Skip to content

Commit 37d306f

Browse files
SylvainCorlaySylvainCorlay
committed
iterate on slides
1 parent c8dc5f7 commit 37d306f

File tree

2 files changed

+120
-12
lines changed

2 files changed

+120
-12
lines changed

‎2017-07-14-Scipy2017-xtensor/src/index.html

+120-12
Original file line numberDiff line numberDiff line change
@@ -664,6 +664,57 @@ <h2>What is xtensor?</h2>
664664
</div>
665665
</div>
666666
</section>
667+
668+
</section>
669+
<section>
670+
<section>
671+
<div class="splitting"></div>
672+
<p>Iteration</p>
673+
<div style="margin-top: 2%;">
674+
<div class="left-panel">
675+
Row-major iteration over the array <code>for x in np.nditer(a)</code>
676+
</div>
677+
<div class="right-panel">
678+
<pre class="panel"><code class="cpp panel" data-trim>
679+
for(auto it=a.xbegin(); it!=a.xend(); ++it)
680+
</code></pre>
681+
</div>
682+
</div>
683+
<div style="margin-top: 2%;">
684+
<div class="left-panel fragment">
685+
Iterating over <code>a</code> with a prescribed broadcasting shape
686+
</div>
687+
<div class="right-panel fragment">
688+
<pre class="panel"><code class="cpp panel" data-trim>
689+
a.xbegin({3, 4})
690+
a.xend({3, 4})
691+
</code></pre>
692+
</div>
693+
</div>
694+
<div style="margin-top:2%;">
695+
<div class="left-panel fragment">
696+
Iterating over <code>a</code> in a column-major fashion
697+
</div>
698+
<div class="right-panel fragment">
699+
<pre class="panel"><code class="cpp panel" data-trim>
700+
a.template xbegin&lt;layout_type::column_major&gt;()
701+
a.template xend&lt;layout_type::column_major&gt;()
702+
</code></pre>
703+
</div>
704+
</div>
705+
<div style="margin-top:2%;">
706+
<div class="left-panel fragment">
707+
Iterating over <code>a</code> in a column-major fashion with a prescribed broadcasting shape
708+
</div>
709+
<div class="right-panel fragment">
710+
<pre class="panel"><code class="cpp panel" data-trim>
711+
a.template xbegin&lt;layout_type::column_major&gt;({3, 4})
712+
a.template xend&lt;layout_type::column_major&gt;({3, 4})
713+
</code></pre>
714+
</div>
715+
</div>
716+
</section>
717+
</section>
667718
</section>
668719
<section>
669720
<section>
@@ -1083,12 +1134,27 @@ <h3>A simple R extension</h3>
10831134
<li>Packaging boilerplate (setup.py, build.jl)</li>
10841135
</ul>
10851136
</section>
1137+
1138+
<section>
1139+
<img alt="xtensor-robot" src="robot.png" width="20%"/>
1140+
<p>Scientific computing in a polyglot world with xtensor</p>
1141+
<ul>
1142+
<li>Write a native extension once, operating on xtensor expressions</li>
1143+
<li>Get free bindings to R, Julia and Python</li>
1144+
<li>For more details about the best practices, see
1145+
<p>
1146+
<a href="http://quantstack.net/c++/2017/05/30/polyglot-scientific-computing-with-xtensor.html">http://quantstack.net/c++/2017/05/30/polyglot-scientific-computing-with-xtensor.html</a>
1147+
</p>
1148+
</li>
1149+
</ul>
1150+
</section>
1151+
10861152
<section>
10871153
<p>Bindings with BLAS libraries</p>
10881154
<img alt="xtensor-blas" src="xtensor-blas.svg" width="55%"/>
1089-
<p>BLAS-based implementation of <code>numpy.linalg</code> module are implemented in xtensor-blas.</p>
1155+
<p>BLAS-based implementation of <code>numpy.linalg</code></p>
10901156
<ul>
1091-
<li>ISO results with numpy.linalg is main goal</li>
1157+
<li>ISO results with <code>numpy.linalg</code> is main goal</li>
10921158
<li>Seeking adoption by Python to C++ compilers (Pythran, Jet)</li>
10931159
<li>Works with any BLAS implementation (openblas, mkl, netlib)</li>
10941160
<li>See the <a href="http://xtensor.readthedocs.io/en/latest/numpy.html">numpy to xtensor cheat sheet</a></li>
@@ -1097,17 +1163,17 @@ <h3>A simple R extension</h3>
10971163

