📦 SGen – Streaming Signal‑Processing Generator

<p align="center"> <img src="img/sgen.png" alt="SGen logo" width="300"/> </p>

SGen is a generator that produces compact, high‑performance hardware designs for a variety of signal‑processing transforms. All generated modules work on streaming data: the dataset is split into chunks that are processed over several cycles, thus allowing a reduced use of resources. The size of these chunks is called the streaming width.

As an example, the figures below represent three discrete Fourier transforms on 8 elements, with a streaming width of 8 (no streaming), 4 and 2.

<p style="display:flex;justify-content:center;"> <img src="img/dft8basic.svg" alt="FFT 8‑point, no streaming" style="margin:0 20px;"> <img src="img/dft8s4basic.svg" alt="FFT 8‑point, streaming width = 4" style="margin:0 20px;"> <img src="img/dft8s2basic.svg" alt="FFT 8‑point, streaming width = 2" style="margin:0 20px;"> </p>

The generator emits a Verilog file ready to be synthesised on any FPGA.

• 📄 Technical overview – https://acl.inf.ethz.ch/research/hardware
• 🐞 Bug reports / feature requests – contact F. Serre

---

🚀 Quick start (requires Java 8+)

1. Download the latest sgen.bat from the releases page.

2. Open a terminal in the folder containing sgen.bat and run:

# Windows
sgen.bat -n 3 wht

# Linux
./sgen.bat -n 3 wht

The command above generates a streaming Walsh‑Hadamard transform on 2³ = 8 points.

---

🖥️ Command‑line interface

A SGen command line consists of a list of options followed by the name of the transform you want to generate:

sgen.bat [options] <transform‑name> [lp matrices...]

Options

| Option | Argument | Meaning | |--------|----------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | -n | <n> | Required – log₂ of the transform size (e.g. -n 3 → 8 elements). | | -k | <k> | log₂ of the streaming width. -k 2 ⇒ 4 input / output ports, one transform every 2^(n‑k) cycles. Omit for full‑parallel (one port per data element). | | -r | <r> | log₂ of the radix (used by DFT/WHT). Must divide n. For compact designs (*compact) it must also satisfy r ≤ k. If omitted, the highest possible radix is chosen. | | -sf | <sf> | Scaling factor for DFTs (applied at every stage). Example: -sf 0.5 idft. | | -o | <file> | Output file name. | | -benchmark | – | Adds a benchmark module in the generated design. | | -rtlgraph | – | Emits a DOT graph of the generated RTL. | | -dualramcontrol | – | Uses independent read/write addresses (uses more resources, but offers more flexibile timing constraints). Implicitly enabled for *compact designs. | | -singleported | – | Uses single‑ported RAM (read = write address). May increase latency. | | -zip | – | Packs the design and all dependencies (e.g. FloPoCo modules) into a zip archive. | | -hw | <repr>| Hardware arithmetic representation of the input data (see the table below). |

Hardware data‑type (-hw)

| Token | Description | |-------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------| | fixedpoint <int> <frac> | Signed fixed‑point: <int> integer bits, <frac> fractional bits. | | signed <size> | Signed integer of <size> bits (alias of fixedpoint <size> 0). | | char, short, int, long | Signed integers of 8, 16, 32, 64 bits. Equivalent of signed ?. | | unsigned <size> | Unsigned integer of <size> bits. | | uchar, ushort, uint, ulong | Unsigned integers of 8, 16, 32, 64 bits. Equivalent of signed ?. | | ieee754 <wE> <wF> | IEEE‑754 format built on top of FloPoCo operators. Unless otherwise specified when generating FloPoCo operators, denormal numbers are flushed to zero. | | float, double, half | IEEE‑754 single, double and half precision. Equivalent of ieee754 ? ?. | | minifloat | 8‑bit floating‑point (4‑bit exponent, 3‑bit mantissa). Equivalent of ieee754 4 3. | | bfloat16 | Brain‑float 16 (8‑bit exponent, 7‑bit mantissa). Equivalent of ieee754 8 7. | | flopoco <wE> <wF> | FloPoCo floating‑point (exponent =wE, mantissa =wF). Requires the corresponding FloPoCo VHDL files in the flopoco/ folder. | | complex <repr> | Cartesian complex number, each component encoded with <repr> and concatenated. |

---

Supported transforms

| Category | Command | Example | |-------------------------------------------------------------------------|------------------------|---------| | Linear permutations | lp | sgen.bat -n 5 -k 2 lp bitrev | | DFT (full‑throughput / compact) | dft / dftcompact | sgen.bat -n 4 -k 2 -hw complex fixedpoint 8 8 dft | | Inverse DFT (full‑throughput / compact) | idft / idftcompact | sgen.bat -n 10 -k 3 -hw complex fixedpoint 8 8 idft | | Walsh‑Hadamard (full‑throughput / compact) | wht / whtcompact | sgen.bat -n 6 -k 3 -hw fixedpoint 8 8 wht | | Inverse WHT (full‑throughput / compact) | iwht / iwhtcompact | sgen.bat -n 6 -k 3 -hw fixedpoint 8 8 iwht |

Linear permutations (lp)

lp expects an invertible bit‑matrix (row‑major order) that describes the permutation.
Convenient shortcuts:

| Shortcut | Meaning | |----------|---------| | bitrev | Bit‑reversal permutation. | | identity | No change. |

Multiple matrices can be listed, separated by spaces. The i‑th matrix will be applied to the i‑th incoming dataset.

This design allows full‑throughput pipelines (no idle cycles between datasets). See this publication for details.

Example: streaming bit‑reversal

# 32‑point bit‑reversal, streamed on 4 ports (k=2)
sgen.bat -n 5 -k 2 lp bitrev

# 8‑point bit‑rev on odd datasets, a custom “half‑rev” on evens.
sgen.bat -n 3 -k 1 lp bitrev 100110111

---

Full-throughput and compact designs

• Full-thoughput designs allow to perform the transform without any delay between two datasets.

• Compact designs have an architecture that reuses several times the same hardware. They require however a delay between two transforms (see generated description).

---

RAM‑control strategies

When n > k (i.e. the design is streamed), memory blocks are required.
Choose the most suitable control scheme with the corresponding flag:

| Mode | Flag | Behaviour | Resource / Flexibility trade‑off | |------|------|----------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Dual‑RAM control | -dualramcontrol | Independent read & write address generators. | Highest flexibility (datasets can start at any time). Implicit for *compact designs. | | Single‑RAM control (defa