An equity calculator for NLH poker players figuring out how much to bet / call that is plus EV.

What makes it so fast?

• Multi-threaded goodness: I use a branching algorithm with memoization synchronized across threads optimized for concurrent reads using DashMap / rwlocks - as the algorithm is read-heavy.
• Lots of bit manipulation: I use the first 52 bits in a u64 int to represent the state of the board. To derive the rank of each hand, it just involves bit manipulations which is fast.
• Keep as much on the stack as possible: related to above, by encoding boards and cards as u64 values, most of the data stays on the stack avoiding runtime heap allocations and reducing memory overhead.
• SIMD acceleration: Many of the bitwise computations are vectorizable, allowing us to use SIMD instructions to evaluate multiple bits in parallel, hence I use SIMD whenever I run the bit manipulation logic to derive a hand's rank.

As a standard benchmark throughout the development of this project, I considered generating the equity of the following state: 2 players, empty board -> 48 cards to select 5 from. The branching algorithm exhaustively checks every possible combination of 5 cards on the board and computes which player won. It then aggregates all these results into a final probability.

Note that 48 choose 5 is approximately ~1.7m -> the algorithm goes through a 5 layer search tree with ~1.7m leaf nodes.

I first built a proof-of-concept in Python - this setup above resulted in a runtime of approximately 60-70 seconds. After migrating to Rust, an unoptimized implementation still nets around the same time as the Python implementation.

After all these optimizations listed above, I was able to bring down the runtime to approximately 400ms. Up to a 175x speed up!

Try it yourself by going into poker-odds-rs/crates/poker-odds-gui and run cargo run. You'll need the nightly build as I am using SIMD instrinsics.