Speaker
Description
Parameter inference via unbinned maximum likelihood fits is a central technique in particle physics. The large data samples available at the HL-LHC and advanced statistical methods require highly efficient fitting solutions. I will present one such solution, MoreFit, and discuss in detail the optimization techniques employed to make it as efficient as possible. MoreFit is based on compute graphs that are automatically optimized and compiled just in time. The inherent parallelism of the likelihood can be exploited on a wide range of platforms: GPUs can be utilized through an OpenCL backend, CPUs through a backend based on LLVM and Clang for single- or multithreaded exectution, which in addition allows for SIMD vectorization. Finally, I will discuss the resulting performance using some illustrative benchmarks and compare with several other fitting frameworks.