fprintf(screen,"Initializing Device and compiling on process 0...");
+ fflush(screen);
+ }
+
+ int init_ok=0;
+ if (world_me==0)
+ init_ok=BCLCSMF.init(ntypes, cutsq, host_rhoinv, host_born1, host_born2,
+ host_born3, host_a, host_c, host_d, sigma, offset,
+ special_lj, inum, nall, 300, maxspecial, cell_size,
+ gpu_split, screen, host_cut_ljsq, host_cut_coulsq,
+ host_special_coul, qqrd2e, g_ewald);
+
+ BCLCSMF.device->world_barrier();
+ if (message)
+ fprintf(screen,"Done.\n");
+
+ for (int i=0; i<procs_per_gpu; i++) {
+ if (message) {
+ if (last_gpu-first_gpu==0)
+ fprintf(screen,"Initializing Device %d on core %d...",first_gpu,i);
+ else
+ fprintf(screen,"Initializing Devices %d-%d on core %d...",first_gpu,
+ last_gpu,i);
+ fflush(screen);
+ }
+ if (gpu_rank==i && world_me!=0)
+ init_ok=BCLCSMF.init(ntypes, cutsq, host_rhoinv, host_born1, host_born2,
+ host_born3, host_a, host_c, host_d, sigma, offset,
+ special_lj, inum, nall, 300, maxspecial, cell_size,
+ gpu_split, screen, host_cut_ljsq, host_cut_coulsq,
+ host_special_coul, qqrd2e, g_ewald);
+
+ BCLCSMF.device->gpu_barrier();
+ if (message)
+ fprintf(screen,"Done.\n");
+ }
+ if (message)
+ fprintf(screen,"\n");
+
+ if (init_ok==0)
+ BCLCSMF.estimate_gpu_overhead();
+ return init_ok;
+}
+
+void bornclcs_gpu_clear() {
+ BCLCSMF.clear();
+}
+
+int** bornclcs_gpu_compute_n(const int ago, const int inum_full,
+ const int nall, double **host_x, int *host_type,
+ double *sublo, double *subhi, tagint *tag, int **nspecial,
+ tagint **special, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ int **ilist, int **jnum, const double cpu_time,
+ bool &success, double *host_q, double *boxlo,
+ double *prd) {
+ return BCLCSMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
+ subhi, tag, nspecial, special, eflag, vflag, eatom,
+ vatom, host_start, ilist, jnum, cpu_time, success,
+ host_q, boxlo, prd);
+}
+
+void bornclcs_gpu_compute(const int ago, const int inum_full, const int nall,
+ double **host_x, int *host_type, int *ilist, int *numj,
+ int **firstneigh, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ const double cpu_time, bool &success, double *host_q,
+ const int nlocal, double *boxlo, double *prd) {
+ BCLCSMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,
+ firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success,
+ host_q,nlocal,boxlo,prd);
+}
+
+double bornclcs_gpu_bytes() {
+ return BCLCSMF.host_memory_usage();
+}
+
+
diff --git a/lib/gpu/lal_coul_long_cs.cpp b/lib/gpu/lal_coul_long_cs.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..7b88647fccc414c532d9709f16470f8194bc23b0
--- /dev/null
+++ b/lib/gpu/lal_coul_long_cs.cpp
@@ -0,0 +1,158 @@
+/***************************************************************************
+ coul_long_cs.cpp
+ -------------------
+ Trung Nguyen (Northwestern)
+
+ Class for acceleration of the coul/long pair style.
+
+ __________________________________________________________________________
+ This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+ __________________________________________________________________________
+
+ begin : June 2018
+ email : trung.nguyen@northwestern.edu
+ ***************************************************************************/
+
+#if defined(USE_OPENCL)
+#include "coul_long_cs_cl.h"
+#elif defined(USE_CUDART)
+const char *coul_long_cs=0;
+#else
+#include "coul_long_cs_cubin.h"
+#endif
+
+#include "lal_coul_long_cs.h"
+#include <cassert>
+using namespace LAMMPS_AL;
+#define CoulLongCST CoulLongCS<numtyp, acctyp>
+
+extern Device<PRECISION,ACC_PRECISION> pair_gpu_device;
+
+template <class numtyp, class acctyp>
+CoulLongCST::CoulLongCS() : BaseCharge<numtyp,acctyp>(), _allocated(false) {
+}
+
+template <class numtyp, class acctyp>
+CoulLongCST::~CoulLongCS() {
+ clear();
+}
+
+template <class numtyp, class acctyp>
+int CoulLongCST::bytes_per_atom(const int max_nbors) const {
+ return this->bytes_per_atom_atomic(max_nbors);
+}
+
+template <class numtyp, class acctyp>
+int CoulLongCST::init(const int ntypes, double **host_scale,
+ const int nlocal, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size,
+ const double gpu_split, FILE *_screen,
+ const double host_cut_coulsq, double *host_special_coul,
+ const double qqrd2e, const double g_ewald) {
+ int success;
+ success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,
+ gpu_split,_screen,coul_long_cs,"k_coul_long_cs");
+ if (success!=0)
+ return success;
+
+ int lj_types=ntypes;
+ shared_types=false;
+ int max_shared_types=this->device->max_shared_types();
+ if (lj_types<=max_shared_types && this->_block_size>=max_shared_types) {
+ lj_types=max_shared_types;
+ shared_types=true;
+ }
+ _lj_types=lj_types;
+
+ // Allocate a host write buffer for data initialization
+ UCL_H_Vec<numtyp> host_write(lj_types*lj_types*32,*(this->ucl_device),
+ UCL_WRITE_ONLY);
+
+ for (int i=0; i<lj_types*lj_types; i++)
+ host_write[i]=0.0;
+
+ scale.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
+ this->atom->type_pack1(ntypes,lj_types,scale,host_write,host_scale);
+
+ sp_cl.alloc(4,*(this->ucl_device),UCL_READ_ONLY);
+ for (int i=0; i<4; i++) {
+ host_write[i]=host_special_coul[i];
+ }
+ ucl_copy(sp_cl,host_write,4,false);
+
+ _cut_coulsq=host_cut_coulsq;
+ _qqrd2e=qqrd2e;
+ _g_ewald=g_ewald;
+
+ _allocated=true;
+ this->_max_bytes=scale.row_bytes()+sp_cl.row_bytes();
+ return 0;
+}
+
+template <class numtyp, class acctyp>
+void CoulLongCST::reinit(const int ntypes, double **host_scale) {
+ UCL_H_Vec<numtyp> hscale(_lj_types*_lj_types,*(this->ucl_device),
+ UCL_WRITE_ONLY);
+ this->atom->type_pack1(ntypes,_lj_types,scale,hscale,host_scale);
+}
+
+template <class numtyp, class acctyp>
+void CoulLongCST::clear() {
