Page Menu
Home
c4science
Search
Configure Global Search
Log In
Files
F64208503
intel_preprocess.h
No One
Temporary
Actions
Download File
Edit File
Delete File
View Transforms
Subscribe
Mute Notifications
Award Token
Subscribers
None
File Metadata
Details
File Info
Storage
Attached
Created
Sat, May 25, 08:16
Size
12 KB
Mime Type
text/x-c
Expires
Mon, May 27, 08:16 (2 d)
Engine
blob
Format
Raw Data
Handle
17549198
Attached To
rLAMMPS lammps
intel_preprocess.h
View Options
/* -*- c++ -*- -------------------------------------------------------------
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
http://lammps.sandia.gov, Sandia National Laboratories
Steve Plimpton, sjplimp@sandia.gov
Copyright (2003) Sandia Corporation. Under the terms of Contract
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
certain rights in this software. This software is distributed under
the GNU General Public License.
See the README file in the top-level LAMMPS directory.
------------------------------------------------------------------------- */
/* ----------------------------------------------------------------------
Contributing author: W. Michael Brown (Intel)
------------------------------------------------------------------------- */
#ifdef __INTEL_OFFLOAD
#ifdef LMP_INTEL_OFFLOAD
#define _LMP_INTEL_OFFLOAD
#endif
#endif
#ifndef LMP_INTEL_PREPROCESS_H
#define LMP_INTEL_PREPROCESS_H
#ifndef LAMMPS_MEMALIGN
#error Please set -DLAMMPS_MEMALIGN=64 in CCFLAGS for your LAMMPS makefile.
#else
#if (LAMMPS_MEMALIGN != 64)
#error Please set -DLAMMPS_MEMALIGN=64 in CCFLAGS for your LAMMPS makefile.
#endif
#endif
#if defined(_OPENMP)
#define _use_omp_pragma(txt) _Pragma(txt)
#else
#define _use_omp_pragma(txt)
#endif
#if defined(__INTEL_COMPILER)
#define _use_simd_pragma(txt) _Pragma(txt)
#else
#define _use_simd_pragma(txt)
#endif
namespace LAMMPS_NS {
enum {LMP_OVERFLOW, LMP_LOCAL_MIN, LMP_LOCAL_MAX, LMP_GHOST_MIN,
LMP_GHOST_MAX};
enum {TIME_PACK, TIME_HOST_NEIGHBOR, TIME_HOST_PAIR, TIME_OFFLOAD_NEIGHBOR,
TIME_OFFLOAD_PAIR, TIME_OFFLOAD_WAIT, TIME_OFFLOAD_LATENCY,
TIME_IMBALANCE};
#define NUM_ITIMERS ( TIME_IMBALANCE + 1 )
#define INTEL_DATA_ALIGN 64
#define INTEL_ONEATOM_FACTOR 2
#define INTEL_MIC_VECTOR_WIDTH 16
#define INTEL_MIC_NBOR_PAD INTEL_MIC_VECTOR_WIDTH
#define INTEL_VECTOR_WIDTH 8
#define INTEL_NBOR_PAD INTEL_VECTOR_WIDTH
#define INTEL_LB_MEAN_WEIGHT 0.1
#define INTEL_BIGP 1e15
#define IP_PRE_get_stride(stride, n, datasize, torque) \
{ \
int blength = n; \
if (torque) blength *= 2; \
const int bytes = blength * datasize; \
stride = INTEL_DATA_ALIGN - (bytes % INTEL_DATA_ALIGN); \
stride = blength + stride / datasize; \
}
#if defined(_OPENMP)
#define IP_PRE_omp_range(ifrom, ito, tid, inum, nthreads) \
{ \
const int idelta = 1 + inum/nthreads; \
ifrom = tid * idelta; \
ito = ((ifrom + idelta) > inum) ? inum : ifrom + idelta; \
}
#define IP_PRE_omp_range_id(ifrom, ito, tid, inum, nthreads) \
{ \
tid = omp_get_thread_num(); \
IP_PRE_omp_range(ifrom, ito, tid, inum, nthreads); \
}
#define IP_PRE_omp_range_align(ifrom, ito, tid, inum, nthreads, \
datasize) \
{ \
int chunk_size = INTEL_DATA_ALIGN / datasize; \
int idelta = static_cast<int>(static_cast<float>(inum) \
/chunk_size/nthreads) + 1; \
idelta *= chunk_size; \
ifrom = tid*idelta; \
ito = ifrom + idelta; \
if (ito > inum) ito = inum; \
}
