<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Andrew Page</title>
    <description>The latest articles on DEV Community by Andrew Page (@andrewofc).</description>
    <link>https://dev.to/andrewofc</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4255818%2F07db6134-b466-43e5-8ac1-7f0392582740.png</url>
      <title>DEV Community: Andrew Page</title>
      <link>https://dev.to/andrewofc</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/andrewofc"/>
    <language>en</language>
    <item>
      <title>Optimizing a Reduction Algorithm with AVX, NEON, SVE, and std::experimental::simd</title>
      <dc:creator>Andrew Page</dc:creator>
      <pubDate>Fri, 20 Nov 2026 14:22:14 +0000</pubDate>
      <link>https://dev.to/andrewofc/optimizing-a-reduction-algorithm-with-avx-neon-sve-and-stdexperimentalsimd-7pg</link>
      <guid>https://dev.to/andrewofc/optimizing-a-reduction-algorithm-with-avx-neon-sve-and-stdexperimentalsimd-7pg</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fawbel29vde0oqlajwt9k.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fawbel29vde0oqlajwt9k.png" alt="speedup" width="640" height="480"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;An example of optimizing a reduction algorithm using AVX-512 compiler intrinsics and std::experimental::native_simd types.&lt;br&gt;
&lt;a href="https://github.com/AndrewOfC/vector_reduction" rel="noopener noreferrer"&gt;https://github.com/AndrewOfC/vector_reduction&lt;/a&gt;&lt;br&gt;
The speedup is evaluated against a trivial implementation:&lt;br&gt;
template&lt;br&gt;
T generic_reduction(T *src, size_t len, T initial)&lt;br&gt;
{&lt;br&gt;
    for(size_t i = 0; i &amp;lt; len; i++)&lt;br&gt;
    {&lt;br&gt;
        initial += src[i];&lt;br&gt;
    }&lt;br&gt;
    return initial;&lt;br&gt;
}&lt;br&gt;
And against the C++ standard library std::reduce.&lt;br&gt;
Three approaches are used, but yield similar results in the 'sweet spot' between 1000 and 6000 samples.&lt;br&gt;
Intrinsics and the stdx:native_simd have the same basic structure of:&lt;br&gt;
NOTE: some blocks are not shown for the sake of brevity; full code is in the public GitHub repo.&lt;br&gt;
while( len &amp;gt;= 8 * 8 ) {&lt;br&gt;
    __m512d r000 = _mm512_load_pd(&amp;amp;src[0]);&lt;br&gt;
    __m512d r001 = _mm512_load_pd(&amp;amp;src[8]);&lt;br&gt;
    __m512d r002 = _mm512_load_pd(&amp;amp;src[16]);&lt;br&gt;
    __m512d r003 = _mm512_load_pd(&amp;amp;src[24]);&lt;br&gt;
    __m512d r004 = _mm512_load_pd(&amp;amp;src[32]);&lt;br&gt;
    __m512d r005 = _mm512_load_pd(&amp;amp;src[40]);&lt;br&gt;
    __m512d r006 = _mm512_load_pd(&amp;amp;src[48]);&lt;br&gt;
    __m512d r007 = _mm512_load_pd(&amp;amp;src[56]);&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;r000 = _mm512_add_pd(r000, r004);
r001 = _mm512_add_pd(r001, r005);
r002 = _mm512_add_pd(r002, r006);
r003 = _mm512_add_pd(r003, r007);

r000 = _mm512_add_pd(r000, r002);
r001 = _mm512_add_pd(r001, r003);

r000 = _mm512_add_pd(r000, r001);

initial += _mm512_reduce_add_pd(r000) ;


len -= 8 * 8 ;
src += 8 * 8  ;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;} &lt;br&gt;
// while&lt;br&gt;
/*&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;p&gt;*/&lt;br&gt;
while( len &amp;gt;= 4 * 8 ) {&lt;br&gt;
__m512d r000 = _mm512_load_pd(&amp;amp;src[0]);&lt;br&gt;
__m512d r001 = _mm512_load_pd(&amp;amp;src[8]);&lt;br&gt;
__m512d r002 = _mm512_load_pd(&amp;amp;src[16]);&lt;br&gt;
__m512d r003 = _mm512_load_pd(&amp;amp;src[24]);&lt;/p&gt;

&lt;p&gt;r000 = _mm512_add_pd(r000, r002);&lt;br&gt;
r001 = _mm512_add_pd(r001, r003);&lt;/p&gt;

&lt;p&gt;r000 = _mm512_add_pd(r000, r001);&lt;/p&gt;

