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

Move common modules to common directory

Browse Source 
Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>
This commit is contained in:
Rafael Ravedutti
2022-08-17 17:56:31 +02:00
parent 47db9e86b0
commit 28d3946072
31 changed files with 40 additions and 1102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
gromacs/includes/allocate.h
View File

@@ -1,29 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <stdlib.h>
#ifndef __ALLOCATE_H_
#define __ALLOCATE_H_
extern void* allocate (int alignment, size_t bytesize);
extern void* reallocate (void* ptr, int alignment, size_t newBytesize, size_t oldBytesize);
#endif

32
gromacs/includes/atom.h
View File

@@ -130,15 +130,31 @@ extern void growAtom(Atom*);
extern void growClusters(Atom*);
#ifdef AOS
#define POS_DATA_LAYOUT "AoS"
#define atom_x(i) atom->x[(i) * 3 + 0]
#define atom_y(i) atom->x[(i) * 3 + 1]
#define atom_z(i) atom->x[(i) * 3 + 2]
# define POS_DATA_LAYOUT "AoS"
# define atom_x(i) atom->x[(i) * 3 + 0]
# define atom_y(i) atom->x[(i) * 3 + 1]
# define atom_z(i) atom->x[(i) * 3 + 2]
/*
# define atom_vx(i) atom->vx[(i) * 3 + 0]
# define atom_vy(i) atom->vx[(i) * 3 + 1]
# define atom_vz(i) atom->vx[(i) * 3 + 2]
# define atom_fx(i) atom->fx[(i) * 3 + 0]
# define atom_fy(i) atom->fx[(i) * 3 + 1]
# define atom_fz(i) atom->fx[(i) * 3 + 2]
*/
#else
#define POS_DATA_LAYOUT "SoA"
#define atom_x(i) atom->x[i]
#define atom_y(i) atom->y[i]
#define atom_z(i) atom->z[i]
# define POS_DATA_LAYOUT "SoA"
# define atom_x(i) atom->x[i]
# define atom_y(i) atom->y[i]
# define atom_z(i) atom->z[i]
#endif
// TODO: allow to switch velocites and forces to AoS
# define atom_vx(i) atom->vx[i]
# define atom_vy(i) atom->vy[i]
# define atom_vz(i) atom->vz[i]
# define atom_fx(i) atom->fx[i]
# define atom_fy(i) atom->fy[i]
# define atom_fz(i) atom->fz[i]
#endif

55
gromacs/includes/eam.h
View File

@@ -1,55 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <stdio.h>
#include <atom.h>
#include <parameter.h>
#ifndef __EAM_H_
#define __EAM_H_
typedef struct {
int nrho, nr;
MD_FLOAT drho, dr, cut, mass;
MD_FLOAT *frho, *rhor, *zr;
} Funcfl;
typedef struct {
MD_FLOAT* fp;
int nmax;
int nrho, nr;
int nrho_tot, nr_tot;
MD_FLOAT dr, rdr, drho, rdrho;
MD_FLOAT *frho, *rhor, *z2r;
MD_FLOAT *rhor_spline, *frho_spline, *z2r_spline;
Funcfl file;
} Eam;
void initEam(Eam* eam, Parameter* param);
void coeff(Eam* eam, Parameter* param);
void init_style(Eam* eam, Parameter *param);
void read_eam_file(Funcfl* file, const char* filename);
void file2array(Eam* eam);
void array2spline(Eam* eam, Parameter* param);
void interpolate(int n, MD_FLOAT delta, MD_FLOAT* f, MD_FLOAT* spline);
void grab(FILE* fptr, int n, MD_FLOAT* list);
#endif

