Convolutional neural networks(CNNs) achieve high accuracy on large datasets but requires significant computation and storage requirement for training/testing. While many applications demand low latency and energy-efficient processing of the images, deploying these complex algorithms on the hardware is a challenging task. This dissertation first presents a compiler-based CNN training accelerator using DDR3 and HBM2 memory. An optimized RTL library is implemented to perform training-specific tasks and an RTL compiler is developed to generate FPGA-synthesizable RTL based on user-defined constraints. High Bandwidth Memory(HBM) provides efficient off-chip communication and improves the training performance. The impact of HBM2 on CNN training workloads is analyzed and compressively compared with DDR3. For training ResNet-20/VGG-like CNNs for the CIFAR-10 dataset, the proposed CNN training accelerator on Stratix-10 GX FPGA(DDR3) demonstrates 479 GOPS performance, and on Stratix-10 MX FPGA(HBM) shows 4.5/9.7 X energy-efficiency improvement compared to Tesla V100 GPU. Next, the FPGA online learning accelerator is presented. Adopting model segmentation techniques from Progressive Segmented Training(PST), the online learning accelerator achieved a 4.2X reduction in training latency. Furthermore, this dissertation presents an 8-bit floating-point (FP8) training processor which implements (1) Highly parallel tensor cores that maintain high PE utilization, (2) Hardware-efficient channel gating for dynamic output activation sparsity (3) Dynamic weight sparsity based on group Lasso (4) Gradient skipping based on FP prediction error. The 28nm prototype chip demonstrates significant improvements in FLOPs reduction (7.3×), energy efficiency (16.4 TFLOPS/W), and overall training latency speedup (4.7×) for both supervised training and self-supervised training tasks. In addition to the training accelerators, this dissertation also presents a CNN inference accelerator on ASIC(FixyNN) and FPGA(FixyFPGA). FixyNN consists of a fixed-weight feature extractor that generates ubiquitous CNN features and a conventional programmable CNN accelerator. In the fixed-weight feature extractor, the network weights are hard-coded into hardware and used as a fixed operand for the multiplication. Experimental results demonstrate FixyNN can achieve very high energy efficiencies up to 26.6 TOPS/W, and FixyFPGA achieves $2.34\times$ higher GOPS on ImageNet classification. In summary, this dissertation comprehensively discusses novel architectures of high-performance and energy-efficient ASIC/FPGA CNN inference/training accelerators.

Coarse-grain reconfigurable architectures (CGRAs) have shown significant improvements as hardware accelerator whilst demanding low power. Such acceleration inherits from the nature of instruction-level parallelism and exploited by many techniques. Modulo scheduling is a popular approach to software pipelining techniques that provides an efficient heuristic to accelerations on loops, repetitive regions of an application. Existing scheduling algorithms for modulo scheduling heuristic persist on loop exiting problems that limit CGRA acceleration to only loops with known trip count and no exit statements. Another notable limitation is the early exit problem, where loops can only terminate after certain iterations as CGRA moves to kernel stage. In attempts to circumvent such obstacles, COMSAT introduces a modified modulo scheduling technique that acts as an external module and can be applied to any existing scheduling/mapping algorithms with minimal hardware changes. Experiments from MiBench and Rodinia benchmark suites have shown that COMSAT achieved an average speedup of 3x in overall benchmarks and 10x speedup in kernel regions. Without COMSAT techniques, only 25% of said loops would have been able to accelerate, reducing benchmark and kernel speedups to 1.25x and 3.63x respectively.

