<html>
    <head>
      <base href="https://bugs.freedesktop.org/">
    </head>
    <body><span class="vcard"><a class="email" href="mailto:jason@jlekstrand.net" title="Jason Ekstrand <jason@jlekstrand.net>"> <span class="fn">Jason Ekstrand</span></a>
</span> changed
          <a class="bz_bug_link 
          bz_status_RESOLVED  bz_closed"
   title="RESOLVED NOTABUG - Invalid subgroupSize for Intel GPU"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=108875">bug 108875</a>
          <br>
             <table border="1" cellspacing="0" cellpadding="8">
          <tr>
            <th>What</th>
            <th>Removed</th>
            <th>Added</th>
          </tr>

         <tr>
           <td style="text-align:right;">Resolution</td>
           <td>---
           </td>
           <td>NOTABUG
           </td>
         </tr>

         <tr>
           <td style="text-align:right;">Status</td>
           <td>NEW
           </td>
           <td>RESOLVED
           </td>
         </tr></table>
      <p>
        <div>
            <b><a class="bz_bug_link 
          bz_status_RESOLVED  bz_closed"
   title="RESOLVED NOTABUG - Invalid subgroupSize for Intel GPU"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=108875#c6">Comment # 6</a>
              on <a class="bz_bug_link 
          bz_status_RESOLVED  bz_closed"
   title="RESOLVED NOTABUG - Invalid subgroupSize for Intel GPU"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=108875">bug 108875</a>
              from <span class="vcard"><a class="email" href="mailto:jason@jlekstrand.net" title="Jason Ekstrand <jason@jlekstrand.net>"> <span class="fn">Jason Ekstrand</span></a>
</span></b>
        <pre>Just because the advertised subgroupSize is 32 doesnn't mean we have to run
with "full" subgroups.  Intel hardware has dispatch widths of 8, 16, and 32. 
In the Vulkan subgroup model, dispatch modes of 8 and 16 are advertised as a
subgroup size of 32 and where only the first 8 or 16 invocations are enabled. 
This is entirely in-line with the spec; there is nothing that guarantees that
local_size_x = subgroupSize will get you a single invocation or that
gl_NumSubgroups = DIV_ROUND_UP(gl_WorkgroupSize.x * gl_WorkgroupSize.y *
gl_WorkgroupSize.z, gl_SubgroupSize).

Besides that, you really don't want to run 32-wide on Intel.  The performance
trade-offs almost always aren't worth it.  We default to 16-wide because that's
tends to be a nice sweet-spot but sometimes 8-wide is even better.</pre>
        </div>
      </p>


      <hr>
      <span>You are receiving this mail because:</span>

      <ul>
          <li>You are the assignee for the bug.</li>
          <li>You are the QA Contact for the bug.</li>
      </ul>
    </body>
</html>