<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
"http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="X-UA-Compatible" content="IE=Edge" />
<meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
<title>User Guide for NVPTX Back-end — LLVM 8 documentation</title>
<link rel="stylesheet" href="_static/llvm-theme.css" type="text/css" />
<link rel="stylesheet" href="_static/pygments.css" type="text/css" />
<script type="text/javascript" id="documentation_options" data-url_root="./" src="_static/documentation_options.js"></script>
<script type="text/javascript" src="_static/jquery.js"></script>
<script type="text/javascript" src="_static/underscore.js"></script>
<script type="text/javascript" src="_static/doctools.js"></script>
<script type="text/javascript" src="_static/language_data.js"></script>
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="User Guide for AMDGPU Backend" href="AMDGPUUsage.html" />
<link rel="prev" title="How To Use Attributes" href="HowToUseAttributes.html" />
<style type="text/css">
table.right { float: right; margin-left: 20px; }
table.right td { border: 1px solid #ccc; }
</style>
</head><body>
<div class="logo">
<a href="index.html">
<img src="_static/logo.png"
alt="LLVM Logo" width="250" height="88"/></a>
</div>
<div class="related" role="navigation" aria-label="related navigation">
<h3>Navigation</h3>
<ul>
<li class="right" style="margin-right: 10px">
<a href="genindex.html" title="General Index"
accesskey="I">index</a></li>
<li class="right" >
<a href="AMDGPUUsage.html" title="User Guide for AMDGPU Backend"
accesskey="N">next</a> |</li>
<li class="right" >
<a href="HowToUseAttributes.html" title="How To Use Attributes"
accesskey="P">previous</a> |</li>
<li><a href="http://llvm.org/">LLVM Home</a> | </li>
<li><a href="index.html">Documentation</a>»</li>
</ul>
</div>
<div class="document">
<div class="documentwrapper">
<div class="body" role="main">
<div class="section" id="user-guide-for-nvptx-back-end">
<h1>User Guide for NVPTX Back-end<a class="headerlink" href="#user-guide-for-nvptx-back-end" title="Permalink to this headline">¶</a></h1>
<div class="contents local topic" id="contents">
<ul class="simple">
<li><a class="reference internal" href="#introduction" id="id11">Introduction</a></li>
<li><a class="reference internal" href="#conventions" id="id12">Conventions</a><ul>
<li><a class="reference internal" href="#marking-functions-as-kernels" id="id13">Marking Functions as Kernels</a></li>
<li><a class="reference internal" href="#address-spaces" id="id14">Address Spaces</a></li>
<li><a class="reference internal" href="#triples" id="id15">Triples</a></li>
</ul>
</li>
<li><a class="reference internal" href="#nvptx-intrinsics" id="id16">NVPTX Intrinsics</a><ul>
<li><a class="reference internal" href="#address-space-conversion" id="id17">Address Space Conversion</a><ul>
<li><a class="reference internal" href="#llvm-nvvm-ptr-to-gen-intrinsics" id="id18">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.*.to.gen</span></code>’ Intrinsics</a></li>
<li><a class="reference internal" href="#llvm-nvvm-ptr-gen-to-intrinsics" id="id19">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.gen.to.*</span></code>’ Intrinsics</a></li>
</ul>
</li>
<li><a class="reference internal" href="#reading-ptx-special-registers" id="id20">Reading PTX Special Registers</a><ul>
<li><a class="reference internal" href="#llvm-nvvm-read-ptx-sreg" id="id21">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.read.ptx.sreg.*</span></code>’</a></li>
</ul>
</li>
<li><a class="reference internal" href="#barriers" id="id22">Barriers</a><ul>
<li><a class="reference internal" href="#llvm-nvvm-barrier0" id="id23">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.barrier0</span></code>’</a></li>
</ul>
</li>
<li><a class="reference internal" href="#other-intrinsics" id="id24">Other Intrinsics</a></li>
</ul>
</li>
<li><a class="reference internal" href="#linking-with-libdevice" id="id25">Linking with Libdevice</a><ul>
<li><a class="reference internal" href="#reflection-parameters" id="id26">Reflection Parameters</a></li>
</ul>
</li>
<li><a class="reference internal" href="#executing-ptx" id="id27">Executing PTX</a></li>
<li><a class="reference internal" href="#common-issues" id="id28">Common Issues</a><ul>
<li><a class="reference internal" href="#ptxas-complains-of-undefined-function-nvvm-reflect" id="id29">ptxas complains of undefined function: __nvvm_reflect</a></li>
</ul>
</li>
<li><a class="reference internal" href="#tutorial-a-simple-compute-kernel" id="id30">Tutorial: A Simple Compute Kernel</a><ul>
<li><a class="reference internal" href="#the-kernel" id="id31">The Kernel</a></li>
<li><a class="reference internal" href="#dissecting-the-kernel" id="id32">Dissecting the Kernel</a><ul>
<li><a class="reference internal" href="#data-layout" id="id33">Data Layout</a></li>
<li><a class="reference internal" href="#target-intrinsics" id="id34">Target Intrinsics</a></li>
<li><a class="reference internal" href="#id10" id="id35">Address Spaces</a></li>
<li><a class="reference internal" href="#kernel-metadata" id="id36">Kernel Metadata</a></li>
</ul>
</li>
<li><a class="reference internal" href="#running-the-kernel" id="id37">Running the Kernel</a></li>
</ul>
</li>
<li><a class="reference internal" href="#tutorial-linking-with-libdevice" id="id38">Tutorial: Linking with Libdevice</a></li>
</ul>
</div>
<div class="section" id="introduction">
<h2><a class="toc-backref" href="#id11">Introduction</a><a class="headerlink" href="#introduction" title="Permalink to this headline">¶</a></h2>
<p>To support GPU programming, the NVPTX back-end supports a subset of LLVM IR
along with a defined set of conventions used to represent GPU programming
concepts. This document provides an overview of the general usage of the back-
end, including a description of the conventions used and the set of accepted
LLVM IR.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">This document assumes a basic familiarity with CUDA and the PTX
assembly language. Information about the CUDA Driver API and the PTX assembly
language can be found in the <a class="reference external" href="http://docs.nvidia.com/cuda/index.html">CUDA documentation</a>.</p>
</div>
</div>
<div class="section" id="conventions">
<h2><a class="toc-backref" href="#id12">Conventions</a><a class="headerlink" href="#conventions" title="Permalink to this headline">¶</a></h2>
<div class="section" id="marking-functions-as-kernels">
<h3><a class="toc-backref" href="#id13">Marking Functions as Kernels</a><a class="headerlink" href="#marking-functions-as-kernels" title="Permalink to this headline">¶</a></h3>
<p>In PTX, there are two types of functions: <em>device functions</em>, which are only
callable by device code, and <em>kernel functions</em>, which are callable by host
code. By default, the back-end will emit device functions. Metadata is used to
declare a function as a kernel function. This metadata is attached to the
<code class="docutils literal notranslate"><span class="pre">nvvm.annotations</span></code> named metadata object, and has the following format:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>!0 = !{<function-ref>, metadata !"kernel", i32 1}
</pre></div>
</div>
<p>The first parameter is a reference to the kernel function. The following
example shows a kernel function calling a device function in LLVM IR. The
