Improved the performance of the code significantly using SIMD and OpenMP

Makefile.Defines

@@ -46,37 +46,35 @@ COLLRESOLVE_HOME = $(ROOT_DIR)/collresolve/
 # DO NOT include in FFLAGS the "-c" option to compile object only
 # this is done explicitly as needed in the Makefile
 ADVIXE_DIR = /apps/cent7/intel/advisor_2019
-ADVIXE_FLAGS = -g -O2 -qopt-report=5 -vec -vecabi=cmdtarget -simd -shared-intel -debug inline-debug-info -DTBB_DEBUG -DTBB_USE_THREADING_TOOLS -fp-model no-except -mp1 -xhost -traceback 
+ADVIXE_FLAGS = -g -O2 -qopt-report=5 -vecabi=cmdtarget -simd -shared-intel -debug inline-debug-info -DTBB_DEBUG -DTBB_USE_THREADING_TOOLS -xhost -traceback 
 
-VTUNE_FLAGS = -g -O2 -vec -simd -shared-intel -qopenmp -debug inline-debug-info -parallel-source-info=2 -parallel -DTBB_DEBUG -DTBB_USE_THREADING_TOOLS -qopenmp -fp-model no-except -mp1 -xhost -traceback 
+VTUNE_FLAGS = -g -O2 -qopt-report=5 -simd -shared-intel -qopenmp -debug inline-debug-info -parallel-source-info=2 -parallel -DTBB_DEBUG -DTBB_USE_THREADING_TOOLS -qopenmp -fp-model no-except -mp1 -xhost -traceback 
 #Be sure to set the environment variable KMP_FORKJOIN_FRAMES=1 for OpenMP debuging in vtune
 
 IDEBUG = -O0 -init=snan,arrays  -nogen-interfaces -no-pie -no-ftz -fpe-all=0 -g -traceback -mp1 -fp-model strict -fpe0 -debug all -align all -pad -ip -prec-div -prec-sqrt -assume protect-parens -CB -no-wrap-margin
-STRICTREAL = -fp-model strict -fp-model no-except -prec-div -prec-sqrt -assume protect-parens 
-SIMDVEC = -simd -xhost -align all -assume contiguous_assumed_shape -vecabi=cmdtarget -prec-div -prec-sqrt -assume protect-parens
+STRICTREAL = -fp-model strict -prec-div -prec-sqrt -assume protect-parens 
+SIMDVEC = -simd -xhost -align all -assume contiguous_assumed_shape -vecabi=cmdtarget -fp-model no-except 
 PAR = -qopenmp #-parallel  #Something goes wrong in SyMBA at the moment with auto-paralellization enabled
-HEAPARR = -heap-arrays 1048576 
+HEAPARR = -heap-arrays 4194304 
 OPTREPORT = -qopt-report=5 
-IPRODUCTION = -init=snan,arrays  -no-wrap-margin -O3 $(STRICTREAL) $(PAR) $(SIMDVEC) $(HEAPARR) 
+IPRODUCTION = -init=snan,arrays  -no-wrap-margin -O3 $(PAR) $(SIMDVEC) -fp-model fast #$(STRICTREAL) #$(HEAPARR)
 
 #gfortran flags
 GDEBUG = -g -Og -fbacktrace -fbounds-check -ffree-line-length-none
 GPAR = -fopenmp #-ftree-parallelize-loops=4
 GMEM =  -fsanitize-address-use-after-scope -fstack-check  -fsanitize=bounds-strict -fsanitize=undefined -fsanitize=signed-integer-overflow -fsanitize=object-size -fstack-protector-all
 GWARNINGS = -Wall -Warray-bounds  -Wimplicit-interface -Wextra -Warray-temporaries
-GPRODUCTION = -O3 -ffree-line-length-none $(GPAR)
+GPRODUCTION = -O2 -ffree-line-length-none $(GPAR)
 