#define IP_PRE_omp_range_id_align(ifrom, ito, tid, inum, \
nthreads, datasize) \
{ \
tid = omp_get_thread_num(); \
IP_PRE_omp_range_align(ifrom, ito, tid, inum, nthreads, \
datasize); \
}
#else
#define IP_PRE_omp_range(ifrom, ito, tid, inum, nthreads) \
{ \
ifrom = 0; \
ito = inum; \
}
#define IP_PRE_omp_range_id(ifrom, ito, tid, inum, nthreads) \
{ \
tid = 0; \
ifrom = 0; \
ito = inum; \
}
#define IP_PRE_omp_range_align(ifrom, ito, tid, inum, nthreads, \
datasize) \
{ \
ifrom = 0; \
ito = inum; \
}
#define IP_PRE_omp_range_id_align(ifrom, ito, tid, inum, \
nthreads, datasize) \
{ \
tid = 0; \
ifrom = 0; \
ito = inum; \
}
#endif
#ifdef _LMP_INTEL_OFFLOAD
#include <sys/time.h>
__declspec( target (mic))
inline double MIC_Wtime() {
double time;
struct timeval tv;
gettimeofday(&tv, NULL);
time = 1.0 * tv.tv_sec + 1.0e-6 * tv.tv_usec;
return time;
}
#define IP_PRE_pack_separate_buffers(fix, buffers, ago, offload, \
nlocal, nall) \
{ \
if (fix->separate_buffers() && ago != 0) { \
fix->start_watch(TIME_PACK); \
if (offload) { \
_use_omp_pragma("omp parallel default(none) shared(buffers,nlocal,nall)") \
{ \
int ifrom, ito, tid; \
int nthreads = comm->nthreads; \
IP_PRE_omp_range_id_align(ifrom, ito, tid, nlocal, \
nthreads, sizeof(flt_t)); \
buffers->thr_pack_cop(ifrom, ito, 0); \
int nghost = nall - nlocal; \
if (nghost) { \
IP_PRE_omp_range_align(ifrom, ito, tid, nall - nlocal, \
nthreads, sizeof(flt_t)); \
buffers->thr_pack_cop(ifrom + nlocal, ito + nlocal, \
fix->offload_min_ghost() - nlocal, \
ago == 1); \
} \
} \
} else { \
buffers->thr_pack_host(fix->host_min_local(), nlocal, 0); \
buffers->thr_pack_host(nlocal, nall, \
fix->host_min_ghost()-nlocal); \
} \
fix->stop_watch(TIME_PACK); \
} \
}
#define IP_PRE_get_transfern(ago, newton, evflag, eflag, vflag, \
buffers, offload, fix, separate_flag, \
x_size, q_size, ev_size, f_stride) \
{ \
separate_flag = 0; \
if (ago == 0) { \
x_size = 0; \
q_size = nall; \
if (offload) { \
if (fix->separate_buffers()) { \
if (lmp->atom->torque) \
separate_flag = 2; \
else \
separate_flag = 1; \
} else \
separate_flag = 3; \
} \
} else { \
x_size = nall; \
q_size = 0; \
} \
ev_size = 0; \
if (evflag) { \
if (eflag) ev_size = 2; \
if (vflag) ev_size = 8; \
} \
int f_length; \
if (newton) \
f_length = nall; \
else \
f_length = nlocal; \
f_length -= minlocal; \
f_stride = buffers->get_stride(f_length); \
}
#define IP_PRE_get_buffers(offload, buffers, fix, tc, f_start, \
ev_global) \
{ \
if (offload) { \
tc = buffers->get_off_threads(); \
f_start = buffers->get_off_f(); \
ev_global = buffers->get_ev_global(); \
} else { \
tc = comm->nthreads; \
f_start = buffers->get_f(); \
fix->start_watch(TIME_HOST_PAIR); \
ev_global = buffers->get_ev_global_host(); \
} \
}
#define IP_PRE_repack_for_offload(newton, separate_flag, nlocal, nall, \
f_stride, x, q) \
{ \
if (separate_flag) { \
if (separate_flag < 3) { \
int all_local = nlocal; \
int ghost_min = overflow[LMP_GHOST_MIN]; \
nlocal = overflow[LMP_LOCAL_MAX] + 1; \
int nghost = overflow[LMP_GHOST_MAX] + 1 - ghost_min; \
if (nghost < 0) nghost = 0; \
nall = nlocal + nghost; \
separate_flag--; \
int flength; \
if (NEWTON_PAIR) flength = nall; \
else flength = nlocal; \
IP_PRE_get_stride(f_stride, flength, sizeof(FORCE_T), \
separate_flag); \
if (nghost) { \
if (nlocal < all_local || ghost_min > all_local) { \
