23
23
In \sectionname~\ref{section:newtex} we proposed an advanced thread mapping scheme for the texture-based kernel. The goal is to keep the pixel to block assignment minimal to preserve performance for the small images. The ALU kernel, however, is aimed on larger image sizes and works with 32-by-32 pixel area at minimum. Therefore, an alternative simpler mapping is utilized for the texture-based kernel if it is executed as part of hybrid approach. The block to pixel assignments are kept in sync with the ALU-based kernel. At each iteration a standard region of 16-by-16 pixels is processed. The thread to pixel assignments follow the mapping described in \sectionname~\ref{section:remap}. Each thread is responsible for 4 to 16 pixels and processes them in the loop. The same texture is used to perform linear interpolation in blocks running texture-based algorithm and to cache data in the shared memory in the blocks executing ALU-based reconstruction. The performance and utilization of GPU subsystems using the different reconstruction modes is reviewed in \tablename~\ref{tbl:utilization}.
27
27
\caption{\label{tbl:utilization} Utilization of functional units in hybrid reconstruction mode}