[GPU] Fix modnet_webcam_portrait_matting low similarity. #33370
Merged
Conversation
This file contains hidden or 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
Regression from openvinotoolkit#32236 Signed-off-by: hyunback <[email protected]>
hyunback
force-pushed
the
fix_modnet_webcam_portarit_matting_bad_acc
branch
from
8ecdb78 to
d2e6e94
Compare
isanghao
changed the title
isanghao
reviewed
Dec 30, 2025
| (fmt_prev == format::b_fs_yx_fsv4 && | ||
| prev_output_layout.feature() % 32 == 0 && | ||
| prev_output_layout.spatial(0) == 1 && | ||
| prev_output_layout.spatial(1) == 1)) && is_input_reorder(prev, next)) |
Contributor
There was a problem hiding this comment.
random spot) the PR description is difficult to understand. Let's discuss offline.
Contributor
There was a problem hiding this comment.
- could you improve the PR description? I think the problem itself(problematic pattern) is not described enough.
isanghao
reviewed
Jan 2, 2026
| (fmt_prev == format::b_fs_yx_fsv4 && | ||
| prev_output_layout.feature() % 32 == 0 && | ||
| prev_output_layout.spatial(0) == 1 && | ||
| prev_output_layout.spatial(1) == 1)) && is_input_reorder(prev, next)) |
Contributor
There was a problem hiding this comment.
- could you improve the PR description? I think the problem itself(problematic pattern) is not described enough.
| if (fmt_next == format::b_fs_yx_fsv16 || fmt_next == format::b_fs_zyx_fsv16 || | ||
| fmt_next == format::bs_fs_yx_bsv16_fsv16 || fmt_next == format::b_fs_yx_fsv4) | ||
| return true; | ||
| } |
Contributor
There was a problem hiding this comment.
I think I'm not following the logic here. If cross-layout quantization is executed, performance drops. Then I think it is better to disable cross-layout. Does OneDNN concat selection happen because of the quantization behavior?
Contributor
Author
There was a problem hiding this comment.
Yes, the different graph because of quantize and reorder fusion causes to run onednn concat. onednn concat bad accuracy is not fixed yet.
I think this PR is for restoring refactoring mistake , so It seems better to proceed separately if need. Honestly, Perf improvement shows only for this model(so special) and basically leaving the reorder seems not good.
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
github-merge-queue bot
pushed a commit
that referenced
this pull request
Jan 6, 2026
Regression from #32236 There are some refactoring code in #32236 (Too many if-else branches with no rules in reorder fusion function, probably just added by many people over many times, so refactoring did clean up by primitive based) But in refactoring, there are some mistake in quantize, it caused to change in reorder-fusion rule. - Previous refactoring: Do reorder fusion for quantize and reorder pattern - Refactoring Do reorder fusion for quantize and reorder pattern only quantize' s input format is plain. Because this reason, the graph has been changed (Reorder is not fused now) and it causes unintentional use of onednn concat. It occurs the poor similarity issue. This PR restores previous reorder-fusion pattern which was before refactoring so address the low similarity. And separate from the issue resolution, the onednn concat accuracy problem which was discovered accidentally due to the use of onednn concat, was issued by [MFDNN-14504 ](https://jira.devtools.intel.com/browse/MFDNN-14504) ### Tickets: - *178661* Signed-off-by: hyunback <[email protected]>
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Regression from #32236
There are some refactoring code in #32236
(Too many if-else branches with no rules in reorder fusion function, probably just added by many people over many times, so refactoring did clean up by primitive based)
But in refactoring, there are some mistake in quantize, it caused to change in reorder-fusion rule.
Do reorder fusion for quantize and reorder pattern
Do reorder fusion for quantize and reorder pattern only quantize' s input format is plain.
Because this reason, the graph has been changed (Reorder is not fused now) and it causes unintentional use of onednn concat. It occurs the poor similarity issue.
This PR restores previous reorder-fusion pattern which was before refactoring so address the low similarity. And separate from the issue resolution, the onednn concat accuracy problem which was discovered accidentally due to the use of onednn concat, was issued by MFDNN-14504
Tickets: