webgpu: Fix bank conflicts #6152
Closed
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
qjia7
force-pushed
the
fix_bank_conflicts
branch
from
2159f25 to
c685ff5
Compare
qjia7
commented
Feb 15, 2022
There was a problem hiding this comment.
@xhcao @axinging @haoyunfeix @gyagp Please take a look, thanks.
| @@ -31,7 +31,7 @@ ENV.registerFlag('WEBGPU_CPU_FORWARD', () => true); | |||
| /** | |||
| * Thread register block size for matmul kernel. | |||
| */ | |||
| ENV.registerFlag('WEBGPU_MATMUL_WORK_PER_THREAD', () => 4); | |||
| ENV.registerFlag('WEBGPU_MATMUL_WORK_PER_THREAD', () => 2); | |||
There was a problem hiding this comment.
I changed the workgroup size from (8, 8, 1) to (16, 16, 1) to better align with bank size 16. However, due to the maximum shared memory size limitation, I have to change the work per thread size to 2.
There was a problem hiding this comment.
Remove this part of changes to make sure this PR only focuses on fixing bank conflicts. Please take another look, thanks.
qjia7
force-pushed
the
fix_bank_conflicts
branch
from
c685ff5 to
0b35bec
Compare
|
close it by #6862 |
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Outdated suggestions cannot be applied.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.
This PR fixed the shared memory bank conflicts issue on MatMulPackedProgram. Previously, it used interleaved loads, which resulted the bank conflict. This PR uses sequential accessing to resolve it.
Before:
After:
With this change, Conv2DDerInputMMProgram/Conv2DMMProgram/MatMulPackedProgram get about >20% improvement. For example, the total time of Conv2DBackpropInput in hand_detector becomes 7.46ms from 11.26ms. The total time Conv2D in AutoML Image becomes 4.05ms from 5.61ms.
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is