+ if (!_allocated)
+ return;
+ _allocated=false;
+
+ scale.clear();
+ sp_cl.clear();
+ this->clear_atomic();
+}
+
+template <class numtyp, class acctyp>
+double CoulLongCST::host_memory_usage() const {
+ return this->host_memory_usage_atomic()+sizeof(CoulLongCS<numtyp,acctyp>);
+}
+
+// ---------------------------------------------------------------------------
+// Calculate energies, forces, and torques
+// ---------------------------------------------------------------------------
+template <class numtyp, class acctyp>
+void CoulLongCST::loop(const bool _eflag, const bool _vflag) {
+ // Compute the block size and grid size to keep all cores busy
+ const int BX=this->block_size();
+ int eflag, vflag;
+ if (_eflag)
+ eflag=1;
+ else
+ eflag=0;
+
+ if (_vflag)
+ vflag=1;
+ else
+ vflag=0;
+
+ int GX=static_cast<int>(ceil(static_cast<double>(this->ans->inum())/
+ (BX/this->_threads_per_atom)));
+
+ int ainum=this->ans->inum();
+ int nbor_pitch=this->nbor->nbor_pitch();
+ this->time_pair.start();
+ if (shared_types) {
+ this->k_pair_fast.set_size(GX,BX);
+ this->k_pair_fast.run(&this->atom->x, &scale, &sp_cl,
+ &this->nbor->dev_nbor, &this->_nbor_data->begin(),
+ &this->ans->force, &this->ans->engv,
+ &eflag, &vflag, &ainum, &nbor_pitch,
+ &this->atom->q, &_cut_coulsq, &_qqrd2e, &_g_ewald,
+ &this->_threads_per_atom);
+ } else {
+ this->k_pair.set_size(GX,BX);
+ this->k_pair.run(&this->atom->x, &scale, &_lj_types, &sp_cl,
+ &this->nbor->dev_nbor, &this->_nbor_data->begin(),
+ &this->ans->force, &this->ans->engv, &eflag, &vflag,
+ &ainum, &nbor_pitch, &this->atom->q, &_cut_coulsq,
+ &_qqrd2e, &_g_ewald, &this->_threads_per_atom);
+ }
+ this->time_pair.stop();
+}
+
+template class CoulLongCS<PRECISION,ACC_PRECISION>;
diff --git a/lib/gpu/lal_coul_long_cs.cu b/lib/gpu/lal_coul_long_cs.cu
new file mode 100644
index 0000000000000000000000000000000000000000..f6d3ef8d44d68ccee49849cdcbed3b9c2cadfd55
--- /dev/null
+++ b/lib/gpu/lal_coul_long_cs.cu
@@ -0,0 +1,359 @@
+// **************************************************************************
+// coul_long_cs.cu
+// -------------------
+// Trung Nguyen (Northwestern)
+//
+// Device code for acceleration of the coul/long/cs pair style
+//
+// __________________________________________________________________________
+// This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+// __________________________________________________________________________
+//
+// begin : June 2018
+// email : trung.nguyen@northwestern.edu
+// ***************************************************************************/
+
+#ifdef NV_KERNEL
+
+#include "lal_aux_fun1.h"
+#ifndef _DOUBLE_DOUBLE
+texture<float4> pos_tex;
+texture<float> q_tex;
+#else
+texture<int4,1> pos_tex;
+texture<int2> q_tex;
+#endif
+
+#else
+#define pos_tex x_
+#define q_tex q_
+#endif
+
+#if (ARCH < 300)
+
+#define store_answers_lq(f, e_coul, virial, ii, inum, tid, \
+ t_per_atom, offset, eflag, vflag, ans, engv) \
+ if (t_per_atom>1) { \
+ __local acctyp red_acc[6][BLOCK_PAIR]; \
+ \
+ red_acc[0][tid]=f.x; \
+ red_acc[1][tid]=f.y; \
+ red_acc[2][tid]=f.z; \
+ red_acc[3][tid]=e_coul; \
+ \
+ for (unsigned int s=t_per_atom/2; s>0; s>>=1) { \
+ if (offset < s) { \
+ for (int r=0; r<4; r++) \
+ red_acc[r][tid] += red_acc[r][tid+s]; \
+ } \
+ } \
+ \
+ f.x=red_acc[0][tid]; \
+ f.y=red_acc[1][tid]; \
+ f.z=red_acc[2][tid]; \
+ e_coul=red_acc[3][tid]; \
+ \
+ if (vflag>0) { \
+ for (int r=0; r<6; r++) \
+ red_acc[r][tid]=virial[r]; \
+ \
+ for (unsigned int s=t_per_atom/2; s>0; s>>=1) { \
+ if (offset < s) { \
+ for (int r=0; r<6; r++) \
+ red_acc[r][tid] += red_acc[r][tid+s]; \
+ } \
+ } \
+ \
+ for (int r=0; r<6; r++) \
+ virial[r]=red_acc[r][tid]; \
+ } \
+ } \
+ \
+ if (offset==0) { \
+ __global acctyp *ap1=engv+ii; \
+ if (eflag>0) { \
+ *ap1=(acctyp)0; \
+ ap1+=inum; \
+ *ap1=e_coul*(acctyp)0.5; \
+ ap1+=inum; \
+ } \
+ if (vflag>0) { \
+ for (int i=0; i<6; i++) { \
+ *ap1=virial[i]*(acctyp)0.5; \
+ ap1+=inum; \
+ } \
+ } \
+ ans[ii]=f; \
+ }
+
+#else
+
+#define store_answers_lq(f, e_coul, virial, ii, inum, tid, \
+ t_per_atom, offset, eflag, vflag, ans, engv) \
+ if (t_per_atom>1) { \
+ for (unsigned int s=t_per_atom/2; s>0; s>>=1) { \
+ f.x += shfl_xor(f.x, s, t_per_atom); \
+ f.y += shfl_xor(f.y, s, t_per_atom); \
+ f.z += shfl_xor(f.z, s, t_per_atom); \
+ e_coul += shfl_xor(e_coul, s, t_per_atom); \
+ } \
+ if (vflag>0) { \
+ for (unsigned int s=t_per_atom/2; s>0; s>>=1) { \
+ for (int r=0; r<6; r++) \
+ virial[r] += shfl_xor(virial[r], s, t_per_atom); \
+ } \
+ } \
+ } \
+ if (offset==0) { \
+ __global acctyp *ap1=engv+ii; \
+ if (eflag>0) { \
+ *ap1=(acctyp)0; \
+ ap1+=inum; \
+ *ap1=e_coul*(acctyp)0.5; \
+ ap1+=inum; \
+ } \
+ if (vflag>0) { \
+ for (int i=0; i<6; i++) { \
+ *ap1=virial[i]*(acctyp)0.5; \
+ ap1+=inum; \
+ } \
+ } \
+ ans[ii]=f; \
+ }
+
+#endif
+
+#define B0 (numtyp)-0.1335096380159268
+#define B1 (numtyp)-2.57839507e-1
+#define B2 (numtyp)-1.37203639e-1
+#define B3 (numtyp)-8.88822059e-3
+#define B4 (numtyp)-5.80844129e-3
+#define B5 (numtyp)1.14652755e-1
+#define EPSILON (numtyp)1.0e-20
+#define EPS_EWALD (numtyp)1.0e-6
+#define EPS_EWALD_SQR (numtyp)1.0e-12
+
+__kernel void k_coul_long_cs(const __global numtyp4 *restrict x_,
+ const __global numtyp *restrict scale,
+ const int lj_types,
+ const __global numtyp *restrict sp_cl_in,
+ const __global int *dev_nbor,
+ const __global int *dev_packed,
+ __global acctyp4 *restrict ans,
+ __global acctyp *restrict engv,
+ const int eflag, const int vflag, const int inum,
+ const int nbor_pitch,
+ const __global numtyp *restrict q_,
+ const numtyp cut_coulsq, const numtyp qqrd2e,
+ const numtyp g_ewald, const int t_per_atom) {
+ int tid, ii, offset;
