Vikram S. Adve

Digital Agriculture: 21st century agriculture presents a number of major challenges and opportunities for researchers in computer science and engineering: heterogeneous data analysis, IOT and sensor networks, computer vision, robotics, edge computing, machine learning with limited and sparse data, and others.  I co-founded and co-lead the Center for Digital Agriculture at UIUC.  CDA is a campus-level organization that brings together multidisciplinary teams of researchers and educators spanning Computer Science, Engineering, Agriculture, Food Sciences, Biology, Sustainability, Ecology and related areas for joint research, education, and outreach. Initial funding for CDA has been provided by the office of the Provost and by the Colleges of Engineering and Agriculture.  I also lead the AIFARMS Institute , one of the national artificial intelligence (AI) research institutes launched under the aegis of the US National AI R&D Strategic Plan: 2019 Update.  AIFARMS, funded primarily by USDA NIFA, is exploring foundational AI challenges and how they can play a role in tackling major agriculture challenges, most importantly, how to sustainably feed a global population projected to grow by 70% by 2050.

CropWizard Project on Generative AI for Agriculture: This project I initiated as part of AIFARMS is exploring ways in which generative AI advances can be useful in modern agriculture, from agronomic advice to on-farm decision making to interactive data analytics.  The CropWizard system is an interactive question-answering service for agricultural professionals based on generative AI and knowledge retrieval, using a large database of over 450K documents from land-grant universities and open-access research. CropWizard supports multimodal inputs, including text and images, and automatically invokes (predefined) computational tools for data-driven decision making. One research direction is developing multiple different QA data sets for training and benchmarking AI-driven automated QA services for agricultural topics. These data sets will form a basis for a AI AgriBench, a public benchmarking consortium recently announced by our group along with external partners. Another direction is on improving the accuracy of image analysis through an enhanced retrieval-based (RAG) pipeline for multimodal data. A third is on a multi-agent framework combining reasoning, knowledge extraction, and computational tools, guided by user interactions to achieve more sophisticated data-driven decision making in precision agriculture for open-ended questions. The CropWizard system has been licensed for commercial uses by multiple companies. The CropWizard project is sponsored by USDA NIFA through the AIFARMS institute, and additional funding has been provided by UIUC through the Center for Digital Agriculture, Intel Corporation, Amazon-Illinois AICE Center, and the state of Illinois through the Discovery Partners Institute (DPI).

Edge Computing with HPVM, ApproxHPVM, and ApproxTuner:  Many important and emerging application domains require increasingly powerful computing capabilities at the “edge of the network,” near the sensors, displays and actuators that interact with the physical world. Heterogenous computing with specialized accelerators and often system-on-chip designs are critical to deliver these computing capabilities under tight constraints on memory capacity, compute capacity, energy, power, weight, and heat dissipation. Building on  the HPVM project, we are exploring how compilers, autotuners, and approximate computing techniques can be used to achieve these computing capabilities under such tight resource constraints. We are exploring the use of automatically tuned approximation algorithms to trade off small amounts of accuracy (or result “quality”) to achieve reduced energy or higher performance or both. We are focusing on three broad edge computing application domains: mobile robots in agriculture, autonomous vehicles, and distributed AR/VR systems.

Hydride and MISAAL: Building Compilers Automatically Using Program Synthesis: Hydride is a novel approach to compiling for complex, emerging hardware architectures. Hydride uses vendor-defined pseudocode specifications of multiple hardware ISAs to automatically design retargetable instructions for AutoLLVM IR, an extensible compiler IR which consists of (formally defined) language-independent and target-independent LLVM IR instructions to compile to those ISAs, and automatically generated instruction selection passes to lower AutoLLVM IR to each of the specified hardware ISAs. Hydride also includes a code synthesizer that automatically generates code generation support for schedule-based languages, such as Halide, to optimally generate AutoLLVM IR.

MISAAL dramatically speeds up Hydride (and improves performance) by moving program synthesis entire offline. MISAAL employs a novel strategy to use formal semantics of hardware instructions to automatically prune a large search space of rewrite rules for modern complex instructions in an offline stage. It then uses term-rewriting process in the online stage (at compile-time) instead of pattern-matching to perform optimized code generation that achieves closer-to-optimal code. MISAAL uses a novel methodology to make (online) term rewriting extremely lightweight so as to enable programs to compile in seconds.

Heterogeneous Parallel Virtual Machine: The slowdown of Moore’s Law and the end of Dennard scaling are putting an end to computing speed increases from general-purpose architectures and processor design techniques. Modern computing systems, ranging from smart speakers to mobile phones to laptops to the cloud, now rely on increasing numbers of specialized computing elements (or “accelerators”), such as GPUs, DSPs, FPGAs, image processors, cryptographic hardware, and — increasingly — tensor accelerators for machine learning. The major drawback of these heterogeneous systems is that they are highly challenging to program. The underlying hardware usually has diverse programming interfaces (both languages and instruction sets), differing forms of parallelism, different memory architectures, and even diverse, mutually incompatible programming tools that prevent developing unified applications that can use all components flexibly. In the HPVM project, we are developing a retargetable compiler infrastructure and parallel program representation for heterogeneous parallel systems, with the goal of making it much easier to write performance-portable parallel programs.  A single program in the HPVM representation can be compiled to diverse hardware compute units in a heterogeneous system, using the HPVM program representation to target parallelism and data movement.  Current targets include GPUs, CPUs with vector extensions, FPGAs, and custom accelerators for convolutional neural networks, FFT and Viterbi.