ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'eam-gpu' of https://github.com/ndtrung81/lammps into eam-gpu

Browse Source
This commit is contained in:
Trung Nguyen
2021-05-19 10:59:47 -05:00
parent 8af6e58a7f 5dbf9f754e
commit fe93603d15
3940 changed files with 112883 additions and 47093 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/gpu/geryon/nvd_device.h
View File

@ -24,10 +24,6 @@
#ifndef NVD_DEVICE
#define NVD_DEVICE
// workaround after GPU package Feb2021 update
// todo: make new neighbor code work with CUDA
#define LAL_USE_OLD_NEIGHBOR
#include <string>
#include <vector>
#include <iostream>

19
lib/gpu/geryon/nvd_kernel.h
View File

@ -108,17 +108,14 @@ class UCL_Program {
std::cerr << log << std::endl
<< "----------------------------------------------------------\n\n";
#endif
if (foutput != NULL) {
fprintf(foutput,"\n\n");
fprintf(foutput,
"----------------------------------------------------------\n");
fprintf(foutput," UCL Error: Error compiling PTX Program...\n");
fprintf(foutput,
"----------------------------------------------------------\n");
fprintf(foutput,"%s\n",log);
fprintf(foutput,
"----------------------------------------------------------\n");
fprintf(foutput,"\n\n");
if (foutput != nullptr) {
fprintf(foutput,"\n\n");
fprintf(foutput, "----------------------------------------------------------\n");
fprintf(foutput, " UCL Error: Error compiling PTX Program...\n");
fprintf(foutput, "----------------------------------------------------------\n");
fprintf(foutput, "%s\n",log->c_str());
fprintf(foutput, "----------------------------------------------------------\n");
fprintf(foutput,"\n\n");
}
return UCL_COMPILE_ERROR;
}

15
lib/gpu/geryon/ocl_device.h
View File

@ -29,7 +29,7 @@
#include <iostream>
#ifndef CL_TARGET_OPENCL_VERSION
#define CL_TARGET_OPENCL_VERSION 210
#define CL_TARGET_OPENCL_VERSION 300
#endif
#ifdef __APPLE__
@ -728,6 +728,9 @@ void UCL_Device::print_all(std::ostream &out) {
out << "\nDevice " << i << ": \"" << name(i).c_str() << "\"\n";
out << " Type of device: "
<< device_type_name(i).c_str() << std::endl;
out << " Supported OpenCL Version: "
<< _properties[i].cl_device_version / 100 << "."
<< _properties[i].cl_device_version % 100 << std::endl;
out << " Is a subdevice: ";
if (is_subdevice(i))
out << "Yes\n";
@ -796,6 +799,16 @@ void UCL_Device::print_all(std::ostream &out) {
out << "Yes\n";
else
out << "No\n";
out << " Subgroup support: ";
if (_properties[i].has_subgroup_support)
out << "Yes\n";
else
out << "No\n";
out << " Shuffle support: ";
if (_properties[i].has_shuffle_support)
out << "Yes\n";
else
out << "No\n";
}
}
}

2
lib/gpu/geryon/ocl_macros.h
View File

@ -5,7 +5,7 @@
#include <cassert>
#ifndef CL_TARGET_OPENCL_VERSION
#define CL_TARGET_OPENCL_VERSION 210
#define CL_TARGET_OPENCL_VERSION 300
#endif
#ifdef __APPLE__

2
lib/gpu/lal_base_atomic.cpp
View File

@ -335,7 +335,7 @@ void BaseAtomicT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_charge.cpp
View File

@ -348,7 +348,7 @@ void BaseChargeT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_dipole.cpp
View File

@ -356,7 +356,7 @@ void BaseDipoleT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_dpd.cpp
View File

@ -356,7 +356,7 @@ void BaseDPDT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_ellipsoid.cpp
View File

@ -554,7 +554,7 @@ void BaseEllipsoidT::compile_kernels(UCL_Device &dev,
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_lj_fast.max_subgroup_size(_block_size);
mx_subgroup_sz = std::min(mx_subgroup_sz, k_ellipsoid.max_subgroup_size(_block_size));
mx_subgroup_sz = std::min(mx_subgroup_sz, k_sphere_ellipsoid.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_three.cpp
View File

@ -461,7 +461,7 @@ void BaseThreeT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair.max_subgroup_size(_block_size);
mx_subgroup_sz = std::min(mx_subgroup_sz, k_three_center.max_subgroup_size(_block_size));
mx_subgroup_sz = std::min(mx_subgroup_sz, k_three_end.max_subgroup_size(_block_size));

