<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=us-ascii">
<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:DengXian;
panose-1:2 1 6 0 3 1 1 1 1 1;}
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
@font-face
{font-family:"\@DengXian";
panose-1:2 1 6 0 3 1 1 1 1 1;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0cm;
font-size:11.0pt;
font-family:"Calibri",sans-serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:#0563C1;
text-decoration:underline;}
span.EmailStyle21
{mso-style-type:personal-compose;
font-family:"Calibri",sans-serif;
color:windowtext;
font-weight:normal;
font-style:normal;}
.MsoChpDefault
{mso-style-type:export-only;
font-size:10.0pt;}
@page WordSection1
{size:612.0pt 792.0pt;
margin:72.0pt 72.0pt 72.0pt 72.0pt;}
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-GB" link="#0563C1" vlink="#954F72" style="word-wrap:break-word">
<div class="WordSection1">
<p class="MsoNormal">Hi Chris,<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">I’ve been following the discussions between you and Cameron and just so I understand<o:p></o:p></p>
<p class="MsoNormal">things correctly it looks like you’re proposing a new way of writing vector literals in IR<o:p></o:p></p>
<p class="MsoNormal">where the start and stride are determined from the first two values, i.e. <1, 3, …><o:p></o:p></p>
<p class="MsoNormal">has a start of 1 and stride of 2. I quite like the idea of a vector literal, but are you<o:p></o:p></p>
<p class="MsoNormal">proposing that internally in C++ we have some sort of ConstantStepVector that contains<o:p></o:p></p>
<p class="MsoNormal">a start/stride pair? Or when parsing this vector literal do you expect us to internally<o:p></o:p></p>
<p class="MsoNormal">represent this as something like:<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">splat(1) + 2 * stepvector<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">where stepvector would be fixed as <0, 1, 2, 3, …>?<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">When emitting LLVM IR with something like “clang -S -emit-llvm” would you also<o:p></o:p></p>
<p class="MsoNormal">expect to see these vector literals being emitted too? I’m just thinking about how<o:p></o:p></p>
<p class="MsoNormal">this would look in vectorised code, for example normally when vectorising an<o:p></o:p></p>
<p class="MsoNormal">induction variable we’d see something like this:<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">vector.ph:<o:p></o:p></p>
<p class="MsoNormal">…<o:p></o:p></p>
<p class="MsoNormal">vector.body:<o:p></o:p></p>
<p class="MsoNormal"> %vec.ind = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %vector.ph ], …<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Are you suggesting the vector literal would be created in it’s own statement<o:p></o:p></p>
<p class="MsoNormal">in the preheader, or directly as a constant like before:<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">vector.body:<o:p></o:p></p>
<p class="MsoNormal"> %vec.ind = phi <4 x i32> [ <i32 0, i32 1, …>, %vector.ph ], …<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Kind Regards,<o:p></o:p></p>
<p class="MsoNormal">David.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm">
<p class="MsoNormal"><b><span lang="EN-US">From:</span></b><span lang="EN-US"> Chris Tetreault <ctetreau@quicinc.com>
<br>
<b>Sent:</b> 20 January 2021 21:33<br>
<b>To:</b> David Sherwood <David.Sherwood@arm.com>; llvm-dev@lists.llvm.org<br>
<b>Cc:</b> Sander De Smalen <Sander.DeSmalen@arm.com>; Paul Walker <Paul.Walker@arm.com>; Eli Friedman <efriedma@quicinc.com><br>
<b>Subject:</b> RE: [RFC] Introduce a new stepvector operation<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"><span lang="EN-US">David,<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> Thanks for writing this up. I’d just like to speak to some concerns I have regarding shufflevector. As many of us know, shufflevector takes two vectors and a constant vector of i32, and does stuff. The constant shuffle
mask can be scalable or fixed width. The shuffle mask is supposed to be an arbitrary constant vector, however for scalable vectors only zeroinitializer or undef are accepted. There are reasonable technical reasons for this state of affairs, but it reveals
