Announcement

Collapse
No announcement yet.

Radeon Software 20.10 vs. Upstream Linux AMD Radeon OpenGL / Vulkan Performance

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #31
    bridgman
    Any chance that we'll get a 20.10 driver for Ubuntu 20.04 any time soon or will that be more of an, I dunno, we'll call it a Radeon 20.20 and Ubuntu 20.04.01 thing?

    ....put the tools on the flash. And I thought I said some dumb stuff....

    Comment


    • #32
      20.10 was supposed to come out earlier in the year; I don't think we would go back and spin a new version of it with the same number unless the difference was just critical bug fixes (and no, adding new features or new kernel/distro support that were not planned for the release does not count as a bug fix).

      Next release should be within a month or so AFAIK.
      Last edited by bridgman; 27 April 2020, 01:10 PM.
      Test signature

      Comment


      • #33
        Sorry for maybe off-topic post, but I really do not know what to try next... I have RX590, I5-2500K, 8GB RAM, Ubuntu 18.04 with 5.6 kernel, and in two Steam games, Black Mesa (native Linux) and Sherlock Holmes Crimes and Punishments (Proton), when I get to a certain point (always the same point in the game), the computer just shuts down. Not restart, but shut down. I have checked the temperatures of the CPU and GPU (by writing into file every 2 seconds), they are way below threshold (~40C for CPU and ~60C for GPU), and syslog, kernel.log and Xorg.0.log show absolutely nothing abnormal when this happens (in all of them the reporting just stops when this happens, there is no any trace of what could be happening). I'll try over the weekend to install Ubuntu 20.04 to see if the problem will persist or not, but in the meantime, does anyone have any advice as what to check?

        Comment


        • #34
          Originally posted by bridgman View Post
          This is really only useful for users of libre distros. For everyone else it would be a step backwards, making them dependent on closed-source VBIOS updates to get latest GPU firmware in the same way that libre distro users are are dependent on closed-source SBIOS updates to get latest CPU firmware.
          We could store a launch-day copy of the firmware in the flash and have libre distro users run with that while everyone else continues to get firmware via distro updates, but then we are back to the challenge of convincing one or more board vendors to take on additional cost and support burden in order to support libre distro users, when it's still an uphill battle getting them to think about Linux at all.
          it is not only usefull for libre linux distro users. i bought new graphic cards many time and always ther ewas a DVD with the windows driver.
          so in fact windows users would benefit to because the launch time drivers for windows could be on the flash chip to.
          today such a flash ship is very cheap. we talk for 1-2€ here.
          i think there could be many more usercases than this.
          you can store all kind of usefull tools on the flash and advertise it to windows users as well.

          Originally posted by bridgman View Post
          Not sure what you mean by "put the tools on the flash" - are you talking about building a full GUI into the VBIOS POST sequence ?
          i am sure you can build many usefull stuff with a flash chip on the graphic card. not just a firmware for libre people.
          just ask your customers or developers about ideas what you can do if a graphic card has a little bit flash.
          you could even store pre-compiled shader code to speed up the startup of games.


          Originally posted by bridgman View Post
          I know the "dropping" part sounds like fun but first there's a lot of work to do - getting the open source drivers to the point where they can not only run all the workstation apps competitively but also be delivered in a certified form the way that binary drivers are.
          well yes you are right. but maybe it is worth it ? i am sure this would result in good PR for amd.

          and the argument that then the people buy cheap gaming cards for workstation and server instead of expensive parts...
          i do not think this is true because if a professional need ECC Vram there is no other option than to buy the PRO cards.

          Originally posted by bridgman View Post
          Ooh, yes, we could give up on the datacenter compute market completely and save a lot of R&D expense. Seriously, Vulkan might one day grow into a general purpose compute solution but even Khronos isn't talking about that any more.
          o well we see it with OpenCL3.0 OpenCL is already death because it is made in a way that nvidia only support 1.2 features but still claim they are OpenCL3.0 compatible because you do not need to support any new features because anything else than openCL1.2 is just optional.

          believe it or not but openCL is death. this means you do Vulkan Compute or you lose all the datacenter compute customers.

          Originally posted by bridgman View Post
          Guessing you mean Lenovo (IBM sold their PC & laptop business to Lenovo) rather than IBM ? Lenovo are building laptops with Fedora and RHEL today.
          yes sure but no i do not talk about cheap made in china Lenevo laptops.

          i really think IBM/redhat/amd should merge into one single company and they should sell super highend super quality laptops. made in usa or made in europe instead of cheap china.



