Add a uniform coarsening algorithm by pratikvn · Pull Request #2016 · ginkgo-project/ginkgo

pratikvn · 2026-05-07T07:20:10Z

This PR adds a uniform coarsening algorithm, an improved version of #1526. The idea is the same as in there, but now there is an option to enforce connectedness that ensures that the graphs in the coarser levels remain connected and there are no disjoint vertices.

While still algebraic, this coarsening strategy is cheaper to generate than PGM, and might be beneficial for matrices with a uniform sparsity pattern.

yhmtsai

any reason to keep prolong/restrict as CSR not using SparsityCSR and RowGatherer?

yhmtsai · 2026-05-19T07:42:41Z

+            if (c > 0) A_data.nonzeros.emplace_back(i, i - 1, -1);
+            if (c < N - 1) A_data.nonzeros.emplace_back(i, i + 1, -1);
+            if (r > 0) A_data.nonzeros.emplace_back(i, i - N, -1);
+            if (r < N - 1) A_data.nonzeros.emplace_back(i, i + N, -1);


Suggested change

if (c > 0) A_data.nonzeros.emplace_back(i, i - 1, -1);

if (c < N - 1) A_data.nonzeros.emplace_back(i, i + 1, -1);

if (r > 0) A_data.nonzeros.emplace_back(i, i - N, -1);

if (r < N - 1) A_data.nonzeros.emplace_back(i, i + N, -1);

if (c > 0) {A_data.nonzeros.emplace_back(i, i - 1, -1);}

if (c < N - 1) {A_data.nonzeros.emplace_back(i, i + 1, -1);}

if (r > 0) {A_data.nonzeros.emplace_back(i, i - N, -1);}

if (r < N - 1) {A_data.nonzeros.emplace_back(i, i + N, -1);}

maybe you can reorder it a bit such that it is already sorted

still miss {}

yhmtsai · 2026-05-19T07:45:21Z

+ * UniformCoarsening is a simple coarse grid generation algorithm. It
+ * selects the coarse matrix from the fine matrix by constant jumps that can be
+ * specified by the user.


It sounds a geometric like?

updated the wording now. Its not really geometric, but its not algebraic either. So, I just say index-based ordering now.

yhmtsai · 2026-05-19T08:13:36Z

+         * When set to `true` (the default), every fine row is mapped to
+         * its nearest coarse row (aggregation-style), so that the Galerkin


add more detail about the nearest coarse row? It is assigned to the floor(id/skip_coarse), right?

yhmtsai · 2026-05-19T08:15:37Z

+    using real_type = gko::remove_complex<value_type>;
+    UniformCoarsening()


Suggested change

using real_type = gko::remove_complex<value_type>;

UniformCoarsening()

using real_type = gko::remove_complex<value_type>;

UniformCoarsening()

yhmtsai · 2026-05-19T08:18:46Z

+        SCOPED_TRACE("Using coarse skip:" + std::to_string(coarse_skip));
+        gko::kernels::reference::uniform_coarsening::fill_incremental_indices(
+            ref, coarse_skip, &c_rows);
+        gko::kernels::GKO_DEVICE_NAMESPACE::uniform_coarsening::
+            fill_incremental_indices(exec, coarse_skip, &d_c_rows);
+        GKO_ASSERT_ARRAY_EQ(c_rows, d_c_rows);


Suggested change

SCOPED_TRACE("Using coarse skip:" + std::to_string(coarse_skip));

gko::kernels::reference::uniform_coarsening::fill_incremental_indices(

ref, coarse_skip, &c_rows);

gko::kernels::GKO_DEVICE_NAMESPACE::uniform_coarsening::

fill_incremental_indices(exec, coarse_skip, &d_c_rows);

GKO_ASSERT_ARRAY_EQ(c_rows, d_c_rows);

SCOPED_TRACE("Using coarse skip:" + std::to_string(coarse_skip));

c_rows.fill(-gko::one<index_type>());

d_c_rows.fill(-gko::one<index_type>());

gko::kernels::reference::uniform_coarsening::fill_incremental_indices(

ref, coarse_skip, &c_rows);

gko::kernels::GKO_DEVICE_NAMESPACE::uniform_coarsening::

fill_incremental_indices(exec, coarse_skip, &d_c_rows);

GKO_ASSERT_ARRAY_EQ(c_rows, d_c_rows);

because it updates the value in each loop

yhmtsai · 2026-05-19T08:53:26Z

+    using VT = value_type;
+    UniformCoarseningAgg()


Suggested change

using VT = value_type;

UniformCoarseningAgg()

using VT = value_type;

UniformCoarseningAgg()

yhmtsai · 2026-05-19T08:55:53Z

+    using SparsityCsr = gko::matrix::SparsityCsr<value_type, index_type>;
+    using MgLevel = gko::multigrid::UniformCoarsening<value_type, index_type>;
+    using RowGatherer = gko::matrix::RowGatherer<index_type>;


Suggested change

using SparsityCsr = gko::matrix::SparsityCsr<value_type, index_type>;

using MgLevel = gko::multigrid::UniformCoarsening<value_type, index_type>;

using RowGatherer = gko::matrix::RowGatherer<index_type>;

using MgLevel = gko::multigrid::UniformCoarsening<value_type, index_type>;

yhmtsai · 2026-05-19T08:57:34Z

+    auto c_rows = gko::array<index_type>(this->exec, 10);
+    c_rows.fill(-gko::one<index_type>());
+
+    gko::kernels::reference::uniform_coarsening::fill_incremental_indices(
+        this->exec, 2, &c_rows);
+    GKO_ASSERT_ARRAY_EQ(c_rows, c2_rows);


