Merge branch 'master' of github.com:RRZE-HPC/MD-Bench

gromacs/force_lj.c

@@ -35,9 +35,12 @@ double computeForceLJ_ref(Parameter *param, Atom *atom, Neighbor *neighbor, Stat
     }
 
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force");
 
-    #pragma omp parallel for
+    #pragma omp for
     for(int ci = 0; ci < atom->Nclusters_local; ci++) {
         int ci_cj0 = CJ0_FROM_CI(ci);
         int ci_cj1 = CJ1_FROM_CI(ci);
@@ -119,6 +122,8 @@ double computeForceLJ_ref(Parameter *param, Atom *atom, Neighbor *neighbor, Stat
     }
 
     LIKWID_MARKER_STOP("force");
+    }
+
     double E = getTimeStamp();
     DEBUG_MESSAGE("computeForceLJ end\n");
     return E-S;
@@ -149,9 +154,12 @@ double computeForceLJ_2xnn_half(Parameter *param, Atom *atom, Neighbor *neighbor
     }
 
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force");
 
-    #pragma omp parallel for
+    #pragma omp for
     for(int ci = 0; ci < atom->Nclusters_local; ci++) {
         int ci_cj0 = CJ0_FROM_CI(ci);
         #if CLUSTER_M > CLUSTER_N
@@ -266,6 +274,8 @@ double computeForceLJ_2xnn_half(Parameter *param, Atom *atom, Neighbor *neighbor
     }
 
     LIKWID_MARKER_STOP("force");
+    }
+
     double E = getTimeStamp();
     DEBUG_MESSAGE("computeForceLJ_2xnn end\n");
     return E-S;
@@ -296,9 +306,12 @@ double computeForceLJ_2xnn_full(Parameter *param, Atom *atom, Neighbor *neighbor
     }
 
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force");
 
-    #pragma omp parallel for
+    #pragma omp for
     for(int ci = 0; ci < atom->Nclusters_local; ci++) {
         int ci_cj0 = CJ0_FROM_CI(ci);
         #if CLUSTER_M > CLUSTER_N
@@ -398,6 +411,8 @@ double computeForceLJ_2xnn_full(Parameter *param, Atom *atom, Neighbor *neighbor
     }
 
     LIKWID_MARKER_STOP("force");
+    }
+
     double E = getTimeStamp();
     DEBUG_MESSAGE("computeForceLJ_2xnn end\n");
     return E-S;
@@ -435,9 +450,12 @@ double computeForceLJ_4xn_half(Parameter *param, Atom *atom, Neighbor *neighbor,
     }
 
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force");
 
-    #pragma omp parallel for
+    #pragma omp for
     for(int ci = 0; ci < atom->Nclusters_local; ci++) {
         int ci_cj0 = CJ0_FROM_CI(ci);
         #if CLUSTER_M > CLUSTER_N
@@ -591,6 +609,8 @@ double computeForceLJ_4xn_half(Parameter *param, Atom *atom, Neighbor *neighbor,
     }
 
     LIKWID_MARKER_STOP("force");
+    }
+
     double E = getTimeStamp();
     DEBUG_MESSAGE("computeForceLJ_4xn end\n");
     return E-S;
@@ -620,9 +640,12 @@ double computeForceLJ_4xn_full(Parameter *param, Atom *atom, Neighbor *neighbor,
     }
 
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force");
 
-    #pragma omp parallel for
+    #pragma omp for
     for(int ci = 0; ci < atom->Nclusters_local; ci++) {
         int ci_cj0 = CJ0_FROM_CI(ci);
         #if CLUSTER_M > CLUSTER_N
@@ -751,6 +774,8 @@ double computeForceLJ_4xn_full(Parameter *param, Atom *atom, Neighbor *neighbor,
     }
 
