openvino
[GPU] Optimize MVN and reorders for nnUNet INT8 5D model
#34949
Merged

[GPU] Optimize MVN and reorders for nnUNet INT8 5D model #34949

sungeunk
sungeunk sungeunk added category: GPU
sungeunk sungeunk marked this pull request as ready for review 64 days ago
sungeunk sungeunk requested a review 64 days ago
sungeunk sungeunk requested a review 64 days ago
sungeunk sungeunk force pushed from c71b37bd to 0caa9041 54 days ago
sungeunk sungeunk force pushed from 0caa9041 to 96bcea7f 53 days ago
sungeunk sungeunk changed the title [GPU] prevent oneDNN deconv from selecting slow ocl:ref kernel [GPU] Optimize oneDNN deconv kernel selection and layout for MVN-adjacent patterns 52 days ago
sungeunk
hyunback
hyunback commented on 2026-04-10
hyunback
hyunback commented on 2026-04-10
sungeunk sungeunk requested a review from hyunback hyunback 47 days ago
ahnyoung-paul
ahnyoung-paul commented on 2026-04-14
sungeunk sungeunk changed the title [GPU] Optimize oneDNN deconv kernel selection and layout for MVN-adjacent patterns [GPU] Enable MVN fsv16/fsv32 cross-layout fusing to eliminate redundant reorders 43 days ago
sungeunk sungeunk requested a review from isanghao isanghao 43 days ago
sungeunk sungeunk requested a review from ahnyoung-paul ahnyoung-paul 43 days ago
hyunback
hyunback commented on 2026-04-17
hyunback
hyunback commented on 2026-04-17
hyunback
hyunback commented on 2026-04-17
ahnyoung-paul
ahnyoung-paul commented on 2026-04-17
sungeunk sungeunk changed the title [GPU] Enable MVN fsv16/fsv32 cross-layout fusing to eliminate redundant reorders [GPU] Optimize MVN and reorders for nnUNet INT8 5D model 40 days ago
sungeunk sungeunk force pushed from 41be3e0e to ad7b116e 39 days ago
hyunback
hyunback commented on 2026-04-24
sungeunk sungeunk force pushed from a8520bc6 to 2bbcf031 32 days ago
sungeunk sungeunk requested a review 32 days ago
sungeunk sungeunk removed review request 32 days ago
sungeunk sungeunk requested a review from ValentinaKats ValentinaKats 32 days ago
github-actions github-actions added category: docs
sungeunk sungeunk requested a review from ahnyoung-paul ahnyoung-paul 32 days ago
sungeunk sungeunk requested a review from hyunback hyunback 32 days ago
sungeunk
ahnyoung-paul
ahnyoung-paul commented on 2026-04-28
sungeunk sungeunk requested a review from ahnyoung-paul ahnyoung-paul 31 days ago
hyunback
sungeunk
hyunback
hyunback approved these changes on 2026-04-30
ahnyoung-paul
ahnyoung-paul approved these changes on 2026-04-30
sungeunk sungeunk force pushed from a6b1ed06 to 166da3b6 25 days ago
sungeunk test(GPU): add deconv quantize fusion and kernel-selection tests
5d9d2dfb
sungeunk [GPU] prevent oneDNN deconv from selecting slow ocl:ref kernel
3123e836
sungeunk test(GPU): add MVN fsv16/fsv32 cross-layout and blocked-format tests
e8162e5f
sungeunk feat(GPU): enable MVN fsv16/fsv32 cross-layout fusing to reduce reorders
1221d60a
sungeunk test(GPU): add dynamic-shape cases for bfyx_to_blocked_format reorder
35baefe8
sungeunk feat(GPU): add dynamic shape support for bfyx_to_blocked_format reord…
e6a5411c
sungeunk test(GPU): add tests for blocked<->blocked fsv reorder kernel
8d3ccfa8
sungeunk feat(GPU): add reorder_data_fsv kernel for blocked<->blocked fsv conv…
79a153d5
sungeunk perf(GPU): add vload/vstore vectorization to reorder_data_fsv kernel
589107b2
sungeunk test(GPU): update MVN/reorder tests for kernel label and dynamic regi…
9d631e8c
sungeunk refactor(GPU): rename MVN _imad kernel and extend dynamic reorder reg…
746f4f66
sungeunk docs(GPU): explain dynamic-shape reorder fusion exception for MVN
29678db3
sungeunk fix(GPU): remove unused intermediate_bytes in MVN fsv16 dynamic multi…
ca1ac703
sungeunk remove unnecessary code
52d4b8d8
sungeunk sungeunk force pushed from 166da3b6 to 52d4b8d8 24 days ago
e-ddykim e-ddykim enabled auto-merge 24 days ago
e-ddykim e-ddykim merged 29da988d into master 23 days ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone