Tao Lv

This proposal aims to introduce a set of graph API into oneDNN. Rendered version: [link](https://github.com/TaoLv/mkl-dnn/blob/lvtao/rfcs/graph-api/rfcs/20220711-graph-api/README.md) cc @jianhui-li @igorsafo @mgouicem

API mismatches with latest TensorFlow

1

Great benchmark for evaluting RNN units with different frameworks. But the world is changing fast and TensorFlow has evolved to version 1.4.0rc0. The benchmark for TenforFlow depends on some deprecated...

This is to propose adding OpenCL GPU runtime support in oneDNN Graph API. It addresses the API limitations mentioned RFC https://github.com/oneapi-src/oneDNN/pull/1745 and also paves the way for OpenVINO integration.

rfcs: graph api: support swish operation

2

The RFC proposes adding support for Swish operation in Graph API. Rendered version: [link](https://github.com/TaoLv/mkl-dnn/blob/lvtao/rfcs/swish/rfcs/20241008-graph-api-swish/README.md) .

rfcs: graph api: support safe softmax for sdpa

11

This is to propose supporting safe softmax for sdpa in graph API. A rendered version can be found at: [link](https://github.com/uxlfoundation/oneDNN/blob/lvtao/rfcs/safe-softmax/rfcs/20250321-graph-api-safe-softmax/README.md).

examples: graph: add int8 sdpa example

1

Draft an example for int8 SDPA. MFDNN-13484.

{examples, benchdnn, graph}: add fp8 sdpa support

1

- Added a fp8 (specifically f8_e4m3) sdpa example, with static quantizations. - Exported the JSON file from the example and added into benchdnn inputs as test cases. - Renamed the...

benchdnn: graph: pass acc mode to ref path and more test cases

1

1. Pass the accumulation mode attribute to matmul driver in reference path. 2. Add test cases where both QK and VS are in f16 accumulation mode. Previously only QK in...

Fixes MFDNN-13011 The op kind is not intended for public usage. But with it in the list, we have seen ABI compatibility issue every time when a new op kind...