High-dimensional computing

In case of high-dimensional cases, there are ways to avoid running out of GPU memory. Note that all of these methods require sufficient CPU RAM and only help with limited GPU memory.

Note

High-dimensional support is only for built-in algorithms, not the forward/backward projection operators. For high-dimensional cases, the user needs to manually handle the data if the projector operators are used.

High-dimensional measurement data

This is especially relevant for PET TOF data, but it can be used with any other data too. By default, all measurement data is transferred to the GPU before the reconstruction. However, it is possible to limit only the current subset to the GPU. This is achieved by setting options.loadTOF to false. Although the name suggests TOF-only, it works with any measurement data. However, this only works if subsets are used. This also works with list-mode/custom coordinate data, where the coordinates are also split according to the subset. The measurement data still has to fit into the host RAM. Note that you can further reduce memory usage by using unsigned 16-bit or 8-bit integer input data whenever applicable. There are no restrictions when using this mode, it can only lead to slower reconstructions if there is sufficient memory on the GPU and the data is transferred on-the-fly instead.

High-dimensional CT/image data

This is specifically designed for CT data and especially for µCT. This is similar to the above, but the image (estimate) is also reconstructed in subsets, i.e. the full image/estimate is not transferred to the GPU. Like above, this depends on the number of subsets as well. Unlike above, which supports any reconstruction algorithm, this method only supports FDK, PDHG, and its variants, and PKMA. Note that FDK still computes only a single backprojection despite using subsets. The rest divide the reconstruction into the subsets as normal. This feature is enabled by setting options.largeDim to true. The setting of loadTOF is irrelevant if largeDim is true. For regularization, only non-local methods, RDP, GGMRF, gradient-based TV, and hyperbolic prior are supported. Diagonal and EM image-based preconditioners are supported as well as filtering-based measurement-based preconditioner. With a sufficient number of subsets, any type of data can be reconstructed with any type of GPU. However, the input measurement data and output reconstruction have to fit to the computer RAM. For PDHG, the intermediate variables, such as the dual estimate, need to fit, too.

Note

High-dimensional support is only for projectors 1, 4 and 14, though you should never use 1 for CT data. Power method also works slightly differently with largeDim when using extended FOV. The forward projections of only the extended regions are summed together, while normally the main volume is included as well. This should still work, but the primal values might differ from the non-large dimensional case in the extended region (main volume is unaffected).