170
gromacs/includes/likwid-marker.h
View File

@@ -1,170 +0,0 @@
/*
* =======================================================================================
*
* Filename: likwid-marker.h
*
* Description: Header File of likwid Marker API
*
* Version: <VERSION>
* Released: <DATE>
*
* Authors: Thomas Gruber (tg), thomas.roehl@googlemail.com
*
* Project: likwid
*
* Copyright (C) 2016 RRZE, University Erlangen-Nuremberg
*
* This program is free software: you can redistribute it and/or modify it under
* the terms of the GNU General Public License as published by the Free Software
* Foundation, either version 3 of the License, or (at your option) any later
* version.
*
* This program is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License along with
* this program. If not, see <http://www.gnu.org/licenses/>.
*
* =======================================================================================
*/
#ifndef LIKWID_MARKER_H
#define LIKWID_MARKER_H
/** \addtogroup MarkerAPI Marker API module
* @{
*/
/*!
\def LIKWID_MARKER_INIT
Shortcut for likwid_markerInit() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_THREADINIT
Shortcut for likwid_markerThreadInit() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_REGISTER(regionTag)
Shortcut for likwid_markerRegisterRegion() with \a regionTag if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_START(regionTag)
Shortcut for likwid_markerStartRegion() with \a regionTag if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_STOP(regionTag)
Shortcut for likwid_markerStopRegion() with \a regionTag if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_GET(regionTag, nevents, events, time, count)
Shortcut for likwid_markerGetResults() for \a regionTag if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_SWITCH
Shortcut for likwid_markerNextGroup() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_RESET(regionTag)
Shortcut for likwid_markerResetRegion() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_MARKER_CLOSE
Shortcut for likwid_markerClose() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/** @}*/
#ifdef LIKWID_PERFMON
#include <likwid.h>
#define LIKWID_MARKER_INIT likwid_markerInit()
#define LIKWID_MARKER_THREADINIT likwid_markerThreadInit()
#define LIKWID_MARKER_SWITCH likwid_markerNextGroup()
#define LIKWID_MARKER_REGISTER(regionTag) likwid_markerRegisterRegion(regionTag)
#define LIKWID_MARKER_START(regionTag) likwid_markerStartRegion(regionTag)
#define LIKWID_MARKER_STOP(regionTag) likwid_markerStopRegion(regionTag)
#define LIKWID_MARKER_CLOSE likwid_markerClose()
#define LIKWID_MARKER_RESET(regionTag) likwid_markerResetRegion(regionTag)
#define LIKWID_MARKER_GET(regionTag, nevents, events, time, count) likwid_markerGetRegion(regionTag, nevents, events, time, count)
#else /* LIKWID_PERFMON */
#define LIKWID_MARKER_INIT
#define LIKWID_MARKER_THREADINIT
#define LIKWID_MARKER_SWITCH
#define LIKWID_MARKER_REGISTER(regionTag)
#define LIKWID_MARKER_START(regionTag)
#define LIKWID_MARKER_STOP(regionTag)
#define LIKWID_MARKER_CLOSE
#define LIKWID_MARKER_GET(regionTag, nevents, events, time, count)
#define LIKWID_MARKER_RESET(regionTag)
#endif /* LIKWID_PERFMON */
/** \addtogroup NvMarkerAPI NvMarker API module (MarkerAPI for Nvidia GPUs)
* @{
*/
/*!
\def LIKWID_NVMARKER_INIT
Shortcut for likwid_gpuMarkerInit() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_THREADINIT
Shortcut for likwid_gpuMarkerThreadInit() if compiled with -DLIKWID_PERFMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_REGISTER(regionTag)
Shortcut for likwid_gpuMarkerRegisterRegion() with \a regionTag if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_START(regionTag)
Shortcut for likwid_gpuMarkerStartRegion() with \a regionTag if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_STOP(regionTag)
Shortcut for likwid_gpuMarkerStopRegion() with \a regionTag if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_GET(regionTag, ngpus, nevents, events, time, count)
Shortcut for likwid_gpuMarkerGetRegion() for \a regionTag if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_SWITCH
Shortcut for likwid_gpuMarkerNextGroup() if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_RESET(regionTag)
Shortcut for likwid_gpuMarkerResetRegion() if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/*!
\def LIKWID_NVMARKER_CLOSE
Shortcut for likwid_gpuMarkerClose() if compiled with -DLIKWID_NVMON. Otherwise no operation is performed
*/
/** @}*/
#ifdef LIKWID_NVMON
#ifndef LIKWID_WITH_NVMON
#define LIKWID_WITH_NVMON
#endif
#include <likwid.h>
#define LIKWID_NVMARKER_INIT likwid_gpuMarkerInit()
#define LIKWID_NVMARKER_THREADINIT likwid_gpuMarkerThreadInit()
#define LIKWID_NVMARKER_SWITCH likwid_gpuMarkerNextGroup()
#define LIKWID_NVMARKER_REGISTER(regionTag) likwid_gpuMarkerRegisterRegion(regionTag)
#define LIKWID_NVMARKER_START(regionTag) likwid_gpuMarkerStartRegion(regionTag)
#define LIKWID_NVMARKER_STOP(regionTag) likwid_gpuMarkerStopRegion(regionTag)
#define LIKWID_NVMARKER_CLOSE likwid_gpuMarkerClose()
#define LIKWID_NVMARKER_RESET(regionTag) likwid_gpuMarkerResetRegion(regionTag)
#define LIKWID_NVMARKER_GET(regionTag, ngpus, nevents, events, time, count) \
likwid_gpuMarkerGetRegion(regionTag, ngpus, nevents, events, time, count)
#else /* LIKWID_NVMON */
#define LIKWID_NVMARKER_INIT
#define LIKWID_NVMARKER_THREADINIT
#define LIKWID_NVMARKER_SWITCH
#define LIKWID_NVMARKER_REGISTER(regionTag)
#define LIKWID_NVMARKER_START(regionTag)
#define LIKWID_NVMARKER_STOP(regionTag)
#define LIKWID_NVMARKER_CLOSE
#define LIKWID_NVMARKER_GET(regionTag, nevents, events, time, count)
#define LIKWID_NVMARKER_RESET(regionTag)
#endif /* LIKWID_NVMON */
#endif /* LIKWID_MARKER_H */