an issue: we don't really handle constant scalable vectors very well. Surely there are other similar issues throughout the codebase, but this is one I struggle with regularly so it sticks out in my mind.<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> However, we probably want to be able to use the stepvector in shufflevector. For instance, if we had a stepvector literal, then we could implement vector concatenation in terms of shufflevector:<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Courier New"">%a_concat_b = shufflevector <4 x i16> %a, <4 x i16> %b, <8 x i32> stepvector<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> In fact, a lot of useful shuffles can be implemented in terms of stepvector multiplied or added to some constants. Pulling from Eli's list in
</span><a href="https://lists.llvm.org/pipermail/llvm-dev/2020-January/138762.html"><span lang="EN-US">https://lists.llvm.org/pipermail/llvm-dev/2020-January/138762.html</span></a><span lang="EN-US">, I can see:<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US">“<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> </span><span lang="EN-US" style="font-family:"Courier New"">%result = shufflevector <vscale x 4 x i32> %v1, <vscale x 4 x i32> %v2, SHUFFLE_NAME<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> SHUFFLE_NAME can be one of the following (with examples of the equivalent <4 x i32> shuffles):<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> concat - Concatenate the two operands (<0, 1, 2, 3, 4, 5, 6, 7>) -> see above<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> split_low - Return the low half of the first operand (<0, 1>) -> stepvector of type <vscale x n/2 x i32><o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> split_high - Return the high half of the first operand (<2, 3>) -> (stepvector + splat(n/2)) of type <vscale x n/2 x i32><o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> zip_low - Zip together the low halves of the two operands (<0, 4, 1, 5>)<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> zip_high - Zip together the high halves of the two operands (<2, 6, 3, 7>)<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> unzip_even - Unzip the even elements of the two operands (<0, 2, 4, 6>) (stepvector + stepvector) of type <vscale x n x i32><o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> unzip_odd - unzip the odd elements of the two operands (<1, 3, 5, 7>) (stepvector + stepvector + splat(1)) of type <vscale x n x i32><o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US">“<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> Unfortunately, all of these cannot be done because shufflevector only supports scalable undef or zeroinitializer. In order to support these cases, we would need to extend shufflevector to support stepvector (for concat),
and arbitrary constant expressions for the rest. Supporting stepvector might not be so hard with the current scheme: if the shuffle is scalable, and the mask is <0, 1, ..., n - 1>, then the input mask was a scalable stepvector. However, I think this illustrates
my proposal: <b>vector pattern literals</b>. They could look like this in IR:<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> </span><b><span lang="EN-US" style="font-family:"Courier New""><0, 1, ...></span></b><span lang="EN-US"> ; stepvector<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> This is also more flexible, because it enables lots of other scalable vector literals:<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> </span><span lang="EN-US" style="font-family:"Courier New""><7, 7, ...></span><span lang="EN-US"> ; splat(7) without writing that horrid insertelement/shufflevector thing<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> </span><span lang="EN-US" style="font-family:"Courier New""><0, 2, ...></span><span lang="EN-US"> ; unzip_even mask<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> </span><span lang="EN-US" style="font-family:"Courier New""><1, 3, ...></span><span lang="EN-US"> ; unzip_odd mask<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> The implementation for shufflevector would be straightforward because the mask for the two currently supported cases of zeroinitializer and undef (<0, 0, ...> <=> zeroinitializer and <undef, undef, ...> <=> undef)
already follow the proposed scheme. This could also have the side benefits of making some IR easier to read (for very wide vectors, the fixed width stepvector could be more than 80 columns wide), and might result in efficiency gains in the compiler (don't
need to walk a very wide vector to see if it is a stepvector; can just canonicalize <0, 1, 2, 3, 4, 5, 6, 7, ..., 2048> once to ConstantPatternVector(0, 1)).<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> Sorry for rambling. I think I’ve personally come around to the idea that a constant would be good. However, a more flexible constant would be best if we’re going to use it to add a bunch of special cases to the codebase.
