Fighting with computers

 Many vector graphic algorithms have a cost proportional to the number of segments of the shapes involved. Thus, many different polygon simplification algorithms have been developed over the years. The main idea is to represent the original shape, but reduce the total number of edges of the shape's contour.  A favourite of mine is the Douglas-Peucker algorithm. It uses an error parameter that allows the user to tune the acceptable error introduced by the simplification. That works well in many scenarios, but as you may guess, it does not work well for some of my needs. Free form 2D nesting packs groups of parts inside bins, in my case, those bins are rectangles. The rules are that any two parts cannot overlap. Nesting libraries will determine where to place each part inside a bin and what rotation to apply to the part. The number of parts and the number of edges per part have a big influence on the time required to perform the nesting operation. So, being able to simplify the...

 While it was nice to be able to get a faster 2D nesting code, its efficiency was far from perfect. So, looking at the nesting results, you could see how it could have been done much better in certain cases.  And what we usually do when we overcome a hurdle is to look for the next one. In my case, I realized one of the heuristics of the code was to place small parts inside the unused space of the bounding box of the larger parts. That is ok, but what could be better is to use all the free space of a bin once the first placement of large parts is over.  This time, I leaned on another Chinese AI tool, called GLM-4.5 from the  z.AI company.  I explained the above idea in the chat, and I waited for the magic to happen. I used Roo Code within Visual Studio Code, accessing GLM-4.5 API. A few minutes later, I had a newer version of the library. I was pleasantly surprised when I saw a significant improvement in its packing ability, proving that the new heuristic was a ...

As I mentioned in my previous entry, I migrated a library from Java to C++ to make it easier to use from Python. I tested its performance on both Linux and OSX, and it was on par with the Java version. Sample file S266.txt would take around 45 seconds on my computer using the Java version, and the C++ version needed thirty-something seconds on a M1 MacBook Pro running OSX, a bit less than that on an i9-13900 server running Linux. From that, I could not claim there was a huge difference in performance compared to the original Java runtime. But I wanted to make a fair comparison, and I wanted to face the trouble of building the library for Windows. It was not a challenge, but it took a while. Let me summarize all the steps I needed to cover: Install Microsoft's Visual Studio (Community Edition) Install Microsoft's vcpkg (C++ package manager) Install CMake. Install dependencies (tbb, gtest, boost, pybind11) Clone the project repository . Build the project. It took me like an hour...

Today, we embarked on an ambitious software engineering journey: migrating a sophisticated 2D bin packing program from its original Java implementation into a high-performance C++ library, complete with Python bindings for ease of use. The goal was to retain 100% of the original complex packing logic while unlocking the raw speed of C++. The heart of this migration involved replacing Java's powerful java.awt.geom.Area class. We chose the world-class Boost.Geometry library as its C++ counterpart, meticulously re-implementing every geometric operation, from simple transformations to complex polygon intersections and subtractions. This formed the foundation upon which we rebuilt the original program's advanced packing heuristics. We successfully ported a multi-stage packing strategy, including bounding-box placement, a "drop" simulation for gravity-fed placement, and iterative compression algorithms. However, real-world testing with complex, high-vertex-count polygons r...

Today, I set out to configure a remote access VPN for my small office. My first choice was L2TP/IPsec with StrongSwan and xl2tpd , mainly because it’s supported natively by Windows. It seemed like a solid option; secure and well-documented. But several hours later, after wading through IKE version mismatches, arcane config options, and ambiguous Windows error codes, I hit a wall. Despite correctly setting up IPsec and L2TP, Windows insisted on trying IKEv2, and StrongSwan repeatedly rejected the connections. Exhausted by endless logs and incompatible defaults, I decided to abandon the legacy stack and try something different: OpenVPN . In under ten minutes, using the widely respected OpenVPN install script by Angristan , I had a working VPN server. The script handled everything—certificates, firewall rules, user config—and generated a ready-to-go .ovpn file. I dropped it into OpenVPN Connect on Windows , hit "Connect", and just like that, it worked . Sometimes the best t...

