Skip to content

oneDNN v3,13 migration - #319

Open
azhai219 wants to merge 110 commits into
rls-v3.13from
dev/stock/rls-v3.13-rc4
Open

oneDNN v3,13 migration#319
azhai219 wants to merge 110 commits into
rls-v3.13from
dev/stock/rls-v3.13-rc4

Conversation

@azhai219

@azhai219 azhai219 commented Aug 10, 2026

Copy link
Copy Markdown
Collaborator

Description

Please include a summary of the change. Please also include relevant motivation and context. See contribution guidelines for more details. If the change fixes an issue not documented in the project's Github issue tracker, please document all steps necessary to reproduce it.

Fixes # (github issue)

Checklist

General

  • Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
  • Have you formatted the code using clang-format?

Performance improvements

  • Have you submitted performance data that demonstrates performance improvements?

New features

  • Have you published an RFC for the new feature?
  • Was the RFC approved?
  • Have you added relevant tests?

Bug fixes

  • Have you included information on how to reproduce the issue (either in a github issue or in this PR)?
  • Have you added relevant regression tests?

RFC PR

  • Does RFC document follow the template?
  • Have you added a link to the rendered document?

luweizhou2016 and others added 30 commits August 10, 2026 14:39
@luweizhou2016 @azhai219
[FORK][FEATURE] Support of strided blobs for [de]convolution and simp…
f3b8fbc 
…le reorder

Relaxed mb strides check for FP32/BF16 Convolutions

(cherry picked from commit 118784c)
dmitrygo @azhai219
[FORK][FEATURE] Updated sse41 jit convolutions to support padded chan…
095687a 
…nels

(cherry picked from commit b758837)
dmitrygo @azhai219
[FORK][FEATURE] Introduced Depthwise and Quantization post ops
c31ce13 
Primitives that supports new post ops:
- Jit Convolutions (FP32,BF16,INT8)
- Jit Deconvolution (INT8)
- Jit,Ref Pooling (INT8)

AMX primitives: explicilty pass dst_type into has_default_values checks

Extended int8 AMX convolutions to support depthwise/quantization post ops

ONEDNN 3.2 migration squashed commits:
   - Correct assert for reg id in dw conv kernel

ONEDNN 3.5 squash list:
[FIX] fix avx2 int8 binary postops reg conflict

(cherry picked from commit 52ddb70)
@alexey-varyzgin @azhai219
[FORK][FEATURE] TBB_AUTO was enabled
d2f5910 
(cherry picked from commit 6cc5b1d)
@alexey-varyzgin @azhai219
[FIX] nchw_pooling dense fix
347475a 
(cherry picked from commit d484a99)
dmitrygo @azhai219
[FORK][FEATURE] Enabled BWD (JIT/GEMM) FP32/BF16 Convoltions + Depthw…
78d0d54 
…ise post ops fusings

(cherry picked from commit fbca2ad)
@ilya-lavrenov @azhai219
[FIX] Fixes for MKLDNN to enable LTO
dea6ad4 
(cherry picked from commit e2e3cd0)
mandrono @azhai219
[FIX] [MSVC] Enabling SIMD functionality for VS2019
a7c030b 
(cherry picked from commit e5c847c)
@AlexPeskov @azhai219
[FIX] Add several uni instruction wrappers into jit_generator
79dbb36 
Remove redundancy of two unified instructions

3.5 squash below:
[FIX] Add uni_vhsubps and uni_vaddsubps into jit_generator (#142)

* Add uni_vhsubps and uni_vaddsubps into jit_generator.hpp

* Add missing namespace

* Adjust indentation and add braces

* Improve formatting

* Replace assertion in uni_vhsubps and uni_vhaddps with possible workaround for SSE 4.1

[FIX] added uni_vcvttps2dq

(cherry picked from commit 7a63835)
@AlexPeskov @azhai219
[FIX] Fix name matching with system struct 'user' in llvm-android too…
cc124a6 
…lchain

Note: It is possible that the change is not applicable anymore

Signed-off-by: Alexander Peskov <alexander.peskov@intel.com>
(cherry picked from commit fd4655d)
dmitrygo @azhai219
[FORK][FEATURE] Added JIT FP32/BF16 Softmax for arbitrary inner_size
24704c5 
[WA] Disabled custom JIT Softmax for ndims == 3

(cherry picked from commit a3ca91c)
@a-sidorova @azhai219
[FORK][FEATURE] Added support of hsigmoid, round_half_to_even, round_…
b969223 
…half_away_from_zero elementwise algorithms

Note: hsigmoid implementation probably can be removed since
there is hardsigmoid implementation

(cherry picked from commit be27d5f)
dmitrygo @azhai219
[FIX] Limit applicability of is_1stconv logic for JIT FP32/BF16 AVX51…
0d01177 
…2 Convolution

(cherry picked from commit f896140)
dmitrygo @azhai219
[FIX] [WA] Removed kernel_outside_src condition on JIT FP32/BF16 Conv…
8cc8d61 
…olutions

(cherry picked from commit 3cee5b9)
dmitrygo @azhai219
[FORK][FEATURE] Added custom vesrion of JIT DW FP32/BF16 Convolution …
f7e981c 
…with 5D input support

