mirror of
https://github.com/sfilippone/psblas3.git
synced 2026-10-06 22:55:08 +00:00
Fixed OpenACC test program mixed with CUDA
This commit is contained in:
sfilippone
+6
-20
@@ -12,9 +12,9 @@ EXEDIR=./runs
|
||||
PSBLAS_LIB= -L$(LIBDIR) -L$(PSBLIBDIR) -lpsb_openacc -lpsb_ext -lpsb_util -lpsb_linsolve -lpsb_prec -lpsb_base
|
||||
LDLIBS=$(PSBGPULDLIBS)
|
||||
|
||||
FINCLUDES=$(FMFLAG)$(MODDIR) $(FMFLAG)$(INCDIR) $(FMFLAG). $(FMFLAG)$(PSBMODDIR) $(FMFLAG)$(PSBINCDIR) $(LIBRSB_DEFINES)
|
||||
FINCLUDES= $(FMFLAG). $(FMFLAG)$(MODDIR) $(FMFLAG)$(INCDIR) $(FMFLAG)$(PSBMODDIR) $(FMFLAG)$(PSBINCDIR) $(LIBRSB_DEFINES)
|
||||
|
||||
FFLAGS=-O3 -march=native $(FCOPENACC) -DOPENACC -DCUDA
|
||||
FFLAGS=-O3 -march=native $(FCOPENACC) -DPSB_OPENACC -DPSB_HAVE_CUDA
|
||||
CFLAGS=-O3 -march=native
|
||||
|
||||
VTC=vectoacc.o
|
||||
@@ -22,9 +22,8 @@ DVT=datavect.o
|
||||
CSRC=timers.c
|
||||
|
||||
OBJS=$(SRCS:.F90=.o) $(CSRC:.c=.o)
|
||||
OADIR=oaobjs
|
||||
|
||||
all: dir psb_d_oacc_pde3d dpdegenmv
|
||||
all: dir dpdegenmv
|
||||
|
||||
#$(OBJS)
|
||||
# $(FC) $(FFLAGS) $(OBJS) -o datavect $(FINCLUDES) $(PSBLAS_LIB) $(LDLIBS)
|
||||
@@ -43,25 +42,12 @@ dir:
|
||||
%.o: %.c
|
||||
$(CC) $(CFLAGS) $(FINCLUDES) -c $< -o $@
|
||||
|
||||
cpyoa:
|
||||
@if test ! -d $(OADIR); then mkdir $(OADIR); fi
|
||||
/bin/cp $(TOPDIR)/openacc/*.o $(OADIR)
|
||||
/bin/cp $(TOPDIR)/openacc/impl/*.o $(OADIR)
|
||||
|
||||
|
||||
psb_d_oacc_pde3d:
|
||||
mpifort -fallow-argument-mismatch -frecursive -g -O3 -frecursive -I../../modules/ -I. -DOPENACC -DHAVE_LAPACK -DHAVE_FLUSH_STMT -DLPK8 -DIPK4 -DMPI_MOD -c psb_d_oacc_pde3d.F90 -o psb_d_oacc_pde3d.o
|
||||
$(FLINK) -fopenacc -DOPENACC psb_d_oacc_pde3d.o -o psb_d_oacc_pde3d $(PSBLAS_LIB) $(LDLIBS)
|
||||
/bin/mv psb_d_oacc_pde3d $(EXEDIR)
|
||||
|
||||
dpdegenmv: dpdegenmv.o cpyoa
|
||||
$(FLINK) $(FCOPENACC) dpdegenmv.o -o dpdegenmv $(OADIR)/*.o $(PSBLAS_LIB) $(LDLIBS)
|
||||
dpdegenmv.o: psb_d_pde3d_mod.o
|
||||
dpdegenmv: dpdegenmv.o psb_d_pde3d_mod.o
|
||||
$(FLINK) $(FCOPENACC) dpdegenmv.o psb_d_pde3d_mod.o -o dpdegenmv -fuse-ld=mold $(PSBLAS_LIB) $(LDLIBS)
|
||||
/bin/mv dpdegenmv $(EXEDIR)
|
||||
|
||||
dpmix: dpmix.o cpyoa
|
||||
$(FLINK) $(FCOPENACC) dpmix.o -o dpmix $(OADIR)/*.o $(PSBLAS_LIB) $(LDLIBS)
|
||||
/bin/mv dpmix $(EXEDIR)
|
||||
|
||||
clean:
|
||||
/bin/rm -fr *.o *.mod $(EXEDIR)/*
|
||||
|
||||
|
||||
+82
-549
@@ -37,531 +37,17 @@
|
||||
! the main PSBLAS library.
|
||||
!
|
||||
!
