oneAPI on CUDA setup script

This script allows you to build and setup the DPC++ compiler, oneMKL and oneDNN with Nvidia GPUs support.

Use

1. If needed, set $DPCPP_HOME and $CUDA_ROOT in the script.
2. Install optionally the Intel(R) OpenCL Runtime with sudo ./opencl.sh
3. Run the script with ./build.sh.

Everything will be installed to $DPCPP_HOME/deploy.

To build with testing support and run the tests for DPC++, oneMKL, oneTBB, oneDNN and Lapack, run: DPCPP_TESTS=ON ./build.sh

If you want to build libc++, use CC=clang CXX=clang++ with another version of clang, gcc won't compile the libc as the asm syntax is not the same. Using this clang/dpc++ won't work either, there's a bug. Then link with -stdlib=libc++.

Environment variables

Once everything was built, add the first four exports of the script in your environment or add them in the .bashenv.

Using the CUDA Backend

1. Use the following selector:

class CUDADeviceSelector : public sycl::device_selector {
public:
    int operator()(const sycl::device &device) const override {
        return device.get_platform().get_backend() == sycl::backend::cuda ? 1 : -1;
    }
};

2. Build with -fsycl -fsycl-targets=nvptx64-nvidia-cuda-sycldevice
3. For oneMKL, link with: -lonemkl_blas_cublas -lonemkl

Intel (R) OpenCL Runtime

For AOT build with -fsycl-targets=spir64_x86_64. .

ESIMD

• libva-devel

Dependencies

• Even though the latest version of CUDA seems to work with DPC++, we need to use CUDA 10.2 as oneMKL uses cuBLAS 10.2. On RHEL: cuda-10-2 libcublas-devel-10-2. For the setup see: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#pre-installation-actions.
• libatomic
• ninja-build cmake libxml2-devel ... and other "build essentials" required for building a compiler.
• libcudnn8 libcudnn8-devel for oneDNN.
• gcc-gfortran or another Fortran compiler for NETLIB Lapack.
• hwloc-devel for oneTBB
• mpfr-devel when building libcxx
• lua-devel python3.-devel libedit-devel

Caveats

• CLion integration: you need to add cidr.compiler.clang.fsycl=true in the registry of CLion see
• oneMKL does not support testing when building for both cuBLAS and cuRAND
• for oneMKL with cuda you should use the namespace oneapi::mkl::blas::column_major:: as cuBLAS is column_major.

Repositories used

• OpenCL Headers+Loaders: https://github.com/KhronosGroup/OpenCL-Headers.git , https://github.com/KhronosGroup/OpenCL-ICD-Loader.git
• DPC++ Compiler: https://github.com/intel/llvm.git
• NETLIB Lapack: https://github.com/Reference-LAPACK/lapack.git
• oneTBB: https://github.com/oneapi-src/oneTBB.git
• oneMKL: https://github.com/oneapi-src/oneMKL.git
• oneDNN: https://github.com/oneapi-src/oneDNN.git