
I'm thrilled to introduce **uα** (pronounced "yu-ah"), an experimental assembler project I've been working on that challenges a fundamental assumption in low-level programming: that assembly language must be architecture-specific.

## The Big Idea

What if you could write assembly code once and compile it for any processor architecture? Not by hiding the hardware behind layers of abstraction, but by defining a truly unified assembly language that translates cleanly to native machine code across wildly different platforms.

That's exactly what uα does. Write your program using the Unified Assembly instruction set, and compile it to run on:

- Intel x86-64 (your desktop PC)
- 32-bit x86 (legacy systems)
- ARM processors (mobile devices, Raspberry Pi)
- ARM64 / Apple Silicon (modern Macs)
- RISC-V (emerging open-source hardware)
- 8051 microcontrollers (embedded systems)

The same source file. Different architectures. No code changes required.

## Version 26: "Awesome Ada"

The current release is **Version 26**, codenamed "Awesome Ada" in honor of Ada Lovelace (1815–1852), widely considered the first computer programmer. She wrote the first algorithm intended for Charles Babbage's Analytical Engine—a fitting inspiration for a language that brings modern structure to low-level programming.

## Why This Matters

Assembly language has always been the ultimate form of control—direct access to the hardware with zero overhead. But that power comes at a price: each processor family has its own instruction set, syntax quirks, and conventions. Want to port your code from x86 to ARM? Start rewriting.

uα bridges that gap. It provides:

**Bare Metal Control** — You still work with registers, memory addresses, and CPU instructions. Nothing is hidden. When you write `MOV R0, R1`, you're moving data between registers, just like traditional assembly.

**Modern Ergonomics** — But you also get namespaces, imports, standard libraries, and conditional compilation. The `@IMPORT` system lets you include functionality like string handling or math operations that feel high-level but execute as pure, cycle-accurate assembly.

**Write Once, Run Everywhere** — The Minimum Viable Instruction Set (MVIS) ensures that 37 core instructions work identically across all architectures. The compiler handles the architectural differences behind the scenes.

## A Taste of the Code

Here's a simple "Hello, World!" in uα:

```assembly
@IMPORT std_io
    LDS  R0, "Hello, World!\n"
    CALL std_io.print
    HLT
```

Compile it for any architecture:

```bash
ua hello.ua -arch x86 -sys linux -o hello        # Linux on x86-64
ua hello.ua -arch arm64 -sys macos -o hello     # Apple Silicon
ua hello.ua -arch mcs51 -o hello.bin           # 8051 microcontroller
```

The same code runs on a modern MacBook Pro and a decades-old 8-bit embedded chip.

## The Philosophy: No Compromises

The project is guided by a few core principles:

**Zero Dependencies** — Written in pure C99, uα requires no external linkers, package managers, or runtime environments. One compiler invocation is all you need to build it. It takes raw text and directly emits executable `.exe`, `.elf`, or `.bin` files.

**Transparency over Magic** — There are no hidden safety nets or runtime checks. If an 8-bit register overflows at 255, it overflows exactly as the hardware dictates. The developer is in absolute command.

**Self-Hosting Vision** — The ultimate goal? A compiler written entirely in uα, capable of compiling itself. A language only truly matures when it can bootstrap itself.

## What's Included

uα ships with a complete standard library—written entirely in uα itself:

- `std_io` — console input/output
- `std_string` — string operations
- `std_math` — mathematical utilities
- `std_array` — fixed-size arrays
- `std_vector` — dynamic vectors
- `std_iostream` — file I/O operations

The compiler includes:
- **37 portable instructions** covering data movement, arithmetic, logic, control flow, and system calls
- **14 architecture-specific opcodes** for specialized operations (like x86's `CPUID` or ARM's `WFI`)
- **Conditional compilation** with `@IF_ARCH` and `@IF_SYS` directives
- **Five output modes**: raw binary, Windows PE, Linux ELF, macOS Mach-O, and JIT execution
- **Six CPU backends** supporting everything from modern 64-bit processors to vintage 8-bit chips

## Try It Yourself

The entire compiler is under 3,000 lines of C99 code. Build it with a single GCC command, then start experimenting. Write some arithmetic, compile it for different architectures, compare the machine code output. See for yourself how the same logical operations translate across wildly different hardware.

uα is about exploring an idea: that we can have both the raw power of assembly and the convenience of portable code. That we don't have to choose between control and accessibility.