+ atom_info(t_per_atom,ii,tid,offset);
+
+ __local numtyp sp_cl[4];
+ sp_cl[0]=sp_cl_in[0];
+ sp_cl[1]=sp_cl_in[1];
+ sp_cl[2]=sp_cl_in[2];
+ sp_cl[3]=sp_cl_in[3];
+
+ acctyp e_coul=(acctyp)0;
+ acctyp4 f;
+ f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
+ acctyp virial[6];
+ for (int i=0; i<6; i++)
+ virial[i]=(acctyp)0;
+
+ if (ii<inum) {
+ int nbor, nbor_end;
+ int i, numj;
+ __local int n_stride;
+ nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
+ n_stride,nbor_end,nbor);
+
+ numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
+ int itype=ix.w;
+ numtyp qtmp; fetch(qtmp,i,q_tex);
+
+ for ( ; nbor<nbor_end; nbor+=n_stride) {
+ int j=dev_packed[nbor];
+
+ numtyp factor_coul;
+ factor_coul = sp_cl[sbmask(j)];
+ j &= NEIGHMASK;
+
+ numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
+ int jtype=jx.w;
+
+ // Compute r12
+ numtyp delx = ix.x-jx.x;
+ numtyp dely = ix.y-jx.y;
+ numtyp delz = ix.z-jx.z;
+ numtyp rsq = delx*delx+dely*dely+delz*delz;
+
+ int mtype=itype*lj_types+jtype;
+ if (rsq < cut_coulsq) {
+ rsq += EPSILON;
+
+ numtyp force,prefactor,_erfc;
+ numtyp r2inv = ucl_recip(rsq);
+ numtyp r = ucl_rsqrt(r2inv);
+ fetch(prefactor,j,q_tex);
+ prefactor *= qqrd2e * scale[mtype] * qtmp;
+ if (factor_coul<(numtyp)1.0) {
+ numtyp grij = g_ewald * (r+EPS_EWALD);
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= (r+EPS_EWALD);
+ force = prefactor * (_erfc + EWALD_F*grij*expm2 - ((numtyp)1.0-factor_coul));
+ r2inv = ucl_recip(rsq + EPS_EWALD_SQR);
+ force *= r2inv;
+ } else {
+ numtyp grij = g_ewald * r;
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= r;
+ force = prefactor*(_erfc + EWALD_F*grij*expm2);
+ force *= r2inv;
+ }
+
+ f.x+=delx*force;
+ f.y+=dely*force;
+ f.z+=delz*force;
+
+ if (eflag>0) {
+ e_coul += prefactor*_erfc;
+ if (factor_coul<(numtyp)1.0)
+ e_coul -= ((numtyp)1.0-factor_coul)*prefactor;
+ }
+ if (vflag>0) {
+ virial[0] += delx*delx*force;
+ virial[1] += dely*dely*force;
+ virial[2] += delz*delz*force;
+ virial[3] += delx*dely*force;
+ virial[4] += delx*delz*force;
+ virial[5] += dely*delz*force;
+ }
+ }
+
+ } // for nbor
+ store_answers_lq(f,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
+ vflag,ans,engv);
+ } // if ii
+}
+
+__kernel void k_coul_long_cs_fast(const __global numtyp4 *restrict x_,
+ const __global numtyp *restrict scale_in,
+ const __global numtyp *restrict sp_cl_in,
+ const __global int *dev_nbor,
+ const __global int *dev_packed,
+ __global acctyp4 *restrict ans,
+ __global acctyp *restrict engv,
+ const int eflag, const int vflag, const int inum,
+ const int nbor_pitch,
+ const __global numtyp *restrict q_,
+ const numtyp cut_coulsq, const numtyp qqrd2e,
+ const numtyp g_ewald, const int t_per_atom) {
+ int tid, ii, offset;
+ atom_info(t_per_atom,ii,tid,offset);
+
+ __local numtyp scale[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
+ __local numtyp sp_cl[4];
+ if (tid<4)
+ sp_cl[tid]=sp_cl_in[tid];
+ if (tid<MAX_SHARED_TYPES*MAX_SHARED_TYPES)
+ scale[tid]=scale_in[tid];
+
+ acctyp e_coul=(acctyp)0;
+ acctyp4 f;
+ f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
+ acctyp virial[6];
+ for (int i=0; i<6; i++)
+ virial[i]=(acctyp)0;
+
+ __syncthreads();
+
+ if (ii<inum) {
+ int nbor, nbor_end;
+ int i, numj;
+ __local int n_stride;
+ nbor_info(dev_nbor,dev_packed,nbor_pitch,t_per_atom,ii,offset,i,numj,
+ n_stride,nbor_end,nbor);
+
+ numtyp4 ix; fetch4(ix,i,pos_tex); //x_[i];
+ numtyp qtmp; fetch(qtmp,i,q_tex);
+ int iw=ix.w;
+ int itype=fast_mul((int)MAX_SHARED_TYPES,iw);
+
+ for ( ; nbor<nbor_end; nbor+=n_stride) {
+ int j=dev_packed[nbor];
+
+ numtyp factor_coul;
+ factor_coul = sp_cl[sbmask(j)];
+ j &= NEIGHMASK;
+
+ numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
+ int mtype=itype+jx.w;
+
+ // Compute r12
+ numtyp delx = ix.x-jx.x;
+ numtyp dely = ix.y-jx.y;
+ numtyp delz = ix.z-jx.z;
+ numtyp rsq = delx*delx+dely*dely+delz*delz;
+
+ if (rsq < cut_coulsq) {
+ rsq += EPSILON;
+
+ numtyp force,prefactor,_erfc;
+ numtyp r2inv = ucl_recip(rsq);
+ numtyp r = ucl_rsqrt(r2inv);
+ fetch(prefactor,j,q_tex);
+ prefactor *= qqrd2e * scale[mtype] * qtmp;
+ if (factor_coul<(numtyp)1.0) {
+ numtyp grij = g_ewald * (r+EPS_EWALD);
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= (r+EPS_EWALD);
+ force = prefactor * (_erfc + EWALD_F*grij*expm2 - ((numtyp)1.0-factor_coul));
+ r2inv = ucl_recip(rsq + EPS_EWALD_SQR);
+ force *= r2inv;
+ } else {
+ numtyp grij = g_ewald * r;
+ numtyp expm2 = ucl_exp(-grij*grij);
+ numtyp t = ucl_recip((numtyp)1.0 + EWALD_P*grij);
+ numtyp u = (numtyp)1.0 - t;
+ _erfc = t * ((numtyp)1.0 + u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= r;
+ force = prefactor * (_erfc + EWALD_F*grij*expm2);
+ force *= r2inv;
+ }
+
+ f.x+=delx*force;
+ f.y+=dely*force;
+ f.z+=delz*force;
+
+ if (eflag>0) {
+ e_coul += prefactor*_erfc;
+ if (factor_coul<(numtyp)1.0)
+ e_coul -= ((numtyp)1.0-factor_coul)*prefactor;
+ }
+ if (vflag>0) {
+ virial[0] += delx*delx*force;
+ virial[1] += dely*dely*force;
+ virial[2] += delz*delz*force;
+ virial[3] += delx*dely*force;
+ virial[4] += delx*delz*force;
+ virial[5] += dely*delz*force;
+ }
+ }
+
+ } // for nbor
+ store_answers_lq(f,e_coul,virial,ii,inum,tid,t_per_atom,offset,eflag,
+ vflag,ans,engv);
+ } // if ii
+}
+
diff --git a/lib/gpu/lal_coul_long_cs.h b/lib/gpu/lal_coul_long_cs.h
new file mode 100644
index 0000000000000000000000000000000000000000..7c7d6fbef16df61bb2ec8c4eb4bdaa5bf12b6cf7
--- /dev/null
+++ b/lib/gpu/lal_coul_long_cs.h
@@ -0,0 +1,82 @@
+/***************************************************************************
+ coul_long_cs.h
+ -------------------
+ Trung Nguyen (Northwestern)
+
+ Class for acceleration of the coul/long/cs pair style.
