<html>
  <head>
    <meta content="text/html; charset=utf-8" http-equiv="Content-Type">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <br>
    <br>
    <div class="moz-cite-prefix">On 2017年05月12日 16:43, Christian König
      wrote:<br>
    </div>
    <blockquote
      cite="mid:b82fc6cd-c19e-85c2-8628-89c2cb10c5c0@vodafone.de"
      type="cite">
      <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
      <div class="moz-cite-prefix">Am 12.05.2017 um 10:37 schrieb
        zhoucm1:<br>
      </div>
      <blockquote cite="mid:59157453.8010308@amd.com" type="cite"> <br>
        <br>
        <div class="moz-cite-prefix">On 2017年05月12日 16:33, Christian
          König wrote:<br>
        </div>
        <blockquote
          cite="mid:21e206ed-8d60-450f-6d23-e01c68c5ad73@vodafone.de"
          type="cite">
          <div class="moz-cite-prefix">Am 12.05.2017 um 10:25 schrieb
            zhoucm1:<br>
          </div>
          <blockquote cite="mid:5915717A.3000209@amd.com" type="cite"> <br>
            <div class="moz-cite-prefix">On 2017年05月10日 05:47,
              Kasiviswanathan, Harish wrote:<br>
            </div>
            <blockquote
cite="mid:DM3PR1201MB10382F282D646D630708246D8CEF0@DM3PR1201MB1038.namprd12.prod.outlook.com"
              type="cite">
              <pre wrap="">Hi,

Please review the patch set that supports amdgpu VM update via CPU. This feature provides improved performance for compute (HSA) where mapping / unmapping is carried out (by Kernel) independent of command submissions (done directly by user space). This version doesn't support shadow copy of VM page tables for CPU based update.</pre>
            </blockquote>
            I think your improved performance is from less waiting for
            cs, generally, SDMA engine updating page table is faster
            than CPU, otherwise we don't need sdma for updating PT.<br>
            So whether your this improvement proves we have some
            redundant sync when mapping / unmapping, if yes, we should
            fix that, then not sure if CPU method is need or not.<br>
          </blockquote>
          <br>
          The problem is that the KFD is designed synchronously for page
          table updates. In other words they need to wait for the update
          to finish and that takes time.<br>
          <br>
          Apart from that your comment is absolutely correct, we found
          that the SDMA is sometimes much faster to do the update than
          the CPU.<br>
        </blockquote>
        If the sdma is faster, even they wait for finish, which time is
        shorter than CPU, isn't it? Of course, the precondition is sdma
        is exclusive. They can reserve a sdma for PT updating.<br>
      </blockquote>
      <br>
      No, if I understood Felix numbers correctly the setup and wait
      time for SDMA is a bit (but not much) longer than doing it with
      the CPU.<br>
      <br>
      What would really help is to fix the KFD design and work with
      async page tables updates there as well.<br>
    </blockquote>
    OK, no problem, just curious.<br>
    <br>
    Regards,<br>
    David Zhou<br>
    <blockquote
      cite="mid:b82fc6cd-c19e-85c2-8628-89c2cb10c5c0@vodafone.de"
      type="cite"> <br>
      Regards,<br>
      Christian.<br>
      <br>
      <blockquote cite="mid:59157453.8010308@amd.com" type="cite"> <br>
        Regards,<br>
        David Zhou<br>
        <blockquote
          cite="mid:21e206ed-8d60-450f-6d23-e01c68c5ad73@vodafone.de"
          type="cite"> <br>
          Regards,<br>
          Christian.<br>
          <br>
          <blockquote cite="mid:5915717A.3000209@amd.com" type="cite"> <br>
            Regards,<br>
            David Zhou<br>
            <blockquote
cite="mid:DM3PR1201MB10382F282D646D630708246D8CEF0@DM3PR1201MB1038.namprd12.prod.outlook.com"
              type="cite">
              <pre wrap="">Best Regards,
Harish

</pre>
              <br>
              <fieldset class="mimeAttachmentHeader"></fieldset>
              <br>
              <pre wrap="">_______________________________________________
amd-gfx mailing list
<a moz-do-not-send="true" class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>
<a moz-do-not-send="true" class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a>
</pre>
            </blockquote>
            <br>
            <br>
            <fieldset class="mimeAttachmentHeader"></fieldset>
            <br>
            <pre wrap="">_______________________________________________
amd-gfx mailing list
<a moz-do-not-send="true" class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>
<a moz-do-not-send="true" class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a>
</pre>
          </blockquote>
          <p><br>
          </p>
        </blockquote>
        <br>
        <br>
        <fieldset class="mimeAttachmentHeader"></fieldset>
        <br>
        <pre wrap="">_______________________________________________
amd-gfx mailing list
<a moz-do-not-send="true" class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>
<a moz-do-not-send="true" class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a>
</pre>
      </blockquote>
      <p><br>
      </p>
    </blockquote>
    <br>
  </body>
</html>