DPBench - Numba/Numba-Dpex/dpnp/DPCPP Benchmarks

• _numba.py : This file contains Numba implementations of the benchmarks. There are three modes: nopython-mode, nopython-mode-parallel and nopython-mode-parallel-range.
• numba_dpex.py : This file contains Numba-Dpex implementations of the benchmarks. There are three modes: kernel-mode, numpy-mode and prange-mode.
• _*dpnp.py : This file contains dpnp implementations of the benchmarks.
• _native_ext/__sycl/_kernel.hpp : This file contains native dpcpp implementations of the benchmarks.

Examples of setting up and running the benchmarks

1. Setting up conda environment and installing dependencies:

    $ conda create -n dpbench-dev
    $ conda activate dpbench-dev
    $ conda install python=3.9
    $ conda install -c intel tbb dpcpp_linux-64
    $ conda install numpy numba cython cmake ninja scikit-build pandas
    $ conda install scipy scikit-learn pybind11
    # do not miss the quotes!
    $ conda install -c pkgs/main libgcc-ng">=11.2.0" libstdcxx-ng">=11.2.0" libgomp">=11.2.0"
    $ conda install -c intel dpctl numba-dpex dpnp
   

2. Build and run DPBench

   • To build:

     $  CC=icx CXX=icpx python setup.py develop -- -Dpybind11_DIR=$(python -m pybind11 --cmakedir) -DDPCTL_MODULE_PATH=$(python -m dpctl --cmakedir)

   • To run, taking black_scholes for example:

     $  python -c "import dpbench; dpbench.run_benchmark(\"black_scholes\")"

3. Device Customization

   If a framework is SYCL based, an extra configuration option sycl_device may be set in the framework JSON file to control what device the framework uses for execution. The sycl_device value should be a legal SYCL device filter string. The dpcpp, dpnp, numba_dpex_kernel, and numba_dpex frameworks support the sycl_device option.

   Here is an example:

       {
           "framework": {
               "simple_name": "dpcpp",
               "full_name": "dpcpp",
               "prefix": "dp",
               "postfix": "dpcpp",
               "class": "DpcppFramework",
               "arch": "gpu",
               "sycl_device": "level_zero:gpu:0"
           }
       }

   NOTE: The arch option is deprecated and not used by dpbench.

   To run with custimized framework JSON file, pass it as an argument to the run_benchmark or run_benchmarks functions.

   $ python -c "import dpbench; dpbench.run_benchmark(\"black_scholes\", "<absolute path to json file>")"