150
|
1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
|
|
2 "http://www.w3.org/TR/html4/strict.dtd">
|
|
3 <!-- Material used from: HTML 4.01 specs: http://www.w3.org/TR/html401/ -->
|
|
4 <html>
|
|
5 <head>
|
|
6 <META http-equiv="Content-Type" content="text/html; charset=ISO-8859-1">
|
|
7 <title>Polly - Polyhedral optimizations for LLVM</title>
|
|
8 <link type="text/css" rel="stylesheet" href="menu.css">
|
|
9 <link type="text/css" rel="stylesheet" href="content.css">
|
|
10 <script src="video-js/video.js" type="text/javascript" charset="utf-8"></script>
|
|
11 <script type="text/javascript">
|
|
12 VideoJS.setupAllWhenReady();
|
|
13 </script>
|
|
14
|
|
15 <!-- Include the VideoJS Stylesheet -->
|
|
16 <link rel="stylesheet" href="video-js/video-js.css" type="text/css" media="screen" title="Video JS">
|
|
17 </head>
|
|
18 <body>
|
|
19 <div id="box">
|
|
20 <!--#include virtual="menu.html.incl"-->
|
|
21 <div id="content">
|
|
22 <!--*********************************************************************-->
|
|
23 <h1>About Polly</h1>
|
|
24 <!--*********************************************************************-->
|
|
25
|
|
26 <p> Polly is a high-level loop and data-locality optimizer and optimization
|
|
27 infrastructure for LLVM. It uses an abstract mathematical representation based
|
|
28 on integer polyhedra to analyze and optimize the memory access pattern of a
|
|
29 program. We currently perform classical loop transformations, especially
|
|
30 tiling and loop fusion to improve data-locality. Polly can also exploit
|
252
|
31 OpenMP level parallelism, expose SIMDization opportunities.</p>
|
150
|
32
|
|
33 For many users, however, it's not the existing optimizations in Polly that are
|
|
34 of most interest, but the new analyses and optimizations enabled by the Polly
|
|
35 infrastructure. At
|
|
36 <a href="https://polyhedral.info">polyhedral.info</a> you can get an idea of
|
|
37 what has already been done and what is possible in the context of polyhedral
|
|
38 compilation.
|
|
39
|
|
40 <!--=====================================================================-->
|
|
41 <h2>News</h2>
|
|
42 <!--=====================================================================-->
|
|
43
|
|
44 <table id="news">
|
|
45 <tr><td><b>2017</b></td></tr>
|
|
46 <tr><td width="120"><p>September</p></td>
|
|
47 <td>
|
|
48 <h4>High-Performance Generalized Matrix Multiplication</h4>
|
|
49 Polly automatically detects and optimizes generalized matrix
|
|
50 multiplication, the computation C ← α ⊗ C ⊕ β
|
|
51 ⊗ A ⊗ B, where A, B, and C are three appropriately sized
|
|
52 matrices, ⊕ and ⊗ operations are originating from the
|
|
53 corresponding matrix semiring, and α and β are constants, and
|
|
54 beta is not equal to zero. It allows to obtain the highly optimized form
|
|
55 structured similar to the expert implementation of GEMM that can be found
|
|
56 in GotoBLAS and its successors.
|
|
57 <h4>The performance evaluation of GEMM</h4>
|
|
58 <img src="images/GEMM_double.png" /><br />
|
|
59 </td>
|
|
60 <tr><td><b>2017</b></td></tr>
|
|
61 <tr><td width="120"><p>January</p></td>
|
|
62 <td>
|
|
63 <a href="http://impact.gforge.inria.fr/impact2017">IMPACT 2017</a> program
|
|
64 announced. Join IMPACT 2017 on January 23rd in Stockholm <a
|
|
65 href="https://www.hipeac.net/2017/stockholm/">@HiPEAC'17</a>.
|
|
66 </td>
|
|
67 </tr>
|
|
68 <tr><td><b>2016</b></td></tr>
|
|
69 <tr><td width="120"><p>August</p></td>
|
|
70 <td>
|
|
71 <a href="http://impact.gforge.inria.fr/impact2017">IMPACT 2017</a> the 7th
|
|
72 International Workshop on Polyhedral Compilation Techniques will take place
|
|
73 at January 23-25, 2017 together with HiPEAC 2017 in Stockholm, Sweden. It is
|
|
74 a great opportunity to discuss and present work on Polyhedral Compilation,
|
|
75 including work on Polly.
|
|
76 </td>
|
|
77 </tr>
|
|
78 <tr><td width="120"><p>April</p></td>
|
|
79 <td>
|
|
80 A source checkout that contains Polly now provides Polly functionality
|
|
81 by default in clang/opt/bugpoint without the need to load an additional
|
|
82 module.
|
|
83 </td>
|
|
84 </tr>
|
|
85 <tr><td><b>2015</b></td></tr>
|
|
86 <tr><td width="120"><p>July</p></td>
|
|
87 <td>
|
|
88 <h4>AST Generation Paper published in TOPLAS</h4>
|
|
89 The July issue of TOPLAS contains a 50 page discussion of the AST
|
|
90 generation techniques used in Polly. This discussion gives not only an
|
|
91 in-depth description of how we (re)generate an imperative AST from our
|
|
92 polyhedral based mathematical program description, but also gives
|
|
93 interesting insights about:
|
|
94 <ul>
|
|
95 <li><b>Schedule trees:</b> A tree-based mathematical program description
|
|
96 that enables us to perform loop transformations on an abstract level,
|
|
97 while issues like the generation of the correct loop structure and loop
|
|
98 bounds will be taken care of by our AST generator.
|
|
99 <li><b>Polyhedral unrolling:</b> We discuss techniques that allow the
|
|
100 unrolling of non-trivial loops in the context of parameteric loop bounds,
|
|
101 complex tile shapes and conditionally executed statements. Such unrolling
|
|
102 support enables the generation of predicated code e.g. in the context of
|
|
103 GPGPU computing.
|
|
104 <li><b>Isolation for full/partial tile separation:</b> We discuss native
|
|
105 support for handling full/partial tile separation and -- in general --
|
|
106 native support for isolation of boundary cases to enable smooth code
|
|
107 generation for core computations.
|
|
108 <li><b>AST generation with modulo constraints:</b> We discuss how modulo
|
|
109 mappings are lowered to efficient C/LLVM code.
|
|
110 <li><b>User-defined constraint sets for run-time checks</b> We discuss how
|
|
111 arbitrary sets of constraints can be used to automatically create run-time
|
236
|
112 checks that ensure a set of constraints actually hold. This feature is
|
150
|
113 very useful to verify at run-time various assumptions that have been taken
|
|
114 program optimization.
|
|
115 </ul>
|
|
116
|
|
117 <a href="https://www.grosser.es#pub-polyhedral-AST-generation">
|
|
118 <em>Polyhedral AST generation is more than scanning polyhedra</em></a><br />
|
|
119 Tobias Grosser, Sven Verdoolaege, Albert Cohen<br />
|
|
120 ACM Transations on Programming Languages and Systems (TOPLAS), 37(4),
|
|
121 July 2015
|
|
122
|
|
123 <br>
|
|
124 <br>
|
|
125 <br>
|
|
126 <br>
|
|
127 </td>
|
|
128 </tr>
|
|
129 <tr><td width="120"><p>February</p></td>
|
|
130 <td>
|
|
131 <h4>Polly allows now non-affine subregions</h4>
|
|
132 Data-dependent or floating point conditionals inside a SCoP can now be
|
|
133 overapproximated in order to increase the applicability on general purpose
|
|
134 code.
|
|
135 </td>
|
|
136 </tr>
|
|
137 <tr><td><b>2014</b></td></tr>
|
|
138 <tr><td width="120"><p>August</p></td>
|
|
139 <td>
|
|
140 <h4>Polly drops the support of ScopLib and the external optimizer PoCC</h4>
|
|
141 The support for ScopLib as an exchange format has been removed as recent
|
|
142 versions of clan, candl and pluto all support the OpenScop exchange format.
