Skip to content

[GPU] Add dropout support for Ref Matmul #4494

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

Jump to bottom
Merged
h-sadia merged 6 commits into from
Jan 7, 2026
Merged

[GPU] Add dropout support for Ref Matmul #4494

h-sadia merged 6 commits into from
Jan 7, 2026

Conversation

@h-sadia
Copy link
Contributor

@h-sadia h-sadia commented Jan 5, 2026

This is the GPU side of the adding s64 offset and seed support for frameworks. The CPU/Arch part was merged before holidays here: #4031

@h-sadia h-sadia requested review from a team as code owners January 5, 2026 23:33
@github-actions github-actions bot added platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel component:tests Codeowner: @oneapi-src/onednn-arch labels Jan 5, 2026
@h-sadia h-sadia force-pushed the hsadia/dropout_mm_main branch 5 times, most recently from 9b5d3dc to 1ac0545 Compare January 6, 2026 01:07
@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 6, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_matmul
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg
enable arch_gpu_xe3-lpg

dzarukin
dzarukin reviewed Jan 6, 2026
View reviewed changes
@h-sadia h-sadia force-pushed the hsadia/dropout_mm_main branch from 1ac0545 to 50c012b Compare January 6, 2026 19:05
@h-sadia h-sadia requested a review from a team as a code owner January 6, 2026 19:05
@h-sadia h-sadia force-pushed the hsadia/dropout_mm_main branch from 50c012b to 4dc9fd9 Compare January 6, 2026 20:02
@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 6, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_matmul
enable arch_gpu_xe-hpc

dzarukin
dzarukin approved these changes Jan 6, 2026
View reviewed changes
Copy link
Contributor

@dzarukin dzarukin left a comment

Choose a reason for hiding this comment

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

LGTM, left some minor comments.

src/gpu/intel/matmul/ref.cl
float po_acc = convert_float(temp);

#if WITH_DROPOUT
#if USE_OFFSET
Copy link
Contributor

@dzarukin dzarukin Jan 6, 2026

Choose a reason for hiding this comment

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

I think this should check if SEED_DATA_T is int or int64_t. If it's not, then the assumption is the offset won't be passed anyway. If it is, then use a new version and if offset wasn't specified, the kernel sets it to 0.

Copy link
Contributor Author

@h-sadia h-sadia Jan 6, 2026

Choose a reason for hiding this comment

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

good catch! the offset will already be set to zero if use_offset = 0, but this condition is misleading.

Copy link
Contributor Author

@h-sadia h-sadia Jan 6, 2026

Choose a reason for hiding this comment

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

https://github.com/uxlfoundation/oneDNN/blob/main/src/cpu/primitive_attr_postops.cpp#L304 - based on this condition use_offset is the precursor between the two philox functions used. Will stick with what I have by adding explicit conversion to seed for philox without offset. Changing condition will mean I will have to add another similar function to cater for s64 philox without seed.

Copy link
Contributor

@kealan-barbieri kealan-barbieri Jan 7, 2026

Choose a reason for hiding this comment

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

Alternatively you could have two versions of philox_4x32 with different type signatures so it would dispatch implicitly based on type.

Copy link
Contributor Author

@h-sadia h-sadia Jan 7, 2026

Choose a reason for hiding this comment

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

We strictly like to keep the code between cpu/gpu same. Additionally if I do add another philox function, it will be just repeating the same code with lesser lines.

@h-sadia h-sadia force-pushed the hsadia/dropout_mm_main branch 3 times, most recently from 115b8aa to 5701436 Compare January 6, 2026 23:49
@h-sadia h-sadia force-pushed the hsadia/dropout_mm_main branch from 5701436 to 417e650 Compare January 7, 2026 01:20
@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 7, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_matmul
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg
enable arch_gpu_xe3-lpg

kealan-barbieri
kealan-barbieri reviewed Jan 7, 2026
View reviewed changes
kealan-barbieri
kealan-barbieri reviewed Jan 7, 2026
View reviewed changes
tests/benchdnn/inputs/matmul/harness_matmul_dropout
--check-ref-impl=false

--attr-dropout=0.5:12345678,0.75:12345678:undef,0.25:843921:any:1238976:true
--attr-dropout=0.5:12345678,0.75:12345678:undef,0.25:843921:any:1238976:true,0.75:111786:any:121716:false
Copy link
Contributor

@kealan-barbieri kealan-barbieri Jan 7, 2026

Choose a reason for hiding this comment

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

are we able to specify s32/s64 seed type from benchdnn? it might be worth verifying both if theyre still supported.

Copy link
Contributor Author

@h-sadia h-sadia Jan 7, 2026

Choose a reason for hiding this comment

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

So benchdnn only supports s64 as a standard. We are keeping s32 for older test-pathways. I checked the s32 datatype here by making the change here: https://github.com/uxlfoundation/oneDNN/blob/main/tests/benchdnn/dnn_types.cpp#L1449-L1454
(and all the tests are passing for harness_matmul_dropout)

@h-sadia h-sadia force-pushed the hsadia/dropout_mm_main branch from 417e650 to d04c1ad Compare January 7, 2026 03:39
@h-sadia h-sadia merged commit 7513f1f into main Jan 7, 2026
32 checks passed
@h-sadia h-sadia deleted the hsadia/dropout_mm_main branch January 7, 2026 05:21
@h-sadia h-sadia mentioned this pull request Jan 7, 2026
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@dzarukin dzarukin dzarukin approved these changes

@kealan-barbieri kealan-barbieri kealan-barbieri approved these changes

Assignees

No one assigned

Labels

component:tests Codeowner: @oneapi-src/onednn-arch platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@h-sadia @dzarukin @kealan-barbieri