The performance of the code generated by a compiler depends on the order in which the optimization passes are applied. In high-level synthesis, the quality of the generated circuit relates directly to the code generated by the front-end compiler. Choosing a good order--often referred to as the phase-ordering problem--is an NP-hard problem. In this paper, we evaluate a new technique to address the phase-ordering problem: deep reinforcement learning. We implement a framework in the context of the LLVM compiler to optimize the ordering for HLS programs and compare the performance of deep reinforcement learning to state-of-the-art algorithms that address the phase-ordering problem. Overall, our framework runs one to two orders of magnitude faster than these algorithms, and achieves a 16% improvement in circuit performance over the -O3 compiler flag.
Using FPGAs to accelerate ConvNets has attracted significant attention in recent years. However, FPGA accelerator design has not leveraged the latest progress of ConvNets. As a result, the key application characteristics such as frames-per-second (FPS) are ignored in favor of simply counting GOPs, and results on accuracy, which is critical to application success, are often not even reported. In this work, we adopt an algorithm-hardware co-design approach to develop a ConvNet accelerator called Synetgy and a novel ConvNet model called DiracDeltaNet$^{\dagger}$. Both the accelerator and ConvNet are tailored to FPGA requirements. DiracDeltaNet, as the name suggests, is a ConvNet with only $1\times 1$ convolutions while spatial convolutions are replaced by more efficient shift operations. DiracDeltaNet achieves competitive accuracy on ImageNet (88.7\% top-5), but with 42$\times$ fewer parameters and 48$\times$ fewer OPs than VGG16. We further quantize DiracDeltaNet's weights to 4-bit and activations to 4-bits, with less than 1\% accuracy loss. These quantizations exploit well the nature of FPGA hardware. In short, DiracDeltaNet's small model size, low computational OP count, low precision and simplified operators allow us to co-design a highly customized computing unit for an FPGA. We implement the computing units for DiracDeltaNet on an Ultra96 SoC system through high-level synthesis. Our accelerator's final top-5 accuracy of 88.1\% on ImageNet, is higher than all the previously reported embedded FPGA accelerators. In addition, the accelerator reaches an inference speed of 66.3 FPS on the ImageNet classification task, surpassing prior works with similar accuracy by at least 11.6$\times$.
A deep-learning inference accelerator is synthesized from a C-language software program parallelized with Pthreads. The software implementation uses the well-known producer/consumer model with parallel threads interconnected by FIFO queues. The LegUp high-level synthesis (HLS) tool synthesizes threads into parallel FPGA hardware, translating software parallelism into spatial parallelism. A complete system is generated where convolution, pooling and padding are realized in the synthesized accelerator, with remaining tasks executing on an embedded ARM processor. The accelerator incorporates reduced precision, and a novel approach for zero-weight-skipping in convolution. On a mid-sized Intel Arria 10 SoC FPGA, peak performance on VGG-16 is 138 effective GOPS.
In order to improve usability and safety, modern unmanned aerial vehicles (UAVs) are equipped with sensors to monitor the environment, such as laser-scanners and cameras. One important aspect in this monitoring process is to detect obstacles in the flight path in order to avoid collisions. Since a large number of consumer UAVs suffer from tight weight and power constraints, our work focuses on obstacle avoidance based on a lightweight stereo camera setup. We use disparity maps, which are computed from the camera images, to locate obstacles and to automatically steer the UAV around them. For disparity map computation we optimize the well-known semi-global matching (SGM) approach for the deployment on an embedded FPGA. The disparity maps are then converted into simpler representations, the so called U-/V-Maps, which are used for obstacle detection. Obstacle avoidance is based on a reactive approach which finds the shortest path around the obstacles as soon as they have a critical distance to the UAV. One of the fundamental goals of our work was the reduction of development costs by closing the gap between application development and hardware optimization. Hence, we aimed at using high-level synthesis (HLS) for porting our algorithms, which are written in C/C++, to the embedded FPGA. We evaluated our implementation of the disparity estimation on the KITTI Stereo 2015 benchmark. The integrity of the overall realtime reactive obstacle avoidance algorithm has been evaluated by using Hardware-in-the-Loop testing in conjunction with two flight simulators.
