{"title":"A Virtual GPU as Developer-Friendly OpenMP Offload Target","authors":"A. Patel, Shilei Tian, J. Doerfert, B. Chapman","doi":"10.1145/3458744.3473356","DOIUrl":"https://doi.org/10.1145/3458744.3473356","url":null,"abstract":"While parallel programming is hard, programming accelerators has always been even more complicated. One fundamental reason is the lack of mature tooling that can be used to inspect a program that executes on two different architectures. As GPU software stacks of different vendors provide vastly different experience for developers, it is clear that the gold standard for debugging is still host (CPU) execution with its myriad of mature tooling options. In this work we present a virtual GPU (VGPU) OpenMP offloading target that allows to emulate a GPU execution environment on the host. In contrast to classical “host offloading”, the VGPU target reuses the same execution model, compilation paths, and runtimes as a physical GPU. While this execution mode is not able to perform as good as host-specific compilation, runtimes, and execution, it provides the developor with a more accurate stand-in for GPU offloading that is still amendable to existing host tooling.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"126 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128055028","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Constraint Solving by Quantum Annealing","authors":"P. Codognet","doi":"10.1145/3458744.3473364","DOIUrl":"https://doi.org/10.1145/3458744.3473364","url":null,"abstract":"We present experiments in solving combinatorial optimization and constraint satisfaction problems by means of Quantum Annealing. We describe how to model classic constraint problems such as N-queens and Magic Square as well as hard combinatorial problems such as the Costas Array Problem or the Quadratic Assignment Problem in terms of QUBO (Quadratic Unconstrained Binary Optimization). QUBO is the input language of quantum computers based on quantum annealing such as the D-Wave systems and of the ”quantum-inspired” but classical devices such as Fujitsu’s Digital Annealing Unit or Hitachi’s CMOS Annealing Machine. We present preliminary results for solving these combinatorial optimization and constraint satisfaction problems by implementation on the D-Wave quantum computer.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"118 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128128140","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Che-Chia Lin, Chao-Lin Lee, Jenq-Kuen Lee, Howard Wang, Ming-Yu Hung
{"title":"Accelerate Binarized Neural Networks with Processing-in-Memory Enabled by RISC-V Custom Instructions","authors":"Che-Chia Lin, Chao-Lin Lee, Jenq-Kuen Lee, Howard Wang, Ming-Yu Hung","doi":"10.1145/3458744.3473351","DOIUrl":"https://doi.org/10.1145/3458744.3473351","url":null,"abstract":"As the speed of processing units grows rapidly, the bottleneck of system’s performance is usually the speed of memory, and the situation is the so-called ”Memory Wall”. There are emerging technologies trying to take down the ”Memory Wall”, and one of them is Processing-in-Memory (PIM). Processing-in-Memory means that the data are processed just inside the memory itself. It does not need to take time to travel between CPU and Memory. Moreover, for very little modifications to memory devices, the memory can do primitive bit-wise operations at the memory side. Binarized Neural Network (BNN), which replaces the convolution’s multiplication and addition operations with bit-wise AND and population count operations, is therefore suited for utilizing PIM to gain performance. This work architects PIM AND, NOT, and population count operations and enables PIM operations working under RISC-V custom instruction encodings. Besides, we also utilize TVM’s support of BNN for application sources. In addition, we offer a new design for BNN’s convolution in which a better memory layout is considered. With our design, the results of the speedup range from 3.7x to 57.3x comparing with CPU-based system for the execution time of end-to-end BNN model inferences.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128130316","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Intra- and Inter- Layer Transformation to Reduce Memory Traffic for CNN Computation","authors":"Pin-Wei Liao, W. Hsu, Shih-Wei Liao","doi":"10.1145/3458744.3473353","DOIUrl":"https://doi.org/10.1145/3458744.3473353","url":null,"abstract":"Edge inference has gained much popularity in recent years. Many AI accelerators have been proposed and extensively studied. Such devices are often packed with a large number of PEs (Processing Elements), and lots of on-chip SRAM. The key to successful AI acceleration is to effectively use the data transferred from off-chip DRAM to the on-chip SRAM. Most prior studies optimize the use of on-chip SRAM for a single convolution layer, they tend to ignore the opportunity of inter-layer data reuse. We have proposed an algorithm to schedule two adjacent layers of CNN operations. Our goal is to reduce traffic between DRAM and local memory more than allocating the buffer to only a single layer. Our cross-layer scheduling effectively reduces the memory traffic. We hav also verified the validity of our memory traffic reduction model on the Gemmini simulator from UC Berkeley.