Skip to content

graph: backend: dnnl: executables: add exec profile verbose for gen_index#4849

Open
TaoLv wants to merge 1 commit intomainfrom
lvtao/main/gen-index-verbose
Open

graph: backend: dnnl: executables: add exec profile verbose for gen_index#4849
TaoLv wants to merge 1 commit intomainfrom
lvtao/main/gen-index-verbose

Conversation

@TaoLv
Copy link
Contributor

@TaoLv TaoLv commented Mar 18, 2026

Fixes MFDNN-14511

Add execute profile verbose for gen_index kernel which currently is not a standard primitive.
It helps for SDPA cases performance profiling and debugging.

The verbose line looks like below:

onednn_verbose,v1,graph,exec,gpu,_gen_index(_gen_index),18446744073709551615,attrs:{axis:2},inputs:{f32:102:strided:undef:1x16x384x384:2359296x147456x384x1},outputs:{s32:1021:strided:undef:1x16x384x384:2359296x147456x384x1},0.193115

@TaoLv TaoLv requested a review from a team as a code owner March 18, 2026 08:06
@github-actions github-actions bot added the component:graph-api Codeowner: @oneapi-src/onednn-graph label Mar 18, 2026
@TaoLv TaoLv force-pushed the lvtao/main/gen-index-verbose branch from 76e3a67 to 8a7c81a Compare March 18, 2026 08:33
@karturov
Copy link
Contributor

@TaoLv, can you please rebase your PR on top the new training support in #4814? I tried to cherry-pick, and it didn't work on my side with compilation errors in gen_index.

@TaoLv
Copy link
Contributor Author

TaoLv commented Mar 19, 2026

@TaoLv, can you please rebase your PR on top the new training support in #4814? I tried to cherry-pick, and it didn't work on my side with compilation errors in gen_index.

#4814 is rebased to the latest main. Now these two PRs can be picked together without conflicts. I prepared a branch for that: lvtao/yixin/sdpa_ukernel_train-genindex_verb.

@TaoLv
Copy link
Contributor Author

TaoLv commented Mar 19, 2026

make test
disable benchdnn_all
enable benchdnn_graph

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

component:graph-api Codeowner: @oneapi-src/onednn-graph

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants