CAT: Cellular Automata on Tensor Cores

CAT is a WIP library to perform Fast Larger than Life (LtL) Cellular Automata (CA)simulations using tensor cores. It's programmed using CUDA and C++.

Requirements

• CUDA 11+
• C++17
• A tensor core NVIDIA GPU

Compilation

• Modify the Makefile to suite your needs. Architecture microcode, Radius, LtL rules, GPU Block size (we recomed using 16x16), and the number of regions to use for CAT (we recomend using NREGIONS_H=2, NREGIONS_V=5 for low shared memory GPUs and NREGIONS_H=1, NREGIONS_V=13 otherwise, although result may vary).
• Perform the compilation make -j 10. If debug information is needed, compile using make debug -j 10.

Execution

Depending on the compilation performed, the main executable should be bin/prog or debug/prog. To execute, simply run bin/prog <method> <n> <steps>, where method is the solver algorithm, n is the side length without halo, and steps is the number of steps to simulate. Below is a complete list of the supported arguments which can be accesed using bin/prog -h:

Usage: CAT: Celular Automata on Tensor Cores [--help] [--version] [--deviceId VAR] [--density VAR] [--seed VAR] [--doVerify] n solver steps

Positional arguments:
  n               Size of the data domain 
  solver          Solver to use:
                        0 - BASE: Global Memory
                        1 - SHARED: Shared Memory
                        2 - CAT: Fast Tensor Core
                        3 - COARSE: Thread Coarsening
                        4 - MCELL: Multiple cells per thread
                        5 - PACK: uint64 Packet Coding 
  steps           Number of steps of the CA simulation 

Optional arguments:
  -h, --help      shows help message and exits 
  -v, --version   prints version information and exits 
  -g, --deviceId  Device ID [nargs=0..1] [default: 0]
  -d, --density   Density of the data domain 
  --seed          Seed for the random number generator [nargs=0..1] [default: 0]
  --doVerify      Verify the results? WARNING: memory requirements double 

Additional Information

All of the kernel code is in src/GPUKernels.cu. In order to measure power, the compile macro MEASURE_POWER must be set to a value other than NO.

Temporal research group from the Universidad Austral de Chile.