function <code class="docutils literal notranslate"><span class="pre">@my_kernel</span></code> is callable from host code, but <code class="docutils literal notranslate"><span class="pre">@my_fmad</span></code> is not.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">define</span> <span class="kt">float</span> <span class="vg">@my_fmad</span><span class="p">(</span><span class="kt">float</span> <span class="nv">%x</span><span class="p">,</span> <span class="kt">float</span> <span class="nv">%y</span><span class="p">,</span> <span class="kt">float</span> <span class="nv">%z</span><span class="p">)</span> <span class="p">{</span>
<span class="nv">%mul</span> <span class="p">=</span> <span class="k">fmul</span> <span class="kt">float</span> <span class="nv">%x</span><span class="p">,</span> <span class="nv">%y</span>
<span class="nv">%add</span> <span class="p">=</span> <span class="k">fadd</span> <span class="kt">float</span> <span class="nv">%mul</span><span class="p">,</span> <span class="nv">%z</span>
<span class="k">ret</span> <span class="kt">float</span> <span class="nv">%add</span>
<span class="p">}</span>
<span class="k">define</span> <span class="kt">void</span> <span class="vg">@my_kernel</span><span class="p">(</span><span class="kt">float</span><span class="p">*</span> <span class="nv">%ptr</span><span class="p">)</span> <span class="p">{</span>
<span class="nv">%val</span> <span class="p">=</span> <span class="k">load</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span><span class="p">*</span> <span class="nv">%ptr</span>
<span class="nv">%ret</span> <span class="p">=</span> <span class="k">call</span> <span class="kt">float</span> <span class="vg">@my_fmad</span><span class="p">(</span><span class="kt">float</span> <span class="nv">%val</span><span class="p">,</span> <span class="kt">float</span> <span class="nv">%val</span><span class="p">,</span> <span class="kt">float</span> <span class="nv">%val</span><span class="p">)</span>
<span class="k">store</span> <span class="kt">float</span> <span class="nv">%ret</span><span class="p">,</span> <span class="kt">float</span><span class="p">*</span> <span class="nv">%ptr</span>
<span class="k">ret</span> <span class="kt">void</span>
<span class="p">}</span>
<span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!1</span><span class="p">}</span>
<span class="nv nv-Anonymous">!1</span> <span class="p">=</span> <span class="p">!{</span><span class="kt">void</span> <span class="p">(</span><span class="kt">float</span><span class="p">*)*</span> <span class="vg">@my_kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>When compiled, the PTX kernel functions are callable by host-side code.</p>
</div>
<div class="section" id="address-spaces">
<span id="id1"></span><h3><a class="toc-backref" href="#id14">Address Spaces</a><a class="headerlink" href="#address-spaces" title="Permalink to this headline">¶</a></h3>
<p>The NVPTX back-end uses the following address space mapping:</p>
<blockquote>
<div><table border="1" class="docutils">
<colgroup>
<col width="37%" />
<col width="63%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd"><th class="head">Address Space</th>
<th class="head">Memory Space</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even"><td>0</td>
<td>Generic</td>
</tr>
<tr class="row-odd"><td>1</td>
<td>Global</td>
</tr>
<tr class="row-even"><td>2</td>
<td>Internal Use</td>
</tr>
<tr class="row-odd"><td>3</td>
<td>Shared</td>
</tr>
<tr class="row-even"><td>4</td>
<td>Constant</td>
</tr>
<tr class="row-odd"><td>5</td>
<td>Local</td>
</tr>
</tbody>
</table>
</div></blockquote>
<p>Every global variable and pointer type is assigned to one of these address
spaces, with 0 being the default address space. Intrinsics are provided which
can be used to convert pointers between the generic and non-generic address
spaces.</p>
<p>As an example, the following IR will define an array <code class="docutils literal notranslate"><span class="pre">@g</span></code> that resides in
global device memory.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="vg">@g</span> <span class="p">=</span> <span class="k">internal</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)</span> <span class="k">global</span> <span class="p">[</span><span class="m">4</span> <span class="k">x</span> <span class="k">i32</span><span class="p">]</span> <span class="p">[</span> <span class="k">i32</span> <span class="m">0</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">,</span> <span class="k">i32</span> <span class="m">2</span><span class="p">,</span> <span class="k">i32</span> <span class="m">3</span> <span class="p">]</span>
</pre></div>
</div>
<p>LLVM IR functions can read and write to this array, and host-side code can
copy data to it by name with the CUDA Driver API.</p>
<p>Note that since address space 0 is the generic space, it is illegal to have
global variables in address space 0. Address space 0 is the default address
space in LLVM, so the <code class="docutils literal notranslate"><span class="pre">addrspace(N)</span></code> annotation is <em>required</em> for global
variables.</p>
</div>
<div class="section" id="triples">
<h3><a class="toc-backref" href="#id15">Triples</a><a class="headerlink" href="#triples" title="Permalink to this headline">¶</a></h3>
<p>The NVPTX target uses the module triple to select between 32/64-bit code
generation and the driver-compiler interface to use. The triple architecture
can be one of <code class="docutils literal notranslate"><span class="pre">nvptx</span></code> (32-bit PTX) or <code class="docutils literal notranslate"><span class="pre">nvptx64</span></code> (64-bit PTX). The
operating system should be one of <code class="docutils literal notranslate"><span class="pre">cuda</span></code> or <code class="docutils literal notranslate"><span class="pre">nvcl</span></code>, which determines the
interface used by the generated code to communicate with the driver. Most
users will want to use <code class="docutils literal notranslate"><span class="pre">cuda</span></code> as the operating system, which makes the
generated PTX compatible with the CUDA Driver API.</p>
<p>Example: 32-bit PTX for CUDA Driver API: <code class="docutils literal notranslate"><span class="pre">nvptx-nvidia-cuda</span></code></p>
<p>Example: 64-bit PTX for CUDA Driver API: <code class="docutils literal notranslate"><span class="pre">nvptx64-nvidia-cuda</span></code></p>
</div>
</div>
<div class="section" id="nvptx-intrinsics">
<span id="id2"></span><h2><a class="toc-backref" href="#id16">NVPTX Intrinsics</a><a class="headerlink" href="#nvptx-intrinsics" title="Permalink to this headline">¶</a></h2>
<div class="section" id="address-space-conversion">
<h3><a class="toc-backref" href="#id17">Address Space Conversion</a><a class="headerlink" href="#address-space-conversion" title="Permalink to this headline">¶</a></h3>
<div class="section" id="llvm-nvvm-ptr-to-gen-intrinsics">
<h4><a class="toc-backref" href="#id18">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.*.to.gen</span></code>’ Intrinsics</a><a class="headerlink" href="#llvm-nvvm-ptr-to-gen-intrinsics" title="Permalink to this headline">¶</a></h4>
<div class="section" id="syntax">
<h5>Syntax:<a class="headerlink" href="#syntax" title="Permalink to this headline">¶</a></h5>
<p>These are overloaded intrinsics. You can use these on any pointer types.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.global.to.gen.p0i8.p1i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)</span>
<span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.shared.to.gen.p0i8.p3i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">3</span><span class="p">)*)</span>
<span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">4</span><span class="p">)*)</span>
<span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.local.to.gen.p0i8.p5i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">5</span><span class="p">)*)</span>
</pre></div>
</div>
</div>
<div class="section" id="overview">
<h5>Overview:<a class="headerlink" href="#overview" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.*.to.gen</span></code>’ intrinsics convert a pointer in a non-generic
address space to a generic address space pointer.</p>
</div>
<div class="section" id="semantics">
<h5>Semantics:<a class="headerlink" href="#semantics" title="Permalink to this headline">¶</a></h5>
<p>These intrinsics modify the pointer value to be a valid generic address space
pointer.</p>
</div>
</div>
<div class="section" id="llvm-nvvm-ptr-gen-to-intrinsics">
<h4><a class="toc-backref" href="#id19">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.gen.to.*</span></code>’ Intrinsics</a><a class="headerlink" href="#llvm-nvvm-ptr-gen-to-intrinsics" title="Permalink to this headline">¶</a></h4>
<div class="section" id="id3">
<h5>Syntax:<a class="headerlink" href="#id3" title="Permalink to this headline">¶</a></h5>
<p>These are overloaded intrinsics. You can use these on any pointer types.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.global.p1i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
<span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">3</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.shared.p3i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
<span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">4</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.constant.p4i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
<span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">5</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.local.p5i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
</pre></div>
</div>
</div>
<div class="section" id="id4">
<h5>Overview:<a class="headerlink" href="#id4" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.gen.to.*</span></code>’ intrinsics convert a pointer in the generic
address space to a pointer in the target address space. Note that these
intrinsics are only useful if the address space of the target address space of
the pointer is known. It is not legal to use address space conversion
intrinsics to convert a pointer from one non-generic address space to another
non-generic address space.</p>
</div>
<div class="section" id="id5">
<h5>Semantics:<a class="headerlink" href="#id5" title="Permalink to this headline">¶</a></h5>
<p>These intrinsics modify the pointer value to be a valid pointer in the target
non-generic address space.</p>
</div>
</div>
</div>
<div class="section" id="reading-ptx-special-registers">
<h3><a class="toc-backref" href="#id20">Reading PTX Special Registers</a><a class="headerlink" href="#reading-ptx-special-registers" title="Permalink to this headline">¶</a></h3>
<div class="section" id="llvm-nvvm-read-ptx-sreg">
<h4><a class="toc-backref" href="#id21">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.read.ptx.sreg.*</span></code>’</a><a class="headerlink" href="#llvm-nvvm-read-ptx-sreg" title="Permalink to this headline">¶</a></h4>
<div class="section" id="id6">
<h5>Syntax:<a class="headerlink" href="#id6" title="Permalink to this headline">¶</a></h5>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ntid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ntid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ntid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ctaid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ctaid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ctaid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.nctaid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.nctaid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.nctaid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.warpsize</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="id7">
<h5>Overview:<a class="headerlink" href="#id7" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.*</span></code>’ intrinsics provide access to the PTX
special registers, in particular the kernel launch bounds. These registers
map in the following way to CUDA builtins:</p>
<blockquote>
<div><table border="1" class="docutils">
<colgroup>
<col width="24%" />
<col width="76%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd"><th class="head">CUDA Builtin</th>
<th class="head">PTX Special Register Intrinsic</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even"><td><code class="docutils literal notranslate"><span class="pre">threadId</span></code></td>
<td><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.tid.*</span></code></td>
</tr>
<tr class="row-odd"><td><code class="docutils literal notranslate"><span class="pre">blockIdx</span></code></td>
<td><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.ctaid.*</span></code></td>
</tr>
<tr class="row-even"><td><code class="docutils literal notranslate"><span class="pre">blockDim</span></code></td>
<td><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.ntid.*</span></code></td>
</tr>
<tr class="row-odd"><td><code class="docutils literal notranslate"><span class="pre">gridDim</span></code></td>
<td><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.nctaid.*</span></code></td>
</tr>
</tbody>
</table>
</div></blockquote>
</div>
</div>
</div>
<div class="section" id="barriers">
<h3><a class="toc-backref" href="#id22">Barriers</a><a class="headerlink" href="#barriers" title="Permalink to this headline">¶</a></h3>
<div class="section" id="llvm-nvvm-barrier0">
<h4><a class="toc-backref" href="#id23">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.barrier0</span></code>’</a><a class="headerlink" href="#llvm-nvvm-barrier0" title="Permalink to this headline">¶</a></h4>
<div class="section" id="id8">
<h5>Syntax:<a class="headerlink" href="#id8" title="Permalink to this headline">¶</a></h5>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="kt">void</span> <span class="vg">@llvm.nvvm.barrier0</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="id9">
<h5>Overview:<a class="headerlink" href="#id9" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.barrier0()</span></code>’ intrinsic emits a PTX <code class="docutils literal notranslate"><span class="pre">bar.sync</span> <span class="pre">0</span></code>
instruction, equivalent to the <code class="docutils literal notranslate"><span class="pre">__syncthreads()</span></code> call in CUDA.</p>
</div>
</div>
</div>
<div class="section" id="other-intrinsics">
<h3><a class="toc-backref" href="#id24">Other Intrinsics</a><a class="headerlink" href="#other-intrinsics" title="Permalink to this headline">¶</a></h3>
<p>For the full set of NVPTX intrinsics, please see the
<code class="docutils literal notranslate"><span class="pre">include/llvm/IR/IntrinsicsNVVM.td</span></code> file in the LLVM source tree.</p>
</div>
</div>
<div class="section" id="linking-with-libdevice">
<span id="libdevice"></span><h2><a class="toc-backref" href="#id25">Linking with Libdevice</a><a class="headerlink" href="#linking-with-libdevice" title="Permalink to this headline">¶</a></h2>
<p>The CUDA Toolkit comes with an LLVM bitcode library called <code class="docutils literal notranslate"><span class="pre">libdevice</span></code> that
implements many common mathematical functions. This library can be used as a
high-performance math library for any compilers using the LLVM NVPTX target.