Deep neural networks (DNNs), as a main-stream algorithm for various AI tasks, achieve higher accuracy at the cost of increased computational complexity and model size, posing great challenges to hardware platforms. This dissertation first tackles the design challenges of resistive random-access-memory (RRAM) based in-memory computing (IMC) architectures. A new metric, model stability from the loss landscape, is proposed to help shed light on accuracy under variations and model compression and guide a novel variation-aware training (VAT) solution. The proposed method effectively improves post-mapping accuracy of multiple datasets. Next, a hybrid RRAM/SRAM IMC DNN inference accelerator is developed, that integrates an RRAM-based IMC macro, a reconfigurable SRAM-based multiply-accumulate (MAC) macro, and a programmable shifter. The hybrid IMC accelerator fully recovers the inference accuracy post the mapping. Furthermore, this dissertation researches on architectural optimizations for high IMC utilization, low on-chip communication cost, and low energy-delay product (EDP), including on-chip interconnect design, PE array utilization, and tile-to-router mapping and scheduling. The optimal choice of on-chip interconnect results in up to 6x improvement in energy-delay-area product for RRAM IMC architectures. Furthermore, the PE and NoC optimizations show up to 62% improvement in PE utilization, 78% reduction in area, and 78% lower energy-area product for a wide range of modern DNNs. Finally, this dissertation proposes a novel chiplet-based IMC benchmarking simulator, SIAM, and a heterogeneous chiplet IMC architecture to address the limitations of a monolithic DNN accelerator. SIAM utilizes model-based and cycle-accurate simulation to provide a scalable and flexible architecture. SIAM is calibrated against a published silicon result, SIMBA, from Nvidia. The heterogeneous architecture utilizes a custom mapping with a bank of big and little chiplets, and a hybrid network-on-package (NoP) to optimize the utilization, interconnect bandwidth, and energy efficiency. The proposed big-little chiplet-based RRAM IMC architecture significantly improves energy efficiency at lower area, compared to conventional GPUs. In summary, this dissertation comprehensively investigates novel methods that encompass device, circuits, architecture, packaging, and algorithm to design scalable high-performance and energy-efficient IMC architectures.

Visual navigation is a useful and important task for a variety of applications. As the preva­lence of robots increase, there is an increasing need for energy-­efficient navigation methods as well. Many aspects of efficient visual navigation algorithms have been implemented in the lit­erature, but there is a lack of work on evaluation of the efficiency of the image sensors. In this thesis, two methods are evaluated: adaptive image sensor quantization for traditional camera pipelines as well as new event­-based sensors for low­-power computer vision.The first contribution in this thesis is an evaluation of performing varying levels of sen­sor linear and logarithmic quantization with the task of visual simultaneous localization and mapping (SLAM). This unconventional method can provide efficiency benefits with a trade­ off between accuracy of the task and energy-­efficiency. A new sensor quantization method, gradient­-based quantization, is introduced to improve the accuracy of the task. This method only lowers the bit level of parts of the image that are less likely to be important in the SLAM algorithm since lower bit levels signify better energy­-efficiency, but worse task accuracy. The third contribution is an evaluation of the efficiency and accuracy of event­-based camera inten­sity representations for the task of optical flow. The results of performing a learning based optical flow are provided for each of five different reconstruction methods along with ablation studies. Lastly, the challenges of an event feature­-based SLAM system are presented with re­sults demonstrating the necessity for high quality and high­ resolution event data. The work in this thesis provides studies useful for examining trade­offs for an efficient visual navigation system with traditional and event vision sensors. The results of this thesis also provide multiple directions for future work.