Recent work has shown that Field-Programmable Gate Arrays (FPGAs) play an important role in the acceleration of Machine Learning applications. Initial specification of machine learning applications are often done using a high-level Python-oriented framework such as Tensorflow, followed by a manual translation to either C or RTL for synthesis using vendor tools. This manual translation step is time-consuming and requires expertise that limit the applicability of FPGAs in this important domain. In this paper, we present an open-source tool-flow that maps numerical computation models written in Tensorflow to synthesizable hardware. Unlike other tools, which are often constrained by a small number of inflexible templates, our flow uses Google's XLA compiler which emits LLVM code directly from a Tensorflow specification. This LLVM code can then be used with a high-level synthesis tool to automatically generate hardware. We show that our flow allows users to generate Deep Neural Networks with very few lines of Python code.
Multiresolution Analysis (MRA) is a mathematical method that is based on working on a problem at different scales. One of its applications is medical imaging where processing at multiple scales, based on the concept of Gaussian and Laplacian image pyramids, is a well-known technique. It is often applied to reduce noise while preserving image detail on different levels of granularity without modifying the filter kernel. In scientific computing, multigrid methods are a popular choice, as they are asymptotically optimal solvers for elliptic Partial Differential Equations (PDEs). As such algorithms have a very high computational complexity that would overwhelm CPUs in the presence of real-time constraints, application-specific processors come into consideration for implementation. Despite of huge advancements in leveraging productivity in the respective fields, designers are still required to have detailed knowledge about coding techniques and the targeted architecture to achieve efficient solutions. Recently, the HIPAcc framework was proposed as a means for automatic code generation of image processing algorithms, based on a Domain-Specific Language (DSL). From the same code base, it is possible to generate code for efficient implementations on several accelerator technologies including different types of Graphics Processing Units (GPUs) as well as reconfigurable logic (FPGAs). In this work, we demonstrate the ability of HIPAcc to generate code for the implementation of multiresolution applications on FPGAs and embedded GPUs.
In the domain of image processing, often real-time constraints are required. In particular, in safety-critical applications, such as X-ray computed tomography in medical imaging or advanced driver assistance systems in the automotive domain, timing is of utmost importance. A common approach to maintain real-time capabilities of compute-intensive applications is to offload those computations to dedicated accelerator hardware, such as Field Programmable Gate Arrays (FPGAs). Programming such architectures is a challenging task, with respect to the typical FPGA-specific design criteria: Achievable overall algorithm latency and resource usage of FPGA primitives (BRAM, FF, LUT, and DSP). High-Level Synthesis (HLS) dramatically simplifies this task by enabling the description of algorithms in well-known higher languages (C/C++) and its automatic synthesis that can be accomplished by HLS tools. However, algorithm developers still need expert knowledge about the target architecture, in order to achieve satisfying results. Therefore, in previous work, we have shown that elevating the description of image algorithms to an even higher abstraction level, by using a Domain-Specific Language (DSL), can significantly cut down the complexity for designing such algorithms for FPGAs. To give the developer even more control over the common trade-off, latency vs. resource usage, we will present an automatic optimization process where these criteria are analyzed and fed back to the DSL compiler, in order to generate code that is closer to the desired design specifications. Finally, we generate code for stereo block matching algorithms and compare it with handwritten implementations to quantify the quality of our results.
Image aesthetics assessment has been challenging due to its subjective nature. Inspired by the scientific advances in the human visual perception and neuroaesthetics, we design Brain-Inspired Deep Networks (BDN) for this task. BDN first learns attributes through the parallel supervised pathways, on a variety of selected feature dimensions. A high-level synthesis network is trained to associate and transform those attributes into the overall aesthetics rating. We then extend BDN to predicting the distribution of human ratings, since aesthetics ratings are often subjective. Another highlight is our first-of-its-kind study of label-preserving transformations in the context of aesthetics assessment, which leads to an effective data augmentation approach. Experimental results on the AVA dataset show that our biological inspired and task-specific BDN model gains significantly performance improvement, compared to other state-of-the-art models with the same or higher parameter capacity.
