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

git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@6053 f3b2605a-c512-4ea7-a41b-209d697bcdaa

Browse Source
This commit is contained in:
sjplimp
2011-05-02 15:02:52 +00:00
parent 2be078632d
commit 5f799182b3
70 changed files with 4489 additions and 2253 deletions
Download Patch File Download Diff File Expand all files Collapse all files

140
lib/gpu/charge_gpu_memory.cpp
View File

@ -23,23 +23,28 @@ extern PairGPUDevice<PRECISION,ACC_PRECISION> pair_gpu_device;
template <class numtyp, class acctyp>
ChargeGPUMemoryT::ChargeGPUMemory() : _compiled(false), _max_bytes(0) {
device=&pair_gpu_device;
ans=new PairGPUAns<numtyp,acctyp>();
nbor=new PairGPUNbor();
}
template <class numtyp, class acctyp>
ChargeGPUMemoryT::~ChargeGPUMemory() {
delete ans;
delete nbor;
}
template <class numtyp, class acctyp>
int ChargeGPUMemoryT::bytes_per_atom_atomic(const int max_nbors) const {
return device->atom.bytes_per_atom()+device->nbor.bytes_per_atom(max_nbors);
return device->atom.bytes_per_atom()+ans->bytes_per_atom()+
nbor->bytes_per_atom(max_nbors);
}
template <class numtyp, class acctyp>
bool ChargeGPUMemoryT::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 char *pair_program) {
int ChargeGPUMemoryT::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 char *pair_program) {
nbor_time_avail=false;
screen=_screen;
@ -48,24 +53,31 @@ bool ChargeGPUMemoryT::init_atomic(const int nlocal, const int nall,
gpu_nbor=true;
int _gpu_host=0;
int host_nlocal=hd_balancer.first_host_count(nlocal,gpu_nbor,gpu_split);
int host_nlocal=hd_balancer.first_host_count(nlocal,gpu_split,gpu_nbor);
if (host_nlocal>0)
_gpu_host=1;
if (!device->init(true,false,nlocal,host_nlocal,nall,maxspecial,gpu_nbor,
_gpu_host,max_nbors,cell_size,false))
return false;
_threads_per_atom=device->threads_per_charge();
if (_threads_per_atom>1 && gpu_nbor==false) {
nbor->packing(true);
_nbor_data=&(nbor->dev_packed);
} else
_nbor_data=&(nbor->dev_nbor);
int success=device->init(*ans,true,false,nlocal,host_nlocal,nall,nbor,
maxspecial,_gpu_host,max_nbors,cell_size,false);
if (success!=0)
return success;
ucl_device=device->gpu;
atom=&device->atom;
nbor=&device->nbor;
_block_size=BLOCK_1D;
if (static_cast<size_t>(_block_size)>ucl_device->group_size())
_block_size=ucl_device->group_size();
_block_size=device->pair_block_size();
_block_bio_size=device->block_bio_pair();
compile_kernels(*ucl_device,pair_program);
// Initialize host-device load balancer
hd_balancer.init(device,gpu_split);
hd_balancer.init(device,gpu_nbor,gpu_split);
// Initialize timers for the selected GPU
time_pair.init(*ucl_device);
@ -74,9 +86,14 @@ bool ChargeGPUMemoryT::init_atomic(const int nlocal, const int nall,
pos_tex.bind_float(atom->dev_x,4);
q_tex.bind_float(atom->dev_q,1);
_max_an_bytes=atom->gpu_bytes()+nbor->gpu_bytes();
_max_an_bytes=ans->gpu_bytes()+nbor->gpu_bytes();
return true;
return success;
}
template <class numtyp, class acctyp>
void ChargeGPUMemoryT::estimate_gpu_overhead() {
device->estimate_gpu_overhead(1,_gpu_overhead,_driver_overhead);
}
template <class numtyp, class acctyp>
@ -84,7 +101,10 @@ void ChargeGPUMemoryT::clear_atomic() {
// Output any timing information
acc_timers();
double avg_split=hd_balancer.all_avg_split();
device->output_times(time_pair,avg_split,_max_bytes+_max_an_bytes,screen);
_gpu_overhead*=hd_balancer.timestep();
_driver_overhead*=hd_balancer.timestep();
device->output_times(time_pair,*ans,*nbor,avg_split,_max_bytes+_max_an_bytes,
_gpu_overhead,_driver_overhead,_threads_per_atom,screen);
if (_compiled) {
k_pair_fast.clear();
@ -109,7 +129,7 @@ int * ChargeGPUMemoryT::reset_nbors(const int nall, const int inum, int *ilist,
nbor_time_avail=true;
int mn=nbor->max_nbor_loop(inum,numj);
int mn=nbor->max_nbor_loop(inum,numj,ilist);
resize_atom(inum,nall,success);
resize_local(inum,mn,success);
if (!success)
@ -117,7 +137,7 @@ int * ChargeGPUMemoryT::reset_nbors(const int nall, const int inum, int *ilist,
nbor->get_host(inum,ilist,numj,firstneigh,block_size());
double bytes=atom->gpu_bytes()+nbor->gpu_bytes();
double bytes=ans->gpu_bytes()+nbor->gpu_bytes();
if (bytes>_max_an_bytes)
_max_an_bytes=bytes;
@ -131,8 +151,8 @@ template <class numtyp, class acctyp>
inline void ChargeGPUMemoryT::build_nbor_list(const int inum,