18
lib/gpu/lal_device.cpp
View File

@ -18,6 +18,7 @@
#include <map>
#include <cmath>
#include <cstdlib>
#include <iostream>
#if (LAL_USE_OMP == 1)
#include <omp.h>
#endif
@ -777,6 +778,7 @@ void DeviceT::output_times(UCL_Timer &time_pair, Answer<numtyp,acctyp> &ans,
#ifdef USE_OPENCL
// Workaround for timing issue on Intel OpenCL
if (times[3] > 80e6) times[3]=0.0;
if (times[5] > 80e6) times[5]=0.0;
#endif
if (replica_me()==0)
@ -1025,6 +1027,22 @@ Device<PRECISION,ACC_PRECISION> global_device;
}
using namespace LAMMPS_AL;
bool lmp_has_gpu_device()
{
UCL_Device gpu;
return (gpu.num_platforms() > 0);
}
std::string lmp_gpu_device_info()
{
std::ostringstream out;
UCL_Device gpu;
if (gpu.num_platforms() > 0)
gpu.print_all(out);
return out.str();
}
int lmp_init_device(MPI_Comm world, MPI_Comm replica, const int ngpu,
const int first_gpu_id, const int gpu_mode,
const double particle_split, const int t_per_atom,

192
lib/gpu/lal_lj_smooth.cpp Normal file
View File

@ -0,0 +1,192 @@
/***************************************************************************
lj_smooth.cpp
-------------------
Gurgen Melikyan (HSE University)
Class for acceleration of the lj/smooth pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : gkmeliyan@edu.hse.ru
***************************************************************************/
#if defined(USE_OPENCL)
#include "lj_smooth_cl.h"
#elif defined(USE_CUDART)
const char *lj_smooth=0;
#else
#include "lj_smooth_cubin.h"
#endif
#include "lal_lj_smooth.h"
#include <cassert>
namespace LAMMPS_AL {
#define LJSMOOTHT LJSMOOTH<numtyp, acctyp>
extern Device<PRECISION,ACC_PRECISION> device;
template <class numtyp, class acctyp>
LJSMOOTHT::LJSMOOTH() : BaseAtomic<numtyp,acctyp>(), _allocated(false) {
}
template <class numtyp, class acctyp>
LJSMOOTHT::~LJSMOOTH() {
clear();
}
template <class numtyp, class acctyp>
int LJSMOOTHT::bytes_per_atom(const int max_nbors) const {
return this->bytes_per_atom_atomic(max_nbors);
}
template <class numtyp, class acctyp>
int LJSMOOTHT::init(const int ntypes,
double **host_cutsq, double **host_lj1,
double **host_lj2, double **host_lj3,
double **host_lj4, 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_ljsw0, double **host_ljsw1, double **host_ljsw2, double **host_ljsw3,
double **host_ljsw4,
double **cut_inner, double **cut_inner_sq) {
const int max_shared_types=this->device->max_shared_types();
int onetype=0;
#ifdef USE_OPENCL
if (maxspecial==0)
for (int i=1; i<ntypes; i++)
for (int j=i; j<ntypes; j++)
if (host_cutsq[i][j]>0) {
if (onetype>0)
onetype=-1;
else if (onetype==0)
onetype=i*max_shared_types+j;
}
if (onetype<0) onetype=0;
#endif
int success;
success=this->init_atomic(nlocal,nall,max_nbors,maxspecial,cell_size,gpu_split,
_screen,lj_smooth,"k_lj_smooth",onetype);
if (success!=0)
return success;
// If atom type constants fit in shared memory use fast kernel
int lj_types=ntypes;
shared_types=false;
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;
