Section: Software and Platforms

Chuba

Participants : Christophe Alias, Alain Darte, Alexandru Plesco [Compsys/Zettice] .

Chuba is a source-level optimizer that improves a C program in the context of the high-level synthesis (HLS) of hardware. Chuba is an implementation of the work described in the PhD thesis of Alexandru Plesco. The optimized program specifies a system of multiple communicating accelerators, which optimize the data transfers with the external DDR memory. The program is divided into blocks of computations obtained thanks to tiling techniques, and, in each block, data are fetched by block to reduce the penalty due to line changes in the DDR accesses. Four accelerators achieve data transfers in a macro-pipeline fashion so that data transfers and computations (performed by a fifth accelerator) are overlapped.

So far, the back-end of Chuba is specific to the HLS tool C2H but the analysis is quite general and adapting Chuba to other HLS tools should be possible. Besides, it is interesting to mention that the program analysis and optimizations implemented in Chuba address a problem that is also very relevant in the context of GPGPUs. The underlying theory and corresponding experiments are described in [4] .

Chuba has been implemented by Christophe Alias, using the compiler infrastructure PoCo (see Section  5.5 ). It represents more than 900 lines of C++. The reduced size of Chuba is mainly due to the high-level abstractions provided by PoCo.