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

Feb2021 GPU Package Update - GPU Package Files

Browse Source
This commit is contained in:
Michael Brown
2021-02-15 08:20:50 -08:00
parent 16004e8f45
commit e7e2d2323b
345 changed files with 13424 additions and 7708 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
lib/gpu/lal_base_dpd.cpp
View File

@ -27,6 +27,9 @@ BaseDPDT::BaseDPD() : _compiled(false), _max_bytes(0) {
nbor=new Neighbor();
pair_program=nullptr;
ucl_device=nullptr;
#if defined(LAL_OCL_EV_JIT)
pair_program_noev=nullptr;
#endif
}
template <class numtyp, class acctyp>
@ -36,6 +39,10 @@ BaseDPDT::~BaseDPD() {
k_pair_fast.clear();
k_pair.clear();
if (pair_program) delete pair_program;
#if defined(LAL_OCL_EV_JIT)
k_pair_noev.clear();
if (pair_program_noev) delete pair_program_noev;
#endif
}
template <class numtyp, class acctyp>
@ -47,9 +54,9 @@ int BaseDPDT::bytes_per_atom_atomic(const int max_nbors) const {
template <class numtyp, class acctyp>
int BaseDPDT::init_atomic(const int nlocal, const int nall,
const int max_nbors, const int maxspecial,
const double cell_size,
const double gpu_split, FILE *_screen,
const void *pair_program, const char *k_name) {
const double cell_size, const double gpu_split,
FILE *_screen, const void *pair_program,
const char *k_name, const int onetype) {
screen=_screen;
int gpu_nbor=0;
@ -63,31 +70,30 @@ int BaseDPDT::init_atomic(const int nlocal, const int nall,
if (host_nlocal>0)
_gpu_host=1;
_threads_per_atom=device->threads_per_charge();
if (_threads_per_atom>1 && gpu_nbor==0) {
nbor->packing(true);
_nbor_data=&(nbor->dev_packed);
} else
_nbor_data=&(nbor->dev_nbor);
_threads_per_atom=device->threads_per_atom();
int success=device->init(*ans,false,false,nlocal,nall,maxspecial,true);
if (success!=0)
return success;
success = device->init_nbor(nbor,nlocal,host_nlocal,nall,maxspecial,_gpu_host,
max_nbors,cell_size,false,_threads_per_atom);
if (success!=0)
return success;
if (ucl_device!=device->gpu) _compiled=false;
ucl_device=device->gpu;
atom=&device->atom;
_block_size=device->pair_block_size();
_block_bio_size=device->block_bio_pair();
compile_kernels(*ucl_device,pair_program,k_name);
compile_kernels(*ucl_device,pair_program,k_name,onetype);
if (_threads_per_atom>1 && gpu_nbor==0) {
nbor->packing(true);
_nbor_data=&(nbor->dev_packed);
} else
_nbor_data=&(nbor->dev_nbor);
success = device->init_nbor(nbor,nlocal,host_nlocal,nall,maxspecial,_gpu_host,
max_nbors,cell_size,false,_threads_per_atom);
if (success!=0)
return success;
// Initialize host-device load balancer
hd_balancer.init(device,gpu_nbor,gpu_split);
@ -167,8 +173,8 @@ inline void BaseDPDT::build_nbor_list(const int inum, const int host_inum,
atom->cast_copy_x(host_x,host_type);
int mn;
nbor->build_nbor_list(host_x, inum, host_inum, nall, *atom, sublo, subhi, tag,
nspecial, special, success, mn);
nbor->build_nbor_list(host_x, inum, host_inum, nall, *atom, sublo, subhi,
tag, nspecial, special, success, mn, ans->error_flag);
double bytes=ans->gpu_bytes()+nbor->gpu_bytes();
if (bytes>_max_an_bytes)
@ -179,16 +185,30 @@ inline void BaseDPDT::build_nbor_list(const int inum, const int host_inum,
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void BaseDPDT::compute(const int f_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, tagint *tag, double **host_v,
const double dtinvsqrt, const int seed, const int timestep,
void BaseDPDT::compute(const int f_ago, const int inum_full, const int nall,
double **host_x, int *host_type, int *ilist, int *numj,
int **firstneigh, const bool eflag_in,
const bool vflag_in, const bool eatom,
