moebiusband/MD-Bench
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Integrate GROMACS GPU implementation into master branch

Browse Source 
Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>
This commit is contained in:
Rafael Ravedutti
2022-11-08 18:33:23 +01:00
parent 493915fe95
commit c70ebce4c1
9 changed files with 480 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
gromacs/includes/atom.h
View File

@@ -11,29 +11,43 @@
#define DELTA 20000
#define CLUSTER_M 4
// Nbnxn layouts (as of GROMACS):
// Simd4xN: M=4, N=VECTOR_WIDTH
// Simd2xNN: M=4, N=(VECTOR_WIDTH/2)
// Cuda: M=8, N=VECTOR_WIDTH
// Simd2xNN (here used for single-precision)
#if VECTOR_WIDTH > CLUSTER_M * 2
# define KERNEL_NAME "Simd2xNN"
# define CLUSTER_N (VECTOR_WIDTH / 2)
# define computeForceLJ computeForceLJ_2xnn
// Simd4xN
#else
# define KERNEL_NAME "Simd4xN"
#ifdef CUDA_TARGET
# undef VECTOR_WIDTH
# define VECTOR_WIDTH 8
# define KERNEL_NAME "CUDA"
# define CLUSTER_M 8
# define CLUSTER_N VECTOR_WIDTH
# define computeForceLJ computeForceLJ_4xn
#endif
#ifdef USE_REFERENCE_VERSION
# undef KERNEL_NAME
# undef computeForceLJ
# define KERNEL_NAME "Reference"
# define computeForceLJ computeForceLJ_ref
# define computeForceLJ computeForceLJ_cuda
# define initialIntegrate cudaInitialIntegrate
# define finalIntegrate cudaFinalIntegrate
# define updatePbc cudaUpdatePbc
#else
# define CLUSTER_M 4
// Simd2xNN (here used for single-precision)
# if VECTOR_WIDTH > CLUSTER_M * 2
# define KERNEL_NAME "Simd2xNN"
# define CLUSTER_N (VECTOR_WIDTH / 2)
# define computeForceLJ computeForceLJ_2xnn
// Simd4xN
# else
# define KERNEL_NAME "Simd4xN"
# define CLUSTER_N VECTOR_WIDTH
# define computeForceLJ computeForceLJ_4xn
# endif
# ifdef USE_REFERENCE_VERSION
# undef KERNEL_NAME
# undef computeForceLJ
# define KERNEL_NAME "Reference"
# define computeForceLJ computeForceLJ_ref
# endif
# define initialIntegrate cpuInitialIntegrate
# define finalIntegrate cpuFinalIntegrate
# define updatePbc cpuUpdatePbc
#endif
#if CLUSTER_M == CLUSTER_N

56
gromacs/includes/integrate.h Normal file
View File

@@ -0,0 +1,56 @@
/*
* Copyright (C) 2022 NHR@FAU, University Erlangen-Nuremberg.
* All rights reserved. This file is part of MD-Bench.
* Use of this source code is governed by a LGPL-3.0
* license that can be found in the LICENSE file.
*/
#include <stdbool.h>
//---
#include <atom.h>
#include <parameter.h>
#include <util.h>
void cpuInitialIntegrate(Parameter *param, Atom *atom) {
DEBUG_MESSAGE("cpuInitialIntegrate start\n");
for(int ci = 0; ci < atom->Nclusters_local; ci++) {
int ci_vec_base = CI_VECTOR_BASE_INDEX(ci);
MD_FLOAT *ci_x = &atom->cl_x[ci_vec_base];
MD_FLOAT *ci_v = &atom->cl_v[ci_vec_base];
MD_FLOAT *ci_f = &atom->cl_f[ci_vec_base];
for(int cii = 0; cii < atom->iclusters[ci].natoms; cii++) {
ci_v[CL_X_OFFSET + cii] += param->dtforce * ci_f[CL_X_OFFSET + cii];
ci_v[CL_Y_OFFSET + cii] += param->dtforce * ci_f[CL_Y_OFFSET + cii];
ci_v[CL_Z_OFFSET + cii] += param->dtforce * ci_f[CL_Z_OFFSET + cii];
ci_x[CL_X_OFFSET + cii] += param->dt * ci_v[CL_X_OFFSET + cii];
ci_x[CL_Y_OFFSET + cii] += param->dt * ci_v[CL_Y_OFFSET + cii];
ci_x[CL_Z_OFFSET + cii] += param->dt * ci_v[CL_Z_OFFSET + cii];
}
}
DEBUG_MESSAGE("cpuInitialIntegrate end\n");
}
void cpuFinalIntegrate(Parameter *param, Atom *atom) {
DEBUG_MESSAGE("cpuFinalIntegrate start\n");
for(int ci = 0; ci < atom->Nclusters_local; ci++) {
int ci_vec_base = CI_VECTOR_BASE_INDEX(ci);
MD_FLOAT *ci_v = &atom->cl_v[ci_vec_base];
MD_FLOAT *ci_f = &atom->cl_f[ci_vec_base];
for(int cii = 0; cii < atom->iclusters[ci].natoms; cii++) {
ci_v[CL_X_OFFSET + cii] += param->dtforce * ci_f[CL_X_OFFSET + cii];
ci_v[CL_Y_OFFSET + cii] += param->dtforce * ci_f[CL_Y_OFFSET + cii];
ci_v[CL_Z_OFFSET + cii] += param->dtforce * ci_f[CL_Z_OFFSET + cii];
}
}
DEBUG_MESSAGE("cpuFinalIntegrate end\n");
}
#ifdef CUDA_TARGET
void cudaInitialIntegrate(Parameter*, Atom*);
void cudaFinalIntegrate(Parameter*, Atom*);
#endif

6
gromacs/includes/pbc.h
View File

@@ -10,7 +10,11 @@
#ifndef __PBC_H_
#define __PBC_H_
extern void initPbc();
extern void updatePbc(Atom*, Parameter*, int);
extern void cpuUpdatePbc(Atom*, Parameter*, int);
extern void updateAtomsPbc(Atom*, Parameter*);
extern void setupPbc(Atom*, Parameter*);
#ifdef CUDA_TARGET
extern void cudaUpdatePbc(Atom*, Parameter*, int);
#endif
#endif