Qianshui-Jiang / zbench

A simple CM kernel launchpad

zbench: A simple CM kernel launchpad

Introduction

• Using dxd12 runtime and intel graphics dev extension to compile and running a CM shader,
• Using Pybind11 to construct the interface, which means all the IO tensors, build options, shader params are setup from python file.
  • All the IO buffers are in Numpy array format. and direct setup as the kernel input arguments.
  • All the Hyper parameters are setup through C/C++ Macros, which means contained in build options in format like -Dxxxx=123
  • the param name and value are flexible, except that buffers should follow the order of your CM kernel input.
  • 3 necessary args should be given in a fixed name: cm_file, build_options, and iter_nums
• Assign the iteration numbers by iter_nums, and usually put it in the last arg of launch_rt_igdext()
• For more information, please find out the examples in zbench/tests:

Option1. Build from source

1. you need have oneAPI base toolkit installed in your windows machine. after installed, you can find setvars.bat in your oneAPI target path, and run following cmd in your terminal shell:

   cmd "/K" '"{Your OneAPI Path}\setvars.bat" && powershell'

2. create a python environment(here using conda), and activate it:

   conda create -n kernel_dev python==3.10.13
   conda activate kernel_dev
   pip install -r requirements

3. clone the repo:

   git clone https://github.com/Qianshui-Jiang/zbench
   git submodule init
   git submodule update 
   

4. entering zbench folder and build or install it:

   cd zbench
   // direct build & install 
   python setup.py install
   // only build
   python setup.py build

5. if you want build python wheels:

   python setup.py bdist_wheel

   the wheel files would be exists in folder zbench/dist

Option2. Install already built python wheel directly

1. create a python environment(here using conda), and activate it:

   conda create -n kernel_dev python==3.10.13
   conda activate kernel_dev
   pip install -r requirements

2. install the zbench wheel:

   pip install zbench-0.0.0-cp310-cp310-win_amd64.whl

Try the example:

• Step 1. exporting reference tensor by dxdispatch

python dxdispatch_dump_data_llama2.py

• Step 2. test MHA with exported ref data

python test_flash_att_llm.py

TODOs:

• Add MHA tests for SD's shape
• Add online softmax tests.
• Add GEMV tests.
• Add GEMM tests.
• More runtime support. （L0 / OCL）
• Hardware performance monitor counter support.