|
||||
module psb_d_pde3d_mod
|
||||
|
||||
|
||||
use psb_base_mod, only : psb_dpk_, psb_ipk_, psb_lpk_, psb_desc_type,&
|
||||
& psb_dspmat_type, psb_d_vect_type, dzero,&
|
||||
& psb_d_base_sparse_mat, psb_d_base_vect_type, &
|
||||
& psb_i_base_vect_type, psb_l_base_vect_type
|
||||
|
||||
interface
|
||||
function d_func_3d(x,y,z) result(val)
|
||||
import :: psb_dpk_
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
real(psb_dpk_) :: val
|
||||
end function d_func_3d
|
||||
end interface
|
||||
|
||||
interface psb_gen_pde3d
|
||||
module procedure psb_d_gen_pde3d
|
||||
end interface psb_gen_pde3d
|
||||
|
||||
contains
|
||||
|
||||
function d_null_func_3d(x,y,z) result(val)
|
||||
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
real(psb_dpk_) :: val
|
||||
|
||||
val = dzero
|
||||
|
||||
end function d_null_func_3d
|
||||
!
|
||||
! functions parametrizing the differential equation
|
||||
!
|
||||
|
||||
!
|
||||
! Note: b1, b2 and b3 are the coefficients of the first
|
||||
! derivative of the unknown function. The default
|
||||
! we apply here is to have them zero, so that the resulting
|
||||
! matrix is symmetric/hermitian and suitable for
|
||||
! testing with CG and FCG.
|
||||
! When testing methods for non-hermitian matrices you can
|
||||
! change the B1/B2/B3 functions to e.g. done/sqrt((3*done))
|
||||
!
|
||||
function b1(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: b1
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
b1=done/sqrt((3*done))
|
||||
end function b1
|
||||
function b2(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: b2
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
b2=done/sqrt((3*done))
|
||||
end function b2
|
||||
function b3(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: b3
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
b3=done/sqrt((3*done))
|
||||
end function b3
|
||||
function c(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: c
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
c=dzero
|
||||
end function c
|
||||
function a1(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: a1
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
a1=done/80
|
||||
end function a1
|
||||
function a2(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: a2
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
a2=done/80
|
||||
end function a2
|
||||
function a3(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: a3
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
a3=done/80
|
||||
end function a3
|
||||
function g(x,y,z)
|
||||
use psb_base_mod, only : psb_dpk_, done, dzero
|
||||
implicit none
|
||||
real(psb_dpk_) :: g
|
||||
real(psb_dpk_), intent(in) :: x,y,z
|
||||
g = dzero
|
||||
if (x == done) then
|
||||
g = done
|
||||
else if (x == dzero) then
|
||||
g = exp(y**2-z**2)
|
||||
end if
|
||||
end function g
|
||||
|
||||
|
||||
!
|
||||
! subroutine to allocate and fill in the coefficient matrix and
|
||||
! the rhs.
|
||||
!
|
||||
subroutine psb_d_gen_pde3d(ctxt,idim,a,bv,xv,desc_a,afmt,info,&
|
||||
& f,amold,vmold,imold,partition,nrl,iv,tnd)
|
||||
use psb_base_mod
|
||||
use psb_util_mod
|
||||
!
|
||||
! Discretizes the partial differential equation
|
||||
!
|
||||
! a1 dd(u) a2 dd(u) a3 dd(u) b1 d(u) b2 d(u) b3 d(u)
|
||||
! - ------ - ------ - ------ + ----- + ------ + ------ + c u = f
|
||||
! dxdx dydy dzdz dx dy dz
|
||||
!
|
||||
! with Dirichlet boundary conditions
|
||||
! u = g
|
||||
!
|
||||
! on the unit cube 0<=x,y,z<=1.
|
||||
!
|
||||
!
|
||||
! Note that if b1=b2=b3=c=0., the PDE is the Laplace equation.
|
||||
!
|
||||
implicit none
|
||||
integer(psb_ipk_) :: idim
|
||||
type(psb_dspmat_type) :: a
|
||||
type(psb_d_vect_type) :: xv,bv
|
||||
type(psb_desc_type) :: desc_a
|
||||
type(psb_ctxt_type) :: ctxt
|
||||
integer(psb_ipk_) :: info
|
||||
character(len=*) :: afmt
|
||||
procedure(d_func_3d), optional :: f
|
||||
class(psb_d_base_sparse_mat), optional :: amold
|
||||
class(psb_d_base_vect_type), optional :: vmold
|
||||
class(psb_i_base_vect_type), optional :: imold
|
||||
integer(psb_ipk_), optional :: partition, nrl,iv(:)
|
||||
logical, optional :: tnd
|
||||
! Local variables.
