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

181
lib/gpu/pair_gpu_device.h
View File

@ -19,11 +19,17 @@
#define PAIR_GPU_DEVICE_H
#include "pair_gpu_atom.h"
#include "pair_gpu_ans.h"
#include "pair_gpu_nbor.h"
#include "pppm_gpu_memory.h"
#include "mpi.h"
#include <sstream>
#include "stdio.h"
#include <string>
#include <queue>
template <class numtyp, class acctyp,
class grdtyp, class grdtyp4> class PPPMGPUMemory;
template <class numtyp, class acctyp>
class PairGPUDevice {
@ -33,10 +39,15 @@ class PairGPUDevice {
/// Initialize the device for use by this process
/** Sets up a per-device MPI communicator for load balancing and initializes
* the device (>=first_gpu and <=last_gpu) that this proc will be using **/
bool init_device(MPI_Comm world, MPI_Comm replica, const int first_gpu,
* the device (>=first_gpu and <=last_gpu) that this proc will be using
* Returns:
* - 0 if successfull
* - -2 if GPU not found
* - -4 if GPU library not compiled for GPU **/
int init_device(MPI_Comm world, MPI_Comm replica, const int first_gpu,
const int last_gpu, const int gpu_mode,
const double particle_split, const int nthreads);
const double particle_split, const int nthreads,
const int t_per_atom);
/// Initialize the device for Atom and Neighbor storage
/** \param rot True if quaternions need to be stored
@ -50,19 +61,67 @@ class PairGPUDevice {
* \param max_nbors Initial number of rows in the neighbor matrix
* \param cell_size cutoff+skin
* \param pre_cut True if cutoff test will be performed in separate kernel
* than the force kernel **/
bool init(const bool charge, const bool rot, const int nlocal,
const int host_nlocal, const int nall, const int maxspecial,
const bool gpu_nbor, const int gpu_host, const int max_nbors,
const double cell_size, const bool pre_cut);
* than the force kernel
* Returns:
* - 0 if successfull
* - -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(PairGPUAns<numtyp,acctyp> &a, const bool charge, const bool rot,
const int nlocal, const int host_nlocal, const int nall,
PairGPUNbor *nbor, const int maxspecial, const int gpu_host,
const int max_nbors, const double cell_size, const bool pre_cut);
/// Initialize the device for Atom storage only
/** \param nlocal Total number of local particles to allocate memory for
* \param nall Total number of local+ghost particles
*
* Returns:
* - 0 if successfull
* - -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(PairGPUAns<numtyp,acctyp> &ans, const int nlocal, const int nall);
/// Output a message for pair_style acceleration with device stats
void init_message(FILE *screen, const char *name,
const int first_gpu, const int last_gpu);
/// Perform charge assignment asynchronously for PPPM
void set_single_precompute(PPPMGPUMemory<numtyp,acctyp,
float,_lgpu_float4> *pppm);
/// Perform charge assignment asynchronously for PPPM
void set_double_precompute(PPPMGPUMemory<numtyp,acctyp,
double,_lgpu_double4> *pppm);
/// Esimate the overhead from GPU calls from multiple procs
/** \param kernel_calls Number of kernel calls/timestep for timing estimated
* overhead
* \param gpu_overhead Estimated gpu overhead per timestep (sec)
* \param driver_overhead Estimated overhead from driver per timestep (s) **/
void estimate_gpu_overhead(const int kernel_calls, double &gpu_overhead,
double &gpu_driver_overhead);
/// Returns true if double precision is supported on card
inline bool double_precision() { return gpu->double_precision(); }
/// Output a message with timing information
void output_times(UCL_Timer &time_pair, const double avg_split,
const double max_bytes, FILE *screen);
void output_times(UCL_Timer &time_pair, PairGPUAns<numtyp,acctyp> &ans,
PairGPUNbor &nbor, const double avg_split,
const double max_bytes, const double gpu_overhead,
const double driver_overhead,
const int threads_per_atom, FILE *screen);
/// Output a message with timing information
void output_kspace_times(UCL_Timer &time_in, UCL_Timer &time_out,
UCL_Timer & time_map, UCL_Timer & time_rho,
UCL_Timer &time_interp,
PairGPUAns<numtyp,acctyp> &ans,
const double max_bytes, const double cpu_time,
const double cpu_idle_time, FILE *screen);
/// Clear all memory on host and device associated with atom and nbor data
void clear();
@ -70,11 +129,37 @@ class PairGPUDevice {
/// Clear all memory on host and device
void clear_device();
/// Add an answer object for putting forces, energies, etc from GPU to LAMMPS
inline void add_ans_object(PairGPUAns<numtyp,acctyp> *ans)
{ ans_queue.push(ans); }
/// Add "answers" (force,energies,etc.) into LAMMPS structures
inline double fix_gpu(double **f, double **tor, double *eatom,
