Adds row scatter functionality to dense #1356

MarcelKoch · 2023-06-30T15:28:59Z

This PR add row scatter functionality to dense. The target rows can be specified either by a plain array<index_type> or by index_set<index_type>. Note that the index_set implies that the target rows are ordered. Signature:

void row_scatter(const array<index_type>*, LinOp*);
void row_scatter(const index_set<index_type>*, LinOp*);

Todo:

dpcpp kernels....
maybe have an additive variant like the gather

yhmtsai

the function description and one test variable

include/ginkgo/core/base/index_set.hpp

include/ginkgo/core/matrix/dense.hpp

test/matrix/dense_kernels.cpp

sonarqubecloud · 2023-08-03T01:27:45Z

Kudos, SonarCloud Quality Gate passed!

0 Bugs
0 Vulnerabilities
1 Security Hotspot
19 Code Smells

83.9% Coverage
4.8% Duplication

yhmtsai · 2023-08-07T08:23:19Z

core/matrix/dense.cpp

+    dim<2> expected_dim{row_idxs->get_num_elems(), this->get_size()[1]};
+    GKO_ASSERT_EQUAL_DIMENSIONS(expected_dim, this);
+    GKO_ASSERT_EQUAL_COLS(this, target);
+    // @todo check that indices are inbounds for target


not sure whether to check the indices here because it will affect the performance. I guess it will have the same runtime as make_row_scatter

Yes, that would also be my expectation. Perhaps it could be handled by setting an error code in the kernel and then check that afterward. I will need to think if that can be realized.

I added an in-kernel check to make sure we don't write out-of-bounds. I'm open to discussions on how to handle that, because I don't think there are many places where we already use something like this. In principle, we could also do this for the gather kernels, although I don't think the issue is as big there, since we are only reading out-of-bounds. Also, we could wrap the check in some #ifdef to disable it perhaps in release builds.

core/matrix/dense.cpp

greole

Not a full review yet, just some quick comments/questions?

greole · 2023-09-21T09:29:28Z

include/ginkgo/core/matrix/dense.hpp

+    /**
+     * Copies this matrix into the given rows of the target matrix.
+     *
+     * @tparam IndexType  the index type, either int32 or int 64


Suggested change

* @tparam IndexType the index type, either int32 or int 64

* @tparam IndexType the index type, either int32 or int64

greole · 2023-09-21T09:33:01Z

include/ginkgo/core/matrix/dense.hpp

@@ -539,6 +543,42 @@ class Dense
                    ptr_param<const LinOp> beta,
                    ptr_param<LinOp> row_collection) const;

+    /**
+     * Copies this matrix into the given rows of the target matrix.


I think this needs to be a bit more expressive in what it means to "copy this matrix into the given rows". Could you add an example here.

TBH, since this is the brief description, I think it's fine. But the detailed description is mangled up in the description of the parameters, so perhaps moving that into a separate description block is better.

greole · 2023-09-21T09:33:37Z

include/ginkgo/core/matrix/dense.hpp

+    /**
+     * Copies this matrix into the given rows of the target matrix.
+     *
+     * @tparam IndexType  the index type, either int32 or int 64


Suggested change

* @tparam IndexType the index type, either int32 or int 64

* @tparam IndexType the index type, either int32 or int64

greole · 2023-09-21T09:33:48Z

include/ginkgo/core/matrix/dense.hpp

+                     ptr_param<LinOp> target) const;
+
+    /**
+     * Copies this matrix into the given rows of the target matrix.


Same comment as above.

greole · 2023-09-21T09:54:10Z

reference/test/matrix/dense_kernels.cpp

+                        l<T>({{3.0, 2.7, 6.5},
+                              {0.7, 1.1, 4.0},


I am confused, looks to me like row 0 and 1 should be switched? Doesn't permute_idxs{exec, {1, 0, 4}} say {3.0, 2.7, 6.5} goes into row 1 ?

An index set is always ordered. So permute_idxs{exec, {1, 0, 4}} is equivalent to permute_idxs{exec, {0, 1, 4}}. I will add a comment to row_scatter

stale

sonarqubecloud · 2023-10-16T16:02:03Z

Kudos, SonarCloud Quality Gate passed!

