Announcement

Collapse
No announcement yet.

Intel Haswell Scheduler Updated In LLVM

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

  • Intel Haswell Scheduler Updated In LLVM

    Phoronix: Intel Haswell Scheduler Updated In LLVM

    Last month in LLVM there was new Sandy Bridge scheduler information to improve the instruction scheduling and other hardware detail changes so LLVM can generate more efficient code for those older CPUs. At that time we learned Intel developers were also planning improvements too for LLVM with newer Haswell / Broadwell / Skylake / Skylake-X CPUs. Improvements have now landed for Haswell...

    http://www.phoronix.com/scan.php?pag...er-Update-LLVM

  • #2
    Probably a stupid question, but would that speed up compiling shaders for radeonsi on a system with a Haswell CPU?

    Comment


    • #3
      Originally posted by geearf View Post
      Probably a stupid question, but would that speed up compiling shaders for radeonsi on a system with a Haswell CPU?
      it shouldn't, this should only make generated (cpu) executables faster

      Comment


      • #4
        Originally posted by davidbepo View Post

        it shouldn't, this should only make generated (cpu) executables faster
        I think this is what he means. It would not speed up the shaders, but it could improve the compilation process.

        Comment


        • #5
          Originally posted by geearf View Post
          Probably a stupid question, but would that speed up compiling shaders for radeonsi on a system with a Haswell CPU?
          It improves code generation (instruction scheduling) FOR haswell cpus.

          Comment


          • #6
            Originally posted by sdack View Post
            I think this is what he means. It would not speed up the shaders, but it could improve the compilation process.
            Correct!
            The compilation of the shader is done by the CPU, and since shader compiling is often pointed out as slow, I'm wondering

            Originally posted by mlau View Post

            It improves code generation (instruction scheduling) FOR haswell cpus.
            Right, so LLVM on my system should be faster, among other things, for compiling no?

            Comment


            • #7
              Originally posted by geearf View Post

              Correct!
              The compilation of the shader is done by the CPU, and since shader compiling is often pointed out as slow, I'm wondering


              Right, so LLVM on my system should be faster, among other things, for compiling no?
              No, the code llvm generates FOR haswell cpu should be faster, and that only if the -mtune haswell switch was used when compiling. So the version of llvm your distio provides, which probably uses generic optimizer flags, won't benefit.

              Comment


              • #8
                Originally posted by Masush5 View Post

                No, the code llvm generates FOR haswell cpu should be faster, and that only if the -mtune haswell switch was used when compiling. So the version of llvm your distio provides, which probably uses generic optimizer flags, won't benefit.
                But if they build llvm on a haswell machine using this new llvm code, they could see slightly faster shader compilation. Of course a shader cache will blow away any improvements on the second run...

                Comment


                • #9
                  Originally posted by Masush5 View Post

                  No, the code llvm generates FOR haswell cpu should be faster, and that only if the -mtune haswell switch was used when compiling. So the version of llvm your distio provides, which probably uses generic optimizer flags, won't benefit.
                  Oooooh, I did not think of that :'/ I don't mind compiling certains packages, but llvm is a bit heavy "/

                  Comment


                  • #10
                    Originally posted by geearf View Post
                    Oooooh, I did not think of that :'/ I don't mind compiling certains packages, but llvm is a bit heavy "/
                    Yeah, one of the main driving factors in my upgrade from a phenom II x6 to a Ryzen 1700 was that it dropped my llvm compile times from 35-45 minutes to 15 or less.

                    Comment

                    Working...
                    X