+
+ __________________________________________________________________________
+ This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+ __________________________________________________________________________
+
+ begin : June 2018
+ email : trung.nguyen@northwestern.edu
+ ***************************************************************************/
+
+#ifndef LAL_COUL_LONG_CS_H
+#define LAL_COUL_LONG_CS_H
+
+#include "lal_base_charge.h"
+
+namespace LAMMPS_AL {
+
+template <class numtyp, class acctyp>
+class CoulLongCS : public BaseCharge<numtyp, acctyp> {
+ public:
+ CoulLongCS();
+ ~CoulLongCS();
+
+ /// Clear any previous data and set up for a new LAMMPS run
+ /** \param max_nbors initial number of rows in the neighbor matrix
+ * \param cell_size cutoff + skin
+ * \param gpu_split fraction of particles handled by device
+ *
+ * Returns:
+ * - 0 if successfull
+ * - -1 if fix gpu not found
+ * - -3 if there is an out of memory error
+ * - -4 if the GPU library was not compiled for GPU
+ * - -5 Double precision is not supported on card **/
+ int init(const int ntypes, double **scale,
+ const int nlocal, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size,
+ const double gpu_split, FILE *screen,
+ const double host_cut_coulsq, double *host_special_coul,
+ const double qqrd2e, const double g_ewald);
+
+ /// Send updated coeffs from host to device (to be compatible with fix adapt)
+ void reinit(const int ntypes, double **scale);
+
+ /// Clear all host and device data
+ /** \note This is called at the beginning of the init() routine **/
+ void clear();
+
+ /// Returns memory usage on device per atom
+ int bytes_per_atom(const int max_nbors) const;
+
+ /// Total host memory used by library for pair style
+ double host_memory_usage() const;
+
+ // --------------------------- TYPE DATA --------------------------
+
+ /// scale
+ UCL_D_Vec<numtyp> scale;
+ /// Special Coul values [0-3]
+ UCL_D_Vec<numtyp> sp_cl;
+
+ /// If atom type constants fit in shared memory, use fast kernels
+ bool shared_types;
+
+ /// Number of atom types
+ int _lj_types;
+
+ numtyp _cut_coulsq, _qqrd2e, _g_ewald;
+
+ private:
+ bool _allocated;
+ void loop(const bool _eflag, const bool _vflag);
+};
+
+}
+
+#endif
diff --git a/lib/gpu/lal_coul_long_cs_ext.cpp b/lib/gpu/lal_coul_long_cs_ext.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..25619a49216587950edf2f732b1259bfc6c29f98
--- /dev/null
+++ b/lib/gpu/lal_coul_long_cs_ext.cpp
@@ -0,0 +1,145 @@
+/***************************************************************************
+ coul_long_cs_ext.cpp
+ -------------------
+ Trung Nguyen (Northwestern)
+
+ Functions for LAMMPS access to coul/long/cs acceleration routines.
+
+ __________________________________________________________________________
+ This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
+ __________________________________________________________________________
+
+ begin : June 2018
+ email : trung.nguyen@northwestern.edu
+ ***************************************************************************/
+
+#include <iostream>
+#include <cassert>
+#include <cmath>
+
+#include "lal_coul_long_cs.h"
+
+using namespace std;
+using namespace LAMMPS_AL;
+
+static CoulLongCS<PRECISION,ACC_PRECISION> CLCSMF;
+
+// ---------------------------------------------------------------------------
+// Allocate memory on host and device and copy constants to device
+// ---------------------------------------------------------------------------
+int clcs_gpu_init(const int ntypes, double **host_scale,
+ const int inum, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size, int &gpu_mode,
+ FILE *screen, double host_cut_coulsq, double *host_special_coul,
+ const double qqrd2e, const double g_ewald) {
+ CLCSMF.clear();
+ gpu_mode=CLCSMF.device->gpu_mode();
+ double gpu_split=CLCSMF.device->particle_split();
+ int first_gpu=CLCSMF.device->first_device();
+ int last_gpu=CLCSMF.device->last_device();
+ int world_me=CLCSMF.device->world_me();
+ int gpu_rank=CLCSMF.device->gpu_rank();
+ int procs_per_gpu=CLCSMF.device->procs_per_gpu();
+
+ CLCSMF.device->init_message(screen,"coul/long/cs",first_gpu,last_gpu);
+
+ bool message=false;
+ if (CLCSMF.device->replica_me()==0 && screen)
+ message=true;
+
+ if (message) {
+ fprintf(screen,"Initializing Device and compiling on process 0...");
+ fflush(screen);
+ }
+
+ int init_ok=0;
+ if (world_me==0)
+ init_ok=CLCSMF.init(ntypes, host_scale, inum, nall, 300, maxspecial,
+ cell_size, gpu_split, screen, host_cut_coulsq,
+ host_special_coul, qqrd2e, g_ewald);
+
+ CLCSMF.device->world_barrier();
+ if (message)
+ fprintf(screen,"Done.\n");
+
+ for (int i=0; i<procs_per_gpu; i++) {
+ if (message) {
+ if (last_gpu-first_gpu==0)
+ fprintf(screen,"Initializing Device %d on core %d...",first_gpu,i);
+ else
+ fprintf(screen,"Initializing Devices %d-%d on core %d...",first_gpu,
+ last_gpu,i);
+ fflush(screen);
+ }
+ if (gpu_rank==i && world_me!=0)
+ init_ok=CLCSMF.init(ntypes, host_scale, inum, nall, 300, maxspecial,
+ cell_size, gpu_split, screen, host_cut_coulsq,
+ host_special_coul, qqrd2e, g_ewald);
+
+ CLCSMF.device->gpu_barrier();
+ if (message)
+ fprintf(screen,"Done.\n");
+ }
+ if (message)
+ fprintf(screen,"\n");
+
+ if (init_ok==0)
+ CLCSMF.estimate_gpu_overhead();
+ return init_ok;
+}
+
+// ---------------------------------------------------------------------------
+// Copy updated coeffs from host to device
+// ---------------------------------------------------------------------------
+void clcs_gpu_reinit(const int ntypes, double **host_scale) {
+ int world_me=CLCSMF.device->world_me();
+ int gpu_rank=CLCSMF.device->gpu_rank();
+ int procs_per_gpu=CLCSMF.device->procs_per_gpu();
+
+ if (world_me==0)
+ CLCSMF.reinit(ntypes, host_scale);
+
+ CLCSMF.device->world_barrier();
+
+ for (int i=0; i<procs_per_gpu; i++) {
+ if (gpu_rank==i && world_me!=0)
+ CLCSMF.reinit(ntypes, host_scale);
+
+ CLCSMF.device->gpu_barrier();
+ }
+}
+
+void clcs_gpu_clear() {
+ CLCSMF.clear();
+}
+
+int** clcs_gpu_compute_n(const int ago, const int inum_full,
+ const int nall, double **host_x, int *host_type,
+ double *sublo, double *subhi, tagint *tag, int **nspecial,
+ tagint **special, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ int **ilist, int **jnum, const double cpu_time,
+ bool &success, double *host_q, double *boxlo,
+ double *prd) {
+ return CLCSMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
+ subhi, tag, nspecial, special, eflag, vflag, eatom,
+ vatom, host_start, ilist, jnum, cpu_time, success,
+ host_q, boxlo, prd);
+}
+
+void clcs_gpu_compute(const int ago, const int inum_full, const int nall,
+ double **host_x, int *host_type, int *ilist, int *numj,
+ int **firstneigh, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ const double cpu_time, bool &success, double *host_q,
+ const int nlocal, double *boxlo, double *prd) {
+ CLCSMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,
+ firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success,
+ host_q,nlocal,boxlo,prd);
+}
+
+double clcs_gpu_bytes() {
+ return CLCSMF.host_memory_usage();
+}
+
+
diff --git a/src/GPU/pair_born_coul_long_cs_gpu.cpp b/src/GPU/pair_born_coul_long_cs_gpu.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..13cf5f643abf8bb1b7003a64910ad6dc70fd371d
--- /dev/null
+++ b/src/GPU/pair_born_coul_long_cs_gpu.cpp
@@ -0,0 +1,340 @@
+/* ----------------------------------------------------------------------
+ LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
+ http://lammps.sandia.gov, Sandia National Laboratories
+ Steve Plimpton, sjplimp@sandia.gov
+
+ Copyright (2003) Sandia Corporation. Under the terms of Contract
+ DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
+ certain rights in this software. This software is distributed under
+ the GNU General Public License.
+
+ See the README file in the top-level LAMMPS directory.
