Requested to use GPU for bundle adjustment, but Ceres was compiled without CUDA support. Falling back to CPU-based dense solvers.

[Deye27]

Subject: Pre-compiled binaries and vcpkg manifest lack CUDA support for Ceres

Description

The pre-compiled Windows binaries for GLOMAP seem to be built against a version of Ceres that does not have CUDA or cuDSS support enabled. When running the mapper, this results in the following warnings, and the process falls back to CPU-based solvers:

W20250822 16:57:58.663830 22028 global_positioning.cc:380] Requested to use GPU for bundle adjustment, but Ceres was compiled without CUDA support. Falling back to CPU-based dense solvers.
W20250822 16:57:58.664532 22028 global_positioning.cc:397] Requested to use GPU for bundle adjustment, but Ceres was compiled without cuDSS support. Falling back to CPU-based sparse solvers.

Root Cause Analysis

I attempted to resolve this by manually compiling ceres.dll with CUDA and cuDSS support and replacing the distributed DLL. However, the warning persisted. This suggests the issue is not with the DLL at runtime, but rather with how the main glomap.exe executable itself was compiled.

Upon inspection of the repository, the root cause appears to be in the vcpkg.json manifest file. The ceres dependency is declared without the cuda feature, meaning vcpkg will build a CPU-only version of Ceres by default.

Current vcpkg.json entry for Ceres:
https://github.com/colmap/glomap/blob/main/vcpkg.json

        {
            "name": "ceres",
            "features": [
                "lapack",
                "schur",
                "suitesparse"
            ]
        },

Proposed Solution & Current Status

To enable GPU acceleration, the cuda feature needs to be added to the Ceres dependency list in vcpkg.json. This ensures that when GLOMAP is built from source using vcpkg, it links against a version of Ceres that is CUDA-aware.

Proposed change in vcpkg.json:

        {
            "name": "ceres",
            "features": [
                "lapack",
                "schur",
                "suitesparse",
                "cuda"
            ]
        },

I am currently recompiling GLOMAP from source with this modification. I expect this to resolve the issue and enable GPU-based bundle adjustment.

[lpanaf]

Hi @Deye27 , thanks for the feedback. Do you have any luck on this already? If so, it would be really nice if you can reflect the changes in a pull request

[Deye27]

Hi @lpanaf, I have an update: I successfully solved the first warning, Requested to use GPU for bundle adjustment, but Ceres was compiled without CUDA support.

The issue was indeed that the pre-compiled binaries did not include CUDA support in the bundled Ceres library. To fix this, I recompiled GLOMAP from source on Windows. Here is a summary of the steps I took, which may be useful for others aiming to build with GPU support:

1. Enable CUDA in Ceres via vcpkg.json

The key step was to modify GLOMAP's root vcpkg.json file. I added "cuda" to the list of features for the ceres dependency:

{
    "name": "ceres",
    "features": [
        "lapack",
        "schur",
        "suitesparse",
        "cuda"
    ]
},

This instructed vcpkg to build Ceres with its CUDA feature enabled.

2. Handle Additional Build Dependencies

During configuration, the COLMAP dependency required two additional libraries that were not in the default list:

• Missing CGAL: Solved by adding "cgal" to the main dependencies list in vcpkg.json.
• Missing Qt5: Solved by disabling the COLMAP GUI build, which isn't required by GLOMAP. This was done by adding the -DGUI_ENABLED=OFF flag to the CMake command.

3. Final Configuration Command

The complete CMake command that led to a successful configuration was:

cmake .. -G "Visual Studio 17 2022" -A x64 -DCMAKE_TOOLCHAIN_FILE="C:/path/to/vcpkg/scripts/buildsystems/vcpkg.cmake" -DGUI_ENABLED=OFF

Current Status & Remaining Issue

After compiling and running the new binaries, the first warning has disappeared.

However, the second warning persists:
W2025... Requested to use GPU for bundle adjustment, but Ceres was compiled without cuDSS support. Falling back to CPU-based sparse solvers.

I am holding off on creating a pull request since the solution is incomplete. I suspect that while the "cuda" feature in vcpkg's Ceres portfile enables CUDA, it might not be configured to also detect and enable cuDSS.

I will continue to investigate, but if you have any insights into how cuDSS could be enabled within this vcpkg-based build process, I would greatly appreciate your input.

[Kation]

Any update?