Results 1 to 6 of 6

Thread: LLVM/Clang 3.3 Performing Against GCC For Old Intel CPU

  1. #1
    Join Date
    Jan 2007
    Posts
    14,378

    Default LLVM/Clang 3.3 Performing Against GCC For Old Intel CPU

    Phoronix: LLVM/Clang 3.3 Performing Against GCC For Old Intel CPU

    Generally when delivering new Linux compiler benchmarks on Phoronix it's from x86/ARM hardware within the past two years. It's the most recent generations of hardware that excites us the most and generally where the professional Linux software developers are focusing their time and resources. However, after seeing the recent LLVM/Clang 3.3 performance improvements for this forthcoming open-source compiler release, we decided to go back a bit in CPU history.

    http://www.phoronix.com/vr.php?view=18671

  2. #2
    Join Date
    Jul 2011
    Posts
    73

    Default

    Thanks for the interesting test!

    I had wished to get a stronger conclusion here: The results are pretty strong, so I think you could say more than just that LLVM is close to on-par with GCC.

    So I generated a table of the results myself (It would be cool if this were already in the conclusion):

    Benchmark results summary.

    Higher is always better (if less was better, I took the inverse)

    Code:
    |                    |              GCC | LLVM (the better one) | relative (GCC/LLVM) |
    |--------------------+------------------+-----------------------+---------------------|
    | HMMer              |  0.0151331719128 |       0.0125786163522 |  1.2030871670677201 |
    | MAFFT              |  0.0292056074766 |       0.0284495021337 |   1.026577102802947 |
    | GM Blur            |               30 |                    34 |  0.8823529411764706 |
    | GM Sharpen         |               16 |                    22 |  0.7272727272727273 |
    | GM Resizing        |               41 |                    40 |               1.025 |
    | GM Color Space     |               55 |                    42 |  1.3095238095238095 |
    | Himeno Pressure    |           672.19 |                636.89 |  1.0554255836957718 |
    | C-Ray              | 5.88373734996e-3 |      5.63920374443e-3 |  1.0433631442686448 |
    | Primesieve         |  2.1435767722e-3 |      1.06682597935e-3 |  2.0093031231823275 |
    | Apache             |          5814.74 |               5943.17 |  0.9783903203172717 |
    |--------------------+------------------+-----------------------+---------------------|
    | Average            |                  |                       |  1.1260295919307688 |
    | Standard deviation |                  |                       | 0.34831009079240405 |
    #+TBLFM: @2$4..@11$4='(/ (float $2) (float $3));N::@12$4='(/ (apply '+ '( @2$4..@11$4 )) 10.0);N::@13$4='(progn (defun std (&rest samples) "Compute the standard deviation of the given samples."  (let ((mean (/ (apply '+ samples) (length samples))) (dev 0)) (loop for sample in samples do (setq dev (+ dev (* (- sample mean) (- sample mean))))) (sqrt (/ dev (max 1 (1- (length samples))))))) (apply 'std'( @2$4..@11$4 )));N
    If you’d do the same evaluation for each LLVM version, you could actually say which of them is better.

    Currently the results show that on average GCC is over 12% faster than LLVM, but that the tests differ widely by area: In Primesieve, database search and Graphics Magic colorspace GCC wins by over 20%, in Graphics Magic Blur and Sharpen it loses by more than 10%. The rest of the differences is small.

    PS: Table calculated using the given table formula in Emacs Org-mode.

    PPS: I left out the PHP compilation because it does not say anything about the efficiency of the result.
    Last edited by ArneBab; 04-24-2013 at 10:00 AM.

  3. #3
    Join Date
    Feb 2008
    Location
    Linuxland
    Posts
    4,995

    Default

    The openbenchmarking site lets you do summaries like that.

  4. #4
    Join Date
    Jul 2011
    Posts
    73

    Default

    Quote Originally Posted by curaga View Post
    The openbenchmarking site lets you do summaries like that.
    I did not find how I can do that easily (when I checked last time). Also that is not the point: The point is, that this information would be additional value for the article: Not only providing raw data but also some interpretation.

  5. #5

    Default

    Quote Originally Posted by ArneBab View Post
    I did not find how I can do that easily (when I checked last time). Also that is not the point: The point is, that this information would be additional value for the article: Not only providing raw data but also some interpretation.
    A concise table won't be added to articles themselves since readers should be clicking on every page anyhow if they wish to learn all the information.

  6. #6
    Join Date
    Jul 2011
    Posts
    73

    Default

    Quote Originally Posted by Michael View Post
    A concise table won't be added to articles themselves since readers should be clicking on every page anyhow if they wish to learn all the information.
    I think that sad, but if it works best for you Iíll have to accept it.

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •