Autopiper: automatic pipeline synthesis

Autopiper is a pipeline synthesis tool that understands high-level constructs and allows for automatic pipelining and automatic inference of low-level microarchitectural ‘plumbing’ in RTL (Verilog). It has well-defined semantics that allow for easier reasoning about digital systems at the level of functional state, and allow leeway for formal transforms to produce correct high-performance implementations.

Autopiper is a research project – this may or may not work! – and is intended for now more as a developed-in-the-open toy than a serious tool.

Autopiper is developed by Chris Fallin <cfallin@c1f.net>. Autopiper’s copyright is assigned to Google Inc., but autopiper is an independent side-project and is not an official Google product in any way.

Current status

The pipeline lowering / synthesis backend (IR to Verilog) has implemented the key transforms, and works with small tests. Currently the frontend is in development (parsing, typechecking/inference, codegen), after which the macro layer will be developed alongside some larger test designs.

In other words: it’s not actually done yet! You can play with the current functionality, i.e., lowering IR to pipeline form, by running autopiper on the IR inputs in tests/.

Building

Autopiper requires CMake, the Boost C++ libraries, and GMP (an arbitrary-precision numeric library).

On Ubuntu:

$ sudo apt-get install cmake libboost-dev libgmp-dev

On Arch Linux:

$ sudo pacman -Syu cmake boost gmp

Then create a build directory and compile the application:

$mkdirbuild\mathrm{/}mkdir\; build/$ cd build/ $cmake\mathrm{.}\mathrm{.}cmake\; ..$ make $ src/autopiper –help