Skip to content

[GPU] Update condition of winograd kernel selection #33681

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
p-durandin merged 1 commit into from
Mar 18, 2026
+2 −0
Merged

[GPU] Update condition of winograd kernel selection #33681

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -500,6 +500,8 @@ bool should_use_winograd_2x3_s1(const convolution_node& node,
|| weights_layout.batch() % 64 != 0 // current algorithm is effective for ofm to be multiply of 64
|| any_not_one(prim->stride) // stride has to be 1x1 by definition
|| any_not_one(prim->dilation) // no support for dilation
|| !all_zeroes(prim->padding_begin) // no padding supported. padding could makes higher accuracy loss.
|| !all_zeroes(prim->padding_end) // no padding supported. padding could makes higher accuracy loss.
Copy link
Contributor

@byungilm byungilm Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this difficult to add relevant test-cases?

Copy link
Contributor Author

@wilson-seok wilson-seok Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@byungilm It's hard to create a proper test case about winograd kernel selection. Because the kernel requires byxf format and byfx format selection in get_preferred_format() depends on not only should_use_winograd_2x3_s1() but also +2 depth of convolution_byxf_opt().

|| output_size_handling_enabled // This condition is weird. Need to revise it and replace with something meaningful
|| (input_layout.count() > 3000000) // limit max input size as winograd consumes more memory
|| (input_layout.count() < 50000) // limit min input size as winograd is not effective for small input
Expand Down
Loading