double **vatom, double *virial, double &ecoul) {
atom.data_unavail();
if (ans_queue.empty()==false) {
stop_host_timer();
double evdw=0.0;
while (ans_queue.empty()==false) {
evdw+=ans_queue.front()->get_answers(f,tor,eatom,vatom,virial,ecoul);
ans_queue.pop();
}
return evdw;
}
return 0.0;
}
/// Start timer on host
inline void start_host_timer() { _cpu_full=MPI_Wtime(); }
inline void start_host_timer()
{ _cpu_full=MPI_Wtime(); _host_timer_started=true; }
/// Stop timer on host
inline void stop_host_timer() { _cpu_full=MPI_Wtime()-_cpu_full; }
inline void stop_host_timer() {
if (_host_timer_started) {
_cpu_full=MPI_Wtime()-_cpu_full;
_host_timer_started=false;
}
}
/// Return host time
inline double host_time() { return _cpu_full; }
@ -114,6 +199,42 @@ class PairGPUDevice {
inline double particle_split() const { return _particle_split; }
/// Return the initialization count for the device
inline int init_count() const { return _init_count; }
/// True if device is being timed
inline bool time_device() const { return _time_device; }
/// Return the number of threads accessing memory simulatenously
inline int num_mem_threads() const { return _num_mem_threads; }
/// Return the number of threads per atom for pair styles
inline int threads_per_atom() const { return _threads_per_atom; }
/// Return the number of threads per atom for pair styles using charge
inline int threads_per_charge() const { return _threads_per_charge; }
/// Return the min of the pair block size or the device max block size
inline int pair_block_size() const { return _block_pair; }
/// Return the maximum number of atom types that can be used with shared mem
inline int max_shared_types() const { return _max_shared_types; }
/// Return the maximum order for PPPM splines
inline int pppm_max_spline() const { return _pppm_max_spline; }
/// Return the block size for PPPM kernels
inline int pppm_block() const { return _pppm_block; }
/// Return the block size for neighbor binning
inline int block_cell_2d() const { return _block_cell_2d; }
/// Return the block size for atom mapping for neighbor builds
inline int block_cell_id() const { return _block_cell_id; }
/// Return the block size for neighbor build kernel
inline int block_nbor_build() const { return _block_nbor_build; }
/// Return the block size for "bio" pair styles
inline int block_bio_pair() const { return _block_bio_pair; }
/// Return the maximum number of atom types for shared mem with "bio" styles
inline int max_bio_shared_types() const { return _max_bio_shared_types; }
// -------------------- SHARED DEVICE ROUTINES --------------------
// Perform asynchronous zero of integer array
void zero(UCL_D_Vec<int> &mem, const int numel) {
int num_blocks=static_cast<int>(ceil(static_cast<double>(numel)/
_block_pair));
k_zero.set_size(num_blocks,_block_pair);
k_zero.run(&mem.begin(),&numel);
}
// -------------------------- DEVICE DATA -------------------------
@ -130,11 +251,30 @@ class PairGPUDevice {
// --------------------------- NBOR DATA ----------------------------
/// Neighbor Data
PairGPUNbor nbor;
PairGPUNborShared _nbor_shared;
// ------------------------ LONG RANGE DATA -------------------------
// Long Range Data
int _long_range_precompute;
PPPMGPUMemory<numtyp,acctyp,float,_lgpu_float4> *pppm_single;
PPPMGPUMemory<numtyp,acctyp,double,_lgpu_double4> *pppm_double;
/// Precomputations for long range charge assignment (asynchronously)
inline void precompute(const int ago, const int nlocal, const int nall,
double **host_x, int *host_type, bool &success,
double *charge, double *boxlo, double *prd) {
if (_long_range_precompute==1)
pppm_single->precompute(ago,nlocal,nall,host_x,host_type,success,charge,
boxlo,prd);
else if (_long_range_precompute==2)
pppm_double->precompute(ago,nlocal,nall,host_x,host_type,success,charge,
boxlo,prd);
}
private:
std::queue<PairGPUAns<numtyp,acctyp> *> ans_queue;
int _init_count;
bool _device_init;
bool _device_init, _host_timer_started, _time_device;
MPI_Comm _comm_world, _comm_replica, _comm_gpu;
int _procs_per_gpu, _gpu_rank, _world_me, _world_size, _replica_me,
_replica_size;
@ -142,6 +282,19 @@ class PairGPUDevice {
double _particle_split;
double _cpu_full;
int _num_mem_threads, _warp_size, _threads_per_atom, _threads_per_charge;
int _pppm_max_spline, _pppm_block;
int _block_pair, _max_shared_types;
int _block_cell_2d, _block_cell_id, _block_nbor_build;
int _block_bio_pair, _max_bio_shared_types;
UCL_Program *dev_program;
UCL_Kernel k_zero, k_info;
bool _compiled;
int compile_kernels();
int _data_in_estimate, _data_out_estimate;
template <class t>
inline std::string toa(const t& in) {
std::ostringstream o;