Optimize fusedL2NN when data is skinny #794
Merged
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
ahendriksen
added
improvement
ahendriksen
force-pushed
the
enh-faster-fusedl2nn
branch
from
67e1328 to
1d46f38
Compare
ahendriksen
force-pushed
the
enh-faster-fusedl2nn
branch
2 times, most recently
from
da8fa38 to
16a6fc9
Compare
|
rerun tests |
There were very few test cases for skinny matrices. They have now been added.
In updateReDucedVal, a single warp can contain multiple rows (in registers). A single thread within the warp uses the first element of each row to update an output array (atomically). In the previous implementation, a shuffle was used to move the head of each row into the first thread of the warp. Unfortunately, this would overwrite the value all other rows. This strategy, however, worked when the number of rows per warp equalled 2. Hence, the bug never triggered. In a recent commit, the number of rows per warp was increased to four in certain situations (skinny matrices). Hence, this bug triggered. In the new implementation, the values are not shuffled into the first thread of the warp any more. Instead, the threads that contain the first element of a row update the output in sequential order. The sequential ordering is necessary to avoid deadlock on Pascal architecture.
In the current implementation, it looks like values from different rows are mixed together in what should be a row-wise warp reduce. All tests do pass however. Just in case, I have added a width parameter to the shuffle so that it only shuffles within a row within the warp.
ahendriksen
force-pushed
the
enh-faster-fusedl2nn
branch
from
16a6fc9 to
1c6fa14
Compare
tfeher
requested changes
Sep 5, 2022
tfeher
reviewed
Sep 5, 2022
There was a problem with defgroup syntax.
ahendriksen
force-pushed
the
enh-faster-fusedl2nn
branch
from
1c6fa14 to
36e5865
Compare
tfeher
approved these changes
Sep 6, 2022
|
@gpucibot merge |
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.
The fusedL2NN kernel uses tiling to maximize performance. The current implementation assumes that the input matrices are at least 32 elements wide. When this is not the case, it performs redundant computations.
This PR adds a policy to apply when the matrix is skinny (less than 32 elements wide). This results in a 1.5 - 2x performance improvement across GPU architectures.