|
|
143
|
236
|
144 The support of the external optimizer PoCC has been dropped in favor of the
|
150
|
145 isl optimizer (default) and the still available pluto support.
|
|
146 </td>
|
|
147 </tr>
|
|
148 <tr><td><b>2014</b></td></tr>
|
|
149 <tr><td width="120"><p>June</p></td>
|
|
150 <td>
|
|
151 <h4>Polly can be built without GPL licensed software</h4> After Sebastian
|
|
152 Pop's
|
|
153 and David Peixotto's (both Qualcomm) recent <a
|
|
154 href="https://repo.or.cz/w/isl.git/commit/60703e3ee89b9d5d4d1afb6a3f611292c0884574">commit</a>
|
|
155 to isl, isl's latest development version can be built with imath instead of
|
236
|
156 GMP. With both CLooG and gmp having become optional, the last obligatory
|
150
|
157 dependency to GPL licensed software has been removed. Now Polly only depends
|
|
158 on isl (and the included imath), which are both MIT licensed.
|
|
159 </td>
|
|
160 </tr>
|
|
161 <tr><td width="120"><p>April</p></td>
|
|
162 <td>
|
|
163 <h4>Polly Phone Call - 23April</h4>
|
|
164 We had a polly phone call about delinearizing array accesses (II)<a
|
|
165 href="https://docs.google.com/document/d/1IZewI8Up0iEkCNIPr6gVtwJxF7RV6KmXkdwOBM_Q5Cs/edit?usp=sharing ">Meeting notes</a> are available online.
|
|
166 <h4>Polly Phone Call - 17th April</h4>
|
|
167 We had a polly phone call about delinearizing array accesses <a
|
|
168 href="https://docs.google.com/document/d/14d3ehkH2MsvBdqsEOSYjysH0Ztyzb75Lp843hnxh2IA/edit?usp=sharing">Meeting notes</a> are available online.
|
|
169 <h4>Polly Phone Call - 10th April</h4>
|
|
170 We had a polly phone call. <a
|
|
171 href="https://docs.google.com/document/d/12W-qZjiZGEQ_lVrob4OzvKJI3EooonC-ap1b9f9KCUE/edit?usp=sharing">Meeting notes</a> are available online.
|
|
172 <h4>Polly Phone Call - 4th April</h4>
|
|
173 We had a polly phone call. <a
|
|
174 href="https://drive.google.com/folderview?id=0B7OMOXTgCYIUWkpJbWVJcW04ams&usp=sharing">Meeting notes</a> are available online.
|
|
175 </td>
|
|
176 </tr>
|
|
177 <tr><td width="120"><p>March</p></td>
|
|
178 <td>
|
|
179 <h4>Static link Polly into tools</h4> Polly can now be configured with 'cmake
|
|
180 -D LINK_POLLY_INTO_TOOLS:Bool=ON' to be statically linked in the tools (opt,
|
|
181 bugpoint, and clang.) This makes it easier to use polly without having to load
|
|
182 a shared library, and it also reduces the complexity of compiling Polly on
|
|
183 Windows.
|
|
184 </td>
|
|
185 </tr>
|
|
186 <tr><td width="120"><p>February</p></td>
|
|
187 <td>
|
|
188 <h4>Polly presentation at FOSDEM 2014</h4> Polly was <a
|
|
189 href="https://fosdem.org/2014/schedule/event/polly/">presented</a> at the
|
|
190 FOSDEM LLVM developer's meeting.
|
|
191 <h4>New LLVM test-suite buildbots</h4>
|
|
192 The set of <a href="http://lab.llvm.org:8011/console?category=polly">Polly
|
|
193 buildbots</a> has been extended. We now have 16 new blades that track
|
|
194 correctness and performance when compiling the LLVM test-suite. For now five
|
|
195 of them are used to provide <a
|
|
196 href="https://llvm.org/perf/db_default/v4/nts/22463">fine granularity
|
|
197 reports</a> (almost per-commit)
|
|
198 for 'clang -O3' (no polly). We also have six machines that track different
|
|
199 configurations of polly.
