On which basis test are split during parallel execution?

Gradle docs say nothing about it, but I would like to know how the tests are split between forks in gradle when I specify maxParallelForks and --parallel options.