|
||||
|
||||
integer(psb_ipk_), parameter :: nb=20
|
||||
type(psb_d_csc_sparse_mat) :: acsc
|
||||
type(psb_d_coo_sparse_mat) :: acoo
|
||||
type(psb_d_csr_sparse_mat) :: acsr
|
||||
real(psb_dpk_) :: zt(nb),x,y,z
|
||||
integer(psb_ipk_) :: nnz,nr,nlr,i,j,ii,ib,k, partition_
|
||||
integer(psb_lpk_) :: m,n,glob_row,nt
|
||||
integer(psb_ipk_) :: ix,iy,iz,ia,indx_owner
|
||||
! For 3D partition
|
||||
! Note: integer control variables going directly into an MPI call
|
||||
! must be 4 bytes, i.e. psb_mpk_
|
||||
integer(psb_mpk_) :: npdims(3), npp, minfo
|
||||
integer(psb_ipk_) :: npx,npy,npz, iamx,iamy,iamz,mynx,myny,mynz
|
||||
integer(psb_ipk_), allocatable :: bndx(:),bndy(:),bndz(:)
|
||||
! Process grid
|
||||
integer(psb_ipk_) :: np, iam
|
||||
integer(psb_ipk_) :: icoeff
|
||||
integer(psb_lpk_), allocatable :: irow(:),icol(:),myidx(:)
|
||||
real(psb_dpk_), allocatable :: val(:)
|
||||
! deltah dimension of each grid cell
|
||||
! deltat discretization time
|
||||
real(psb_dpk_) :: deltah, sqdeltah, deltah2
|
||||
real(psb_dpk_), parameter :: rhs=dzero,one=done,zero=dzero
|
||||
real(psb_dpk_) :: t0, t1, t2, t3, tasb, talc, ttot, tgen, tcdasb
|
||||
integer(psb_ipk_) :: err_act
|
||||
procedure(d_func_3d), pointer :: f_
|
||||
logical :: tnd_
|
||||
character(len=20) :: name, ch_err,tmpfmt
|
||||
|
||||
info = psb_success_
|
||||
name = 'create_matrix'
|
||||
call psb_erractionsave(err_act)
|
||||
|
||||
call psb_info(ctxt, iam, np)
|
||||
|
||||
|
||||
if (present(f)) then
|
||||
f_ => f
|
||||
else
|
||||
f_ => d_null_func_3d
|
||||
end if
|
||||
|
||||
deltah = done/(idim+2)
|
||||
sqdeltah = deltah*deltah
|
||||
deltah2 = (2*done)* deltah
|
||||
|
||||
if (present(partition)) then
|
||||
if ((1<= partition).and.(partition <= 3)) then
|
||||
partition_ = partition
|
||||
else
|
||||
write(*,*) 'Invalid partition choice ',partition,' defaulting to 3'
|
||||
partition_ = 3
|
||||
end if
|
||||
else
|
||||
partition_ = 3
|
||||
end if
|
||||
|
||||
! initialize array descriptor and sparse matrix storage. provide an
|
||||
! estimate of the number of non zeroes
|
||||
|
||||
m = (1_psb_lpk_*idim)*idim*idim
|
||||
n = m
|
||||
nnz = ((n*7)/(np))
|
||||
if(iam == psb_root_) write(psb_out_unit,'("Generating Matrix (size=",i0,")...")')n
|
||||
t0 = psb_wtime()
|
||||
select case(partition_)
|
||||
case(1)
|
||||
! A BLOCK partition
|
||||
if (present(nrl)) then
|
||||
nr = nrl
|
||||
else
|
||||
!
|
||||
! Using a simple BLOCK distribution.
|
||||
!
|
||||
nt = (m+np-1)/np
|
||||
nr = max(0,min(nt,m-(iam*nt)))
|
||||
end if
|
||||
|
||||
nt = nr
|
||||
call psb_sum(ctxt,nt)
|
||||
if (nt /= m) then
|
||||
write(psb_err_unit,*) iam, 'Initialization error ',nr,nt,m
|
||||
info = -1
|
||||
call psb_barrier(ctxt)
|
||||
call psb_abort(ctxt)
|
||||
return
|
||||
end if
|
||||
|
||||
!
|
||||
! First example of use of CDALL: specify for each process a number of
|
||||
! contiguous rows
|
||||
!
|
||||
call psb_cdall(ctxt,desc_a,info,nl=nr)
|
||||
myidx = desc_a%get_global_indices()
|
||||
nlr = size(myidx)
|
||||
|
||||
case(2)
|
||||
! A partition defined by the user through IV
|
||||
|
||||
if (present(iv)) then
|
||||
if (size(iv) /= m) then
|
||||
write(psb_err_unit,*) iam, 'Initialization error: wrong IV size',size(iv),m
|
||||
info = -1
|
||||
call psb_barrier(ctxt)
|
||||
call psb_abort(ctxt)
|
||||
return
|
||||
end if
|
||||
else
|
||||
write(psb_err_unit,*) iam, 'Initialization error: IV not present'
|
||||
info = -1
|
||||
call psb_barrier(ctxt)
|
||||
call psb_abort(ctxt)
|
||||
return
|
||||
end if
|
||||
|
||||
!
|
||||
! Second example of use of CDALL: specify for each row the
|
||||
! process that owns it
|
||||
!
|
||||
call psb_cdall(ctxt,desc_a,info,vg=iv)
|
||||
myidx = desc_a%get_global_indices()
|
||||
nlr = size(myidx)
|
||||
|
||||
case(3)
|
||||
! A 3-dimensional partition
|
||||
|
||||
! A nifty MPI function will split the process list
|
||||
npdims = 0
|
||||
call mpi_dims_create(np,3,npdims,info)
|
||||
npx = npdims(1)
|
||||
npy = npdims(2)
|
||||
npz = npdims(3)
|
||||
|
||||
allocate(bndx(0:npx),bndy(0:npy),bndz(0:npz))
|
||||
! We can reuse idx2ijk for process indices as well.
|
||||
call idx2ijk(iamx,iamy,iamz,iam,npx,npy,npz,base=0)
|
||||
! Now let's split the 3D cube in hexahedra
|
||||
call dist1Didx(bndx,idim,npx)
|
||||
mynx = bndx(iamx+1)-bndx(iamx)
|
||||
call dist1Didx(bndy,idim,npy)
|
||||
myny = bndy(iamy+1)-bndy(iamy)
|
||||
call dist1Didx(bndz,idim,npz)
|
||||
mynz = bndz(iamz+1)-bndz(iamz)
|
||||
|
||||
! How many indices do I own?
|
||||
nlr = mynx*myny*mynz
|
||||
allocate(myidx(nlr))
|
||||
! Now, let's generate the list of indices I own
|
||||
nr = 0
|
||||
do i=bndx(iamx),bndx(iamx+1)-1
|
||||
do j=bndy(iamy),bndy(iamy+1)-1
|
||||
do k=bndz(iamz),bndz(iamz+1)-1
|
||||
nr = nr + 1
|
||||
call ijk2idx(myidx(nr),i,j,k,idim,idim,idim)
|
||||
end do
|
||||
end do
|
||||
end do
|
||||
if (nr /= nlr) then
|
||||
write(psb_err_unit,*) iam,iamx,iamy,iamz, 'Initialization error: NR vs NLR ',&
|
||||
& nr,nlr,mynx,myny,mynz
|
||||
info = -1
|
||||
call psb_barrier(ctxt)
|
||||
call psb_abort(ctxt)
|
||||
end if
|
||||
|
||||
!
|
||||
! Third example of use of CDALL: specify for each process
|
||||
! the set of global indices it owns.
|
||||
!
|
||||
call psb_cdall(ctxt,desc_a,info,vl=myidx)
|
||||
|
||||
case default
|
||||
write(psb_err_unit,*) iam, 'Initialization error: should not get here'
|
||||
info = -1
|
||||
call psb_barrier(ctxt)
|
||||
call psb_abort(ctxt)
|
||||
return
|
||||
end select
|
||||
|
||||
|
||||
if (info == psb_success_) call psb_spall(a,desc_a,info,nnz=nnz,&
|
||||
& dupl=psb_dupl_err_)
|
||||
! define rhs from boundary conditions; also build initial guess
|
||||
if (info == psb_success_) call psb_geall(xv,desc_a,info)
|
||||
if (info == psb_success_) call psb_geall(bv,desc_a,info)
|
||||
|
||||
call psb_barrier(ctxt)
|
||||
talc = psb_wtime()-t0
|
||||
|
||||
if (info /= psb_success_) then
|
||||
info=psb_err_from_subroutine_
|
||||
ch_err='allocation rout.'
|
||||
call psb_errpush(info,name,a_err=ch_err)
|
||||
goto 9999
|
||||
end if
|
||||
|
||||
! we build an auxiliary matrix consisting of one row at a
|
||||
! time; just a small matrix. might be extended to generate
|
||||
! a bunch of rows per call.
|
||||
!
|
||||
allocate(val(20*nb),irow(20*nb),&
|
||||
&icol(20*nb),stat=info)
|
||||
if (info /= psb_success_ ) then
|
||||
info=psb_err_alloc_dealloc_
|
||||
call psb_errpush(info,name)
|
||||
goto 9999
|
||||
endif
|
||||
|
||||
|
||||
! loop over rows belonging to current process in a block
|
||||
! distribution.
|
||||
|
||||
call psb_barrier(ctxt)
|
||||
t1 = psb_wtime()
|
||||
do ii=1, nlr,nb
|
||||
ib = min(nb,nlr-ii+1)
|
||||
icoeff = 1
|
||||
do k=1,ib
|
||||
i=ii+k-1
|
||||
! local matrix pointer
|
||||
glob_row=myidx(i)
|
||||
! compute gridpoint coordinates
|
||||
call idx2ijk(ix,iy,iz,glob_row,idim,idim,idim)
|
||||
! x, y, z coordinates
|
||||
x = (ix-1)*deltah
|
||||
y = (iy-1)*deltah
|
||||
z = (iz-1)*deltah
|
||||
zt(k) = f_(x,y,z)
|
||||
! internal point: build discretization
|
||||
!