## What's Next?

This is an experimental project, and there's still much to explore. The instruction set continues to evolve, more standard library functions are being added, and the optimization pipeline is getting smarter.

But the core vision is clear: one assembly language, understood everywhere, giving developers bare-metal control without tying them to a single architecture.

If you've ever wondered what it would be like to program "close to the metal" but without being locked into one platform, uα is worth a look.

---

*uα is released under the GNU General Public License v3. Full documentation, examples, and source code are available in the project repository.*



uα - Unified Assembler: One Assembly Language for Every Architecture


When I first released **uα** (Unified Assembler), the vision was clear: one assembly language that works across every architecture. Version 26 "Awesome Ada" proved the concept—37 portable instructions running on everything from x86-64 to 8051 microcontrollers. But there was always a deeper goal waiting on the horizon.

Today, I'm excited to share the roadmap for **uα Version 27: "Amazing Grace"**—and it's going to change everything.

## The Ultimate Test: A Compiler Written in Itself

Here's the challenge that defines a truly mature language: can it compile itself?

A Fortran compiler written in Fortran. A Rust compiler written in Rust. A C compiler written in C. This self-hosting capability isn't just a technical milestone—it's proof that a language is powerful enough, expressive enough, and complete enough to handle its most demanding use case.

**In 2027, uα will compile uα.**

The entire compiler—lexer, parser, code generator, optimizer, and all six architecture backends—will be rewritten in Unified Assembly. No more C99 dependency. Just pure uα code, running on bare metal, producing native executables for any platform.

This is the moment when uα transitions from **alpha to beta**.

## Version 27: "Amazing Grace"

The codename honors **Rear Admiral Grace Hopper** (1906–1992), the pioneering computer scientist who created the first compiler and championed the idea that programming languages should be readable and portable. She famously said, "It's easier to ask forgiveness than it is to get permission"—a philosophy that applies perfectly to building an experimental assembler that dares to unify six different CPU architectures.

Grace Hopper believed programming should be accessible without sacrificing power. That's exactly what uα strives for.

## Building the Foundation: Libraries, Libraries, Libraries

Self-hosting doesn't happen in a vacuum. A compiler needs infrastructure—data structures, memory management, file I/O, string processing, symbol tables, hash maps, error handling. The works.

**The work has already begun.**

You can see the progress in this active merge request: [github.com/el-dockerr/UAlang/pull/1](https://github.com/el-dockerr/UAlang/pull/1)

### Dynamic Memory Management

The first major piece is now functional: **`std_malloc`**, a complete dynamic memory allocator written entirely in uα.

```assembly
@IMPORT std_malloc

    ; Allocate 256 bytes on the heap
    MOV R0, 256
    CALL std_malloc.alloc
    ; R0 now contains pointer to allocated memory
    
    ; Use the memory...
    
    ; Free it when done
    CALL std_malloc.free
```

This isn't a wrapper around the system's `malloc()`. It's a real allocator—managing free blocks, coalescence, fragmentation—implemented in pure assembly. It runs identically on x86, ARM, RISC-V, and even 8-bit 8051 controllers (with appropriate memory model adjustments).

The fact that this works, and works *portably*, proves something crucial: **uα is ready for serious systems programming.**

## The Roadmap to Beta

Between now and 2027, the standard library will grow substantially. Here's what's coming:

**Core Data Structures**
- `std_list` — linked lists
- `std_hashmap` — hash tables for symbol lookup
- `std_tree` — binary trees for syntax parsing
- `std_stack` — stack management
- `std_queue` — FIFO queues

**Advanced String Handling**
- Pattern matching and lexical analysis
- UTF-8 support
- String builders for code generation

**File System & I/O**
- Enhanced `std_iostream` with buffering
- Directory traversal
- Binary file manipulation

Each library will be written in uα, tested across all supported architectures, and documented with examples. By late 2027, we'll have everything needed to rewrite the compiler itself.

## Why Self-Hosting Matters

You might wonder: why go through all this effort? The current C99 compiler works. Why rewrite it?