memmove(x + nlocal, x + ghost_min, \
(nall - nlocal) * sizeof(ATOM_T)); \
if (q != 0) \
memmove((void *)(q + nlocal), (void *)(q + ghost_min), \
(nall - nlocal) * sizeof(flt_t)); \
} \
} \
} \
x[nall].x = INTEL_BIGP; \
x[nall].y = INTEL_BIGP; \
x[nall].z = INTEL_BIGP; \
} \
}
#else
#define MIC_Wtime MPI_Wtime
#define IP_PRE_pack_separate_buffers(fix, buffers, ago, offload, \
nlocal, nall)
#define IP_PRE_get_transfern(ago, newton, evflag, eflag, vflag, \
buffers, offload, fix, separate_flag, \
x_size, q_size, ev_size, f_stride) \
{ \
separate_flag = 0; \
int f_length; \
if (newton) \
f_length = nall; \
else \
f_length = nlocal; \
f_stride = buffers->get_stride(f_length); \
}
#define IP_PRE_get_buffers(offload, buffers, fix, tc, f_start, \
ev_global) \
{ \
tc = comm->nthreads; \
f_start = buffers->get_f(); \
fix->start_watch(TIME_HOST_PAIR); \
ev_global = buffers->get_ev_global_host(); \
}
#define IP_PRE_repack_for_offload(newton, separate_flag, nlocal, nall, \
f_stride, x, q)
#endif
#define IP_PRE_ev_tally_nbor(vflag, ev_pre, fpair, delx, dely, delz) \
{ \
if (vflag == 1) { \
sv0 += ev_pre * delx * delx * fpair; \
sv1 += ev_pre * dely * dely * fpair; \
sv2 += ev_pre * delz * delz * fpair; \
sv3 += ev_pre * delx * dely * fpair; \
sv4 += ev_pre * delx * delz * fpair; \
sv5 += ev_pre * dely * delz * fpair; \
} \
}
#define IP_PRE_ev_tally_atom(evflag, eflag, vflag, f, fwtmp) \
{ \
if (evflag) { \
if (eflag) { \
f[i].w += fwtmp; \
oevdwl += sevdwl; \
} \
if (vflag == 1) { \
ov0 += sv0; \
ov1 += sv1; \
ov2 += sv2; \
ov3 += sv3; \
ov4 += sv4; \
ov5 += sv5; \
} \
} \
}
#define IP_PRE_ev_tally_atomq(evflag, eflag, vflag, f, fwtmp) \
{ \
if (evflag) { \
if (eflag) { \
f[i].w += fwtmp; \
oevdwl += sevdwl; \
oecoul += secoul; \
} \
if (vflag == 1) { \
ov0 += sv0; \
ov1 += sv1; \
ov2 += sv2; \
ov3 += sv3; \
ov4 += sv4; \
ov5 += sv5; \
} \
} \
}
#define IP_PRE_fdotr_acc_force(newton, evflag, eflag, vflag, eatom, \
nall, nlocal, minlocal, nthreads, \
f_start, f_stride, x) \
{ \
int o_range; \
if (newton) \
o_range = nall; \
else \
o_range = nlocal; \
if (offload == 0) o_range -= minlocal; \
IP_PRE_omp_range_align(iifrom, iito, tid, o_range, nthreads, \
sizeof(acc_t)); \
\
int t_off = f_stride; \
if (eflag && eatom) { \
for (int t = 1; t < nthreads; t++) { \
_use_simd_pragma("vector nontemporal") \
_use_simd_pragma("novector") \
for (int n = iifrom; n < iito; n++) { \
f_start[n].x += f_start[n + t_off].x; \
f_start[n].y += f_start[n + t_off].y; \
f_start[n].z += f_start[n + t_off].z; \
f_start[n].w += f_start[n + t_off].w; \
} \
t_off += f_stride; \
} \
} else { \
for (int t = 1; t < nthreads; t++) { \
_use_simd_pragma("vector nontemporal") \
_use_simd_pragma("novector") \
for (int n = iifrom; n < iito; n++) { \
f_start[n].x += f_start[n + t_off].x; \
f_start[n].y += f_start[n + t_off].y; \
f_start[n].z += f_start[n + t_off].z; \
} \
t_off += f_stride; \
} \
} \
\
if (evflag) { \
if (vflag == 2) { \
const ATOM_T * _noalias const xo = x + minlocal; \
_use_simd_pragma("vector nontemporal") \
_use_simd_pragma("novector") \
for (int n = iifrom; n < iito; n++) { \
ov0 += f_start[n].x * xo[n].x; \
ov1 += f_start[n].y * xo[n].y; \
ov2 += f_start[n].z * xo[n].z; \
ov3 += f_start[n].y * xo[n].x; \
ov4 += f_start[n].z * xo[n].x; \
ov5 += f_start[n].z * xo[n].y; \
} \
} \
} \
}
}
#endif
Event Timeline
Log In to Comment