Implementation of a Heterogeneous System for Image Processing on an FPGA
Heterogeneous systems combine a general-purpose host processor with domain-specific Programmable Many-Core Accelerators (PMCAs). Such systems are highly versatile, due to their host processor capabilities, while having high performance and energy efficiency through their PMCAs. HERO is a FPGA-based research platform developed at IIS that combines a PMCA composed by RISC-V cores, implemented as soft cores on an FPGA fabric, with a hard ARM Cortex-A multicore host processor.
Heterogeneous systems have a complex programming model, which lead to significant effort to develop tools to retain a high programmer productivity. Halide is domain specific programming language designed to write fast image processing algorithms. More specifically, it is a C++ dialect with a functional programming paradigm. It's aim is to separate the function applied to the image (pipeline), and the sequence in which the algorithm is executed (schedule). For example, the schedule encompasses how the algorithm is parallelized, if the image is tiled, processed in column or row major order, if solutions required by multiple threads are shared or recomputed, if parts of the computation is offloaded to an accelerator, and so on. This allows a programmer to write a functional description of the image processing algorithm and then explore ways of scheduling the execution with only a couple of lines of code, and without modifying the algorithm. Furthermore, the same algorithm can be run efficiently on multiple different architectures by only changing the schedule. To have Halide generate efficient code, the specific architecture requires to have an efficient Halide runtime implementation, and good compiler support, as Halide is tightly coupled with the compiler.
The goal of this project is to bring up Halide on HERO, using Ariane, a 64-bit RV64GC core, as a host processor. Ariane would manage Halide's frontend, while the image processing tasks would execute on 32-bit cores in the cluster. The final goal of this thesis is to have Halide programmed image processing kernels running on an HERO system implemented on an FPGA.
The project can be done by as one or two semester thesis. The project consists of three parts:
1. Familiarizing with the Halide language and the architecture of HERO (~2 person weeks).
2. Add a RISC-V target to Halide's frontend (~3 person weeks).
3. Test up the Halide environment on an FPGA with a set of custom image processing kernels (~1 person week)
4. Documentation and report writing (~1 person week)
To work on this project, you will need:
- to have worked in the past with at least one RTL language (SystemVerilog or Verilog or VHDL). Having followed the VLSI 1 course is recommended.
- to have prior knowlegde of the C++ programming language
- to have prior knowledge of hardware design and computer architecture
- to be motivated to work hard on a super cool open-source project
Status: In progress
Meetings & Presentations
The students and advisor(s) agree on weekly meetings to discuss all relevant decisions and decide on how to proceed. Of course, additional meetings can be organized to address urgent issues.
Around the middle of the project there is a design review, where senior members of the lab review your work (bring all the relevant information, such as prelim. specifications, block diagrams, synthesis reports, testing strategy, ...) to make sure everything is on track and decide whether further support is necessary. They also make the definite decision on whether the chip is actually manufactured (no reason to worry, if the project is on track) and whether more chip area, a different package, ... is provided. For more details refer to (1).
At the end of the project, you have to present/defend your work during a 15 min. presentation and 5 min. of discussion as part of the IIS Colloquium.
- Andreas Kurth, Pirmin Vogel, Alessandro Capotondi, Andrea Marongiu, Luca Benini. HERO: Heterogeneous Embedded Research Platform for Exploring RISC-V Manycore Accelerators on FPGA. CARRV' 2017. link
- Jonathan Ragan-Kelley, Andrew Adams, Sylvain Paris, Marc Levoy, Saman Amarasinghe, Frédo Durand. Decoupling Algorithms from Schedules for Easy Optimization of Image Processing Pipelines. SIGGRAPH 2012. link