Section: New Results

Program Analysis and Communication Optimizations for HLS

Participants : Christophe Alias, Alain Darte, Alexandru Plesco.

High-level synthesis (HLS) tools are now getting more mature for generating hardware accelerators with an optimized internal structure, thanks to efficient scheduling techniques, resource sharing, and finite-state machines generation. However, interfacing them with the outside world, i.e., integrating the automatically-generated hardware accelerators within the complete design, with optimized communications, so that they achieve the best throughput, remains a very hard task, reserved to expert designers. The goal of our research on HLS is to study and to develop source-to-source strategies to improve the design of these interfaces, trying to consider the HLS tool as a back-end for more advanced front-end transformations.

Using the C2H HLS tool from Altera, which can synthesize hardware accelerators communicating to an external DDR-SDRAM memory, we showed that it is possible to automatically restructure the application code, to generate adequate communication processes in C, and to compile them all with C2H, so that the resulting application is highly-optimized, with full usage of the memory bandwidth.

These transformations and optimizations, which combine techniques such as double buffering, array contraction, loop tiling, software pipelining, among others, were incorporated in an automatic source-to-source transformation tool, called Chuba (see Section  5.7 ), based on the polyhedral model representation. Our study shows that HLS tools can indeed be used as back-end optimizers for front-end optimizations, as it is the case for standard compilation with high-level transformations developed on top of assembly-code optimizers. We believe this is the way to go for making HLS tools viable. The complete automation of the process will be presented at PPoPP'12 [3] and Impact'12 [15] .

We also showed how to extend this method to programs with irregular control and array accesses. The main difficulty arises when some data may be redefined in the accelerator but this is not sure. We showed that techniques based on parametric polyhedral optimizations can be used to generate the sets of data to be loaded (resp. stored) just before (resp. after) each tile. An interesting feature is that the previous method appears nicely as a particular case when no approximation is needed. This work is fully described in a research report [23] , but is not yet published in a conference or journal.