69
gromacs/includes/parameter.h
View File

@@ -1,69 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#ifndef __PARAMETER_H_
#define __PARAMETER_H_
#if PRECISION == 1
#define MD_FLOAT float
#else
#define MD_FLOAT double
#endif
typedef struct {
int force_field;
char* param_file;
char* input_file;
char* vtk_file;
char* xtc_file;
MD_FLOAT epsilon;
MD_FLOAT sigma;
MD_FLOAT sigma6;
MD_FLOAT temp;
MD_FLOAT rho;
MD_FLOAT mass;
int ntypes;
int ntimes;
int nstat;
int reneigh_every;
int prune_every;
int x_out_every;
int v_out_every;
int half_neigh;
int nx, ny, nz;
MD_FLOAT dt;
MD_FLOAT dtforce;
MD_FLOAT cutforce;
MD_FLOAT skin;
MD_FLOAT cutneigh;
MD_FLOAT lattice;
MD_FLOAT xlo, xhi, ylo, yhi, zlo, zhi;
MD_FLOAT xprd, yprd, zprd;
double proc_freq;
char* eam_file;
} Parameter;
void initParameter(Parameter*);
void readParameter(Parameter*, const char*);
void printParameter(Parameter*);
#endif

69
gromacs/includes/simd.h
View File

@@ -1,69 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <immintrin.h>
#ifndef NO_ZMM_INTRIN
# include <zmmintrin.h>
#endif
#ifndef CLUSTER_M
# define CLUSTER_M 1
#endif
#ifndef CLUSTER_N
# define CLUSTER_N 1
#endif
#ifdef AVX512
# if PRECISION == 2
# include "simd/avx512_double.h"
# else
# include "simd/avx512_float.h"
# endif
#else
# if PRECISION == 2
# include "simd/avx_avx2_double.h"
# else
# include "simd/avx_avx2_float.h"
# endif
#endif
#define SIMD_PRINT_REAL(a) simd_print_real(#a, a);
#define SIMD_PRINT_MASK(a) simd_print_mask(#a, a);
static inline void simd_print_real(const char *ref, MD_SIMD_FLOAT a) {
double x[VECTOR_WIDTH];
memcpy(x, &a, sizeof(x));
fprintf(stdout, "%s: ", ref);
for(int i = 0; i < VECTOR_WIDTH; i++) {
fprintf(stdout, "%f ", x[i]);
}
fprintf(stdout, "\n");
}
static inline void simd_print_mask(const char *ref, MD_SIMD_MASK a) { fprintf(stdout, "%s: %x\n", ref, simd_mask_to_u32(a)); }

