Migrate linear algebra code from faer-rs to nalgebra #13665
Labels
priority: medium
Rust
This PR or issue is related to Rust code in the repository
type: feature request
New feature or request
Milestone
Comments
mtreinish
added
priority: medium
Rust
mtreinish
added a commit
to mtreinish/qiskit-core
that referenced
this issue
Jan 21, 2025
As part of Qiskit#13665 this migrates the faer-rs usage in quantum_volume.rs to just use nalgebra instead. The only faer usage there was for the qr decomposition and everything else was done with fixed sized arrays or Array2. This migration is fairly simple but it's a good start towards the goal of eventually removing our faer-rs usage in Qiskit. It might improve performance because we remove one layer of heap allocated objects by moving from Array2 to Matrix4 for rust operations. Although we do need an equivalent heap allocation for generating the numpy arrays used in `UnitaryGate` so it might not really change anything significantly.
mtreinish
added a commit
to mtreinish/qiskit-core
that referenced
this issue
Jan 23, 2025
This commit stops using faer and ndarray for the uc_gate.rs module and switches to use nalgebra instead. The uc_gate.rs module works all in terms of 2x2 matrices and moving to nalgebra brings the advantage of having everything be fixed size stack allocated for the arrays used. While this code path isn't performance critical this should improve performance, despite the goal here being to remove a dependency on faer. The tradeoff here is that nalgebra doesn't seem to have a general eigensolver natively. There is functionality for this in the nalgebra-lapack crate, but as the name implies this depends on linking against lapack to work. The reason we originally started using faer is because it's pure rust and doesn't complicate our build system. In this specific case this is not a big deal because the matrices we need to compute the eigenvectors and eigenvalues for is a 2x2 and it's trivial to do by hand, which this commit does. This is something to keep in mind for the future if we need this functionality more generally we will have to make a choice between faer and linking against lapack. Part of Qiskit#13665
mtreinish
added a commit
to mtreinish/qiskit-core
that referenced
this issue
Jan 23, 2025
This commit stops using faer and ndarray for the uc_gate.rs module and switches to use nalgebra instead. The uc_gate.rs module works all in terms of 2x2 matrices and moving to nalgebra brings the advantage of having everything be fixed size stack allocated for the arrays used. While this code path isn't performance critical this should improve performance, despite the goal here being to remove a dependency on faer. The tradeoff here is that nalgebra doesn't seem to have a general eigensolver natively. There is functionality for this in the nalgebra-lapack crate, but as the name implies this depends on linking against lapack to work. The reason we originally started using faer is because it's pure rust and doesn't complicate our build system. In this specific case this is not a big deal because the matrices we need to compute the eigenvectors and eigenvalues for is a 2x2 so this commit includes a function to compute that directly. This lack of functionality is something to keep in mind for the future if we need this functionality more generally we will have to make a choice between faer and linking against lapack. Part of Qiskit#13665
mtreinish
added a commit
to mtreinish/qiskit-core
that referenced
this issue
Jan 23, 2025
This commit stops using faer and ndarray for the uc_gate.rs module and switches to use nalgebra instead. The uc_gate.rs module works all in terms of 2x2 matrices and moving to nalgebra brings the advantage of having everything be fixed size stack allocated for the arrays used. While this code path isn't performance critical this should improve performance, despite the goal here being to remove a dependency on faer. The tradeoff here is that nalgebra doesn't seem to have a general eigensolver natively. There is functionality for this in the nalgebra-lapack crate, but as the name implies this depends on linking against lapack to work. The reason we originally started using faer is because it's pure rust and doesn't complicate our build system. In this specific case this is not a big deal because the matrices we need to compute the eigenvectors and eigenvalues for is a 2x2 so this commit includes a function to compute that directly. This lack of functionality is something to keep in mind for the future if we need this functionality more generally we will have to make a choice between faer and linking against lapack. Part of Qiskit#13665
mtreinish
added a commit
to mtreinish/qiskit-core
that referenced
this issue
Jan 23, 2025
This commit stops using faer and ndarray for the uc_gate.rs module and switches to use nalgebra instead. The uc_gate.rs module works all in terms of 2x2 matrices and moving to nalgebra brings the advantage of having everything be fixed size stack allocated for the arrays used. While this code path isn't performance critical this should improve performance, despite the goal here being to remove a dependency on faer. The tradeoff here is that nalgebra doesn't seem to have a general eigensolver natively. There is functionality for this in the nalgebra-lapack crate, but as the name implies this depends on linking against lapack to work. The reason we originally started using faer is because it's pure rust and doesn't complicate our build system. In this specific case this is not a big deal because the matrices we need to compute the eigenvectors and eigenvalues for is a 2x2 so this commit includes a function to compute that directly. This lack of functionality is something to keep in mind for the future if we need this functionality more generally we will have to make a choice between faer and linking against lapack. Part of Qiskit#13665
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
What should we add?
In qiskit 1.0 we started using the faer-rs library for linear algebra from rust. This enabled us to migrate a lot of python/numpy numeric code into Rust during the development of 1.x which greatly accelerated a lot of transpiler code in the 1.x cycle. But with the recent releases of the faer it's become clear the project isn't really in a place where we can rely on it as a stable upstream dependency for Qiskit. The development of the library is very rapid and it's hard to keep track of what is going on, and the MSRV changes are incompatible with our requirements for Qiskit.
For our use cases most of the linear algebra we do is with small matrices, especially in the transpiler, typically with 2x2 and 4x4 matrices, and nalgebra is a good fit for that because we can used fixed size stack allocated types for the arrays. This will also provide better performance in most cases for what we're currently using. It's also a more stable/established library that is better aligned with our requirements for dependencies in Qiskit.
The text was updated successfully, but these errors were encountered: