<html>
  <head>
    <meta content="text/html; charset=windows-1252"
      http-equiv="Content-Type">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">On 2016-03-23 10:42, Mark Abraham
      wrote:<br>
    </div>
    <blockquote
cite="mid:CAMNuMAT8Wgv4ZroixK-JqNxvyYhpZzWNr7R8AK4rgetgFeEjrQ@mail.gmail.com"
      type="cite">
      <meta http-equiv="Content-Type" content="text/html;
        charset=windows-1252">
      <div dir="ltr">Hi,<br>
        <br>
        <div class="gmail_quote">
          <div dir="ltr">On Wed, Mar 23, 2016 at 9:44 AM Berk Hess &lt;<a
              moz-do-not-send="true" href="mailto:hess@kth.se"><a class="moz-txt-link-abbreviated" href="mailto:hess@kth.se">hess@kth.se</a></a>&gt;
            wrote:<br>
          </div>
          <blockquote class="gmail_quote" style="margin:0 0 0
            .8ex;border-left:1px #ccc solid;padding-left:1ex">Hi,<br>
            <br>
            Luckily Szilard does thorough testing and noticed a
            performance<br>
            degradation in change set 25 of <a moz-do-not-send="true"
              href="https://gerrit.gromacs.org/#/c/5232/"
              rel="noreferrer" target="_blank">https://gerrit.gromacs.org/#/c/5232/</a>
            The<br>
            only signifcant change with respect to previous sets is
            replacing C<br>
            pointers by std::vector. I traced the performance difference
            back to a<br>
            single loop, which must have become several factors slower
            to explain<br>
            the time difference. I get the performance back when
            replacing the<br>
            vector by a pointer extracted with .data(), see below. I
            looked at the<br>
            assembly code from gcc 5.3.1 and the vector case generated
            200 extra<br>
            instructions, which makes it difficult to see what the
            actual difference<br>
            is. The pointer case uses a lot of vmovss and vaddss, which
            the vector<br>
            one does much less, but this is only serial SIMD
            instruction. I thought<br>
            that [] in vector might does bounds checking,</blockquote>
          <div><br>
          </div>
          <div>Definitely not in release builds.</div>
          <div> </div>
          <blockquote class="gmail_quote" style="margin:0 0 0
            .8ex;border-left:1px #ccc solid;padding-left:1ex"> but it
            seems it does not.<br>
            Can anyone explain why the vector case can be so slow?</blockquote>
          <blockquote class="gmail_quote" style="margin:0 0 0
            .8ex;border-left:1px #ccc solid;padding-left:1ex">
            If this is a general issue (with RVec or more?), we need to
            always extra<br>
            a pointer with .data() for use in all inner-loops. This is
            pretty<br>
            annoying and difficult to enforce.<br>
            <br>
            Cheers,<br>
            <br>
            Berk<br>
            <br>
                                    const std::vector&lt;RVec&gt;
            f_foreign =<br>
            idt_foreign-&gt;force<br>
          </blockquote>
          <div><br>
          </div>
          <div>This does a copy of the vector, and doesn't seem to be in
            any version of this patch in gerrit. Is this what you meant
            to write?</div>
        </div>
      </div>
    </blockquote>
    I tried this. But my original "vectorized" patch set took a pointer
    from idt_foreign and did not copy the vector, that gives the same,
    slow, performance.<br>
    <br>
    Berk<br>
    <blockquote
cite="mid:CAMNuMAT8Wgv4ZroixK-JqNxvyYhpZzWNr7R8AK4rgetgFeEjrQ@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_quote">
          <div><br>
          </div>
          <div>Mark</div>
          <div><br>
          </div>
          <blockquote class="gmail_quote" style="margin:0 0 0
            .8ex;border-left:1px #ccc solid;padding-left:1ex">
            or<br>
                                    const RVec               *f_foreign 
             =<br>
            idt_foreign-&gt;force.data();<br>
            <br>
                                     int natom =
            atomList-&gt;atom.size();<br>
                                     for (int i = 0; i &lt; natom; i++)<br>
                                     {<br>
                                         int ind = atomList-&gt;atom[i];<br>
                                         rvec_inc(f[ind],
            f_foreign[ind]);<br>
                                     }<br>
            --<br>
            Gromacs Developers mailing list<br>
            <br>
            * Please search the archive at <a moz-do-not-send="true"
              href="http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List"
              rel="noreferrer" target="_blank">http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List</a>
            before posting!<br>
            <br>
            * Can't post? Read <a moz-do-not-send="true"
              href="http://www.gromacs.org/Support/Mailing_Lists"
              rel="noreferrer" target="_blank">http://www.gromacs.org/Support/Mailing_Lists</a><br>
            <br>
            * For (un)subscribe requests visit<br>
            <a moz-do-not-send="true"
href="https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers"
              rel="noreferrer" target="_blank">https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers</a>
            or send a mail to <a moz-do-not-send="true"
              href="mailto:gmx-developers-request@gromacs.org"
              target="_blank">gmx-developers-request@gromacs.org</a>.<br>
          </blockquote>
        </div>
      </div>
      <br>
      <fieldset class="mimeAttachmentHeader"></fieldset>
      <br>
    </blockquote>
    <br>
  </body>
</html>