|
||||
! term depending on (x-1,y,z)
|
||||
!
|
||||
val(icoeff) = -a1(x,y,z)/sqdeltah-b1(x,y,z)/deltah2
|
||||
if (ix == 1) then
|
||||
zt(k) = g(dzero,y,z)*(-val(icoeff)) + zt(k)
|
||||
else
|
||||
call ijk2idx(icol(icoeff),ix-1,iy,iz,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
endif
|
||||
! term depending on (x,y-1,z)
|
||||
val(icoeff) = -a2(x,y,z)/sqdeltah-b2(x,y,z)/deltah2
|
||||
if (iy == 1) then
|
||||
zt(k) = g(x,dzero,z)*(-val(icoeff)) + zt(k)
|
||||
else
|
||||
call ijk2idx(icol(icoeff),ix,iy-1,iz,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
endif
|
||||
! term depending on (x,y,z-1)
|
||||
val(icoeff)=-a3(x,y,z)/sqdeltah-b3(x,y,z)/deltah2
|
||||
if (iz == 1) then
|
||||
zt(k) = g(x,y,dzero)*(-val(icoeff)) + zt(k)
|
||||
else
|
||||
call ijk2idx(icol(icoeff),ix,iy,iz-1,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
endif
|
||||
|
||||
! term depending on (x,y,z)
|
||||
val(icoeff)=(2*done)*(a1(x,y,z)+a2(x,y,z)+a3(x,y,z))/sqdeltah &
|
||||
& + c(x,y,z)
|
||||
call ijk2idx(icol(icoeff),ix,iy,iz,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
! term depending on (x,y,z+1)
|
||||
val(icoeff)=-a3(x,y,z)/sqdeltah+b3(x,y,z)/deltah2
|
||||
if (iz == idim) then
|
||||
zt(k) = g(x,y,done)*(-val(icoeff)) + zt(k)
|
||||
else
|
||||
call ijk2idx(icol(icoeff),ix,iy,iz+1,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
endif
|
||||
! term depending on (x,y+1,z)
|
||||
val(icoeff)=-a2(x,y,z)/sqdeltah+b2(x,y,z)/deltah2
|
||||
if (iy == idim) then
|
||||
zt(k) = g(x,done,z)*(-val(icoeff)) + zt(k)
|
||||
else
|
||||
call ijk2idx(icol(icoeff),ix,iy+1,iz,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
endif
|
||||
! term depending on (x+1,y,z)
|
||||
val(icoeff)=-a1(x,y,z)/sqdeltah+b1(x,y,z)/deltah2
|
||||
if (ix==idim) then
|
||||
zt(k) = g(done,y,z)*(-val(icoeff)) + zt(k)
|
||||
else
|
||||
call ijk2idx(icol(icoeff),ix+1,iy,iz,idim,idim,idim)
|
||||
irow(icoeff) = glob_row
|
||||
icoeff = icoeff+1
|
||||
endif
|
||||
|
||||
end do
|
||||
call psb_spins(icoeff-1,irow,icol,val,a,desc_a,info)
|
||||
if(info /= psb_success_) exit
|
||||
call psb_geins(ib,myidx(ii:ii+ib-1),zt(1:ib),bv,desc_a,info)
|
||||
if(info /= psb_success_) exit
|
||||
zt(:)=dzero
|
||||
call psb_geins(ib,myidx(ii:ii+ib-1),zt(1:ib),xv,desc_a,info)
|
||||
if(info /= psb_success_) exit
|
||||
end do
|
||||
|
||||
tgen = psb_wtime()-t1
|
||||
if(info /= psb_success_) then
|
||||
info=psb_err_from_subroutine_
|
||||
ch_err='insert rout.'
|
||||
call psb_errpush(info,name,a_err=ch_err)
|
||||
goto 9999
|
||||
end if
|
||||
|
||||
deallocate(val,irow,icol)
|
||||
|
||||
call psb_barrier(ctxt)
|
||||
t1 = psb_wtime()
|
||||
call psb_cdasb(desc_a,info,mold=imold)
|
||||
tcdasb = psb_wtime()-t1
|
||||
call psb_barrier(ctxt)
|
||||
t1 = psb_wtime()
|
||||
if (info == psb_success_) then
|
||||
if (present(amold)) then
|
||||
call psb_spasb(a,desc_a,info,mold=amold,bld_and=tnd)
|
||||
else
|
||||
call psb_spasb(a,desc_a,info,afmt=afmt,bld_and=tnd)
|
||||
end if
|
||||
end if
|
||||
call psb_barrier(ctxt)
|
||||
if(info /= psb_success_) then
|
||||
info=psb_err_from_subroutine_
|
||||
ch_err='asb rout.'
|
||||
call psb_errpush(info,name,a_err=ch_err)
|
||||
goto 9999
|
||||
end if
|
||||
if (info == psb_success_) call psb_geasb(xv,desc_a,info,mold=vmold)
|
||||
if (info == psb_success_) call psb_geasb(bv,desc_a,info,mold=vmold)
|
||||
if(info /= psb_success_) then
|
||||
info=psb_err_from_subroutine_
|
||||
ch_err='asb rout.'