Detection of moving objects in videos is a crucial step towards successful surveillance and monitoring applications. A key component for such tasks is called background subtraction and tries to extract regions of interest from the image background for further processing or action. For this reason, its accuracy and real-time performance is of great significance. Although, effective background subtraction methods have been proposed, only a few of them take into consideration the special characteristics of thermal imagery. In this work, we propose a background subtraction scheme, which models the thermal responses of each pixel as a mixture of Gaussians with unknown number of components. Following a Bayesian approach, our method automatically estimates the mixture structure, while simultaneously it avoids over/under fitting. The pixel density estimate is followed by an efficient and highly accurate updating mechanism, which permits our system to be automatically adapted to dynamically changing operation conditions. We propose a reference implementation of our method in reconfigurable hardware achieving both adequate performance and low power consumption. Adopting a High Level Synthesis design, demanding floating point arithmetic operations are mapped in reconfigurable hardware; demonstrating fast-prototyping and on-field customization at the same time.
Deep Convolutional Neural Networks have become a Swiss knife in solving critical artificial intelligence tasks. However, deploying deep CNN models for latency-critical tasks remains to be challenging because of the complex nature of CNNs. Recently, FPGA has become a favorable device to accelerate deep CNNs thanks to its high parallel processing capability and energy efficiency. In this work, we explore different fast convolution algorithms including Winograd and Fast Fourier Transform (FFT), and find an optimal strategy to apply them together on different types of convolutions. We also propose an optimization scheme to exploit parallelism on novel CNN architectures such as Inception modules in GoogLeNet. We implement a configurable IP-based face recognition acceleration system based on FaceNet using High-Level Synthesis. Our implementation on a Xilinx Ultrascale device achieves 3.75x latency speedup compared to a high-end NVIDIA GPU and surpasses previous FPGA results significantly.
Recent results at the Large Hadron Collider (LHC) have pointed to enhanced physics capabilities through the improvement of the real-time event processing techniques. Machine learning methods are ubiquitous and have proven to be very powerful in LHC physics, and particle physics as a whole. However, exploration of the use of such techniques in low-latency, low-power FPGA hardware has only just begun. FPGA-based trigger and data acquisition (DAQ) systems have extremely low, sub-microsecond latency requirements that are unique to particle physics. We present a case study for neural network inference in FPGAs focusing on a classifier for jet substructure which would enable, among many other physics scenarios, searches for new dark sector particles and novel measurements of the Higgs boson. While we focus on a specific example, the lessons are far-reaching. We develop a package based on High-Level Synthesis (HLS) called hls4ml to build machine learning models in FPGAs. The use of HLS increases accessibility across a broad user community and allows for a drastic decrease in firmware development time. We map out FPGA resource usage and latency versus neural network hyperparameters to identify the problems in particle physics that would benefit from performing neural network inference with FPGAs. For our example jet substructure model, we fit well within the available resources of modern FPGAs with a latency on the scale of 100 ns.
Deep convolutional neural networks have dominated the pattern recognition scene by providing much more accurate solutions in computer vision problems such as object recognition and object detection. Most of these solutions come at a huge computational cost, requiring billions of multiply-accumulate operations and, thus, making their use quite challenging in real-time applications that run on embedded mobile (resource-power constrained) hardware. This work presents the architecture, the high-level synthesis design, and the implementation of SqueezeJet, an FPGA accelerator for the inference phase of the SqueezeNet DCNN architecture, which is designed specifically for use in embedded systems. Results show that SqueezeJet can achieve 15.16 times speed-up compared to the software implementation of SqueezeNet running on an embedded mobile processor with less than 1% drop in top-5 accuracy.
Software-compilable C programs routinely fail to complete the four-stage pipeline of a high-level synthesis (HLS) toolchain -- compilation, C simulation (CSim), synthesis, and C/RTL co-simulation (CoSim) -- because HLS accepts only a synthesizable subset of C (HLS-C). Yet most existing large language model (LLM) systems built for HLS code repair only cover the early pipeline stages and feed raw tool logs directly to the model, yielding brittle and hard-to-reproduce fixes. We formulate C-to-HLS-C conversion as a closed-loop generation-verification-diagnosis-repair problem on an HLS tool (Xilinx Vitis), contributing three components: an end-to-end workflow of cooperating agents closed by the four-stage verifier under strict evidence isolation; a Progressive Mismatch Localization Chain (PMLC) that localizes CSim/CoSim mismatches through log normalization, AST backward slicing, and dual-trace instrumentation; and a typed-query, two-stage evidence RAG backed by a self-evolving, family-routed repair-card pool. Experimental results show that the proposed workflow substantially outperforms all comparable state-of-the-art models.