[Sparse] Support Diag sparse format in C++ #5432

czkkkkkk · 2023-03-07T07:55:50Z

Description

This is one of the PR for the issue #5367 .

Checklist

Please feel free to remove inapplicable items for your PR.

The PR title starts with [$CATEGORY] (such as [NN], [Model], [Doc], [Feature]])
I've leverage the tools to beautify the python and c++ code.
The PR is complete and small, read the Google eng practice (CL equals to PR) to understand more about small PR. In DGL, we consider PRs with less than 200 lines of core code change are small (example, test and documentation could be exempted).
All changes have test coverage
Code is well-documented
To the best of my knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change
Related issue is referred in this PR
If the PR is for a new model/paper, I've updated the example index here.

Changes

dgl-bot · 2023-03-07T07:56:19Z

To trigger regression tests:

@dgl-bot run [instance-type] [which tests] [compare-with-branch];
For example: @dgl-bot run g4dn.4xlarge all dmlc/master or @dgl-bot run c5.9xlarge kernel,api dmlc/master

dgl-bot · 2023-03-07T07:57:45Z

Commit ID: b0d2b55

Build ID: 1

Status: ❌ CI test failed in Stage [Lint Check].

Report path: link

Full logs path: link

frozenbugs · 2023-03-07T09:16:17Z

dgl_sparse/include/sparse/sparse_format.h

@@ -90,6 +95,21 @@ std::shared_ptr<CSR> COOToCSC(const std::shared_ptr<COO>& coo);
 /** @brief Convert a CSR format to CSC format. */
 std::shared_ptr<CSR> CSRToCSC(const std::shared_ptr<CSR>& csr);

+/** @brief Convert a Diag format to COO format. */
+std::shared_ptr<COO> DiagToCOO(


Why do we need this 3 conversion? In which case, they will be used?

They are used for operators that do not have implementation on Diag format, e.g., SpMM.

What's the implication of the performance here if we convert diag to COO/CSR/CSV for operators?
Will it strongly decrease the spmm performance?

This solution looks good to me, but let's make sure we understand the trade-off we are making here.

I simply follow our current implementation to ensure no performance regression. Currently, we also convert the DiagMatrix to SparseMatrix for SpMM on the Python side.

dgl-bot · 2023-03-07T09:47:47Z

Commit ID: b6d2f92

Build ID: 2

Status: ✅ CI test succeeded.

Report path: link

Full logs path: link

dgl-bot · 2023-03-08T07:14:43Z

Commit ID: f49b4ca6b7ac65fdaa0467c414e4024503003224

Build ID: 3

Status: ✅ CI test succeeded.

Report path: link

Full logs path: link

jermainewang · 2023-03-08T13:23:44Z

dgl_sparse/src/elemenwise_op.cc

+  if (A->HasDiag() && B->HasDiag()) {
+    return SparseMatrix::FromDiagPointer(
+        A->DiagPtr(), A->value() + B->value(), A->shape());
+  }


nit: I will usually prefer parallel if-else branch, e.g.,

if (...) { ... } else { ... }

@frozenbugs what would you think?

The current implementation is better.
For all functions, it could have multiple early return checks, so it is highly recommended to do:

if (...) { return ... } if (...) { return ... } blablabla blablabla blablabla return ...

jermainewang · 2023-03-08T13:48:21Z

dgl_sparse/src/sparse_format.cc

+    const std::shared_ptr<Diag>& diag,
+    const c10::TensorOptions& indices_options) {
+  int64_t nnz = std::min(diag->num_rows, diag->num_cols);
+  auto indptr = torch::arange(nnz + 1, indices_options);


Will using the inplace operation torch::arange_out be cleaner? This way you can first create an array of length diag->num_rows+1 with all values being nnz, then fill the front part with arange. This also avoids creating multiple intermediate buffers.

jermainewang · 2023-03-08T13:51:57Z

dgl_sparse/src/spspmm.cc

+    const c10::intrusive_ptr<SparseMatrix>& lhs_mat,
+    const c10::intrusive_ptr<SparseMatrix>& rhs_mat) {
+  if (lhs_mat->HasDiag()) {
+    if (rhs_mat->HasDiag()) {