-#FFLAGS  = $(IDEBUG) $(HEAPARR) $(SIMDVEC) $(PAR)
-FFLAGS  = $(IPRODUCTION) $(OPTREPORT)
+#FFLAGS  = $(IPRODUCTION) $(OPTREPORT)
+#FFLAGS = $(IPRODUCTION) $(OPTREPORT) $(ADVIXE_FLAGS)
+FFLAGS = -O3 $(PAR)  $(OPTREPORT)  $(SIMDVEC) -shared-intel  -debug inline-debug-info -DTBB_DEBUG -DTBB_USE_THREADING_TOOLS -traceback -g
 FORTRAN = ifort
 #AR     = xiar
 
 #FORTRAN = gfortran
 #FFLAGS =  $(GDEBUG) $(GMEM) $(GPAR)
-#FFLAGS =  $(GPRODUCTION) -g -fbacktrace #-fcheck=all #-Wall 
-AR     = ar
-
-# DO NOT include in CFLAGS the "-c" option to compile object only
+#FFLAGS =  $(GPRODUCTION) -g -fbacktrace #-fcheck=all #-Wall AR     = ar # DO NOT include in CFLAGS the "-c" option to compile object only
 # this is done explicitly as needed in the Makefile
 
 CC              = icc

docs/src/rmvs_encounter_check.f90

@@ -35,10 +35,7 @@ module function rmvs_encounter_check_tp(self, system, dt) result(lencounter)
                   if ((.not.tp%lmask(i)).or.(tp%plencP(i) /= 0)) cycle
                   xr(:) = tp%xh(:, i) - pl%xbeg(:, j)
                   vr(:) = tp%vh(:, i) - pl%vbeg(:, j)
-                  r2 = dot_product(xr(:), xr(:))
-                  v2 = dot_product(vr(:), vr(:))
-                  vdotr = dot_product(vr(:), xr(:))
-                  lflag = rmvs_chk_ind(r2, v2, vdotr, dt, r2crit(j))
+                  lflag = rmvs_chk_ind(xr(1), xr(2), xr(3), vr(1), vr(2), vr(3), dt, r2crit(j))
                   if (lflag) tp%plencP(i) = j
                end do
                pl%nenc(j) = count(tp%plencP(:) == j)

docs/src/symba_encounter_check.f90

@@ -202,20 +202,20 @@ module pure elemental subroutine symba_encounter_check_one(xr, yr, zr, vxr, vyr,
       !! Adapted from Hal Levison's Swift routine symba5_chk.f
       implicit none
       ! Arguments
-      real(DP),     intent(in)  :: xr, yr, zr, vxr, vyr, vzr
-      real(DP),     intent(in)  :: rhill1, rhill2, dt
-      integer(I4B), intent(in)  :: irec
-      logical,      intent(out) :: lencounter, lvdotr
+      real(DP),     intent(in)  :: xr, yr, zr     !! Relative distance vector components
+      real(DP),     intent(in)  :: vxr, vyr, vzr  !! Relative velocity vector components
+      real(DP),     intent(in)  :: rhill1, rhill2 !! Hill spheres of the two bodies
+      real(DP),     intent(in)  :: dt             !! Step size
+      integer(I4B), intent(in)  :: irec           !! Current SyMBA recursion level
+      real(DP),     intent(in)  :: r2crit         !! Square of the critical encounter distance
+      logical,      intent(out) :: lencounter     !! Flag indicating that an encounter has occurred
+      logical,      intent(out) :: lvdotr         !! Logical flag indicating the direction of the v .dot. r vector
       ! Internals
-      real(DP)     :: r2, v2, rcrit, r2crit, vdotr
-
-      rcrit = (rhill1 + rhill2)*RHSCALE*(RSHELL**(irec))
-      r2crit = rcrit**2
-      r2 = xr**2 + yr**2 + zr**2
-      v2 = vxr**2 + vyr**2 + vzr**2
-      vdotr = xr * vxr + yr * vyr + zr * vzr
-      lencounter = rmvs_chk_ind(r2, v2, vdotr, dt, r2crit)
-      lvdotr = (vdotr < 0.0_DP)
+      real(DP)     :: r2crit
+
+      r2crit = (rhill1 + rhill2)*RHSCALE*(RSHELL**(irec))
+      r2crit = r2crit**2
+      call rmvs_chk_ind(xr, yr, zr, vxr, vyr, vzr, dt, r2crit, lencounter, lvdotr)
 
       return
    end subroutine symba_encounter_check_one