From 43f9d043d2b8b92363964e22b0c907806011a170 Mon Sep 17 00:00:00 2001 From: Martin Diehl Date: Mon, 14 Jan 2019 12:56:46 +0100 Subject: [PATCH] removed time syncing --- src/crystallite.f90 | 286 ++------------------------------------------ src/numerics.f90 | 5 - 2 files changed, 7 insertions(+), 284 deletions(-) diff --git a/src/crystallite.f90 b/src/crystallite.f90 index 50757cb29..86348db4e 100644 --- a/src/crystallite.f90 +++ b/src/crystallite.f90 @@ -80,8 +80,6 @@ module crystallite logical, dimension(:,:), allocatable, private :: & crystallite_clearToWindForward, & !< description not available crystallite_clearToCutback, & !< description not available - crystallite_syncSubFrac, & !< description not available - crystallite_syncSubFracCompleted, & !< description not available crystallite_neighborEnforcedCutback !< description not available enum, bind(c) @@ -248,8 +246,6 @@ subroutine crystallite_init allocate(crystallite_todo(cMax,iMax,eMax), source=.false.) allocate(crystallite_converged(cMax,iMax,eMax), source=.true.) allocate(crystallite_clearToWindForward(iMax,eMax), source=.true.) - allocate(crystallite_syncSubFrac(iMax,eMax), source=.false.) - allocate(crystallite_syncSubFracCompleted(iMax,eMax), source=.false.) allocate(crystallite_clearToCutback(iMax,eMax), source=.true.) allocate(crystallite_neighborEnforcedCutback(iMax,eMax), source=.false.) allocate(crystallite_output(maxval(crystallite_Noutput), & @@ -444,8 +440,7 @@ subroutine crystallite_stressAndItsTangent(updateJaco) use numerics, only: & subStepMinCryst, & subStepSizeCryst, & - stepIncreaseCryst, & - numerics_timeSyncing + stepIncreaseCryst #ifdef DEBUG use debug, only: & debug_level, & @@ -600,235 +595,6 @@ subroutine crystallite_stressAndItsTangent(updateJaco) if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & write(6,'(a,i6)') '<< CRYST >> crystallite iteration ',NiterationCrystallite #endif - timeSyncing1: if (any(.not. crystallite_localPlasticity) .and. numerics_timeSyncing) then - - ! Time synchronization can only be used for nonlocal calculations, and only there it makes sense. - ! The idea is that in nonlocal calculations often the vast majority of the ips - ! converges in one iteration whereas a small fraction of ips has to do a lot of cutbacks. - ! Hence, we try to minimize the computational effort by just doing a lot of cutbacks - ! in the vicinity of the "bad" ips and leave the easily converged volume more or less as it is. - ! However, some synchronization of the time step has to be done at the border between "bad" ips - ! and the ones that immediately converged. - - if (any(crystallite_syncSubFrac)) then - - ! Just did a time synchronization. - ! If all synchronizers converged, then do nothing else than winding them forward. - ! If any of the synchronizers did not converge, something went completely wrong - ! and its not clear how to fix this, so all nonlocals become terminally ill. - - if (any(crystallite_syncSubFrac .and. .not. crystallite_converged(1,:,:))) then -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) then - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if (crystallite_syncSubFrac(i,e) .and. .not. crystallite_converged(1,i,e)) & - write(6,'(a,i8,1x,i2)') '<< CRYST >> time synchronization: failed at el,ip ',e,i - enddo - enddo - endif -#endif - crystallite_syncSubFrac = .false. - where(.not. crystallite_localPlasticity) - crystallite_substep = 0.0_pReal - crystallite_todo = .false. - endwhere - else - !