0 Bugs
0 Vulnerabilities
1 Security Hotspot
18 Code Smells

82.5% Coverage
0.0% Duplication

The version of Java (11.0.3) you have used to run this analysis is deprecated and we will stop accepting it soon. Please update to at least Java 17.
Read more here

sonarqubecloud · 2023-12-15T21:51:06Z

Quality Gate passed

The SonarCloud Quality Gate passed, but some issues were introduced.

18 New issues
1 Security Hotspot
83.5% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarCloud

upsj

LGTM in general, though I would consider making this more strongly typed than just an array, then we also have a well-defined place where we can check for invalid data #770. I think we should already keep an eye out for a potential future scatter_add interface, which may profit from some preprocessing to avoid atomics, which would also encourage a dedicated type.

common/cuda_hip/matrix/dense_kernels.hpp.inc

common/unified/matrix/dense_kernels.template.cpp

core/matrix/dense.cpp

MarcelKoch · 2024-10-23T14:55:43Z

@upsj I guess one way to have a more specific type would be to use index_set. But I think the array is the more natural way to describe the indices, so I would just go with it for now. If we decide on a more sophisticated type, I would expect that we can construct that type from an array, so it would not be difficult to add an overload.

upsj · 2024-10-23T15:06:12Z

@MarcelKoch my goal is to deprecate the array interfaces in favor of Permutation and related classes in GInkgo 2.0, for consistency it would seem sensible to resolve this topic before adding new interfaces which might need to be deprecated shortly after. index_set IMO is an overly specific type for this purpose, as its storage layout favors a specific kind of subset, which doesn't give any performance benefit in the use cases for row scattering I can think of (namely halo exchange) on general meshes.

MarcelKoch · 2024-10-23T15:14:26Z

Ok, that makes more sense. I will look into providing a Scatter class instead.

MarcelKoch · 2024-10-24T07:34:48Z

After looking into it a bit, I don't think adding an extra class for scattering is the right choice at the moment. Without the scatter_add this is a fairly simple operation, so an extra class would be overkill. E.g. if the scatter is not done repeatedly, users would have to write:

RowScatterer<int32>::create(exec, size, idxs)->apply(a, b);

which to me seems to be quite cumbersome, compared to a->row_scatter(idxs, b).

upsj · 2024-10-24T07:47:21Z

To me this is an abstraction level mismatch - array is a low-level type with no interpretation of the data associated with it, Dense is a high-level type that gives an interpretation to the underlying data. The array you pass into row_scatter could be anything. Otherwise the same kind of argument could be made for Permutation. IMO we should aim for interfaces that don't mix different abstraction levels.

As for the discussion between scatter_add and scatter, I believe we should consider making scatter_add the default, since otherwise the operation can produce race conditions and non-deterministic behavior trivially.

MarcelKoch · 2024-10-24T08:12:36Z

Making scatter_add the default will require quite a bit of work. Since I don't think this feature is urgently needed, I will remove this PR from the upcoming release.

currently version with index_set as input is not implemented for devices Co-authored-by: Yu-Hsiang M. Tsai <[email protected]> Signed-off-by: Marcel Koch <[email protected]>

Signed-off-by: Marcel Koch <[email protected]>

Co-authored-by: Yu-Hsiang M. Tsai <[email protected]>

Co-authored-by: Tobias Ribizel <[email protected]>

- add todo - more specific error message Co-authored-by: Tobias Ribizel <[email protected]>

Signed-off-by: Marcel Koch <[email protected]>

ginkgo-bot added reg:testing This is related to testing. type:matrix-format This is related to the Matrix formats mod:all This touches all Ginkgo modules. labels Jun 30, 2023

MarcelKoch self-assigned this Jun 30, 2023

MarcelKoch added the 1:ST:ready-for-review This PR is ready for review label Jul 3, 2023

yhmtsai requested changes Jul 10, 2023

View reviewed changes

MarcelKoch force-pushed the row-scatter branch from dcaabd9 to 4908ff5 Compare August 2, 2023 11:19

yhmtsai previously approved these changes Aug 7, 2023

View reviewed changes

upsj requested review from upsj and greole September 18, 2023 12:33

greole reviewed Sep 21, 2023

View reviewed changes

MarcelKoch force-pushed the row-scatter branch from bb553d5 to 7d52f8c Compare October 13, 2023 14:32

MarcelKoch force-pushed the row-scatter branch from d6223ed to 39ae497 Compare October 16, 2023 06:58

MarcelKoch force-pushed the row-scatter branch 2 times, most recently from caab3d0 to ecc1b8c Compare December 15, 2023 09:52

