merge

src/Makefile

@@ -60,7 +60,7 @@ ifeq ($(F90), gfortran)
 		LDFLAGS += -fopenmp
 	endif
 	ifeq ($(debug), 0)
-		CFLAGS += -O3 -ffast-math
+		CFLAGS += -O2
 	else
 		CFLAGS += -O0 -Wall -fbounds-check -ftrapv -fbacktrace -fdump-core \
 			-Waliasing -Walign-commons -Wampersand -Wintrinsic-shadow \
@@ -105,7 +105,7 @@ ifeq ($(F90), pgf90)
 		LDFLAGS += -acc -ta=tesla:cc70,managed -Mcuda -lnvToolsExt
 	endif
         ifeq ($(debug), 0)
-                CFLAGS += -O3 -fast
+                CFLAGS += -O2
         else
                CFLAGS += -O0 -Mbounds -Mchkstk -traceback
                LDFLAGS +=
@@ -125,7 +125,7 @@ ifeq ($(F90), f90)
 # -fpover=yes: Provide run-time overflow check during READ
 # -xcheck=%all: Enable all run-time check
 	ifeq ($(debug), 0)
-		CFLAGS += -O3 -r8const -fast -xloopinfo -s -xipo
+		CFLAGS += -O2 -r8const -xloopinfo -s -xipo
 	else
 		CFLAGS += -w3 -r8const -C -fpover=yes -xcheck=%all
 	endif

src/dt_mass.f90

@@ -145,7 +145,7 @@ subroutine dt_mass
 enddo
 
 dt = min(dt_elastic, dt_maxwell)
-
+!$ACC update device(dt, dt_elastic, dt_maxwell, dtmax_therm)
 return
 end
 

src/fl_move.f90

@@ -274,13 +274,15 @@ end subroutine resurface
 
 
 subroutine add_marker_at_top(i, dz_ratio, time, kph, nmarkers)
+  use myrandom_mod
   use marker_data
   use arrays
   include 'precision.inc'
 
+  iseed = nloop
   do while(.true.)
-     call random_number(r1)
-     call random_number(r2)
+     call myrandom(iseed, r1)
+     call myrandom(iseed, r2)
      j = 1
 
      ! (x1, y1) and (x2, y2)

src/fl_srate.f90

@@ -1,7 +1,6 @@
 
 ! Calculation of strain rates from velocities 
 ! Mixed of volumetric strain_rates (for incompressible flow)
-
 subroutine fl_srate
 use arrays
 use params
@@ -17,6 +16,10 @@ subroutine fl_srate
 !$OMP         vx1,vy1,vx2,vy2,vx3,vy3,vx4,vy4, &
 !$OMP         em,eda,edb,s11,s22,s12, &
 !$OMP         srII,srI,srs2,stII)
+!$ACC parallel loop collapse(2) private(i,j,x1,y1,x2,y2,x3,y3,x4,y4, &
+!$ACC         vx1,vy1,vx2,vy2,vx3,vy3,vx4,vy4, &
+!$ACC         em,eda,edb,s11,s22,s12, &
+!$ACC         srII,srI,srs2,stII)
 do 2  i = 1,nx-1
     do 2  j = 1,nz-1
 
@@ -109,10 +112,14 @@ subroutine fl_srate
 !$OMP end parallel do
 ! following block is needed for averaging
 dtavg = dtavg + dt
-
+!$ACC update device(dtavg)
 ! re-initialisation after navgsr steps
 if( nsrate .eq. ifreq_avgsr ) then
 !$OMP parallel do
+!$ACC parallel loop collapse(2) private(i,j,x1,y1,x2,y2,x3,y3,x4,y4, &
+!$ACC         vx1,vy1,vx2,vy2,vx3,vy3,vx4,vy4, &
+!$ACC         em,eda,edb,s11,s22,s12, &
+!$ACC         srII,srI,srs2,stII)
     do i = 1,nx-1
         do j = 1, nz-1
             e2sr(j,i) = se2sr(j,i) / dtavg
@@ -126,6 +133,10 @@ subroutine fl_srate
     nsrate = 0
 elseif( nsrate .eq. -1 ) then
 !$OMP parallel do
+!$ACC parallel loop collapse(2) private(i,j,x1,y1,x2,y2,x3,y3,x4,y4, &
+!$ACC         vx1,vy1,vx2,vy2,vx3,vy3,vx4,vy4, &
+!$ACC         em,eda,edb,s11,s22,s12, &
+!$ACC         srII,srI,srs2,stII)
     do i = 1,nx-1
         do j = 1, nz-1
             e2sr(j,i) = se2sr(j,i) / dtavg

src/fl_therm.f90

@@ -2,6 +2,8 @@
 ! Calculate thermal field in explict form
 
 subroutine fl_therm
+!$ACC routine(Eff_cp) seq
+!$ACC routine(Eff_conduct) seq
 use arrays
 use params
 include 'precision.inc'
@@ -25,7 +27,7 @@ subroutine fl_therm
 !  | /         / |
 !  2         2---3
 
-
+!$ACC kernels
 ! saving old temperature
 if (istress_therm.gt.0) temp0(:,:) = temp(:,:)
 
@@ -244,6 +246,7 @@ subroutine fl_therm
 end do
 !$OMP end do
 !$OMP end parallel
+!$ACC end kernels
 
 return
 end 

src/flac.f90

@@ -56,10 +56,7 @@ subroutine flac
 
 ! Adjust inertial masses or time step due to deformations
 call nvtxStartRange('dt_mass')
-if( mod(nloop,ifreq_imasses) .eq. 0 ) then
-    call dt_mass
-!$ACC update device(dt)
-end if
+if( mod(nloop,ifreq_imasses) .eq. 0 ) call dt_mass
 call nvtxEndRange()
 
 500 continue