Adjust ISA options and improve output

Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>

Makefile

@@ -64,8 +64,12 @@ ifneq ($(VECTOR_WIDTH),)
     DEFINES += -DVECTOR_WIDTH=$(VECTOR_WIDTH)
 endif
 
-ifeq ($(strip $(NO_AVX2)),true)
+ifeq ($(strip $(MASK_REGISTERS)),true)
-    DEFINES += -DNO_AVX2
+    DEFINES += -DMASK_REGISTERS
+endif
+
+ifeq ($(strip $(SIMD_KERNEL_AVAILABLE)),true)
+    DEFINES += -DSIMD_KERNEL_AVAILABLE
 endif
 
 ifeq ($(strip $(AVX512)),true)

config.mk

@@ -2,6 +2,9 @@
 TAG ?= ICC
 # Instruction set (SSE/AVX/AVX2/AVX512)
 ISA ?= AVX512
+# Use mask registers (AVX512 must be available in the target CPU)
+# This is always true when ISA is set to AVX512
+MASK_REGISTERS ?= true
 # Optimization scheme (lammps/gromacs/clusters_per_bin)
 OPT_SCHEME ?= lammps
 # Enable likwid (true or false)

gromacs/includes/simd/avx_avx2_double.h

@@ -28,10 +28,10 @@
 #define MD_SIMD_FLOAT       __m256d
 #define MD_SIMD_INT         __m128i
 
-#ifdef NO_AVX2
+#ifdef MASK_REGISTERS
-#   define MD_SIMD_MASK     __m256d
-#else
 #   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); }

include_ISA.mk

@@ -2,12 +2,13 @@ ifeq ($(strip $(ISA)), SSE)
 _VECTOR_WIDTH=2
 else ifeq ($(strip $(ISA)), AVX)
 # Vector width is 4 but AVX2 instruction set is not supported
-NO_AVX2=true
 _VECTOR_WIDTH=4
 else ifeq ($(strip $(ISA)), AVX2)
+#SIMD_KERNEL_AVAILABLE=true
 _VECTOR_WIDTH=4
 else ifeq ($(strip $(ISA)), AVX512)
 AVX512=true
+SIMD_KERNEL_AVAILABLE=true
 _VECTOR_WIDTH=8
 endif
 

lammps/force_dem.c

@@ -21,16 +21,14 @@
  * =======================================================================================
  */
 #include <math.h>
+//---
+#include <atom.h>
 #include <likwid-marker.h>
-
-#include <timing.h>
 #include <neighbor.h>
 #include <parameter.h>
-#include <atom.h>
 #include <stats.h>
+#include <timing.h>
 
-// TODO: Joint common files for gromacs and lammps variants
-#include "../gromacs/includes/simd.h"
 
 double computeForceDemFullNeigh(Parameter *param, Atom *atom, Neighbor *neighbor, Stats *stats) {
     int Nlocal = atom->Nlocal;

lammps/force_lj.c

@@ -20,16 +20,20 @@
  *   with MD-Bench.  If not, see <https://www.gnu.org/licenses/>.
  * =======================================================================================
  */
+#include <stdio.h>
+#include <stdlib.h>
+//---
+#include <atom.h>
 #include <likwid-marker.h>
-
-#include <timing.h>
 #include <neighbor.h>
 #include <parameter.h>
-#include <atom.h>
 #include <stats.h>
+#include <timing.h>
 
 // TODO: Joint common files for gromacs and lammps variants
+#ifdef SIMD_KERNEL_AVAILABLE
 #include "../gromacs/includes/simd.h"
+#endif
 
 double computeForceLJFullNeigh_plain_c(Parameter *param, Atom *atom, Neighbor *neighbor, Stats *stats) {
     int Nlocal = atom->Nlocal;
@@ -194,11 +198,6 @@ double computeForceLJFullNeigh_simd(Parameter *param, Atom *atom, Neighbor *neig
     MD_FLOAT cutforcesq = param->cutforce * param->cutforce;
     MD_FLOAT sigma6 = param->sigma6;
     MD_FLOAT epsilon = param->epsilon;
-    MD_SIMD_FLOAT cutforcesq_vec = simd_broadcast(cutforcesq);
-    MD_SIMD_FLOAT sigma6_vec = simd_broadcast(sigma6);
-    MD_SIMD_FLOAT eps_vec = simd_broadcast(epsilon);
-    MD_SIMD_FLOAT c48_vec = simd_broadcast(48.0);
-    MD_SIMD_FLOAT c05_vec = simd_broadcast(0.5);
 
     for(int i = 0; i < Nlocal; i++) {
         atom_fx(i) = 0.0;
@@ -209,10 +208,16 @@ double computeForceLJFullNeigh_simd(Parameter *param, Atom *atom, Neighbor *neig
     double S = getTimeStamp();
     LIKWID_MARKER_START("force");
 
