<!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>llvm-exegesis - LLVM Machine Instruction Benchmark — 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="llvm-pdbutil - PDB File forensics and diagnostics" href="llvm-pdbutil.html" />
<link rel="prev" title="llvm-build - LLVM Project Build Utility" href="llvm-build.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="llvm-pdbutil.html" title="llvm-pdbutil - PDB File forensics and diagnostics"
accesskey="N">next</a> |</li>
<li class="right" >
<a href="llvm-build.html" title="llvm-build - LLVM Project Build Utility"
accesskey="P">previous</a> |</li>
<li><a href="http://llvm.org/">LLVM Home</a> | </li>
<li><a href="../index.html">Documentation</a>»</li>
<li class="nav-item nav-item-1"><a href="index.html" accesskey="U">LLVM Command Guide</a> »</li>
</ul>
</div>
<div class="document">
<div class="documentwrapper">
<div class="body" role="main">
<div class="section" id="llvm-exegesis-llvm-machine-instruction-benchmark">
<h1>llvm-exegesis - LLVM Machine Instruction Benchmark<a class="headerlink" href="#llvm-exegesis-llvm-machine-instruction-benchmark" title="Permalink to this headline">¶</a></h1>
<div class="section" id="synopsis">
<h2>SYNOPSIS<a class="headerlink" href="#synopsis" title="Permalink to this headline">¶</a></h2>
<p><strong class="program">llvm-exegesis</strong> [<em>options</em>]</p>
</div>
<div class="section" id="description">
<h2>DESCRIPTION<a class="headerlink" href="#description" title="Permalink to this headline">¶</a></h2>
<p><strong class="program">llvm-exegesis</strong> is a benchmarking tool that uses information available
in LLVM to measure host machine instruction characteristics like latency or port
decomposition.</p>
<p>Given an LLVM opcode name and a benchmarking mode, <strong class="program">llvm-exegesis</strong>
generates a code snippet that makes execution as serial (resp. as parallel) as
possible so that we can measure the latency (resp. uop decomposition) of the
instruction.
The code snippet is jitted and executed on the host subtarget. The time taken
(resp. resource usage) is measured using hardware performance counters. The
result is printed out as YAML to the standard output.</p>
<p>The main goal of this tool is to automatically (in)validate the LLVM’s TableDef
scheduling models. To that end, we also provide analysis of the results.</p>
<p><strong class="program">llvm-exegesis</strong> can also benchmark arbitrary user-provided code
snippets.</p>
</div>
<div class="section" id="example-1-benchmarking-instructions">
<h2>EXAMPLE 1: benchmarking instructions<a class="headerlink" href="#example-1-benchmarking-instructions" title="Permalink to this headline">¶</a></h2>
<p>Assume you have an X86-64 machine. To measure the latency of a single
instruction, run:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>$ llvm-exegesis -mode<span class="o">=</span>latency -opcode-name<span class="o">=</span>ADD64rr
</pre></div>
</div>
<p>Measuring the uop decomposition of an instruction works similarly:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>$ llvm-exegesis -mode<span class="o">=</span>uops -opcode-name<span class="o">=</span>ADD64rr
</pre></div>
</div>
<p>The output is a YAML document (the default is to write to stdout, but you can
redirect the output to a file using <cite>-benchmarks-file</cite>):</p>
<div class="highlight-none notranslate"><div class="highlight"><pre><span></span>---
key:
opcode_name: ADD64rr
mode: latency
config: ''
cpu_name: haswell
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 10000
measurements:
- { key: latency, value: 1.0058, debug_string: '' }
error: ''
info: 'explicit self cycles, selecting one aliasing configuration.
Snippet:
ADD64rr R8, R8, R10
'
...
</pre></div>
</div>
<p>To measure the latency of all instructions for the host architecture, run:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span><span class="ch">#!/bin/bash</span>
<span class="nb">readonly</span> <span class="nv">INSTRUCTIONS</span><span class="o">=</span><span class="k">$(($(</span>grep INSTRUCTION_LIST_END build/lib/Target/X86/X86GenInstrInfo.inc <span class="p">|</span> cut -f2 -d<span class="o">=</span><span class="k">)</span> <span class="o">-</span> <span class="m">1</span><span class="k">))</span>
<span class="k">for</span> INSTRUCTION in <span class="k">$(</span>seq <span class="m">1</span> <span class="si">${</span><span class="nv">INSTRUCTIONS</span><span class="si">}</span><span class="k">)</span><span class="p">;</span>
<span class="k">do</span>
./build/bin/llvm-exegesis -mode<span class="o">=</span>latency -opcode-index<span class="o">=</span><span class="si">${</span><span class="nv">INSTRUCTION</span><span class="si">}</span> <span class="p">|</span> sed -n <span class="s1">'/---/,$p'</span>
<span class="k">done</span>
</pre></div>
</div>
<p>FIXME: Provide an <strong class="program">llvm-exegesis</strong> option to test all instructions.</p>
</div>
<div class="section" id="example-2-benchmarking-a-custom-code-snippet">
<h2>EXAMPLE 2: benchmarking a custom code snippet<a class="headerlink" href="#example-2-benchmarking-a-custom-code-snippet" title="Permalink to this headline">¶</a></h2>
<p>To measure the latency/uops of a custom piece of code, you can specify the
<cite>snippets-file</cite> option (<cite>-</cite> reads from standard input).</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>$ <span class="nb">echo</span> <span class="s2">"vzeroupper"</span> <span class="p">|</span> llvm-exegesis -mode<span class="o">=</span>uops -snippets-file<span class="o">=</span>-
</pre></div>
</div>
<p>Real-life code snippets typically depend on registers or memory.
