Adapting oneDNN for KDNN

This section describes how to adapt oneDNN for KDNN. Misoperations may introduce undefined behaviors. Exercise caution when performing this operation.

Adaptation Procedure

Obtain the oneDNN adaptation code. Assume that /path/to is the directory where the source code is cloned.

git clone -b v3.0.0 https://gitee.com/openeuler/kail_dnn_adapter
cd kail_dnn_adapter
git submodule update --init --recursive
cd oneDNN-open
ulimit -n 262144
patch -p1 < ../0001-kdnn-adapter.patch

Access /path/to/kail_dnn_adapter and compile oneDNN.
- For Kunpeng 920 7280Z:
```
cd /path/to/kail_dnn_adapter
sh build.sh --use_static_kdnn=off
```
  The --use_static_kdnn=on/off option specifies whether to use the KDNN static library or dynamic library during compilation. The default value is off.
- For new Kunpeng 920 processor model (the following example is based on the BiSheng Compiler):
```
cd /path/to/kail_dnn_adapter
sh build.sh --compiler=clang
```
Path of the compilation product:

Path of libdnnl.so: out/oneDNN-open/build/src/
Paths to the dependency libraries:
- Path to .so files related to ACL: out/ComputeLibrary-23.11/build/
- Path to .so files related to KAIL: /usr/local/kdnn/lib/omp/libkdnn.so
You can use all interfaces of oneDNN v3.4.0 by linking against libdnnl.so.

In the compile command, --compiler=clang indicates that the BiSheng Compiler is used. By default, GCC is used.

Verification After Adaptation

After compiling oneDNN, use the test cases included in the software package to verify the adaptation.

Go to the /path/to/kail_dnn_adapter/out/llt/scripts directory.
```
cd /path/to/kail_dnn_adapter/out/llt/scripts
```
Run the test cases.
```
python run_daily_build.py --working_dir=../../oneDNN-open/build/tests/benchdnn
```
If all results returned passed and the following information is displayed, oneDNN is successfully adapted.

Parent topic: Using KAIL