The library can be found under <code class="docutils literal notranslate"><span class="pre">nvvm/libdevice/</span></code> in the CUDA Toolkit and
there is a separate version for each compute architecture.</p>
<p>For a list of all math functions implemented in libdevice, see
<a class="reference external" href="http://docs.nvidia.com/cuda/libdevice-users-guide/index.html">libdevice Users Guide</a>.</p>
<p>To accommodate various math-related compiler flags that can affect code
generation of libdevice code, the library code depends on a special LLVM IR
pass (<code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code>) to handle conditional compilation within LLVM IR. This
pass looks for calls to the <code class="docutils literal notranslate"><span class="pre">@__nvvm_reflect</span></code> function and replaces them
with constants based on the defined reflection parameters. Such conditional
code often follows a pattern:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="kt">float</span> <span class="nf">my_function</span><span class="p">(</span><span class="kt">float</span> <span class="n">a</span><span class="p">)</span> <span class="p">{</span>
<span class="k">if</span> <span class="p">(</span><span class="n">__nvvm_reflect</span><span class="p">(</span><span class="s">"FASTMATH"</span><span class="p">))</span>
<span class="k">return</span> <span class="n">my_function_fast</span><span class="p">(</span><span class="n">a</span><span class="p">);</span>
<span class="k">else</span>
<span class="k">return</span> <span class="n">my_function_precise</span><span class="p">(</span><span class="n">a</span><span class="p">);</span>
<span class="p">}</span>
</pre></div>
</div>
<p>The default value for all unspecified reflection parameters is zero.</p>
<p>The <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass should be executed early in the optimization
pipeline, immediately after the link stage. The <code class="docutils literal notranslate"><span class="pre">internalize</span></code> pass is also
recommended to remove unused math functions from the resulting PTX. For an
input IR module <code class="docutils literal notranslate"><span class="pre">module.bc</span></code>, the following compilation flow is recommended:</p>
<ol class="arabic simple">
<li>Save list of external functions in <code class="docutils literal notranslate"><span class="pre">module.bc</span></code></li>
<li>Link <code class="docutils literal notranslate"><span class="pre">module.bc</span></code> with <code class="docutils literal notranslate"><span class="pre">libdevice.compute_XX.YY.bc</span></code></li>
<li>Internalize all functions not in list from (1)</li>
<li>Eliminate all unused internal functions</li>
<li>Run <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass</li>
<li>Run standard optimization pipeline</li>
</ol>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last"><code class="docutils literal notranslate"><span class="pre">linkonce</span></code> and <code class="docutils literal notranslate"><span class="pre">linkonce_odr</span></code> linkage types are not suitable for the
libdevice functions. It is possible to link two IR modules that have been
linked against libdevice using different reflection variables.</p>
</div>
<p>Since the <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass replaces conditionals with constants, it will
often leave behind dead code of the form:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="nl">entry:</span>
<span class="p">..</span>
<span class="k">br</span> <span class="k">i1</span> <span class="k">true</span><span class="p">,</span> <span class="kt">label</span> <span class="nv">%foo</span><span class="p">,</span> <span class="kt">label</span> <span class="nv">%bar</span>
<span class="nl">foo:</span>
<span class="p">..</span>
<span class="nl">bar:</span>
<span class="c">; Dead code</span>
<span class="p">..</span>
</pre></div>
</div>
<p>Therefore, it is recommended that <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> is executed early in the
optimization pipeline before dead-code elimination.</p>
<p>The NVPTX TargetMachine knows how to schedule <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> at the beginning
of your pass manager; just use the following code when setting up your pass
manager:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="n">std</span><span class="o">::</span><span class="n">unique_ptr</span><span class="o"><</span><span class="n">TargetMachine</span><span class="o">></span> <span class="n">TM</span> <span class="o">=</span> <span class="p">...;</span>
<span class="n">PassManagerBuilder</span> <span class="nf">PMBuilder</span><span class="p">(...);</span>
<span class="k">if</span> <span class="p">(</span><span class="n">TM</span><span class="p">)</span>
<span class="n">TM</span><span class="o">-></span><span class="n">adjustPassManager</span><span class="p">(</span><span class="n">PMBuilder</span><span class="p">);</span>
</pre></div>
</div>
<div class="section" id="reflection-parameters">
<h3><a class="toc-backref" href="#id26">Reflection Parameters</a><a class="headerlink" href="#reflection-parameters" title="Permalink to this headline">¶</a></h3>
<p>The libdevice library currently uses the following reflection parameters to
control code generation:</p>
<table border="1" class="docutils">
<colgroup>
<col width="27%" />
<col width="73%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd"><th class="head">Flag</th>
<th class="head">Description</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even"><td><code class="docutils literal notranslate"><span class="pre">__CUDA_FTZ=[0,1]</span></code></td>
<td>Use optimized code paths that flush subnormals to zero</td>
</tr>
</tbody>
</table>
<p>The value of this flag is determined by the “nvvm-reflect-ftz” module flag.
The following sets the ftz flag to 1.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="nv">!llvm.module.flag</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="k">i32</span> <span class="m">4</span><span class="p">,</span> <span class="nv">!"nvvm-reflect-ftz"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>(<code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">4</span></code> indicates that the value set here overrides the value in another
module we link with. See the <cite>LangRef <LangRef.html#module-flags-metadata></cite>
for details.)</p>
</div>
</div>
<div class="section" id="executing-ptx">
<h2><a class="toc-backref" href="#id27">Executing PTX</a><a class="headerlink" href="#executing-ptx" title="Permalink to this headline">¶</a></h2>
<p>The most common way to execute PTX assembly on a GPU device is to use the CUDA
Driver API. This API is a low-level interface to the GPU driver and allows for
JIT compilation of PTX code to native GPU machine code.</p>
<p>Initializing the Driver API:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="n">CUdevice</span> <span class="n">device</span><span class="p">;</span>
<span class="n">CUcontext</span> <span class="n">context</span><span class="p">;</span>
<span class="c1">// Initialize the driver API</span>
<span class="n">cuInit</span><span class="p">(</span><span class="mi">0</span><span class="p">);</span>
<span class="c1">// Get a handle to the first compute device</span>
<span class="n">cuDeviceGet</span><span class="p">(</span><span class="o">&</span><span class="n">device</span><span class="p">,</span> <span class="mi">0</span><span class="p">);</span>
<span class="c1">// Create a compute device context</span>
<span class="n">cuCtxCreate</span><span class="p">(</span><span class="o">&</span><span class="n">context</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">device</span><span class="p">);</span>
</pre></div>
</div>
<p>JIT compiling a PTX string to a device binary:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="n">CUmodule</span> <span class="n">module</span><span class="p">;</span>
<span class="n">CUfunction</span> <span class="n">function</span><span class="p">;</span>
<span class="c1">// JIT compile a null-terminated PTX string</span>
<span class="n">cuModuleLoadData</span><span class="p">(</span><span class="o">&</span><span class="n">module</span><span class="p">,</span> <span class="p">(</span><span class="kt">void</span><span class="o">*</span><span class="p">)</span><span class="n">PTXString</span><span class="p">);</span>
<span class="c1">// Get a handle to the "myfunction" kernel function</span>
<span class="n">cuModuleGetFunction</span><span class="p">(</span><span class="o">&</span><span class="n">function</span><span class="p">,</span> <span class="n">module</span><span class="p">,</span> <span class="s">"myfunction"</span><span class="p">);</span>