**Because it proves the language is real.**

Right now, uα is a fascinating experiment. After self-hosting, it becomes a legitimate tool. If the language is powerful enough to compile itself—handling complex parsing, optimization passes, and multi-architecture code generation—then it's powerful enough for anything.

It also creates a virtuous cycle: improvements to the language benefit the compiler, which makes development faster, which enables more improvements. The snake eats its tail, and suddenly you have a self-improving ecosystem.

## What This Means for You

If you're already experimenting with uα, these new libraries will dramatically expand what you can build. Want to write a text editor? `std_string` and `std_iostream` have you covered. A networked embedded device? The memory management and data structures are coming.

If you're watching from the sidelines, 2027 is when uα becomes undeniably real. A self-hosted, cross-platform assembler with a rich standard library isn't a toy project—it's a serious tool for systems programming.

## Join the Journey

The merge request is live. The code is open. If you're interested in low-level programming, compiler design, or just want to see how far we can push portable assembly, now is the perfect time to get involved.

Write a library. Port an algorithm. Test the limits. Break things. Report bugs. Suggest optimizations.

By this time next year, when we release **uα 27: "Amazing Grace"**, we'll have hit a milestone that most experimental languages never reach: a fully self-hosting, production-ready compiler that works on six different architectures.

One assembly language. Everywhere. Written in itself.

The future of bare-metal programming is being built right now, one instruction at a time.

---

*Follow the progress on GitHub: [github.com/el-dockerr/UAlang](https://github.com/el-dockerr/UAlang). The project is released under the GNU General Public License v3.*


The Road to uα 27: Self-Hosting and the Beta Milestone


Sometimes the best ideas come from constraints. This weekend, I built something that I'm genuinely excited about: a **hardware-efficient approximation algorithm** that reduces matrix multiplication complexity by 33% while maintaining over 99% accuracy. And the best part? It's specifically designed for the kind of environments where every milliwatt and every gram counts—**satellites, drones, and edge AI systems**.

## The Problem: FPGAs Are Hungry for DSP Blocks

If you've ever worked with FPGAs for vision processing or neural networks, you know the pain: **multipliers are expensive**. Each multiplication requires a DSP block, and DSP blocks consume significant chip area, power, and cost. In a typical 3×3 convolution kernel (the workhorse of computer vision), you need **3 multiplications** per dot product.

Now scale that to real-time image processing on a satellite or surveillance drone. You're doing millions of these operations per second. The power budget? Maybe 5-10 watts. The weight budget? Every gram matters when you're fighting gravity or extending flight time.

Traditional algorithms like Strassen or Coppersmith-Winograd are theoretically faster, but they're "galactic algorithms"—beautiful on paper, impractical in hardware. They require recursive decomposition, massive constant factors, and high-precision floating-point arithmetic. Not exactly what you want when you're routing wires on an FPGA.

## The Solution: Machine Learning Meets Hardware Design

Here's the key insight: **what if we could trade a tiny bit of mathematical precision for a massive reduction in hardware complexity?**

Using a machine-learning-driven search for low-rank coefficients, I found a way to approximate the standard 3-element dot product:

$$\mathbf{a} \cdot \mathbf{b} = a_0 b_0 + a_1 b_1 + a_2 b_2$$

Using only **2 multiplications instead of 3**, by decomposing it as:

$$\mathbf{a} \cdot \mathbf{b} \approx \sum_{k=0}^{1} w_k \cdot \left( \sum_{i=0}^{2} u_{k,i} a_i \right) \cdot \left( \sum_{j=0}^{2} v_{k,j} b_j \right)$$

The magic? All the coefficients ($u_{k,i}$, $v_{k,j}$, $w_k$) are **powers of two**. That means no actual multiplications—just bit-shifts and additions, which are essentially free in hardware!