Suggested change

auto c_rows = gko::array<index_type>(this->exec, 10);

c_rows.fill(-gko::one<index_type>());

gko::kernels::reference::uniform_coarsening::fill_incremental_indices(

this->exec, 2, &c_rows);

GKO_ASSERT_ARRAY_EQ(c_rows, c2_rows);

auto c_rows = gko::array<index_type>(this->exec, 10);

c_rows.fill(-gko::one<index_type>());

gko::kernels::reference::uniform_coarsening::fill_incremental_indices(

this->exec, 2, &c_rows);

GKO_ASSERT_ARRAY_EQ(c_rows, c2_rows);

It is more to keep the consistent as the following.

yhmtsai · 2026-05-19T09:10:09Z

+}
+
+
+}  // namespace


same here

Suggested change

} // namespace

yhmtsai · 2026-05-19T09:22:42Z

+    auto uc_factory =
+        uc::build().with_coarse_skip(2).with_aggregation(false).on(this->exec);


I think we need another test for the diag matrix size is not divisible by coarse_skip like coarse_skip = 3

yhmtsai · 2026-05-21T15:00:19Z

+            if (c > 0) A_data.nonzeros.emplace_back(i, i - 1, -1);
+            if (c < N - 1) A_data.nonzeros.emplace_back(i, i + 1, -1);
+            if (r > 0) A_data.nonzeros.emplace_back(i, i - N, -1);
+            if (r < N - 1) A_data.nonzeros.emplace_back(i, i + N, -1);


still miss {}

yhmtsai · 2026-06-08T06:48:36Z

+                              array<IndexType>* coarse_rows)
+{
+    const size_type num_elems = coarse_rows->get_size();
+    const size_type num_iters = (num_elems + coarse_skip - 1) / coarse_skip;


use ceildiv

yhmtsai · 2026-06-08T06:49:41Z

+
+
+/**
+ * UniformCoarsening is a simple algebraic coarse grid generation algorithm.


Suggested change

* UniformCoarsening is a simple algebraic coarse grid generation algorithm.

* UniformCoarsening is a simple geometric coarse grid generation algorithm.

To me it is more like geometric?

stale

yhmtsai · 2026-07-20T12:30:14Z

@pratikvn after rebase, you do not need EnableLinOp anymore, but you will need to change it to LinOp and corresponding constructor

ginkgo-bot · 2026-07-22T15:59:32Z

Error: The following files need to be formatted:

include/ginkgo/core/multigrid/uniform_coarsening.hpp

You can find a formatting patch under Artifacts here or run format! if you have write access to Ginkgo

pratikvn requested review from MarcelKoch and yhmtsai May 7, 2026 07:20

pratikvn self-assigned this May 7, 2026

pratikvn added is:new-feature A request or implementation of a feature that does not exist yet. 1:ST:ready-for-review This PR is ready for review labels May 7, 2026

yhmtsai requested changes May 15, 2026

View reviewed changes

pratikvn force-pushed the unif-coarsening branch from fd10b08 to 7d5f6cc Compare May 17, 2026 23:23

pratikvn requested a review from yhmtsai May 18, 2026 09:04

yhmtsai requested changes May 19, 2026

View reviewed changes

yhmtsai previously requested changes Jun 8, 2026

View reviewed changes

pratikvn force-pushed the unif-coarsening branch from 5f54885 to ceef8f9 Compare June 16, 2026 19:32

pratikvn requested a review from yhmtsai June 16, 2026 19:34

pratikvn force-pushed the unif-coarsening branch from ceef8f9 to 7e13493 Compare July 13, 2026 09:00

pratikvn added 8 commits July 22, 2026 17:24

add UniformCoarsening core and config

8ab0cf5

add core tests

7149aa3

add reference kernels and tests

cb84155

add common kernels, tests and mpi tests

469be1c

enforce_connectedness by default

778466c

review updates, rename to aggregation

2d3c195

remove dep on spgemm

61a4e1a

update example, deprec functions

706272f

pratikvn added 6 commits July 22, 2026 17:24

add off-diag aggregation test

2ddadbb

fix injection imap bug, update example

ad24367

review updates

b6e7019

add TODOs on unification with PGM

552bb14

review updates

97226d1

cloneable fixes

cef6b44

pratikvn force-pushed the unif-coarsening branch from 7e13493 to cef6b44 Compare July 22, 2026 15:56

		* When set to `true` (the default), every fine row is mapped to
		* its nearest coarse row (aggregation-style), so that the Galerkin

		using real_type = gko::remove_complex<value_type>;
		UniformCoarsening()

		auto uc_factory =
		uc::build().with_coarse_skip(2).with_aggregation(false).on(this->exec);



		/**
		* UniformCoarsening is a simple algebraic coarse grid generation algorithm.

	* UniformCoarsening is a simple algebraic coarse grid generation algorithm.
	* UniformCoarsening is a simple geometric coarse grid generation algorithm.

Uh oh!

Conversation

pratikvn commented May 7, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

yhmtsai left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yhmtsai commented Jul 20, 2026

Uh oh!

ginkgo-bot commented Jul 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants