<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Gpu on Alfonso Subiotto</title><link>http://asubiotto.com/tags/gpu/</link><description>Recent content in Gpu on Alfonso Subiotto</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Wed, 24 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="http://asubiotto.com/tags/gpu/index.xml" rel="self" type="application/rss+xml"/><item><title>My Journey Optimizing a CUDA Kernel with Polar Signals</title><link>http://asubiotto.com/writing/optimizing-fsst-cuda/</link><pubDate>Wed, 24 Jun 2026 00:00:00 +0000</pubDate><guid>http://asubiotto.com/writing/optimizing-fsst-cuda/</guid><description>&lt;p&gt;An account of building and optimizing an FSST string decompression CUDA kernel for Vortex, and using the Polar Signals GPU profiler to find out why the first version was slower than the CPU implementation.&lt;/p&gt;&#10;&lt;p&gt;Memory access patterns dominate GPU performance. Buffering decompressed bytes in registers to enable wider, aligned writes instead of storing a byte at a time ended up making decompression 3x faster than comparable solutions.&lt;/p&gt;</description></item></channel></rss>