<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<meta name="Generator" content="Microsoft Word 15 (filtered medium)">
<style><!--
/* Font Definitions */
@font-face
{font-family:"MS Mincho";
panose-1:2 2 6 9 4 2 5 8 3 4;}
@font-face
{font-family:"Cambria Math";
panose-1:2 4 5 3 5 4 6 3 2 4;}
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
@font-face
{font-family:Consolas;
panose-1:2 11 6 9 2 2 4 3 2 4;}
@font-face
{font-family:"\@MS Mincho";
panose-1:2 2 6 9 4 2 5 8 3 4;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0in;
margin-bottom:.0001pt;
font-size:12.0pt;
font-family:"Times New Roman",serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:blue;
text-decoration:underline;}
a:visited, span.MsoHyperlinkFollowed
{mso-style-priority:99;
color:purple;
text-decoration:underline;}
pre
{mso-style-priority:99;
mso-style-link:"HTML Preformatted Char";
margin:0in;
margin-bottom:.0001pt;
font-size:10.0pt;
font-family:"Courier New";}
span.HTMLPreformattedChar
{mso-style-name:"HTML Preformatted Char";
mso-style-priority:99;
mso-style-link:"HTML Preformatted";
font-family:Consolas;}
span.EmailStyle19
{mso-style-type:personal-reply;
font-family:"Calibri",sans-serif;
color:#1F497D;}
.MsoChpDefault
{mso-style-type:export-only;
font-family:"Calibri",sans-serif;}
@page WordSection1
{size:8.5in 11.0in;
margin:1.0in 1.0in 1.0in 1.0in;}
div.WordSection1
{page:WordSection1;}
--></style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
</head>
<body lang="EN-US" link="blue" vlink="purple">
<div class="WordSection1">
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal">>But you're right we could also just detect the reduction and add two halves.<o:p></o:p></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">If we take this approach, we also need to teach vector type legalizer about that the last add of the following is reduction sum and thus add to the same temp
instead of creating the second temp. Should not be a big deal --- just calling it out.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal" style="margin-left:.5in"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">full vector unit stride load of A[],<br>
full vector unit stride load of B[],<br>
sign extend both, (this makes it 2x full vector, on the surface)<br>
multiply<br>
add<br>
<br>
</span><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Hideki<o:p></o:p></span></p>
<p class="MsoNormal"><a name="_MailEndCompose"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></a></p>
<p class="MsoNormal"><a name="_____replyseparator"></a><b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif">From:</span></b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif"> Craig Topper [mailto:craig.topper@gmail.com]
<br>
<b>Sent:</b> Tuesday, July 24, 2018 10:05 AM<br>
<b>To:</b> Hal Finkel <hfinkel@anl.gov><br>
<b>Cc:</b> Saito, Hideki <hideki.saito@intel.com>; estotzer@ti.com; Nemanja Ivanovic <nemanja.i.ibm@gmail.com>; Adam Nemet <anemet@apple.com>; graham.hunter@arm.com; Michael Kuperstein <mkuper@google.com>; Sanjay Patel <spatel@rotateright.com>; Simon Pilgrim
<llvm-dev@redking.me.uk>; ashutosh.nema@amd.com; llvm-dev <llvm-dev@lists.llvm.org><br>
<b>Subject:</b> Re: [LoopVectorizer] Improving the performance of dot product reduction loop<o:p></o:p></span></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal">On Tue, Jul 24, 2018 at 6:10 AM Hal Finkel <<a href="mailto:hfinkel@anl.gov">hfinkel@anl.gov</a>> wrote:<o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">On 07/23/2018 06:37 PM, Craig Topper wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal"><br clear="all">
<o:p></o:p></p>
<div>
<div>
<p class="MsoNormal">~Craig<o:p></o:p></p>
</div>
</div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal">On Mon, Jul 23, 2018 at 4:24 PM Hal Finkel <<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>> wrote:<o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">On 07/23/2018 05:22 PM, Craig Topper wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal">Hello all,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">This code <a href="https://godbolt.org/g/tTyxpf" target="_blank">
https://godbolt.org/g/tTyxpf</a> is a dot product reduction loop multipying sign extended 16-bit values to produce a 32-bit accumulated result. The x86 backend is currently not able to optimize it as well as gcc and icc. The IR we are getting from the loop
vectorizer has several v8i32 adds and muls inside the loop. These are fed by v8i16 loads and sexts from v8i16 to v8i32. The x86 backend recognizes that these are addition reductions of multiplication so we use the vpmaddwd instruction which calculates 32-bit
products from 16-bit inputs and does a horizontal add of adjacent pairs. A vpmaddwd given two v8i16 inputs will produce a v4i32 result.<o:p></o:p></p>
</div>
</div>
</blockquote>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">That godbolt link seems wrong. It wasn't supposed to be clang IR. This should be right.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">In the example code, because we are reducing the number of elements from 8->4 in the vpmaddwd step we are left with a width mismatch between vpmaddwd and the vpaddd instruction that we use to sum with the results from the previous loop
