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

Updating all code to use full neighbor lists for all host forces

Browse Source 
computations. Changed the load balancing scheme. Added timing estimates
for GPU and driver overhead, cpu idle time.
This commit is contained in:
W. Michael Brown
2011-03-28 13:18:39 -04:00
parent c831733b31
commit 71ff7b8096
28 changed files with 168 additions and 161 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
lib/gpu/atomic_gpu_memory.cpp
View File

@ -68,6 +68,9 @@ bool AtomicGPUMemoryT::init_atomic(const int nlocal, const int nall,
_block_size=ucl_device->group_size();
compile_kernels(*ucl_device,pair_program);
// Initialize host-device load balancer
hd_balancer.init(device,gpu_nbor,gpu_split);
// Initialize timers for the selected GPU
time_pair.init(*ucl_device);
time_pair.zero();
@ -76,9 +79,6 @@ bool AtomicGPUMemoryT::init_atomic(const int nlocal, const int nall,
_max_an_bytes=ans->gpu_bytes()+nbor->gpu_bytes();
// Initialize host-device load balancer
hd_balancer.init(device,gpu_nbor,gpu_split);
return true;
}
@ -204,52 +204,6 @@ void AtomicGPUMemoryT::compute(const int f_ago, const int inum_full,
hd_balancer.stop_timer();
}
// ---------------------------------------------------------------------------
// Reneighbor on GPU if necessary and then compute forces, virials, energies
// ---------------------------------------------------------------------------
template <class numtyp, class acctyp>
int * AtomicGPUMemoryT::compute(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,
const bool vflag, const bool eatom,
const bool vatom, int &host_start,
const double cpu_time, bool &success) {
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);
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);
if (!success)
return NULL;
hd_balancer.start_timer();
} else {
atom->cast_x_data(host_x,host_type);
hd_balancer.start_timer();
atom->add_x_data(host_x,host_type);
}
loop(eflag,vflag);
ans->copy_answers(eflag,vflag,eatom,vatom);
device->add_ans_object(ans);
hd_balancer.stop_timer();
return nbor->host_nbor.begin();
}
// ---------------------------------------------------------------------------
// Reneighbor on GPU if necessary and then compute forces, virials, energies
// ---------------------------------------------------------------------------