If you’ve worked with 3D meshes in Python, you’ve probably heard the name PyMeshLab . It’s the “Python face” of the legendary open-source MeshLab—famous for its ability to clean up, repair, and process even the messiest 3D models. And truly, PyMeshLab is a fantastic tool : powerful, open, scriptable, and (for most operations) robust. But—let’s be honest—getting started with PyMeshLab sometimes feels a bit like entering a quirky, magical forest: full of treasures, but you might trip on some roots along the way . Here’s my little “cautionary tale” and field guide to the weirdness you may encounter, and how to laugh it off while getting things done. 1. “Is That a Snake or a Camel?!”: The Naming Circus You’ll quickly notice that the naming of filters and methods in PyMeshLab is... creative . Sometimes you call a mesh processing function by a method like compute_normal_for_point_sets() . Other times, the same operation is invoked using apply_filter('compute_normal_for_poin...

 I have recently started learning Chen-style Tai Chi. I like the activity and my instructor, but as I am finding out, it requires more than just a few minutes of practice to improve performance. I have started looking for sources for additional help. Recorded videos are very helpful as they can be a tool for practising at home, but the computer nerd in me wonders if there is a better way.  What I came up with is to create 3D animations in a similar way to a video recording, but with the added benefit that the user could set the point of view of the scene at will. So I started to look around to see how that could be achieved, as my knowledge of the field is second to none.  I learned that motion capture could be used to obtain the motion pattern from a human performer, and that piece of information could be later used to animate a 3D character in a virtual environment. However, the way this is done sometimes requires expensive equipment that may also interfere with the ran...

 I have been looking at the motion problem beyond the snap to realize that its derivative, crackle, was not continuous in my case .  So I went ahead and looked for another expression; this time, I used chatGPT, and I asked for a sinusoidal function that: Be continuous. Started and ended at zero. Its definite integral was zero. Consisted of two positive pulses, with a negative pulse in between. It was suggested this function could do the work: j ( t ) = sin ( π t/T ​ ) sin ( 3 π t/T ​ ) So, after some back and forth with the integration constants, I got the signals of the chart you can see on the right.  How to use it? Turning that into a look-up table may be the more practical approach, depending on the computing power. In my application, using stepper motors with a step/direction interface requires the determination of the precise moment for each step pulse. That calculation might be done before starting the movement or, iteratively, calculating the time of the next step...

I am retiring from academia, so I had to burn some remaining cash and settled on building a new computer. After the latest troubles with Intel processors, I decided to look around for an AMD processor and motherboard. I wanted to have a generous amount of RAM as I have been playing a lot with different IA projects. So, I chose a 4nm Ryzen 7 9700X processor (8 cores, 40 MB cache), and I trusted the good reviews of the inexpensive Asus PRIME B650-PLUS , which I checked supported 128 HB of DDR5 RAM.  The build was easy, and I was quickly running Windows 11 and Ubuntu 24.04, but something fishy was going on: My DDR5 DRAM was only running at 3600 MHz, which the manufacturer marketed as 5200 MHz. I contacted the seller's support, and I was told that while the board could do 128 GB and 5200MHz memory, it could not do both at the same time, and the BIOS was choosing a more conservative frequency for the memory if I was using the four DRAM SIMMs (which I was). I could have taken the loss ...

Java Network Programming uses the DatagramPacket class as the data representation for UDP datagrams. Whether you want to send or receive it, it will be a DatagramPacket . In a recent exam, we asked our students to create a simple UDP echo server.   The logic of such a server is that it waits endlessly until it receives a UDP datagram and then sends it back to the same address (IP and port number) from which it came. Nothing complicated, but the devil is in the details, and some students wrote a program that would create a new DatagramPacket for each datagram received. It is not wrong, as it works, but it is a bit wasteful, as it places more stress on memory allocation and garbage collection than needed.  However, I was surprised when chatGPT commented on the code I got from DeepSeek-R1 as a sample solution, where a call to setLength() was made after each iteration. It did not make sense to me, but chatGPT went on for quite a while, reasoning why it was needed. Eventually...

QUESTION: I f the center of a hexagon is located at a (0,0) coordinate, which are the coordinates of each vertex if the length of the side of the hexagon is 10 mm? Analyzing and reasoning phase Okay, so the user wants to find the coordinates of each vertex of a hexagon centered at (0,0) with a side length of 10 mm. Let me think about how to approach this. First, I remember that a regular hexagon can be inscribed in a circle. Each vertex is equally spaced around the circumference. The key here is relating the side length to the radius of the circumscribed circle. Wait, is the radius the same as the side length? Hmm, no, I think that's only true for a hexagon with a side length equal to the radius. Wait, actually, in a regular hexagon, the side length is equal to the radius of the circumscribed circle. Let me confirm that. In a regular hexagon, all sides are equal, and all internal angles are 120 degrees. The radius of the circumscribed circle (distance from center to a vertex) is ...