MarcelKoch requested review from greole and yhmtsai December 15, 2023 10:05

MarcelKoch force-pushed the row-scatter branch from ecc1b8c to 3247327 Compare December 15, 2023 10:47

MarcelKoch added this to the Ginkgo 1.9.0 milestone Aug 26, 2024

upsj reviewed Oct 18, 2024

View reviewed changes

common/cuda_hip/matrix/dense_kernels.hpp.inc Outdated Show resolved Hide resolved

common/unified/matrix/dense_kernels.template.cpp Outdated Show resolved Hide resolved

core/matrix/dense.cpp Outdated Show resolved Hide resolved

MarcelKoch force-pushed the row-scatter branch from 3247327 to 4ad1dfe Compare October 23, 2024 14:56

MarcelKoch removed this from the Ginkgo 1.9.0 milestone Oct 24, 2024

MarcelKoch added 1:ST:WIP This PR is a work in progress. Not ready for review. and removed 1:ST:ready-for-review This PR is ready for review labels Oct 24, 2024

MarcelKoch marked this pull request as draft October 24, 2024 08:13

MarcelKoch force-pushed the row-scatter branch from cbc4aec to 83bc786 Compare February 12, 2025 16:29

MarcelKoch and others added 11 commits February 21, 2025 10:46

adds Dense::row_scatter

91979f5

currently version with index_set as input is not implemented for devices Co-authored-by: Yu-Hsiang M. Tsai <[email protected]> Signed-off-by: Marcel Koch <[email protected]>

adds device index set kernels

ea6825c

Signed-off-by: Marcel Koch <[email protected]>

adds in-bounds check to kernels

eca9b97

Co-authored-by: Yu-Hsiang M. Tsai <[email protected]>

remove index-set overload

0efaf63

Co-authored-by: Tobias Ribizel <[email protected]>

review updates:

14d04ed

- add todo - more specific error message Co-authored-by: Tobias Ribizel <[email protected]>

wip: use row-scatterer type

0a831d7

Signed-off-by: Marcel Koch <[email protected]>

move row_scatter kernels into own file

4a4610d

update tests to new interface

19cf8f1

add reference kernel test

5ce5b3d

use additive row scatter

9b9fa61

adds advanced scatter operation

d5da55e

MarcelKoch force-pushed the row-scatter branch from 83bc786 to d5da55e Compare February 21, 2025 15:33

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adds row scatter functionality to dense #1356

Adds row scatter functionality to dense #1356

MarcelKoch commented Jun 30, 2023 •

edited

Loading

yhmtsai left a comment

sonarqubecloud bot commented Aug 3, 2023

yhmtsai Aug 7, 2023

MarcelKoch Aug 7, 2023

MarcelKoch Aug 7, 2023

greole left a comment

greole Sep 21, 2023

greole Sep 21, 2023

MarcelKoch Sep 21, 2023

greole Sep 21, 2023

greole Sep 21, 2023

greole Sep 21, 2023

MarcelKoch Sep 21, 2023

sonarqubecloud bot commented Oct 16, 2023

sonarqubecloud bot commented Dec 15, 2023

upsj left a comment

MarcelKoch commented Oct 23, 2024

upsj commented Oct 23, 2024

MarcelKoch commented Oct 23, 2024

MarcelKoch commented Oct 24, 2024

upsj commented Oct 24, 2024

MarcelKoch commented Oct 24, 2024

	* @tparam IndexType the index type, either int32 or int 64
	* @tparam IndexType the index type, either int32 or int64

Adds row scatter functionality to dense #1356

Are you sure you want to change the base?

Adds row scatter functionality to dense #1356

Conversation

MarcelKoch commented Jun 30, 2023 • edited Loading

yhmtsai left a comment

Choose a reason for hiding this comment

sonarqubecloud bot commented Aug 3, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

greole left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sonarqubecloud bot commented Oct 16, 2023

sonarqubecloud bot commented Dec 15, 2023

Quality Gate passed

upsj left a comment

Choose a reason for hiding this comment

MarcelKoch commented Oct 23, 2024

upsj commented Oct 23, 2024

MarcelKoch commented Oct 23, 2024

MarcelKoch commented Oct 24, 2024

upsj commented Oct 24, 2024

MarcelKoch commented Oct 24, 2024

MarcelKoch commented Jun 30, 2023 •

edited

Loading