Context & Goals
Most kernels of interest in machine learning and high-performance computing manipulate sparse tensors. Sparse codes are highly irregular and make use of array indirections and dynamic control which jeopardize static automatic parallelization algorithms.
The overall objective of this postdoctoral fellowship is to investigate compiler and runtime algorithms to delay the specialization of the dense code at runtime when the sparse structure is known.
From a dense specification, we seek to compile a code able to specialize itself on the sparse input data. The specialization will involve a set of sub-computations, which are expected to be achievable by standard linear algebra routines (e.g. gemm), using state-of-the art linear algebra libraries. Several issues must be investigated:
- How to specialize the code? In particular, how propagate efficiently the sparsity along the computation flow?
- How to detect library kernels on the specialized code?
- How to enforce a proper scheduling for the parallel runtime?
Points 1 and 2 have been partially addressed by a PhD student.
The postdoctoral fellow will:
- Propose code optimizations and data structures for scaling sparse propagation (point 1)
- Address runtime scheduling (point 3) by relying on existing parallel runtimes
- Validate the complete approach (points 1, 2 and 3) on scientific benchmarks by using sparse tensors from the Florida sparse matrix collection as well as machine learning applications.
Environment & Salary
This position is funded by the French prioritary research program for exascale computing in France (PEPR NumPEx).
This postdoctoral fellowship will last 14 months and be held at Laboratoire de l’Informatique du Parallélisme at Ecole Normale Supérieure de Lyon, France.
The gross salary is 2788 euros/month.
Apply at :https://jobs.inria.fr/public/classic/en/offres/2026-10485
