GPU Virtualization in Embedded Automotive Solutions
Download as PPT, PDF3 likes2,289 views
The document discusses GPU virtualization in embedded automotive solutions, outlining the motivations for virtualization such as isolation, parallelism, and safety. It reviews various GPU virtualization approaches, including API remoting, driver split, full emulation, and mediated pass-through, analyzing their strengths and weaknesses. Additionally, it presents key virtualization techniques and performance illustrations while referencing various related academic works and industry sources.
GPU Virtualization in Embedded Automotive Solutions
- 1. ©2014 GlobalLogic Inc. GPU Virtualization in Embedded Automotive Solutions Sergiy Kibrik GlobalLogic Embedded TechTalk July 22, 2015
- 2. 2 Why One Wants to Virtualize?Why One Wants to Virtualize? VSVS
- 3. 3 ARM implemented virtualization support on some of ARM v7 processorsARM implemented virtualization support on some of ARM v7 processors
- 4. 4 Key Virtualization TargetsKey Virtualization Targets • isolationisolation • parallelismparallelism • safetysafety • robustnessrobustness • performanceperformance
- 5. 5 Common GPU Virtualization ApproachesCommon GPU Virtualization Approaches OpenGL/DirectX/OpenCL etcOpenGL/DirectX/OpenCL etc API RemotingAPI Remoting match:match: isolation, parallelism, safetyisolation, parallelism, safety no match:no match: robustness, performancerobustness, performance on embedded systemson embedded systems Vendors: Xen (VMGL), Chromium,Vendors: Xen (VMGL), Chromium, Android Emluator,Android Emluator, presumably MentorGraphics & RedBendpresumably MentorGraphics & RedBend
- 6. 6 Common GPU Virtualization ApproachesCommon GPU Virtualization Approaches Driver SplitDriver Split Vendors: VMWare SVGA2Vendors: VMWare SVGA2 match:match: isolation, parallelism, safetyisolation, parallelism, safety no match:no match: robustness, performancerobustness, performance on embedded systemson embedded systems + high cost+ high cost
- 7. 7 Do It The Hard WayDo It The Hard WayTMTM Full GPU Emulation!Full GPU Emulation! Registers & VRAMRegisters & VRAM virtualizedvirtualized Every single oneEvery single one access is trapped &access is trapped & emulatedemulated
- 8. 8 Common GPU Virtualization ApproachesCommon GPU Virtualization Approaches Direct pass-through:Direct pass-through: match:match: safety, performance,safety, performance, robustnessrobustness no match:no match: isolation, parallelismisolation, parallelism Vendors: NVIDIA GRID vGPUVendors: NVIDIA GRID vGPU
- 9. 9 And the Right Approach :)And the Right Approach :) match:match: isolation, parallelism,isolation, parallelism, safety, robustnesssafety, robustness not quite match:not quite match: performanceperformance Mediated pass-throughMediated pass-through
- 10. 10 A bit of GPU Pipeline…A bit of GPU Pipeline…
- 11. 11 ……and what this stuff actually doesand what this stuff actually does
- 12. 12 Real World Example: GeForce 8800Real World Example: GeForce 8800 (2006)(2006)
- 13. 13 Embedded Case: PowerVR Graphics PipelineEmbedded Case: PowerVR Graphics Pipeline • tile-based renderingtile-based rendering • hidden surface removalhidden surface removal
- 14. 14 PowerVR SGX5xx HardwarePowerVR SGX5xx Hardware TI OMAP5770/DRA7/Jacinto6:TI OMAP5770/DRA7/Jacinto6: •2 SGX544 cores2 SGX544 cores •4 USSE processors each4 USSE processors each •4 pipes & 16 threads each4 pipes & 16 threads each •128 parallel threads128 parallel threads •runs OS: uKernelruns OS: uKernel
- 15. 15 Key Virtualization Techniques: MMU Pass-ThroughKey Virtualization Techniques: MMU Pass-Through • trap all MMU accessestrap all MMU accesses • introduce shadow PD/PTintroduce shadow PD/PT • hypercall to Xen for translationhypercall to Xen for translation
- 16. 16 Key Virtualization Techniques: control hijackingKey Virtualization Techniques: control hijacking • un-map IO registers for allun-map IO registers for all domainsdomains • trap each register accesstrap each register access • if domain is active repeat‒if domain is active repeat‒ actionaction • remember action otherwiseremember action otherwise