## The Math (For Those Who Care)

The learned coefficients quantize beautifully to hardware-friendly values:

**Block 0:**
$$\text{term}_A = \left(\frac{a_0}{2} - \frac{a_1}{2} + \frac{a_2}{8}\right), \quad \text{term}_B = \left(-b_0 + b_1 - \frac{b_2}{16}\right)$$

**Block 1:**
$$\text{term}_A = \left(-\frac{a_0}{2} - \frac{a_1}{2} - \frac{a_2}{2}\right), \quad \text{term}_B = \left(\frac{b_0}{2} + \frac{b_1}{2} + b_2\right)$$

**Final result:**
$$\text{result} = -(\text{product}_0) - (\text{product}_1)$$

In C/C++, this becomes:

```cpp
int fast_dot_product(int a[3], int b[3]) {
    int result = 0;
    
    // Block 1 - MULTIPLICATION #1
    int term_a_0 = ((a[0] >> 1) - (a[1] >> 1) + (a[2] >> 3));
    int term_b_0 = (- b[0] + b[1] - (b[2] >> 4));
    result += -(term_a_0 * term_b_0);

    // Block 2 - MULTIPLICATION #2
    int term_a_1 = (-(a[0] >> 1) - (a[1] >> 1) - (a[2] >> 1));
    int term_b_1 = ((b[0] >> 1) + (b[1] >> 1) + b[2]);
    result += -(term_a_1 * term_b_1);

    return result;
}
```

## Hardware Impact: The Numbers That Matter

Let's talk about what this means in the real world:

### FPGA Resource Savings
- **DSP Blocks:** 3 → 2 (33% reduction)
- **Power Consumption:** ~33% reduction (multipliers dominate power budget)
- **Chip Area:** ~20% reduction
- **Throughput:** 50% increase per DSP block (you can fit more parallel operations)

### The Energy Wall
In edge AI and space applications, you're battling the "energy wall." Every computation drains your battery or solar panel budget. By operating at **33% of the energy cost** of standard multiplication, you're effectively extending mission duration, increasing frame rates, or reducing thermal dissipation requirements.

For a satellite doing real-time Earth observation, this could mean:
- **Higher resolution imaging** with the same power budget
- **Longer operational lifetime** with reduced thermal stress
- **Lower launch costs** due to reduced power/cooling requirements

For surveillance drones:
- **Extended flight time** with the same battery
- **Real-time 4K processing** instead of 1080p
- **Smaller form factor** with fewer cooling requirements

## Accuracy: Good Enough Is Perfect

The algorithm achieves **>99.1% Structural Similarity (SSIM)** on real-world image data. Here's what's fascinating: while the Mean Absolute Error might be around 10-13%, the error manifests as a **systematic DC-offset** (brightness shift) rather than destructive noise.

Why does this matter? Because modern neural networks use **Batch Normalization**, which normalizes away constant offsets. The approximation becomes essentially **transparent to the neural network**—edges, gradients, and feature maps are preserved perfectly.

Tested on the classic Lenna and Astronaut test images, the visual difference is imperceptible. The algorithm preserves structural information (what CNNs care about) while sacrificing numerical precision (what CPUs care about).

## Beating Theoretical Limits (Sort Of)

Here's where it gets interesting. Let's compare multiplication counts for full matrix operations:

| Matrix Size | Standard ($N^3$) | My Approach | Strassen/Winograd | Savings |
|-------------|------------------|-------------|-------------------|---------|
| 3×3         | 27               | 18          | 23                | **33%** |
| 6×6         | 216              | 144         | ~150              | **33%** |
| 9×9         | 729              | 486         | ~480              | **33%** |

For the standard 3×3 CNN kernel (the backbone of computer vision), we're achieving **33% savings consistently**, matching or beating the multiplication count of theoretical algorithms like Laderman's method—but with **hardware-native implementation** that actually runs on silicon.

Coppersmith-Winograd is asymptotically faster ( $O(n^{2.373})$ vs $O(n^3)$ ), but it's a "galactic algorithm"—only faster when $n > 10^{15}$. For real-world kernel sizes, linear complexity with a better constant factor wins every time.

## Built in a Weekend (With Years of Preparation)

I won't lie—this project came together in a fraction of a weekend. But that's because I've spent years working with FPGA constraints, studying low-rank decompositions, and understanding the intersection of machine learning and hardware design.

