OpenVX is a computer vision framework that enables embedded and real-time applications to optimize computer vision processing for performance and power. OpenVX addresses system-level optimizations by making use of a graph-based computational API. Although this gives a clear advantage over other traditional computer vision libraries such as OpenCV, which mainly addresses kernel-level optimizations, OpenVX still relies on vendor implementations to optimize individual built-in kernels. OpenVX implements several computer vision kernels but in order to increase adoption and user flexibility, OpenVX added support for C based user-kernels, which by default are single-threaded and there is no particular way to accelerate kernels or offload the computation to an accelerator such us a GPU. The user has to do the heavy lifting of supporting a multi-threaded implementation. We propose two different OpenVX API extensions to allow developers deploy accelerated user-kernels using OpenCL.
{"title":"OpenCL Interoperability with OpenVX Graphs","authors":"Ben Ashbaugh, A. Bernal","doi":"10.1145/3078155.3078183","DOIUrl":"https://doi.org/10.1145/3078155.3078183","url":null,"abstract":"OpenVX is a computer vision framework that enables embedded and real-time applications to optimize computer vision processing for performance and power. OpenVX addresses system-level optimizations by making use of a graph-based computational API. Although this gives a clear advantage over other traditional computer vision libraries such as OpenCV, which mainly addresses kernel-level optimizations, OpenVX still relies on vendor implementations to optimize individual built-in kernels. OpenVX implements several computer vision kernels but in order to increase adoption and user flexibility, OpenVX added support for C based user-kernels, which by default are single-threaded and there is no particular way to accelerate kernels or offload the computation to an accelerator such us a GPU. The user has to do the heavy lifting of supporting a multi-threaded implementation. We propose two different OpenVX API extensions to allow developers deploy accelerated user-kernels using OpenCL.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"91 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115105581","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}
In the machine learning domain, machine learning frameworks are predominantly written and maintained in NVIDIA® CUDA™ language. There have been attempts to port these frameworks to OpenCL®, notably the ports of Caffe framework by Gu et al; Tschopp; and Engel; and of Torch framework by Perkins. The authors of these frameworks found merging their work into the mainstream framework challenging, and maintain their forks as separate branches or repositories. CUDA-on-CL addresses this problem by leaving the reference implementation entirely in NVIDIA CUDA, both host-side and device-side, and providing a compiler and a runtime component, so that any CUDA C++11 application can in theory be compiled and run on any OpenCL 1.2 device. We use Tensorflow framework as a case-study, and demonstrate the ability to run unary, binary and reduction Tensorflow and Eigen kernels, with no modification to the original CUDA source-code. Performance studies are undertaken, using the Tensorflow kernels. For buffer sizes of 1MB or more, performance is comparable between CUDA and CUDA-on-CL, across unary operations, binary operations and single-axis reductions. Full reduction is around 14 times slower on CUDA-on-CL than on CUDA. We think this may be because of the absence of the low-level hardware shfl operation. The asymptotic time for zero buffer sizes is double that of CUDA, possibly because of the overhead of additional kernel boilerplate needed to workaround limitations in the OpenCL 1.2 standard.