          Phantom circuit Sequence Reducer Dyslexia

          Comment


          • #35
            Originally posted by Qaridarium View Post
            it is not only usefull for libre linux distro users. i bought new graphic cards many time and always ther ewas a DVD with the windows driver.
            so in fact windows users would benefit to because the launch time drivers for windows could be on the flash chip to.
            today such a flash ship is very cheap. we talk for 1-2€ here.

            i think there could be many more usercases than this. you can store all kind of usefull tools on the flash and advertise it to windows users as well.

            i am sure you can build many usefull stuff with a flash chip on the graphic card. not just a firmware for libre people.
            just ask your customers or developers about ideas what you can do if a graphic card has a little bit flash.
            you could even store pre-compiled shader code to speed up the startup of games.
            Not sure where you are looking for pricing, but I don't think a 1-2 euro flash chip is big enough for something like a driver install package and windows tools yet. It is an interesting idea though... at some point the cost of burning and packaging/handling a DVD-ROM might cross-over with the cost of a flash chip on board. I am seeing very inexpensive 16GB USB keys, so the time may not be that far off.

            Originally posted by Qaridarium View Post
            well yes you are right. but maybe it is worth it ? i am sure this would result in good PR for amd.
            Sure, and we've been working on that for a while. The point I'm trying to make is that we can't just drop the current workstation GL driver and then ask customers to wait a couple of years until the open source GL driver can provide a replacement - we have to do the work up front including testing & certification before we can even think about dropping the current driver.

            Originally posted by Qaridarium View Post
            o well we see it with OpenCL3.0 OpenCL is already death because it is made in a way that nvidia only support 1.2 features but still claim they are OpenCL3.0 compatible because you do not need to support any new features because anything else than openCL1.2 is just optional.

            believe it or not but openCL is death. this means you do Vulkan Compute or you lose all the datacenter compute customers.
            Just to be clear, we are not thinking that OpenCL is the future of datacenter computing. Our investment is going into HIP and higher level libraries/frameworks. Vulkan is even less well suited than OpenCL, except for consumer applications where Vulkan's pervasiveness (is that a word ?) is more important than its limited computing capabilities.

            Originally posted by Qaridarium View Post
            yes sure but no i do not talk about cheap made in china Lenevo laptops. i really think IBM/redhat/amd should merge into one single company and they should sell super highend super quality laptops. made in usa or made in europe instead of cheap china.
            You know the old joke about how to make a billion dollar company ? Start with a 10 billion dollar company, then mess with the business plan and wait a few years.

            We are starting to see AMD processors being used in high end laptops, which is a nice change - Picasso opened the door and hopefully got the market ready for Renoir.
            Last edited by bridgman; 28 April 2020, 02:46 AM.
            Test signature

            Comment


            • #36
              Originally posted by bridgman View Post
              Not sure where you are looking for pricing, but I don't think a 1-2 euro flash chip is big enough for something like a driver install package and windows tools yet. It is an interesting idea though... at some point the cost of burning and packaging/handling a DVD-ROM might cross-over with the cost of a flash chip on board. I am seeing very inexpensive 16GB USB keys, so the time may not be that far off.
              most modern computer even my own threadripper system do not have any CD/DVD/Bluray drive. this means it is not only about the "Cost" of a flash chip in comparison of a DVD disc... no modern computer do not have a optical drive. for this fact alone AMD should put a 8GB or 16gb flash chip on the GPU card.
              this is not all new idea people here in the phoronix forum write to you to tell amd to add flash chip for firmware and more.
              but 1 thing changed: in the past flash chips per GB where expensive today they are very cheap.
              today on geizhals.at a 8GB USB stick is 2,90€ https://geizhals.de/transcend-jetfla...loc=at&hloc=de
              and 16GB usb stick is at 3,95€ https://geizhals.de/intenso-rainbow-...loc=at&hloc=de
              so lets face it the argument of the costs are not anymore valid because as you said "at some point the cost of burning and packaging/handling a DVD-ROM might cross-over with the cost of a flash chip on board. I am seeing very inexpensive 16GB USB keys, so the time may not be that far off."
              but the costs are not the main argument the main argument should be: modern computer do not have Optical drive anymore and Libre linux distro need firmware on the flash chip.
              also i think you can do amazing stuff if your gpu cards do have a 16gb flash chip from shader cache to other stuff.

