Bundle

Bundle is a hardware architecture designed to accelerate array computing on FPGA. It is similar to NumExpr and has the same use case: it can speed up computations when applied to large arrays. NumPy arrays are split into chunks and offloaded to the FPGA which can operate in parallel on each chunk.

Right now, the goal is to map this hardware to the PYNQ-Z1 board. But if you don't have this board and want to see how things work, the whole system can be simulated entirely in Python (it is also very convenient for exploring new architectures). For that I have created a small Hardware Description Language called PyClk. It can generate this kind of waveforms:

Architecture

The hardware is intentionally kept simple, and consists of:

• memories: they are used to cache chunks of the host arrays in the FPGA before making operations on them, and write chunks of the result of these operations before copying them to the host DDR memory.
• DDR <-> FPGA streamers: they handle data copy from the DDR memory to the FPGA memories (to get the function arguments) and from the FPGA memories to the DDR memory (to return the function results).
• iterators: they stream data from FPGA memories, feed it to functions, and stream the function's result to another FPGA memory.
• functions: they take data in and produce a result, e.g. the sum of two numbers.

Streamers, iterators and functions can connect to any memory, but the hardware doesn't have a clue what it is doing: it is just a bundle of wires. All the knowledge comes from the software, which maps a high-level expression down to low-level basic operations.

The software is responsible for orchestrating the evaluation of an expression, e.g. a * b + c. It breaks down the expression into a sequence of functions (here tmp = a * b then tmp + c). It allocates FPGA memory for the evaluation, copies chunks of the data from the host DDR memory to the FPGA memory, schedules operations, and copies back the result to the host DDR memory. Since the FPGA can execute many operations in parallel, the scheduler uses asynchronous programming (and asyncio in particular) to manage memory copy and computation tasks.

Install

You will need PyClk to simulate the hardware:

pip install git+https://github.com/davidbrochart/pyclk

Bundle can be installed directly through GitHub: pip install git+https://github.com/davidbrochart/bundle

Or you can clone this repository and pip install -e bundle