在机器学习领域，机器学习框架主要使用NVIDIA®CUDA™语言编写和维护。已经有人尝试将这些框架移植到OpenCL®，特别是Gu等人对Caffe框架的移植;Tschopp;和恩格尔;以及Perkins的Torch框架。这些框架的作者发现将他们的工作合并到主流框架中是一项挑战，并将他们的分支作为单独的分支或存储库进行维护。CUDA-on- cl通过将参考实现完全保留在NVIDIA CUDA(主机端和设备端)中解决了这个问题，并提供了编译器和运行时组件，因此任何CUDA c++ 11应用程序理论上都可以在任何OpenCL 1.2设备上编译和运行。我们使用Tensorflow框架作为案例研究，并演示了在不修改原始CUDA源代码的情况下运行一元，二进制和约简Tensorflow和特征核的能力。使用Tensorflow核进行性能研究。对于1MB或更大的缓冲区大小，CUDA和CUDA-on- cl之间的性能在一元操作、二进制操作和单轴缩减方面是相当的。完全还原在CUDA-on- cl上比在CUDA上慢14倍左右。我们认为这可能是因为缺少底层硬件shfl操作。零缓冲区大小的渐近时间是CUDA的两倍，可能是因为需要额外的内核样板的开销来解决OpenCL 1.2标准中的限制。
{"title":"CUDA-on-CL: a compiler and runtime for running NVIDIA® CUDA™ C++11 applications on OpenCL™ 1.2 Devices","authors":"Hugh Perkins","doi":"10.1145/3078155.3078156","DOIUrl":"https://doi.org/10.1145/3078155.3078156","url":null,"abstract":"In the machine learning domain, machine learning frameworks are predominantly written and maintained in NVIDIA® CUDA™ language. There have been attempts to port these frameworks to OpenCL®, notably the ports of Caffe framework by Gu et al; Tschopp; and Engel; and of Torch framework by Perkins. The authors of these frameworks found merging their work into the mainstream framework challenging, and maintain their forks as separate branches or repositories. CUDA-on-CL addresses this problem by leaving the reference implementation entirely in NVIDIA CUDA, both host-side and device-side, and providing a compiler and a runtime component, so that any CUDA C++11 application can in theory be compiled and run on any OpenCL 1.2 device. We use Tensorflow framework as a case-study, and demonstrate the ability to run unary, binary and reduction Tensorflow and Eigen kernels, with no modification to the original CUDA source-code. Performance studies are undertaken, using the Tensorflow kernels. For buffer sizes of 1MB or more, performance is comparable between CUDA and CUDA-on-CL, across unary operations, binary operations and single-axis reductions. Full reduction is around 14 times slower on CUDA-on-CL than on CUDA. We think this may be because of the absence of the low-level hardware shfl operation. The asymptotic time for zero buffer sizes is double that of CUDA, possibly because of the overhead of additional kernel boilerplate needed to workaround limitations in the OpenCL 1.2 standard.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130494139","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}
Khronos open source components, including the ICD and Clang compiler, require significant time and effort to manually download, build, and install. Source code updates to these components require recompilation, and developers must repeat error-prone steps to build new test environments. Ideally developers should be able to use a tool that automatically obtains, builds, and installs OpenCL codes, libraries, and tools. The Windsor Build and Testing Framework (WBTF) is a tool that has been developed at the University of Windsor that does this. This paper will discuss how the WBTF works, demonstrate how it is used, will show how OpenCL C and C++ programs can be built, run, and/or used to perform various header-only, link, and/or various conformance-style tests using OpenCL reference, host-installed, or using device-installed header and libraries. Those interested in OpenCL C/C++ development, the Khronos OpenCL Clang compiler, and in writing conformance tests will be interested in this framework.
{"title":"The Windsor Build and Testing Framework","authors":"Shane M. Peelar, P. Preney","doi":"10.1145/3078155.3078184","DOIUrl":"https://doi.org/10.1145/3078155.3078184","url":null,"abstract":"Khronos open source components, including the ICD and Clang compiler, require significant time and effort to manually download, build, and install. Source code updates to these components require recompilation, and developers must repeat error-prone steps to build new test environments. Ideally developers should be able to use a tool that automatically obtains, builds, and installs OpenCL codes, libraries, and tools. The Windsor Build and Testing Framework (WBTF) is a tool that has been developed at the University of Windsor that does this. This paper will discuss how the WBTF works, demonstrate how it is used, will show how OpenCL C and C++ programs can be built, run, and/or used to perform various header-only, link, and/or various conformance-style tests using OpenCL reference, host-installed, or using device-installed header and libraries. Those interested in OpenCL C/C++ development, the Khronos OpenCL Clang compiler, and in writing conformance tests will be interested in this framework.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"15 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133213902","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}
Agent-based modeling (ABM) is a bottom-up modeling approach, where each entity of the system being modeled is uniquely represented as a self-determining agent. Large scale emergent behavior in ABMs is population sensitive. As such, it is advisable that the number of agents in a simulation is able to reflect the reality of the system being modeled. This means that in domains such as social modeling, ecology, and biology, systems can contain millions or billions of individuals. Such large scale simulations are only feasible in non-distributed scenarios when the computational power of commodity processors, such as GPUs and multi-core CPUs, is fully exploited. In this paper we evaluate the feasibility of using CPU-oriented OpenCL for high-performance simulations of agent-based models. We compare a CPU-oriented OpenCL implementation of a reference ABM against a parallel Java version of the same model. We show that there are considerable gains in using CPU-based OpenCL for developing and implementing ABMs, with speedups up to 10x over the parallel Java version on a 10-core hyper-threaded CPU.