iterations. We rely on the fact that a 128-bit vpmaddwd zeros the upper bits of the register so that we can use a 256-bit vpaddd instruction so that the upper elements can keep going around the loop without being disturbed in case they weren't initialized
to 0. But this still means the vpmaddwd instruction is doing half the amount of work the CPU is capable of if we had been able to use a 256-bit vpmaddwd instruction. Additionally, future x86 CPUs will be gaining an instruction that can do VPMADDWD and VPADDD
in one instruction, but that width mismatch makes that instruction difficult to utilize.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">In order for the backend to handle this better it would be great if we could have something like two v32i8 loads, two shufflevectors to extract the even elements and the odd elements to create four v16i8 pieces.<o:p></o:p></p>
</div>
</div>
</blockquote>
<p class="MsoNormal"><br>
Why v*i8 loads? I thought that we have 16-bit and 32-bit types here?<o:p></o:p></p>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Oops that should have been v16i16. Mixed up my 256-bit types.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal">Sign extend each of those pieces. Multiply the two even pieces and the two odd pieces separately, sum those results with a v8i32 add. Then another v8i32 add to accumulate the previous loop iterations.<o:p></o:p></p>
</div>
</div>
</blockquote>
</div>
</blockquote>
</div>
</div>
</blockquote>
<p class="MsoNormal"><br>
I'm still missing something. Why do you want to separate out the even and odd parts instead of just adding up the first half of the numbers and the second half?<o:p></o:p></p>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Doing even/odd matches up with a pattern I already have to support for the code in <a href="https://reviews.llvm.org/D49636">https://reviews.llvm.org/D49636</a>. I wouldn't even need to detect is as a reduction to do the reassocation since
even/odd exactly matches the behavior of the instruction. But you're right we could also just detect the reduction and add two halves.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<p class="MsoNormal"><br>
Thanks again,<br>
Hal<br>
<br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal">Then ensures that no pieces exceed the target vector width and the final operation is correctly sized to go around the loop in one register. All but the last add can then be pattern matched to vpmaddwd as proposed in <a href="https://reviews.llvm.org/D49636" target="_blank">https://reviews.llvm.org/D49636</a>.
And for the future CPU the whole thing can be matched to the new instruction.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Do other targets have a similar instruction or a similar issue to this? Is this something we can solve in the loop vectorizer? Or should we have a separate IR transformation that can recognize this pattern and generate the new sequence?
As a separate pass we would need to pair two vector loads together, remove a reduction step outside the loop and remove half the phis assuming the loop was partially unrolled. Or if there was only one add/mul inside the loop we'd have to reduce its width and
the width of the phi.<o:p></o:p></p>
</div>
</div>
</blockquote>
<p class="MsoNormal"><br>
Can you explain how the desired code from the vectorizer differs from the code that the vectorizer produces if you add '#pragma clang loop vectorize(enable) vectorize_width(16)' above the loop? I tried it in your godbolt example and the generated code looks
very similar to the icc-generated code.<o:p></o:p></p>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">It's similar, but the vpxor %xmm0, %xmm0, %xmm0 is being unnecessarily carried across the loop. It's then redundantly added twice in the reduction after the loop despite it being 0. This happens because we basically tricked the backend
into generating a 256-bit vpmaddwd concated with a 256-bit zero vector going into a 512-bit vaddd before type legalization. The 512-bit concat and vpaddd get split during type legalization, and the high half of the add gets constant folded away. I'm guessing
we probably finished with 4 vpxors before the loop but MachineCSE(or some other pass?) combined two of them when it figured out the loop didn't modify them.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<p class="MsoNormal"><br>
Thanks again,<br>
Hal<br>
<br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal">Thanks,<br clear="all">
<o:p></o:p></p>
<div>
<div>
<p class="MsoNormal">~Craig<o:p></o:p></p>
</div>
</div>
</div>
</blockquote>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<pre>-- <o:p></o:p></pre>
<pre>Hal Finkel<o:p></o:p></pre>
<pre>Lead, Compiler Technology and Programming Languages<o:p></o:p></pre>
<pre>Leadership Computing Facility<o:p></o:p></pre>
<pre>Argonne National Laboratory<o:p></o:p></pre>
</div>
</blockquote>
</div>
</div>
</blockquote>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<pre>-- <o:p></o:p></pre>
<pre>Hal Finkel<o:p></o:p></pre>
<pre>Lead, Compiler Technology and Programming Languages<o:p></o:p></pre>
<pre>Leadership Computing Facility<o:p></o:p></pre>
<pre>Argonne National Laboratory<o:p></o:p></pre>
</div>
</blockquote>
</div>
</div>
</div>
</body>
</html>