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

Misc Improvements to GPU Package

Browse Source 
- Optimizations for molecular systems
-   Improved kernel performance and greater CPU overlap
- Reduced GPU to CPU communications for discrete devices
- Switch classic Intel makefiles to use LLVM-based compilers
- Prefetch optimizations supported for OpenCL
- Optimized data repack for quaternions
This commit is contained in:
W. Michael Brown
2023-03-05 21:03:12 -08:00
parent 142876a59b
commit 37f22c8627
151 changed files with 1085 additions and 617 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
lib/gpu/lal_atom.h
View File

@ -52,6 +52,12 @@ using namespace ucl_cudadr;
namespace LAMMPS_AL {
struct EllipsoidBonus {
double shape[3];
double quat[4];
int ilocal;
};
template <class numtyp, class acctyp>
class Atom {
public:
@ -306,8 +312,8 @@ class Atom {
if (_x_avail==false) {
double t=MPI_Wtime();
#ifdef GPU_CAST
memcpy(host_x_cast.begin(),host_ptr[0],_nall*3*sizeof(double));
memcpy(host_type_cast.begin(),host_type,_nall*sizeof(int));
memcpy(x_cast.host.begin(),host_ptr[0],_nall*3*sizeof(double));
memcpy(type_cast.host.begin(),host_type,_nall*sizeof(int));
#else
vec3d *host_p=reinterpret_cast<vec3d*>(&(host_ptr[0][0]));
vec4d_t *xp=reinterpret_cast<vec4d_t*>(&(x[0]));
@ -351,6 +357,24 @@ class Atom {
add_x_data(host_ptr,host_type);
}
// Cast mu data to write buffer (stored in quat)
template<class cpytyp>
inline void cast_mu_data(cpytyp *host_ptr) {
if (_quat_avail==false) {
double t=MPI_Wtime();
if (sizeof(numtyp)==sizeof(double))
memcpy(quat.host.begin(),host_ptr,_nall*4*sizeof(numtyp));
else
#if (LAL_USE_OMP == 1) && (LAL_USE_OMP_SIMD == 1)
#pragma omp parallel for simd schedule(static)
#elif (LAL_USE_OMP_SIMD == 1)
#pragma omp simd
#endif
for (int i=0; i<_nall*4; i++) quat[i]=host_ptr[i];
_time_cast+=MPI_Wtime()-t;
}
}
// Cast charges to write buffer
template<class cpytyp>
inline void cast_q_data(cpytyp *host_ptr) {
@ -384,22 +408,24 @@ class Atom {
}
// Cast quaternions to write buffer
template<class cpytyp>
inline void cast_quat_data(cpytyp *host_ptr) {
inline void cast_quat_data(const int *ellipsoid,
const EllipsoidBonus *bonus) {
if (_quat_avail==false) {
double t=MPI_Wtime();
if (_host_view) {
quat.host.view((numtyp*)host_ptr,_nall*4,*dev);
quat.device.view(quat.host);
} else if (sizeof(numtyp)==sizeof(double))
memcpy(quat.host.begin(),host_ptr,_nall*4*sizeof(numtyp));
else
#if (LAL_USE_OMP == 1) && (LAL_USE_OMP_SIMD == 1)
#pragma omp parallel for simd schedule(static)
#elif (LAL_USE_OMP_SIMD == 1)
#pragma omp simd
#endif
for (int i=0; i<_nall*4; i++) quat[i]=host_ptr[i];
#if (LAL_USE_OMP == 1) && (LAL_USE_OMP_SIMD == 1)
#pragma omp parallel for simd schedule(static)
#elif (LAL_USE_OMP_SIMD == 1)
#pragma omp simd
#endif
for (int i=0; i<_nall; i++) {
int qi = ellipsoid[i];
if (qi > -1) {
quat[i*4] = bonus[qi].quat[0];
quat[i*4+1] = bonus[qi].quat[1];
quat[i*4+2] = bonus[qi].quat[2];
quat[i*4+3] = bonus[qi].quat[3];
}
}
_time_cast+=MPI_Wtime()-t;
}
}
@ -419,10 +445,6 @@ class Atom {
inline void cast_v_data(double **host_ptr, const tagint *host_tag) {
if (_v_avail==false) {
double t=MPI_Wtime();
#ifdef GPU_CAST
memcpy(host_v_cast.begin(),host_ptr[0],_nall*3*sizeof(double));
memcpy(host_tag_cast.begin(),host_tag,_nall*sizeof(int));
#else
vec3d *host_p=reinterpret_cast<vec3d*>(&(host_ptr[0][0]));
vec4d_t *vp=reinterpret_cast<vec4d_t*>(&(v[0]));
#if (LAL_USE_OMP == 1)
@ -434,7 +456,6 @@ class Atom {
vp[i].z=host_p[i].z;
vp[i].w=host_tag[i];
}
#endif
_time_cast+=MPI_Wtime()-t;
}
}
@ -444,16 +465,7 @@ class Atom {
inline void add_v_data(double ** /*host_ptr*/, tagint * /*host_tag*/) {
time_vel.start();
if (_v_avail==false) {
#ifdef GPU_CAST
v_cast.update_device(_nall*3,true);
tag_cast.update_device(_nall,true);
int block_size=64;
int GX=static_cast<int>(ceil(static_cast<double>(_nall)/block_size));
k_cast_x.set_size(GX,block_size);
k_cast_x.run(&v, &v_cast, &tag_cast, &_nall);
#else
v.update_device(_nall*4,true);
#endif
_v_avail=true;
}
time_vel.stop();
@ -519,7 +531,7 @@ class Atom {
UCL_Vector<numtyp4,numtyp4> extra;
#ifdef GPU_CAST
UCL_Vector<numtyp,numtyp> x_cast;
UCL_Vector<double,double> x_cast;
UCL_Vector<int,int> type_cast;
#endif