|
||||
call psb_errpush(info,name,a_err=ch_err)
|
||||
goto 9999
|
||||
end if
|
||||
tasb = psb_wtime()-t1
|
||||
call psb_barrier(ctxt)
|
||||
ttot = psb_wtime() - t0
|
||||
|
||||
call psb_amx(ctxt,talc)
|
||||
call psb_amx(ctxt,tgen)
|
||||
call psb_amx(ctxt,tasb)
|
||||
call psb_amx(ctxt,ttot)
|
||||
if(iam == psb_root_) then
|
||||
tmpfmt = a%get_fmt()
|
||||
write(psb_out_unit,'("The matrix has been generated and assembled in ",a3," format.")')&
|
||||
& tmpfmt
|
||||
write(psb_out_unit,'("-allocation time : ",es12.5)') talc
|
||||
write(psb_out_unit,'("-coeff. gen. time : ",es12.5)') tgen
|
||||
write(psb_out_unit,'("-desc asbly time : ",es12.5)') tcdasb
|
||||
write(psb_out_unit,'("- mat asbly time : ",es12.5)') tasb
|
||||
write(psb_out_unit,'("-total time : ",es12.5)') ttot
|
||||
|
||||
end if
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(ctxt,err_act)
|
||||
|
||||
return
|
||||
end subroutine psb_d_gen_pde3d
|
||||
|
||||
|
||||
end module psb_d_pde3d_mod
|
||||
|
||||
|
||||
program pdgenmv
|
||||
use psb_base_mod
|
||||
use psb_util_mod
|
||||
use psb_ext_mod
|
||||
#ifdef OPENACC
|
||||
use psb_d_pde3d_mod
|
||||
#ifdef PSB_OPENACC
|
||||
use psb_oacc_mod
|
||||
#endif
|
||||
use psb_d_pde3d_mod
|
||||
#ifdef PSB_HAVE_CUDA
|
||||
use psb_cuda_mod
|
||||
#endif
|
||||
implicit none
|
||||
|
||||
! input parameters
|
||||
@@ -580,10 +66,6 @@ program pdgenmv
|
||||
type(psb_desc_type) :: desc_a
|
||||
! dense matrices
|
||||
type(psb_d_vect_type), target :: xv, bv, xg, bg
|
||||
#ifdef OPENACC
|
||||
type(psb_d_vect_oacc) :: vmold
|
||||
type(psb_i_vect_oacc) :: imold
|
||||
#endif
|
||||
real(psb_dpk_), allocatable :: x1(:), x2(:), x0(:)
|
||||
! blacs parameters
|
||||
type(psb_ctxt_type) :: ctxt
|
||||
@@ -599,12 +81,29 @@ program pdgenmv
|
||||
type(psb_d_hll_sparse_mat), target :: ahll
|
||||
type(psb_d_dia_sparse_mat), target :: adia
|
||||
type(psb_d_hdia_sparse_mat), target :: ahdia
|
||||
#ifdef OPENACC
|
||||
type(psb_d_oacc_ell_sparse_mat), target :: aelg
|
||||
type(psb_d_oacc_csr_sparse_mat), target :: acsrg
|
||||
type(psb_d_oacc_hll_sparse_mat), target :: ahlg
|
||||
type(psb_d_base_vect_type), target :: dvect
|
||||
type(psb_i_base_vect_type), target :: ivect
|
||||
class(psb_d_base_sparse_mat), pointer :: acmold => acsr
|
||||
class(psb_d_base_vect_type), pointer :: vmold => dvect
|
||||
class(psb_i_base_vect_type), pointer :: imold => ivect
|
||||
class(psb_d_base_vect_type), pointer :: vgmold => dvect
|
||||
class(psb_i_base_vect_type), pointer :: igmold => ivect
|
||||
#if defined(PSB_HAVE_CUDA)
|
||||
type(psb_d_cuda_hlg_sparse_mat), target :: ahlg
|
||||
type(psb_d_cuda_hdiag_sparse_mat), target :: ahdiag
|
||||
type(psb_d_cuda_csrg_sparse_mat), target :: acsrg
|
||||
type(psb_d_cuda_elg_sparse_mat), target :: aelg
|
||||
type(psb_d_vect_cuda), target :: dvgpu
|
||||
type(psb_i_vect_cuda), target :: ivgpu
|
||||
#endif
|
||||
class(psb_d_base_sparse_mat), pointer :: agmold, acmold
|
||||
#if defined(PSB_OPENACC)
|
||||
type(psb_d_oacc_hll_sparse_mat), target :: ahlo
|
||||
type(psb_d_oacc_csr_sparse_mat), target :: acsro
|
||||
type(psb_d_oacc_ell_sparse_mat), target :: aelo
|
||||
type(psb_d_vect_oacc), target :: dvoac
|
||||
type(psb_i_vect_oacc), target :: ivoac
|
||||
#endif
|
||||
class(psb_d_base_sparse_mat), pointer :: agmold
|
||||
! other variables
|
||||
logical, parameter :: dump=.false.