lj1.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack4(ntypes,lj_types,lj1,host_write,host_lj1,host_lj2,
host_cutsq, cut_inner_sq);
lj3.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack4(ntypes,lj_types,lj3,host_write,host_lj3,host_lj4,
host_offset);
ljsw.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack4(ntypes,lj_types,ljsw,host_write,host_ljsw1,host_ljsw2,
host_ljsw3,host_ljsw4);
ljsw0.alloc(lj_types*lj_types,*(this->ucl_device),UCL_READ_ONLY);
this->atom->type_pack2(ntypes,lj_types,ljsw0,host_write,host_ljsw0,cut_inner);
UCL_H_Vec<double> dview;
sp_lj.alloc(4,*(this->ucl_device),UCL_READ_ONLY);
dview.view(host_special_lj,4,*(this->ucl_device));
ucl_copy(sp_lj,dview,false);
_allocated=true;
this->_max_bytes=lj1.row_bytes()+lj3.row_bytes()+ljsw.row_bytes()+ljsw0.row_bytes()+sp_lj.row_bytes();
return 0;
}
template <class numtyp, class acctyp>
void LJSMOOTHT::reinit(const int ntypes, double **host_cutsq, double **host_lj1,
double **host_lj2, double **host_lj3,
double **host_lj4, double **host_offset,
double **host_ljsw0, double **host_ljsw1, double **host_ljsw2, double **host_ljsw3,
double **host_ljsw4,
double **cut_inner, double **cut_inner_sq) {
// 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;
this->atom->type_pack4(ntypes,_lj_types,lj1,host_write,host_lj1,host_lj2,
host_cutsq,cut_inner_sq);
this->atom->type_pack4(ntypes,_lj_types,lj3,host_write,host_lj3,host_lj4,
host_offset);
this->atom->type_pack4(ntypes,_lj_types,ljsw,host_write,host_ljsw1,host_ljsw2,
host_ljsw3,host_ljsw4);
this->atom->type_pack2(ntypes,_lj_types,ljsw0,host_write,host_ljsw0,cut_inner);
}
template <class numtyp, class acctyp>
void LJSMOOTHT::clear() {
if (!_allocated)
return;
_allocated=false;
lj1.clear();
lj3.clear();
ljsw.clear();
ljsw0.clear();
sp_lj.clear();
this->clear_atomic();
}
template <class numtyp, class acctyp>
double LJSMOOTHT::host_memory_usage() const {
return this->host_memory_usage_atomic()+sizeof(LJSMOOTH<numtyp,acctyp>);
}
// ---------------------------------------------------------------------------
// Calculate energies, forces, and torques
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int LJSMOOTHT::loop(const int eflag, const int vflag) {
// 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 ainum=this->ans->inum();
int nbor_pitch=this->nbor->nbor_pitch();
this->time_pair.start();
if (shared_types) {
this->k_pair_sel->set_size(GX,BX);
this->k_pair_sel->run(&this->atom->x, &lj1, &lj3, &ljsw, &ljsw0, &sp_lj,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag,
&vflag, &ainum, &nbor_pitch,
&this->_threads_per_atom);
} else {
this->k_pair.set_size(GX,BX);
this->k_pair.run(&this->atom->x, &lj1, &lj3, &ljsw, &ljsw0, &_lj_types, &sp_lj,
&this->nbor->dev_nbor, &this->_nbor_data->begin(),
&this->ans->force, &this->ans->engv, &eflag, &vflag,
&ainum, &nbor_pitch, &this->_threads_per_atom);
}
this->time_pair.stop();
return GX;
}
template class LJSMOOTH<PRECISION,ACC_PRECISION>;
}

