Announcement

Collapse
No announcement yet.

"Mega Drivers" Being Proposed For A Faster Mesa

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

  • "Mega Drivers" Being Proposed For A Faster Mesa

    Phoronix: "Mega Drivers" Being Proposed For A Faster Mesa

    Eric Anholt of Intel is currently working on some experimental Mesa code for shipping "Mega drivers", or building all of the Mesa/Gallium3D drivers together as a single shared object library file. There's some promise to this mega drivers concept in enhancing performance due to compiler/linker optimizations...

    http://www.phoronix.com/vr.php?view=MTQzMTY

  • #2
    this can be good as you can easily work bugs out (Like in the Linux Kernal) but hmm? i see a few things that can go bad doing it this way the major one is ptaches

    Comment


    • #3
      the performace gain from this seems to be very small, but yah every bit helps.
      the time for CD distros is over go with a dvd, no one uses it anyway they use a usb :P

      Comment


      • #4
        Originally posted by LinuxGamer View Post
        this can be good as you can easily work bugs out (Like in the Linux Kernal) but hmm? i see a few things that can go bad doing it this way the major one is ptaches
        I don't think this is supposed to actually unify the code, but instead it just changes a few things to be able to build all of mesa as a single object, and then makes some changes on which symbols are exported and which symbols are just for mesa.

        Comment


        • #5
          Originally posted by LinuxGamer View Post
          this can be good as you can easily work bugs out (Like in the Linux Kernal) but hmm? i see a few things that can go bad doing it this way the major one is ptaches
          as mrugiero stated this a compilation angle things and not code related beyond the fixes needed to make it compile, this technique will allow better runtime performance since the symbol lookup tables get greatly reduced and could help to reduce it more making public only those symbols really needed instead of expose mesa internal symbols.

          additional side effects is protecting mesa internal could reduce the surface of action of an potential security attack in some obscure scenarios

          Comment


          • #6
            6% performance improvement across the board? Sounds like the next step. What does this mean for out-of-tree Direct3D and CUDA state trackers?

            Comment


            • #7
              Originally posted by MrTheSoulz View Post
              the performace gain from this seems to be very small, but yah every bit helps.
              the time for CD distros is over go with a dvd, no one uses it anyway they use a usb :P
              USB? Who installs using those? TFTP/PXE/NFS is where it's at!

              Comment


              • #8
                Originally posted by Serge View Post
                USB? Who installs using those? TFTP/PXE/NFS is where it's at!
                I bet that 99% of all Home users never heard from TFTP/PXE/NFS. The Big Problem is that the most routers can't ship an tftp server ip with dhcp.
                Last edited by Nille; 08-09-2013, 03:15 AM.

                Comment


                • #9
                  -0.798709% +/- 0.333703%

                  Error in same order of magnitude as the results? We call it insignificant (or nothing).

                  I also can't understand those trailing digits (these after 0.7 in -0.798709%) if we have an error of 0.3%. they absolutely don't have any value or meaning.

                  Comment


                  • #10
                    only for classic drivers

                    libdricore is only used by classic mesa drivers (mainly only intel chips). Gallium drivers are not affected.

                    Comment


                    • #11
                      Originally posted by arashbm View Post
                      Error in same order of magnitude as the results? We call it insignificant (or nothing).

                      I also can't understand those trailing digits (these after 0.7 in -0.798709%) if we have an error of 0.3%. they absolutely don't have any value or meaning.
                      The results look significant to me at alpha=5% (non null hypothesis).

                      Comment


                      • #12
                        Originally posted by MrTheSoulz View Post
                        the performace gain from this seems to be very small, but yah every bit helps.
                        the time for CD distros is over go with a dvd, no one uses it anyway they use a usb :P
                        Ubuntu, Kubuntu, Xubuntu and Mint can't be put a CD already since a few months ago.

                        Comment


                        • #13
                          Originally posted by ethana2 View Post
                          6% performance improvement across the board? Sounds like the next step. What does this mean for out-of-tree Direct3D and CUDA state trackers?
                          Seems that it's only when hardware acceleration disabled (see INTEL_NO_HW=1 in article). If so, perfomance grow should be lesser in normal situation, i.e when acceleration enabled.

                          Comment


                          • #14
                            I see two downsides:
                            - you can't remove or update drivers separately
                            - the dropped DRI drivers that one could previously build from an earlier release, would this break them?

                            Comment


                            • #15
                              Originally posted by arashbm View Post
                              Error in same order of magnitude as the results? We call it insignificant (or nothing).

                              I also can't understand those trailing digits (these after 0.7 in -0.798709%) if we have an error of 0.3%. they absolutely don't have any value or meaning.
                              Those numbers aren't the interesting ones!
                              These are:
                              On a megadrivers+LTO compared to non-megadrivers, non-LTO, the difference was -6.35008% +/- 0.675067% (n=10).

                              Comment

                              Working...
                              X