126
gromacs/includes/simd/avx512_double.h
View File

@@ -1,126 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <immintrin.h>
#ifndef NO_ZMM_INTRIN
# include <zmmintrin.h>
#endif
#define MD_SIMD_FLOAT __m512d
#define MD_SIMD_MASK __mmask8
#define MD_SIMD_INT __m256i
static inline MD_SIMD_FLOAT simd_broadcast(MD_FLOAT scalar) { return _mm512_set1_pd(scalar); }
static inline MD_SIMD_FLOAT simd_zero() { return _mm512_set1_pd(0.0); }
static inline MD_SIMD_FLOAT simd_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_add_pd(a, b); }
static inline MD_SIMD_FLOAT simd_sub(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_sub_pd(a, b); }
static inline MD_SIMD_FLOAT simd_mul(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_mul_pd(a, b); }
static inline MD_SIMD_FLOAT simd_fma(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_FLOAT c) { return _mm512_fmadd_pd(a, b, c); }
static inline MD_SIMD_FLOAT simd_reciprocal(MD_SIMD_FLOAT a) { return _mm512_rcp14_pd(a); }
static inline MD_SIMD_FLOAT simd_masked_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_MASK m) { return _mm512_mask_add_pd(a, m, a, b); }
static inline MD_SIMD_MASK simd_mask_and(MD_SIMD_MASK a, MD_SIMD_MASK b) { return _kand_mask8(a, b); }
static inline MD_SIMD_MASK simd_mask_cond_lt(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_cmp_pd_mask(a, b, _CMP_LT_OQ); }
static inline MD_SIMD_MASK simd_mask_from_u32(unsigned int a) { return _cvtu32_mask8(a); }
static inline unsigned int simd_mask_to_u32(MD_SIMD_MASK a) { return _cvtmask8_u32(a); }
static inline MD_SIMD_FLOAT simd_load(MD_FLOAT *p) { return _mm512_load_pd(p); }
static inline void simd_store(MD_FLOAT *p, MD_SIMD_FLOAT a) { _mm512_store_pd(p, a); }
static inline MD_SIMD_FLOAT select_by_mask(MD_SIMD_FLOAT a, MD_SIMD_MASK m) { return _mm512_mask_mov_pd(_mm512_setzero_pd(), m, a); }
static inline MD_FLOAT simd_h_reduce_sum(MD_SIMD_FLOAT a) {
MD_SIMD_FLOAT x = _mm512_add_pd(a, _mm512_shuffle_f64x2(a, a, 0xee));
x = _mm512_add_pd(x, _mm512_shuffle_f64x2(x, x, 0x11));
x = _mm512_add_pd(x, _mm512_permute_pd(x, 0x01));
return *((MD_FLOAT *) &x);
}
static inline MD_FLOAT simd_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1, MD_SIMD_FLOAT v2, MD_SIMD_FLOAT v3) {
__m512d t0, t2;
__m256d t3, t4;
t0 = _mm512_add_pd(v0, _mm512_permute_pd(v0, 0x55));
t2 = _mm512_add_pd(v2, _mm512_permute_pd(v2, 0x55));
t0 = _mm512_mask_add_pd(t0, simd_mask_from_u32(0xaa), v1, _mm512_permute_pd(v1, 0x55));
t2 = _mm512_mask_add_pd(t2, simd_mask_from_u32(0xaa), v3, _mm512_permute_pd(v3, 0x55));
t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0x4e));
t0 = _mm512_mask_add_pd(t0, simd_mask_from_u32(0xF0), t2, _mm512_shuffle_f64x2(t2, t2, 0x4e));
t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0xb1));
t0 = _mm512_mask_shuffle_f64x2(t0, simd_mask_from_u32(0x0C), t0, t0, 0xee);
t3 = _mm512_castpd512_pd256(t0);
t4 = _mm256_load_pd(m);
t4 = _mm256_add_pd(t4, t3);
_mm256_store_pd(m, t4);
t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4e));
t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xb1));
return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
}
static inline MD_SIMD_FLOAT simd_load_h_duplicate(const MD_FLOAT *m) {
return _mm512_broadcast_f64x4(_mm256_load_pd(m));
}
static inline MD_SIMD_FLOAT simd_load_h_dual(const MD_FLOAT *m) {
return _mm512_insertf64x4(_mm512_broadcastsd_pd(_mm_load_sd(m)), _mm256_broadcastsd_pd(_mm_load_sd(m + 1)), 1);
}
static inline MD_FLOAT simd_h_dual_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1) {
__m512d t0;
__m256d t2, t3;
t0 = _mm512_add_pd(v0, _mm512_permutex_pd(v0, 0x4e));
t0 = _mm512_mask_add_pd(t0, simd_mask_from_u32(0xccul), v1, _mm512_permutex_pd(v1, 0x4e));
t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xb1));
t0 = _mm512_mask_shuffle_f64x2(t0, simd_mask_from_u32(0xaaul), t0, t0, 0xee);
t2 = _mm512_castpd512_pd256(t0);
t3 = _mm256_load_pd(m);
t3 = _mm256_add_pd(t3, t2);
_mm256_store_pd(m, t3);
t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4e));
t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xb1));
return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
}
inline void simd_h_decr(MD_FLOAT *m, MD_SIMD_FLOAT a) {
__m256d t;
a = _mm512_add_pd(a, _mm512_shuffle_f64x2(a, a, 0xee));
t = _mm256_load_pd(m);
t = _mm256_sub_pd(t, _mm512_castpd512_pd256(a));
_mm256_store_pd(m, t);
}
static inline void simd_h_decr3(MD_FLOAT *m, MD_SIMD_FLOAT a0, MD_SIMD_FLOAT a1, MD_SIMD_FLOAT a2) {
simd_h_decr(m, a0);
simd_h_decr(m + CLUSTER_N, a1);
simd_h_decr(m + CLUSTER_N * 2, a2);
}
// Functions used in LAMMPS kernel
static inline MD_SIMD_FLOAT simd_gather(MD_SIMD_INT vidx, const MD_FLOAT *m, int s) { return _mm512_i32gather_pd(vidx, m, s); }
static inline MD_SIMD_INT simd_int_broadcast(int scalar) { return _mm256_set1_epi32(scalar); }
static inline MD_SIMD_INT simd_int_zero() { return _mm256_setzero_si256(); }
static inline MD_SIMD_INT simd_int_seq() { return _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0); }
static inline MD_SIMD_INT simd_int_load(const int *m) { return _mm256_load_epi32(m); }
static inline MD_SIMD_INT simd_int_add(MD_SIMD_INT a, MD_SIMD_INT b) { return _mm256_add_epi32(a, b); }
static inline MD_SIMD_INT simd_int_mul(MD_SIMD_INT a, MD_SIMD_INT b) { return _mm256_mul_epi32(a, b); }
static inline MD_SIMD_INT simd_int_mask_load(const int *m, MD_SIMD_MASK k) { return _mm256_mask_load_epi32(simd_int_zero(), k, m); }
static inline MD_SIMD_MASK simd_mask_int_cond_lt(MD_SIMD_INT a, MD_SIMD_INT b) { return _mm256_cmp_epi32_mask(a, b, _MM_CMPINT_LT); }

