Accelerate adjacency matrix to CSR conversion for DBSCAN #4803
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
This PR is a follow up to PR #4776, which also involves DBSCAN. |
ahendriksen
added
improvement
tfeher
requested changes
Jul 7, 2022
There was a problem hiding this comment.
Thanks @ahendriksen for this work! I have marked a few issue that shall be addressed. Overall it looks good, and I am excited to see this improvement.
ahendriksen
force-pushed
the
fea-2387-improve-perf-dbscan-adj-csr
branch
from
167dfd6 to
f68e6fa
Compare
For large data sizes, the batch size of the DBSCAN algorithm is small in order to fit the distance matrix in memory. This results in a matrix that has dimensions num_points x batch_size, both for the distance and adjacency matrix. The conversion of the boolean adjacency matrix to CSR format is performed in the 'adjgraph' step. This step was slow when the batch size was small, as described in issue rapidsai#2387. In this commit, the adjgraph step is sped up. This is done in two ways: 1. The adjacency matrix is now stored in row-major batch_size x num_points format --- it was transposed before. This required changes in the vertexdeg step. 2. The csr_row_op kernel has been replaced by the adj_to_csr kernel. This kernel can divide the work over multiple blocks even when the number of rows (batch size) is small. It makes optimal use of memory bandwidth because rows of the matrix are laid out contiguously in memory.
Temporary commit to check if CI passes with new raft warp-aggregated atomics.
ahendriksen
force-pushed
the
fea-2387-improve-perf-dbscan-adj-csr
branch
from
f68e6fa to
f6caeab
Compare
cjnolet
requested changes
Jul 14, 2022
There was a problem hiding this comment.
@ahendriksen overall this good to me, minus moving the csr_to_adj back over to RAFT. I also left a little reminder to update the raft pin now that the PR for the atomic utilty function has been moved.
|
rerun tests |
cjnolet
approved these changes
Jul 22, 2022
tfeher
approved these changes
Jul 22, 2022
tfeher
changed the title
|
@gpucibot merge |
Codecov Report
@@ Coverage Diff @@
## branch-22.08 #4803 +/- ##
=============================================
Coverage 77.62% 77.62%
=============================================
Files 180 180
Lines 11382 11384 +2
=============================================
+ Hits 8835 8837 +2
Misses 2547 2547
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
|
@gpucibot merge |
jakirkham
pushed a commit
to jakirkham/cuml
that referenced
this pull request
Feb 27, 2023
Fixes issue rapidsai#2387. For large data sizes, the batch size of the DBSCAN algorithm is small in order to fit the distance matrix in memory. This results in a matrix that has dimensions num_points x batch_size, both for the distance and adjacency matrix. The conversion of the boolean adjacency matrix to CSR format is performed in the 'adjgraph' step. This step was slow when the batch size was small, as described in issue rapidsai#2387. In this commit, the adjgraph step is sped up. This is done in two ways: 1. The adjacency matrix is now stored in row-major batch_size x num_points format --- it was transposed before. This required changes in the vertexdeg step. 2. The csr_row_op kernel has been replaced by the adj_to_csr kernel. This kernel can divide the work over multiple blocks even when the number of rows (batch size) is small. It makes optimal use of memory bandwidth because rows of the matrix are laid out contiguously in memory. Authors: - Allard Hendriksen (https://github.com/ahendriksen) - Corey J. Nolet (https://github.com/cjnolet) Approvers: - Corey J. Nolet (https://github.com/cjnolet) - Tamas Bela Feher (https://github.com/tfeher) URL: rapidsai#4803
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes issue #2387.
For large data sizes, the batch size of the DBSCAN algorithm is small in order to fit the distance matrix in memory.
This results in a matrix that has dimensions num_points x batch_size, both for the distance and adjacency matrix.
The conversion of the boolean adjacency matrix to CSR format is performed in the 'adjgraph' step. This step was slow when the batch size was small, as described in issue #2387.
In this commit, the adjgraph step is sped up. This is done in two ways:
The adjacency matrix is now stored in row-major batch_size x num_points format --- it was transposed before. This required changes in the vertexdeg step.
The csr_row_op kernel has been replaced by the adj_to_csr kernel. This kernel can divide the work over multiple blocks even when the number of rows (batch size) is small. It makes optimal use of memory bandwidth because rows of the matrix are laid out contiguously in memory.