- 17. 17 Key Virtualization Techniques: Global (firmware) context switchKey Virtualization Techniques: Global (firmware) context switch • we can inject commandwe can inject command • uKernel has power commanduKernel has power command • forward IRQ to active domain onlyforward IRQ to active domain only
- 18. 18 Key Virtualization Techniques: VGPU SchedulingKey Virtualization Techniques: VGPU Scheduling • GPU tasks non-preemptibleGPU tasks non-preemptible • long context switch timelong context switch time • GPU scheduled separatelyGPU scheduled separately • simple FIFO scheduler for nowsimple FIFO scheduler for now
- 19. 19
- 20. 20 Similar Approach: XenGTSimilar Approach: XenGT • our inspirationour inspiration • has some kind of independent channelshas some kind of independent channels • Intel onlyIntel only • heavily abstractedheavily abstracted
- 21. 21 Some Performance IllustrationSome Performance Illustration TestTest RateRate GL Benchmark EgyptGL Benchmark Egypt 22.5 fps22.5 fps NenaMark2NenaMark2 27.3 fps27.3 fps
- 22. 22 Further Improvements:Further Improvements: • virtualization support in uKernel (VE)virtualization support in uKernel (VE) • smart GPU schedulersmart GPU scheduler Materials:Materials: • Kun Tian , Yaozu Dong , David Cowperthwaite, A full GPU virtualization solution with mediated pass-Kun Tian , Yaozu Dong , David Cowperthwaite, A full GPU virtualization solution with mediated pass- through, Proceedings of the 2014 USENIX conference on USENIX Annual Technical Conference, Junethrough, Proceedings of the 2014 USENIX conference on USENIX Annual Technical Conference, June 19-20, 2014, Philadelphia, PA19-20, 2014, Philadelphia, PA • Yusuke Suzuki, Shinpei Kato, Hiroshi Yamada, and Kenji Kono. 2014.Yusuke Suzuki, Shinpei Kato, Hiroshi Yamada, and Kenji Kono. 2014. GPUvm: why not virtualizingGPUvm: why not virtualizing GPUs at the hypervisor?.GPUs at the hypervisor?. InIn Proceedings of the 2014 USENIX conference on USENIX Annual TechnicalProceedings of the 2014 USENIX conference on USENIX Annual Technical ConferenceConference (USENIX ATC'14), Garth Gibson and Nickolai Zeldovich (Eds.).(USENIX ATC'14), Garth Gibson and Nickolai Zeldovich (Eds.). USENIX Association,USENIX Association, Berkeley, CA, USA, 109-120.Berkeley, CA, USA, 109-120. • Micah Dowty and Jeremy Sugerman. 2008. GPU virtualization on VMware's hosted I/O architecture. InMicah Dowty and Jeremy Sugerman. 2008. GPU virtualization on VMware's hosted I/O architecture. In Proceedings of the First conference on I/O virtualizationProceedings of the First conference on I/O virtualization (WIOV'08).(WIOV'08). USENIX Association, Berkeley,USENIX Association, Berkeley, CA, USA, 7-7CA, USA, 7-7 • NVIDIA, “Fermi: The First Complete GPU Computing Architecture”, 2009NVIDIA, “Fermi: The First Complete GPU Computing Architecture”, 2009 • J. Nickolls and D. Kirk, ‘‘Graphics and Computing GPUs,’’ Computer Organization and Design: TheJ. Nickolls and D. Kirk, ‘‘Graphics and Computing GPUs,’’ Computer Organization and Design: The Hardware/Software Interface, D.A. Patterson and J.L. Hennessy, 4th ed., Morgan Kaufmann, 2009,Hardware/Software Interface, D.A. Patterson and J.L. Hennessy, 4th ed., Morgan Kaufmann, 2009, pp.pp. A2-A77.A2-A77. • Imagination, “PowerVR Series5 Architecture Guide for Developers”, 2014Imagination, “PowerVR Series5 Architecture Guide for Developers”, 2014 • Imagination, “PowerVR Hardware Architecture Overview for Developers”, 2015Imagination, “PowerVR Hardware Architecture Overview for Developers”, 2015 • H.AndrésLagar-Cavilla, “VMGL: VMM-Independent Graphics Acceleration”, undatedH.AndrésLagar-Cavilla, “VMGL: VMM-Independent Graphics Acceleration”, undated • OpenSynergy, “Shared GPU using Virtualization”, undatedOpenSynergy, “Shared GPU using Virtualization”, undated
- 23. ©2014 GlobalLogic Inc. Thank You