Most special cases for scalable undef and zeroinitializer can be replaced with equivalent code that also handles vector pattern literals.<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US">Thanks,<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"> Christopher Tetreault<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm">
<p class="MsoNormal"><b><span lang="EN-US">From:</span></b><span lang="EN-US"> David Sherwood <</span><a href="mailto:David.Sherwood@arm.com"><span lang="EN-US">David.Sherwood@arm.com</span></a><span lang="EN-US">>
<br>
<b>Sent:</b> Wednesday, January 20, 2021 8:04 AM<br>
<b>To:</b> </span><a href="mailto:llvm-dev@lists.llvm.org"><span lang="EN-US">llvm-dev@lists.llvm.org</span></a><span lang="EN-US"><br>
<b>Cc:</b> Sander De Smalen <</span><a href="mailto:Sander.DeSmalen@arm.com"><span lang="EN-US">Sander.DeSmalen@arm.com</span></a><span lang="EN-US">>; Paul Walker <</span><a href="mailto:Paul.Walker@arm.com"><span lang="EN-US">Paul.Walker@arm.com</span></a><span lang="EN-US">>;
Chris Tetreault <</span><a href="mailto:ctetreau@quicinc.com"><span lang="EN-US">ctetreau@quicinc.com</span></a><span lang="EN-US">><br>
<b>Subject:</b> [EXT] [RFC] Introduce a new stepvector operation<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><span lang="EN-US"><o:p> </o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">Hi,<o:p></o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"><o:p> </o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">As part of adding support for scalable vectorization we need to update llvm::InnerLoopVectorizer::getStepVector for scalable vectors. Currently this just returns a constant vector
with the sequence <0, 1, 2, 3, ..>, however this assumes we are using fixed length vectors. For scalable vectors we need an operation that does the same thing, but without having to explicitly initalise all the elements. Any new stepvector operation we provide
could also be used for fixed length vectors too if desired.</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">I believe the desirable properties of the operation should be:</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">1. The operation requires no arguments and simply returns a vector with the numeric sequence <0, 1, 2, …></span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">2. For types with a large number of elements, e.g. <vscale x 32 x i8> (vscale = 16), there is the possibility of the sequence value exceeding the limit of the type midway through
the vector. In such cases we define the operation such that those elements are undefined or poison values.</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">A simple ‘stepvector’ operation (however we choose to implement it) with the properties described above can then be used together with additional ‘mul’ and ‘add’ instructions to
create any arbitrary linear sequence, i.e. <0, 2, 4, 6, …> or <1, 3, 5, 7, …></span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">The first possible implementation with the properties as described above involves using a new llvm.stepvector() intrinsic with no arguments that simply returns a vector sequence
<0, 1, 2, …> of the requested type, i.e.</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> declare <vscale x 4 x i32> @llvm.stepvector.nxv4i32()</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">Introducing a new intrinsic is simple to implement and we can easily come up with an appropriate cost model – cheap for fixed width vectors or scalable vectors using SVE where we
have the ‘index’ instruction.</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">However, since such an intrinsic is effectively returning a constant vector sequence we could instead implement it using a new ‘stepvector’ constant in a similar way to how ‘zeroinitializer’
works. This would be done using a new ConstantStepVector class similar to ConstantAggregateZero and would return a vector with the numeric sequence <0, 1, 2, …>. The main advantages of using a constant over an intrinsic are:</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">1. It is easy to write tests in LLVM IR since ‘stepvector’ would work in the same way as ‘zeroinitializer’, i.e. “%1 = add <4 x i32> %0, stepvector”</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">2. Creation of the node is easy with the simple interface:</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> static Constant *ConstantStepVector::get(Type Ty)</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">3. It is easy to do optimisations, e.g. CSE, and pattern matching in IR.</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">The main disadvantages are:</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"> </span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">1. A scalable constant cannot be represented as well in the .data section, although we can still create a constant based on the architectural maximum for vscale. It’s worth pointing
out that this problem also exists for zeroinitializer too – we’re just more likely to have cheap instructions to do the job.</span><o:p></o:p></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">2. Harder to fit into the cost model due to it being a constant.<o:p></o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">3. There are some concerns that we might then have to support stepvector as a constant in the shufflevector operation too and that it should be restricted to zeroinitializer only.<o:p></o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"><o:p> </o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"><o:p> </o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">Any thoughts or feedback you have would be much appreciated!<o:p></o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black"><o:p> </o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">Kind Regards,<o:p></o:p></span></p>
<p style="margin:0cm"><span style="font-family:"Arial",sans-serif;color:black">David Sherwood.<o:p></o:p></span></p>
<p style="margin:0cm"><o:p> </o:p></p>
</div>
</body>
</html>