Comment on page
Architecture
zkLLVM is designed to be adjustable for different use cases with modular and extendable architecture. Its pipeline can be divided into 4 main components:
- Frontends are responsible for parsing the input high-level language and generating the AST. Currently, we support C/C++ and Rust as frontends which are based on Clang and Rustc respectively. The main difference between the vanilla LLVM project frontends and zkLLVM frontends is that the latter generates the AST in a way that is compatible with the zkSNARKs circuit generation process - it contains additional types and metadata that are required for the circuit generation process.
circifiergets AST and converts it into the intermediate representation (IR) applying circuit-specific optimizations. The output of the circifier is a circuit-specific LLVM IR.assignertakes the circuit-specific LLVM IR and generates assignment table, using description of the components corresponding to the circuit-specific LLVM IR instructions.assigneralso generates the arithmetization-specific constraints for the provided LLVM IR instructions. The output of theassigneris a set of constraints and the assignment table.- Prover and Verifier are responsible for the proof generation and verification. They are not part of the zkLLVM project, but are provided by the proof system you are using. The Verifier can be on-chain or off-chain depending on the use case.
We are going split the assigner logic into two parts: one will be responsible for the assignment table generation and the other will be responsible for the constraints generation. The reason for this is that the constraints generation process can be don one time for the whole circuit, while the assignment table generation process should be done for each input.
There is also an SDKs layer upon the pipeline, which is responsible for the integration of the zkLLVM pipeline into the projects. It consist of the library full of circuit-friendly algorithms implementations and of some additional circuit development tools (tools are not fully implemented yet).
In order to proceed it may make sense to explain what we assume by the term
arithmetization, since in different context it might have different meanings. Here we use the term to describe the limitations and the rules of some particular proof system input type. The list of currently existing arithmetizations contains but is not limited to:- R1CS (Rank-1 Constraint System)
- PLONK (Permutation-based zkSNARK)
- PLONK with custom gates
- AIR
Each aritmetization has its own limitation, which have to be taken into account when designing the circuit. Most algorithms can be implemented in different ways, and the different ways can be efficient for different arithmetizations.
As you may notice, arithmetizations are not limited to one proof system. The code compiled to vanilla PLONK arithmetization can be used with different proof systems.
As mentioned above, zkLLVM is designed to be extendable. In order to add support for a new arithmetization, one has to implement only the
assigner component of the system. It can be implemented in any language, since the core of the assigner component is an LLVM IR parser and can be easily ported to any language. Current assigner implementation is written in C++ and can be found here and is only for PLONK with custom gates arithmetization. We are working on the new version of the assigner component, which will be written in Rust and will be more generic and will support more arithmetizations. Let's look how the implementation process of this new version will look:- First, we will implement the LLVM IR parser in Rust. This can be done from scratch or by porting the existing C++ implementation. Or we can use the existing LLVM IR parser from the inkwell and extend it with the custom types and built-ins.
- Then the rest of assigner logic needs to be implemented. This can be done using an existing RUST PLONK components library, like the components from halo2. You need to solve the following tasks in order to implement minimal assigner for your arithmetization:
- Assignment table generation
- Constraints generation
- Constraints serialization
Apart from the basic functionality, there is a lot of things
assigner can do to improve the performance of the resulting circuit. For example, it can apply different optimizations to the circuit, like:- Adjust the order of the components
- Adjust the size of the components in the assigment table
- Group together the components with common constraints (makes sense for PLONKish arithmetizations)
Optimizations in the
assigner might not be very effective for small circuits, but it can give a significant performance boost for the large ones, such as complex zk-Bridges, zk-VMs or zk-EVMs. More about how to implement assigner the optimizations and how to customize its behaviour to your needs can be found optimizations section.An arithmetization is not limited to only one proof system. For example, there are a lot of proof systems with PLONKish arithmetizaion, and the same circuit can be used with all of them. But implementation of the custom
assigner may be required in case if you work on your own proof system or optimize the existing one. In that case the parameters of the arithmetization might change and the circuit generation has to be adjusted accordingly.You may have noticed, that we only mentioned the
assigner component, but not the circifier. The reason for that is that the circifier is not arithmetization-specific and can be used for any arithmetization. It generates the intermediate representation for any circuit generation related process, and the assigner component is responsible for the conversion of the intermediate representation into the arithmetization-specific constraints and assignment table.