The methodology:
1. **Analytical initialization** of the decomposition
2. **PyTorch optimization** using a Straight-Through Estimator (STE)
3. **Power-of-Two constraints** during training (ensuring zero multipliers)
4. **Validation on non-linear image data** (Lenna/Astronaut)

The result is a **MIT-licensed** algorithm that anyone can use in their FPGA designs, satellite missions, or drone control systems.

## Perfect for SWaP-C Environments

SWaP-C stands for **Size, Weight, Power, and Cost**—the four constraints that dominate aerospace and defense applications.

This algorithm is designed specifically for these environments:

✅ **Space missions:** Lower power = longer mission duration  
✅ **Surveillance drones:** Reduced weight = extended flight time  
✅ **Edge AI devices:** Smaller chip area = lower cost  
✅ **Satellite imagery:** Higher throughput = better temporal resolution  
✅ **Radar pre-processing:** Real-time processing with limited power  

## Limitations (Because Honesty Matters)

Let me be clear about what this is **NOT** good for:

❌ **Flight control laws:** Don't approximate safety-critical calculations  
❌ **Cryptography:** Stochastic errors are unacceptable here  
❌ **Financial calculations:** You need exact precision  

This is a **perception algorithm**—perfect for CNNs, optical flow, radar pre-processing, and feature extraction where **throughput > precision**.

## Benchmark Results

Running 1,000,000 iterations with random test vectors:

```
MULTIPLICATIONS:
  Standard method:    3,000,000 MULs (3 per iteration)
  Approximate method: 2,000,000 MULs (2 per iteration)
  Saved:              1,000,000 MULs (33.3%)

ACCURACY:
  Average error:      0.90%
  SSIM on images:     >99.1%

[SUCCESS] Hardware savings: 33% fewer multipliers 
          = less FPGA area & power
```

## Get the Code

The full implementation, benchmarks, and hardware analysis are available on GitHub:

👉 [Low-Rank Hardware Approximation Repository](https://github.com/el-dockerr/Low-Rank_Hardware_Approximation)

The repository includes:
- **C/C++ implementation** for embedded systems
- **Benchmark suite** with accuracy analysis
- **Visual validation** on real-world images
- **Theoretical complexity analysis**

## What's Next?

This is just the beginning. I'm exploring:
- **5×5 and 7×7 kernel optimizations** for deeper CNNs
- **INT8 quantization** for even more hardware efficiency
- **Hybrid precision schemes** (full precision for critical layers, approximation for perception)
- **Adaptive switching** based on scene complexity

If you're working on FPGA-based vision systems, satellite processing, or drone perception, I'd love to hear from you. Open an issue, start a discussion, or reach out directly.

## The Big Picture

We live in an era where **energy efficiency is the ultimate constraint**. Moore's Law is slowing down. Dennard Scaling is dead. The only way forward is **smarter algorithms** that understand hardware constraints from the ground up.

This project represents a philosophy: **co-design algorithms and hardware together**. Don't just implement existing algorithms on new hardware—redesign the mathematics to fit the silicon.

Because sometimes, 99% accuracy with 33% less power is worth infinitely more than 100% accuracy that drains your battery.

---

*Have questions about FPGA implementation, accuracy analysis, or hardware constraints? Open an issue on the [GitHub repo](https://github.com/el-dockerr/Low-Rank_Hardware_Approximation) or reach out!*

**Technical Note:** For mathematical details on the complexity analysis, see Appendix A (Standard Algorithm: $O(n^3)$) and Appendix B (Coppersmith-Winograd: $O(n^{2.373})$) in the [repository documentation](https://github.com/el-dockerr/Low-Rank_Hardware_Approximation).


Low-Rank Hardware Approximation: 33% Faster Matrix Math for Space & Drones


Another update to **GMlib**, the C++ library that compresses memory in RAM using clever paging! Version 1.1.0 focuses on security.

For those new to GhostMem: it's designed to dramatically reduce your application's RAM footprint through intelligent compression and paging. With optional on-disk swap support, your programs can run with minimal RAM requirements—GhostMem handles the heavy lifting of keeping data compressed and paging it efficiently.

## 🔍 Encryption: See What's Happening Under the Hood

It's not that complex. I added a stream encryption (ChaCha). This is perfect, because ew make use of pages, and we cannot encrypt the whole file to pick up memory from the swap file. The key will be stored in memory and vanish once the program quits.
This encryption shall only make sure that the left swapping file cannot be unzipped, since it might contains sensitive data.
This makes it perfect for IO Devices with low RAM.