{"title":"Assessing the feasibility of OpenCL CPU implementations for agent-based simulations","authors":"Nuno Fachada, A. Rosa","doi":"10.1145/3078155.3078174","DOIUrl":"https://doi.org/10.1145/3078155.3078174","url":null,"abstract":"Agent-based modeling (ABM) is a bottom-up modeling approach, where each entity of the system being modeled is uniquely represented as a self-determining agent. Large scale emergent behavior in ABMs is population sensitive. As such, it is advisable that the number of agents in a simulation is able to reflect the reality of the system being modeled. This means that in domains such as social modeling, ecology, and biology, systems can contain millions or billions of individuals. Such large scale simulations are only feasible in non-distributed scenarios when the computational power of commodity processors, such as GPUs and multi-core CPUs, is fully exploited. In this paper we evaluate the feasibility of using CPU-oriented OpenCL for high-performance simulations of agent-based models. We compare a CPU-oriented OpenCL implementation of a reference ABM against a parallel Java version of the same model. We show that there are considerable gains in using CPU-based OpenCL for developing and implementing ABMs, with speedups up to 10x over the parallel Java version on a 10-core hyper-threaded CPU.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124142591","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}
In this technical session we present the open architectural design of the debugger and how it fits into the OpenCL JIT compilation flow. We demonstrate a show case on how to natively work with the debugger to solve functional bugs, as-well-as low-level debugging techniques on SIMD thread level which help to solve complex issues such as misaligned or out of range accesses to local/global memory, stack overflows, Illegal instructions, etc. Finally, we cover the challenges in debugging.
{"title":"Challenges and Opportunities in Native GPU Debugging","authors":"Jeff McAllister, Uri Levy","doi":"10.1145/3078155.3078158","DOIUrl":"https://doi.org/10.1145/3078155.3078158","url":null,"abstract":"In this technical session we present the open architectural design of the debugger and how it fits into the OpenCL JIT compilation flow. We demonstrate a show case on how to natively work with the debugger to solve functional bugs, as-well-as low-level debugging techniques on SIMD thread level which help to solve complex issues such as misaligned or out of range accesses to local/global memory, stack overflows, Illegal instructions, etc. Finally, we cover the challenges in debugging.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115260310","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}
A. Ling, U. Aydonat, Shane O'Connell, D. Capalija, Gordon R. Chiu
After decades of research, High-Level Synthesis has finally caught on as a mainstream design technique for FPGAs. However, achieving performance results that are comparable to designing at a hardware description level still remains a challenge. In this talk, we illustrate how we achieve world class performance results on HPC applications by using OpenCL. Specifically, we show how we achieve 1Tflop of performance on a matrix multiply and over 1.3Tflops on a CNN application, run on Intel's 20nm Arria 10 FPGA device. By leveraging specific coding styles, we show how you can achieve peak performance on the FPGA without having to resort to tedious hardware design languages. Finally, we will describe spatial coding techniques that lead to efficient structures, such as systolic-arrays, to ensure that the FPGA runs efficiently.
{"title":"Creating High Performance Applications with Intel's FPGA OpenCL™ SDK","authors":"A. Ling, U. Aydonat, Shane O'Connell, D. Capalija, Gordon R. Chiu","doi":"10.1145/3078155.3078169","DOIUrl":"https://doi.org/10.1145/3078155.3078169","url":null,"abstract":"After decades of research, High-Level Synthesis has finally caught on as a mainstream design technique for FPGAs. However, achieving performance results that are comparable to designing at a hardware description level still remains a challenge. In this talk, we illustrate how we achieve world class performance results on HPC applications by using OpenCL. Specifically, we show how we achieve 1Tflop of performance on a matrix multiply and over 1.3Tflops on a CNN application, run on Intel's 20nm Arria 10 FPGA device. By leveraging specific coding styles, we show how you can achieve peak performance on the FPGA without having to resort to tedious hardware design languages. Finally, we will describe spatial coding techniques that lead to efficient structures, such as systolic-arrays, to ensure that the FPGA runs efficiently.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121932909","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}
The Production-CL library for iterative scientific calculations with OpenCL is presented. The main goal is to get rid of long repeating lines of standard code which slow down the development process, and realize the typical workflow elements for simulation of physics problems. Main entities of PCL library are: (i) kernel (called with single line resembling CUDA kernel invocation) and (ii) batch of kernels (to help constructing complex step of each iteration). In addition, PCL realizes the procedures standard for scientific calculations 'in production': typical cycle of iterations with main step and regular save/load the whole state, to save work. As an example of library application, we show and compare several projects developed with different approaches.
{"title":"Production-CL library for iterative scientific calculations","authors":"P. Kartsev","doi":"10.1145/3078155.3078162","DOIUrl":"https://doi.org/10.1145/3078155.3078162","url":null,"abstract":"The Production-CL library for iterative scientific calculations with OpenCL is presented. The main goal is to get rid of long repeating lines of standard code which slow down the development process, and realize the typical workflow elements for simulation of physics problems. Main entities of PCL library are: (i) kernel (called with single line resembling CUDA kernel invocation) and (ii) batch of kernels (to help constructing complex step of each iteration). In addition, PCL realizes the procedures standard for scientific calculations 'in production': typical cycle of iterations with main step and regular save/load the whole state, to save work. As an example of library application, we show and compare several projects developed with different approaches.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129256312","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}
MATLAB is a high-level language used in various scientific and engineering fields. Deployment of well-tested MATLAB code to production would be highly desirable, but in practice a number of obstacles prevent this, notably performance and portability. Although MATLAB-to-C compilers exist, the performance of the generated C code may not be sufficient and thus it is important to research alternatives, such as CPU parallelism, GPGPU computing and FPGAs. OpenCL is an API and programming language that allows targeting these devices, hence the motivation for MATLAB-to-OpenCL compilation. In this paper, we describe our recent efforts on offloading code to OpenCL devices in the context of our MATLAB to C/OpenCL compiler.
MATLAB是一种用于各种科学和工程领域的高级语言。将经过良好测试的MATLAB代码部署到生产环境中是非常可取的，但是在实践中，许多障碍阻碍了这一点，特别是性能和可移植性。虽然存在MATLAB-to-C编译器，但生成的C代码的性能可能不够，因此研究替代方案很重要，例如CPU并行性，GPGPU计算和fpga。OpenCL是一种API和编程语言，允许针对这些设备，因此matlab到OpenCL编译的动机。在本文中，我们描述了我们最近在MATLAB to C/OpenCL编译器的背景下将代码卸载到OpenCL设备上的努力。
{"title":"Compiler Techniques for Efficient MATLAB to OpenCL Code Generation","authors":"Luís Reis, João Bispo, João MP Cardoso","doi":"10.1145/3078155.3078186","DOIUrl":"https://doi.org/10.1145/3078155.3078186","url":null,"abstract":"MATLAB is a high-level language used in various scientific and engineering fields. Deployment of well-tested MATLAB code to production would be highly desirable, but in practice a number of obstacles prevent this, notably performance and portability. Although MATLAB-to-C compilers exist, the performance of the generated C code may not be sufficient and thus it is important to research alternatives, such as CPU parallelism, GPGPU computing and FPGAs. OpenCL is an API and programming language that allows targeting these devices, hence the motivation for MATLAB-to-OpenCL compilation. In this paper, we describe our recent efforts on offloading code to OpenCL devices in the context of our MATLAB to C/OpenCL compiler.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129303035","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}
In the current landscape of C++ applications, there is an increasing need of including different levels of support for heterogeneous platforms, where multiple specialised devices collaborate to execute an application. In this context, the SYCL standard[8] has been published by Khronos, providing a C++ abstraction layer on top of OpenCL[9] that enables single-source programming for a large number of heterogeneous devices. SYCL single-source programming and task data-flow approach enable developers to leverage modern programming techniques on heterogeneous platforms. In this paper, we present SYCL-BLAS, a BLAS implementation using SYCL that uses Expression Tree templates to generate BLAS kernels. This technique is then used to demonstrate seamless kernel fusion via composition of tree nodes. We also demonstrate how SYCL can be used to quickly develop libraries for heterogeneous systems by providing sufficient levels of abstraction.
