Thallo (https://www.thallo-lang.org/) is a new version of Opt (optlang.org), a new language in which a user simply writes energy functions over image- or graph-structured unknowns, and a compiler automatically generates state-of-the-art GPU optimization kernels. Real-world energy functions compile directly into highly optimized GPU solver implementations with performance competitive with the best published hand-tuned, application-specific GPU solvers.
Thallo introduces scheduling to this problem domain, as discussed in the accompanying publication: Thallo – Scheduling for High-Performance Large-scale Non-linear Least-Squares Solvers
This is a pre-alpha release of the software. An alpha release, with reproducible build instructions and scripts, docker images, and documentation is now slated for a September 2021 release.
As a pre-alpha release, there are no guarantees made; the alpha release will seek to support all popular runtime environments for modern Cuda (such as cloud V100 machines); it is known that some combinations of versions of LLVM/Terra/Cuda/GPUs/Driver Versions produce incorrect PTX; this will be resolved and fully documented in the alpha release.
Open an issue on (https://github.com/thallolang/thallo) if you see improvements that could be made that are not covered by the above.
Thallo is composed of a library libThallo.a and a header file Thallo.h. An application links Thallo and uses its API to define and solve optimization problems. Thallo's high-level energy functions behave like shaders in OpenGL. They are loaded as your application runs using the Thallo_ProblemDefine API.
See the Makefiles in the examples for instructions on how to link Thallo into your applications.