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

128
lib/gpu/lal_base_charge.cpp
View File

@ -27,6 +27,9 @@ BaseChargeT::BaseCharge() : _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 @@ BaseChargeT::~BaseCharge() {
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>
@ -64,21 +71,11 @@ int BaseChargeT::init_atomic(const int nlocal, const int nall,
_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);
int success=device->init(*ans,true,false,nlocal,nall,maxspecial);
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;
@ -88,6 +85,17 @@ int BaseChargeT::init_atomic(const int nlocal, const int nall,
_block_bio_size=device->block_bio_pair();
compile_kernels(*ucl_device,pair_program,k_name);
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);
@ -104,8 +112,8 @@ int BaseChargeT::init_atomic(const int nlocal, const int nall,
}
template <class numtyp, class acctyp>
void BaseChargeT::estimate_gpu_overhead() {
device->estimate_gpu_overhead(1,_gpu_overhead,_driver_overhead);
void BaseChargeT::estimate_gpu_overhead(const int add_kernels) {
device->estimate_gpu_overhead(1+add_kernels,_gpu_overhead,_driver_overhead);
}
template <class numtyp, class acctyp>
@ -166,8 +174,8 @@ inline void BaseChargeT::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)
@ -179,14 +187,28 @@ inline void BaseChargeT::build_nbor_list(const int inum, const int host_inum,
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void BaseChargeT::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, double *host_q,
const int nlocal, double *boxlo, double *prd) {
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, double *host_q,
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
@ -215,8 +237,8 @@ void BaseChargeT::compute(const int f_ago, const int inum_full,
device->precompute(f_ago,nlocal,nall,host_x,host_type,success,host_q,
boxlo, prd);
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();
}
@ -226,15 +248,29 @@ void BaseChargeT::compute(const int f_ago, const int inum_full,
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int** BaseChargeT::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,
const bool vatom, int &host_start,
int **ilist, int **jnum,
const double cpu_time, bool &success,
double *host_q, double *boxlo, double *prd) {
const int nall, double **host_x, int *host_type,
double *sublo, double *subhi, tagint *tag,
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,
double *host_q, 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
@ -269,8 +305,8 @@ int** BaseChargeT::compute(const int ago, const int inum_full,
device->precompute(ago,inum_full,nall,host_x,host_type,success,host_q,
boxlo, prd);
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();
@ -292,13 +328,37 @@ void BaseChargeT::compile_kernels(UCL_Device &dev, const void *pair_str,
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";
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");
q_tex.get_texture(*pair_program,"q_tex");
#if defined(LAL_OCL_EV_JIT)
oclstring = device->compile_string()+" -DEVFLAG=0";
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 BaseCharge<PRECISION,ACC_PRECISION>;