|
|
200 </td>
|
|
201 </tr>
|
|
202 <tr><td width="120"><p>January</p></td>
|
|
203 <td>
|
|
204 <h4>islplot released</h4>
|
|
205 <a href="https://github.com/tobig/islplot">islplot</a> is a library that
|
|
206 generates illustrations of integer sets and maps. It relies on <a
|
|
207 href="https://repo.or.cz/w/isl.git">isl</a> to model the integer sets and uses the <a
|
|
208 href="https://pypi.python.org/pypi/islpy">islpy</a> Python bindings to access
|
|
209 them. Plotting is performed with <a
|
|
210 href="https://matplotlib.org">matplotlib</a>. The following <a
|
|
211 href="https://nbviewer.ipython.org/github/tobig/islplot/blob/master/notebooks/islplot-examples.ipynb">
|
|
212 Examples</a> show its use.
|
|
213 </td>
|
|
214 </tr>
|
|
215 <tr><td><b>2013</b></td></tr>
|
|
216 <tr><td width="120"><p>November</p></td>
|
|
217 <td>
|
|
218 <h4>Loop optimization BoF at upcoming LLVM conference</h4>
|
|
219 At the upcoming <a href="https://llvm.org/devmtg/2013-11/#bof5">LLVM conference
|
|
220 </a> there will be a loop optimization BoF discussing Polly and other high
|
|
221 level loop optimizers.
|
|
222 </td>
|
|
223 </tr>
|
|
224 <tr><td width="120"><p>October</p></td>
|
|
225 <td>
|
|
226 <h4>Automatic code coverage and static analysis tests</h4>
|
|
227 Sylvestre Ledru set up automatic tests for <a
|
|
228 href="https://llvm.org/reports/coverage/">code coverage</a> and
|
|
229 <a href="https://llvm.org/reports/scan-build/">static analysis</a>
|
|
230 which run at least once a day and which include results for Polly.
|
|
231 <h4>Move to CLooG 0.18.1 and isl 0.12.1</h4>
|
|
232 With the move to an isl 0.12 version Polly can be compiled without the
|
|
233 need to link directly to GMP (if isl is used for code generation). Currently
|
|
234 isl is still internally using GMP, but private patches exist to also remove
|
|
235 this dependency. Without the use of GMP, a <b>GPL free</b> version of Polly
|
|
236 is possible.
|
|
237 </td></tr>
|
|
238
|
|
239 <tr><td><b>2012</b></td></tr>
|
|
240 <tr><td width="120"><p>December</p></td>
|
|
241 <td>
|
|
242 <h4> New publication in the PPL Journal
|
|
243 </h4>
|
|
244
|
|
245 We published a journal version of the Polly paper named
|
|
246 <em>
|
|
247 Polly - Performing polyhedral optimizations on a low-level intermediate
|
|
248 representation</em> in the Parallel Processing Letters 2012.
|
|
249 </td></tr>
|
|
250 <tr><td width="120"><p>September</p></td>
|
|
251 <td>
|
|
252 <h4>Experimental support for the <b>new isl code generator</b></h4>
|
|
253 The code generator can be parameterized on a fine-grained
|
|
254 level. It gives direct control for example over unrolling, the amount of
|
|
255 control overhead and the code size. It can also be used to
|
|
256 create loops to handle border conditions or to perform full-partial tile
|
|
257 separation.<br />
|
|
258 We also relicensed isl under the <b>MIT license</b>. This means, with the
|
|
259 exception of GMP (LGPL), there is no other (L)GPL licensed software used in
|
|
260 Polly. The
|
|
261 use of GMP is limited to a well defined interface. Replacing it with
|
|
262 a BSD licensed replacement is a tractable engineering project we would
|
|
263 be very interested in. For more information about isl see the <a
|
|
264 href="http://www.kotnet.org/~skimo/isl/manual.pdf">isl manual</a>.