+------------------------------------------------------------------------- */
+
+/* ----------------------------------------------------------------------
+ Contributing author: Trung Dac Nguyen (Northwestern)
+------------------------------------------------------------------------- */
+
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include "pair_born_coul_long_cs_gpu.h"
+#include "atom.h"
+#include "atom_vec.h"
+#include "comm.h"
+#include "force.h"
+#include "neighbor.h"
+#include "neigh_list.h"
+#include "integrate.h"
+#include "math_const.h"
+#include "memory.h"
+#include "error.h"
+#include "neigh_request.h"
+#include "universe.h"
+#include "update.h"
+#include "domain.h"
+#include "kspace.h"
+#include "gpu_extra.h"
+
+using namespace LAMMPS_NS;
+using namespace MathConst;
+
+#define EWALD_F 1.12837917
+#define EWALD_P 9.95473818e-1
+#define B0 -0.1335096380159268
+#define B1 -2.57839507e-1
+#define B2 -1.37203639e-1
+#define B3 -8.88822059e-3
+#define B4 -5.80844129e-3
+#define B5 1.14652755e-1
+
+#define EPSILON 1.0e-20
+#define EPS_EWALD 1.0e-6
+#define EPS_EWALD_SQR 1.0e-12
+
+// External functions from cuda library for atom decomposition
+
+int bornclcs_gpu_init(const int ntypes, double **cutsq, double **host_rhoinv,
+ double **host_born1, double **host_born2,
+ double **host_born3, double **host_a,
+ double **host_c, double **host_d,
+ double **sigma, double **offset, double *special_lj,
+ const int inum, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size,
+ int &gpu_mode, FILE *screen, double **host_cut_ljsq,
+ double host_cut_coulsq, double *host_special_coul,
+ const double qqrd2e, const double g_ewald);
+void bornclcs_gpu_clear();
+int** bornclcs_gpu_compute_n(const int ago, const int inum_full, const int nall,
+ double **host_x, int *host_type, double *sublo,
+ double *subhi, tagint *tag, int **nspecial,
+ tagint **special, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ int **ilist, int **jnum, const double cpu_time,
+ bool &success, double *host_q, double *boxlo,
+ double *prd);
+void bornclcs_gpu_compute(const int ago, const int inum_full, const int nall,
+ double **host_x, int *host_type, int *ilist, int *numj,
+ int **firstneigh, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ const double cpu_time, bool &success, double *host_q,
+ const int nlocal, double *boxlo, double *prd);
+double bornclcs_gpu_bytes();
+
+/* ---------------------------------------------------------------------- */
+
+PairBornCoulLongCSGPU::PairBornCoulLongCSGPU(LAMMPS *lmp) :
+ PairBornCoulLongCS(lmp), gpu_mode(GPU_FORCE)
+{
+ respa_enable = 0;
+ reinitflag = 0;
+ cpu_time = 0.0;
+ GPU_EXTRA::gpu_ready(lmp->modify, lmp->error);
+}
+
+/* ----------------------------------------------------------------------
+ free all arrays
+------------------------------------------------------------------------- */
+
+PairBornCoulLongCSGPU::~PairBornCoulLongCSGPU()
+{
+ bornclcs_gpu_clear();
+}
+
+/* ---------------------------------------------------------------------- */
+
+void PairBornCoulLongCSGPU::compute(int eflag, int vflag)
+{
+ if (eflag || vflag) ev_setup(eflag,vflag);
+ else evflag = vflag_fdotr = 0;
+
+ int nall = atom->nlocal + atom->nghost;
+ int inum, host_start;
+
+ bool success = true;
+ int *ilist, *numneigh, **firstneigh;
+ if (gpu_mode != GPU_FORCE) {
+ inum = atom->nlocal;
+ firstneigh = bornclcs_gpu_compute_n(neighbor->ago, inum, nall, atom->x,
+ atom->type, domain->sublo, domain->subhi,
+ atom->tag, atom->nspecial, atom->special,
+ eflag, vflag, eflag_atom, vflag_atom,
+ host_start, &ilist, &numneigh, cpu_time,
+ success, atom->q, domain->boxlo,
+ domain->prd);
+ } else {
+ inum = list->inum;
+ ilist = list->ilist;
+ numneigh = list->numneigh;
+ firstneigh = list->firstneigh;
+ bornclcs_gpu_compute(neighbor->ago, inum, nall, atom->x, atom->type,
+ ilist, numneigh, firstneigh, eflag, vflag, eflag_atom,
+ vflag_atom, host_start, cpu_time, success, atom->q,
+ atom->nlocal, domain->boxlo, domain->prd);
+ }
+ if (!success)
+ error->one(FLERR,"Insufficient memory on accelerator");
+
+ if (host_start<inum) {
+ cpu_time = MPI_Wtime();
+ cpu_compute(host_start, inum, eflag, vflag, ilist, numneigh, firstneigh);
+ cpu_time = MPI_Wtime() - cpu_time;
+ }
+}
+
+/* ----------------------------------------------------------------------
+ init specific to this pair style
+------------------------------------------------------------------------- */
+
+void PairBornCoulLongCSGPU::init_style()
+{
+ if (!atom->q_flag)
+ error->all(FLERR,
+ "Pair style born/coul/long/cs/gpu requires atom attribute q");
+ if (force->newton_pair)
+ error->all(FLERR,
+ "Cannot use newton pair with born/coul/long/cs/gpu pair style");
+
+ // Repeat cutsq calculation because done after call to init_style
+ double maxcut = -1.0;
+ double cut;
+ for (int i = 1; i <= atom->ntypes; i++) {
+ for (int j = i; j <= atom->ntypes; j++) {
+ if (setflag[i][j] != 0 || (setflag[i][i] != 0 && setflag[j][j] != 0)) {
+ cut = init_one(i,j);
+ cut *= cut;
+ if (cut > maxcut)
+ maxcut = cut;
+ cutsq[i][j] = cutsq[j][i] = cut;
+ } else
+ cutsq[i][j] = cutsq[j][i] = 0.0;
+ }
+ }
+ double cell_size = sqrt(maxcut) + neighbor->skin;
+
+ cut_coulsq = cut_coul * cut_coul;
+
+ // insure use of KSpace long-range solver, set g_ewald
+
+ if (force->kspace == NULL)
+ error->all(FLERR,"Pair style requires a KSpace style");
+ g_ewald = force->kspace->g_ewald;
+
+ int maxspecial=0;
+ if (atom->molecular)
+ maxspecial=atom->maxspecial;
+ int success = bornclcs_gpu_init(atom->ntypes+1, cutsq, rhoinv,
+ born1, born2, born3, a, c, d, sigma,
+ offset, force->special_lj, atom->nlocal,
+ atom->nlocal+atom->nghost, 300, maxspecial,
+ cell_size, gpu_mode, screen, cut_ljsq,
+ cut_coulsq, force->special_coul,
+ force->qqrd2e, g_ewald);
+
+ GPU_EXTRA::check_flag(success,error,world);
+
+ if (gpu_mode == GPU_FORCE) {
+ int irequest = neighbor->request(this,instance_me);
+ neighbor->requests[irequest]->half = 0;
+ neighbor->requests[irequest]->full = 1;
+ }
+}
+
+/* ---------------------------------------------------------------------- */
+
+double PairBornCoulLongCSGPU::memory_usage()
+{
+ double bytes = Pair::memory_usage();
+ return bytes + bornclcs_gpu_bytes();
+}
+
+/* ---------------------------------------------------------------------- */
+
+void PairBornCoulLongCSGPU::cpu_compute(int start, int inum, int eflag,