259
lib/gpu/lal_lj_smooth.cu Normal file
View File

@ -0,0 +1,259 @@
// **************************************************************************
// lj_smooth.cu
// -------------------
// Gurgen Melikyan (HSE University)
//
// Device code for acceleration of the lj/smooth pair style
//
// __________________________________________________________________________
// This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
// __________________________________________________________________________
//
// begin :
// email : gkmeliyan@edu.hse.ru
// ***************************************************************************
#if defined(NV_KERNEL) || defined(USE_HIP)
#include "lal_aux_fun1.h"
#ifndef _DOUBLE_DOUBLE
_texture( pos_tex,float4);
#else
_texture_2d( pos_tex,int4);
#endif
#else
#define pos_tex x_
#endif
__kernel void k_lj_smooth(const __global numtyp4 *restrict x_,
const __global numtyp4 *restrict lj1,
const __global numtyp4 *restrict lj3,
const __global numtyp4 *restrict ljsw,
const __global numtyp2 *restrict ljsw0,
const int lj_types,
const __global numtyp *restrict sp_lj,
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 int t_per_atom) {
int tid, ii, offset;
atom_info(t_per_atom,ii,tid,offset);
int n_stride;
local_allocate_store_pair();
acctyp4 f;
f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
acctyp energy, virial[6];
if (EVFLAG) {
energy=(acctyp)0;
for (int i=0; i<6; i++) virial[i]=(acctyp)0;
}
if (ii<inum) {
int i, numj, nbor, nbor_end;
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 force, r6inv, factor_lj, forcelj;
numtyp r, t, tsq, fskin;
for ( ; nbor<nbor_end; nbor+=n_stride) {
int j=dev_packed[nbor];
factor_lj = sp_lj[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<lj1[mtype].z) {
numtyp r2inv=ucl_recip(rsq);
if (rsq < lj1[mtype].w) {
r6inv = r2inv*r2inv*r2inv;
forcelj = r6inv*(lj1[mtype].x*r6inv-lj1[mtype].y);
}
else {
r = sqrt(rsq);
t = r - ljsw0[mtype].y;
tsq = t*t;
fskin = ljsw[mtype].x + ljsw[mtype].y*t +
ljsw[mtype].z*tsq + ljsw[mtype].w*tsq*t;
forcelj = fskin*r;
}
force = factor_lj*r2inv*forcelj;
f.x+=delx*force;
f.y+=dely*force;
f.z+=delz*force;
if (EVFLAG && eflag) {
numtyp e;
if (rsq < lj1[mtype].w)
e = r6inv * (lj3[mtype].x*r6inv - lj3[mtype].y) - lj3[mtype].z;
else
e = ljsw0[mtype].x - ljsw[mtype].x*t -
ljsw[mtype].y*tsq/2.0 - ljsw[mtype].z*tsq*t/3.0 -
ljsw[mtype].w*tsq*tsq/4.0 - lj3[mtype].z;
//numtyp e=r6inv*(lj3[mtype].x*r6inv-lj3[mtype].y);
energy+=factor_lj*e;
}
if (EVFLAG && vflag) {
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
} // if ii
store_answers(f,energy,virial,ii,inum,tid,t_per_atom,offset,eflag,vflag,
ans,engv);
}
__kernel void k_lj_smooth_fast(const __global numtyp4 *restrict x_,
const __global numtyp4 *restrict lj1_in,
const __global numtyp4 *restrict lj3_in,
const __global numtyp4 *restrict ljsw,
const __global numtyp2 *restrict ljsw0,
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 int t_per_atom) {
int tid, ii, offset;
atom_info(t_per_atom,ii,tid,offset);
#ifndef ONETYPE
__local numtyp4 lj1[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
__local numtyp4 lj3[MAX_SHARED_TYPES*MAX_SHARED_TYPES];
__local numtyp sp_lj[4];
if (tid<4)
sp_lj[tid]=sp_lj_in[tid];
if (tid<MAX_SHARED_TYPES*MAX_SHARED_TYPES) {
lj1[tid]=lj1_in[tid];
if (EVFLAG && eflag)
lj3[tid]=lj3_in[tid];
}
__syncthreads();
#else
const numtyp lj1x=lj1_in[ONETYPE].x;
const numtyp lj1y=lj1_in[ONETYPE].y;
const numtyp cutsq=lj1_in[ONETYPE].z;
numtyp lj3x, lj3y, lj3z;
if (EVFLAG && eflag) {
lj3x=lj3_in[ONETYPE].x;
lj3y=lj3_in[ONETYPE].y;
lj3z=lj3_in[ONETYPE].z;
}
#endif
int n_stride;
local_allocate_store_pair();
acctyp4 f;
f.x=(acctyp)0; f.y=(acctyp)0; f.z=(acctyp)0;
acctyp energy, virial[6];
if (EVFLAG) {
energy=(acctyp)0;
for (int i=0; i<6; i++) virial[i]=(acctyp)0;
}
if (ii<inum) {
int i, numj, nbor, nbor_end;
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];
#ifndef ONETYPE
int iw=ix.w;
int itype=fast_mul((int)MAX_SHARED_TYPES,iw);
numtyp force, r6inv, factor_lj, forcelj;
numtyp r, t, tsq, fskin;
#endif
NOUNROLL
for ( ; nbor<nbor_end; nbor+=n_stride) {
int j=dev_packed[nbor];
#ifndef ONETYPE
factor_lj = sp_lj[sbmask(j)];
j &= NEIGHMASK;
#endif
numtyp4 jx; fetch4(jx,j,pos_tex); //x_[j];
#ifndef ONETYPE
int mtype=itype+jx.w;
#endif
// 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<lj1[mtype].z) {
numtyp r2inv=ucl_recip(rsq);
if (rsq < lj1[mtype].w) {
r6inv = r2inv*r2inv*r2inv;
forcelj = r6inv*(lj1[mtype].x*r6inv-lj1[mtype].y);
}
else {
r = sqrt(rsq);
t = r - ljsw0[mtype].y; //?
tsq = t*t;
fskin = ljsw[mtype].x + ljsw[mtype].y*t +
ljsw[mtype].z*tsq + ljsw[mtype].w*tsq*t;
forcelj = fskin*r;
}
force = factor_lj*r2inv*forcelj;
f.x+=delx*force;
f.y+=dely*force;
f.z+=delz*force;
if (EVFLAG && eflag) {
numtyp e;
if (rsq < lj1[mtype].w)
e = r6inv * (lj3[mtype].x*r6inv - lj3[mtype].y) - lj3[mtype].z;
else
e = ljsw0[mtype].x - ljsw[mtype].x*t -
ljsw[mtype].y*tsq/2.0 - ljsw[mtype].z*tsq*t/3.0 -
ljsw[mtype].w*tsq*tsq/4.0 - lj3[mtype].z; //???
energy+=factor_lj*e;
}
if (EVFLAG && vflag) {
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
} // if ii
store_answers(f,energy,virial,ii,inum,tid,t_per_atom,offset,eflag,vflag,
ans,engv);
}

