Résumé
Due to non-associativity of floating-point operations and dynamic scheduling on parallel architectures, getting a bit-wise reproducible floating-point result for multiple executions of the same code on different or even similar parallel architectures is challenging. In this paper, we address the problem of reproducibility in the context of matrix multiplication and propose an algorithm that yields both reproducible and accurate results. This algorithm is composed of two main stages: a filtering stage that uses fast vectorized floating-point expansions in conjunction with error-free transformations; an accumulation stage based on Kulisch long accumulators in a high-radix carry-save representation. Finally, we provide implementations and performance results in parallel environments like GPUs.