101
gromacs/includes/simd/avx512_float.h
View File

@@ -1,101 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <stdlib.h>
#include <string.h>
#include <immintrin.h>
#include <zmmintrin.h>
#define MD_SIMD_FLOAT __m512
#define MD_SIMD_MASK __mmask16
static inline MD_SIMD_FLOAT simd_broadcast(float scalar) { return _mm512_set1_ps(scalar); }
static inline MD_SIMD_FLOAT simd_zero() { return _mm512_set1_ps(0.0f); }
static inline MD_SIMD_FLOAT simd_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_add_ps(a, b); }
static inline MD_SIMD_FLOAT simd_sub(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_sub_ps(a, b); }
static inline MD_SIMD_FLOAT simd_mul(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_mul_ps(a, b); }
static inline MD_SIMD_FLOAT simd_fma(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_FLOAT c) { return _mm512_fmadd_ps(a, b, c); }
static inline MD_SIMD_FLOAT simd_reciprocal(MD_SIMD_FLOAT a) { return _mm512_rcp14_ps(a); }
static inline MD_SIMD_FLOAT simd_masked_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_MASK m) { return _mm512_mask_add_ps(a, m, a, b); }
static inline MD_SIMD_MASK simd_mask_and(MD_SIMD_MASK a, MD_SIMD_MASK b) { return _kand_mask16(a, b); }
static inline MD_SIMD_MASK simd_mask_cond_lt(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm512_cmp_ps_mask(a, b, _CMP_LT_OQ); }
static inline MD_SIMD_MASK simd_mask_from_u32(unsigned int a) { return _cvtu32_mask16(a); }
static inline unsigned int simd_mask_to_u32(MD_SIMD_MASK a) { return _cvtmask16_u32(a); }
static inline MD_SIMD_FLOAT simd_load(MD_FLOAT *p) { return _mm512_load_ps(p); }
static inline void simd_store(MD_FLOAT *p, MD_SIMD_FLOAT a) { _mm512_store_ps(p, a); }
static inline MD_SIMD_FLOAT select_by_mask(MD_SIMD_FLOAT a, MD_SIMD_MASK m) { return _mm512_mask_mov_ps(_mm512_setzero_ps(), m, a); }
static inline MD_FLOAT simd_h_reduce_sum(MD_SIMD_FLOAT a) {
// This would only be called in a Mx16 configuration, which is not valid in GROMACS
fprintf(stderr, "simd_h_reduce_sum(): Called with AVX512 intrinsics and single-precision which is not valid!\n");
exit(-1);
return 0.0;
}
static inline MD_FLOAT simd_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1, MD_SIMD_FLOAT v2, MD_SIMD_FLOAT v3) {
// This would only be called in a Mx16 configuration, which is not valid in GROMACS
fprintf(stderr, "simd_h_reduce_sum(): Called with AVX512 intrinsics and single-precision which is not valid!\n");
exit(-1);
return 0.0;
}
static inline MD_SIMD_FLOAT simd_load_h_duplicate(const float* m) {
return _mm512_castpd_ps(_mm512_broadcast_f64x4(_mm256_load_pd((const double *)(m))));
}
static inline MD_SIMD_FLOAT simd_load_h_dual(const float* m) {
return _mm512_shuffle_f32x4(_mm512_broadcastss_ps(_mm_load_ss(m)), _mm512_broadcastss_ps(_mm_load_ss(m + 1)), 0x44);
}
static inline MD_FLOAT simd_h_dual_incr_reduced_sum(float* m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1) {
__m512 t0, t1;
__m128 t2, t3;
t0 = _mm512_shuffle_f32x4(v0, v1, 0x88);
t1 = _mm512_shuffle_f32x4(v0, v1, 0xdd);
t0 = _mm512_add_ps(t0, t1);
t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0x4e));
t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xb1));
t0 = _mm512_maskz_compress_ps(simd_mask_from_u32(0x1111ul), t0);
t3 = _mm512_castps512_ps128(t0);
t2 = _mm_load_ps(m);
t2 = _mm_add_ps(t2, t3);
_mm_store_ps(m, t2);
t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4e));
t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xb1));
return _mm_cvtss_f32(t3);
}
inline void simd_h_decr(MD_FLOAT *m, MD_SIMD_FLOAT a) {
__m256 t;
a = _mm512_add_ps(a, _mm512_shuffle_f32x4(a, a, 0xee));
t = _mm256_load_ps(m);
t = _mm256_sub_ps(t, _mm512_castps512_ps256(a));
_mm256_store_ps(m, t);
}
static inline void simd_h_decr3(MD_FLOAT *m, MD_SIMD_FLOAT a0, MD_SIMD_FLOAT a1, MD_SIMD_FLOAT a2) {
simd_h_decr(m, a0);
simd_h_decr(m + CLUSTER_N, a1);
simd_h_decr(m + CLUSTER_N * 2, a2);
}