$OMP PARALLEL DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - crystallite_clearToWindForward(i,e) = crystallite_localPlasticity(1,i,e) .or. crystallite_syncSubFrac(i,e) - crystallite_clearToCutback(i,e) = crystallite_localPlasticity(1,i,e) - enddo - enddo - !$OMP END PARALLEL DO -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a,i6)') '<< CRYST >> time synchronization: wind forward' -#endif - endif - - elseif (any(crystallite_syncSubFracCompleted)) then - - ! Just completed a time synchronization. - ! Make sure that the ips that synchronized their time step start non-converged - - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if (crystallite_syncSubFracCompleted(i,e)) crystallite_converged(1,i,e) = .false. - crystallite_syncSubFracCompleted(i,e) = .false. - crystallite_clearToWindForward(i,e) = crystallite_localPlasticity(1,i,e) - crystallite_clearToCutback(i,e) = crystallite_localPlasticity(1,i,e) .or. .not. crystallite_converged(1,i,e) - enddo - enddo -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a,i6)') '<< CRYST >> time synchronization: done, proceed with cutback' -#endif - else - - ! Normal calculation. - ! If all converged and are at the end of the time increment, then just do a final wind forward. - ! If all converged, but not all reached the end of the time increment, then we only wind - ! those forward that are still on their way, all others have to wait. - ! If some did not converge and all are still at the start of the time increment, - ! then all non-convergers force their converged neighbors to also do a cutback. - ! In case that some ips have already wound forward to an intermediate time (subfrac), - ! then all those ips that converged in the first iteration, but now have a non-converged neighbor - ! have to synchronize their time step to the same intermediate time. If such a synchronization - ! takes place, all other ips have to wait and only the synchronizers do a cutback. In the next - ! iteration those will do a wind forward while all others still wait. - - !$OMP PARALLEL DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - crystallite_clearToWindForward(i,e) = crystallite_localPlasticity(1,i,e) - crystallite_clearToCutback(i,e) = crystallite_localPlasticity(1,i,e) - enddo - enddo - !$OMP END PARALLEL DO - if (all(crystallite_localPlasticity .or. crystallite_converged)) then - if (all(crystallite_localPlasticity .or. crystallite_subStep + crystallite_subFrac >= 1.0_pReal)) then - crystallite_clearToWindForward = .true. ! final wind forward -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a,i6)') '<< CRYST >> final wind forward' -#endif - else - !$OMP PARALLEL DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - crystallite_clearToWindForward(i,e) = crystallite_localPlasticity(1,i,e) .or. crystallite_subStep(1,i,e) < 1.0_pReal - enddo - enddo - !$OMP END PARALLEL DO -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a,i6)') '<< CRYST >> wind forward' -#endif - endif - else - subFracIntermediate = maxval(crystallite_subFrac, mask=.not.crystallite_localPlasticity) - if (dNeq0(subFracIntermediate)) then - crystallite_neighborEnforcedCutback = .false. ! look for ips that require a cutback because of a nonconverged neighbor - !$OMP PARALLEL - !