|
|
265 </p>
|
|
266 </td></tr>
|
|
267 <tr><td width="120"><p>July</p></td>
|
|
268 <td>
|
|
269 <p> Polly can now be directly linked to the <a
|
|
270 href="http://pluto-compiler.sourceforge.net/">Pluto optimizer</a>. We were
|
|
271 already able to perform Pluto-like optimizations with Polly, as a similar
|
|
272 algorithm was added to isl half a year ago. However, being able to directly
|
|
273 compare with the original implementation will not only bring in competition in
|
|
274 the optimizer field. It will also allow new experiments with a cutting edge
|
|
275 research tool.<br \>
|
|
276 This support was on of the outcomes of the 1-day Polly workshop and the
|
|
277 following week of joint work at IISC Bangalore and in cooperation with
|
|
278 AMD India.
|
|
279 </td></tr>
|
|
280 <td>
|
|
281 </td></tr>
|
|
282 <tr><td width="120"><p>February</p></td>
|
|
283 <td>
|
|
284 <p>Polly is an official LLVM project, reachable at <a
|
|
285 href="https://polly.llvm.org">https://polly.llvm.org</a></p>
|
|
286 </td></tr>
|
|
287 <tr><td width="120"><p>January</p></td>
|
|
288 <td>
|
|
289 <p>Improved support for the isl scheduling optimizer</p>
|
|
290 Polly can now automatically optimize all <a
|
|
291 href="https://web.cse.ohio-state.edu/~pouchet.2/software/polybench/">polybench
|
|
292 2.0</a> kernels without the help of
|
|
293 an external optimizer. The compile time is reasonable and we can show
|
|
294 notable speedups for various kernels.
|
|
295 </td></tr>
|
|
296
|
|
297 <tr>
|
|
298 <tr><td><b><br/>2011</b></td></tr>
|
|
299 <tr><td width="120"><p>November</p></td>
|
|
300 <td>
|
|
301 <p>
|
|
302 Talk at the <a href="https://llvm.org/devmtg/2011-11/">
|
|
303 LLVM Developer Meeting 2011</a></p>
|
|
304 New SCEV parser<br>
|
|
305 (Allows parameters in array subscript and max/signextend)
|
|
306 </td></tr>
|
|
307
|
|
308 <tr>
|
|
309 <td><p>October</p></td>
|
|
310 <td>
|
|
311 <p>Polly can use the isl schedule optimizer<br>
|
|
312 (The optimizer is similar to the one in Pluto, but it is part of isl)
|
|
313 </p>
|
|
314 </td></tr>
|
|
315
|
|
316 <tr>
|
|
317 <td><p>August</p></td>
|
|
318 <td>
|
|
319 <p>
|
|
320 <a href="example_load_Polly_into_clang.html">Use Polly as
|
|
321 clang plugin</a></p>
|
|
322 </td>
|
|
323 </tr>
|
|
324
|
|
325 <tr>
|
|
326 <td><p>July</p></td>
|
|
327 <td>
|
|
328 <p> Polly builder as part of the <a
|
|
329 href="http://lab.llvm.org:8011/console">LLVM Buildbots</a>
|
|
330 </p>
|
|
331 </td>
|
|
332 </tr>
|
|
333
|
|
334 <tr>
|
|
335 <td><p>June</p></td>
|
|
336 <td>
|
|
337 <p><a href="https://www.grosser.es">Tobias</a> is founded for
|
|
338 three years by a <a
|
|
339 href="https://ai.google/research/outreach/phd-fellowship/recipients/?category=2011">
