TensorFlow Data Backend¶

Contents

Why a TensorFlow backend?
Challenges

Note

This project was completed as part of GSoC 2021 3.

QuTiP’s data layer provides the mathematical operations needed to work with quantum states and operators, i.e. Qobj, inside QuTiP. As part of Google Summer of Code 2020, the data layer was rewritten to allow new backends to be added more easily and for different backends to interoperate with each other. Backends using in-memory spares and dense matrices already exist, and we would like to add a backend that implements the necessary operations using TensorFlow 1.

Why a TensorFlow backend?¶

TensorFlow supports distributing matrix operations across multiple GPUs and multiple machines, and abstracts away some of the complexities of doing so efficiently. We hope that by using TensorFlow we might enable QuTiP to scale to bigger quantum systems (e.g. more qubits) and decrease the time taken to simulate them.

There is particular interest in trying the new backend with the BoFiN HEOM (Hierarchical Equations of Motion) solver 2.

Challenges ¶

TensorFlow is a very different kind of computational framework to the existing dense and sparse matrix backends. It uses flow graphs to describe operations, and to work efficiently. Ideally large graphs of operations need to be executed together in order to efficiently compute results.

The QuTiP data layer might need to be adjusted to accommodate these differences, and it is possible that this will prove challenging or even that we will not find a reasonable way to achieve the desired performance.

Expected outcomes ¶

Add a qutip.core.data.tensorflow data type.
Implement specialisations for some important operations (e.g. add, mul, matmul, eigen, etc).
Write a small benchmark to show how Qobj operations scale on the new backend in comparison to the existing backends. Run the benchmark both with and without using a GPU.
Implement enough for a solver to run on top of the new TensorFlow data backend and benchmark that (stretch goal).