Troubleshooting and FAQ#

If your problem is not covered here, search the issues and discussions. When you report a bug, include a minimal, complete example that someone else can run to see the error, along with your operating system, GPU, CUDA version and TIGRE version.

Installing#

TIGRE does not compile. Follow the installation guide for your language; it covers the usual causes (CUDA not found, wrong compiler, missing Visual Studio on Windows).

nvcc fatal: Unsupported gpu architecture 'compute_XX'. Your CUDA version cannot build for one of the GPU architectures in the build flags. The Python build asks nvcc which architectures it supports, and the MATLAB build chooses them from your CUDA version, so this should be rare. If you still see it, remove the unsupported -gencode entries from the COMPFLAGS line of your MATLAB mex_CUDA_*.xml file, and please open an issue so the build can be fixed.

There are many warnings during compilation. They are expected and do not affect results or speed.

Can I compile with MinGW on Windows? No. CUDA on Windows requires the Microsoft Visual Studio compiler. The free Community edition is enough.

Running#

Input data should be float32 (Python) or Image should be single type (MATLAB). TIGRE does not convert arrays for you. Use img.astype(np.float32) or single(img).

A shape error from Ax, Atb or an algorithm. The image must have shape geo.nVoxel, and the projections (len(angles), *geo.nDetector) in Python. Remember that Python orders the axes (z, y, x) and (v, u); see Describing your scanner: the geometry.

nVoxel*dVoxel is not equal to sVoxel. When you change one of nVoxel, dVoxel or sVoxel, update the others to match. The same applies to the detector fields.

texture memory fail or invalid device function. TIGRE was not built for your GPU. Rebuild it on the computer you use, with a CUDA version that supports your GPU.

One (or more) of your GPUs is being heavily used by another program. TIGRE needs at least half of each selected GPU’s memory to be free. Close programs using the GPU, or select another GPU; see GPUs and memory.

the launch timed out and was terminated or unspecified launch failure on large images. The operating system stopped a GPU kernel because it ran too long, which happens on GPUs that also drive a display. On Windows, raise the TDR delay or turn TDR off (for example in NVIDIA Nsight Monitor, under WDDM TDR Delay and WDDM TDR Enabled). Data-centre GPUs can be put in TCC mode instead.

After a CUDA error, every call fails. A failed CUDA call can leave the GPU context in a bad state. Restart MATLAB or the Python process.

The reconstruction is blurred, doubled or shaded. This is almost always the geometry, not the algorithm. Work through the checklist at the end of Describing your scanner: the geometry.

A Python algorithm ignores an option. Python prints a warning, rather than raising an error, for keyword arguments it does not recognise. Check the spelling: for example the step size is lmbda, not lambda.

Using TIGRE#

Can I use TIGRE for my project, including commercial work? Yes. TIGRE is released under the permissive BSD 3-clause licence. If you publish work that uses it, please cite TIGRE.

How can I contribute? See the contributing guide. Bug fixes, new algorithms, scanner loaders and documentation are all welcome.