$OMP DO PRIVATE(neighboring_e,neighboring_i) - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if (.not. crystallite_localPlasticity(1,i,e) .and. crystallite_converged(1,i,e)) then - do n = 1_pInt,FE_NipNeighbors(FE_celltype(FE_geomtype(mesh_element(2,e)))) - neighboring_e = mesh_ipNeighborhood(1,n,i,e) - neighboring_i = mesh_ipNeighborhood(2,n,i,e) - if (neighboring_e > 0_pInt .and. neighboring_i > 0_pInt) then - if (.not. crystallite_localPlasticity(1,neighboring_i,neighboring_e) & - .and. .not. crystallite_converged(1,neighboring_i,neighboring_e)) then - crystallite_neighborEnforcedCutback(i,e) = .true. -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a12,i5,1x,i2,a,i5,1x,i2)') '<< CRYST >> ', neighboring_e,neighboring_i, & - ' enforced cutback at ',e,i -#endif - exit - endif - endif - enddo - endif - enddo - enddo - !$OMP END DO - !$OMP DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if(crystallite_neighborEnforcedCutback(i,e)) crystallite_converged(1,i,e) = .false. - enddo - enddo - !$OMP END DO - !$OMP END PARALLEL - else - crystallite_syncSubFrac = .false. ! look for ips that have to do a time synchronization because of a nonconverged neighbor - !$OMP PARALLEL - !$OMP DO PRIVATE(neighboring_e,neighboring_i) - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if (.not. crystallite_localPlasticity(1,i,e) .and. dNeq0(crystallite_subFrac(1,i,e))) then - do n = 1_pInt,FE_NipNeighbors(FE_celltype(FE_geomtype(mesh_element(2,e)))) - neighboring_e = mesh_ipNeighborhood(1,n,i,e) - neighboring_i = mesh_ipNeighborhood(2,n,i,e) - if (neighboring_e > 0_pInt .and. neighboring_i > 0_pInt) then - if (.not. crystallite_localPlasticity(1,neighboring_i,neighboring_e) & - .and. .not. crystallite_converged(1,neighboring_i,neighboring_e)) then - crystallite_syncSubFrac(i,e) = .true. -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a12,i5,1x,i2,a,i5,1x,i2)') '<< CRYST >> ',neighboring_e,neighboring_i, & - ' enforced time synchronization at ',e,i -#endif - exit - endif - endif - enddo - endif - enddo - enddo - !$OMP END DO - !$OMP DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if(crystallite_syncSubFrac(i,e)) crystallite_converged(1,i,e) = .false. - enddo - enddo - !$OMP END DO - !$OMP END PARALLEL - endif - where(.not. crystallite_localPlasticity .and. crystallite_subStep < 1.0_pReal) & - crystallite_converged = .false. - if (any(crystallite_syncSubFrac)) then ! have to do syncing now, so all wait except for the synchronizers which do a cutback - !$OMP PARALLEL DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - crystallite_clearToWindForward(i,e) = crystallite_localPlasticity(1,i,e) - crystallite_clearToCutback(i,e) = crystallite_localPlasticity(1,i,e) .or. crystallite_syncSubFrac(i,e) - enddo - enddo - !$OMP END PARALLEL DO -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a,i6)') '<< CRYST >> time synchronization: cutback' -#endif - else - !$OMP PARALLEL DO - do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if(.not. crystallite_converged(1,i,e)) crystallite_clearToCutback(i,e) = .true. - enddo - enddo - !$OMP END PARALLEL DO -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) & - write(6,'(a,i6)') '<< CRYST >> cutback' -#endif - endif - endif - endif - - ! Make sure that all cutbackers start with the same substep - - where(.not. crystallite_localPlasticity .and. .not. crystallite_converged) & - crystallite_subStep = minval(crystallite_subStep, mask=.not. crystallite_localPlasticity & - .and. .not. crystallite_converged) - - ! Those that do neither wind forward nor cutback are not to do - - !$OMP PARALLEL DO - elementLooping2: do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - if(.not. crystallite_clearToWindForward(i,e) .and. .not. crystallite_clearToCutback(i,e)) & - crystallite_todo(1,i,e) = .false. - enddo - enddo elementLooping2 - !