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"39 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125873681","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Advancing OpenMP Offload Debugging Capabilities in LLVM","authors":"J. Doerfert, Joseph Huber, Melanie Cornelius","doi":"10.1145/3458744.3473358","DOIUrl":"https://doi.org/10.1145/3458744.3473358","url":null,"abstract":"Debugging an application is famously twice as hard as writing the application in the first place. While this sentiment predates modern GPU programming by decades, it is all the more true when the application has to manage computation and memory across different architectures, memory spaces, and execution modes. Any subtle error, whether in the application, the compiler, or runtime system, can lead to unexpected behavior that is hard to understand from the program output alone. While some tooling solutions for GPU debugging exist, their maturity and usefulness varies gravely between vendors. Furthermore, as OpenMP offloading puts an abstraction layer between the programmer and the underlying hardware, the information from a native GPU driver (debugging tool) is not always transferable to the OpenMP programming model. As the OpenMP Tooling [12] (OMPT) and Debug [4] (OMPD) interfaces are still not ready to debug OpenMP offloading code in production, developers have a hard time to comprehend the implementation state, error sources, and interplay of the OpenMP world with the foreign device runtimes, e.g., CUDA. In this work we present our efforts towards better debug information for OpenMP offloading through the LLVM compiler. Our developer-centric extensions allow us to inspect the state of the OpenMP offloading runtimes and their interaction with the foreign runtimes. In contrast to partial solutions employed before, we provide the ability to select the reported information through a fine-granular and flexible system without introducing substantial overheads. Furthermore, we present a software-managed stack trace feature for OpenMP offload kernels that works even without the notoriously fragile debug information in the GPU code.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"16 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123098220","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Tarindu Jayatilaka, Hideto Ueno, G. Georgakoudis, Eunjung Park, J. Doerfert
{"title":"Towards Compile-Time-Reducing Compiler Optimization Selection via Machine Learning","authors":"Tarindu Jayatilaka, Hideto Ueno, G. Georgakoudis, Eunjung Park, J. Doerfert","doi":"10.1145/3458744.3473355","DOIUrl":"https://doi.org/10.1145/3458744.3473355","url":null,"abstract":"Compilers come with a multitude of optimizations to choose from, and the chosen optimizations significantly affect the performance of the code being optimized. Selecting the optimal set of optimizations to apply and determining the order to run them is non-trivial. All of these optimizations closely interact with each other, and an optimization’s ability to improve the code heavily depends on how the previous optimizations modified it. The current approach to solve this is to use a one-size-fits-all optimization sequence, that is designed to perform reasonably well for any given source code. In other words, they are not designed to optimize depending on the nature of the code, which usually results in sub-optimal performance. In this paper, we present preliminary work tackling the problem from the perspective of compile-time by adapting the optimization sequence to cater to different program types. We start by analyzing how the source code interacts with the passes, as well as how the passes interact with each other. We use this information and propose two potential methods driven by machine learning to run customized optimization sequences on the source code. First, we look at how we can use a neural network to predict and skip passes that do not optimize the code to improve compilation time. Second, we look at how we can use clustering and predictive models to select custom pass pipelines. We believe that our approach has the potential to replace the current one-size-fits-all approach, with better optimization sequences that are tailored to perform better depending on the code. At the same time, it will allow testing the potential pipelines thoroughly, a practical requirement to gain confidence in the correctness of the compiler.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"22 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131026790","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Samit Shivadekar, Jayalakshmi Mangalagiri, P. Nguyen, David Chapman, M. Halem, Rahul Gite
{"title":"An Intelligent Parallel Distributed Streaming Framework for near Real-time Science Sensors and High-Resolution Medical Images","authors":"Samit Shivadekar, Jayalakshmi Mangalagiri, P. Nguyen, David Chapman, M. Halem, Rahul Gite","doi":"10.1145/3458744.3474039","DOIUrl":"https://doi.org/10.1145/3458744.3474039","url":null,"abstract":"Our goals are to address challenges such as latency, scalability, throughput and heterogeneous data sources of streaming analytics and deep learning pipelines in science sensors and medical imaging applications. We present a prototype Intelligent Parallel Distributed Streaming Framework (IPDSF) that is capable of distributed streaming processing as well as performing distributed deep training in batch mode. IPDSF is designed to run streaming Artificial Intelligent (AI) analytic tasks using data parallelism including partitions of multiple streams of short time sensing data and high-resolution 3D medical images, and fine grain tasks distribution. We will show the implementation of IPDSF for two real world applications, (i) an Air Quality Index based on near real time streaming of aerosol Lidar backscatter and (ii) data generation of Covid-19 Computing Tomography (CT) scans using deep learning. We evaluate the latency, throughput, scalability, and quantitative evaluation of training and prediction compared against a baseline single instance. As the results, IPDSF scales to process thousands of streaming science sensors in parallel for Air Quality Index application. IPDSF uses novel 3D conditional Generative Adversarial Network (cGAN) training using parallel distributed Graphic Processing Units (GPU) nodes to generate realistic 3D high resolution Computed Tomography scans of Covid-19 patient lungs. We will show that IPDSF can reduce cGAN training time linearly with the number of GPUs.