How to get GPU acceleration working

[DSSolo]

Hi I have downloaded the latest release and when I try to run it i get the next message

Running global positioning ...

I20251119 09:52:30.736499 6744 image_undistorter.cc:20] Undistorting images..
I20251119 09:52:30.736564 6744 image_undistorter.cc:44] Image undistortion done
I20251119 09:52:30.737263 6744 global_positioning.cc:52] Setting up the global positioner problem
W20251119 09:52:32.141412 6744 global_positioning.cc:512] Requested to use GPU for bundle adjustment, but Ceres was compiled without CUDA support. Falling back to CPU-based dense solvers.
W20251119 09:52:32.141913 6744 global_positioning.cc:529] Requested to use GPU for bundle adjustment, but Ceres was compiled without cuDSS support. Falling back to CPU-based sparse solvers.
I20251119 09:52:32.144541 6744 global_positioning.cc:79] Solving the global positioner problem

I have downloaded the cuda glomap variant .

[Kation]

Same problem

[nocnestudio]

I managed to successfully build COLMAP 3.13 from the main Git repository on Ubuntu 24.04, using Ceres Solver 2.3, GloMAP 1.2, cuDSS 0.5, and CUDA 12.6.3. I can help and show steps... Interested ? Running bundle adjustment ...

I20260202 14:05:48.157963 10109 global_mapper.cc:196] Bundle adjustment start
I20260202 14:06:50.844521 10109 bundle_adjustment.cc:103] Ceres Solver Report: Iterations: 30, Initial cost: 6.935733e+07, Final cost: 4.987982e+07, Termination: CONVERGENCE
I20260202 14:06:50.844617 10109 global_mapper.cc:212] Global bundle adjustment iteration 1 / 3, stage 1 finished (position only)

[nocnestudio]

SOLVED !!!
I managed to successfully build COLMAP 3.13 from the main Git repository on Ubuntu 24.04, using Ceres Solver 2.3, GloMAP 1.2, cuDSS 0.5, and CUDA 12.6.3. I can help and show steps... Interested ? Running bundle adjustment ...

I20260202 14:05:48.157963 10109 global_mapper.cc:196] Bundle adjustment start
I20260202 14:06:50.844521 10109 bundle_adjustment.cc:103] Ceres Solver Report: Iterations: 30, Initial cost: 6.935733e+07, Final cost: 4.987982e+07, Termination: CONVERGENCE
I20260202 14:06:50.844617 10109 global_mapper.cc:212] Global bundle adjustment iteration 1 / 3, stage 1 finished (position only)

[Kation]

Can you build a win version?

[nocnestudio]

I think this is the same way.
0. Clean or Uninstall drivers.
2. Install Nvidia Cuda 12.6.3
3. Install Nvidia Cudss 0.5
4. Install (Build) Ceres Solver 2.3 - http://ceres-solver.org/installation.html
git clone https://ceres-solver.googlesource.com/ceres-solver
cd ceres-solver
mkdir build && cd build
cmake ..
-DUSE_CUDA=ON
-Dcudss_DIR=/lib/x86_64-linux-gnu/cmake/cudss \ # Cmake/Cudss Location ( files inside like: cudss-config.cmake )
-DBUILD_TESTING=OFF
-DCMAKE_BUILD_TYPE=Release
make -j8
make install

5. Install (Build) Colmap 3.13.0
   Download Colmap 3.13.0 - https://github.com/colmap/colmap/archive/refs/tags/3.13.0.zip
   cd colmap
   mkdir build && cd build
   cmake ..
   -DCMAKE_BUILD_TYPE=Release
   -DCeres_DIR=/usr/local/lib/cmake/Ceres \ # Cmake/Ceres Location ( files inside like: CeresConfig.cmake)
   -DGUI_ENABLED=ON
   -GNinja
   -DBUILD_TESTING=OFF
   -DCGAL_ENABLED=OFF
   ninja
   sudo ninja install

6. Install (Build) Glomap 1.2
   git clone https://github.com/colmap/glomap.git
   cd glomap
   mkdir build && cd build
   cmake ..
   -DCMAKE_BUILD_TYPE=Release
   -DCeres_DIR=/usr/local/lib/cmake/Ceres \ # Cmake/Ceres Location ( files inside like: CeresConfig.cmake)
   -DGUI_ENABLED=ON
   -GNinja
   -DBUILD_TESTING=OFF
   -DCGAL_ENABLED=OFF
   ninja
   sudo ninja install

[nocnestudio]

Can you build a win version?

The easiest way to install CUDA

The simplest way to install CUDA is to use WSL (Windows Subsystem for Linux). I highly recommend choosing Ubuntu 24.04.

It works great because:

Native experience: The installation process is exactly the same as on a real Linux system.

Full GPU support: WSL fully supports your graphics card, so you can run Linux apps on Windows without any issues.

Performance: It is fast, stable, and very easy to set up.

Why I recommend Linux over Windows

I also tried installing everything directly on Windows, but it was just too much of a headache. It is much easier on Linux.

Installing all the necessary libraries for PyTorch, Python, and various coding environments is significantly smoother. You save yourself a lot of stress because things just work the way they are supposed to.

[nocnestudio]

It's working with Cuda 13.1 and Cudss 0.7 Ubuntu 24.04 Colmap 3.13.0 !! ;-)