https://vinayakdsci.github.io/posts/ Recent content in Posts on vinayakdsci Hugo en-us vinayakdev.sci@gmail.com vinayakdev.sci@gmail.com Mon, 01 Dec 2025 00:00:00 +0000 https://vinayakdsci.github.io/posts/2025-12-01/ Mon, 01 Dec 2025 00:00:00 +0000vinayakdev.sci@gmail.com https://vinayakdsci.github.io/posts/2025-12-01/ <h3 id="why-a-separate-post-to-introduce-the-library">Why a separate post to introduce the library?</h3> <p>In the previous post in the BLAS series, we looked at how <code>NumPy</code> significantly outperformed a naive <code>IJK</code> ordered three-loop C++ implementation on the same matrix sizes. The post was an announcement for both the blog series and the library we will implement as the series progresses. Over the past two weeks, I spent some time setting up the library and plugging in our naive C++ implementation into the codebase.</p> https://vinayakdsci.github.io/posts/2025-11-17/ Mon, 17 Nov 2025 00:00:00 +0000vinayakdev.sci@gmail.com https://vinayakdsci.github.io/posts/2025-11-17/ <h3 id="the-motivation">The Motivation</h3> <p>For years, I’ve relied on highly optimized BLAS libraries like OpenBLAS, BLIS, and MKL without ever really understanding the machinery inside them.</p> <p>Recently, while working on some multithreaded code, I found myself benchmarking NumPy operations against naive C++ implementations. The results, though not surprising, were drastic enough to trigger a deeper curiosity about how NumPy achieves such speed. We all know NumPy is backed by C, and often links against BLAS libraries like OpenBLAS, but that only raises a more interesting question:</p>