<!DOCTYPE html> <html> <head> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> </head> <body> Hi, Oak, <div class="moz-cite-prefix">On 1/30/24 01:21, Zeng, Oak wrote: </div> <blockquote type="cite" cite="mid:SA1PR11MB6991DF41C958CA46DD715E9F927D2@SA1PR11MB6991.namprd11.prod.outlook.com"> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style>@font-face {font-family:"Cambria Math"; panose-1:2 4 5 3 5 4 6 3 2 4;}@font-face {font-family:DengXian; panose-1:2 1 6 0 3 1 1 1 1 1;}@font-face {font-family:Calibri; panose-1:2 15 5 2 2 2 4 3 2 4;}@font-face {font-family:Consolas; panose-1:2 11 6 9 2 2 4 3 2 4;}@font-face {font-family:"\@DengXian"; panose-1:2 1 6 0 3 1 1 1 1 1;}p.MsoNormal, li.MsoNormal, div.MsoNormal {margin:0cm; font-size:11.0pt; font-family:"Calibri",sans-serif;}a:link, span.MsoHyperlink {mso-style-priority:99; color:blue; text-decoration:underline;}pre {mso-style-priority:99; mso-style-link:"HTML Preformatted Char"; margin:0cm; margin-bottom:.0001pt; font-size:10.0pt; font-family:"Courier New";}p.MsoListParagraph, li.MsoListParagraph, div.MsoListParagraph {mso-style-priority:34; margin-top:0cm; margin-right:0cm; margin-bottom:0cm; margin-left:36.0pt; font-size:11.0pt; font-family:"Calibri",sans-serif;}span.HTMLPreformattedChar {mso-style-name:"HTML Preformatted Char"; mso-style-priority:99; mso-style-link:"HTML Preformatted"; font-family:Consolas;}span.EmailStyle21 {mso-style-type:personal-reply; font-family:"Calibri",sans-serif; color:windowtext;}.MsoChpDefault {mso-style-type:export-only; font-size:10.0pt; mso-ligatures:none;}div.WordSection1 {page:WordSection1;}ol {margin-bottom:0cm;}ul {margin-bottom:0cm;}</style> <div class="WordSection1"> The example you used to prove that KFD is a design failure, is against *any* design which utilize system allocator and hmm. The way that one proxy process running on host to handle many guest processes, doesn’t fit into the concept of “share address space b/t cpu and gpu”. The shared address space has to be within one process. Your proxy process represent many guest processes. It is a fundamental conflict.<o:p></o:p> <o:p> </o:p> Also your userptr proposal does’t solve this problem either:<o:p></o:p> Imagine you have guest process1 mapping CPU address range A…B to GPU address range C…D<o:p></o:p> And you have guest process 2 mapping CPU address range A…B to GPU address range C…D, since process 1 and 2 are two different process, it is legal for process 2 to do the exact same mapping.<o:p></o:p> Now when gpu shader access address C…D, a gpu page fault happens, what does your proxy process do? Which guest process will this fault be directed to and handled? Except you have extra information/API to tell proxy process and GPU HW, there is no way to figure out.<o:p></o:p> <o:p> </o:p> Compared to the shared virtual address space concept of HMM, the userptr design is nothing new except it allows CPU and GPU to use different address to access the same object. If you replace above C…D with A…B, above description becomes a description of the “problem” of HMM/shared virtual address design. <o:p></o:p> <o:p> </o:p> Both design has the same difficulty with your example of the special virtualization environment setup.<o:p></o:p> <o:p> </o:p> As said, we spent effort scoped the userptr solution some time ago. The problem we found enabling userptr with migration were:<o:p></o:p> <ol style="margin-top:0cm" type="1" start="1"> <li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1">The user interface of userptr is not as convenient as system allocator. With the userptr solution, user need to call userptr_ioctl and vm_bind for *every* single cpu pointer that he want to use in a gpu program. While with system allocator, programmer just use any cpu pointer directly in gpu program without any extra driver ioctls.</li> </ol> </div> </blockquote> No, the augmented userptr (lets call it "hmmptr" to distinguish here) would typically only be bound once when the VM is created. It's just a different way to expose the whole SVM mapping to user-space. It's sparsely populated and is not backed by a bo, and it is per-device so UMD would have to replicate the SVM setup and attribute settings on each device. <blockquote type="cite" cite="mid:SA1PR11MB6991DF41C958CA46DD715E9F927D2@SA1PR11MB6991.namprd11.prod.outlook.com"> <div class="WordSection1"> <ol style="margin-top:0cm" type="1" start="1"> <li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1"><o:p></o:p> </li> <li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1">We don’t see the real benefit of using a different Gpu address C…D than the A..B, except you can prove my above reasoning is wrong. In most use cases, you can make GPU C…D == CPU A…B, why bother then?