|
||||
integer(psb_ipk_) :: info, i, j, nr, nrg
|
||||
@@ -618,9 +117,12 @@ program pdgenmv
|
||||
call psb_init(ctxt)
|
||||
call psb_info(ctxt,iam,np)
|
||||
|
||||
#ifdef OPENACC
|
||||
#ifdef PSB_OPENACC
|
||||
call psb_oacc_init(ctxt)
|
||||
#endif
|
||||
#ifdef PSB_HAVE_CUDA
|
||||
call psb_cuda_init(ctxt)
|
||||
#endif
|
||||
|
||||
if (iam < 0) then
|
||||
! This should not happen, but just in case
|
||||
@@ -636,9 +138,13 @@ program pdgenmv
|
||||
write(*,*) 'Welcome to PSBLAS version: ',psb_version_string_
|
||||
write(*,*) 'This is the ',trim(name),' sample program'
|
||||
end if
|
||||
#ifdef OPENACC
|
||||
#ifdef PSB_OPENACC
|
||||
!!$ write(*,*) 'Process ',iam,' running on device: ', psb_oacc_getDevice(),' out of', psb_oacc_getDeviceCount()
|
||||
!!$ write(*,*) 'Process ',iam,' device ', psb_oacc_getDevice(),' is a: ', trim(psb_oacc_DeviceName())
|
||||
#endif
|
||||
#if defined(PSB_HAVE_CUDA)
|
||||
write(*,*) 'Process ',iam,' running on device: ', psb_cuda_getDevice(),' out of', psb_cuda_getDeviceCount()
|
||||
write(*,*) 'Process ',iam,' device ', psb_cuda_getDevice(),' is a: ', trim(psb_cuda_DeviceName())
|
||||
#endif
|
||||
!
|
||||
! get parameters
|
||||
@@ -650,7 +156,7 @@ program pdgenmv
|
||||
!
|
||||
call psb_barrier(ctxt)
|
||||
t1 = psb_wtime()
|
||||
call psb_gen_pde3d(ctxt,idim,a,bv,xv,desc_a,'CSR ',info,partition=3,tnd=tnd)
|
||||
call psb_d_gen_pde3d(ctxt,idim,a,bv,xv,desc_a,'CSR ',info,partition=3,tnd=tnd)
|
||||
call psb_barrier(ctxt)
|
||||
t2 = psb_wtime() - t1
|
||||
if(info /= psb_success_) then
|
||||
@@ -691,29 +197,54 @@ program pdgenmv
|
||||
stop
|
||||
end if
|
||||
|
||||
#ifdef OPENACC
|
||||
select case(psb_toupper(agfmt))
|
||||
#if defined(PSB_HAVE_CUDA)
|
||||
case('ELG')
|
||||
agmold => aelg
|
||||
vgmold => dvgpu
|
||||
igmold => ivgpu
|
||||
case('HLG')
|
||||
call psi_set_hksz(32)
|
||||
agmold => ahlg
|
||||
vgmold => dvgpu
|
||||
igmold => ivgpu
|
||||
case('HDIAG')
|
||||
agmold => ahdiag
|
||||
vgmold => dvgpu
|
||||
igmold => ivgpu
|
||||
case('CSRG')
|
||||
agmold => acsrg
|
||||
case default
|
||||
write(*,*) 'Unknown format defaulting to HLG'
|
||||
agmold => ahlg
|
||||
end select
|
||||
vgmold => dvgpu
|
||||
igmold => ivgpu
|
||||
#endif
|
||||
#if defined(PSB_OPENACC)
|
||||
case('ELO')
|
||||
agmold => aelo
|
||||
vgmold => dvoac
|
||||
igmold => ivoac
|
||||
case('HLO')
|
||||
call psi_set_hksz(32)
|
||||
agmold => ahlo
|
||||
vgmold => dvoac
|
||||
igmold => ivoac
|
||||
!!$ case('HDIAG')
|
||||
!!$ amold => ahdiag
|
||||
case('CSRO')
|
||||
agmold => acsro
|
||||
vgmold => dvoac
|
||||
igmold => ivoac
|
||||
#endif
|
||||
end select
|
||||
call a%cscnv(agpu,info,mold=agmold)
|
||||
if ((info /= 0).or.(psb_get_errstatus()/=0)) then
|
||||
write(0,*) 'From cscnv ',info
|
||||
call psb_error()
|
||||
stop
|
||||
end if
|
||||
call desc_a%cnv(mold=imold)
|
||||
call desc_a%cnv(mold=igmold)
|
||||
|
||||
call psb_geasb(bg,desc_a,info,scratch=.true.,mold=vmold)
|
||||
call psb_geasb(xg,desc_a,info,scratch=.true.,mold=vmold)
|
||||
call psb_geasb(bg,desc_a,info,scratch=.true.,mold=vgmold)
|
||||
call psb_geasb(xg,desc_a,info,scratch=.true.,mold=vgmold)
|
||||
nr = desc_a%get_local_rows()
|
||||
nrg = desc_a%get_global_rows()
|
||||
call psb_geall(x0,desc_a,info)
|
||||
@@ -738,11 +269,10 @@ program pdgenmv
|
||||
call xv%bld(x0)
|
||||
call psb_geasb(bv,desc_a,info,scratch=.true.)