+ int vflag, int *ilist, int *numneigh,
+ int **firstneigh)
+{
+ int i,j,ii,jj,jnum,itable,itype,jtype;
+ double qtmp,xtmp,ytmp,ztmp,delx,dely,delz,evdwl,ecoul,fpair;
+ double fraction,table;
+ double r,rsq,rexp,r2inv,r6inv,forcecoul,forceborn,factor_coul,factor_lj;
+ double grij,expm2,prefactor,t,erfc,u;
+ int *jlist;
+
+ evdwl = ecoul = 0.0;
+
+ double **x = atom->x;
+ double **f = atom->f;
+ double *q = atom->q;
+ int *type = atom->type;
+ double *special_coul = force->special_coul;
+ double *special_lj = force->special_lj;
+ double qqrd2e = force->qqrd2e;
+
+ // loop over neighbors of my atoms
+
+ for (ii = start; ii < inum; ii++) {
+ i = ilist[ii];
+ qtmp = q[i];
+ xtmp = x[i][0];
+ ytmp = x[i][1];
+ ztmp = x[i][2];
+ itype = type[i];
+ jlist = firstneigh[i];
+ jnum = numneigh[i];
+
+ for (jj = 0; jj < jnum; jj++) {
+ j = jlist[jj];
+ factor_lj = special_lj[sbmask(j)];
+ factor_coul = special_coul[sbmask(j)];
+ j &= NEIGHMASK;
+
+ delx = xtmp - x[j][0];
+ dely = ytmp - x[j][1];
+ delz = ztmp - x[j][2];
+ rsq = delx*delx + dely*dely + delz*delz;
+ jtype = type[j];
+
+ if (rsq < cutsq[itype][jtype]) {
+
+ if (rsq < cut_coulsq) {
+ rsq += EPSILON; // Add Epsilon for case: r = 0; Interaction must be removed by special bond;
+ r2inv = 1.0/rsq;
+ if (!ncoultablebits || rsq <= tabinnersq) {
+ r = sqrt(rsq);
+ prefactor = qqrd2e * qtmp*q[j];
+ if (factor_coul < 1.0) {
+ // When bonded parts are being calculated a minimal distance (EPS_EWALD)
+ // has to be added to the prefactor and erfc in order to make the
+ // used approximation functions for the Ewald correction valid
+ grij = g_ewald * (r+EPS_EWALD);
+ expm2 = exp(-grij*grij);
+ t = 1.0 / (1.0 + EWALD_P*grij);
+ u = 1.0 - t;
+ erfc = t * (1.+u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= (r+EPS_EWALD);
+ forcecoul = prefactor * (erfc + EWALD_F*grij*expm2 - (1.0-factor_coul));
+ // Additionally r2inv needs to be accordingly modified since the later
+ // scaling of the overall force shall be consistent
+ r2inv = 1.0/(rsq + EPS_EWALD_SQR);
+ } else {
+ grij = g_ewald * r;
+ expm2 = exp(-grij*grij);
+ t = 1.0 / (1.0 + EWALD_P*grij);
+ u = 1.0 - t;
+ erfc = t * (1.+u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= r;
+ forcecoul = prefactor * (erfc + EWALD_F*grij*expm2);
+ }
+ } else {
+ union_int_float_t rsq_lookup;
+ rsq_lookup.f = rsq;
+ itable = rsq_lookup.i & ncoulmask;
+ itable >>= ncoulshiftbits;
+ fraction = (rsq_lookup.f - rtable[itable]) * drtable[itable];
+ table = ftable[itable] + fraction*dftable[itable];
+ forcecoul = qtmp*q[j] * table;
+ if (factor_coul < 1.0) {
+ table = ctable[itable] + fraction*dctable[itable];
+ prefactor = qtmp*q[j] * table;
+ forcecoul -= (1.0-factor_coul)*prefactor;
+ }
+ }
+
+ forcecoul *= r2inv;
+
+ } else forcecoul = 0;
+
+ r2inv = 1.0/rsq;
+ r = sqrt(rsq);
+ if (rsq < cut_ljsq[itype][jtype]) {
+ r6inv = r2inv*r2inv*r2inv;
+ rexp = exp((sigma[itype][jtype]-r)*rhoinv[itype][jtype]);
+ forceborn = born1[itype][jtype]*r*rexp - born2[itype][jtype]*r6inv
+ + born3[itype][jtype]*r2inv*r6inv;
+ } else forceborn = 0.0;
+
+ fpair = forcecoul + factor_lj*forceborn * r2inv;
+
+ f[i][0] += delx*fpair;
+ f[i][1] += dely*fpair;
+ f[i][2] += delz*fpair;
+
+ if (eflag) {
+ if (rsq < cut_coulsq) {
+ ecoul = prefactor*erfc;
+ if (factor_coul < 1.0) ecoul -= (1.0-factor_coul)*prefactor;
+ } else ecoul = 0.0;
+ if (rsq < cut_ljsq[itype][jtype]) {
+ evdwl = a[itype][jtype]*rexp - c[itype][jtype]*r6inv
+ + d[itype][jtype]*r6inv*r2inv - offset[itype][jtype];
+ evdwl *= factor_lj;
+ } else evdwl = 0.0;
+ }
+
+ if (evflag) ev_tally_full(i,evdwl,ecoul,fpair,delx,dely,delz);
+ }
+ }
+ }
+}
diff --git a/src/GPU/pair_born_coul_long_cs_gpu.h b/src/GPU/pair_born_coul_long_cs_gpu.h
new file mode 100644
index 0000000000000000000000000000000000000000..65c80157f2a6c334e229d4d072cb3acfa02d5d75
--- /dev/null
+++ b/src/GPU/pair_born_coul_long_cs_gpu.h
@@ -0,0 +1,66 @@
+/* -*- c++ -*- ----------------------------------------------------------
+ LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
+ http://lammps.sandia.gov, Sandia National Laboratories
+ Steve Plimpton, sjplimp@sandia.gov
+
+ Copyright (2003) Sandia Corporation. Under the terms of Contract
+ DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
+ certain rights in this software. This software is distributed under
+ the GNU General Public License.
+
+ See the README file in the top-level LAMMPS directory.
+------------------------------------------------------------------------- */
+
+#ifdef PAIR_CLASS
+
+PairStyle(born/coul/long/cs/gpu,PairBornCoulLongCSGPU)
+
+#else
+
+#ifndef LMP_PAIR_BORN_COUL_LONG_CS_GPU_H
+#define LMP_PAIR_BORN_COUL_LONG_CS_GPU_H
+
+#include "pair_born_coul_long_cs.h"
+
+namespace LAMMPS_NS {
+
+class PairBornCoulLongCSGPU : public PairBornCoulLongCS {
+ public:
+ PairBornCoulLongCSGPU(LAMMPS *lmp);
+ ~PairBornCoulLongCSGPU();
+ void cpu_compute(int, int, int, int, int *, int *, int **);
+ void compute(int, int);
+ void init_style();
+ double memory_usage();
+
+ enum { GPU_FORCE, GPU_NEIGH, GPU_HYB_NEIGH };
+
+ private:
+ int gpu_mode;
+ double cpu_time;
+};
+
+}
+#endif
+#endif
+
+/* ERROR/WARNING messages:
+
+E: Insufficient memory on accelerator
+
+There is insufficient memory on one of the devices specified for the gpu
+package
+
+E: Pair style born/coul/long/cs/gpu requires atom attribute q
+
+The atom style defined does not have this attribute.
+
+E: Cannot use newton pair with born/coul/long/cs/gpu pair style
+
+Self-explanatory.
+
+E: Pair style requires a KSpace style
+
+No kspace style is defined.
+
+*/
diff --git a/src/GPU/pair_coul_long_cs_gpu.cpp b/src/GPU/pair_coul_long_cs_gpu.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..cbb9c73ceb0fa0cc371d62781647a427b338a130
--- /dev/null
+++ b/src/GPU/pair_coul_long_cs_gpu.cpp
@@ -0,0 +1,320 @@
+/* ----------------------------------------------------------------------
+ LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
+ http://lammps.sandia.gov, Sandia National Laboratories
+ Steve Plimpton, sjplimp@sandia.gov
+
+ Copyright (2003) Sandia Corporation. Under the terms of Contract
+ DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
+ certain rights in this software. This software is distributed under
+ the GNU General Public License.
+
+ See the README file in the top-level LAMMPS directory.
