Rodinia Benchmarks for DPC++

Rodinia benchmarks for CUDA translated to DPC++ using the Intel DPC++ Compatibility Tool. These benchmarks can run in CPU, and NVIDIA/Intel GPUs.

Requirements

Before to run the benchmarks you have to install some dependencies.

• CUDA Toolkit 11.4 → Mandatory to run in NVIDIA GPUs.
• Intel LLVM open source compiler → It will run the code on all the devices.
• Intel oneAPI Base Toolkit. The commercial "dpcpp" compiler cohabits with the previous compiler. However, it does not support NVIDIA GPUs.
• In order to run the "Hybridsort" test, you have to install some OpenGL libraries:

  sudo apt-get install freeglut3 freeglut3-dev
  sudo apt-get install binutils-goldc
  sudo apt-get install libglew1.5

Project Configuration

Once you installed all the requirements, you have to edit the file "common/make.config", changing the value of the following variables:

• CUDA_DIR → set the path where you have installed the CUDA Toolkit (e.g. /usr/local/cuda)
• LLVM_DIR → set the location where you have installed the LLVM compiler (e.g. ~/sycl_workspace/llvm/build)
• ONEAPI_DIR → set the location where you have installed the oneAPI Base Toolkit (e.g. /opt/intel/oneapi)

Compilation

At this point, there are two Makefiles to build the benchmarks, one of them for CUDA benchmarks and another for DPC++ benchmarks.

CUDA Benchmark Compilation

Move to cuda folder and invoke the make command with the following arguments:

• time=<0,1> → Prints the time consumption of the device.

Example:

cd cuda
make time=1

DPC++ Benchmark Compilation

Move to dpcpp folder and invoke the make command with the following arguments:

• time=<0,1> → Prints the time consumption of the device.
• DPCPP_ENV=<clang,oneapi> → The "clang" option uses the LLVM compiler, while the "oneapi" uses the oneAPI compiler.
• DEVICE=<CPU,INTEL_GPU,NVIDIA_GPU> → Selects the device where the code runs. The "NVIDIA_GPU" option just works selecting the variable "DPCPP_ENV=clang".

The following example compiles the benchmarks using the LLVM compiler, selects the NVIDIA GPU, and choose to show the GPU time consumption:

cd dpcpp
make DPCPP_ENV=clang DEVICE=NVIDIA_GPU time=1

How to run it?

You can run them one by one, or use the scripts we provide ("time_cuda.sh", "time_dpcpp.sh"), which save the kernel time in a "timing" folder. For that, you had to compile them with the "time=1" argument.

Known Issues

Benchmarks not working in DPC++

The following benchmarks does not work in DPC++:

• Hybridsort
• Kmeans
• Leukocyte
• Mummergpu

Publications

• Germán Castaño, Youssef Faqir-Rhazoui, Carlos García and Manuel Prieto-Matías (2022). "Evaluation of Intel’s DPC++ Compatibility Tool in Heterogeneous Computing." Journal of Parallel and Distributed Computing, volume 165, pages 120-129.
  • Free available here.

Acknowledgements

This work has been supported by the EU (FEDER) and the Spanish MINECO and CM under grants S2018/TCS-4423 and RTI2018-093684-B-I00.