91
lib/gpu/lal_lj_smooth.h Normal file
View File

@ -0,0 +1,91 @@
/***************************************************************************
lj_smooth.h
-------------------
Gurgen Melikyan (HSE University)
Class for acceleration of the lj/smooth pair style.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : gkmelikyan@edu.hse.ru
***************************************************************************/
#ifndef LAL_LJ_SMOOTH_H
#define LAL_LJ_SMOOTH_H
#include "lal_base_atomic.h"
namespace LAMMPS_AL {
template <class numtyp, class acctyp>
class LJSMOOTH : public BaseAtomic<numtyp, acctyp> {
public:
LJSMOOTH();
~LJSMOOTH();
/// 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 successful
* - -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_lj1, double **host_lj2, double **host_lj3,
double **host_lj4, 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_ljsw0, double **host_ljsw1, double **host_ljsw2,
double **host_ljsw3, double **host_ljsw4,
double **cut_inner, double **cut_inner_sq);
/// Send updated coeffs from host to device (to be compatible with fix adapt)
void reinit(const int ntypes, double **host_cutsq,
double **host_lj1, double **host_lj2, double **host_lj3,
double **host_lj4, double **host_offset,
double **host_ljsw0, double **host_ljsw1, double **host_ljsw2,
double **host_ljsw3, double **host_ljsw4,
double **cut_inner, double **cut_inner_sq);
/// 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 --------------------------
/// lj1.x = lj1, lj1.y = lj2, lj1.z = cutsq, lj1.w = cut_inner_sq
UCL_D_Vec<numtyp4> lj1;
/// lj3.x = lj3, lj3.y = lj4, lj3.z = offset
UCL_D_Vec<numtyp4> lj3;
/// ljsw.x = ljsw1, ljsw.y = ljsw2, ljsw.z = ljsw3, ljsw.w = ljsw4
UCL_D_Vec<numtyp4> ljsw;
/// ljsw0.x = ljsw0 ljsw0.y = cut_inner
UCL_D_Vec<numtyp2> ljsw0;
/// Special LJ values
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;
private:
bool _allocated;
int loop(const int _eflag, const int _vflag);
};
}
#endif