(cherry picked from commit c05a57c)
dmitrygo @azhai219
[FORK][FEATURE] Asymmetric quntization for activations
77e79ce 
cpu: fix: 1x1 jit conv error
(cherry picked from commit 49fdb83)

fix warning format: [FORK][FEATURE] Asymmetric quntization for activations
dmitrygo @azhai219
[FORK][FEATURE] Added 3D DW case support for JIT INT8 Convolutions
eebb22a 
(cherry picked from commit 05a021b)
dmitrygo @azhai219
[FORK][FEATURE] Added JIT AVX512/AVX2 FP32 Planar Convolution impleme…
3abf859 
…ntation
dmitrygo @azhai219
[FORK][FEATURE] Binary networks support
f61ec03 
(cherry picked from commit 37c8b47)
@myshevts @azhai219
[FIX] Accommodating oneTBB (with hybrid cores support) that
a44b805 
1) needs no custom config on windows
2)TBB_INTERFACE_VERSION defined in the oneapi/tbb folder, while original "tbb" folder comes with stubs

NB: this pre-production ver of the TBB comes only in the VS2019 (vc14) flavor!
(cherry picked from commit 5df277e)
dmitrygo @azhai219
[FIX] [WA] Fixed fallback on ref conv in case exceeding scratchpad limit
cd81419 
(cherry picked from commit edd39a3)
@antonvor @azhai219
[FORK][FEATURE] Returned old behavior for fp32 avx2 1x1 conv with dw …
0857712 
…conv fusing

ONEDNN 3.2 migration squashed commits:
 - Fix jit_avx2_1x1_convolution_with_dw_conv

(cherry picked from commit 18ff7bf)
@a-sidorova @azhai219
[FIX] Updated SoftPlus
2535b05 
(cherry picked from commit 4a737f5)
@IvanNovoselov @azhai219
[FIX] Disable reorder JIT if both inputs and outputs are batch-strided.
cec8af4 
(cherry picked from commit 5d21f96)
@AlexPeskov @azhai219
[FIX] Include TBB headers as system
03c4f08 
There is a problem with compilation with -wundef warning flag.
Include tbb directory by --system alow to avoid checker passes
for TBB headers.

(cherry picked from commit 0b3f4b8)
@luweizhou2016 @azhai219
[FORK][FEATURE] nspc layout support for convolutions
e98ded5 
(cherry picked from commit b25ed00)
@antonvor @azhai219
[FIX] set scale = 1.f in case of signed input on platforms without vnni
949a330 
(cherry picked from commit f55944c)
@maxnick @azhai219
[FIX] Memory descriptor dynamism related changes
9fc2062 
Includes:
- Memory desc wrapper extended with permutation exis extraction
- Fill blocked algorithm has been extracted to a separate funtion
- Moved fill_blocked into memory_desc_wrapper.hpp
- Similar_to changes

(cherry picked from commit fef6e00)
@antonvor @azhai219
[FORK][FEATURE] Added prelu as binary post op
2326e39 
(cherry picked from commit 0f4b3c7)
@antonvor @azhai219
[FORK][FEATURE] Depthwise and Quantization post ops for Gemm Convolut…
1c05b72 
…ions

3.5 squash list:

[FORK][FIX] Fix onednn gemm JIT kernel accuracy issue caused by dest scale.

[FORK][FIX] gemm convolution: fixed bias offset

[FORK][FIX][x64] gemm conv fix for avx2 platforms

(cherry picked from commit 7b499e0)
@azhai219 azhai219 mentioned this pull request Aug 10, 2026
Merged
@azhai219
azhai219 force-pushed the dev/stock/rls-v3.13-rc4 branch from 1b98db9 to 3a924c2 Compare August 10, 2026 07:07
@azhai219
ci: restore upstream cpu github actions (.github + cmake only)
d1501a4 
(cherry picked from commit c98066b)
Excluded: src/ and tests/ changes

ci: restore dnnl_debug_autogenerated.cpp from c98066b

(cherry picked from commit c98066b -- src/common/dnnl_debug_autogenerated.cpp only)

[FORK] Restore DNNL_USE_ACL flag
@azhai219
azhai219 force-pushed the dev/stock/rls-v3.13-rc4 branch from a3c3221 to b489ebe Compare August 12, 2026 01:40
@allnes allnes mentioned this pull request Aug 13, 2026
Merged
@azhai219
azhai219 force-pushed the dev/stock/rls-v3.13-rc4 branch from cb4152d to 1289c3b Compare August 14, 2026 02:12
@allnes allnes mentioned this pull request Aug 17, 2026
Merged
deepak0x pushed a commit to deepak0x/openvino that referenced this pull request Aug 18, 2026
@azhai219 @alvoron
[CPU] migrate onednn rls-v3.13 rc2 (openvinotoolkit#37160)
9860d84 
### Details:
- *forked oneDNN:
https://github.com/openvinotoolkit/oneDNN/tree/dev/stock/rls-v3.13-rc4*
 - *oneDNN PR: openvinotoolkit/oneDNN#319
- *legacy pr openvinotoolkit#36844 is
dropped.*

### Tickets:
 - *191001*

### AI Assistance:
 - *AI assistance used: no / yes*
- *If yes, summarize how AI was used and what human validation was
performed (build/tests/manual checks).*

---------

Co-authored-by: Aleksandr Voron <aleksandr.voron@intel.com>
@azhai219 azhai219 mentioned this pull request Aug 19, 2026
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@allnes allnes allnes approved these changes

Assignees

No one assigned

Labels

backport component:api component:build component:common component:graph-api component:tests devops documentation Improvements or additions to documentation platform:cpu-aarch64 platform:cpu-x64 third_party

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.