$OMP END PARALLEL DO - - endif timeSyncing1 !$OMP PARALLEL DO PRIVATE(formerSubStep) elementLooping3: do e = FEsolving_execElem(1),FEsolving_execElem(2) @@ -856,13 +622,7 @@ subroutine crystallite_stressAndItsTangent(updateJaco) sourceState(phaseAt(c,i,e))%p(mySource)%state( :,phasememberAt(c,i,e)) enddo crystallite_subTstar0_v(1:6,c,i,e) = crystallite_Tstar_v(1:6,c,i,e) ! ...2nd PK stress - if (crystallite_syncSubFrac(i,e)) then ! if we just did a synchronization of states, then we wind forward without any further time integration - crystallite_syncSubFracCompleted(i,e) = .true. - crystallite_syncSubFrac(i,e) = .false. - crystallite_todo(c,i,e) = .false. - else - crystallite_todo(c,i,e) = .true. - endif + crystallite_todo(c,i,e) = .true. #ifdef DEBUG if (iand(debug_level(debug_crystallite),debug_levelBasic) /= 0_pInt & .and. ((e == debug_e .and. i == debug_i .and. c == debug_g) & @@ -878,11 +638,7 @@ subroutine crystallite_stressAndItsTangent(updateJaco) ! --- cutback --- elseif (.not. crystallite_converged(c,i,e) .and. crystallite_clearToCutback(i,e)) then - if (crystallite_syncSubFrac(i,e)) then ! synchronize time - crystallite_subStep(c,i,e) = subFracIntermediate - else - crystallite_subStep(c,i,e) = subStepSizeCryst * crystallite_subStep(c,i,e) ! cut step in half and restore... - endif + crystallite_subStep(c,i,e) = subStepSizeCryst * crystallite_subStep(c,i,e) ! cut step in half and restore... crystallite_Fp(1:3,1:3,c,i,e) = crystallite_subFp0(1:3,1:3,c,i,e) ! ...plastic def grad crystallite_invFp(1:3,1:3,c,i,e) = math_inv33(crystallite_Fp(1:3,1:3,c,i,e)) crystallite_Fi(1:3,1:3,c,i,e) = crystallite_subFi0(1:3,1:3,c,i,e) ! ...intermediate def grad @@ -933,29 +689,6 @@ subroutine crystallite_stressAndItsTangent(updateJaco) enddo elementLooping3 !$OMP END PARALLEL DO - timeSyncing2: if(numerics_timeSyncing) then - if (any(.not. crystallite_localPlasticity .and. .not. crystallite_todo .and. .not. crystallite_converged & - .and. crystallite_subStep <= subStepMinCryst)) then ! no way of rescuing a nonlocal ip that violated the lower time step limit, ... -#ifdef DEBUG - if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) then - elementLooping4: do e = FEsolving_execElem(1),FEsolving_execElem(2) - do i = FEsolving_execIP(1,e),FEsolving_execIP(2,e) - do c = 1,homogenization_Ngrains(mesh_element(3,e)) - if (.not. crystallite_localPlasticity(c,i,e) .and. .not. crystallite_todo(c,i,e) & - .and. .not. crystallite_converged(c,i,e) .and. crystallite_subStep(c,i,e) <= subStepMinCryst) & - write(6,'(a,i8,1x,i2,1x,i3)') '<< CRYST >> nonlocal violated minimum subStep at el ip ipc ',e,i,c - enddo - enddo - enddo elementLooping4 - endif -#endif - where(.not. crystallite_localPlasticity) - crystallite_todo = .false. ! ... so let all nonlocal ips die peacefully - crystallite_subStep = 0.0_pReal - endwhere - endif - endif timeSyncing2 - #ifdef DEBUG if (iand(debug_level(debug_crystallite),debug_levelExtensive) /= 0_pInt) then write(6,'(/,a,f8.5)') '<< CRYST >> min(subStep) ',minval(crystallite_subStep) @@ -2268,8 +2001,6 @@ subroutine integrateStateEuler() debug_levelExtensive, & debug_levelSelective #endif - use numerics, only: & - numerics_timeSyncing use FEsolving, only: & FEsolving_execElem, & FEsolving_execIP @@ -2340,7 +2071,7 @@ eIter = FEsolving_execElem(1:2) NaN = NaN .or. any(IEEE_is_NaN(sourceState(p)%p(mySource)%dotState(:,c))) enddo if (NaN) then ! NaN occured in any dotState - if (.not. crystallite_localPlasticity(g,i,e) .and. .not. numerics_timeSyncing) then ! if broken non-local... + if (.not. crystallite_localPlasticity(g,i,e)) then ! if broken non-local... !$OMP CRITICAL (checkTodo) crystallite_todo = crystallite_todo .and. crystallite_localPlasticity ! ...all non-locals skipped !$OMP END CRITICAL (checkTodo) @@ -2397,8 +2128,7 @@ eIter = FEsolving_execElem(1:2) if (crystallite_todo(g,i,e) .and. .not. crystallite_converged(g,i,e)) then crystallite_todo(g,i,e) = stateJump(g,i,e) !$OMP FLUSH(crystallite_todo) - if (.not. crystallite_todo(g,i,e) .and. .not. crystallite_localPlasticity(g,i,e) & ! if broken non-local... - .and. .not. numerics_timeSyncing) then + if (.not. crystallite_todo(g,i,e) .and. .not. crystallite_localPlasticity(g,i,e)) then ! if broken non-local... !$OMP CRITICAL (checkTodo) crystallite_todo = crystallite_todo .and. crystallite_localPlasticity ! ...all non-locals skipped !$OMP END CRITICAL (checkTodo) @@ -2431,8 +2161,7 @@ eIter = FEsolving_execElem(1:2) if (crystallite_todo(g,i,e) .and. .not. crystallite_converged(g,i,e)) then crystallite_todo(g,i,e) = integrateStress(g,i,e) !$OMP FLUSH(crystallite_todo) - if (.not. crystallite_todo(g,i,e) .and. .not. crystallite_localPlasticity(g,i,e) & ! if broken non-local... - .and. .not. numerics_timeSyncing) then + if (.not. crystallite_todo(g,i,e) .and. .not. crystallite_localPlasticity(g,i,e)) then ! if broken non-local... !$OMP CRITICAL (checkTodo) crystallite_todo = crystallite_todo .and. crystallite_localPlasticity ! ...all non-locals skipped !$OMP END CRITICAL (checkTodo) @@ -2456,8 +2185,7 @@ eIter = FEsolving_execElem(1:2) ! --- CHECK NON-LOCAL CONVERGENCE --- if (.not. singleRun) then ! if not requesting Integration of just a single IP - if (any(.not. crystallite_converged .and. .not. crystallite_localPlasticity) & ! any non-local not yet converged (or broken)... - .and. .not. numerics_timeSyncing) & + if (any(.not. crystallite_converged .and. .not. crystallite_localPlasticity) ) & ! any non-local not yet converged (or broken)... crystallite_converged = crystallite_converged .and. crystallite_localPlasticity ! ...restart all non-local as not converged endif diff --git a/src/numerics.f90 b/src/numerics.f90 index e4ceec622..9e585dda7 100644 --- a/src/numerics.f90 +++ b/src/numerics.f90 @@ -276,8 +276,6 @@ subroutine numerics_init numerics_integrator = IO_intValue(line,chunkPos,2_pInt) case ('usepingpong') usepingpong = IO_intValue(line,chunkPos,2_pInt) > 0_pInt - case ('timesyncing') - numerics_timeSyncing = IO_intValue(line,chunkPos,2_pInt) > 0_pInt case ('unitlength') numerics_unitlength = IO_floatValue(line,chunkPos,2_pInt) @@ -454,8 +452,6 @@ subroutine numerics_init end select #endif - numerics_timeSyncing = numerics_timeSyncing .and. all(numerics_integrator==2_pInt) ! timeSyncing only allowed for explicit Euler integrator - !-------------------------------------------------------------------------------------------------- ! writing parameters to output write(6,'(a24,1x,es8.1)') ' relevantStrain: ',relevantStrain @@ -476,7 +472,6 @@ subroutine numerics_init write(6,'(a24,1x,es8.1)') ' rTol_crystalliteStress: ',rTol_crystalliteStress write(6,'(a24,1x,es8.1)') ' aTol_crystalliteStress: ',aTol_crystalliteStress write(6,'(a24,2(1x,i8))') ' integrator: ',numerics_integrator - write(6,'(a24,1x,L8)') ' timeSyncing: ',numerics_timeSyncing write(6,'(a24,1x,L8)') ' use ping pong scheme: ',usepingpong write(6,'(a24,1x,es8.1,/)')' unitlength: ',numerics_unitlength