Load balancing of GPU subsystems to accelerate back projection for synchrotron tomography

Start Date
01-10-2018 14:00
Auditorium, Central Building
Speaker's name
Speaker's institute
Karlsruhe Institute of Technology
Contact name
Sarah Brulé
Host name
Alessandro Mirone
Add event to calendar
iCal | vCal

Back-Projection is the major algorithm in Computed Tomography to reconstruct images from a set of recorded projections. It is used for both fast analytical methods and high-quality iterative techniques. While existing implementations rely heavily on the parallel hardware to speed-up the image reconstruction, specific features of the GPU architecture are rarely considered. On the other hand, there are important differences. To provide high performance, GPU architecture include multiple components that are operating independently. Texture fetches, memory operations, several types of arithmetic instructions are executed by the different blocks of a GPU in parallel. Hence, the execution time is not determined by the sum of all operations, but rather by the slowest execution pipeline. The performance balance between these components varies between GPU architectures significantly. Our strategy to implement an efficient algorithm is to balance operations between the available GPU blocks and to minimize the time required to execute the slowest pipeline. I will discuss GPU architectures and present several optimizations to improve resource utilization and to balance the load between texture engine, computational and special function units, as well as different types of memory. The proposed optimizations maximize the utilization of all GPU subsystems and depending on the architecture result 2-7 times speedup as compared to the state-of-the-art algorithms used in standard reconstructions codes.

Visitors from off-site please contact Sarah Brulé tel +33 (0)4 76 88 20 11 to arrange for a gate pass.
Requests made by e-mail will be confirmed.
If you do not receive a confirmation e-mail, please contact us by phone.