     LIKWID_MARKER_STOP("force");
+    }
+
     double E = getTimeStamp();
     DEBUG_MESSAGE("computeForceLJ_4xn end\n");
     return E-S;

include_ICX.mk

@@ -11,7 +11,7 @@ endif
 
 ifeq ($(ISA),AVX2)
 OPTS     = -Ofast -xCORE-AVX2  $(PROFILE)
-#OPTS     = -Ofast -xAVX2  $(PROFILE)
+#OPTS     = -Ofast -xHost  $(PROFILE)
 #OPTS     = -Ofast -march=core-avx2 $(PROFILE)
 endif
 

lammps/force_eam.c

@@ -31,8 +31,12 @@ double computeForceEam(Eam* eam, Parameter* param, Atom *atom, Neighbor *neighbo
     int nrho = eam->nrho; int nrho_tot = eam->nrho_tot;
     double S = getTimeStamp();
 
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force_eam_fp");
-    #pragma omp parallel for
+
+    #pragma omp for
     for(int i = 0; i < Nlocal; i++) {
         neighs = &neighbor->neighbors[i * neighbor->maxneighs];
         int numneighs = neighbor->numneigh[i];
@@ -95,13 +99,19 @@ double computeForceEam(Eam* eam, Parameter* param, Atom *atom, Neighbor *neighbo
     }
 
     LIKWID_MARKER_STOP("force_eam_fp");
+    }
 
     // We still need to update fp for PBC atoms
     for(int i = 0; i < atom->Nghost; i++) {
         fp[Nlocal + i] = fp[atom->border_map[i]];
     }
 
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force_eam");
+
+    #pragma omp for
     for(int i = 0; i < Nlocal; i++) {
         neighs = &neighbor->neighbors[i * neighbor->maxneighs];
         int numneighs = neighbor->numneigh[i];
@@ -192,6 +202,8 @@ double computeForceEam(Eam* eam, Parameter* param, Atom *atom, Neighbor *neighbo
     }
 
     LIKWID_MARKER_STOP("force_eam");
+    }
+
     double E = getTimeStamp();
     return E-S;
 }

lammps/force_lj.c

@@ -36,9 +36,12 @@ double computeForceLJFullNeigh_plain_c(Parameter *param, Atom *atom, Neighbor *n
         atom_fz(i) = 0.0;
     }
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("force");
 
-    #pragma omp parallel for
+    #pragma omp for
     for(int i = 0; i < Nlocal; i++) {
         neighs = &neighbor->neighbors[i * neighbor->maxneighs];
         int numneighs = neighbor->numneigh[i];
@@ -92,6 +95,8 @@ double computeForceLJFullNeigh_plain_c(Parameter *param, Atom *atom, Neighbor *n
     }
 
     LIKWID_MARKER_STOP("force");
+    }
+
     double E = getTimeStamp();
     return E-S;
 }
@@ -112,8 +117,12 @@ double computeForceLJHalfNeigh(Parameter *param, Atom *atom, Neighbor *neighbor,
     }
 
     double S = getTimeStamp();
+
+    #pragma omp parallel
+    {
     LIKWID_MARKER_START("forceLJ-halfneigh");
 
+    #pragma omp for
     for(int i = 0; i < Nlocal; i++) {
         neighs = &neighbor->neighbors[i * neighbor->maxneighs];
         int numneighs = neighbor->numneigh[i];
@@ -173,6 +182,8 @@ double computeForceLJHalfNeigh(Parameter *param, Atom *atom, Neighbor *neighbor,
     }
 
     LIKWID_MARKER_STOP("forceLJ-halfneigh");
+    }
+
     double E = getTimeStamp();
     return E-S;
 }
@@ -191,7 +202,6 @@ double computeForceLJFullNeigh_simd(Parameter *param, Atom *atom, Neighbor *neig
     }
 
     double S = getTimeStamp();
-    LIKWID_MARKER_START("force");
 
