Announcement

Collapse
No announcement yet.

GCC 4.8 vs. LLVM/Clang 3.3 On Intel's Core i7 4770K

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

  • GCC 4.8 vs. LLVM/Clang 3.3 On Intel's Core i7 4770K

    Phoronix: GCC 4.8 vs. LLVM/Clang 3.3 On Intel's Core i7 4770K

    Complementing the earlier Phoronix article about optimized binaries for Intel Haswell CPUs via the "-march=core-avx2" Haswell compiler optimizations, in this article is a comparison of the GCC and LLVM/Clang compilers when targeting the new Core i7 4770K CPU. GCC 4.7.3, GCC 4.8.1, LLVM Clang 3.2, and LLVM Clang 3.3 were the tested compilers under Ubuntu Linux when seeing how well these different compilers optimized for Haswell.

    http://www.phoronix.com/vr.php?view=18807

  • #2
    What is with the insistance of using tests that are designed to compile with OpenMP or in C-Ray's case use -ffast-math that isn't even supported in LLVM/Clang to show comparison results?

    One would expect to toss those out until LLVM/Clang have OpenMP 4 [that is coming] ready for testing on trunk.

    Wake me when C-Ray makefiles are tied to GCC.

    Comment


    • #3
      Originally posted by Marc Driftmeyer View Post
      What is with the insistance of using tests that are designed to compile with OpenMP
      Well it's more interesting than doing a test where one compiler uses openmp and the other doesn't, that's totally pointless. Now atleast he doesn't use openmp.

      Originally posted by Marc Driftmeyer View Post
      or in C-Ray's case use -ffast-math that isn't even supported in LLVM/Clang to show comparison results?
      The specified commandline for the C-Ray test was:

      -lm -lpthread -O3 -march=core-avx2

      -O3 does not enable -ffast-math, -Ofast does, so this test should be without -ffast-math. For the record -ffast-math is a very useful optimization for software which doesn't need the precision, but again, it's not enabled by the settings used in this benchmark.

      The tests which are worthless are the Scimark ones as for some reason Michael won't declare the optimizations used for those.

      For all we know they could be done using -O0, -O1, -O2, which doesn't say anything really as there is no standard correlation of what is enabled in these optimization settings between compilers.

      This is why if you only use one option and want to compare for performance of generated code, you will use -O3 as it is the option which is supposed to generate the fastest code across these compilers.

      Originally posted by Marc Driftmeyer View Post
      Wake me when C-Ray makefiles are tied to GCC.
      ?

      Comment


      • #4
        Originally posted by Marc Driftmeyer View Post
        What is with the insistance of using tests that are designed to compile with OpenMP or in C-Ray's case use -ffast-math that isn't even supported in LLVM/Clang to show comparison results?
        +1. Also, what is with the insistence of using CPUs most people don't even have yet. I'd expect Ivy/Sandy Bridge is what interests people the most at this time.

        Comment


        • #5
          Originally posted by RealNC View Post
          +1. Also, what is with the insistence of using CPUs most people don't even have yet. I'd expect Ivy/Sandy Bridge is what interests people the most at this time.
          I thought that this test report was much better than the other ones: Less biased and less spiced with invalid comparisions.

          @Michael: Thanks!

          Comment


          • #6
            Originally posted by ArneBab View Post
            I thought that this test report was much better than the other ones: Less biased and less spiced with invalid comparisions.

            @Michael: Thanks!
            Agreed. +1

            Comment


            • #7
              Originally posted by XorEaxEax View Post
              The tests which are worthless are the Scimark ones as for some reason Michael won't declare the optimizations used for those.
              It's not that I won't declare them, it's that PTS wasn't able to auto-determine and record what optimizations are used by SciMark.
              Michael Larabel
              http://www.michaellarabel.com/

              Comment


              • #8
                Originally posted by Marc Driftmeyer View Post
                What is with the insistance of using tests that are designed to compile with OpenMP
                Mostly just as a reminder to the point that LLVM doesn't yet handle OpenMP, as an important point for many users.
                Michael Larabel
                http://www.michaellarabel.com/

                Comment


                • #9
                  What about tsx-ni

                  Can we get some tsx-ni comparison benchmarks? 4770k doesn't have it, and for me it's the most exciting feature of haswell.
                  http://ark.intel.com/products/75123/...up-to-3_90-GHz

                  Perhaps do a 4770k vs 4770 comparison, on applications that should benefit from tsx-ni?

                  Comment


                  • #10
                    Originally posted by RealNC View Post
                    +1. Also, what is with the insistence of using CPUs most people don't even have yet. I'd expect Ivy/Sandy Bridge is what interests people the most at this time.
                    I would like to see how the FX 8350 does myself

                    Comment

                    Working...
                    X