<o:p></o:p></li> <li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1">Looked into implementation details, since hmm fundamentally assume a shared virtual address space b/t cpu and device, for the userptr solution to leverage hmm, you need perform address space conversion every time you calls into hmm functions. </li> </ol> </div> </blockquote> I think very much focus lands on the A..B -> C..D mapping in the discussion. It's just an added flexibility with little or no implementation cost. Although I must admit I'm not fully clear about the actual use-case. In a para-virtualized environment like virGL or vmware's vmx/renderers I could imagine C..D being the guest virtual addresses including compute kernel pointers, A..B being the host renderer's CPU virtual addresses. (Host creates the VM's, and then this translation is needed. I'm not sure para-virtualized SVM exists ATM, but forcing A==C, B==D in the uAPI would rule out such a beast in the future?) /Thomas <blockquote type="cite" cite="mid:SA1PR11MB6991DF41C958CA46DD715E9F927D2@SA1PR11MB6991.namprd11.prod.outlook.com"> <div class="WordSection1"> <ol style="margin-top:0cm" type="1" start="1"> <li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1"><o:p></o:p> </li> </ol> <o:p> </o:p> In summary, GPU device is just a piece of HW to accelerate your CPU program. If HW allows, it is more convenient to use shared address space b/t cpu and GPU. On old HW (example, no gpu page fault support, or gpu only has a very limited address space), we can disable system allocator/SVM. If you use different address space on modern GPU, why don’t you use different address space on different CPU cores?<o:p></o:p> <o:p> </o:p> Regards,<o:p></o:p> Oak<o:p></o:p> <div style="border:none;border-left:solid blue 1.5pt;padding:0cm 0cm 0cm 4.0pt"> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm"> From: dri-devel <a class="moz-txt-link-rfc2396E" href="mailto:dri-devel-bounces@lists.freedesktop.org"><dri-devel-bounces@lists.freedesktop.org></a> On Behalf Of Christian König Sent: Monday, January 29, 2024 5:20 AM To: Zeng, Oak <a class="moz-txt-link-rfc2396E" href="mailto:oak.zeng@intel.com"><oak.zeng@intel.com></a>; Thomas Hellström <a class="moz-txt-link-rfc2396E" href="mailto:thomas.hellstrom@linux.intel.com"><thomas.hellstrom@linux.intel.com></a>; Daniel Vetter <a class="moz-txt-link-rfc2396E" href="mailto:daniel@ffwll.ch"><daniel@ffwll.ch></a>; Dave Airlie <a class="moz-txt-link-rfc2396E" href="mailto:airlied@redhat.com"><airlied@redhat.com></a> Cc: Brost, Matthew <a class="moz-txt-link-rfc2396E" href="mailto:matthew.brost@intel.com"><matthew.brost@intel.com></a>; Felix Kuehling <a class="moz-txt-link-rfc2396E" href="mailto:felix.kuehling@amd.com"><felix.kuehling@amd.com></a>; Welty, Brian <a class="moz-txt-link-rfc2396E" href="mailto:brian.welty@intel.com"><brian.welty@intel.com></a>; <a class="moz-txt-link-abbreviated" href="mailto:dri-devel@lists.freedesktop.org">dri-devel@lists.freedesktop.org</a>; Ghimiray, Himal Prasad <a class="moz-txt-link-rfc2396E" href="mailto:himal.prasad.ghimiray@intel.com"><himal.prasad.ghimiray@intel.com></a>; Bommu, Krishnaiah <a class="moz-txt-link-rfc2396E" href="mailto:krishnaiah.bommu@intel.com"><krishnaiah.bommu@intel.com></a>; Gupta, saurabhg <a class="moz-txt-link-rfc2396E" href="mailto:saurabhg.gupta@intel.com"><saurabhg.gupta@intel.com></a>; Vishwanathapura, Niranjana <a class="moz-txt-link-rfc2396E" href="mailto:niranjana.vishwanathapura@intel.com"><niranjana.vishwanathapura@intel.com></a>; <a class="moz-txt-link-abbreviated" href="mailto:intel-xe@lists.freedesktop.org">intel-xe@lists.freedesktop.org</a>; Danilo Krummrich <a class="moz-txt-link-rfc2396E" href="mailto:dakr@redhat.com"><dakr@redhat.com></a> Subject: Re: Making drm_gpuvm work across gpu devices<o:p></o:p> </div> </div> <o:p> </o:p> Well Daniel and Dave noted it as well, so I'm just repeating it: Your design choices are not an argument to get something upstream. It's the job of the maintainers and at the end of the Linus to judge of something is acceptable or not. As far as I can see a good part of this this idea has been exercised lengthy with KFD and it turned out to not be the best approach. So from what I've seen the design you outlined is extremely unlikely to go upstream. Regards, Christian.