Extending support for binary primitive

Open TejaX-Alaghari opened this issue 3 years ago • 2 comments

Description

This PR extends binary SYCL kernel support for non-uniform group sizes. This includes a new logic for work-item config in kernel launch and handling the trailing portions of workspace. In addition, the PR adds support for common scales and handles saturation and rounding for vectors.

Checklist

General

[x] Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit? test_binary_all.txt
[x] Have you formatted the code using clang-format?

Performance improvements

[x] Have you submitted performance data that demonstrates performance improvements? oneDNN Performance Tracker.xlsx

Jul 28 '22 00:07 TejaX-Alaghari

THank you for the PR @TejaX-Alaghari . Could you share on which platform(s) you validated this change?

Sep 02 '22 08:09 mgouicem

THank you for the PR @TejaX-Alaghari . Could you share on which platform(s) you validated this change?

Validation is performed on Nvidia Tesla T4 GPU. Attached the clinfo and nvidia-smi logs for ref. Let me know if any further info is required. nvidia_clinfo.txt nvidia_smi_info.txt

Sep 02 '22 10:09 TejaX-Alaghari

Closing this outdated PR and raised a new PR https://github.com/oneapi-src/oneDNN/pull/1612.

Mar 25 '23 09:03 TejaX-Alaghari