Skip to content
FPGA Haskell machine with game changing performance. Reduceron is Matthew Naylor, Colin Runciman and Jason Reich's high performance FPGA softcore for running lazy functional programs, including hardware garbage collection. Reduceron has been implemented on various FPGAs with clock frequency ranging from 60 to 150 MHz depending on the FPGA. A hig…
Branch: master
Clone or download
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
docs Collect some of the relevant papers and presentations Jan 19, 2018
emulator EMU: usage message Sep 4, 2017
flite
fpga
ghc-flite ghc-flite: a GHC based front-end Feb 27, 2015
hf 2015 Haloween maintenance Nov 2, 2015
programs
red-lava Improved support for parallel builds and test (including Verilog sim) Nov 19, 2017
semantics semantics: Variants now extracts readable versions Nov 14, 2015
Makefile
README.md Clang can't build C code from Flite, give workaround Feb 9, 2019
STATUS

README.md

Reduceron, an efficient processor for functional programs

WHAT IS REDUCERON?

Reduceron is a high performance FPGA softcore for running lazy functional programs, complete with hardware garbage collection. Reduceron has been implemented on various FPGAs with clock frequency ranging from 60 to 150 MHz depending on the FPGA. A high degree of parallelism allows Reduceron to implement graph evaluation very efficiently.

Reduceron is the work of Matthew Naylor, Colin Runciman and Jason Reich, who have kindly made their work available for others to use. Please see http://www.cs.york.ac.uk/fp/reduceron for supporting articles, memos, and original distribution.

OK, WHAT'S THIS THEN?

The present is a fork of the original distribution which intends to take the (York) Reduceron from the research prototype to the point where it can be useful for embedded projects and more.

The York Reduceron needs the following enhancements to meet our needs:

  1. The heap and program must (for the most parts) be kept in external memory, with FPGA block memory used for the stacks and heap and program caches.

    This simultaneously enables smaller and less expensive FPGAs to be used as well as allows for a much larger heap and larger programs.

  2. Access to memory mapped IO devices (and optionally, RAM).

  3. Richer set of primitives, including multiplication, shifts, logical and, or, ...

  4. Support for 32-bit integers - this greatly simplifies interfacing to existing IO devices and simplifies various numerical computations.

  5. Stack, update stack, [and case table stack?] should overflow into/underflow from external, allowing for orders of magnitude larger structures.

While Reduceron technically refers to the FPGA implementation, it is supported by

  • Flite: the F-lite to Red translator.
  • A Red emulator in C
  • Red Lava: Reduceron is a Red Lava program, which generate Verilog
  • Support for Verilog simulation and synthesis for various FPGA boards

As much of the history as was available has been gathered and Reduceron, Lava, and the Flite distribution have been merged into one repository.

HOW DO I USE IT?

The was last tested with Glasgow Haskell Compiler, Version 8.4.4 on macOS 10.14.3 and Linux, 64-bit.

Optionally: just run make in the toplevel directory and a large regression run will start. The Verilog simulation part will take weeks to finish.

To build:

make

Or run a specific test suite:

make -C programs $X

where $X is one of regress-emu, regress-flite-sim, regress-flite-comp, or regress-red-verilog-sim.

Note: the code generated by the C backend for Flite (used in the regress-flite-comp) depends on GCC features, such as nested functions. To build on macOS, install real gcc (say via Mac Homebrew) and invoke make as make CC=gcc-7 (assuming you installed version 7 of gcc).

To build a hardware version of a given test

cd fpga; make && flite -r ../programs/$P | ./Red -v

where $P is one of the programs (.hs). Next, build a Reduceron system for an FPGA board, fx the BeMicroCV A9:

make -C Reduceron/BeMicroCV-A9

Unfortunately programs can't currently be loaded dynamically but are baked into the FPGA image. It's a high priority goal to change that.

WHERE IS THIS GOING?

Plan

  1. Port to Verilog and remove Xilinx-isms. DONE!

  2. Shrink to fit mid-sized FPGA kits (eg. DE2-115 and BeMicroCV-A9). DONE!

  3. Rework Lava and the Reduceron implementation to be more composable and elastic; this means fewer or no global assumptions about timing. ONGOING!

  4. Support load/store to an external bus (the key difficulty is stalling while waiting on the bus).

  5. Use the program memory as a cache, making programs dynamically loadable and dramatically raise the size limits.

Eventual Plan

  • Move the heap [and tospace] to external memory
  • Add a heap cache/newspace memory
  • Implement the emu-32.c representation for the external heap
  • Much richer primitives
  • Haskell front-end

Long Term Plan

  • Research the design space; explore parallelism

OPEN QUESTIONS, with answers from Matthew:

Q1: Currently there doesn't seem an efficient way to handle toplevel variable bindings (CAFs). What did the York team have in mind there or does it require an extension? (Obviously one can treat them all other functional arguments, but that would mean a lot of parameters to pass around).

A1: "Some mechanism would be needed to construct graphs at a specified location on the heap at the beginning of program execution. The initial (unevaluated) graphs have constant size so can be linked to at compile time."

Q2: Why does Flite default to 0 for the MAXREGS parameter? Eg, why is

  redDefaults = CompileToRed 6 4 2 1 0

A2: (Historical reasons it would appear).

Q3: What happend to Memo 24?

A3: "I'd like to say it was our best kept secret, but in reality it probably got trashed :)"

tip for next commit

You can’t perform that action at this time.