<o:p></o:p> <div> Am 27.01.24 um 03:21 schrieb Zeng, Oak:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> Regarding the idea of expanding userptr to support migration, we explored this idea long time ago. It provides similar functions of the system allocator but its interface is not as convenient as system allocator. Besides the shared virtual address space, another benefit of a system allocator is, you can offload cpu program to gpu easier, you don’t need to call driver specific API (such as register_userptr and vm_bind in this case) for memory allocation.<o:p></o:p> <o:p></o:p> We also scoped the implementation. It turned out to be big, and not as beautiful as hmm. Why we gave up this approach. <o:p></o:p> <o:p></o:p> <div style="border:none;border-left:solid blue 1.5pt;padding:0cm 0cm 0cm 4.0pt"> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm"> From: Christian König <a href="mailto:christian.koenig@amd.com" moz-do-not-send="true"><christian.koenig@amd.com></a> Sent: Friday, January 26, 2024 7:52 AM To: Thomas Hellström <a href="mailto:thomas.hellstrom@linux.intel.com" moz-do-not-send="true"><thomas.hellstrom@linux.intel.com></a>; Daniel Vetter <a href="mailto:daniel@ffwll.ch" moz-do-not-send="true"><daniel@ffwll.ch></a> Cc: Brost, Matthew <a href="mailto:matthew.brost@intel.com" moz-do-not-send="true"><matthew.brost@intel.com></a>; Felix Kuehling <a href="mailto:felix.kuehling@amd.com" moz-do-not-send="true"><felix.kuehling@amd.com></a>; Welty, Brian <a href="mailto:brian.welty@intel.com" moz-do-not-send="true"><brian.welty@intel.com></a>; Ghimiray, Himal Prasad <a href="mailto:himal.prasad.ghimiray@intel.com" moz-do-not-send="true"><himal.prasad.ghimiray@intel.com></a>; Zeng, Oak <a href="mailto:oak.zeng@intel.com" moz-do-not-send="true"><oak.zeng@intel.com></a>; Gupta, saurabhg <a href="mailto:saurabhg.gupta@intel.com" moz-do-not-send="true"><saurabhg.gupta@intel.com></a>; Danilo Krummrich <a href="mailto:dakr@redhat.com" moz-do-not-send="true"><dakr@redhat.com></a>; <a href="mailto:dri-devel@lists.freedesktop.org" moz-do-not-send="true">dri-devel@lists.freedesktop.org</a>; Bommu, Krishnaiah <a href="mailto:krishnaiah.bommu@intel.com" moz-do-not-send="true"><krishnaiah.bommu@intel.com></a>; Dave Airlie <a href="mailto:airlied@redhat.com" moz-do-not-send="true"><airlied@redhat.com></a>; Vishwanathapura, Niranjana <a href="mailto:niranjana.vishwanathapura@intel.com" moz-do-not-send="true"><niranjana.vishwanathapura@intel.com></a>; <a href="mailto:intel-xe@lists.freedesktop.org" moz-do-not-send="true">intel-xe@lists.freedesktop.org</a> Subject: Re: Making drm_gpuvm work across gpu devices<o:p></o:p> </div> </div> <o:p></o:p> Am 26.01.24 um 09:21 schrieb Thomas Hellström: <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Hi, all<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>On Thu, 2024-01-25 at 19:32 +0100, Daniel Vetter wrote:<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>On Wed, Jan 24, 2024 at 09:33:12AM +0100, Christian König wrote:<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Am 23.01.24 um 20:37 schrieb Zeng, Oak:<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>[SNIP]<o:p></o:p></pre> <pre>Yes most API are per device based.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>One exception I know is actually the kfd SVM API. If you look at<o:p></o:p></pre> <pre>the svm_ioctl function, it is per-process based. Each kfd_process<o:p></o:p></pre> <pre>represent a process across N gpu devices.<o:p></o:p></pre> </blockquote> <pre> <o:p></o:p></pre> <pre>Yeah and that was a big mistake in my opinion. We should really not<o:p></o:p></pre> <pre>do that<o:p></o:p></pre> <pre>ever again.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Need to say, kfd SVM represent a shared virtual address space<o:p></o:p></pre> <pre>across CPU and all GPU devices on the system. This is by the<o:p></o:p></pre> <pre>definition of SVM (shared virtual memory). This is very different<o:p></o:p></pre> <pre>from our legacy gpu *device* driver which works for only one<o:p></o:p></pre> <pre>device (i.e., if you want one device to access another device's<o:p></o:p></pre> <pre>memory, you will have to use dma-buf export/import etc).