&lt;p&gt;initial += _mm512_reduce_add_pd(r000) ;&lt;/p&gt;

&lt;p&gt;len -= 4 * 8 ;&lt;br&gt;
src += 4 * 8  ;&lt;br&gt;
} &lt;br&gt;
// while&lt;br&gt;
The stdx::native_simd allows the use of templates for other datatypes.&lt;br&gt;
template&lt;br&gt;
using vector_t = std::experimental::native_simd;&lt;/p&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;while( len &amp;gt;= 8 * sz ) {&lt;br&gt;
    vector_t r00, r01, r02, r03, r04, r05, r06, r07 ;&lt;br&gt;
    r00.copy_from(&amp;amp;src[0 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r01.copy_from(&amp;amp;src[1 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r02.copy_from(&amp;amp;src[2 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r03.copy_from(&amp;amp;src[3 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r04.copy_from(&amp;amp;src[4 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r05.copy_from(&amp;amp;src[5 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r06.copy_from(&amp;amp;src[6 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r07.copy_from(&amp;amp;src[7 * sz], stdx::vector_aligned) ;&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;r00 += r04 ;
r01 += r05 ;
r02 += r06 ;
r03 += r07 ;

r00 += r02 ;
r01 += r03 ;

r00 += r01 ;

initial += stdx::reduce(r00) ;

len -= 8 * sz ;
src += 8 * sz ;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;} // while&lt;/p&gt;

&lt;p&gt;while( len &amp;gt;= 4 * sz ) {&lt;br&gt;
    vector_t r00, r01, r02, r03 ;&lt;br&gt;
    r00.copy_from(&amp;amp;src[0 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r01.copy_from(&amp;amp;src[1 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r02.copy_from(&amp;amp;src[2 * sz], stdx::vector_aligned) ;&lt;br&gt;
    r03.copy_from(&amp;amp;src[3 * sz], stdx::vector_aligned) ;&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;r00 += r02 ;
r01 += r03 ;

r00 += r01 ;

initial += stdx::reduce(r00) ;

len -= 4 * sz ;
src += 4 * sz ;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;} // while&lt;br&gt;
There's a third approach that uses a recursive C++ template expressing the reduction algorithm.   Suprisingly this formed as well as the intrinsic implementation and templated simd, where the trivial loop did not.&lt;br&gt;
template&lt;br&gt;
inline static vector_t reduce(vector_t a, T *srcB) &lt;br&gt;
// terminal case for recursion&lt;br&gt;
{&lt;br&gt;
    vector_t b ;&lt;br&gt;
    b.copy_from(srcB, stdx::&lt;br&gt;
vector_aligned&lt;br&gt;
);&lt;br&gt;
    return a + b ;&lt;br&gt;
}&lt;/p&gt;

&lt;p&gt;inline static auto reduce(auto a, auto srcB, auto... args) &lt;br&gt;
// consume argument and recurse&lt;br&gt;
{&lt;br&gt;
    return reduce(reduce(a, srcB), args...) ;&lt;br&gt;
}&lt;/p&gt;

&lt;p&gt;vinitial = 0 ;&lt;br&gt;
while( len &amp;gt;= 8 * sz ) {&lt;br&gt;
    vinitial += reduce(vector_t(0), &amp;amp;src[0 * sz], &amp;amp;src[1 * sz], &amp;amp;src[2 * sz], &amp;amp;src[3 * sz], &amp;amp;src[4 * sz], &amp;amp;src[5 * sz], &amp;amp;src[6 * sz], &amp;amp;src[7 * sz]) ;&lt;br&gt;
    len -= 8 * sz ;&lt;br&gt;
    src += 8 * sz ;&lt;br&gt;
} &lt;br&gt;
// while&lt;/p&gt;

&lt;p&gt;initial += stdx::reduce(vinitial) ;&lt;br&gt;
/*&lt;br&gt;
 *&lt;br&gt;
 */&lt;br&gt;
vinitial = 0 ;&lt;br&gt;
while( len &amp;gt;= 4 * sz ) {&lt;br&gt;
    vinitial += reduce(vector_t(0), &amp;amp;src[0 * sz], &amp;amp;src[1 * sz], &amp;amp;src[2 * sz], &amp;amp;src[3 * sz]) ;&lt;br&gt;
    len -= 4 * sz ;&lt;br&gt;
    src += 4 * sz ;&lt;br&gt;
} &lt;br&gt;
// while&lt;/p&gt;

&lt;p&gt;initial += stdx::reduce(vinitial) ;&lt;/p&gt;

</description>
      <category>simd</category>
      <category>cpp</category>
      <category>std</category>
    </item>
  </channel>
</rss>
