<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 26/12/17 05:32, Sagar Arun Kamble
      wrote:<br>
    </div>
    <blockquote type="cite"
      cite="mid:04eca028-3705-5a28-b500-089ca19e712c@intel.com">
      <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
      <p><br>
      </p>
      <br>
      <div class="moz-cite-prefix">On 12/22/2017 3:46 PM, Lionel
        Landwerlin wrote:<br>
      </div>
      <blockquote type="cite"
        cite="mid:30446e83-3d0f-ae39-5c0d-a23bc4c89557@intel.com">
        <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">
          <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.</blockquote>
      Could you share details about platform, scenario in which 2ms
      drift per second is being seen with timecounter.<br>
      I did not observe this on SKL.<br>
    </blockquote>
    <br>
    The 2ms drift was on SKL GT4.<br>
    <br>
    With the patch above, I'm seeing only a ~40us drift over ~7seconds
    of recording both perf tracepoints & i915 perf reports.<br>
    I'm tracking the kernel tracepoints adding gem requests and the i915
    perf reports.<br>
    Here a screenshot at the beginning of the 7s recording :
    <a class="moz-txt-link-freetext" href="https://i.imgur.com/hnexgjQ.png">https://i.imgur.com/hnexgjQ.png</a> (you can see the gem request add
    before the work starts in the i915 perf reports).<br>
    At the end of the recording, the gem requests appear later than the
    work in the i915 perf report : <a class="moz-txt-link-freetext" href="https://i.imgur.com/oCd0C9T.png">https://i.imgur.com/oCd0C9T.png</a><br>
    <br>
    I'll try to prepare some IGT tests that show the drift using perf
    & i915 perf, so we can run those on different platforms.<br>
    I tend to mostly test on a SKL GT4 & KBL GT2, but BXT definitely
    needs more attention...<br>
    <br>
    <blockquote type="cite"
      cite="mid:04eca028-3705-5a28-b500-089ca19e712c@intel.com">
      <blockquote type="cite"
        cite="mid:30446e83-3d0f-ae39-5c0d-a23bc4c89557@intel.com"> Could
        we be using it wrong?<br>
        <br>
      </blockquote>
      if we use two changes highlighted above with timecounter maybe we
      will get same results as your current implementation.<br>
      <blockquote type="cite"
        cite="mid:30446e83-3d0f-ae39-5c0d-a23bc4c89557@intel.com"> 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>
      </blockquote>
      timecounter interface takes care of fractional parts so that
      should help us.<br>
      we can either go with timecounter or our own implementation
      provided conversions are precise.<br>
    </blockquote>
    <br>
    Looking at clocks_calc_mult_shift(), it seems clear to me that there
    is less precision when using timecounter :<br>
    <br>
     /*<br>
      * Find the conversion shift/mult pair which has the best<br>
      * accuracy and fits the maxsec conversion range:<br>
      */<br>
    <br>
    On the other hand, there is a performance penalty for doing a div64
    for every report.<br>
    <br>
    <blockquote type="cite"
      cite="mid:04eca028-3705-5a28-b500-089ca19e712c@intel.com">
      <blockquote type="cite"
        cite="mid:30446e83-3d0f-ae39-5c0d-a23bc4c89557@intel.com"> We
        can probably do better by always computing the clock using the
        entire delta rather than the accumulated delta.<br>
        <br>
      </blockquote>
      issue is that the reported clock cycles in the OA report is 32bits
      LSB of GPU TS whereas counter is 36bits. Hence we will need to<br>
      accumulate the delta. ofc there is assumption that two reports
      can't be spaced with count value of 0xffffffff apart.<br>
    </blockquote>
    <br>
    You're right :)<br>
    I thought maybe we could do this : <br>
    <br>
    Look at teduhe opening period parameter, if it's superior to the
    period of timestamps wrapping, make sure we schle some work on
    kernel context to generate a context switch report (like at least
    once every 6 minutes on gen9).<br>
    <br>
    <blockquote type="cite"
      cite="mid:04eca028-3705-5a28-b500-089ca19e712c@intel.com">
      <blockquote type="cite"
        cite="mid:30446e83-3d0f-ae39-5c0d-a23bc4c89557@intel.com">
        <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>
      </blockquote>
      <br>
    </blockquote>
    <p><br>
    </p>
  </body>
</html>