<o:p></o:p></pre> </blockquote> <pre> <o:p></o:p></pre> <pre>Exactly that thinking is what we have currently found as blocker<o:p></o:p></pre> <pre>for a<o:p></o:p></pre> <pre>virtualization projects. Having SVM as device independent feature<o:p></o:p></pre> <pre>which<o:p></o:p></pre> <pre>somehow ties to the process address space turned out to be an<o:p></o:p></pre> <pre>extremely bad<o:p></o:p></pre> <pre>idea.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>The background is that this only works for some use cases but not<o:p></o:p></pre> <pre>all of<o:p></o:p></pre> <pre>them.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>What's working much better is to just have a mirror functionality<o:p></o:p></pre> <pre>which says<o:p></o:p></pre> <pre>that a range A..B of the process address space is mapped into a<o:p></o:p></pre> <pre>range C..D<o:p></o:p></pre> <pre>of the GPU address space.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>Those ranges can then be used to implement the SVM feature required<o:p></o:p></pre> <pre>for<o:p></o:p></pre> <pre>higher level APIs and not something you need at the UAPI or even<o:p></o:p></pre> <pre>inside the<o:p></o:p></pre> <pre>low level kernel memory management.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>When you talk about migrating memory to a device you also do this<o:p></o:p></pre> <pre>on a per<o:p></o:p></pre> <pre>device basis and *not* tied to the process address space. If you<o:p></o:p></pre> <pre>then get<o:p></o:p></pre> <pre>crappy performance because userspace gave contradicting information<o:p></o:p></pre> <pre>where to<o:p></o:p></pre> <pre>migrate memory then that's a bug in userspace and not something the<o:p></o:p></pre> <pre>kernel<o:p></o:p></pre> <pre>should try to prevent somehow.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>[SNIP]<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>I think if you start using the same drm_gpuvm for multiple<o:p></o:p></pre> <pre>devices you<o:p></o:p></pre> <pre>will sooner or later start to run into the same mess we have<o:p></o:p></pre> <pre>seen with<o:p></o:p></pre> <pre>KFD, where we moved more and more functionality from the KFD to<o:p></o:p></pre> <pre>the DRM<o:p></o:p></pre> <pre>render node because we found that a lot of the stuff simply<o:p></o:p></pre> <pre>doesn't work<o:p></o:p></pre> <pre>correctly with a single object to maintain the state.<o:p></o:p></pre> </blockquote> <pre>As I understand it, KFD is designed to work across devices. A<o:p></o:p></pre> <pre>single pseudo /dev/kfd device represent all hardware gpu devices.<o:p></o:p></pre> <pre>That is why during kfd open, many pdd (process device data) is<o:p></o:p></pre> <pre>created, each for one hardware device for this process.<o:p></o:p></pre> </blockquote> <pre> <o:p></o:p></pre> <pre>Yes, I'm perfectly aware of that. And I can only repeat myself that<o:p></o:p></pre> <pre>I see<o:p></o:p></pre> <pre>this design as a rather extreme failure. And I think it's one of<o:p></o:p></pre> <pre>the reasons<o:p></o:p></pre> <pre>why NVidia is so dominant with Cuda.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>This whole approach KFD takes was designed with the idea of<o:p></o:p></pre> <pre>extending the<o:p></o:p></pre> <pre>CPU process into the GPUs, but this idea only works for a few use<o:p></o:p></pre> <pre>cases and<o:p></o:p></pre> <pre>is not something we should apply to drivers in general.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>A very good example are virtualization use cases where you end up<o:p></o:p></pre> <pre>with CPU<o:p></o:p></pre> <pre>address != GPU address because the VAs are actually coming from the<o:p></o:p></pre> <pre>guest VM<o:p></o:p></pre> <pre>and not the host process.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>SVM is a high level concept of OpenCL, Cuda, ROCm etc.. This should<o:p></o:p></pre> <pre>not have<o:p></o:p></pre> <pre>any influence on the design of the kernel UAPI.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>If you want to do something similar as KFD for Xe I think you need<o:p></o:p></pre> <pre>to get<o:p></o:p></pre> <pre>explicit permission to do this from Dave and Daniel and maybe even<o:p></o:p></pre> <pre>Linus.<o:p></o:p></pre> </blockquote> <pre> <o:p></o:p></pre> <pre>I think the one and only one exception where an SVM uapi like in kfd<o:p></o:p></pre> <pre>makes<o:p></o:p></pre> <pre>sense, is if the _hardware_ itself, not the software stack defined<o:p></o:p></pre> <pre>semantics that you've happened to build on top of that hw, enforces a<o:p></o:p></pre> <pre>1:1<o:p></o:p></pre> <pre>mapping with the cpu process address space.