{"title":"SYCL-BLAS: Leveraging Expression Trees for Linear Algebra","authors":"J. Aliaga, Ruymán Reyes, M. Goli","doi":"10.1145/3078155.3078189","DOIUrl":"https://doi.org/10.1145/3078155.3078189","url":null,"abstract":"In the current landscape of C++ applications, there is an increasing need of including different levels of support for heterogeneous platforms, where multiple specialised devices collaborate to execute an application. In this context, the SYCL standard[8] has been published by Khronos, providing a C++ abstraction layer on top of OpenCL[9] that enables single-source programming for a large number of heterogeneous devices. SYCL single-source programming and task data-flow approach enable developers to leverage modern programming techniques on heterogeneous platforms. In this paper, we present SYCL-BLAS, a BLAS implementation using SYCL that uses Expression Tree templates to generate BLAS kernels. This technique is then used to demonstrate seamless kernel fusion via composition of tree nodes. We also demonstrate how SYCL can be used to quickly develop libraries for heterogeneous systems by providing sufficient levels of abstraction.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"35 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129733810","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}
Anastasios Doumoulakis, R. Keryell, Kenneth O'Brien
Heterogeneous computing is required in systems ranging from low-end embedded systems up to the high-end HPC systems to reach high-performance while keeping power consumption low. Having more and more accelerators and CPUs also creates challenges for the programmer, requiring even more expertise of them. Fortunately, new modern C++-based domain-specific languages, such as the SYCL open standard from Khronos Group, simplify the programming at the full system level while keeping high performance. SYCL is a single-source programming model providing a task graph of heterogeneous kernels that can be run on various accelerators or even just the CPU. The memory heterogeneity is abstracted through buffer objects and the memory usage is abstracted with accessor objects. From these accessors, the task graph is implicitly constructed, the synchronizations and the data movements across the various physical memories are done automatically, by opposition to OpenCL or CUDA. Sometimes, some applications or libraries already exist using the OpenCL standard or some OpenCL kernels are provided, either as OpenCL kernel source code or even as built-in OpenCL kernels written in RTL for extreme optimization on FPGA. SYCL provides an OpenCL interoperability mode to reuse existing OpenCL code while keeping the higher level task graph programming model without needing explicit memory transfers. We present some experiments on two applications on GPU and FPGA with the triSYCL open-source implementation to show the benefits of this OpenCL interoperability mode.
{"title":"SYCL C++ and OpenCL interoperability experimentation with triSYCL","authors":"Anastasios Doumoulakis, R. Keryell, Kenneth O'Brien","doi":"10.1145/3078155.3078188","DOIUrl":"https://doi.org/10.1145/3078155.3078188","url":null,"abstract":"Heterogeneous computing is required in systems ranging from low-end embedded systems up to the high-end HPC systems to reach high-performance while keeping power consumption low. Having more and more accelerators and CPUs also creates challenges for the programmer, requiring even more expertise of them. Fortunately, new modern C++-based domain-specific languages, such as the SYCL open standard from Khronos Group, simplify the programming at the full system level while keeping high performance. SYCL is a single-source programming model providing a task graph of heterogeneous kernels that can be run on various accelerators or even just the CPU. The memory heterogeneity is abstracted through buffer objects and the memory usage is abstracted with accessor objects. From these accessors, the task graph is implicitly constructed, the synchronizations and the data movements across the various physical memories are done automatically, by opposition to OpenCL or CUDA. Sometimes, some applications or libraries already exist using the OpenCL standard or some OpenCL kernels are provided, either as OpenCL kernel source code or even as built-in OpenCL kernels written in RTL for extreme optimization on FPGA. SYCL provides an OpenCL interoperability mode to reuse existing OpenCL code while keeping the higher level task graph programming model without needing explicit memory transfers. We present some experiments on two applications on GPU and FPGA with the triSYCL open-source implementation to show the benefits of this OpenCL interoperability mode.","PeriodicalId":267581,"journal":{"name":"Proceedings of the 5th International Workshop on OpenCL","volume":"65 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2017-05-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125850670","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}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：