+------------------------------------------------------------------------- */
+
+/* ----------------------------------------------------------------------
+ Contributing author: Trung Nguyen (Northwestern)
+------------------------------------------------------------------------- */
+
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include "pair_coul_long_cs_gpu.h"
+#include "atom.h"
+#include "atom_vec.h"
+#include "comm.h"
+#include "force.h"
+#include "neighbor.h"
+#include "neigh_list.h"
+#include "integrate.h"
+#include "memory.h"
+#include "error.h"
+#include "neigh_request.h"
+#include "universe.h"
+#include "update.h"
+#include "domain.h"
+#include "kspace.h"
+#include "gpu_extra.h"
+
+using namespace LAMMPS_NS;
+
+#define EWALD_F 1.12837917
+#define EWALD_P 9.95473818e-1
+#define B0 -0.1335096380159268
+#define B1 -2.57839507e-1
+#define B2 -1.37203639e-1
+#define B3 -8.88822059e-3
+#define B4 -5.80844129e-3
+#define B5 1.14652755e-1
+
+#define EPSILON 1.0e-20
+#define EPS_EWALD 1.0e-6
+#define EPS_EWALD_SQR 1.0e-12
+
+// External functions from cuda library for atom decomposition
+
+int clcs_gpu_init(const int ntypes, double **scale,
+ const int nlocal, const int nall, const int max_nbors,
+ const int maxspecial, const double cell_size, int &gpu_mode,
+ FILE *screen, double host_cut_coulsq, double *host_special_coul,
+ const double qqrd2e, const double g_ewald);
+void clcs_gpu_reinit(const int ntypes, double **scale);
+void clcs_gpu_clear();
+int ** clcs_gpu_compute_n(const int ago, const int inum,
+ const int nall, double **host_x, int *host_type,
+ double *sublo, double *subhi, tagint *tag,
+ int **nspecial, tagint **special, const bool eflag,
+ const bool vflag, const bool eatom, const bool vatom,
+ int &host_start, int **ilist, int **jnum,
+ const double cpu_time, bool &success, double *host_q,
+ double *boxlo, double *prd);
+void clcs_gpu_compute(const int ago, const int inum, const int nall,
+ double **host_x, int *host_type, int *ilist, int *numj,
+ int **firstneigh, const bool eflag, const bool vflag,
+ const bool eatom, const bool vatom, int &host_start,
+ const double cpu_time, bool &success, double *host_q,
+ const int nlocal, double *boxlo, double *prd);
+double clcs_gpu_bytes();
+
+/* ---------------------------------------------------------------------- */
+
+PairCoulLongCSGPU::PairCoulLongCSGPU(LAMMPS *lmp) :
+ PairCoulLongCS(lmp), gpu_mode(GPU_FORCE)
+{
+ respa_enable = 0;
+ cpu_time = 0.0;
+ GPU_EXTRA::gpu_ready(lmp->modify, lmp->error);
+}
+
+/* ----------------------------------------------------------------------
+ free all arrays
+------------------------------------------------------------------------- */
+
+PairCoulLongCSGPU::~PairCoulLongCSGPU()
+{
+ clcs_gpu_clear();
+}
+
+/* ---------------------------------------------------------------------- */
+
+void PairCoulLongCSGPU::compute(int eflag, int vflag)
+{
+ if (eflag || vflag) ev_setup(eflag,vflag);
+ else evflag = vflag_fdotr = 0;
+
+ int nall = atom->nlocal + atom->nghost;
+ int inum, host_start;
+
+ bool success = true;
+ int *ilist, *numneigh, **firstneigh;
+ if (gpu_mode != GPU_FORCE) {
+ inum = atom->nlocal;
+ firstneigh = clcs_gpu_compute_n(neighbor->ago, inum, nall, atom->x,
+ atom->type, domain->sublo, domain->subhi,
+ atom->tag, atom->nspecial, atom->special,
+ eflag, vflag, eflag_atom, vflag_atom,
+ host_start, &ilist, &numneigh, cpu_time,
+ success, atom->q, domain->boxlo,
+ domain->prd);
+ } else {
+ inum = list->inum;
+ ilist = list->ilist;
+ numneigh = list->numneigh;
+ firstneigh = list->firstneigh;
+ clcs_gpu_compute(neighbor->ago, inum, nall, atom->x, atom->type,
+ ilist, numneigh, firstneigh, eflag, vflag, eflag_atom,
+ vflag_atom, host_start, cpu_time, success, atom->q,
+ atom->nlocal, domain->boxlo, domain->prd);
+ }
+ if (!success)
+ error->one(FLERR,"Insufficient memory on accelerator");
+
+ if (host_start<inum) {
+ cpu_time = MPI_Wtime();
+ cpu_compute(host_start, inum, eflag, vflag, ilist, numneigh, firstneigh);
+ cpu_time = MPI_Wtime() - cpu_time;
+ }
+}
+
+/* ----------------------------------------------------------------------
+ init specific to this pair style
+------------------------------------------------------------------------- */
+
+void PairCoulLongCSGPU::init_style()
+{
+ cut_respa = NULL;
+
+ if (!atom->q_flag)
+ error->all(FLERR,"Pair style coul/long/cs/gpu requires atom attribute q");
+ if (force->newton_pair)
+ error->all(FLERR,"Cannot use newton pair with coul/long/cs/gpu pair style");
+
+ // Call init_one calculation make sure scale is correct
+ for (int i = 1; i <= atom->ntypes; i++) {
+ for (int j = i; j <= atom->ntypes; j++) {
+ if (setflag[i][j] != 0 || (setflag[i][i] != 0 && setflag[j][j] != 0)) {
+ double cut = init_one(i,j);
+ }
+ }
+ }
+ double cell_size = cut_coul + neighbor->skin;
+
+ cut_coulsq = cut_coul * cut_coul;
+
+ // insure use of KSpace long-range solver, set g_ewald
+
+ if (force->kspace == NULL)
+ error->all(FLERR,"Pair style requires a KSpace style");
+ g_ewald = force->kspace->g_ewald;
+
+ // setup force tables
+
+ if (ncoultablebits) init_tables(cut_coul,cut_respa);
+
+ int maxspecial=0;
+ if (atom->molecular)
+ maxspecial=atom->maxspecial;
+ int success = clcs_gpu_init(atom->ntypes+1, scale,
+ atom->nlocal, atom->nlocal+atom->nghost, 300,
+ maxspecial, cell_size, gpu_mode, screen, cut_coulsq,
+ force->special_coul, force->qqrd2e, g_ewald);
+
+ GPU_EXTRA::check_flag(success,error,world);
+
+ if (gpu_mode == GPU_FORCE) {
+ int irequest = neighbor->request(this,instance_me);
+ neighbor->requests[irequest]->half = 0;
+ neighbor->requests[irequest]->full = 1;
+ }
+}
+
+/* ---------------------------------------------------------------------- */
+
+void PairCoulLongCSGPU::reinit()
+{
+ Pair::reinit();
+
+ clcs_gpu_reinit(atom->ntypes+1, scale);
+}
+
+/* ---------------------------------------------------------------------- */
+
+double PairCoulLongCSGPU::memory_usage()
+{
+ double bytes = Pair::memory_usage();
+ return bytes + clcs_gpu_bytes();
+}
+
+/* ---------------------------------------------------------------------- */
+
+void PairCoulLongCSGPU::cpu_compute(int start, int inum, int eflag,