<strong class="program">llvm-exegesis</strong> checks the liveliness of registers (i.e. any register
use has a corresponding def or is a “live in”). If your code depends on the
value of some registers, you have two options:</p>
<ul class="simple">
<li>Mark the register as requiring a definition. <strong class="program">llvm-exegesis</strong> will
automatically assign a value to the register. This can be done using the
directive <cite>LLVM-EXEGESIS-DEFREG <reg name> <hex_value></cite>, where <cite><hex_value></cite>
is a bit pattern used to fill <cite><reg_name></cite>. If <cite><hex_value></cite> is smaller than
the register width, it will be sign-extended.</li>
<li>Mark the register as a “live in”. <strong class="program">llvm-exegesis</strong> will benchmark
using whatever value was in this registers on entry. This can be done using
the directive <cite>LLVM-EXEGESIS-LIVEIN <reg name></cite>.</li>
</ul>
<p>For example, the following code snippet depends on the values of XMM1 (which
will be set by the tool) and the memory buffer passed in RDI (live in).</p>
<div class="highlight-none notranslate"><div class="highlight"><pre><span></span># LLVM-EXEGESIS-LIVEIN RDI
# LLVM-EXEGESIS-DEFREG XMM1 42
vmulps (%rdi), %xmm1, %xmm2
vhaddps %xmm2, %xmm2, %xmm3
addq $0x10, %rdi
</pre></div>
</div>
</div>
<div class="section" id="example-3-analysis">
<h2>EXAMPLE 3: analysis<a class="headerlink" href="#example-3-analysis" title="Permalink to this headline">¶</a></h2>
<p>Assuming you have a set of benchmarked instructions (either latency or uops) as
YAML in file <cite>/tmp/benchmarks.yaml</cite>, you can analyze the results using the
following command:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span> $ llvm-exegesis -mode<span class="o">=</span>analysis <span class="se">\</span>
-benchmarks-file<span class="o">=</span>/tmp/benchmarks.yaml <span class="se">\</span>
-analysis-clusters-output-file<span class="o">=</span>/tmp/clusters.csv <span class="se">\</span>
-analysis-inconsistencies-output-file<span class="o">=</span>/tmp/inconsistencies.html
</pre></div>
</div>
<p>This will group the instructions into clusters with the same performance
characteristics. The clusters will be written out to <cite>/tmp/clusters.csv</cite> in the
following format:</p>
<div class="highlight-none notranslate"><div class="highlight"><pre><span></span>cluster_id,opcode_name,config,sched_class
...
2,ADD32ri8_DB,,WriteALU,1.00
2,ADD32ri_DB,,WriteALU,1.01
2,ADD32rr,,WriteALU,1.01
2,ADD32rr_DB,,WriteALU,1.00
2,ADD32rr_REV,,WriteALU,1.00
2,ADD64i32,,WriteALU,1.01
2,ADD64ri32,,WriteALU,1.01
2,MOVSX64rr32,,BSWAP32r_BSWAP64r_MOVSX64rr32,1.00
2,VPADDQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.02
2,VPSUBQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.01
2,ADD64ri8,,WriteALU,1.00
2,SETBr,,WriteSETCC,1.01
...
</pre></div>
</div>
<p><strong class="program">llvm-exegesis</strong> will also analyze the clusters to point out
inconsistencies in the scheduling information. The output is an html file. For
example, <cite>/tmp/inconsistencies.html</cite> will contain messages like the following :</p>
<img alt="../_images/llvm-exegesis-analysis.png" class="align-center" src="../_images/llvm-exegesis-analysis.png" />
<p>Note that the scheduling class names will be resolved only when
<strong class="program">llvm-exegesis</strong> is compiled in debug mode, else only the class id will
be shown. This does not invalidate any of the analysis results though.</p>
</div>
<div class="section" id="options">
<h2>OPTIONS<a class="headerlink" href="#options" title="Permalink to this headline">¶</a></h2>
<dl class="option">
<dt id="cmdoption-help">
<code class="descname">-help</code><code class="descclassname"></code><a class="headerlink" href="#cmdoption-help" title="Permalink to this definition">¶</a></dt>
<dd><p>Print a summary of command line options.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-opcode-index">
<code class="descname">-opcode-index</code><code class="descclassname">=<LLVM opcode index></code><a class="headerlink" href="#cmdoption-opcode-index" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the opcode to measure, by index. See example 1 for details.
