Announcement

Collapse
No announcement yet.

AMD Releases Optimizing C/C++ Compiler For Ryzen

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

  • #21
    Originally posted by juno View Post

    Why would marketing claim it is an "AMD compiler" if it's just upstream llvm instead of advertising that the company is contributing to free software?
    1. We're talking about marketing teams
    2. We're talking about AMD's marketing team

    Comment


    • #22
      Originally posted by bridgman View Post
      What part of "we have been pushing changes upstream for a quite a while" was not clear ?
      Should we expect no differences in performance of the generated binaries then?

      Comment


      • #23
        Originally posted by andrei_me View Post
        Should we expect no differences in performance of the generated binaries then?
        Depends on timing relative to cleaning up work for upstreaming. There's usually a certain amount of "let's try this... hey that works.... now how the heck do we fit that into the upstream project patterns ?".

        I don't think we have enough people working on the compiler to maintain an IP-safe but non-upstream tree yet, similar to what we recently started publishing for gfx drivers (amd-staging-4.9 / 4.11).
        Test signature

        Comment


        • #24
          Originally posted by peppercats View Post

          yeahhh I don't think this is gonna be upstreamed
          If it's worth enough, they'll probably find a way. It does seem a bit overly protective but who really benefits from compiler enhancements. Answer: everyone!

          Comment


          • #25
            Originally posted by ermo View Post

            Erh ... surely people who turn on the relevant architecture flag will benefit?

            Or are you assuming that the people who want the extra performance won't recompile from source for some unspecified reason?
            And who on earth should enable a specific architecture flag despite Gentoo users? All the software out there is compiled more generically ...

            Comment


            • #26
              GCC could also use a little more love in the Zen department... the initial znver1 patches submitted almost 2 years ago use the cost model of carrizo/excavator, which
              produces worse results than e.g. using the haswell cost model.

              Comment


              • #27
                just a thought this zen ryzenmicroarchitecture must have brought a bunch of new instructions along with it. as far as i know these instructions can vary between manufacturers thinking of intel's vtx vs amds. will not gcc be updated to take benefits of these instructions while generating binaries for ryzen platform? did i get it right? pardon my lack of knowledge

                Comment


                • #28
                  Originally posted by Kemosabe View Post

                  And who on earth should enable a specific architecture flag despite Gentoo users? All the software out there is compiled more generically ...
                  Extra performance has its place ( video editing software, math heavy software, game engines, etc. ). GUI is not such software. Be sure most of the heavy software have different code paths for different processors ( even from the same vendor ), even for different cache size/hierarchies for that matter.

                  Comment


                  • #29
                    Shouldn't they just merge this into Clang/LLVM eventually? Along with SPIR and everything else?

                    Comment


                    • #30
                      Originally posted by bridgman View Post



                      Huh ? The compiler teams have been submitting Ryzen support upstream into LLVM for quite a while now.

                      This is just a binary release with the standard binary EULA as far as I know.

                      Why would we try to push a binary upstream ?
                      Seemed like you guys were just taking clang, adding patches, and holding Zen optimization hostage in a proprietary compiler.
                      Good to know that's not what it is.

                      Comment

                      Working...
                      X