<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:"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;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0in;
margin-bottom:.0001pt;
font-size:11.0pt;
font-family:"Calibri",sans-serif;
color:black;}
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";
color:black;}
p.msonormal0, li.msonormal0, div.msonormal0
{mso-style-name:msonormal;
mso-margin-top-alt:auto;
margin-right:0in;
mso-margin-bottom-alt:auto;
margin-left:0in;
font-size:11.0pt;
font-family:"Calibri",sans-serif;
color:black;}
span.HTMLPreformattedChar
{mso-style-name:"HTML Preformatted Char";
mso-style-priority:99;
mso-style-link:"HTML Preformatted";
font-family:Consolas;
color:black;}
span.EmailStyle21
{mso-style-type:personal-reply;
font-family:"Calibri",sans-serif;
color:windowtext;}
.MsoChpDefault
{mso-style-type:export-only;
font-size:10.0pt;}
@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 bgcolor="white" lang="EN-US" link="blue" vlink="purple">
<div class="WordSection1">
<p class="MsoNormal"><span style="color:windowtext"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="color:windowtext"><o:p> </o:p></span></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal" style="margin-left:.5in"><b><span style="color:windowtext">From:</span></b><span style="color:windowtext"> Hal Finkel <hfinkel@anl.gov>
<br>
<b>Sent:</b> Tuesday, July 24, 2018 5:05 AM<br>
<b>To:</b> Craig Topper <craig.topper@gmail.com>; 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>; Nema, Ashutosh <Ashutosh.Nema@amd.com>; llvm-dev <llvm-dev@lists.llvm.org><br>
<b>Subject:</b> Re: [llvm-dev] [LoopVectorizer] Improving the performance of dot product reduction loop<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal" style="margin-left:.5in"><o:p> </o:p></p>
<p style="margin-left:.5in"><o:p> </o:p></p>
<p class="MsoNormal" style="margin-left:.5in"><o:p> </o:p></p>
<div>
<p class="MsoNormal" style="margin-left:.5in">On 07/23/2018 06:23 PM, Hal Finkel via llvm-dev wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal" style="margin-left:.5in"><o:p> </o:p></p>
<div>
<p class="MsoNormal" style="margin-left:.5in">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" style="margin-left:.5in">Hello all,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal" style="margin-left:.5in"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal" style="margin-left:.5in">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>
<p class="MsoNormal" style="margin-left:.5in"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal" style="margin-left:.5in">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" style="margin-left:.5in"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal" style="margin-left:.5in">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" style="margin-left:.5in"><br>
Why v*i8 loads? I thought that we have 16-bit and 32-bit types here?<br>
<br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal" style="margin-left:.5in">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. 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">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" style="margin-left:.5in"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal" style="margin-left:.5in">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" style="margin-left:.5in"><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>
</blockquote>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Vectorizer considers the largest data type size in the loop body and considers the maximum possible VF as 8, hence in this example it generates the 128 bits vpmaddwd.<o:p></o:p></p>
<p class="MsoNormal">Like to share my observation where instead of forcing vector factor to 16 (by using pragma) tried option “vectorizer-maximize-bandwidth”.<o:p></o:p></p>
<p class="MsoNormal">“vectorizer-maximize-bandwidth” considers the smallest data type size in the loop body and allows the possible VF up to 16, but LV selects the VF as 8 though VF16 has the same cost.<o:p></o:p></p>
<p class="MsoNormal">LV: Vector loop of width 8 costs: 1.<o:p></o:p></p>
<p class="MsoNormal">LV: Vector loop of width 16 costs: 1.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">It’s because of below check in LV:<o:p></o:p></p>
<p class="MsoNormal">LoopVectorizationCostModel::selectVectorizationFactor() {<o:p></o:p></p>
<p class="MsoNormal">…<o:p></o:p></p>
<p class="MsoNormal"> if (VectorCost < Cost) {<o:p></o:p></p>
<p class="MsoNormal"> Cost = VectorCost;<o:p></o:p></p>
<p class="MsoNormal"> Width = i;<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">I don’t know the history behind this change but wondering why it’s like this, at least for “vectorizer-maximize-bandwidth” it should be (VectorCost <= Cost).<o:p></o:p></p>
<p class="MsoNormal">By forcing the vector factor to 16 it generates 256 bits vpmaddwd.<span style="color:windowtext"><o:p></o:p></span></p>
<p class="MsoNormal"><span style="color:windowtext"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="color:windowtext">Regards,<o:p></o:p></span></p>
<p class="MsoNormal"><span style="color:windowtext">Ashutosh</span><o:p></o:p></p>
</div>
</body>
</html>