## Get Started

If you haven't tried GhostMem yet, check out the [GitHub repository](https://github.com/el-dockerr/ghostmem) to get started. The new documentation will have you up and running in minutes!

For existing users, upgrading is straightforward—the new verbose logging feature is optional and won't affect your current configuration.

Happy coding, and here's to running more with less RAM! 🚀

---

*Have feedback or questions about GhostMem? Feel free to open an issue on GitHub or reach out through the project's communication channels.*


GMlib 1.1.0: Add encryption to disk swapping


I'm excited to share the first update to **GhostMem**, the C++ library that compresses memory in RAM using clever paging! Version 1.0.1 focuses on developer experience with a new debugging feature, better documentation, and support for project sustainability.

For those new to GhostMem: it's designed to dramatically reduce your application's RAM footprint through intelligent compression and paging. With optional on-disk swap support, your programs can run with minimal RAM requirements—GhostMem handles the heavy lifting of keeping data compressed and paging it efficiently.

## 🔍 Verbose Logging: See What's Happening Under the Hood

The headline feature of this release is **verbose logging**. Ever wondered what GhostMem is doing as it compresses and pages your memory? Now you can watch it in action!

Simply enable it in your `GhostConfig`:

```cpp
GhostConfig config;
config.enable_verbose_logging = true;
```

With verbose logging enabled, GhostMem will output detailed debug information about:
- Memory compression and decompression operations
- Paging events and swap activity
- RAM usage optimization in real-time
- Disk swap operations (when enabled)
- Performance metrics and compression ratios

This is incredibly useful during development and troubleshooting. You can observe how effectively your data is being compressed, monitor paging behavior, and fine-tune your configuration for optimal performance.

## 📚 Documentation Gets a Major Upgrade

Good documentation makes all the difference. This release includes significant improvements across all documentation files:

- **Better Navigation**: The README now features badges and a well-organized link structure to help you find what you need faster
- **Quick Reference Table**: A new table in the API reference gives you instant access to all configuration options
- **Enhanced Cross-References**: "See Also" sections throughout the docs help you discover related features and concepts
- **Updated Examples**: All code examples now include the new verbose logging feature, so you can learn by seeing it in action

Whether you're just getting started or diving deep into advanced features, the documentation is now more accessible and easier to navigate.

## 💝 Supporting the Project

Open source thrives on community support. This release introduces a funding configuration file (`.github/FUNDING.yml`), making it easy for users who find GhostMem valuable to support its continued development.

If you're using GhostMem in your projects and want to see it grow, sponsorship options are now available. Every contribution helps maintain the project, add new features, and improve documentation.

## What's Next?

This update lays the groundwork for more developer-friendly features. Future releases will continue to focus on:
- Add encryption during use of I/O swap (disk swap)
- Smarter paging strategies
- More granular control over RAM vs. disk swap balance
- Additional debugging tools for analyzing compression efficiency

## Get Started

If you haven't tried GhostMem yet, check out the [GitHub repository](https://github.com/el-dockerr/ghostmem) to get started. The new documentation will have you up and running in minutes!

For existing users, upgrading is straightforward—the new verbose logging feature is optional and won't affect your current configuration.

Happy coding, and here's to running more with less RAM! 🚀

---

*Have feedback or questions about GhostMem? Feel free to open an issue on GitHub or reach out through the project's communication channels.*


GhostMem 1.0.1: Verbose Logging and Better Documentation


# Welcome

This is my first blog post. This blog is intentionally minimal and fast - no bloat, no tracking, just content.

I'll be writing about:
- C++ development
- Software engineering
- Personal projects
- Technical documentation

Stay tuned for more updates!


dockerr.blog

Blog

uα - Unified Assembler: One Assembly Language for Every Architecture

The Road to uα 27: Self-Hosting and the Beta Milestone

Low-Rank Hardware Approximation: 33% Faster Matrix Math for Space & Drones

GMlib 1.1.0: Add encryption to disk swapping

GhostMem 1.0.1: Verbose Logging and Better Documentation

Welcome to my blog