ContextQMD
Back to Open3d

Open3D SYCL

3rdparty/README_SYCL.md

0.19.04.5 KB
Original Source

Open3D SYCL

Note: this file will be moved to the docs folder once SYCL support is mature.

Concepts

  • OpenCL: Low-level programming interface for heterogeneous platforms.
  • SYCL: High-level programming interface for heterogeneous platforms. SYCL was originally based on OpenCL but now SYCL is an independent standard.
  • DPC++: Intel's implementation of the SYCL standard.
  • oneAPI: A collection of open standards, including DPC++, oneMKL, oneTBB, etc.

Open3D's SYCL support runs on DPC++ and requires oneAPI dependencies. For convenience, in the source code, we use the term "SYCL" although we may be referring to "DPC++".

Setup

  1. Install oneAPI for Linux: install via apt-get
  2. Install Intel GPGPU driver: GPGPU installation guide
  3. Prepare environment
    bash
    # Source environments
source /opt/intel/oneapi/setvars.sh

# We'll be using oneAPI's distribution of conda and Python
# Python 3.6+ will work
conda create -n sycl python=3.10
conda activate sycl
  4. Check your environment
    bash
    which icx        # /opt/intel/oneapi/compiler/<version>/linux/bin/icx
which icpx       # /opt/intel/oneapi/compiler/<version>/linux/bin/icpx
which python     # ${HOME}/.conda/envs/sycl/bin/python
python --version # Python 3.10.0 :: Intel Corporation
  5. Config and build
    bash
    cmake -DBUILD_SYCL_MODULE=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DBUILD_UNIT_TESTS=ON ..
make -j$(nproc)
  6. Run demos We provide several ways to run SYCL demos. This ensures the linking are correct and all run-time dependencies are satisfied.
    bash
    # C++ test
make tests -j$(nproc)
./bin/tests --gtest_filter="*SYCLDemo*"

# C++ example
make SYCLDemo -j$(nproc)
./bin/examples/SYCLDemo

# Python
make install-pip-package -j$(nproc)
pytest ../python/test/core/test_sycl_utils.py -s

Available SYCL devices

After activating the SYCL environment, you may run sycl-ls to list available SYCL devices, for example:

bash
$ sycl-ls
[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM)
[opencl:cpu:1] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K
[opencl:gpu:2] Intel(R) OpenCL HD Graphics, Intel(R) UHD Graphics 770
[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) UHD Graphics 770
[host:host:0] SYCL host platform, SYCL host device

Open3D is designed to make use of the SYCL GPU devices.

  • core::Device("SYCL:0") maps the default SYCL GPU returned by sycl::gpu_selector().
    • For Intel GPUs, the "level-zero GPU" (ext_oneapi_level_zero:gpu:0 in the example above) backend will be used by default instead of the "OpenCL GPU" (opencl:gpu:2 in the example above).
    • Currently we only support one SYCL GPU. Multi-GPU support will be added in the future.
  • If SYCL GPU device is not available (e.g. when you don't have an Intel integrated or discrete GPU, or when the GPGPU driver is not properly installed), SYCL:0 will fall back to the SYCL host device. The SYCL host device is used for debugging only.

Known limitations/requirement

  • Limitations (not implemented yet)
    • Only supports Linux
    • Only supports BUILD_CUDA_MODULE=OFF
    • ML ops not supported (C++17 cannot be enabled for TF)
  • Requirements (required by DPC++)
    • Only supports GLIBCXX_USE_CXX11_ABI=ON
    • Only supports set(CMAKE_CXX_STANDARD 17)

List of oneAPI Python packages

To make pip install open3d works out-of-the box on SYCL-enabled platforms, we can utilize runtime libraries released via PyPI. This feature needs to be implemented.

User:

Libraries: