.. meta:: :description: installing GSplat for ROCm :keywords: installation instructions, GSplat, AMD, ROCm, Gaussian splatting .. _gsplat-on-rocm-installation: ******************************************************************** GSplat on ROCm installation ******************************************************************** System requirements ==================================================================== To use GSplat (Gaussian splatting) `1.5.3b2 `__, you need the following prerequisites: - **ROCm version:** `6.4.3 `__ , `7.0.0 `__ (recommended) - **Operating system:** Ubuntu 22.04, 24.04 - **GPU platforms:** AMD Instinct™ MI325X, MI300X - **PyTorch:** `2.6 `__, `2.8 `__ (ROCm-enabled) - **Python:** `3.10 `__, `3.12 `__ Install GSplat ==================================================================== To install GSplat for ROCm, you have the following options: * :ref:`using-docker-image-with-pytorch` **(recommended)** * :ref:`build-from-source` .. _using-docker-image-with-pytorch: Use a base PyTorch Docker image and install with pip -------------------------------------------------------------------- You need to install PyTorch to begin. The easiest way is to use the official ROCm-enabled PyTorch Docker image. Docker is the recommended method to set up your environment, as it avoids potential installation issues. 1. Pull a ROCm PyTorch Docker image with a supported configuration (see `Docker Hub `__ to browse available images). For example: .. tab-set:: .. tab-item:: ROCm 7.0.0 :sync: rocm7 .. tab-set:: .. tab-item:: Ubuntu 24.04 :sync: ubuntu-24 .. code-block:: shell docker pull rocm/pytorch:rocm7.0_ubuntu24.04_py3.12_pytorch_release_2.8.0 See `rocm/pytorch:rocm7.0_ubuntu24.04_py3.12_pytorch_release_2.8.0 `__ on Docker Hub. .. tab-item:: Ubuntu 22.04 :sync: ubuntu-22 .. code-block:: shell docker pull rocm/pytorch:rocm7.0_ubuntu22.04_py3.10_pytorch_release_2.8.0 See `rocm/pytorch:rocm7.0_ubuntu22.04_py3.10_pytorch_release_2.8.0 `__ on Docker Hub. .. tab-item:: ROCm 6.4.3 :sync: rocm6 .. tab-set:: .. tab-item:: Ubuntu 24.04 :sync: ubuntu-24 .. code-block:: shell docker pull rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0 See `rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0 `__ on Docker Hub. .. tab-item:: Ubuntu 22.04 :sync: ubuntu-22 .. code-block:: shell docker pull rocm/pytorch:rocm6.4.3_ubuntu22.04_py3.10_pytorch_release_2.6.0 See `rocm/pytorch:rocm6.4.3_ubuntu22.04_py3.10_pytorch_release_2.6.0 `__ on Docker Hub. 2. Launch and connect to the Docker container: .. tab-set:: .. tab-item:: ROCm 7.0.0 :sync: rocm7 .. tab-set:: .. tab-item:: Ubuntu 24.04 :sync: ubuntu-24 .. code-block:: shell docker run -it \ --cap-add=SYS_PTRACE \ --ipc=host \ --privileged=true \ --shm-size=128GB \ --network=host \ --device=/dev/kfd \ --device=/dev/dri \ --group-add video \ -v $HOME:$HOME \ --name rocm_pytorch \ rocm/pytorch:rocm7.0_ubuntu24.04_py3.12_pytorch_release_2.8.0 .. tab-item:: Ubuntu 22.04 :sync: ubuntu-22 .. code-block:: shell docker run -it \ --cap-add=SYS_PTRACE \ --ipc=host \ --privileged=true \ --shm-size=128GB \ --network=host \ --device=/dev/kfd \ --device=/dev/dri \ --group-add video \ -v $HOME:$HOME \ --name rocm_pytorch \ rocm/pytorch:rocm7.0_ubuntu22.04_py3.10_pytorch_release_2.8.0 .. tab-item:: ROCm 6.4.3 :sync: rocm6 .. tab-set:: .. tab-item:: Ubuntu 24.04 :sync: ubuntu-24 .. code-block:: shell docker run -it \ --cap-add=SYS_PTRACE \ --ipc=host \ --privileged=true \ --shm-size=128GB \ --network=host \ --device=/dev/kfd \ --device=/dev/dri \ --group-add video \ -v $HOME:$HOME \ --name rocm_pytorch \ rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0 .. tab-item:: Ubuntu 22.04 :sync: ubuntu-22 .. code-block:: shell docker run -it \ --cap-add=SYS_PTRACE \ --ipc=host \ --privileged=true \ --shm-size=128GB \ --network=host \ --device=/dev/kfd \ --device=/dev/dri \ --group-add video \ -v $HOME:$HOME \ --name rocm_pytorch \ rocm/pytorch:rocm6.4.3_ubuntu22.04_py3.10_pytorch_release_2.6.0 3. After setting up the container, you can install GSplat along with all supported dependencies from the AMD-hosted PyPI repository: .. tab-set:: .. tab-item:: ROCm 7.0.0 :sync: rocm7 .. code-block:: bash pip install -r https://github.com/AMD-Ecosystem/gsplat/releases/download/1.5.3b2/requirements-gsplat-rocm-7.0.0.txt --require-hashes --index-url https://pypi.org/simple .. tab-item:: ROCm 6.4.3 :sync: rocm6 .. code-block:: bash pip install -r https://github.com/AMD-Ecosystem/gsplat/releases/download/1.5.3b2/requirements-gsplat-rocm-6.4.3.txt --require-hashes --index-url https://pypi.org/simple 4. Verify the installation: .. code-block:: bash pip show gsplat 5. The output should show as follows: .. code-block:: text Name: gsplat Version: 1.5.3+6f2f06e Summary: Python package for differentiable rasterization of gaussians Home-page: https://github.com/AMD-Ecosystem/gsplat Author: AMD Corporation .. _build-from-source: Build GSplat from source -------------------------------------------------------------------- You can choose to setup the development environment and build the GSplat library manually with the following steps. Prerequisites ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Before building GSplat from source, ensure you have one of the following installed: - `ROCm 7.0.0 `__ and `PyTorch with ROCm 7.0.0 `__ - `ROCm 6.4.3 `__ and `PyTorch with ROCm 6.4.3 `__ 1. To verify your setup, check the ROCm installation and GPU architecture: .. code-block:: bash rocminfo | grep gfx 2. The output should be as follows: .. code-block:: bash Name: gfx942 Name: amdgcn-amd-amdhsa--gfx942:sramecc+:xnack- Name: amdgcn-amd-amdhsa--gfx9-4-generic:sramecc+:xnack- 3. Confirm that PyTorch is installed and accessible: .. code-block:: bash python3 -c 'import torch' 2> /dev/null && echo 'Success' || echo 'Failure' 4. Confirm that the GPU is accessible from PyTorch: .. code-block:: bash python3 -c 'import torch; print(torch.cuda.is_available())' Build steps ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1. Clone the repository with submodules: .. code-block:: bash git clone --recurse-submodules https://github.com/AMD-Ecosystem/gsplat.git 2. Install the GLM dependency: .. code-block:: bash cd gsplat/cuda/csrc/third_party/glm cmake \ -DGLM_BUILD_TESTS=OFF \ -DBUILD_SHARED_LIBS=OFF \ -B build . \ -DCMAKE_INSTALL_PREFIX:PATH=~/.local cmake --build build -- all cmake --build build -- install 3. Build the GSplat wheel: .. code-block:: bash cd ../../../.. python setup.py bdist_wheel 4. Install the wheel: .. code-block:: bash pip install dist/gsplat*.whl 5. Verify the installation: .. code-block:: bash pip show gsplat Run unit tests and verify the installation ==================================================================== The ``tests/`` folder contains automated test scripts that verify that the GPU implementations match those of PyTorch. They primarily focus on validating the functionality of the Gaussian Splatting (GS) and 3D rendering components. These tests ensure the correctness, performance, and stability of the core features implemented in the GSplat library. 1. To run the tests, first install the `nerfacc `__ dependency: .. code-block:: bash git clone https://github.com/AMD-Ecosystem/nerfacc.git cd nerfacc python setup.py bdist_wheel pip install dist/nerfacc*.whl 2. Run specific tests from the ``tests`` directory: .. code-block:: bash cd tests pytest -s -v test_2dgs.py::test_fully_fused_projection_packed_2dgs pytest -s -v test_2dgs.py pytest -s -v test_basic.py Run a GSplat example ==================================================================== A set of examples is available to help you get started. See :doc:`Run a GSplat example <../examples/gsplat-examples>` for more details. Benchmarking and evaluation ==================================================================== The `https://github.com/AMD-Ecosystem/gsplat `_ repository includes a standalone script that reproduces the official Gaussian Splatting results with matching performance on ``PSNR``, ``SSIM``, ``LPIPS``, and the converged number of Gaussians. See :doc:`benchmarks <../reference/benchmark-evaluation>` for more details. Limitations ------------------------------------------------------------------- - Compression of ``splat parameters`` (``positions``, ``scales``, ``rotations``, ``colors``, and ``features``) using PNG image encoding is not supported. - Usage of ``Fused Bilagrid`` in the simple trainer is not supported (torch-based implementations are used instead). - Distributed training with the ``packed`` option is not supported.