Announcement

Collapse
No announcement yet.

Performance Work Coming Up For Mesa 7.11

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

  • #41
    Originally posted by agd5f View Post
    If we knew exactly what needed to be fixed we'd do it. It comes down to lots of profiling. [..]
    Sounds like an interesting task. What exactly has to be profiled, how is it done and is any special ability required to do that despite of a lot of patience? :-)

    Comment


    • #42
      Originally posted by hal2k1 View Post
      Does the description above agree with this new article?
      http://www.phoronix.com/scan.php?pag...item&px=OTAwMg
      I'm not quite sure what you are asking. Radeon pageflipping support was merged for 2.6.38 and as I noted, it's vblank synced, but does reduce memory bandwidth requirements and pipeline latency since it avoids the additional blit required for the bufferswap which has a factor in performance. The frame rate if refresh rate limited, but the memory bandwidth requirements are lower. For non-vblank-synced pageflipping, the dri2 code in the xserver needs some additional work.

      Comment


      • #43
        Originally posted by smitty3268 View Post
        Yes, and for good reason.

        This has been discussed to death over and over again. The current limiting factor in getting better OSS drivers is developer manpower, so anything that can help them out should be the #1 priority. The only people hurt by this are the devs at AMD and NVidia, and they have enough manpower to overcome the problem. Everyone else does not.
        You say "hurt by this", what do you mean by "this"? The lack of an ABI?

        On another note (start rant, hehe):
        Even though they have manpower (and that they are "huge evil corporations"TM), they usually don't have an infinite amount of money. Especially not AMD I assume...
        The manpower they do have, they're probably trying to do the best they can with. Sadly that "best" usually doesn't align with what we think is the best for the Linux community. But of course I don't think the developers are to blame for this, it's usually decisions made far above their heads. They're probably just as frustrated with their managers as us other coders are, hehe

        Comment


        • #44
          Originally posted by LibertyZero View Post
          Sounds like an interesting task. What exactly has to be profiled, how is it done and is any special ability required to do that despite of a lot of patience? :-)
          Basically profile a 3D app whose performance you want to improve and see where most of the time is spent. If a lot of time per frame is spent copying vertexes, you might look at those paths in the driver. rinse and repeat.

          Comment


          • #45
            Heh... Stable ABI... "Hey Gallium dudes, forget about LLVM; stick to TGSL for ten years! Oh and Limug guys; forget about a good kernel; do it like NT!" Yeah that'll be the day, morons, when the reason for closed ddrivers is simply speed. Speed that FLOSS doesn't have because of a manpower problem and not an awesomeness problem.

            Comment


            • #46
              Originally posted by V!NCENT View Post
              Heh... Stable ABI... "Hey Gallium dudes, forget about LLVM; stick to TGSL for ten years! Oh and Limug guys; forget about a good kernel; do it like NT!" Yeah that'll be the day, morons, when the reason for closed ddrivers is simply speed. Speed that FLOSS doesn't have because of a manpower problem and not an awesomeness problem.
              I'm not saying anything has to be static for 10 years, there's difference between breaking ABIs and developing ABIs. But my question was sincere, I wasn't ironic (nor was it a rhetorical question).

              Comment


              • #47
                Originally posted by smitty3268 View Post
                "So you're calling stupid anyone who wishes for a driver ABI."

                Yes, and for good reason.

                This has been discussed to death over and over again. The current limiting factor in getting better OSS drivers is developer manpower, so anything that can help them out should be the #1 priority. The only people hurt by this are the devs at AMD and NVidia, and they have enough manpower to overcome the problem. Everyone else does not.
                That may be valid points, but the pro-ABI people also have valid points. You can't call people idiots just because you disagree with them. Name calling means it's you who doesn't have arguments.

                Comment


                • #48
                  Originally posted by RealNC View Post
                  No, it doesn't. The GPL has a "NO WARRANTY" (yes, in capitals) section.

                  There goes your "responsibility."
                  They talk the talk of no responsibility but they walk the walk of responsibility. Reverse liars compared to any other group of people.

                  Comment


                  • #49
                    Originally posted by agd5f View Post
                    I'm not quite sure what you are asking. Radeon pageflipping support was merged for 2.6.38 and as I noted, it's vblank synced, but does reduce memory bandwidth requirements and pipeline latency since it avoids the additional blit required for the bufferswap which has a factor in performance. The frame rate if refresh rate limited, but the memory bandwidth requirements are lower. For non-vblank-synced pageflipping, the dri2 code in the xserver needs some additional work.
                    If my understanding is correct, pageflipping for vblank-synched works by changing the pointers to the back buffer during the vblank period.

                    How does it work for non-vblank-synced?

                    Comment


                    • #50
                      Originally posted by Danny View Post
                      If my understanding is correct, pageflipping for vblank-synched works by changing the pointers to the back buffer during the vblank period.

                      How does it work for non-vblank-synced?
                      The pointers are swapped immediately rather than during the vblank period so it will tear, but you'll get really fast frame rates.

                      Comment

                      Working...
                      X