+ int vflag, int *ilist, int *numneigh,
+ int **firstneigh)
+{
+ int i,j,ii,jj,jnum,itable,itype,jtype;
+ double qtmp,xtmp,ytmp,ztmp,delx,dely,delz,ecoul,fpair;
+ double fraction,table;
+ double r,r2inv,forcecoul,factor_coul;
+ double grij,expm2,prefactor,t,erfc,u;
+ int *jlist;
+ double rsq;
+
+ ecoul = 0.0;
+
+ double **x = atom->x;
+ double **f = atom->f;
+ double *q = atom->q;
+ int *type = atom->type;
+ double *special_coul = force->special_coul;
+ double qqrd2e = force->qqrd2e;
+
+ // loop over neighbors of my atoms
+
+ for (ii = start; ii < inum; ii++) {
+ i = ilist[ii];
+ qtmp = q[i];
+ xtmp = x[i][0];
+ ytmp = x[i][1];
+ ztmp = x[i][2];
+ itype = type[i];
+ jlist = firstneigh[i];
+ jnum = numneigh[i];
+
+ for (jj = 0; jj < jnum; jj++) {
+ j = jlist[jj];
+ factor_coul = special_coul[sbmask(j)];
+ j &= NEIGHMASK;
+
+ delx = xtmp - x[j][0];
+ dely = ytmp - x[j][1];
+ delz = ztmp - x[j][2];
+ rsq = delx*delx + dely*dely + delz*delz;
+ jtype = type[j];
+
+ if (rsq < cut_coulsq) {
+ rsq += EPSILON; // Add Epsilon for case: r = 0; Interaction must be removed by special bond;
+ r2inv = 1.0/rsq;
+ if (!ncoultablebits || rsq <= tabinnersq) {
+ r = sqrt(rsq);
+ prefactor = qqrd2e * scale[itype][jtype] * qtmp*q[j];
+ if (factor_coul < 1.0) {
+ // When bonded parts are being calculated a minimal distance (EPS_EWALD)
+ // has to be added to the prefactor and erfc in order to make the
+ // used approximation functions for the Ewald correction valid
+ grij = g_ewald * (r+EPS_EWALD);
+ expm2 = exp(-grij*grij);
+ t = 1.0 / (1.0 + EWALD_P*grij);
+ u = 1.0 - t;
+ erfc = t * (1.+u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= (r+EPS_EWALD);
+ forcecoul = prefactor * (erfc + EWALD_F*grij*expm2 - (1.0-factor_coul));
+ // Additionally r2inv needs to be accordingly modified since the later
+ // scaling of the overall force shall be consistent
+ r2inv = 1.0/(rsq + EPS_EWALD_SQR);
+ } else {
+ grij = g_ewald * r;
+ expm2 = exp(-grij*grij);
+ t = 1.0 / (1.0 + EWALD_P*grij);
+ u = 1.0 - t;
+ erfc = t * (1.+u*(B0+u*(B1+u*(B2+u*(B3+u*(B4+u*B5)))))) * expm2;
+ prefactor /= r;
+ forcecoul = prefactor * (erfc + EWALD_F*grij*expm2);
+ }
+ } else {
+ union_int_float_t rsq_lookup;
+ rsq_lookup.f = rsq;
+ itable = rsq_lookup.i & ncoulmask;
+ itable >>= ncoulshiftbits;
+ fraction = (rsq_lookup.f - rtable[itable]) * drtable[itable];
+ table = ftable[itable] + fraction*dftable[itable];
+ forcecoul = scale[itype][jtype] * qtmp*q[j] * table;
+ if (factor_coul < 1.0) {
+ table = ctable[itable] + fraction*dctable[itable];
+ prefactor = scale[itype][jtype] * qtmp*q[j] * table;
+ forcecoul -= (1.0-factor_coul)*prefactor;
+ }
+ }
+
+ fpair = forcecoul * r2inv;
+
+ f[i][0] += delx*fpair;
+ f[i][1] += dely*fpair;
+ f[i][2] += delz*fpair;
+
+ if (eflag) {
+ if (rsq < cut_coulsq) {
+ if (!ncoultablebits || rsq <= tabinnersq)
+ ecoul = prefactor*erfc;
+ else {
+ table = etable[itable] + fraction*detable[itable];
+ ecoul = scale[itype][jtype] * qtmp*q[j] * table;
+ }
+ if (factor_coul < 1.0) ecoul -= (1.0-factor_coul)*prefactor;
+ } else ecoul = 0.0;
+ }
+
+ if (evflag) ev_tally_full(i,0.0,ecoul,fpair,delx,dely,delz);
+ }
+ }
+ }
+}
diff --git a/src/GPU/pair_coul_long_cs_gpu.h b/src/GPU/pair_coul_long_cs_gpu.h
new file mode 100644
index 0000000000000000000000000000000000000000..95ed8d018227aa1246d85f3764b22cd38642059a
--- /dev/null
+++ b/src/GPU/pair_coul_long_cs_gpu.h
@@ -0,0 +1,67 @@
+/* -*- c++ -*- ----------------------------------------------------------
+ LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
+ http://lammps.sandia.gov, Sandia National Laboratories
+ Steve Plimpton, sjplimp@sandia.gov
+
+ Copyright (2003) Sandia Corporation. Under the terms of Contract
+ DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
+ certain rights in this software. This software is distributed under
+ the GNU General Public License.
+
+ See the README file in the top-level LAMMPS directory.
+------------------------------------------------------------------------- */
+
+#ifdef PAIR_CLASS
+
+PairStyle(coul/long/cs/gpu,PairCoulLongCSGPU)
+
+#else
+
+#ifndef LMP_PAIR_COUL_LONG_CS_GPU_H
+#define LMP_PAIR_COUL_LONG_CS_GPU_H
+
+#include "pair_coul_long_cs.h"
+
+namespace LAMMPS_NS {
+
+class PairCoulLongCSGPU : public PairCoulLongCS {
+ public:
+ PairCoulLongCSGPU(LAMMPS *lmp);
+ ~PairCoulLongCSGPU();
+ void cpu_compute(int, int, int, int, int *, int *, int **);
+ void compute(int, int);
+ void init_style();
+ void reinit();
+ double memory_usage();
+
+ enum { GPU_FORCE, GPU_NEIGH, GPU_HYB_NEIGH };
+
+ private:
+ int gpu_mode;
+ double cpu_time;
+};
+
+}
+#endif
+#endif
+
+/* ERROR/WARNING messages:
+
+E: Insufficient memory on accelerator
+
+There is insufficient memory on one of the devices specified for the gpu
+package
+
+E: Pair style coul/long/cs/gpu requires atom attribute q
+
+The atom style defined does not have these attributes.
+
+E: Cannot use newton pair with coul/long/cs/gpu pair style
+
+Self-explanatory.
+
+E: Pair style requires a KSpace style
+
+No kspace style is defined.
+
+*/
diff --git a/src/USER-MISC/pair_lj_expand_coul_long.cpp b/src/USER-MISC/pair_lj_expand_coul_long.cpp
index 528d62209bbd8ce31669c1807f6e1bf1fa03f68f..79a3c2a49791346e5e22e5f3f5fbc9ca22294316 100644
--- a/src/USER-MISC/pair_lj_expand_coul_long.cpp
+++ b/src/USER-MISC/pair_lj_expand_coul_long.cpp
@@ -217,7 +217,7 @@ void PairLJExpandCoulLong::compute_inner()
int i,j,ii,jj,inum,jnum,itype,jtype;
double qtmp,xtmp,ytmp,ztmp,delx,dely,delz,fpair;
double rsq,r2inv,r6inv,forcecoul,forcelj,factor_coul,factor_lj;
- double rsw,rsq,rshift,rshiftsq,rshift2inv;
+ double rsw,r,rshift,rshiftsq,rshift2inv;
int *ilist,*jlist,*numneigh,**firstneigh;
double **x = atom->x;
@@ -306,8 +306,8 @@ void PairLJExpandCoulLong::compute_middle()
{
int i,j,ii,jj,inum,jnum,itype,jtype;
double qtmp,xtmp,ytmp,ztmp,delx,dely,delz,fpair;
- double rsq,r2inv,r6inv,forcecoul,forcelj,factor_coul,factor_lj;
- double rsw,rsq,rshift,rshiftsq,rshift2inv;
+ double rsq,r,r2inv,r6inv,forcecoul,forcelj,factor_coul,factor_lj;
+ double rsw,rshift,rshiftsq,rshift2inv;
int *ilist,*jlist,*numneigh,**firstneigh;
double **x = atom->x;