</pre></div>
</div>
<p>For full examples of executing PTX assembly, please see the <a class="reference external" href="https://developer.nvidia.com/cuda-downloads">CUDA Samples</a> distribution.</p>
</div>
<div class="section" id="common-issues">
<h2><a class="toc-backref" href="#id28">Common Issues</a><a class="headerlink" href="#common-issues" title="Permalink to this headline">¶</a></h2>
<div class="section" id="ptxas-complains-of-undefined-function-nvvm-reflect">
<h3><a class="toc-backref" href="#id29">ptxas complains of undefined function: __nvvm_reflect</a><a class="headerlink" href="#ptxas-complains-of-undefined-function-nvvm-reflect" title="Permalink to this headline">¶</a></h3>
<p>When linking with libdevice, the <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass must be used. See
<a class="reference internal" href="#libdevice"><span class="std std-ref">Linking with Libdevice</span></a> for more information.</p>
</div>
</div>
<div class="section" id="tutorial-a-simple-compute-kernel">
<h2><a class="toc-backref" href="#id30">Tutorial: A Simple Compute Kernel</a><a class="headerlink" href="#tutorial-a-simple-compute-kernel" title="Permalink to this headline">¶</a></h2>
<p>To start, let us take a look at a simple compute kernel written directly in
LLVM IR. The kernel implements vector addition, where each thread computes one
element of the output vector C from the input vectors A and B. To make this
easier, we also assume that only a single CTA (thread block) will be launched,
and that it will be one dimensional.</p>
<div class="section" id="the-kernel">
<h3><a class="toc-backref" href="#id31">The Kernel</a><a class="headerlink" href="#the-kernel" title="Permalink to this headline">¶</a></h3>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
<span class="k">target</span> <span class="k">triple</span> <span class="p">=</span> <span class="s">"nvptx64-nvidia-cuda"</span>
<span class="c">; Intrinsic to read X component of thread ID</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="k">define</span> <span class="kt">void</span> <span class="vg">@kernel</span><span class="p">(</span><span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">)</span> <span class="p">{</span>
<span class="nl">entry:</span>
<span class="c">; What is my ID?</span>
<span class="nv">%id</span> <span class="p">=</span> <span class="k">tail</span> <span class="k">call</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="c">; Compute pointers into A, B, and C</span>
<span class="nv">%ptrA</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrB</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrC</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="c">; Read A, B</span>
<span class="nv">%valA</span> <span class="p">=</span> <span class="k">load</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrA</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="nv">%valB</span> <span class="p">=</span> <span class="k">load</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrB</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="c">; Compute C = A + B</span>
<span class="nv">%valC</span> <span class="p">=</span> <span class="k">fadd</span> <span class="kt">float</span> <span class="nv">%valA</span><span class="p">,</span> <span class="nv">%valB</span>
<span class="c">; Store back to C</span>
<span class="k">store</span> <span class="kt">float</span> <span class="nv">%valC</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrC</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="k">ret</span> <span class="kt">void</span>
<span class="p">}</span>
<span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="kt">void</span> <span class="p">(</span><span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)*</span> <span class="vg">@kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>We can use the LLVM <code class="docutils literal notranslate"><span class="pre">llc</span></code> tool to directly run the NVPTX code generator:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span># llc -mcpu=sm_20 kernel.ll -o kernel.ptx
</pre></div>
</div>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">If you want to generate 32-bit code, change <code class="docutils literal notranslate"><span class="pre">p:64:64:64</span></code> to <code class="docutils literal notranslate"><span class="pre">p:32:32:32</span></code>
in the module data layout string and use <code class="docutils literal notranslate"><span class="pre">nvptx-nvidia-cuda</span></code> as the
target triple.</p>
</div>
<p>The output we get from <code class="docutils literal notranslate"><span class="pre">llc</span></code> (as of LLVM 3.4):</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>//
// Generated by LLVM NVPTX Back-End
//
.version 3.1
.target sm_20
.address_size 64
// .globl kernel
// @kernel
.visible .entry kernel(
.param .u64 kernel_param_0,
.param .u64 kernel_param_1,
.param .u64 kernel_param_2
)
{
.reg .f32 %f<4>;
.reg .s32 %r<2>;
.reg .s64 %rl<8>;
// %bb.0: // %entry
ld.param.u64 %rl1, [kernel_param_0];
mov.u32 %r1, %tid.x;
mul.wide.s32 %rl2, %r1, 4;
add.s64 %rl3, %rl1, %rl2;
ld.param.u64 %rl4, [kernel_param_1];
add.s64 %rl5, %rl4, %rl2;
ld.param.u64 %rl6, [kernel_param_2];
add.s64 %rl7, %rl6, %rl2;
ld.global.f32 %f1, [%rl3];
ld.global.f32 %f2, [%rl5];
add.f32 %f3, %f1, %f2;
st.global.f32 [%rl7], %f3;
ret;
}
</pre></div>
</div>
</div>
<div class="section" id="dissecting-the-kernel">
<h3><a class="toc-backref" href="#id32">Dissecting the Kernel</a><a class="headerlink" href="#dissecting-the-kernel" title="Permalink to this headline">¶</a></h3>
<p>Now let us dissect the LLVM IR that makes up this kernel.</p>
<div class="section" id="data-layout">
<h4><a class="toc-backref" href="#id33">Data Layout</a><a class="headerlink" href="#data-layout" title="Permalink to this headline">¶</a></h4>
<p>The data layout string determines the size in bits of common data types, their
ABI alignment, and their storage size. For NVPTX, you should use one of the
following:</p>
<p>32-bit PTX:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
</pre></div>
</div>
<p>64-bit PTX:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
</pre></div>
</div>
</div>
<div class="section" id="target-intrinsics">
<h4><a class="toc-backref" href="#id34">Target Intrinsics</a><a class="headerlink" href="#target-intrinsics" title="Permalink to this headline">¶</a></h4>
<p>In this example, we use the <code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.tid.x</span></code> intrinsic to
read the X component of the current thread’s ID, which corresponds to a read
of register <code class="docutils literal notranslate"><span class="pre">%tid.x</span></code> in PTX. The NVPTX back-end supports a large set of
intrinsics. A short list is shown below; please see
<code class="docutils literal notranslate"><span class="pre">include/llvm/IR/IntrinsicsNVVM.td</span></code> for the full list.</p>
<table border="1" class="docutils">
<colgroup>
<col width="71%" />
<col width="29%" />
</colgroup>
<thead valign="bottom">
<tr class="row-odd"><th class="head">Intrinsic</th>
<th class="head">CUDA Equivalent</th>
</tr>
</thead>
<tbody valign="top">
<tr class="row-even"><td><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.tid.{x,y,z}</span></code></td>
<td>threadIdx.{x,y,z}</td>
</tr>
<tr class="row-odd"><td><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.ctaid.{x,y,z}</span></code></td>
<td>blockIdx.{x,y,z}</td>
</tr>
<tr class="row-even"><td><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.ntid.{x,y,z}</span></code></td>
<td>blockDim.{x,y,z}</td>
</tr>
<tr class="row-odd"><td><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.nctaid.{x,y,z}</span></code></td>
<td>gridDim.{x,y,z}</td>
</tr>
<tr class="row-even"><td><code class="docutils literal notranslate"><span class="pre">void</span> <span class="pre">@llvm.nvvm.barrier0()</span></code></td>
<td>__syncthreads()</td>
</tr>
</tbody>
</table>
</div>
<div class="section" id="id10">
<h4><a class="toc-backref" href="#id35">Address Spaces</a><a class="headerlink" href="#id10" title="Permalink to this headline">¶</a></h4>
<p>You may have noticed that all of the pointer types in the LLVM IR example had
an explicit address space specifier. What is address space 1? NVIDIA GPU
devices (generally) have four types of memory:</p>
<ul class="simple">
<li>Global: Large, off-chip memory</li>
<li>Shared: Small, on-chip memory shared among all threads in a CTA</li>
<li>Local: Per-thread, private memory</li>
<li>Constant: Read-only memory shared across all threads</li>
</ul>
<p>These different types of memory are represented in LLVM IR as address spaces.
