[HOTFIX] Remove duplicated group divide when calculation k per group (SWDEV-291574) #989
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
atamazov
changed the title
atamazov
changed the title
|
Regression test is missing. I am not going to delay merging this PR (and cherry-picking it to 4.3), for the sake of speed. But if this PR does not contain a regression test, then I'll open a high priority ticket for this. |
This comment has been minimized.
This comment has been minimized.
atamazov
pushed a commit
that referenced
this pull request
Jun 20, 2021
atamazov
pushed a commit
that referenced
this pull request
Jul 22, 2021
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Outdated suggestions cannot be applied.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 resolve SWDEV-291574 issue by removing the duplicated
groupdivide while calculating k_per_block, sincekalready divided bygroupin above L796.After this fix, the original problematic config
./bin/MIOpenDriver convfp16 -n 64 -c 256 -H 56 -W 56 -k 256 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1 -l 1 -j 1 -m conv -g 32 -F 2 -t 1should not be valid for this asm solver, sincec_per_bloc = 256/32 = 8,k_per_bloc = 256/32 = 8, this is too small for NCHW asm igemm to support (NCHW fp16 does not support pack c or k)Beside, if reduce group size to like 2, 4, 8, 16, the solver can successfully calculate the result.
./bin/MIOpenDriver convfp16 -n 64 -c 256 -H 56 -W 56 -k 256 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1 -l 1 -j 1 -m conv -g 4 -F 2 -t 1./bin/MIOpenDriver convfp16 -n 64 -c 256 -H 56 -W 56 -k 256 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1 -l 1 -j 1 -m conv -g 8 -F 2 -t 1./bin/MIOpenDriver convfp16 -n 64 -c 256 -H 56 -W 56 -k 256 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1 -l 1 -j 1 -m conv -g 16 -F 2 -t 1