GPU accelerated Counterfactuar Regret Minimization

Source code for thesis project "GPU accelerated Counterfactual Regret Minimization" at KTH by Gustaf Bergmark. The repository contains two branches:

fixed-flop

This branch contains the source code used for measuring the efficiency of GPU and CPU implementations on a fixed flop subgame.

full-game

This branch contains the source code used for evalutaing the hybrid recall abstraction and scalability of the GPU version.

Requirements

A GPU capable of running CUDA 12.2, and 64 GB of RAM if running full-game

Quickstart

To run the CPU implementation, use:

cargo run --release

To run the GPU implementation, use:

cargo run --release --features GPU