144
lib/gpu/lal_lj_smooth_ext.cpp Normal file
View File

@ -0,0 +1,144 @@
/***************************************************************************
lj_smooth_ext.cpp
-------------------
Gurgen Melikyan (HSE University)
Functions for LAMMPS access to lj/smooth acceleration routines.
__________________________________________________________________________
This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
__________________________________________________________________________
begin :
email : gkmeliyan@edu.hse.ru
***************************************************************************/
#include <iostream>
#include <cassert>
#include <cmath>
#include "lal_lj_smooth.h"
using namespace std;
using namespace LAMMPS_AL;
static LJSMOOTH<PRECISION,ACC_PRECISION> LJSMTMF;
// ---------------------------------------------------------------------------
// Allocate memory on host and device and copy constants to device
// ---------------------------------------------------------------------------
int ljsmt_gpu_init(const int ntypes, double **cutsq, double **host_lj1,
double **host_lj2, double **host_lj3, double **host_lj4,
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_ljsw0, double **host_ljsw1, double **host_ljsw2, double **host_ljsw3,
double **host_ljsw4, double **cut_inner, double **cut_inner_sq) {
LJSMTMF.clear();
gpu_mode=LJSMTMF.device->gpu_mode();
double gpu_split=LJSMTMF.device->particle_split();
int first_gpu=LJSMTMF.device->first_device();
int last_gpu=LJSMTMF.device->last_device();
int world_me=LJSMTMF.device->world_me();
int gpu_rank=LJSMTMF.device->gpu_rank();
int procs_per_gpu=LJSMTMF.device->procs_per_gpu();
LJSMTMF.device->init_message(screen,"lj/smooth",first_gpu,last_gpu);
bool message=false;
if (LJSMTMF.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=LJSMTMF.init(ntypes, cutsq, host_lj1, host_lj2, host_lj3,
host_lj4, offset, special_lj, inum, nall, max_nbors,
maxspecial, cell_size, gpu_split, screen,
host_ljsw0, host_ljsw1, host_ljsw2, host_ljsw3, host_ljsw4, cut_inner, cut_inner_sq);
LJSMTMF.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=LJSMTMF.init(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4,
offset, special_lj, inum, nall, max_nbors, maxspecial,
cell_size, gpu_split, screen, host_ljsw0, host_ljsw1, host_ljsw2, host_ljsw3,
host_ljsw4, cut_inner, cut_inner_sq);
LJSMTMF.device->gpu_barrier();
if (message)
fprintf(screen,"Done.\n");
}
if (message)
fprintf(screen,"\n");
if (init_ok==0)
LJSMTMF.estimate_gpu_overhead();
return init_ok;
}
// ---------------------------------------------------------------------------
// Copy updated coeffs from host to device
// ---------------------------------------------------------------------------
void ljsmt_gpu_reinit(const int ntypes, double **cutsq, double **host_lj1,
double **host_lj2, double **host_lj3, double **host_lj4,
double **offset, double **host_ljsw0, double **host_ljsw1, double **host_ljsw2, double **host_ljsw3,
double **host_ljsw4, double **cut_inner, double **cut_inner_sq) {
int world_me=LJSMTMF.device->world_me();
int gpu_rank=LJSMTMF.device->gpu_rank();
int procs_per_gpu=LJSMTMF.device->procs_per_gpu();
if (world_me==0)
LJSMTMF.reinit(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4, offset, host_ljsw0, host_ljsw1, host_ljsw2, host_ljsw3, host_ljsw4, cut_inner, cut_inner_sq);
LJSMTMF.device->world_barrier();
for (int i=0; i<procs_per_gpu; i++) {
if (gpu_rank==i && world_me!=0)
LJSMTMF.reinit(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4, offset, host_ljsw0, host_ljsw1, host_ljsw2, host_ljsw3, host_ljsw4, cut_inner, cut_inner_sq);
LJSMTMF.device->gpu_barrier();
}
}
void ljsmt_gpu_clear() {
LJSMTMF.clear();
}
int ** ljsmt_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) {
return LJSMTMF.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);
}
void ljsmt_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) {
LJSMTMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,
firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success);
}
double ljsmt_gpu_bytes() {
return LJSMTMF.host_memory_usage();
}

4
lib/gpu/lal_neighbor.h
View File

@ -26,8 +26,8 @@
#if !defined(USE_OPENCL) && !defined(USE_HIP)
#ifndef LAL_USE_OLD_NEIGHBOR
// Issue with incorrect results with CUDA 11.2
#if (CUDA_VERSION > 11019) && (CUDA_VERSION < 11030)
// Issue with incorrect results with CUDA >= 11.2
#if (CUDA_VERSION > 11019)
#define LAL_USE_OLD_NEIGHBOR
#endif
#endif

6
lib/gpu/lal_neighbor_gpu.cu
View File

@ -34,8 +34,8 @@ _texture_2d( pos_tex,int4);
#endif
#ifdef NV_KERNEL
#if (__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ == 2)
// Issue with incorrect results in CUDA 11.2
#if (__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ >= 2)
// Issue with incorrect results in CUDA >= 11.2
#define LAL_USE_OLD_NEIGHBOR
#endif
#endif
@ -115,7 +115,7 @@ __kernel void kernel_calc_cell_counts(const unsigned *restrict cell_id,
#define tagint int
#endif
#ifdef LAMMPS_BIGBIG
#define tagint long long int
#define tagint long
#endif
#ifdef LAMMPS_SMALLSMALL
#define tagint int