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>Which means your hardware is using PASID, IOMMU based translation,<o:p></o:p></pre> <pre>PCI-ATS<o:p></o:p></pre> <pre>(address translation services) or whatever your hw calls it and has<o:p></o:p></pre> <pre>_no_<o:p></o:p></pre> <pre>device-side pagetables on top. Which from what I've seen all devices<o:p></o:p></pre> <pre>with<o:p></o:p></pre> <pre>device-memory have, simply because they need some place to store<o:p></o:p></pre> <pre>whether<o:p></o:p></pre> <pre>that memory is currently in device memory or should be translated<o:p></o:p></pre> <pre>using<o:p></o:p></pre> <pre>PASID. Currently there's no gpu that works with PASID only, but there<o:p></o:p></pre> <pre>are<o:p></o:p></pre> <pre>some on-cpu-die accelerator things that do work like that.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>Maybe in the future there will be some accelerators that are fully<o:p></o:p></pre> <pre>cpu<o:p></o:p></pre> <pre>cache coherent (including atomics) with something like CXL, and the<o:p></o:p></pre> <pre>on-device memory is managed as normal system memory with struct page<o:p></o:p></pre> <pre>as<o:p></o:p></pre> <pre>ZONE_DEVICE and accelerator va -> physical address translation is<o:p></o:p></pre> <pre>only<o:p></o:p></pre> <pre>done with PASID ... but for now I haven't seen that, definitely not<o:p></o:p></pre> <pre>in<o:p></o:p></pre> <pre>upstream drivers.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>And the moment you have some per-device pagetables or per-device<o:p></o:p></pre> <pre>memory<o:p></o:p></pre> <pre>management of some sort (like using gpuva mgr) then I'm 100% agreeing<o:p></o:p></pre> <pre>with<o:p></o:p></pre> <pre>Christian that the kfd SVM model is too strict and not a great idea.<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>Cheers, Sima<o:p></o:p></pre> </blockquote> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>I'm trying to digest all the comments here, The end goal is to be able<o:p></o:p></pre> <pre>to support something similar to this here:<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre><a href="https://developer.nvidia.com/blog/simplifying-gpu-application-development-with-heterogeneous-memory-management/" moz-do-not-send="true" class="moz-txt-link-freetext">https://developer.nvidia.com/blog/simplifying-gpu-application-development-with-heterogeneous-memory-management/</a><o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>Christian, If I understand you correctly, you're strongly suggesting<o:p></o:p></pre> <pre>not to try to manage a common virtual address space across different<o:p></o:p></pre> <pre>devices in the kernel, but merely providing building blocks to do so,<o:p></o:p></pre> <pre>like for example a generalized userptr with migration support using<o:p></o:p></pre> <pre>HMM; That way each "mirror" of the CPU mm would be per device and<o:p></o:p></pre> <pre>inserted into the gpu_vm just like any other gpu_vma, and user-space<o:p></o:p></pre> <pre>would dictate the A..B -> C..D mapping by choosing the GPU_VA for the<o:p></o:p></pre> <pre>vma.<o:p></o:p></pre> </blockquote> Exactly that, yes. <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre> <o:p></o:p></pre> <pre>Sima, it sounds like you're suggesting to shy away from hmm and not<o:p></o:p></pre> <pre>even attempt to support this except if it can be done using IOMMU sva<o:p></o:p></pre> <pre>on selected hardware?<o:p></o:p></pre> </blockquote> I think that comment goes more into the direction of: If you have ATS/ATC/PRI capable hardware which exposes the functionality to make memory reads and writes directly into the address space of the CPU then yes an SVM only interface is ok because the hardware can't do anything else. But as long as you have something like GPUVM then please don't restrict yourself. Which I totally agree on as well. The ATS/ATC/PRI combination doesn't allow using separate page tables device and CPU and so also not separate VAs. This was one of the reasons why we stopped using this approach for AMD GPUs. Regards, Christian. <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Could you clarify a bit?<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre>Thanks,<o:p></o:p></pre> <pre>Thomas<o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> <pre> <o:p></o:p></pre> </blockquote> <o:p></o:p> </div> </blockquote> <o:p> </o:p> </div> </div> </blockquote> </body> </html>