CPU and GPU Implementation

In the case of performing back-projecting calculations on CPU Cobra can automatically utilize up to 16 cores.

GPU (graphics process unit) implementation supports graphics cards nVIDIA Geforce of 8000, 9000, 200, and 400 series.  In most cases running a reconstruction with GPU support takes 2-4 times less than running the same task on CPU. A particular speed-up factor depends on corrections applied.  If there are no iterative corrections involved (like metal artifact reduction) a typical GPU speed-up is factor 3-4. Utilized technologies are nVIDIA CUDA and Microsoft DirectX

Reconstructing jumbo volumes

During back-projecting Cobra is using 4-byte accumulation buffer so there are limits on the buffer size that might be allocated in RAM. These limits correspond approximately 700x700x700 cube for CPU and 600x600x600 (Geforce 400 series) or smaller for GPU. However in many cases a desired volume may be larger and much larger. In Cobra this problem is resolved by arranging (as called) multi-pass process when a jumbo cube are automatically divided into horizontal slabs.

A size of back-projecting buffer is defined automatically or can be set up explicitly by a user.

Draft and fine reconstruction

CPU reconstruction implements “voxel driven” approach and can be performed in two modes: draft and fine.  Technically draft mode implements nearest neighbor mapping technique; fine mode uses a bilinear interpolation. As a consequence a draft mode is faster (approximately by factor 1.5) but images may look a little bit noisier. When a detector resolution is substantially higher than a cube resolution, a difference of draft vs. fine may not be distinguished or minor (or non-relevant).

Arranging saving slices as a background process

If requested Cobra supports saving slices to the disk as a background process. It is relevant for jumbo cubes being reconstructed in multiple passes (slabs) and cluster solutions when transferring slices through network may take long. When background slice saving is active then the reconstruction of the slab N+1 is being performed simultaneously with saving slices of the slab N. Obviously it requires an additional memory allocated and back-projecting buffer set manually.