The Block Vecchia Algorithm is a computational method for efficient spatial statistics calculations. This implementation is designed for high-performance computing environments, leveraging GPU acceleration and optimized libraries.
Our computational harness is built using:
- GCC version 10.2.0 (12.2.0)
- CUDA version 11.4 (11.8)
- Intel MKL 2022.2.1
- MAGMA 2.7.2
- GSL 2.6
- NLopt v2.7.1 optimization libraries
Ensure you have the following libraries installed and their paths included in your system's PATH and LD_LIBRARY_PATH:
- GCC 10.2.0 or later
- CUDA 11.4 or later
- Intel MKL 2022.2.1 or later
- MAGMA 2.7.0 or later
- GSL 2.6 or later
- NLopt v2.7.1 or later
The main executable for the Block Vecchia Algorithm is test_dvecchia_batch. Here are some example use cases:
To see all available options:
./bin/test_dvecchia_batch --help
To run a performance test and monitor execution time or calculate intermediate results in KL divergence:
./bin/test_dvecchia_batch --ikernel 1.5:0.1:0.5 --kernel univariate_matern_stationary_no_nugget --num_loc 20000 --perf --vecchia_cs 300 --vecchia_bc 1500 --knn --seed 0
To run the algorithm on simulated data or a real dataset:
./bin/test_dvecchia_batch --ikernel ?:?:? --vecchia_bc 300 --kernel univariate_matern_stationary_no_nugget --num_loc 20000 --vecchia_cs 150 --knn --xy_path /path/to/locations --obs_path /path/to/observations
Optional parameters:
--kernel_init 0.1:0.1:0.1 --tol 4 --omp_threads 40
Results are stored in the ./log file after execution.
This project is licensed under the terms of the Apache license. See the LICENSE file for details.