We investigate applying general-purpose join algorithms to the triangle listing problem on heterogeneous systems that feature a multi-core CPU and multiple GPUs. In particular, we consider an out-of-core context where graph data are available on secondary storage such as a solid-state disk (SSD) and may not fit in the CPU main memory or GPU device memory. We focus on Leapfrog Triejoin (LFTJ), a recently proposed, worst-case optimal algorithm and present "boxing": a novel yet conceptually simple approach for partitioning and feeding out-of-core input data to LFTJ. The "boxing" algorithm integrates well with a GPU-Optimized LFTJ algorithm for triangle listing. We achieve significant performance gains on a heterogeneous system comprised of GPUs and CPU by utilizing the massive-parallel computation capability of GPUs. Our experimental evaluations on real-world and synthetic data sets (some of which containing more than 1.2 billion edges) show that out-of-core LFTJ is competitive with specialized graph algorithms for triangle listing. By using one or two GPUs, we achieve additional speedups on the same graphs.