     #ifndef __SIMD_KERNEL__
     fprintf(stderr, "Error: SIMD kernel not implemented for specified instruction set!");
@@ -203,7 +213,12 @@ double computeForceLJFullNeigh_simd(Parameter *param, Atom *atom, Neighbor *neig
     MD_SIMD_FLOAT c48_vec = simd_broadcast(48.0);
     MD_SIMD_FLOAT c05_vec = simd_broadcast(0.5);
 
-    #pragma omp parallel for
+
+    #pragma omp parallel
+    {
+    LIKWID_MARKER_START("force");
+
+    #pragma omp for
     for(int i = 0; i < Nlocal; i++) {
         neighs = &neighbor->neighbors[i * neighbor->maxneighs];
         int numneighs = neighbor->numneigh[i];
@@ -244,9 +259,11 @@ double computeForceLJFullNeigh_simd(Parameter *param, Atom *atom, Neighbor *neig
         atom_fy(i) += simd_h_reduce_sum(fiy);
         atom_fz(i) += simd_h_reduce_sum(fiz);
     }
-    #endif
 
     LIKWID_MARKER_STOP("force");
+    }
+    #endif
+
     double E = getTimeStamp();
     return E-S;
 }

util/evaluate_latency_and_cfd.sh

@@ -1,16 +1,36 @@
 #!/bin/bash
 
-# Adjustable variables
+[[ -z "$1" ]] && echo "Use: $0 <binary> [-c <core>] [-f <freq>] [-n <nruns>] [-l <log>] [-s]" && exit
-TAG="${TAG:-ICC}"
+[[ ! -f "$1" ]] && echo "Binary file not found, make sure to use 'make'" && exit
-OPT_SCHEME="${OPT_SCHEME:-lammps}"
+[[ ! -f "$1-stub" ]] && echo "Binary file for stubbed case not found, make sure to use 'make VARIANT=stub'" && exit
+
+MDBENCH_BIN=$1
+BIN_INFO="${MDBENCH_BIN#*-}" # $OPT_SCHEME-$TAG-$ISA-$PREC
+OPT_SCHEME="${BIN_INFO%%-*}"
+PREC="${BIN_INFO##*-}"
+BIN_INFO="${BIN_INFO#*-}" # $TAG-$ISA-$PREC
+BIN_INFO="${BIN_INFO%-*}" # $TAG-$ISA
+TAG="${BIN_INFO%%-*}"
+ISA="${BIN_INFO##*-}"
 CORE="${CORE:-0}"
 FREQ="${FREQ:-2.4}"
 NRUNS="${NRUNS:-3}"
 LOG="${LOG:-latencies_and_cfds.log}"
 STUB_ONLY="${STUB_ONLY:-false}"
 
+OPTIND=2
+while getopts "c:f:n:l:s" flag; do
+    case "${flag}" in
+        c) CORE=${OPTARG};;
+        f) FREQ=${OPTARG};;
+        n) NRUNS=${OPTARG};;
+        l) LOG=${OPTARG};;
+        s) STUB_ONLY=true;;
+    esac
+done
+
 # Other useful variables
-MDBENCH_BIN=./MDBench-$TAG-$OPT_SCHEME
+MDBENCH_BIN=./MDBench-$OPT_SCHEME-$TAG-$ISA-$PREC
 FIXED_PARAMS="--freq $FREQ"
 CPU_VENDOR=$(lscpu | grep "Vendor ID" | tr -s ' ' | cut -d ' ' -f3)
 
@@ -46,6 +66,8 @@ function run_benchmark() {
 
 echo "Tag: $TAG" | tee -a $LOG
 echo "Optimization scheme: $OPT_SCHEME" | tee -a $LOG
+echo "Instruction set: $ISA" | tee -a $LOG
+echo "Precision: $PREC" | tee -a $LOG
 echo "Binary: $MDBENCH_BIN(-stub)" | tee -a $LOG
 echo "Frequency: $FREQ" | tee -a $LOG
 echo "Number of runs: $NRUNS" | tee -a $LOG