const bool vatom, int &host_start,
const double cpu_time, bool &success, tagint *tag,
double **host_v, const double dtinvsqrt,
const int seed, const int timestep,
const int nlocal, double *boxlo, double *prd) {
acc_timers();
int eflag, vflag;
if (eatom) eflag=2;
else if (eflag_in) eflag=1;
else eflag=0;
if (vatom) vflag=2;
else if (vflag_in) vflag=1;
else vflag=0;
#ifdef LAL_NO_BLOCK_REDUCE
if (eflag) eflag=2;
if (vflag) vflag=2;
#endif
set_kernel(eflag,vflag);
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
@ -218,8 +238,8 @@ void BaseDPDT::compute(const int f_ago, const int inum_full,
_seed = seed;
_timestep = timestep;
loop(eflag,vflag);
ans->copy_answers(eflag,vflag,eatom,vatom,ilist);
const int red_blocks=loop(eflag,vflag);
ans->copy_answers(eflag_in,vflag_in,eatom,vatom,ilist,red_blocks);
device->add_ans_object(ans);
hd_balancer.stop_timer();
}
@ -231,8 +251,8 @@ template <class numtyp, class acctyp>
int** BaseDPDT::compute(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,
int **nspecial, tagint **special, const bool eflag_in,
const bool vflag_in, const bool eatom,
const bool vatom, int &host_start,
int **ilist, int **jnum,
const double cpu_time, bool &success,
@ -240,6 +260,20 @@ int** BaseDPDT::compute(const int ago, const int inum_full,
const int seed, const int timestep,
double *boxlo, double *prd) {
acc_timers();
int eflag, vflag;
if (eatom) eflag=2;
else if (eflag_in) eflag=1;
else eflag=0;
if (vatom) vflag=2;
else if (vflag_in) vflag=1;
else vflag=0;
#ifdef LAL_NO_BLOCK_REDUCE
if (eflag) eflag=2;
if (vflag) vflag=2;
#endif
set_kernel(eflag,vflag);
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
@ -275,8 +309,8 @@ int** BaseDPDT::compute(const int ago, const int inum_full,
_seed = seed;
_timestep = timestep;
loop(eflag,vflag);
ans->copy_answers(eflag,vflag,eatom,vatom);
const int red_blocks=loop(eflag,vflag);
ans->copy_answers(eflag_in,vflag_in,eatom,vatom,red_blocks);
device->add_ans_object(ans);
hd_balancer.stop_timer();
@ -291,20 +325,48 @@ double BaseDPDT::host_memory_usage_atomic() const {
template <class numtyp, class acctyp>
void BaseDPDT::compile_kernels(UCL_Device &dev, const void *pair_str,
const char *kname) {
if (_compiled)
const char *kname, const int onetype) {
if (_compiled && _onetype==onetype)
return;
_onetype=onetype;
std::string s_fast=std::string(kname)+"_fast";
if (pair_program) delete pair_program;
pair_program=new UCL_Program(dev);
pair_program->load_string(pair_str,device->compile_string().c_str());
std::string oclstring = device->compile_string()+" -DEVFLAG=1";
if (_onetype) oclstring+=" -DONETYPE="+device->toa(_onetype);
pair_program->load_string(pair_str,oclstring.c_str(),nullptr,screen);
k_pair_fast.set_function(*pair_program,s_fast.c_str());
k_pair.set_function(*pair_program,kname);
pos_tex.get_texture(*pair_program,"pos_tex");
vel_tex.get_texture(*pair_program,"vel_tex");
#if defined(LAL_OCL_EV_JIT)
oclstring = device->compile_string()+" -DEVFLAG=0";
if (_onetype) oclstring+=" -DONETYPE="+device->toa(_onetype);
if (pair_program_noev) delete pair_program_noev;
pair_program_noev=new UCL_Program(dev);
pair_program_noev->load_string(pair_str,oclstring.c_str(),nullptr,screen);
k_pair_noev.set_function(*pair_program_noev,s_fast.c_str());
#else
k_pair_sel = &k_pair_fast;
#endif
_compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) {
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));
#endif
if (_threads_per_atom > mx_subgroup_sz)
_threads_per_atom = mx_subgroup_sz;
device->set_simd_size(mx_subgroup_sz);
}
#endif
}
template class BaseDPD<PRECISION,ACC_PRECISION>;