145
gromacs/includes/simd/avx_avx2_double.h
View File

@@ -1,145 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <stdlib.h>
#include <string.h>
#include <immintrin.h>
#define MD_SIMD_FLOAT __m256d
#define MD_SIMD_INT __m128i
#ifdef MASK_REGISTERS
# define MD_SIMD_MASK __mmask8
#else
# define MD_SIMD_MASK __m256d
#endif
static inline MD_SIMD_FLOAT simd_broadcast(MD_FLOAT scalar) { return _mm256_set1_pd(scalar); }
static inline MD_SIMD_FLOAT simd_zero() { return _mm256_set1_pd(0.0); }
static inline MD_SIMD_FLOAT simd_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_add_pd(a, b); }
static inline MD_SIMD_FLOAT simd_sub(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_sub_pd(a, b); }
static inline MD_SIMD_FLOAT simd_mul(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_mul_pd(a, b); }
static inline MD_SIMD_FLOAT simd_load(MD_FLOAT *p) { return _mm256_load_pd(p); }
static inline void simd_store(MD_FLOAT *p, MD_SIMD_FLOAT a) { _mm256_store_pd(p, a); }
static inline MD_SIMD_FLOAT simd_load_h_duplicate(const MD_FLOAT *m) {
MD_SIMD_FLOAT ret;
fprintf(stderr, "simd_load_h_duplicate(): Not implemented for AVX/AVX2 with double precision!");
exit(-1);
return ret;
}
static inline MD_SIMD_FLOAT simd_load_h_dual(const MD_FLOAT *m) {
MD_SIMD_FLOAT ret;
fprintf(stderr, "simd_load_h_dual(): Not implemented for AVX/AVX2 with double precision!");
exit(-1);
return ret;
}
static inline MD_FLOAT simd_h_dual_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1) {
fprintf(stderr, "simd_h_dual_incr_reduced_sum(): Not implemented for AVX/AVX2 with double precision!");
exit(-1);
return 0.0;
}
static inline MD_FLOAT simd_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1, MD_SIMD_FLOAT v2, MD_SIMD_FLOAT v3) {
__m256d t0, t1, t2;
__m128d a0, a1;
t0 = _mm256_hadd_pd(v0, v1);
t1 = _mm256_hadd_pd(v2, v3);
t2 = _mm256_permute2f128_pd(t0, t1, 0x21);
t0 = _mm256_add_pd(t0, t2);
t1 = _mm256_add_pd(t1, t2);
t0 = _mm256_blend_pd(t0, t1, 0b1100);
t1 = _mm256_add_pd(t0, _mm256_load_pd(m));
_mm256_store_pd(m, t1);
t0 = _mm256_add_pd(t0, _mm256_permute_pd(t0, 0b0101));
a0 = _mm256_castpd256_pd128(t0);
a1 = _mm256_extractf128_pd(t0, 0x1);
a0 = _mm_add_sd(a0, a1);
return *((MD_FLOAT *) &a0);
}
#ifdef NO_AVX2
static inline MD_SIMD_FLOAT select_by_mask(MD_SIMD_FLOAT a, MD_SIMD_MASK m) { return _mm256_and_pd(a, m); }
static inline MD_SIMD_FLOAT simd_reciprocal(MD_SIMD_FLOAT a) { return _mm256_cvtps_pd(_mm_rcp_ps(_mm256_cvtpd_ps(a))); }
static inline MD_SIMD_FLOAT simd_fma(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_FLOAT c) { return simd_add(simd_mul(a, b), c); }
static inline MD_SIMD_FLOAT simd_masked_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_MASK m) { return simd_add(a, _mm256_and_pd(b, m)); }
static inline MD_SIMD_MASK simd_mask_cond_lt(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_cmp_pd(a, b, _CMP_LT_OQ); }
static inline MD_SIMD_MASK simd_mask_and(MD_SIMD_MASK a, MD_SIMD_MASK b) { return _mm256_and_pd(a, b); }
// TODO: Initialize all diagonal cases and just select the proper one (all bits set or diagonal) based on cond0