Edge computing applications have recently gained prominence as the world of internet-of-things becomes increasingly embedded into people's lives. Performing computations at the edge addresses multiple issues, such as memory bandwidth-latency bottlenecks, exposure of sensitive data to external attackers, etc. It is important to protect the data collected and processed by edge devices, and also to prevent unauthorized access to such data. It is also important to ensure that the computing hardware fits well within the tight energy and area budgets for the edge devices which are being progressively scaled-down in size. Firstly, a novel low-power smart security prototype chip that combines multiple entropy sources, such as real-time electrocardiogram (ECG) data, and SRAM-based physical unclonable functions (PUF), for authentication and cryptography applications is proposed. Up to ~12X improvement in the equal error rate compared to a prior ECG-only authentication system is achieved by combining feature vectors obtained from ECG, heart rate variability, and SRAM PUF. The resulting vectors can also be utilized for secure cryptography applications. Secondly, a novel in-memory computing (IMC) hardware noise-aware training algorithms that make DNNs more robust to hardware noise is developed and evaluated. Up to 17% accuracy was recovered in deep neural networks (DNNs) deployed on IMC prototype hardware. The noise-aware training principles are also used to improve the adversarial robustness of DNNs, and successfully defend against both adversarial input and weight attacks. Up to ~10\% improvement in robustness against adversarial input attacks, and up to 33% improvement in robustness against adversarial weight attacks are achieved. Finally, a DNN training algorithm that pursues and optimises both activation and weight sparsity simultaneously is proposed and evaluated to obtain highly compressed DNNs. This lead to up to 4.7x reduction in the total number of flops required to perform complex image recognition tasks. A custom sparse inference accelerator is designed and synthesized to evaluate the benefits of the above flop reduction. A speedup of 4.24x is achieved. In summary, this dissertation contains innovative algorithm and hardware design techniques aided by machine learning, which enhance the security and efficiency of edge computing applications.

Coarse-Grained Reconfigurable Arrays (CGRAs) are emerging accelerators that promise low-power acceleration of compute-intensive loops in applications. The acceleration achieved by CGRA relies on the efficient mapping of the compute-intensive loops by the CGRA compiler onto the CGRA. The CGRA mapping problem, being NP-complete, is performed in a two-step process, scheduling, and mapping. The scheduling algorithm allocates timeslots to the nodes of the DFG, and the mapping algorithm maps the scheduled nodes onto the PEs of the CGRA. On a mapping failure, the initiation interval (II) is increased, and a new schedule is obtained for the increased II. Most previous mapping techniques use the Iterative Modulo Scheduling algorithm (IMS) to find a schedule for a given II. Since IMS generates a resource-constrained ASAP (as-soon-as-possible) scheduling, even with increased II, it tends to generate a similar schedule that is not mappable and does not explore the schedule space effectively. The problems encountered by IMS-based scheduling algorithms are explored and an improved randomized scheduling algorithm for scheduling of the application loop to be accelerated is proposed. When encountering a mapping failure for a given schedule, existing mapping algorithms either exit and retry the mapping anew, or recursively remove the previously mapped node to find a valid mapping (backtrack).Abandoning the mapping is extreme, but even backtracking may not be the best choice, since the root of the problem may not be the previous node. The challenges in existing algorithms are systematically analyzed and a failure-aware mapping algorithm is presented. The loops in general-purpose applications are often complicated loops, i.e., loops with perfect and imperfect nests and loops with nested if-then-else's (conditionals). The existing hardware-software solutions to execute branches and conditions are inefficient. A co-design approach that efficiently executes complicated loops on CGRA is proposed. The compiler transforms complex loops, maps them to the CGRA, and lays them out in the memory in a specific manner, such that the hardware can fetch and execute the instructions from the right path at runtime. Finally, a CGRA compilation simulator open-source framework is presented. This open-source CGRA simulation framework is based on LLVM and gem5 to extract the loop, map them onto the CGRA architecture, and execute them as a co-processor to an ARM CPU.

Heterogenous SoCs are in development that marry multiple architectural patterns together. In order for software to be run on such a platform, it must be broken down into its constituent parts, kernels, and scheduled for execution on the hardware. Although this can be done by hand, it would be arduous and time consuming; rather, a tool should be developed that analyzes the source binary, extracts the kernels, schedules the kernels, and optimizes the scheduled kernels for their target component. This dissertation proposes a decidable kernel definition that enables an algorithmic approach to detecting kernels from arbitrary programs. This definition is built upon four constraints that can be tested using basic graph theory. In addition, two algorithms are proposed that successfully extract kernels based upon runtime information. The first utilizes dynamic traces, which are generated using a collection of novel optimizations. The second utilizes a simple affinity matrix, which has no runtime overhead during program execution. Finally, a Dense Neural Network is proposed that is capable of detecting a kernel's archetype based upon only the composition of the source program and the number of times individual basic blocks execute. The contributions proposed in this dissertation provide the necessary infrastructure to perform a litany of other optimizations on kernels. By detecting kernels algorithmically, any program can be analyzed and optimized with techniques that have heretofore required kernels be written in a compatible form. Computational kernels can be extracted from any program with no constraints. The innovations describes here will form the foundation for automated kernel optimization in the future, helping optimize the code of the future.

