Sections

Sections is the fundamental concept of the HTTomo’s framework which is related to how the I/O operations and processing of data is organised.

Note

The main purpose of a section is to organise the data input/output workflow, as well as, chain together the methods so that the constructed pipeline is computationally efficient.

To better understand the purpose of the section it is also useful to read information about Chunks, Blocks and Memory Estimators.

Bellow we present different situations that can lead to the sections being organised in a specific manner.

Fig. 2 Here is a typical pipeline with a loader (L), 5 methods (M), and 4 data transfer operations (T) between methods.

Sections are created when:

1. Re-slicing is needed, which is related to the change of pattern.

2. The output of the method needs to be saved to the disk.

3. The Side outputs is required by one of the methods.

Example 1: Sections with re-slice

Fig. 3 Let us say that the pattern in methods M_1-3 is projection and methods in M_4-5 belong to sinogram pattern. This will result in two sections created and also Re-slicing operation in the data transfer T₃ layer.

Example 2 : Sections with re-slice and data saving

Fig. 4 In addition Example 1 situation, let us assume that we want to save the result of M₂ method to the disk. This means that even though M_1-3 methods can be performed on the GPU, the data will be transferred to CPU. The pipeline will be further fragmented to introduce another section, so that the data transfer T₂ layer also saves the data on the disk, as well as, taking care to return the data back on the GPU for the method M₃.

Example 3 : Sections with side outputs

Fig. 5 Consider, again, Example 1. Here, however, M₅ requests the Side outputs of the method M₄. This, for example, can be because the reconstruction method M₅ requires the Centre of Rotation (CoR) value of M₄, where this value is calculated. This divides M₄ and M₅ into separate sections. Also notice that M₁ needs the data to be saved on disk, so in total, it is a pipeline with 4 sections in it.

Note

It can be seen that creating more sections in pipelines is to be avoided when building an efficient pipeline. Creating a section usually leads to synchronisation of all processes on the CPU and potentially, if not enough memory, through-disk operations.