static inline MD_SIMD_MASK simd_mask_from_u32(unsigned int a) {
const unsigned long long int all = 0xFFFFFFFFFFFFFFFF;
const unsigned long long int none = 0x0;
return _mm256_castsi256_pd(_mm256_set_epi64x((a & 0x8) ? all : none, (a & 0x4) ? all : none, (a & 0x2) ? all : none, (a & 0x1) ? all : none));
}
// TODO: Implement this, althrough it is just required for debugging
static inline int simd_mask_to_u32(MD_SIMD_MASK a) { return 0; }
static inline MD_FLOAT simd_h_reduce_sum(MD_SIMD_FLOAT a) {
__m128d a0, a1;
a = _mm256_add_pd(a, _mm256_permute_pd(a, 0b0101));
a0 = _mm256_castpd256_pd128(a);
a1 = _mm256_extractf128_pd(a, 0x1);
a0 = _mm_add_sd(a0, a1);
return *((MD_FLOAT *) &a0);
}
#else // AVX2
static inline MD_SIMD_FLOAT select_by_mask(MD_SIMD_FLOAT a, MD_SIMD_MASK m) { return _mm256_mask_mov_pd(_mm256_setzero_pd(), m, a); }
static inline MD_SIMD_FLOAT simd_reciprocal(MD_SIMD_FLOAT a) { return _mm256_rcp14_pd(a); }
static inline MD_SIMD_FLOAT simd_fma(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_FLOAT c) { return _mm256_fmadd_pd(a, b, c); }
static inline MD_SIMD_FLOAT simd_masked_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_MASK m) { return _mm256_mask_add_pd(a, m, a, b); }
static inline MD_SIMD_MASK simd_mask_cond_lt(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_cmp_pd_mask(a, b, _CMP_LT_OQ); }
static inline MD_SIMD_MASK simd_mask_and(MD_SIMD_MASK a, MD_SIMD_MASK b) { return _kand_mask8(a, b); }
static inline MD_SIMD_MASK simd_mask_from_u32(unsigned int a) { return _cvtu32_mask8(a); }
static inline unsigned int simd_mask_to_u32(MD_SIMD_MASK a) { return _cvtmask8_u32(a); }
static inline MD_FLOAT simd_h_reduce_sum(MD_SIMD_FLOAT a) {
__m128d a0, a1;
// test with shuffle & add as an alternative to hadd later
a = _mm256_hadd_pd(a, a);
a0 = _mm256_castpd256_pd128(a);
a1 = _mm256_extractf128_pd(a, 0x1);
a0 = _mm_add_sd(a0, a1);
return *((MD_FLOAT *) &a0);
}
static inline void simd_h_decr3(MD_FLOAT *m, MD_SIMD_FLOAT a0, MD_SIMD_FLOAT a1, MD_SIMD_FLOAT a2) {
fprintf(stderr, "simd_h_decr3(): Not implemented for AVX/AVX2 with double precision!");
exit(-1);
}
#endif
// Functions used in LAMMPS kernel
static inline MD_SIMD_FLOAT simd_gather(MD_SIMD_INT vidx, const MD_FLOAT *m, int s) { return _mm256_i32gather_pd(m, vidx, s); }
static inline MD_SIMD_INT simd_int_broadcast(int scalar) { return _mm_set1_epi32(scalar); }
static inline MD_SIMD_INT simd_int_zero() { return _mm_setzero_si128(); }
static inline MD_SIMD_INT simd_int_seq() { return _mm_set_epi32(3, 2, 1, 0); }
static inline MD_SIMD_INT simd_int_load(const int *m) { return _mm_load_si128((__m128i const *) m); }
static inline MD_SIMD_INT simd_int_add(MD_SIMD_INT a, MD_SIMD_INT b) { return _mm_add_epi32(a, b); }
static inline MD_SIMD_INT simd_int_mul(MD_SIMD_INT a, MD_SIMD_INT b) { return _mm_mul_epi32(a, b); }
static inline MD_SIMD_INT simd_int_mask_load(const int *m, MD_SIMD_MASK k) { return simd_int_load(m) & _mm256_cvtpd_epi32(k); }
static inline MD_SIMD_MASK simd_mask_int_cond_lt(MD_SIMD_INT a, MD_SIMD_INT b) { return _mm256_cvtepi32_pd(_mm_cmplt_epi32(a, b)); }