Due to high DRAM access latency and energy, several convolutional neural network(CNN) accelerators face performance and energy efficiency challenges, which are critical for embedded implementations. As these applications exploit larger datasets, memory accesses of these emerging applications are increasing. As a result, it is difficult to predict the combined dynamic random access memory (DRAM) workload behavior, which can sabotage memory optimizations in software. To understand the impact of external memory access on CNN accelerators which reduces the high DRAMaccess latency and energy, simulators such as RAMULATOR and VAMPIRE have been proposed in prior work. In this work, we utilize these simulators to benchmark external memory access in CNN accelerators. Experiments are performed generating trace files based on the number of parameters and data precision and also using trace file generated for CNN Accelerator Altera Arria 10 GX 1150 FPGA data to complete the end to end workflow using the mentioned simulators. Besides that, certain modifications were made in the default VAMPIRE code to implement certain functionalities such as PREA(Precharge All) and REF(Refresh). Then, precalculated energies were computed for DDR3, DDR4, and HBM based on the micron model to mention it in the dram specification file inputted to the VAMPIRE tool. An experimental study was performed and a comparison is made between DDR3, DDR4, and HBM, it was proved that DDR4 is nearly 31% more energy-efficient than DDR3 and HBMis 54% energy-efficient than DDR3. Performed modeling and experimental analysis on a large set of data and then split it into a set of data and compared the results of the small sets multiplied with the number of sets and the large data set and concluded that the results were nearly the same. Finally, a GUI is developed by wrapping both the simulators. GUI provides user-friendly access and can analyze the parameters without much prior knowledge and understanding of the working.

QR decomposition (QRD) of a matrix is one of the most common linear algebra operationsused for the decomposition of a square
on-square matrix. It has a wide range
of applications especially in Multiple Input-Multiple Output (MIMO) communication
systems. Unfortunately it has high computation complexity { for matrix size of nxn,
QRD has O(n3) complexity and back substitution, which is used to solve a system
of linear equations, has O(n2) complexity. Thus, as the matrix size increases, the
hardware resource requirement for QRD and back substitution increases signicantly.
This thesis presents the design and implementation of a
exible QRD and back substitution accelerator using a folded architecture. It can support matrix sizes of
4x4, 8x8, 12x12, 16x16, and 20x20 with low hardware resource requirement.
The proposed architecture is based on the systolic array implementation of the
Givens algorithm for QRD. It is built with three dierent types of computation blocks
which are connected in a 2-D array structure. These blocks are controlled by a
scheduler which facilitates reusability of the blocks to perform computation for any
input matrix size which is a multiple of 4. These blocks are designed using two
basic programming elements which support both the forward and backward paths to
compute matrix R in QRD and column-matrix X in back substitution computation.
The proposed architecture has been mapped to Xilinx Zynq Ultrascale+ FPGA
(Field Programmable Gate Array), ZCU102. All inputs are complex with precision
of 40 bits (38 fractional bits and 1 signed bit). The architecture can be clocked at
50 MHz. The synthesis results of the folded architecture for dierent matrix sizes
are presented. The results show that the folded architecture can support QRD and
back substitution for inputs of large sizes which otherwise cannot t on an FPGA
when implemented using a
at architecture. The memory sizes required for dierent
matrix sizes are also presented.