There is also a fifth address space used by the NVPTX code generator that
corresponds to the “generic” address space. This address space can represent
addresses in any other address space (with a few exceptions). This allows
users to write IR functions that can load/store memory using the same
instructions. Intrinsics are provided to convert pointers between the generic
and non-generic address spaces.</p>
<p>See <a class="reference internal" href="#address-spaces"><span class="std std-ref">Address Spaces</span></a> and <a class="reference internal" href="#nvptx-intrinsics"><span class="std std-ref">NVPTX Intrinsics</span></a> for more information.</p>
</div>
<div class="section" id="kernel-metadata">
<h4><a class="toc-backref" href="#id36">Kernel Metadata</a><a class="headerlink" href="#kernel-metadata" title="Permalink to this headline">¶</a></h4>
<p>In PTX, a function can be either a <cite>kernel</cite> function (callable from the host
program), or a <cite>device</cite> function (callable only from GPU code). You can think
of <cite>kernel</cite> functions as entry-points in the GPU program. To mark an LLVM IR
function as a <cite>kernel</cite> function, we make use of special LLVM metadata. The
NVPTX back-end will look for a named metadata node called
<code class="docutils literal notranslate"><span class="pre">nvvm.annotations</span></code>. This named metadata must contain a list of metadata that
describe the IR. For our purposes, we need to declare a metadata node that
assigns the “kernel” attribute to the LLVM IR function that should be emitted
as a PTX <cite>kernel</cite> function. These metadata nodes take the form:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>!{<function ref>, metadata !"kernel", i32 1}
</pre></div>
</div>
<p>For the previous example, we have:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="kt">void</span> <span class="p">(</span><span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)*</span> <span class="vg">@kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>Here, we have a single metadata declaration in <code class="docutils literal notranslate"><span class="pre">nvvm.annotations</span></code>. This
metadata annotates our <code class="docutils literal notranslate"><span class="pre">@kernel</span></code> function with the <code class="docutils literal notranslate"><span class="pre">kernel</span></code> attribute.</p>
</div>
</div>
<div class="section" id="running-the-kernel">
<h3><a class="toc-backref" href="#id37">Running the Kernel</a><a class="headerlink" href="#running-the-kernel" title="Permalink to this headline">¶</a></h3>
<p>Generating PTX from LLVM IR is all well and good, but how do we execute it on
a real GPU device? The CUDA Driver API provides a convenient mechanism for
loading and JIT compiling PTX to a native GPU device, and launching a kernel.
The API is similar to OpenCL. A simple example showing how to load and
execute our vector addition code is shown below. Note that for brevity this
code does not perform much error checking!</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">You can also use the <code class="docutils literal notranslate"><span class="pre">ptxas</span></code> tool provided by the CUDA Toolkit to offline
compile PTX to machine code (SASS) for a specific GPU architecture. Such
binaries can be loaded by the CUDA Driver API in the same way as PTX. This
can be useful for reducing startup time by precompiling the PTX kernels.</p>
</div>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="cp">#include</span> <span class="cpf"><iostream></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf"><fstream></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf"><cassert></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf">"cuda.h"</span><span class="cp"></span>
<span class="kt">void</span> <span class="nf">checkCudaErrors</span><span class="p">(</span><span class="n">CUresult</span> <span class="n">err</span><span class="p">)</span> <span class="p">{</span>
<span class="n">assert</span><span class="p">(</span><span class="n">err</span> <span class="o">==</span> <span class="n">CUDA_SUCCESS</span><span class="p">);</span>
<span class="p">}</span>
<span class="c1">/// main - Program entry point</span>
<span class="kt">int</span> <span class="nf">main</span><span class="p">(</span><span class="kt">int</span> <span class="n">argc</span><span class="p">,</span> <span class="kt">char</span> <span class="o">**</span><span class="n">argv</span><span class="p">)</span> <span class="p">{</span>
<span class="n">CUdevice</span> <span class="n">device</span><span class="p">;</span>
<span class="n">CUmodule</span> <span class="n">cudaModule</span><span class="p">;</span>
<span class="n">CUcontext</span> <span class="n">context</span><span class="p">;</span>
<span class="n">CUfunction</span> <span class="n">function</span><span class="p">;</span>
<span class="n">CUlinkState</span> <span class="n">linker</span><span class="p">;</span>
<span class="kt">int</span> <span class="n">devCount</span><span class="p">;</span>
<span class="c1">// CUDA initialization</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuInit</span><span class="p">(</span><span class="mi">0</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceGetCount</span><span class="p">(</span><span class="o">&</span><span class="n">devCount</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceGet</span><span class="p">(</span><span class="o">&</span><span class="n">device</span><span class="p">,</span> <span class="mi">0</span><span class="p">));</span>
<span class="kt">char</span> <span class="n">name</span><span class="p">[</span><span class="mi">128</span><span class="p">];</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceGetName</span><span class="p">(</span><span class="n">name</span><span class="p">,</span> <span class="mi">128</span><span class="p">,</span> <span class="n">device</span><span class="p">));</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Using CUDA Device [0]: "</span> <span class="o"><<</span> <span class="n">name</span> <span class="o"><<</span> <span class="s">"</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="kt">int</span> <span class="n">devMajor</span><span class="p">,</span> <span class="n">devMinor</span><span class="p">;</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceComputeCapability</span><span class="p">(</span><span class="o">&</span><span class="n">devMajor</span><span class="p">,</span> <span class="o">&</span><span class="n">devMinor</span><span class="p">,</span> <span class="n">device</span><span class="p">));</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Device Compute Capability: "</span>
<span class="o"><<</span> <span class="n">devMajor</span> <span class="o"><<</span> <span class="s">"."</span> <span class="o"><<</span> <span class="n">devMinor</span> <span class="o"><<</span> <span class="s">"</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">if</span> <span class="p">(</span><span class="n">devMajor</span> <span class="o"><</span> <span class="mi">2</span><span class="p">)</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cerr</span> <span class="o"><<</span> <span class="s">"ERROR: Device 0 is not SM 2.0 or greater</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">return</span> <span class="mi">1</span><span class="p">;</span>
<span class="p">}</span>
<span class="n">std</span><span class="o">::</span><span class="n">ifstream</span> <span class="n">t</span><span class="p">(</span><span class="s">"kernel.ptx"</span><span class="p">);</span>
<span class="k">if</span> <span class="p">(</span><span class="o">!</span><span class="n">t</span><span class="p">.</span><span class="n">is_open</span><span class="p">())</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cerr</span> <span class="o"><<</span> <span class="s">"kernel.ptx not found</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">return</span> <span class="mi">1</span><span class="p">;</span>
<span class="p">}</span>
<span class="n">std</span><span class="o">::</span><span class="n">string</span> <span class="n">str</span><span class="p">((</span><span class="n">std</span><span class="o">::</span><span class="n">istreambuf_iterator</span><span class="o"><</span><span class="kt">char</span><span class="o">></span><span class="p">(</span><span class="n">t</span><span class="p">)),</span>
<span class="n">std</span><span class="o">::</span><span class="n">istreambuf_iterator</span><span class="o"><</span><span class="kt">char</span><span class="o">></span><span class="p">());</span>
<span class="c1">// Create driver context</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuCtxCreate</span><span class="p">(</span><span class="o">&</span><span class="n">context</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">device</span><span class="p">));</span>