10981164
<section>
10991165
<p>SIMD acceleration</p>
1100-
<img alt="xsimd" src="xsimd.svg" width="55%"/>
1166+
<img alt="xsimd" src="xsimd.svg" width="35%"/>
11011167
<ul>
11021168
<li>Unified high-level API to AVX, SSE, Neon with arithmetic operations</li>
11031169
<li>Accelerated mathematical functions</li>
11041170
<li>Aligned memory allocation</li>
11051171
</ul>
11061172
</section>
11071173
<section>
1108-
<p>Assuming that AVX is supported</p>
1174+
<p>xsimd example (Assuming that AVX is supported)</p>
11091175
<pre class="panel">
1110-
<code class="julia">
1176+
<code class="cpp">
11111177
#include &lt;iostream&gt;
11121178
#include "xsimd/xsimd.hpp"
11131179

@@ -1122,11 +1188,11 @@ <h3>A simple R extension</h3>
11221188
return 0;
11231189
}
11241190
</code>
1125-
<pre>
1191+
</pre>
11261192
</section>
11271193
<section>
1128-
<p>Accelerating loops</p>
1129-
<p>This applies to processing including basic math functions from cmath.</p>
1194+
<p>xsimd: accelerating loops</p>
1195+
<p>(This applies to processing including basic math functions from cmath.)</p>
11301196
<pre class="panel">
11311197
<code class="julia">
11321198
#include &lt;cstddef&gt;
@@ -1156,17 +1222,18 @@ <h3>A simple R extension</h3>
11561222
}
11571223
}
11581224
</code>
1159-
<pre>
1225+
</pre>
11601226
</section>
11611227

11621228
<section>
11631229
<p>What is coming?</p>
11641230
<div style="margin-top: 5%; margin-bottom: 10%;">
11651231
<ul>
1166-
<li>Pluging xsimd into xtensor for SIMD acceleration</li>
1167-
<li class="fragment" style="margin-top: 5%;">Bindings with Apache Arrow</li>
1232+
<li>Plugging xsimd into xtensor for SIMD acceleration</li>
1233+
<li class="fragment" style="margin-top: 5%;">Bindings with Apache Arrow, HDF5, FITS, NetCDF</li>
11681234
<li class="fragment" style="margin-top: 5%;">GPU acceleration</li>
1169-
<li class="fragment" style="margin-top: 5%;">ODBC-databases backed expressions</li>
1235+
<li class="fragment" style="margin-top: 5%;">Expressions backed wit ODBC-databases</li>
1236+
<li class="fragment" style="margin-top: 5%;">Arrays with labeled dimensions</li>
11701237
<li class="fragment" style="margin-top: 5%;">World domination</li>
11711238
</ul>
11721239
</div>
@@ -1240,6 +1307,47 @@ <h3>A simple R extension</h3>
12401307
</div>
12411308
</div>
12421309
</section>
1310+
<section>
1311+
<h2>The End</h2>
1312+
</section>
1313+
<section>
1314+
<h3>Aliasing (Appendix)</h3>
1315+
<div style="width: 70%; margin-left: auto; margin-right: auto; text-align: left;">
1316+
<p>In an assignment, aliasing occurs when the LHS refers to the same memory as a member of the RHS expression</p>
1317+
<pre class="panel">
1318+
<code class="cpp">
1319+
a = b + c.dot(a);
1320+
</code>
1321+
</pre>
1322+
<p>By default, xtensor will solve that issue by assigning the result to a temporary variable that is then <em>moved</em> to the left-hand side.</p>
1323+
<p>If the user <em>knows</em> that no aliasing accurs in an assignment, we can save this extra-allocation with the <code>noalias</code> function.</p>
1324+
<pre class="panel">
1325+
<code class="cpp">
1326+
noalias(a) = b + c.dot(d);
1327+
</code>
1328+
</pre>
1329+
</section>
1330+
<section>
1331+
<h3>Static and dynamic dimensionality (Appendix)</h3>
1332+
<div style="width: 70%; margin-left: auto; margin-right: auto; text-align: left; font-size: 30px;">
1333+
<p>
1334+
The xtensor library provides too container types:
1335+
</p>
1336+
<ul>
1337+
<li><code>xarray<double></code></li>
1338+
<li><code>xtensor<double, N></code></li>
1339+
</ul>
1340+
<p>The latter encodes the number of dimensions in the type:</p>
1341+
<ul>
1342+
<li>Stack-allocated shapes and strides.</li>
1343+
<li>Faster iteration.</li>
1344+
</ul>
1345+
<p>What about expressions?</p>
1346+
<ul>
1347+
<li>The compile-time <code>promote_shape</code> mechanism selects the optimal shape type for an expression at compile time.</li>
1348+
</ul>
1349+
</div>
1350+
</section>
12431351
</div>
12441352
</div>
12451353

‎2017-07-14-Scipy2017-xtensor/src/robot.png

124 KB
Loading

0 commit comments

Comments
 (0)