|
Graphical Processing Unit (GPU) Software to Accelerate Underwater Acoustic Autonomous Modeling and Processing
Navy SBIR 2013.2 - Topic N132-119 NAVSEA - Mr. Dean Putnam - [email protected] Opens: May 24, 2013 - Closes: June 26, 2013 N132-119 TITLE: Graphical Processing Unit (GPU) Software to Accelerate Underwater Acoustic Autonomous Modeling and Processing. TECHNOLOGY AREAS: Sensors, Electronics, Battlespace ACQUISITION PROGRAM: PMS485, Maritime Surveillance Systems, Distributed Sensors Group – Shallow RESTRICTION ON PERFORMANCE BY FOREIGN CITIZENS (i.e., those holding non-U.S. Passports): This topic is "ITAR Restricted". The information and materials provided pursuant to or resulting from this topic are restricted under the International Traffic in Arms Regulations (ITAR), 22 CFR Parts 120 - 130, which control the export of defense-related material and services, including the export of sensitive technical data. Foreign Citizens may perform work under an award resulting from this topic only if they hold the "Permanent Resident Card", or are designated as "Protected Individuals" as defined by 8 U.S.C. 1324b(a)(3). If a proposal for this topic contains participation by a foreign citizen who is not in one of the above two categories, the proposal will be rejected. OBJECTIVE: This topic seeks innovation to develop a massively parallel architecture for autonomous acoustic processing to increase significantly the processing capability for new Navy autonomous underwater sensing systems. DESCRIPTION: The Navy seeks to reduce operating and maintenance costs of high-endurance autonomous platforms, which have tremendous potential for persistent monitoring of surface and subsurface acoustic targets. The Navy already has the capability of collecting large volumes of passive acoustic data; however, the data is not analyzed until weeks after collection. Today, tactical and strategic capabilities would benefit not only from immediate processing, classification, and reporting capabilities, but also from reducing the cost of highly skilled manpower associated with evaluation of passive acoustic data. To adequately process low-frequency passive acoustic data in a way that matches the actual three-dimensional (3-D) physical environment, past processing methods either relied on high- power, large footprint hardware systems or gave up physical resolution to enable lower power, smaller footprint distributed-type systems. Various hardware architectures have been used to maximize computations per watt, including field-programmable gate arrays (FPGAs) ("reconfigurable computing"), Digital Signal Processing (DSP) chips, Application Specific Integrated Circuits (ASIC), and hybrids of each. Such architectures often present implementation challenges, leading to high cost of initial implementation and limited flexibility to add incremental capabilities. Parallel processing is well known to be a practical way to optimize GFLOPs/watt, (G for Giga or billions; FLOPs for floating point operations per second), but requires innovative algorithmic approaches to fully exploit the new parallel architectures, particularly in orchestrating the movement of data through a hierarchy of large, slow, off-chip memory to small, fast, on-chip memory (see Ref. 1). Programming Graphical Processing Units (GPUs) for scientific computations used to require coding in graphic shader languages (see Ref. 2 for an early example of a parabolic equation model, implemented with shaders). However, GPUs developed to meet the needs of the Personal Computer (PC) gaming world have emerged as a disruptive technology in high performance computing. These include the advent of new high-level language software interfaces to GPU hardware, such as Compute Unified Device Architecture (CUDA) or Open Computing Language (OpenCL) (see Refs 3 and 4). Massively Parallel Processing is a multiprocessing architecture that uses many processors and a different programming paradigm than the common symmetric multiprocessing (SMP) found in today's computer systems. Today's GPUs are massively parallel computers that are ubiquitous in modestly priced desktops, laptops, and even mobile phones and tablets. This greatly increased computing power offers the potential for realizing innovative new capabilities in Navy underwater acoustic modeling and processing, previously infeasible due to the limitations of having only a few processing cores to commit to the task. To realize this potential will require development of a new multiprocessing architecture and creative adaptation of existing algorithms (acoustic signal processing and automation or acoustic modeling) to take advantage of the distinct and evolving software and hardware architecture being developed for GPUs. This effort seeks to significantly increase (by a factor of at least 5) the amount of computations performed, while maintaining or decreasing the power requirement as compared to conventional processing hardware for either acoustic signal processing and automation or acoustic performance prediction modeling. Examples of such applications that are pushing acoustic signal processing and automation include adaptive beam forming, environmentally adaptive signal processing, broader frequency bandwidths, and multiple detection surfaces, matched field processing, and increased frequency resolution. Similarly, such applications that are pushing the acoustic performance prediction modeling envelope include broader frequency bandwidths, higher and lower frequencies, range-dependence, 3D waveguides (bathymetric canyons, internal waves), multi-static reverberation, platform and scattering surface motion. All of these phenomena have a significant impact on sonar system performance (Ref. 5). The Navy is seeking innovative multiprocessor architectures and processing that is enabled by the increased computational power provided by GPUs. The specific effort would be focused on developing a massively parallel processing architecture and then adapting acoustic signal processing and automation or acoustic performance prediction modeling algorithms for this architecture and then implementing them to run on GPU hardware. Goals are to 1) accelerate the selected application by at least 5-times (unit-processor execution time/parallel execution time); and 2) demonstrate the scalability (the ability to maintain performance gain when system and problem size increase) of the solution. PHASE I: The company will develop concepts for autonomous acoustic signal processing for contact classification or acoustic performance prediction modeling that exploits the parallelism provided by GPUs. The company will develop the concepts for functional and physical implementation in multiprocessor architectures using one or more Commercial-of-the-Shelf (COTS) GPUs. Feasibility of this architecture to meet the technical requirements listed previously will be verified via analysis and modeling. The company will provide a Phase II development plan that addresses technical risk reduction and provides performance goals and key technical milestones. PHASE II: Based on the results of Phase I and the Phase II development plan, the company will develop a prototype autonomous acoustic processing system implemented in a multiprocessor architecture using one or more Commercial-of-the-Shelf (COTS) GPUs identified in Phase I. The company will demonstrate it with Government furnished data. The company will evaluate the system to determine its capability to meet the performance goals defined in the Phase II development plan. The company will also prepare a Phase III development plan to transition the technology to Navy use. PHASE III: In Phase III, the company will support transitioning the technology to Navy use. The company will continue to refine the GPU processing suite and algorithms for evaluation in an operationally relevant environment. Specifically, the prototype developed in Phase II will be integrated with a suitable surveillance sonar system and used in a proof-of-concept test to determine overall systems effectiveness including power consumption, form-factor, reliability, and persistence. PRIVATE SECTOR COMMERCIAL POTENTIAL/DUAL-USE APPLICATIONS: Many of the acoustic processing tasks that will be successfully accelerated on GPUs through this SBIR effort will find use in the commercial sector in physics based modeling, gaming theory, and visual rendering. REFERENCES: 2. Gunderson, Steinar. "GPUwave." 19 Jan. 2007. 13 Nov. 2012. <http://gpuwave.sesse.net/>. 3. "NVIDIA CUDA Zone". NVIDIA Corporation. 13 Nov. 2012. <http://developer.nvidia.com/category/zone/cuda-zone>. 4. "OpenCL – the open standard for parallel programming of heterogeneous systems". Khronos Group: Connecting Software to Silicon. Khronos Group. 13 Nov. 2012. <http://www.khronos.org/opencl>. 5. Jensen, Finn; Kuperman, William; Porter, Michael; and Schmidt, Henrik. Computational Ocean Acoustics, Second Edition. New York, NY: Springer Science and Business Media, June 14, 2011. KEYWORDS: Graphical Processing Unit; passive acoustic processing; parallel processing; field-programmable gate arrays; OpenCL; acoustic performance predictions; bathymetric canyons
|