|
||||
|
||||
#ifdef OPENACC
|
||||
|
||||
#if defined(PSB_OPENACC)||defined(PSB_HAVE_CUDA)
|
||||
call aux_a%cscnv(agpu,info,mold=acoo)
|
||||
call xg%bld(x0,mold=vmold)
|
||||
call psb_geasb(bg,desc_a,info,scratch=.true.,mold=vmold)
|
||||
call xg%bld(x0,mold=vgmold)
|
||||
call psb_geasb(bg,desc_a,info,scratch=.true.,mold=vgmold)
|
||||
call psb_barrier(ctxt)
|
||||
t1 = psb_wtime()
|
||||
call agpu%cscnv(info,mold=agmold)
|
||||
@@ -765,7 +295,7 @@ program pdgenmv
|
||||
t2 = psb_wtime() - t1
|
||||
call psb_amx(ctxt,t2)
|
||||
|
||||
#ifdef OPENACC
|
||||
#if defined(PSB_OPENACC)||defined(PSB_HAVE_CUDA)
|
||||
call xg%set(x0)
|
||||
|
||||
! FIXME: cache flush needed here
|
||||
@@ -861,7 +391,7 @@ program pdgenmv
|
||||
tflops = flops
|
||||
gflops = flops * ngpu
|
||||
write(psb_out_unit,'("Storage type for A: ",a)') a%get_fmt()
|
||||
#ifdef OPENACC
|
||||
#if defined(PSB_OPENACC)||defined(PSB_HAVE_CUDA)
|
||||
write(psb_out_unit,'("Storage type for AGPU: ",a)') agpu%get_fmt()
|
||||
write(psb_out_unit,'("Time to convert A from COO to CPU (1): ",F20.9)')&
|
||||
& tcnvc1
|
||||
@@ -891,7 +421,7 @@ program pdgenmv
|
||||
& t2*1.d3/(1.d0*ntests)
|
||||
write(psb_out_unit,'("MFLOPS (CPU) : ",F20.3)')&
|
||||
& flops/1.d6
|
||||
#ifdef OPENACC
|
||||
#if defined(PSB_OPENACC)||defined(PSB_HAVE_CUDA)
|
||||
write(psb_out_unit,'("Time for ",i6," products (s) (xGPU) : ",F20.3)')&
|
||||
& ntests, tt2
|
||||
write(psb_out_unit,'("Time per product (ms) (xGPU) : ",F20.3)')&
|
||||
@@ -916,7 +446,7 @@ program pdgenmv
|
||||
bdwdth = ntests*nbytes/(t2*1.d6)
|
||||
write(psb_out_unit,*)
|
||||
write(psb_out_unit,'("MBYTES/S sust. effective bandwidth (CPU) : ",F20.3)') bdwdth
|
||||
#ifdef OPENACC
|
||||
#if defined(PSB_OPENACC)||defined(PSB_HAVE_CUDA)
|
||||
bdwdth = ngpu*ntests*nbytes/(gt2*1.d6)
|
||||
write(psb_out_unit,'("MBYTES/S sust. effective bandwidth (GPU) : ",F20.3)') bdwdth
|
||||
!!$ bdwdth = psb_oacc_MemoryPeakBandwidth()
|
||||
@@ -941,8 +471,11 @@ program pdgenmv
|
||||
call psb_errpush(info,name,a_err=ch_err)
|
||||
goto 9999
|
||||
end if
|
||||
#ifdef OPENACC
|
||||
#ifdef PSB_OPENACC
|
||||
call psb_oacc_exit()
|
||||
#endif
|
||||
#ifdef PSB_HAVE_CUDA
|
||||
call psb_cuda_exit()
|
||||
#endif
|
||||
call psb_exit(ctxt)
|
||||
stop
|
||||
@@ -967,7 +500,7 @@ contains
|
||||
if (iam == 0) then
|
||||
write(*,*) 'CPU side format?'
|
||||
read(psb_inp_unit,*) acfmt
|
||||
write(*,*) 'OACC side format?'
|
||||
write(*,*) 'DEVICE side format?'
|
||||
read(psb_inp_unit,*) agfmt
|
||||
write(*,*) 'Size of discretization cube?'
|
||||
read(psb_inp_unit,*) idim
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user