357894b539
The bug that happened was that the last send of a register memory bus interaction was removed due to having multiplicity zero. In the end the bus was left unbalanced. Therefore, we need to run the solver optimizer before the reg mem opt, so that these multiplicities are removed in pairs before the apc optimizer runs and keeps only the first receive and last send, because those multiplicities need to match. This is the main solution implemented here in `build`. I also added some checks that the reg mem interactions need to be even per address and in total. A side effect of this fix is that we can no longer easily remove the timestamp columns of removed mem int receives, the constraint that decomposes it into bytes, the decomposing columns, and the range checks on them. The pattern matching becomes too brittle and would be even hackier if we wanted to keep supporting it that way. That's what causes the regression in the test, and all the code that is removed in this PR, since we won't do this anymore. The ultimate fix is to rely on the optimizer to remove these columns, since: - The reg mem int receive is removed - Then the timestamp column becomes a free variable in its constraint - The constraint can be removed - The limb decomposition columns become unused in polynomial constraints and only appear in range checks - Therefore the range checks and columns can also be removed At this point the result should be the same as the pattern matching alternative removed by this PR. I anticipate that we'll also need a fixpoint loop because it may be the case that we need a few iterations to propagate everything possible when this is implemented.
powdr
WARNING: This codebase is experimental and has not been audited. DO NOT USE FOR PRODUCTION!
For detailed documentation please visit the powdr book.
If you have any questions or want to contribute, feel free to write us in our Matrix Chat.
powdr is a toolkit that helps build zkVMs and similar proof frameworks.
It has two main components:
- powdr-asm: an extensible assembly IR language to perform dynamic executions.
- powdr-PIL: a low level constraint language that allows you to define arithmetic constraints, lookups, etc. It includes a functional meta-constraint language to describe how constraints are generated.
Both frontend and backend are highly flexible.
The powdrVM is a zkVM implementation that builds on powdr. Users can write code in std Rust which is compiled to RISCV, then to powdr-asm and finally to powdr-PIL.
powdr-PIL can be used to generate proofs using multiple backends, such as:
All stages are fully automatic, which means you do not need to write any additional code for witness generation besides your Rust code. All witnesses are automatically inferred from the constraints. Since the witnesses are inferred, powdr can ensure that the system is not underconstrained, i.e., there are no additional unwanted witnesses.
All artifacts from the compilation pipeline are human-readable. You can inspect the powdr-asm IR, the compiled powdr-PIL file, and its final optimized version.
The assembly language is designed to be extensible and does not have a single native instruction. Instead, all instructions are user-defined and because of that, it is easy to adapt powdr assembly to any VM.
Notes on Efficiency
The current focus of the project is VM support and developer experience. The compiler generates many unnecessary columns. We will soon start writing optimizer steps that should bring performance closer to existing production systems.
Project structure
For an overview of the project structure, run:
cargo doc --workspace --no-deps --open
Contributing
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as below, without any additional terms or conditions.
License
This project is licensed under either of
at your option.