|
|
340 Google Europe Fellowship in Efficient Computing</a>.
|
|
341 </p>
|
|
342 </td>
|
|
343 </tr>
|
|
344
|
|
345 <tr>
|
|
346 <td><p>May </p></td>
|
|
347 <td><p><a href="https://www.grosser.es">Tobias</a>' diploma thesis and
|
|
348 Raghesh's master thesis. See our <a
|
|
349 href="publications.html">list of publications</a>.</p></td>
|
|
350 </tr>
|
|
351
|
|
352 <tr>
|
|
353 <td><p>April</p></td>
|
|
354 <td><p>Polly moves to the LLVM infrastructure (svn, bugtracker)</p></td>
|
|
355 </tr>
|
|
356
|
|
357 <tr>
|
|
358 <td><p>March</p></td>
|
|
359 <td><p>Presentation at <a
|
|
360 href="http://impact2011.inrialpes.fr/">CGO/IMPACT</a></p>
|
|
361 <p>Polly can compile
|
|
362 polybench 2.0 with vectorization and OpenMP code generation</p>
|
|
363 </td>
|
|
364 </tr>
|
|
365 <tr>
|
|
366 <td><p>Februar</p></td>
|
|
367 <td><p>pollycc - a script to automatically compile with
|
|
368 polyhedral optimizations </p></td>
|
|
369 </tr>
|
|
370
|
|
371 <tr>
|
|
372 <td><p> Januar</p></td>
|
|
373 <td><p> Basic OpenMP support, Alias analysis integration,
|
|
374 Pluto/POCC support </p></td>
|
|
375 </tr>
|
|
376
|
|
377 <tr><td><b><br>2010</b></td></tr>
|
|
378 <tr>
|
236
|
379 <td><p> December </p></td>
|
150
|
380 <td><p>Basic vectorization support </p></td>
|
|
381 </tr>
|
|
382
|
|
383 <tr>
|
|
384 <td><p> November </p></td>
|
|
385 <td><p>Talk at the <a
|
|
386 href="https://llvm.org/devmtg/2010-11/">LLVM Developer Meeting</a> </p></td>
|
|
387 </tr>
|
|
388
|
|
389 <tr>
|
|
390 <td><p>October</p></td>
|
|
391 <td><p>Dependency analysis </p>
|
|
392 <p>Finished Phase 1 - Get something working </p>
|
|
393 <p>Support scalar dependences and sequential SCoPs </p>
|
|
394 </td>
|
|
395 </tr>
|
|
396
|
|
397 <tr>
|
|
398 <td><p>August</p></td>
|
|
399 <td><p>RegionInfo pass committed to LLVM</p>
|
|
400 <p>llvm-test suite compiles </p>
|
|
401 </td>
|
|
402 </tr>
|
|
403
|
|
404 <tr>
|
|
405 <td><p>July</p></td>
|
|
406 <td><p>Code generation works for normal SCoPs. </p></td>
|
|
407 </tr>
|
|
408
|
|
409 <tr>
|
|
410 <td><p>May</p></td>
|
|
411 <td><p>The CLooG AST can be parsed.</p>
|
|
412 </td>
|
|
413 </tr>
|
|
414
|
|
415 <tr>
|
|
416 <td><p>April</p></td>
|
|
417 <td><p>SCoPs can automatically be detected. </p></td>
|
|
418 </tr>
|
|
419
|
|
420 <tr>
|
|
421 <td><p>March</p></td>
|
|
422 <td><p>The RegionInfo framework is almost completed. </p></td>
|
|
423 </tr>
|
|
424
|
|
425 <tr>
|
|
426 <td><p>February</p></td>
|
|
427 <td><p>Translate a simple loop to Polly-IR and regenerate a loop structure
|
|
428 with CLooG works. </p>
|
|
429 <p>ISL and CLooG are integrated. </p></td>
|
|
430 </tr>
|
|
431
|
|
432 </tr>
|
|
433
|
|
434 <tr>
|
|
435 <td><p>January</p></td>
|
|
436 <td><p>The RegionInfo pass is finished. </p></td>
|
|
437 </tr>
|
|
438
|
|
439 <tr><td><b><br>2009</b></td></tr>
|
|
440 <tr>
|
|
441 <td><p>End of the year</p></td>
|
|
442 <td><p>Work on the infrastructure started. </p></td>
|
|
443 </tr>
|
|
444 </table>
|
|
445 </ul>
|
|
446 </div>
|
|
447 </div>
|
|
448 </body>
|
|
449 </html>
|