              Originally posted by bridgman View Post
              Sure, and we've been working on that for a while. The point I'm trying to make is that we can't just drop the current workstation GL driver and then ask customers to wait a couple of years until the open source GL driver can provide a replacement - we have to do the work up front including testing & certification before we can even think about dropping the current driver.
              yes sure. but talk about future roadmap is important if your customers know that this closed source driver is legacy only and the floss driver is the future.

              Originally posted by bridgman View Post
              Just to be clear, we are not thinking that OpenCL is the future of datacenter computing. Our investment is going into HIP and higher level libraries/frameworks. Vulkan is even less well suited than OpenCL, except for consumer applications where Vulkan's pervasiveness (is that a word ?) is more important than its limited computing capabilities.
              thats really strange to her... other people report much better compute performance with vulkan and the ACO compiler...

              benchmarks show vulkan is much better for compute..: https://translate.google.com/transla...edup-2x-aco%2F

              https://blog.coelacanth-dream.com/po...peedup-2x-aco/

              Originally posted by bridgman View Post
              You know the old joke about how to make a billion dollar company ? Start with a 10 billion dollar company, then mess with the business plan and wait a few years.
              We are starting to see AMD processors being used in high end laptops, which is a nice change - Picasso opened the door and hopefully got the market ready for Renoir.
              i know intel customers they tell me they admit amd chips means cpu and gpu are better but they point out that there is no highend Laptop to buy
              not a single one with for example 4900H(56watt)+5700M ... not a single one with 4900H+nvidia2080
              the best amd laptop you can get today is a 3950X 16core +nvidia 2070

              intel and nvidia customers admit that amd gpu is better in FPS per watt and better in FPS per dollar and the cpus in low-end sector have higher singlecore performance and higher multicore performance and in multicore higher per core performance for example intel 65watt 10core loses against 4900h 8core means AMD have the higher per core multicore performance.

              all this means: amd could sell more if they would have highend laotop brand to sell for example 4900H(56watt)+5700M with a IBM brand or something like this.

              about lenevo... yes lenevo sells 4800/4900 but yet no 5700m and no 4900H+nvidia2080
              Phantom circuit Sequence Reducer Dyslexia

              Comment


              • #37
                Originally posted by smitty3268 View Post

                The defining feature of that Heaven benchmark is that it's heavily limited by tessellation performance, much more than any actual games that have been released or any other demos.

                I do suspect AMD pretty heavily optimized their driver for it, though, and they likely do have an app profile setup.

                It's also possible that the open source drivers haven't bothered optimizing tessellation performance quite as much, since most games don't tend to use it as much.
                Speed of tessellation on AMD (OSS) was one of the shortcomings since Marek implemented it.
                Compare Unigine Heaven/Valley with enabled/disabled (on Polaris 20, here).

                Comment


                • #38
                  Originally posted by bridgman View Post

                  Each GPU has a separate instance of amdgpu, while all GPUs are managed by a single instance of amdkfd. It wouldn't be a stretch to think of amdkfd as the cross-GPU component of the GPU stack, a role it partially shares with drm.

                  All hardware accesses are made by amdgpu, with the exception of a couple of hardware features added specifically for HSA/ROCm that are owned by amdkfd, primarily the HIQ command queue and all the associated functionality.

                  The ROCm stack looks like:

                  Frameworks -> Libraries -> HIP/HCC compilers/runtimes -> ROC runtime (ROCR) -> libhsakmt (ROCT) -> amdkfd (then amdkfd calls into amdgpu when needed)

                  The graphics stack uses different userspace components as agd5f said:

                  Userspace drivers (Mesa or proprietary, mostly using winsys/PAL) -> libdrm -> drm -> amdgpu

                  If you see references to "ROCK" (Radeon Open Compute Kernel drivers) that's just the combination of amdgpu and amdkfd, along with (IIRC) a slightly tweaked version of ttm.
                  .
                  Thanks for explanation. Currently my radeon machine is disassembled, but I wonder if it's normal that there's no mention of amdkfd in /proc/config.gz, dmesg(after loading amdgpu) ad there's no amdkfd.ko in /lib/modules? Is it merged into amdgpu.ko or maybe it's out-of-tree module and I need some dkms package?

                  Comment


                  • #39
                    I think we build amdkfd into amdgpu.ko these days. Doing that allowed amdkfd to access amdgpu data structures and keep the interface simple.

                    I would still expect to see a couple of amdkfd messages in dmesg output though, can you post it somewhere ? It's possible that your distro disabled KFD by default.
                    Test signature

                    Comment


                    • #40
                      Yes, kfd is part of amdgpu now. There is a separate kconfig item for KFD however (CONFIG_HSA_AMD).

                      Comment

                      Working...
                      X