-    #ifdef NO_AVX2
+    #ifndef SIMD_KERNEL_AVAILABLE
-    fprintf(stderr, "Error: SIMD kernel implemented for AVX2 and AVX512 only!");
+    fprintf(stderr, "Error: SIMD kernel not implemented for specified instruction set!");
     exit(-1);
     #else
+    MD_SIMD_FLOAT cutforcesq_vec = simd_broadcast(cutforcesq);
+    MD_SIMD_FLOAT sigma6_vec = simd_broadcast(sigma6);
+    MD_SIMD_FLOAT eps_vec = simd_broadcast(epsilon);
+    MD_SIMD_FLOAT c48_vec = simd_broadcast(48.0);
+    MD_SIMD_FLOAT c05_vec = simd_broadcast(0.5);
+
     #pragma omp parallel for
     for(int i = 0; i < Nlocal; i++) {
         neighs = &neighbor->neighbors[i * neighbor->maxneighs];

lammps/includes/util.h

@@ -26,12 +26,15 @@
 #ifndef MIN
 #   define MIN(x,y) ((x)<(y)?(x):(y))
 #endif
+
 #ifndef MAX
 #   define MAX(x,y) ((x)>(y)?(x):(y))
 #endif
+
 #ifndef ABS
 #   define ABS(a) ((a) >= 0 ? (a) : -(a))
 #endif
+
 #ifdef DEBUG
 #   define DEBUG_MESSAGE   printf
 #else
@@ -46,6 +49,12 @@
 #define FF_EAM  1
 #define FF_DEM  2
 
+#if PRECISION == 1
+#   define PRECISION_STRING     "single"
+#else
+#   define PRECISION_STRING     "double"
+#endif
+
 extern double myrandom(int*);
 extern void random_reset(int *seed, int ibase, double *coord);
 extern int str2ff(const char *string);

lammps/main.c

@@ -226,6 +226,7 @@ int main(int argc, char** argv) {
     param.cutneigh = param.cutforce + param.skin;
     setup(&param, &eam, &atom, &neighbor, &stats);
     printParameter(&param);
+    printf(HLINE);
 
     printf("step\ttemp\t\tpressure\n");
     computeThermo(0, &param, &atom);
@@ -272,14 +273,6 @@ int main(int argc, char** argv) {
     timer[TOTAL] = getTimeStamp() - timer[TOTAL];
     computeThermo(-1, &param, &atom);
 
-    printf(HLINE);
-    printf("Force field: %s\n", ff2str(param.force_field));
-    printf("Data layout for positions: %s\n", POS_DATA_LAYOUT);
-    #if PRECISION == 1
-    printf("Using single precision floating point.\n");
-    #else
-    printf("Using double precision floating point.\n");
-    #endif
     printf(HLINE);
     printf("System: %d atoms %d ghost atoms, Steps: %d\n", atom.Natoms, atom.Nghost, param.ntimes);
     printf("TOTAL %.2fs FORCE %.2fs NEIGH %.2fs REST %.2fs\n",

lammps/parameter.c

@@ -24,6 +24,7 @@
 #include <stdlib.h>
 #include <string.h>
 //---
+#include <atom.h>
 #include <parameter.h>
 #include <util.h>
 
@@ -154,6 +155,9 @@ void printParameter(Parameter *param) {
     }
 
     printf("\tForce field: %s\n", ff2str(param->force_field));
+    printf("\tKernel: %s\n", KERNEL_NAME);
+    printf("\tData layout: %s\n", POS_DATA_LAYOUT);
+    printf("\tFloating-point precision: %s\n", PRECISION_STRING);
     printf("\tUnit cells (nx, ny, nz): %d, %d, %d\n", param->nx, param->ny, param->nz);
     printf("\tDomain box sizes (x, y, z): %e, %e, %e\n", param->xprd, param->yprd, param->zprd);
     printf("\tPeriodic (x, y, z): %d, %d, %d\n", param->pbc_x, param->pbc_y, param->pbc_z);
@@ -175,5 +179,5 @@ void printParameter(Parameter *param) {
     printf("\tCutoff radius: %e\n", param->cutforce);
     printf("\tSkin: %e\n", param->skin);
     printf("\tHalf neighbor lists: %d\n", param->half_neigh);
-    printf("\tProcessor frequency (GHz): %.4f\n\n", param->proc_freq);
+    printf("\tProcessor frequency (GHz): %.4f\n", param->proc_freq);
 }