https://github.com/stuarthayhurst/battleships
Battleships opponent and compute experiments, with AVX2 / AVX-512
https://github.com/stuarthayhurst/battleships
avx2 avx512 battleships compute opponent
Last synced: 2 months ago
JSON representation
Battleships opponent and compute experiments, with AVX2 / AVX-512
- Host: GitHub
- URL: https://github.com/stuarthayhurst/battleships
- Owner: stuarthayhurst
- License: mit
- Created: 2022-03-14T16:54:57.000Z (over 4 years ago)
- Default Branch: master
- Last Pushed: 2026-02-25T17:22:09.000Z (4 months ago)
- Last Synced: 2026-02-25T20:39:30.082Z (4 months ago)
- Topics: avx2, avx512, battleships, compute, opponent
- Language: Python
- Homepage:
- Size: 110 KB
- Stars: 0
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
## battleships
- A repository to experiment with the battleships board game
- I don't particularly like battleships, but it makes for interesting problem solving
## Sub-projects:
- Battleships computer opponent: A computer opponent to play battleships against
- Battleships board compute: Find every valid battleships layout, according to grid size and a set of ship lengths
- Includes AVX2 and AVX-512 accelerated C implementations
## Battleships computer opponent:
- `battleships.py` runs a game of battleships, with options to play with 2 players, against the computer, or watch the computer play against a random number generator
- The computer opponent has been designed to be as hard as is possible
- The opponent plays completely fairly, and can't see the locations of your ships (read the source code if you don't believe me)
- The opponents are stored in `opponents/`, and can be benchmarked with `./benchmark.py`
## Battleships board compute:
- `compute/countBoards.py` will calculate the number of valid battleships layouts from a grid size and list of ship lengths
- This code has been written with `pypy3` in mind, and is strongly suggested to be used (~5x performance improvement)
- This script takes around 13.5 seconds to run using a Ryzen 7 7700X and `pypy3`
- With `n` being the number of ships and `w` being with width of the board, the time complexity scales with `O((w^2 * 2)^n * n!)`
- This is the unoptimised time complexity, if every board was checked
- In reality, it likely won't scale this way, as most boards are discarded early
- Using 5 ships and a width of 7, this gives ~1.1 trillion combinations to try
- **Alternatively**, `compute/countBoards.c` is a C implementation of the same algorithm
- This runs in about 0.5 seconds, using a Ryzen 7 7700X
- However, optimisation work has only been done on Zen 3, Zen 3+ and Zen 4 systems
- Compile: `make -C compute`
- Supports `DEBUG=[true/false]` to enable debug support and verbose build output
- Supports `VERBOSE=[true/false]` to enable verbose build output
- Supports `ARCH=[microarchitecture]` to target a specific microarchitecture
- Defaults to using `-march=native`
- `ARCH=x86-64` could be helpful to run on any x86-64 CPU, if being used to benchmark
- Supports `AVX2=[true/false]` to enable AVX2 optimisations
- Supports `AVX512=[true/false]` to enable AVX-512 optimisations
- Supports `AVX512_SHORT=[true/false]` to enable AVX-512 optimisations, with 8-bit elements
- AVX2, AVX-512 and AVX-512 8-bit optimisations are enabled by default
- They'll be used in order of AVX-512 8-bit -> AVX-512 -> AVX2 -> scalar
- The Makefile selects which options the code is allowed to use
- The code tests for support at compile time and makes a decision
- Supports `BOARD_TYPE_SIZE=[integer]` to force a specific board element size
- Defaults to `32`
- If `AVX512_SHORT` is enabled and supported, this will be overridden
- Supports `BOARD_WIDTH=[integer]` to force a select board width
- Defaults to `7`
- Run: `./compute/countBoards`
- These programs don't save the boards, but could easily be modified to save or print them
- Comparison of implementation performance (Ryzen 7 7700X):
- Scalar code is free to be auto-vectorised by the compiler
| Runner + version | Runtime | Valid boards / s |
|:----------------------|:--------|:-----------------|
| Python (3.12) | 71.56s | 1,743,000 |
| Pypy3 (3.10 / 7.3.16) | 13.28s | 9,394,000 |
| C (Scalar) (GCC-15) | 0.50s | 250,900,000 |
| C (AVX2) (GCC-15) | 0.50s | 251,700,000 |
| C (AVX-512) (GCC-15) | 0.53s | 236,500,000 |
| C (AVX-512S) (GCC-15) | 0.42s | 299,500,000 |
- Runtime is rounded to 2 decimal places
- Number of valid boards per second is rounded to 4 significant figures
- Comparison of SIMD selection performance by runtime (Ryzen 7 9700X, GCC-14):
| SIMD | 9x9 | 10x10 | 11x11 | 12x12 | 13x13 |
|:--------|:-------|:--------|:--------|:---------|:---------|
| Scalar | 18.74s | 67.56s | 379.26s | 1396.50s | |
| AVX2 | 26.94s | 89.10s | 531.56s | 1633.51s | |
| AVX512 | 19.71s | 121.06s | 363.69s | 1330.86s | |
| AVX512S | 16.02s | 83.26s | 234.88s | 786.12s | 2087.83s |