edif2qmasm makes it possible to run Verilog or VHDL programs on a D-Wave quantum annealer.
Why quantum annealing? The primary reason to target a quantum annealer—really, any hardware annealer—is that programs can be run in either the forward or backward direction. One can in fact specify any combination of inputs and outputs and solve for the other, unspecified, values. This benefits both
- expressiveness, as some programs are easier to write in the A → B direction than in the B → A direction, and
- performance, as some classical algorithms run quickly in the A → B direction but slowly in the B → A direction. Consider verifying a solution to an NP-complete problem (fast) versus producing a solution to an NP-complete problem (slow).
Why Verilog/VHDL? Some of the advantages of using a hardware-description language as a D-Wave programming language are that it
- supports basic programming-language features such as conditionals, loops, multi-bit constants and variables, assignments, arithmetic operations, and modules,
- provides precise control over bit widths, which reduces the number of wasted qubits (a precious resource in contemporary D-Wave systems),
- enables exploiting the code optimizations and debugging support provided by synthesis tools.
More precisely, edif2qmasm converts from the EDIF netlist format, which can be output by various synthesis tools, to the QMASM quantum macro assembly language. To date, edif2qmasm has been tested only with Verilog because there exist open-source compilers that convert Verilog to EDIF, and I don't know of an equivalent open-source tool that can convert VHDL to EDIF.
In addition to this file, two other sources of documentation are
There also exists a peer-reviewed academic publication on edif2qmasm that describes the entire process of compiling Verilog programs to a D-Wave Hamiltonian function, explains how edif2qmasm's standard-cell library was constructed, presents some use cases, and even includes a bit of analysis:
Scott Pakin. "Targeting Classical Code to a Quantum Annealer". In Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS 2019), 13–17 April 2019, Providence, Rhode Island, USA, pp. 529–543. ACM, New York, New York, USA. ISBN: 978-1-4503-6240-5, DOI: 10.1145/3297858.3304071.
Associated with the above is a 2-minute "lightning talk" video, which is essentially an advertisement to attend the talk (which was 16 April 2019) and read the paper.
edif2qmasm is written in Go and therefore depends upon a Go compiler to build. See the INSTALL file for build instructions.
edif2qmasm usage is straightforward:
edif2qmasm -o myfile.qmasm myfile.edifIf no input file is specified, edif2qmasm will read from the standard input device. Run edif2qmasm --help for a list of available command-line options.
To run the generated code with QMASM, you'll need to point it to the edif2qmasm standard-cell library. In Bash, enter
export QMASMPATH=/usr/local/share/edif2qmasm:$QMASMPATHreplacing /usr/local with whatever installation prefix you used.
edif2qmasm has only limited support for sequential logic. Sequential logic is implemented by replicating the entire circuit once per clock cycle for a compile-time specified number of clock cycles (cf. the --cycles command-line option). Clocked flip-flops are supported, but unclocked latches are not.
The resulting QMASM programs are not very robust in that the minimum-energy solutions do not consistently represent a correct execution when run on D-Wave hardware. Running QMASM with --postproc=opt helps substantially. Other suggestions on how to improve robustness are welcome.
edif2qmasm is provided under a BSD-ish license with a "modifications must be indicated" clause. See the LICENSE file for the full text.
edif2qmasm is part of the Hybrid Quantum-Classical Computing suite, known internally as LA-CC-16-032.
Scott Pakin, pakin@lanl.gov