const int host_inum,
const int nall, double **host_x,
int *host_type, double *boxlo,
double *boxhi, int *tag,
int *host_type, double *sublo,
double *subhi, int *tag,
int **nspecial, int **special,
bool &success) {
nbor_time_avail=true;
@ -145,10 +165,10 @@ inline void ChargeGPUMemoryT::build_nbor_list(const int inum,
atom->cast_copy_x(host_x,host_type);
int mn;
nbor->build_nbor_list(inum, host_inum, nall, *atom, boxlo, boxhi, tag,
nbor->build_nbor_list(inum, host_inum, nall, *atom, sublo, subhi, tag,
nspecial, special, success, mn);
double bytes=atom->gpu_bytes()+nbor->gpu_bytes();
double bytes=ans->gpu_bytes()+nbor->gpu_bytes();
if (bytes>_max_an_bytes)
_max_an_bytes=bytes;
}
@ -157,24 +177,26 @@ inline void ChargeGPUMemoryT::build_nbor_list(const int inum,
// Copy nbor list from host if necessary and then calculate forces, virials,..
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
void ChargeGPUMemoryT::compute(const int timestep, 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) {
void ChargeGPUMemoryT::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) {
acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
resize_atom(0,nall,success);
zero_timers();
return;
}
int ago=hd_balancer.ago_first(f_ago);
int inum=hd_balancer.balance(timestep,ago,inum_full,cpu_time,
nbor->gpu_nbor());
atom->inum(inum);
int inum=hd_balancer.balance(ago,inum_full,cpu_time);
ans->inum(inum);
host_start=inum;
if (ago==0) {
@ -187,10 +209,14 @@ void ChargeGPUMemoryT::compute(const int timestep, const int f_ago,
atom->cast_q_data(host_q);
hd_balancer.start_timer();
atom->add_x_data(host_x,host_type);
atom->add_other_data();
atom->add_q_data();
device->precompute(f_ago,nlocal,nall,host_x,host_type,success,host_q,
boxlo, prd);
loop(eflag,vflag);
atom->copy_answers(eflag,vflag,eatom,vatom,ilist);
ans->copy_answers(eflag,vflag,eatom,vatom,ilist);
device->add_ans_object(ans);
hd_balancer.stop_timer();
}
@ -198,30 +224,33 @@ void ChargeGPUMemoryT::compute(const int timestep, const int f_ago,
// Reneighbor on GPU if necessary and then compute forces, virials, energies
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int * ChargeGPUMemoryT::compute(const int timestep, const int ago,
const int inum_full, const int nall,
double **host_x, int *host_type, double *boxlo,
double *boxhi, int *tag, int **nspecial,
int **special, const bool eflag,
int** ChargeGPUMemoryT::compute(const int ago, const int inum_full,
const int nall, double **host_x, int *host_type,
double *sublo, double *subhi, int *tag,
int **nspecial, int **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 *host_q, double *boxlo, double *prd) {
acc_timers();
if (inum_full==0) {
host_start=0;
// Make sure textures are correct if realloc by a different hybrid style
resize_atom(0,nall,success);
zero_timers();
return NULL;
}
hd_balancer.balance(cpu_time,nbor->gpu_nbor());
int inum=hd_balancer.get_gpu_count(timestep,ago,inum_full);
atom->inum(inum);
hd_balancer.balance(cpu_time);
int inum=hd_balancer.get_gpu_count(ago,inum_full);
ans->inum(inum);
host_start=inum;
// Build neighbor list on GPU if necessary
if (ago==0) {
build_nbor_list(inum, inum_full-inum, nall, host_x, host_type,
boxlo, boxhi, tag, nspecial, special, success);
sublo, subhi, tag, nspecial, special, success);
if (!success)
return NULL;
atom->cast_q_data(host_q);
@ -232,20 +261,25 @@ int * ChargeGPUMemoryT::compute(const int timestep, const int ago,
hd_balancer.start_timer();
atom->add_x_data(host_x,host_type);
}
atom->add_other_data();
atom->add_q_data();
*ilist=nbor->host_ilist.begin();
*jnum=nbor->host_acc.begin();
device->precompute(ago,inum_full,nall,host_x,host_type,success,host_q,
boxlo, prd);
loop(eflag,vflag);
atom->copy_answers(eflag,vflag,eatom,vatom);
ans->copy_answers(eflag,vflag,eatom,vatom);
device->add_ans_object(ans);
hd_balancer.stop_timer();
return device->nbor.host_nbor.begin();
return nbor->host_jlist.begin()-host_start;
}
template <class numtyp, class acctyp>
double ChargeGPUMemoryT::host_memory_usage_atomic() const {
return device->atom.host_memory_usage()+
device->nbor.host_memory_usage()+4*sizeof(numtyp)+
sizeof(ChargeGPUMemory<numtyp,acctyp>);
return device->atom.host_memory_usage()+nbor->host_memory_usage()+
4*sizeof(numtyp)+sizeof(ChargeGPUMemory<numtyp,acctyp>);
}
template <class numtyp, class acctyp>