<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">On 22/12/17 09:30, Sagar Arun Kamble
      wrote:<br>
    </div>
    <blockquote type="cite"
      cite="mid:523e5349-2db7-747a-bea0-774227913592@intel.com">
      <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
      <p><br>
      </p>
      <br>
      <div class="moz-cite-prefix">On 12/21/2017 6:29 PM, Lionel
        Landwerlin wrote:<br>
      </div>
      <blockquote type="cite"
        cite="mid:62813062-eba1-0fa2-1959-6abf19e3dcae@intel.com">
        <div class="moz-cite-prefix">Some more findings I made while
          playing with this series & GPUTop.<br>
          Turns out the 2ms drift per second is due to timecounter.
          Adding the delta this way :<br>
          <br>
          <a class="moz-txt-link-freetext"
href="https://github.com/djdeath/linux/commit/7b002cb360483e331053aec0f98433a5bd5c5c3f#diff-9b74bd0cfaa90b601d80713c7bd56be4R607"
            moz-do-not-send="true">https://github.com/djdeath/linux/commit/7b002cb360483e331053aec0f98433a5bd5c5c3f#diff-9b74bd0cfaa90b601d80713c7bd56be4R607</a><br>
          <br>
          Eliminates the drift.</div>
      </blockquote>
      I see two imp. changes 1. approximation of start time during
      init_timecounter 2. overflow handling in delta accumulation.<br>
      With these incorporated, I guess timecounter should also work in
      same fashion.<br>
    </blockquote>
    <br>
    I think the arithmetic in timecounter is inherently lossy and that's
    why we're seeing a drift. Could we be using it wrong?<br>
    <br>
    In the patch above, I think there is still a drift because of the
    potential fractional part loss at every delta we add.<br>
    But it should only be a fraction of a nanosecond multiplied by the
    number of reports over a period of time.<br>
    With a report every 1us, that should still be much less than a 1ms
    of drift over 1s.<br>
    <br>
    We can probably do better by always computing the clock using the
    entire delta rather than the accumulated delta.<br>
    <br>
    <blockquote type="cite"
      cite="mid:523e5349-2db7-747a-bea0-774227913592@intel.com">
      <blockquote type="cite"
        cite="mid:62813062-eba1-0fa2-1959-6abf19e3dcae@intel.com">
        <div class="moz-cite-prefix"> Timelines of perf i915 tracepoints
          & OA reports now make a lot more sense.<br>
          <br>
          There is still the issue that reading the CPU clock & the
          RCS timestamp is inherently not atomic. So there is a delta
          there.<br>
          I think we should add a new i915 perf record type to express
          the delta that we measure this way :<br>
          <br>
          <a class="moz-txt-link-freetext"
href="https://github.com/djdeath/linux/commit/7b002cb360483e331053aec0f98433a5bd5c5c3f#diff-9b74bd0cfaa90b601d80713c7bd56be4R2475"
            moz-do-not-send="true">https://github.com/djdeath/linux/commit/7b002cb360483e331053aec0f98433a5bd5c5c3f#diff-9b74bd0cfaa90b601d80713c7bd56be4R2475</a><br>
          <br>
          So that userspace knows there might be a global offset between
          the 2 times and is able to present it.<br>
        </div>
      </blockquote>
      agree on this. Delta ns1-ns0 can be interpreted as max drift.<br>
      <blockquote type="cite"
        cite="mid:62813062-eba1-0fa2-1959-6abf19e3dcae@intel.com">
        <div class="moz-cite-prefix"> Measurement on my KBL system were
          in the order of a few microseconds (~30us).<br>
          I guess we might be able to setup the correlation point better
          (masking interruption?) to reduce the delta.<br>
        </div>
      </blockquote>
      already using spin_lock. Do you mean NMI?<br>
    </blockquote>
    <br>
    I don't actually know much on this point.<br>
    if spin_lock is the best we can do, then that's it :)<br>
    <br>
    <blockquote type="cite"
      cite="mid:523e5349-2db7-747a-bea0-774227913592@intel.com">
      <blockquote type="cite"
        cite="mid:62813062-eba1-0fa2-1959-6abf19e3dcae@intel.com">
        <div class="moz-cite-prefix"> <br>
          Thanks,<br>
          <br>
          -<br>
          Lionel<br>
          <br>
          <br>
          On 07/12/17 00:57, Robert Bragg wrote:<br>
        </div>
        <blockquote type="cite"
cite="mid:CAMou1-2Z7=A_GBcD9a5AvjRGM3_bG-ezoZJnGYvXkrCqqrmT1w@mail.gmail.com">
          <div dir="ltr"><br>
            <div class="gmail_extra"><br>
              <div class="gmail_quote">On Thu, Dec 7, 2017 at 12:48 AM,
                Robert Bragg <span dir="ltr"><<a
                    href="mailto:robert@sixbynine.org" target="_blank"
                    moz-do-not-send="true">robert@sixbynine.org</a>></span>
                wrote:<br>
                <blockquote class="gmail_quote" style="margin:0 0 0
                  .8ex;border-left:1px #ccc solid;padding-left:1ex">
                  <div dir="ltr"><br>
                  </div>
                </blockquote>
                <blockquote class="gmail_quote" style="margin:0 0 0
                  .8ex;border-left:1px #ccc solid;padding-left:1ex">
                  <div dir="ltr">
                    <div class="gmail_extra">
                      <div class="gmail_quote">
                        <div> at least from what I wrote back then it
                          looks like I was seeing a drift of a few
                          milliseconds per second on SKL. I vaguely
                          recall it being much worse given the frequency
                          constants we had for Haswell.<br>
                        </div>
                      </div>
                    </div>
                  </div>
                </blockquote>
                <div><br>
                </div>
                <div>Sorry I didn't actually re-read my own message
                  properly before referencing it :) Apparently the 2ms
                  per second drift was for Haswell, so presumably not
                  quite so bad for SKL. <br>
                </div>
                <div><br>
                </div>
                <div>- Robert<br>
                </div>
              </div>
              <br>
            </div>
          </div>
          <br>
          <fieldset class="mimeAttachmentHeader"></fieldset>
          <br>
          <pre wrap="">_______________________________________________
Intel-gfx mailing list
<a class="moz-txt-link-abbreviated" href="mailto:Intel-gfx@lists.freedesktop.org" moz-do-not-send="true">Intel-gfx@lists.freedesktop.org</a>
<a class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/intel-gfx" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/intel-gfx</a>
</pre>
        </blockquote>
        <p><br>
        </p>
      </blockquote>
      <br>
    </blockquote>
    <p><br>
    </p>
  </body>
</html>