Skip to content

Commit

Permalink
GMRESMLMG: Support Multi-level composite solve (#4271)
Browse files Browse the repository at this point in the history 
We also need to use smoother at the bottom level of the mlmg
preconditioner, because our cg solvers are non-deterministic. What that
means is with cg solvers the preconditioner matrix is not fixed. That
can result in wrong solutions in GMRES.
  • Loading branch information
@WeiqunZhang
WeiqunZhang authored Feb 14, 2025
1 parent 275f55f commit fab0411
Show file tree
Hide file tree
Showing 24 changed files with 1,002 additions and 274 deletions.
201 changes: 201 additions & 0 deletions Src/Base/AMReX_FabArrayUtility.H
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1602,6 +1602,207 @@ Dot (FabArray<FAB> const& x, int xcomp, FabArray<FAB> const& y, int ycomp, int n
return sm;
}

/**
* \brief Compute dot product of FabArray with itself
*
* \param x FabArray
* \param xcomp starting component of x
* \param ncomp number of components
* \param nghost number of ghost cells
* \param local If true, MPI communication is skipped.
*/
template <typename FAB, std::enable_if_t<IsBaseFab<FAB>::value,int> FOO = 0>
typename FAB::value_type
Dot (FabArray<FAB> const& x, int xcomp, int ncomp, IntVect const& nghost, bool local = false)
{
BL_ASSERT(x.nGrowVect().allGE(nghost));

BL_PROFILE("amrex::Dot()");

using T = typename FAB::value_type;
auto sm = T(0.0);
#ifdef AMREX_USE_GPU
if (Gpu::inLaunchRegion()) {
auto const& xma = x.const_arrays();
sm = ParReduce(TypeList<ReduceOpSum>{}, TypeList<T>{}, x, nghost,
[=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k) noexcept -> GpuTuple<T>
{
auto t = T(0.0);
auto const& xfab = xma[box_no];
for (int n = 0; n < ncomp; ++n) {
auto v = xfab(i,j,k,xcomp+n);
t += v*v;
}
return t;
});
} else
#endif
{
#ifdef AMREX_USE_OMP
#pragma omp parallel if (!system::regtest_reduction) reduction(+:sm)
#endif
for (MFIter mfi(x,true); mfi.isValid(); ++mfi)
{
Box const& bx = mfi.growntilebox(nghost);
auto const& xfab = x.const_array(mfi);
AMREX_LOOP_4D(bx, ncomp, i, j, k, n,
{
auto v = xfab(i,j,k,xcomp+n);
sm += v*v;
});
}
}

if (!local) {
ParallelAllReduce::Sum(sm, ParallelContext::CommunicatorSub());
}

return sm;
}

/**
* \brief Compute dot product of two FabArrays in region that mask is true
*
* \param mask mask
* \param x first FabArray
* \param xcomp starting component of x
* \param y second FabArray
* \param ycomp starting component of y
* \param ncomp number of components
* \param nghost number of ghost cells
* \param local If true, MPI communication is skipped.
*/
template <typename IFAB, typename FAB,
std::enable_if_t<IsBaseFab<FAB>::value && IsBaseFab<IFAB>::value,int> FOO = 0>
typename FAB::value_type
Dot (FabArray<IFAB> const& mask, FabArray<FAB> const& x, int xcomp,
FabArray<FAB> const& y, int ycomp, int ncomp, IntVect const& nghost,
bool local = false)
{
BL_ASSERT(x.boxArray() == y.boxArray() && x.boxArray() == mask.boxArray());
BL_ASSERT(x.DistributionMap() == y.DistributionMap() && x.DistributionMap() == mask.DistributionMap());
BL_ASSERT(x.nGrowVect().allGE(nghost) && y.nGrowVect().allGE(nghost) &&
mask.nGrowVect().allGE(nghost));

BL_PROFILE("amrex::Dot()");

using T = typename FAB::value_type;
auto sm = T(0.0);
#ifdef AMREX_USE_GPU
if (Gpu::inLaunchRegion()) {
auto const& mma = mask.const_arrays();
auto const& xma = x.const_arrays();
auto const& yma = y.const_arrays();
sm = ParReduce(TypeList<ReduceOpSum>{}, TypeList<T>{}, x, nghost,
[=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k) noexcept -> GpuTuple<T>
{
auto t = T(0.0);
auto m = T(mma[box_no](i,j,k));
if (m != 0) {
auto const& xfab = xma[box_no];
auto const& yfab = yma[box_no];
for (int n = 0; n < ncomp; ++n) {
t += xfab(i,j,k,xcomp+n) * yfab(i,j,k,ycomp+n);
}
}
return t*m;
});
} else
#endif
{
#ifdef AMREX_USE_OMP
#pragma omp parallel if (!system::regtest_reduction) reduction(+:sm)
#endif
for (MFIter mfi(x,true); mfi.isValid(); ++mfi)
{
Box const& bx = mfi.growntilebox(nghost);
auto const& mfab = mask.const_array(mfi);
auto const& xfab = x.const_array(mfi);
auto const& yfab = y.const_array(mfi);
AMREX_LOOP_4D(bx, ncomp, i, j, k, n,
{
auto m = T(mfab(i,j,k));
sm += m * xfab(i,j,k,xcomp+n) * yfab(i,j,k,ycomp+n);
});
}
}