<span class="c1">// Create module for object</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuModuleLoadDataEx</span><span class="p">(</span><span class="o">&</span><span class="n">cudaModule</span><span class="p">,</span> <span class="n">str</span><span class="p">.</span><span class="n">c_str</span><span class="p">(),</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">));</span>
<span class="c1">// Get kernel function</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuModuleGetFunction</span><span class="p">(</span><span class="o">&</span><span class="n">function</span><span class="p">,</span> <span class="n">cudaModule</span><span class="p">,</span> <span class="s">"kernel"</span><span class="p">));</span>
<span class="c1">// Device data</span>
<span class="n">CUdeviceptr</span> <span class="n">devBufferA</span><span class="p">;</span>
<span class="n">CUdeviceptr</span> <span class="n">devBufferB</span><span class="p">;</span>
<span class="n">CUdeviceptr</span> <span class="n">devBufferC</span><span class="p">;</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemAlloc</span><span class="p">(</span><span class="o">&</span><span class="n">devBufferA</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemAlloc</span><span class="p">(</span><span class="o">&</span><span class="n">devBufferB</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemAlloc</span><span class="p">(</span><span class="o">&</span><span class="n">devBufferC</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="kt">float</span><span class="o">*</span> <span class="n">hostA</span> <span class="o">=</span> <span class="k">new</span> <span class="kt">float</span><span class="p">[</span><span class="mi">16</span><span class="p">];</span>
<span class="kt">float</span><span class="o">*</span> <span class="n">hostB</span> <span class="o">=</span> <span class="k">new</span> <span class="kt">float</span><span class="p">[</span><span class="mi">16</span><span class="p">];</span>
<span class="kt">float</span><span class="o">*</span> <span class="n">hostC</span> <span class="o">=</span> <span class="k">new</span> <span class="kt">float</span><span class="p">[</span><span class="mi">16</span><span class="p">];</span>
<span class="c1">// Populate input</span>
<span class="k">for</span> <span class="p">(</span><span class="kt">unsigned</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="mi">16</span><span class="p">;</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
<span class="n">hostA</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="n">i</span><span class="p">;</span>
<span class="n">hostB</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="kt">float</span><span class="p">)(</span><span class="mi">2</span><span class="o">*</span><span class="n">i</span><span class="p">);</span>
<span class="n">hostC</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="mf">0.0f</span><span class="p">;</span>
<span class="p">}</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemcpyHtoD</span><span class="p">(</span><span class="n">devBufferA</span><span class="p">,</span> <span class="o">&</span><span class="n">hostA</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemcpyHtoD</span><span class="p">(</span><span class="n">devBufferB</span><span class="p">,</span> <span class="o">&</span><span class="n">hostB</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="kt">unsigned</span> <span class="n">blockSizeX</span> <span class="o">=</span> <span class="mi">16</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">blockSizeY</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">blockSizeZ</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">gridSizeX</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">gridSizeY</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">gridSizeZ</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="c1">// Kernel parameters</span>
<span class="kt">void</span> <span class="o">*</span><span class="n">KernelParams</span><span class="p">[]</span> <span class="o">=</span> <span class="p">{</span> <span class="o">&</span><span class="n">devBufferA</span><span class="p">,</span> <span class="o">&</span><span class="n">devBufferB</span><span class="p">,</span> <span class="o">&</span><span class="n">devBufferC</span> <span class="p">};</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Launching kernel</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="c1">// Kernel launch</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuLaunchKernel</span><span class="p">(</span><span class="n">function</span><span class="p">,</span> <span class="n">gridSizeX</span><span class="p">,</span> <span class="n">gridSizeY</span><span class="p">,</span> <span class="n">gridSizeZ</span><span class="p">,</span>
<span class="n">blockSizeX</span><span class="p">,</span> <span class="n">blockSizeY</span><span class="p">,</span> <span class="n">blockSizeZ</span><span class="p">,</span>
<span class="mi">0</span><span class="p">,</span> <span class="nb">NULL</span><span class="p">,</span> <span class="n">KernelParams</span><span class="p">,</span> <span class="nb">NULL</span><span class="p">));</span>
<span class="c1">// Retrieve device data</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemcpyDtoH</span><span class="p">(</span><span class="o">&</span><span class="n">hostC</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">devBufferC</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Results:</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">for</span> <span class="p">(</span><span class="kt">unsigned</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="mi">16</span><span class="p">;</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="n">hostA</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o"><<</span> <span class="s">" + "</span> <span class="o"><<</span> <span class="n">hostB</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o"><<</span> <span class="s">" = "</span> <span class="o"><<</span> <span class="n">hostC</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o"><<</span> <span class="s">"</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="p">}</span>
<span class="c1">// Clean up after ourselves</span>
<span class="k">delete</span> <span class="p">[]</span> <span class="n">hostA</span><span class="p">;</span>
<span class="k">delete</span> <span class="p">[]</span> <span class="n">hostB</span><span class="p">;</span>
<span class="k">delete</span> <span class="p">[]</span> <span class="n">hostC</span><span class="p">;</span>
<span class="c1">// Clean-up</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemFree</span><span class="p">(</span><span class="n">devBufferA</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemFree</span><span class="p">(</span><span class="n">devBufferB</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemFree</span><span class="p">(</span><span class="n">devBufferC</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuModuleUnload</span><span class="p">(</span><span class="n">cudaModule</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuCtxDestroy</span><span class="p">(</span><span class="n">context</span><span class="p">));</span>
<span class="k">return</span> <span class="mi">0</span><span class="p">;</span>
<span class="p">}</span>
</pre></div>
</div>
<p>You will need to link with the CUDA driver and specify the path to cuda.h.</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span># clang++ sample.cpp -o sample -O2 -g -I/usr/local/cuda-5.5/include -lcuda
</pre></div>
</div>
<p>We don’t need to specify a path to <code class="docutils literal notranslate"><span class="pre">libcuda.so</span></code> since this is installed in a
system location by the driver, not the CUDA toolkit.</p>
<p>If everything goes as planned, you should see the following output when
running the compiled program:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>Using CUDA Device [0]: GeForce GTX 680
Device Compute Capability: 3.0
Launching kernel
Results:
0 + 0 = 0
1 + 2 = 3
2 + 4 = 6
3 + 6 = 9
4 + 8 = 12
5 + 10 = 15
6 + 12 = 18
7 + 14 = 21
8 + 16 = 24
9 + 18 = 27
10 + 20 = 30
11 + 22 = 33
12 + 24 = 36
13 + 26 = 39
14 + 28 = 42
15 + 30 = 45
</pre></div>
</div>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">You will likely see a different device identifier based on your hardware</p>
</div>
</div>
</div>
<div class="section" id="tutorial-linking-with-libdevice">
<h2><a class="toc-backref" href="#id38">Tutorial: Linking with Libdevice</a><a class="headerlink" href="#tutorial-linking-with-libdevice" title="Permalink to this headline">¶</a></h2>
<p>In this tutorial, we show a simple example of linking LLVM IR with the
libdevice library. We will use the same kernel as the previous tutorial,
except that we will compute <code class="docutils literal notranslate"><span class="pre">C</span> <span class="pre">=</span> <span class="pre">pow(A,</span> <span class="pre">B)</span></code> instead of <code class="docutils literal notranslate"><span class="pre">C</span> <span class="pre">=</span> <span class="pre">A</span> <span class="pre">+</span> <span class="pre">B</span></code>.