Either <cite>opcode-index</cite>, <cite>opcode-name</cite> or <cite>snippets-file</cite> must be set.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-opcode-name">
<code class="descname">-opcode-name</code><code class="descclassname">=<opcode name 1>,<opcode name 2>,...</code><a class="headerlink" href="#cmdoption-opcode-name" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the opcode to measure, by name. Several opcodes can be specified as
a comma-separated list. See example 1 for details.
Either <cite>opcode-index</cite>, <cite>opcode-name</cite> or <cite>snippets-file</cite> must be set.</p>
<dl class="option">
<dt id="cmdoption-snippets-file">
<code class="descname">-snippets-file</code><code class="descclassname">=<filename></code><a class="headerlink" href="#cmdoption-snippets-file" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the custom code snippet to measure. See example 2 for details.
Either <cite>opcode-index</cite>, <cite>opcode-name</cite> or <cite>snippets-file</cite> must be set.</p>
</dd></dl>
</dd></dl>
<dl class="option">
<dt id="cmdoption-mode">
<code class="descname">-mode</code><code class="descclassname">=[latency|uops|analysis]</code><a class="headerlink" href="#cmdoption-mode" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the run mode.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-num-repetitions">
<code class="descname">-num-repetitions</code><code class="descclassname">=<Number of repetition></code><a class="headerlink" href="#cmdoption-num-repetitions" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the number of repetitions of the asm snippet.
Higher values lead to more accurate measurements but lengthen the benchmark.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-benchmarks-file">
<code class="descname">-benchmarks-file</code><code class="descclassname">=</path/to/file></code><a class="headerlink" href="#cmdoption-benchmarks-file" title="Permalink to this definition">¶</a></dt>
<dd><p>File to read (<cite>analysis</cite> mode) or write (<cite>latency</cite>/<cite>uops</cite> modes) benchmark
results. “-” uses stdin/stdout.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-analysis-clusters-output-file">
<code class="descname">-analysis-clusters-output-file</code><code class="descclassname">=</path/to/file></code><a class="headerlink" href="#cmdoption-analysis-clusters-output-file" title="Permalink to this definition">¶</a></dt>
<dd><p>If provided, write the analysis clusters as CSV to this file. “-” prints to
stdout.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-analysis-inconsistencies-output-file">
<code class="descname">-analysis-inconsistencies-output-file</code><code class="descclassname">=</path/to/file></code><a class="headerlink" href="#cmdoption-analysis-inconsistencies-output-file" title="Permalink to this definition">¶</a></dt>
<dd><p>If non-empty, write inconsistencies found during analysis to this file. <cite>-</cite>
prints to stdout.</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-analysis-numpoints">
<code class="descname">-analysis-numpoints</code><code class="descclassname">=<dbscan numPoints parameter></code><a class="headerlink" href="#cmdoption-analysis-numpoints" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the numPoints parameters to be used for DBSCAN clustering
(<cite>analysis</cite> mode).</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-analysis-espilon">
<code class="descname">-analysis-espilon</code><code class="descclassname">=<dbscan epsilon parameter></code><a class="headerlink" href="#cmdoption-analysis-espilon" title="Permalink to this definition">¶</a></dt>
<dd><p>Specify the numPoints parameters to be used for DBSCAN clustering
(<cite>analysis</cite> mode).</p>
</dd></dl>
<dl class="option">
<dt id="cmdoption-ignore-invalid-sched-class">
<code class="descname">-ignore-invalid-sched-class</code><code class="descclassname">=false</code><a class="headerlink" href="#cmdoption-ignore-invalid-sched-class" title="Permalink to this definition">¶</a></dt>
<dd><p>If set, ignore instructions that do not have a sched class (class idx = 0).</p>
<dl class="option">
<dt id="cmdoption-mcpu">
<code class="descname">-mcpu</code><code class="descclassname">=<cpu name></code><a class="headerlink" href="#cmdoption-mcpu" title="Permalink to this definition">¶</a></dt>
<dd><p>If set, measure the cpu characteristics using the counters for this CPU. This
is useful when creating new sched models (the host CPU is unknown to LLVM).</p>
</dd></dl>
</dd></dl>
</div>
<div class="section" id="exit-status">
<h2>EXIT STATUS<a class="headerlink" href="#exit-status" title="Permalink to this headline">¶</a></h2>
<p><strong class="program">llvm-exegesis</strong> returns 0 on success. Otherwise, an error message is
printed to standard error, and the tool returns a non 0 value.</p>
</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="llvm-pdbutil.html" title="llvm-pdbutil - PDB File forensics and diagnostics"
>next</a> |</li>
<li class="right" >
<a href="llvm-build.html" title="llvm-build - LLVM Project Build Utility"
>previous</a> |</li>
<li><a href="http://llvm.org/">LLVM Home</a> | </li>
<li><a href="../index.html">Documentation</a>»</li>
<li class="nav-item nav-item-1"><a href="index.html" >LLVM Command Guide</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
>
1978