if (!local) {
ParallelAllReduce::Sum(sm, ParallelContext::CommunicatorSub());
}

return sm;
}

/**
* \brief Compute dot product of FabArray with itself in region that mask is true
*
* \param mask mask
* \param x FabArray
* \param xcomp starting component of x
* \param ncomp number of components
* \param nghost number of ghost cells
* \param local If true, MPI communication is skipped.
*/
template <typename IFAB, typename FAB,
std::enable_if_t<IsBaseFab<FAB>::value && IsBaseFab<IFAB>::value,int> FOO = 0>
typename FAB::value_type
Dot (FabArray<IFAB> const& mask, FabArray<FAB> const& x, int xcomp, int ncomp,
IntVect const& nghost, bool local = false)
{
BL_ASSERT(x.boxArray() == mask.boxArray());
BL_ASSERT(x.DistributionMap() == mask.DistributionMap());
BL_ASSERT(x.nGrowVect().allGE(nghost) && mask.nGrowVect().allGE(nghost));

BL_PROFILE("amrex::Dot()");

using T = typename FAB::value_type;
auto sm = T(0.0);
#ifdef AMREX_USE_GPU
if (Gpu::inLaunchRegion()) {
auto const& mma = mask.const_arrays();
auto const& xma = x.const_arrays();
sm = ParReduce(TypeList<ReduceOpSum>{}, TypeList<T>{}, x, nghost,
[=] AMREX_GPU_DEVICE (int box_no, int i, int j, int k) noexcept -> GpuTuple<T>
{
auto t = T(0.0);
auto m = T(mma[box_no](i,j,k));
if (m != 0) {
auto const& xfab = xma[box_no];
for (int n = 0; n < ncomp; ++n) {
auto v = xfab(i,j,k,xcomp+n);
t += v*v;
}
}
return t*m;
});
} else
#endif
{
#ifdef AMREX_USE_OMP
#pragma omp parallel if (!system::regtest_reduction) reduction(+:sm)
#endif
for (MFIter mfi(x,true); mfi.isValid(); ++mfi)
{
Box const& bx = mfi.growntilebox(nghost);
auto const& mfab = mask.const_array(mfi);
auto const& xfab = x.const_array(mfi);
AMREX_LOOP_4D(bx, ncomp, i, j, k, n,
{
auto m = T(mfab(i,j,k));
auto v = xfab(i,j,k,xcomp+n);
sm += m*v*v;
});
}
}

if (!local) {
ParallelAllReduce::Sum(sm, ParallelContext::CommunicatorSub());
}

return sm;
}

//! dst = val
template <class MF, std::enable_if_t<IsMultiFabLike_v<MF>,int> = 0>
void setVal (MF& dst, typename MF::value_type val)
Expand Down
18 changes: 18 additions & 0 deletions Src/Base/AMReX_Vector.H
View file
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,15 @@ namespace amrex
return r;
}

template <class T, std::size_t N, typename = typename T::FABType>
[[nodiscard]] Vector<Array<T,N>*> GetVecOfPtrs (Vector<Array<T,N>>& a)
{
Vector<Array<T,N>*> r;
r.reserve(a.size());
for (auto& x : a) { r.push_back(&x); }
return r;
}

template <class T>
[[nodiscard]] Vector<T*> GetVecOfPtrs (const Vector<std::unique_ptr<T> >& a)
{
Expand All @@ -86,6 +95,15 @@ namespace amrex
return r;
}

template <class T, std::size_t N, typename = typename T::FABType>
[[nodiscard]] Vector<Array<T,N> const*> GetVecOfConstPtrs (Vector<Array<T,N>> const& a)
{
Vector<Array<T,N> const*> r;
r.reserve(a.size());
for (auto& x : a) { r.push_back(&x); }
return r;
}

template <class T>
[[nodiscard]] Vector<const T*> GetVecOfConstPtrs (const Vector<std::unique_ptr<T> >& a)
{
Expand Down
Loading

0 comments on commit fab0411

Please sign in to comment.