Libdevice provides an <code class="docutils literal notranslate"><span class="pre">__nv_powf</span></code> function that we will use.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
<span class="k">target</span> <span class="k">triple</span> <span class="p">=</span> <span class="s">"nvptx64-nvidia-cuda"</span>
<span class="c">; Intrinsic to read X component of thread ID</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="c">; libdevice function</span>
<span class="k">declare</span> <span class="kt">float</span> <span class="vg">@__nv_powf</span><span class="p">(</span><span class="kt">float</span><span class="p">,</span> <span class="kt">float</span><span class="p">)</span>
<span class="k">define</span> <span class="kt">void</span> <span class="vg">@kernel</span><span class="p">(</span><span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">)</span> <span class="p">{</span>
<span class="nl">entry:</span>
<span class="c">; What is my ID?</span>
<span class="nv">%id</span> <span class="p">=</span> <span class="k">tail</span> <span class="k">call</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="c">; Compute pointers into A, B, and C</span>
<span class="nv">%ptrA</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrB</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrC</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="c">; Read A, B</span>
<span class="nv">%valA</span> <span class="p">=</span> <span class="k">load</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrA</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="nv">%valB</span> <span class="p">=</span> <span class="k">load</span> <span class="kt">float</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrB</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="c">; Compute C = pow(A, B)</span>
<span class="nv">%valC</span> <span class="p">=</span> <span class="k">call</span> <span class="kt">float</span> <span class="vg">@__nv_powf</span><span class="p">(</span><span class="kt">float</span> <span class="nv">%valA</span><span class="p">,</span> <span class="kt">float</span> <span class="nv">%valB</span><span class="p">)</span>
<span class="c">; Store back to C</span>
<span class="k">store</span> <span class="kt">float</span> <span class="nv">%valC</span><span class="p">,</span> <span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrC</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="k">ret</span> <span class="kt">void</span>
<span class="p">}</span>
<span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="kt">void</span> <span class="p">(</span><span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="kt">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)*</span> <span class="vg">@kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>To compile this kernel, we perform the following steps:</p>
<ol class="arabic simple">
<li>Link with libdevice</li>
<li>Internalize all but the public kernel function</li>
<li>Run <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> and set <code class="docutils literal notranslate"><span class="pre">__CUDA_FTZ</span></code> to 0</li>
<li>Optimize the linked module</li>
<li>Codegen the module</li>
</ol>
<p>These steps can be performed by the LLVM <code class="docutils literal notranslate"><span class="pre">llvm-link</span></code>, <code class="docutils literal notranslate"><span class="pre">opt</span></code>, and <code class="docutils literal notranslate"><span class="pre">llc</span></code>
tools. In a complete compiler, these steps can also be performed entirely
programmatically by setting up an appropriate pass configuration (see
<a class="reference internal" href="#libdevice"><span class="std std-ref">Linking with Libdevice</span></a>).</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span># llvm-link t2.bc libdevice.compute_20.10.bc -o t2.linked.bc
# opt -internalize -internalize-public-api-list=kernel -nvvm-reflect-list=__CUDA_FTZ=0 -nvvm-reflect -O3 t2.linked.bc -o t2.opt.bc
# llc -mcpu=sm_20 t2.opt.bc -o t2.ptx
</pre></div>
</div>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">The <code class="docutils literal notranslate"><span class="pre">-nvvm-reflect-list=_CUDA_FTZ=0</span></code> is not strictly required, as any
undefined variables will default to zero. It is shown here for evaluation
purposes.</p>
</div>
<p>This gives us the following PTX (excerpt):</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>//
// Generated by LLVM NVPTX Back-End
//
.version 3.1
.target sm_20
.address_size 64
// .globl kernel
// @kernel
.visible .entry kernel(
.param .u64 kernel_param_0,
.param .u64 kernel_param_1,
.param .u64 kernel_param_2
)
{
.reg .pred %p<30>;
.reg .f32 %f<111>;
.reg .s32 %r<21>;
.reg .s64 %rl<8>;
// %bb.0: // %entry
ld.param.u64 %rl2, [kernel_param_0];
mov.u32 %r3, %tid.x;
ld.param.u64 %rl3, [kernel_param_1];
mul.wide.s32 %rl4, %r3, 4;
add.s64 %rl5, %rl2, %rl4;
ld.param.u64 %rl6, [kernel_param_2];
add.s64 %rl7, %rl3, %rl4;
add.s64 %rl1, %rl6, %rl4;
ld.global.f32 %f1, [%rl5];
ld.global.f32 %f2, [%rl7];
setp.eq.f32 %p1, %f1, 0f3F800000;
setp.eq.f32 %p2, %f2, 0f00000000;
or.pred %p3, %p1, %p2;
@%p3 bra BB0_1;
bra.uni BB0_2;
BB0_1:
mov.f32 %f110, 0f3F800000;
st.global.f32 [%rl1], %f110;
ret;
BB0_2: // %__nv_isnanf.exit.i
abs.f32 %f4, %f1;
setp.gtu.f32 %p4, %f4, 0f7F800000;
@%p4 bra BB0_4;
// %bb.3: // %__nv_isnanf.exit5.i
abs.f32 %f5, %f2;
setp.le.f32 %p5, %f5, 0f7F800000;
@%p5 bra BB0_5;
BB0_4: // %.critedge1.i
add.f32 %f110, %f1, %f2;
st.global.f32 [%rl1], %f110;
ret;
BB0_5: // %__nv_isinff.exit.i
...
BB0_26: // %__nv_truncf.exit.i.i.i.i.i
mul.f32 %f90, %f107, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f91, %f90;
mov.f32 %f92, 0fBF317200;
fma.rn.f32 %f93, %f91, %f92, %f107;
mov.f32 %f94, 0fB5BFBE8E;
fma.rn.f32 %f95, %f91, %f94, %f93;
mul.f32 %f89, %f95, 0f3FB8AA3B;
// inline asm
ex2.approx.ftz.f32 %f88,%f89;
// inline asm
add.f32 %f96, %f91, 0f00000000;
ex2.approx.f32 %f97, %f96;
mul.f32 %f98, %f88, %f97;
setp.lt.f32 %p15, %f107, 0fC2D20000;
selp.f32 %f99, 0f00000000, %f98, %p15;
setp.gt.f32 %p16, %f107, 0f42D20000;
selp.f32 %f110, 0f7F800000, %f99, %p16;
setp.eq.f32 %p17, %f110, 0f7F800000;
@%p17 bra BB0_28;
// %bb.27:
fma.rn.f32 %f110, %f110, %f108, %f110;
BB0_28: // %__internal_accurate_powf.exit.i
setp.lt.f32 %p18, %f1, 0f00000000;
setp.eq.f32 %p19, %f3, 0f3F800000;
and.pred %p20, %p18, %p19;
@!%p20 bra BB0_30;
bra.uni BB0_29;
BB0_29:
mov.b32 %r9, %f110;
xor.b32 %r10, %r9, -2147483648;
mov.b32 %f110, %r10;
BB0_30: // %__nv_powf.exit
st.global.f32 [%rl1], %f110;
ret;
}
</pre></div>
</div>
</div>
</div>
</div>
</div>
<div class="clearer"></div>
</div>
<div class="related" role="navigation" aria-label="related navigation">
<h3>Navigation</h3>
<ul>
<li class="right" style="margin-right: 10px">
<a href="genindex.html" title="General Index"
>index</a></li>
<li class="right" >
<a href="AMDGPUUsage.html" title="User Guide for AMDGPU Backend"
>next</a> |</li>
<li class="right" >
<a href="HowToUseAttributes.html" title="How To Use Attributes"
>previous</a> |</li>
<li><a href="http://llvm.org/">LLVM Home</a> | </li>
<li><a href="index.html">Documentation</a>»</li>
</ul>
</div>
<div class="footer" role="contentinfo">
© Copyright 2003-2020, LLVM Project.
Last updated on 2020-09-07.
Created using <a href="http://sphinx-doc.org/">Sphinx</a> 1.8.4.
</div>
</body>
</html>
distrib
>
Mageia
>
7
>
i586
>
media
>
core-updates
>
by-pkgid
>
d635a8cd705396ade48f1d2b830a115d
>
files
>
2037