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129030164","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Implementing Arbitrary/Common Concurrent Writes of CRCW PRAM","authors":"Fady Ghanim, W. Elwasif, D. Bernholdt","doi":"10.1145/3458744.3474041","DOIUrl":"https://doi.org/10.1145/3458744.3474041","url":null,"abstract":"The Parallel Random Access Machines (PRAM) abstraction is the simplest and most elegant algorithmic model for the design and analysis of parallel algorithms. It consists of different models categorized based on the underlying memory access mode used, the most powerful of which is the Concurrent Read Concurrent Write (CRCW) model. A PRAM algorithm describes a series of rounds, each of which consists of a collection of operations that can be executed concurrently within the same time step. However, the lack of support for concurrent memory accesses and the prevalence of asynchronous programming models led to the belief that implementing CRCW PRAM algorithms is unattainable and prompted many to avoid this model except for theoretical studies of optimal performance. In this work, we study the arbitrary and common concurrent writes in the CRCW PRAM model and explore implementation challenges on general-purpose systems. Moreover, we examine current practices for implementing common/arbitrary concurrent writes and propose a new efficient lightweight and thread-safe method to implement concurrent writes through leveraging atomic instructions. To demonstrate the efficacy of our method, we developed OpenMP kernels for classical CRCW PRAM algorithms and provide experimental results and comparisons based on run time performance measured over the x86 multicore architecture. Our results show a performance speedup compared to current practices up to 4.5x across all our benchmarks.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"102 7","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"120982937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Warp-centric K-Nearest Neighbor Graphs construction on GPU","authors":"B. Meyer, A. Pozo, Wagner M. Nunan Zola","doi":"10.1145/3458744.3474053","DOIUrl":"https://doi.org/10.1145/3458744.3474053","url":null,"abstract":"Recent advances and applications of machine learning algorithms are becoming more common in different fields. It is expected that some applications require the processing of large datasets with those algorithms, which leads to high computational costs. Massively parallel GPU methods can be applied to surpass this limitation and reduce the execution time of these algorithms. The construction of approximate K-Nearest Neighbor Graphs (K-NNG) is frequently required for similarity search or other applications such as the t-SNE dimensionality reduction technique. The K-NNG represents the K closest points (neighbors) for each point in a set. In this paper, we propose and analyze an all-points K-Nearest Neighbor Graph construction algorithm on GPU called Warp-centric K-NNG (w-KNNG), which is based on the Random Projection Forest method. Usually, the construction or search for k-NN sets for high dimensional points presents challenges for its implementation on many-core processing units, due to the space limitation in maintaining these sets in high speed shared memory. We present three warp-centric approaches for our algorithm that efficiently search and maintain the k-NN high dimensional point sets in global memory. In our experiments, the new methods allows the algorithm to achieve up to 639% faster execution when compared to the state-of-the-art FAISS library, considering an equivalent accuracy of approximate K-NNG. One of the new strategies (w-KNNG atomic) is more successful when applied to a smaller number of dimensions, while the tiled w-KNNG approach was successful in general scenarios for higher dimensional points.","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"26 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121200355","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Eduardo Romero-Gainza, Christopher Stewart, Angela Li, Kyle C. Hale, Nathaniel Morris
{"title":"Memory Mapping and Parallelizing Random Forests for Speed and Cache Efficiency","authors":"Eduardo Romero-Gainza, Christopher Stewart, Angela Li, Kyle C. Hale, Nathaniel Morris","doi":"10.1145/3458744.3474052","DOIUrl":"https://doi.org/10.1145/3458744.3474052","url":null,"abstract":"Memory mapping enhances decision tree implementations by enabling constant-time statistical inference, and is particularly effective when memory mapped tables fit in processor cache. However, memory mapping is more challenging when applied to random forests—ensembles of many trees—as the table sizes can easily outstrip cache capacity. We argue that careful system design for parallel and cache efficiency can make memory mapping effective for random forests. Our preliminary results show memory-mapped forests can speed up inference latency by a factor of up to 30 × .","PeriodicalId":204366,"journal":{"name":"50th International Conference on Parallel Processing Workshop","volume":"60 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126262053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}