I usually don't like nested if as it is difficult to read. I recommend change it to:

if (lhs_mat->HasDiag() && rhs_mat->HasDiag()) { ... } else if (lhs_mat->HasDiag() && !rhs_mat->HasDiag()) { ... } else { ... }

jermainewang · 2023-03-08T14:02:38Z

dgl_sparse/src/spspmm.cc

+      // Diag @ Sparse
+      auto row = rhs_mat->Indices().index({0});
+      auto val = lhs_mat->value().index_select(0, row) * rhs_mat->value();
+      return SparseMatrix::ValLike(rhs_mat, val);


Is this implementation correct? The index_select should use A.row according to /~https://github.com/dmlc/dgl/blob/master/python/dgl/sparse/matmul.py#L172.

I think so. rhs_mat->Indices().index({0}) returns the row coordinates of the SparseMatrix without data copy.

I see. rhs_mat->Indices() returns the COO index. How is it different with rhs_mat->COOPointer()->indices? Do we want to have two interfaces?

They are the same. But we need Indices() exposed to python to avoid data copy of the COO tensor.

dgl_sparse/src/spspmm.cc

dgl-bot · 2023-03-09T04:38:39Z

Commit ID: bd0b8d124df34e8042355ce185b688f1775d56f8

Build ID: 4

Status: ✅ CI test succeeded.

Report path: link

Full logs path: link

* [Sparse] Support Diag sparse format in C++ * update * Update

czkkkkkk requested review from jermainewang and frozenbugs March 7, 2023 07:55

czkkkkkk mentioned this pull request Mar 7, 2023

[Sparse] Remove Python DiagMatrix class and represent it as a SparseMatrix. #5367

Closed

[Sparse] Support Diag sparse format in C++

b6d2f92

czkkkkkk force-pushed the kdiag branch from b0d2b55 to b6d2f92 Compare March 7, 2023 08:56

frozenbugs reviewed Mar 7, 2023

View reviewed changes

frozenbugs approved these changes Mar 8, 2023

View reviewed changes

update

85042cc

jermainewang requested changes Mar 8, 2023

View reviewed changes

Update

eef2dcd

jermainewang approved these changes Mar 9, 2023

View reviewed changes

jermainewang merged commit a03dec0 into dmlc:master Mar 9, 2023

czkkkkkk added a commit that referenced this pull request Apr 19, 2023

[Sparse] Support Diag sparse format in C++ (#5432)

8b16597

* [Sparse] Support Diag sparse format in C++ * update * Update

DominikaJedynak pushed a commit to DominikaJedynak/dgl that referenced this pull request Mar 12, 2024

[Sparse] Support Diag sparse format in C++ (dmlc#5432)

a538e2e

* [Sparse] Support Diag sparse format in C++ * update * Update

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Sparse] Support Diag sparse format in C++ #5432

[Sparse] Support Diag sparse format in C++ #5432

czkkkkkk commented Mar 7, 2023

dgl-bot commented Mar 7, 2023

dgl-bot commented Mar 7, 2023

frozenbugs Mar 7, 2023

czkkkkkk Mar 8, 2023

frozenbugs Mar 8, 2023

czkkkkkk Mar 8, 2023

dgl-bot commented Mar 7, 2023

dgl-bot commented Mar 8, 2023

jermainewang Mar 8, 2023

frozenbugs Mar 8, 2023

jermainewang Mar 8, 2023

czkkkkkk Mar 9, 2023

jermainewang Mar 8, 2023

frozenbugs Mar 8, 2023

jermainewang Mar 8, 2023

czkkkkkk Mar 9, 2023

jermainewang Mar 9, 2023

czkkkkkk Mar 9, 2023

dgl-bot commented Mar 9, 2023

[Sparse] Support Diag sparse format in C++ #5432

[Sparse] Support Diag sparse format in C++ #5432

Conversation

czkkkkkk commented Mar 7, 2023

Description

Checklist

Changes

dgl-bot commented Mar 7, 2023

dgl-bot commented Mar 7, 2023

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

dgl-bot commented Mar 7, 2023

dgl-bot commented Mar 8, 2023

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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dgl-bot commented Mar 9, 2023