101
gromacs/includes/simd/avx_avx2_float.h
View File

@@ -1,101 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <immintrin.h>
#include <zmmintrin.h>
#define MD_SIMD_FLOAT __m256
#define MD_SIMD_MASK __mmask8
static inline MD_SIMD_FLOAT simd_broadcast(MD_FLOAT scalar) { return _mm256_set1_ps(scalar); }
static inline MD_SIMD_FLOAT simd_zero() { return _mm256_set1_ps(0.0); }
static inline MD_SIMD_FLOAT simd_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_add_ps(a, b); }
static inline MD_SIMD_FLOAT simd_sub(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_sub_ps(a, b); }
static inline MD_SIMD_FLOAT simd_mul(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_mul_ps(a, b); }
static inline MD_SIMD_FLOAT simd_load(MD_FLOAT *p) { return _mm256_load_ps(p); }
static inline void simd_store(MD_FLOAT *p, MD_SIMD_FLOAT a) { _mm256_store_ps(p, a); }
static inline MD_SIMD_FLOAT select_by_mask(MD_SIMD_FLOAT a, MD_SIMD_MASK m) { return _mm256_mask_mov_ps(_mm256_setzero_ps(), m, a); }
static inline MD_SIMD_FLOAT simd_reciprocal(MD_SIMD_FLOAT a) { return _mm256_rcp14_ps(a); }
static inline MD_SIMD_FLOAT simd_fma(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_FLOAT c) { return _mm256_fmadd_ps(a, b, c); }
static inline MD_SIMD_FLOAT simd_masked_add(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b, MD_SIMD_MASK m) { return _mm256_mask_add_ps(a, m, a, b); }
static inline MD_SIMD_MASK simd_mask_cond_lt(MD_SIMD_FLOAT a, MD_SIMD_FLOAT b) { return _mm256_cmp_ps_mask(a, b, _CMP_LT_OQ); }
static inline MD_SIMD_MASK simd_mask_and(MD_SIMD_MASK a, MD_SIMD_MASK b) { return _kand_mask8(a, b); }
static inline MD_SIMD_MASK simd_mask_from_u32(unsigned int a) { return _cvtu32_mask8(a); }
static inline unsigned int simd_mask_to_u32(MD_SIMD_MASK a) { return _cvtmask8_u32(a); }
static inline MD_FLOAT simd_h_reduce_sum(MD_SIMD_FLOAT a) {
__m128 t0;
t0 = _mm_add_ps(_mm256_castps256_ps128(a), _mm256_extractf128_ps(a, 0x1));
t0 = _mm_add_ps(t0, _mm_permute_ps(t0, _MM_SHUFFLE(1, 0, 3, 2)));
t0 = _mm_add_ss(t0, _mm_permute_ps(t0, _MM_SHUFFLE(0, 3, 2, 1)));
return *((MD_FLOAT *) &t0);
}
static inline MD_FLOAT simd_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1, MD_SIMD_FLOAT v2, MD_SIMD_FLOAT v3) {
__m128 t0, t2;
v0 = _mm256_hadd_ps(v0, v1);
v2 = _mm256_hadd_ps(v2, v3);
v0 = _mm256_hadd_ps(v0, v2);
t0 = _mm_add_ps(_mm256_castps256_ps128(v0), _mm256_extractf128_ps(v0, 0x1));
t2 = _mm_add_ps(t0, _mm_load_ps(m));
_mm_store_ps(m, t2);
t0 = _mm_add_ps(t0, _mm_permute_ps(t0, _MM_SHUFFLE(1, 0, 3, 2)));
t0 = _mm_add_ss(t0, _mm_permute_ps(t0, _MM_SHUFFLE(0, 3, 2, 1)));
return *((MD_FLOAT *) &t0);
}
static inline MD_SIMD_FLOAT simd_load_h_duplicate(const MD_FLOAT *m) {
return _mm256_broadcast_ps((const __m128 *)(m));
}
static inline MD_SIMD_FLOAT simd_load_h_dual(const MD_FLOAT *m) {
__m128 t0, t1;
t0 = _mm_broadcast_ss(m);
t1 = _mm_broadcast_ss(m + 1);
return _mm256_insertf128_ps(_mm256_castps128_ps256(t0), t1, 0x1);
}
static inline MD_FLOAT simd_h_dual_incr_reduced_sum(MD_FLOAT *m, MD_SIMD_FLOAT v0, MD_SIMD_FLOAT v1) {
__m128 t0, t1;
v0 = _mm256_hadd_ps(v0, v1);
t0 = _mm256_extractf128_ps(v0, 0x1);
t0 = _mm_hadd_ps(_mm256_castps256_ps128(v0), t0);
t0 = _mm_permute_ps(t0, _MM_SHUFFLE(3, 1, 2, 0));
t1 = _mm_add_ps(t0, _mm_load_ps(m));
_mm_store_ps(m, t1);
t0 = _mm_add_ps(t0, _mm_permute_ps(t0, _MM_SHUFFLE(1, 0, 3, 2)));
t0 = _mm_add_ss(t0, _mm_permute_ps(t0, _MM_SHUFFLE(0, 3, 2, 1)));
return *((MD_FLOAT *) &t0);
}
inline void simd_h_decr(MD_FLOAT *m, MD_SIMD_FLOAT a) {
__m128 asum = _mm_add_ps(_mm256_castps256_ps128(a), _mm256_extractf128_ps(a, 0x1));
_mm_store_ps(m, _mm_sub_ps(_mm_load_ps(m), asum));
}
static inline void simd_h_decr3(MD_FLOAT *m, MD_SIMD_FLOAT a0, MD_SIMD_FLOAT a1, MD_SIMD_FLOAT a2) {
simd_h_decr(m, a0);
simd_h_decr(m + CLUSTER_N, a1);
simd_h_decr(m + CLUSTER_N * 2, a2);
}

31
gromacs/includes/thermo.h
View File

@@ -1,31 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#include <parameter.h>
#include <atom.h>
#ifndef __THERMO_H_
#define __THERMO_H_
extern void setupThermo(Parameter*, int);
extern void computeThermo(int, Parameter*, Atom*);
extern void adjustThermo(Parameter*, Atom*);
#endif

11
gromacs/includes/timers.h
View File

@@ -1,11 +0,0 @@
#ifndef __TIMERS_H_
#define __TIMERS_H_
typedef enum {
TOTAL = 0,
NEIGH,
FORCE,
NUMTIMER
} timertype;
#endif

30
gromacs/includes/timing.h
View File

@@ -1,30 +0,0 @@
/*
* =======================================================================================
*
* Author: Jan Eitzinger (je), jan.eitzinger@fau.de
* Copyright (c) 2020 RRZE, University Erlangen-Nuremberg
*
* This file is part of MD-Bench.
*
* MD-Bench is free software: you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published
* by the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* MD-Bench is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
* PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License along
* with MD-Bench. If not, see <https://www.gnu.org/licenses/>.
* =======================================================================================
*/
#ifndef __TIMING_H_
#define __TIMING_H_
extern double getTimeStamp();
extern double getTimeResolution();
extern double getTimeStamp_();
#endif