Announcement

Collapse
No announcement yet.

Mesa's Rusticl Achieves Official OpenCL 3.0 Conformance

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

  • #11
    Congrats👍
    I just watched your interesting xdc presentation. You mentioned that performance (in darktable?) is not better than cpu, due to overhead. Did I get it wrong? Do actually use GL compute for CL?
    Also i woul be interested in memtestcl since i have one semi-faulty rx480.

    Comment


    • #12
      My main question is: can it run DaVinci Resolve for Linux?

      I'm not sure if DaVinci changed it, but when I tried to run it last (might have been version 16) there was a hard coded check for the Clover driver and would bail if present. I would be happy to see it working with Rusticl.

      Comment


      • #13
        Originally posted by tsuru View Post
        My main question is: can it run DaVinci Resolve for Linux?

        I'm not sure if DaVinci changed it, but when I tried to run it last (might have been version 16) there was a hard coded check for the Clover driver and would bail if present. I would be happy to see it working with Rusticl.
        I tried it briefly before rusticl got merged, and it at least gets into resolve, I think basic editing worked, but anything revolving around the display was busted when using llvmpipe, could be worth trying again I suppose

        Comment


        • #14
          Originally posted by Jannik2099 View Post
          Certainly neat, just a decade too late. OpenCL is a dead API by now, as Nvidia blocked OpenCL 2.0 (aka the actually useful OpenCL) because it would eat into CUDA.
          Funny, I use it daily across all my hosts. 98% of all BOINC projects that offer gpu applications use it. OpenCL 2.0 being a requirement on some applications. No impediment to using Nvidia cards either. I've been using Nvidia cards exclusively for over 20 years for distributed computing. Never any issues. The stability of the Nvidia drivers over the AMD drivers in Linux is also a major benefit.

          Comment


          • #15
            Originally posted by gsedej View Post
            Congrats👍
            I just watched your interesting xdc presentation. You mentioned that performance (in darktable?) is not better than cpu, due to overhead. Did I get it wrong? Do actually use GL compute for CL?
            Also i woul be interested in memtestcl since i have one semi-faulty rx480.
            Im not sure what he said, since I completely forget now, and im not even that sure as to what you mean, but typically the difference between APIs are a lot smaller then most people think while OCL and OGL and even VK are completely different beasts, in the end, they are just apis oriented around controlling gpus.

            Zink is/(was?) a "Graphics" driver but even then there has been work to get rusticl working on it (meaning it would be a compute driver too). there is a degree of crossover, and in some cases, a lot of crossover

            Comment


            • #16
              Originally posted by tsuru View Post
              My main question is: can it run DaVinci Resolve for Linux?

              I'm not sure if DaVinci changed it, but when I tried to run it last (might have been version 16) there was a hard coded check for the Clover driver and would bail if present. I would be happy to see it working with Rusticl.
              it needs two extensions which aren't implemented yet: cl_khr_gl_sharing and cl_khr_image2d_from_buffer

              Comment


              • #17
                Originally posted by karolherbst View Post

                it needs two extensions which aren't implemented yet: cl_khr_gl_sharing and cl_khr_image2d_from_buffer
                How soon do you think to an implementation?

                Comment


                • #18
                  Originally posted by Keith Myers View Post

                  Funny, I use it daily across all my hosts. 98% of all BOINC projects that offer gpu applications use it. OpenCL 2.0 being a requirement on some applications. No impediment to using Nvidia cards either. I've been using Nvidia cards exclusively for over 20 years for distributed computing. Never any issues. The stability of the Nvidia drivers over the AMD drivers in Linux is also a major benefit.
                  TL;DR BOINC is awesome. Agreed, OpenCL is not dead. aCtUAlLy Nvidia does not support OpenCL 2x. Proprietary/ROCm RaNtiNg... Finally Rustcl saves the day.

                  BOINC projects are awesome! I ran (MilkyWay, Einstien, and Rosetta) for many years but sadly have had to stop for various reasons.

                  I would normally not recommend using one product or vendor exclusively (e.g. Apple/Mac). Even when you look at tiny basic devices like a Raspberry Pi you will find that things stop working when you are not using Raspberry Pi OS. I have the same problem with most proprietary software which is why I try my best to isolate from it. Similarly Nvidia's proprietary drivers on Ubuntu works without any issues but on Arch with Wayland it's (or was) a mess. Another related reason why I stay away from proprietary is being subject to a vendor's control. Take Nvidia's virtual market segmentation change with the LHR controversy and the corruption surrounding it. Limitations around Nvidia GRID Without proprietary drivers in Linux Nvidia is still an unreliable piece of junk in terms of desktop-usage stability (at least with modern cards). Since compute is not a priority for me I tend to stick with Intel or AMD until Nvidia's open source desktop support improves. I am honestly happy that proprietary compute is working out for you. BOINC can use all the help it can get. Which distros are you using and are you using virtualization at all?

                  A correction about OpenCL 2.0 - Neither Nvidia nor Intel have implemented CL_DEVICE_SVM_FINE_GRAIN_SYSTEM (a feature of OpenCL 2.0) I doubt either of them ever will implement that exact feature. AMD has provided proprietary support for it since 2014 and later open source support followed https://www.phoronix.com/news/MTc5NjU maybe the applications you are talking about used some basic OpenCL 2.0 features or did not report their requirements correctly? If you can provide me with the name of the application I can look into it for you.

                  AMD is breaking records in the HPC/supercomputing market. Frontier is an absolute beast being the first to break the 1 exaflop barrier, placing 1st in the top500 and 6th in green500 supercomputer lists (as of Nov 2022). It's almost the opposite in the consumer side. ROCm is a mess and only unofficially supports a handful of consumer cards. I completely understand why you have stuck with Nvidia's proprietary stack for at least the past decade. Fortunately for us consumers I'm looking forward to that changing and it's why I'm so happy about Rustcl! I agree with you OpenCL is not dead at all.

                  Comment


                  • #19
                    Congrats.

                    Your name still sounds like what the aging process does to balls of steel.

                    Comment


                    • #20
                      Originally posted by Jabberwocky View Post

                      TL;DR BOINC is awesome. Agreed, OpenCL is not dead. aCtUAlLy Nvidia does not support OpenCL 2x. Proprietary/ROCm RaNtiNg... Finally Rustcl saves the day.

                      A correction about OpenCL 2.0 - Neither Nvidia nor Intel have implemented CL_DEVICE_SVM_FINE_GRAIN_SYSTEM (a feature of OpenCL 2.0) I doubt either of them ever will implement that exact feature. AMD has provided proprietary support for it since 2014 and later open source support followed https://www.phoronix.com/news/MTc5NjU maybe the applications you are talking about used some basic OpenCL 2.0 features or did not report their requirements correctly? If you can provide me with the name of the application I can look into it for you.
                      I completely understand why you have stuck with Nvidia's proprietary stack for at least the past decade. Fortunately for us consumers I'm looking forward to that changing and it's why I'm so happy about Rustcl! I agree with you OpenCL is not dead at all.
                      The application is the Einstein FGRPB1G 1.28 (FGRPopencl2Pup-nvidia)

                      Application uses some OpenCL 2.0 calls which are not available in the prior OpenCL 1.2 conformant drivers.
                      AMD users also had to update drivers to full OpenCL 2.0 conformant drivers as well as with the card hardware.
                      Requires minimum 465+ drivers to pick up OpenCL 3.0 capabilities which covers OpenCL 2.0 requirements for the application.
                      Our Team developer helped develop the application who then passed it on to the Einstein admins for it to be a stock project application.
                      https://developer.nvidia.com/blog/nv...-0-conformant/
                      Prior drivers only covered OpenCL 1.2 applications.
                      Last edited by Keith Myers; 15 November 2022, 01:48 PM.

                      Comment

                      Working...
                      X