Announcement

Collapse
No announcement yet.

AMD Releases OpenCL 2.0 Catalyst Linux Driver

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

  • #21
    Originally posted by Sdar View Post
    Compiles a lot faster (16 seconds vs 184 seconds) but the render times are the same and it has problems with the latest 2.72 build kernel but it could be a blender bug albeit intel opencl compiles and runs fine with 2.72 too...

    Maybe this year we can see cycles working on the Foss Opencl implementation, I doubt it'll ever run correctly on the proprietary one...

    Anyway I'll change to another vendor next time, I'm just too tired of waiting for a fix.
    I believe that until Blender devs will make some adjustments to OpenCL 2.0 code inside Blender nothing will change. More over it was told many years that they need smaller kernel to work on GPU and only after it was told by and AMD guy to Ton directly they thought about that.

    Comment


    • #22
      Originally posted by sunweb View Post
      I believe that until Blender devs will make some adjustments to OpenCL 2.0 code inside Blender nothing will change. More over it was told many years that they need smaller kernel to work on GPU and only after it was told by and AMD guy to Ton directly they thought about that.
      If AMD is the odd man out - if it works on Nvidia and Intel GPUs...

      Comment


      • #23
        Originally posted by curaga View Post
        If AMD is the odd man out - if it works on Nvidia and Intel GPUs...
        We all know that software side isn't strong with AMD comparing to the other 2. What we do know is that the guy from Kronos group said that OpenCL wasn't designed to eat up big kernels because GPU arhitecture was not well suited for it anyway. It obviously some smart hack done by Nvidia and Intel that make it seem as a usual thing. Nvidia had CUDA for years and influenced OpenCL alot as well as Cycles creation, Intel is known for lightning fast compilers for years.
        But i was talking about OpenCL 2.0 code inside the Blender. AMD could very well solve the trouble but only for OpenCL 2.0, we won't know it unless Cycles will have some changes in it. Besides that:
        AMD has released its first fully-functional OpenCL 2.0 driver aimed at early adopter developers. We?re not broadly promoting this driver because we have a performance-optimized driver coming out shortly and we suggest you wait for that version.


        There must be triage from both sides to solve the problem. If BF would've made adjutsments to Cycles as it was said by Kronos and AMD people for years and it wouldn't work than we could b**ch about bad AMD but right now BF is no better. More over Brecht could've made those changes at the start of the project when there were alittle of code and it would save the trouble but chose not to.

        Comment


        • #24
          Originally posted by Marc Driftmeyer View Post
          If you follow LLVM you'll know the R600 with the RadeonSI driver using OpenCL with Blender shows the folks at Blender using the Mesa OpenCL Stack need to learn more about LLVM and SSA when dealing with their architecture.



          They don't know how LLVM works enough to do the work.


          So, to answer your question, yes, the PHINode does satisfy the SSA
          requirements.
          Try reading the whole thread next time.

          Comment


          • #25
            Originally posted by sunweb View Post
            We all know that software side isn't strong with AMD comparing to the other 2. What we do know is that the guy from Kronos group said that OpenCL wasn't designed to eat up big kernels because GPU arhitecture was not well suited for it anyway. It obviously some smart hack done by Nvidia and Intel that make it seem as a usual thing. Nvidia had CUDA for years and influenced OpenCL alot as well as Cycles creation, Intel is known for lightning fast compilers for years.
            But i was talking about OpenCL 2.0 code inside the Blender. AMD could very well solve the trouble but only for OpenCL 2.0, we won't know it unless Cycles will have some changes in it. Besides that:



            There must be triage from both sides to solve the problem. If BF would've made adjutsments to Cycles as it was said by Kronos and AMD people for years and it wouldn't work than we could b**ch about bad AMD but right now BF is no better. More over Brecht could've made those changes at the start of the project when there were alittle of code and it would save the trouble but chose not to.
            According to http://www.blender.org/news/siggraph-2014-report/
            AMD?s OpenCL progress is still fuzzy ? but the general message is that we should try to split up the render kernel in smaller parts ? graphics cards just don?t like it and we might run into similar issues with CUDA sooner than later anyway.
            I had a long chat with Khronos? president on this topic too ? according to him we shouldn?t give up so easily, an industry compatible OpenCL compiler shouldn?t have such problems with building Cycles. To be continued?
            AMD says they should split the kernel, while Khronos president says a compiler should be able to build it... or am I reading this incorrectly.

            Comment


            • #26
              No support for the R9 280 series. Installation failed on my Ubuntu 14.04.1 setup

              Comment


              • #27
                Originally posted by idanny View Post
                No support for the R9 280 series. Installation failed on my Ubuntu 14.04.1 setup
                FWIW, the public, general-purpose, all-cards supporting Catalyst 14.10 Beta driver release is scheduled for next week.

                Comment


                • #28
                  Originally posted by Sdar View Post
                  http://lists.cs.uiuc.edu/pipermail/l...er/077059.html



                  Try reading the whole thread next time.
                  I did read it and if you read the actual paper: http://www.cs.utexas.edu/~pingali/CS.../ssaCytron.pdf

                  You'll understand the end of the thread that the guy didn't fully understand what was being developed wrt to LLVM and it's architecture.

                  Comment


                  • #29
                    Originally posted by log0 View Post
                    According to http://www.blender.org/news/siggraph-2014-report/

                    AMD says they should split the kernel, while Khronos president says a compiler should be able to build it... or am I reading this incorrectly.
                    Please quote the part where AMD says something. AMD said something only on their support resource. The person that Ton spoke with was from Khronos group and its funny that formating on the site changed and it looks like those are different point when original formating had it as the same point. So the "general message" suggest that Khronos guy said that same.

                    Comment


                    • #30
                      Originally posted by d2kx View Post
                      FWIW, the public, general-purpose, all-cards supporting Catalyst 14.10 Beta driver release is scheduled for next week.
                      Good news! Looking forward to try this one out. Till then i am using the default Wrapper xorg thing in Ubuntu.

                      Comment

                      Working...
                      X