Investigating mixed-precision for AGATA pulse-shape analysis
Résumé
The AGATA project (1) aims at building a 4pi gamma-ray spectrometer consisting of 180 germanium crystals, each crystal being divided into 36 segments. Each gamma ray produces an electrical signal within several neighbouring segments, which is compared with a data base of reference signals, enabling to locate the interaction. This step is called Pulse-Shape Analysis (PSA).
In the execution chain leading to the PSA, we observe successive data conversions : the original 14 bits integers given by the electronics are finally converted to 32-bit floats. This made us wonder about the real numerical accuracy of the results, and investigate the use of shorter floats, with the hope to speedup the computation, and also reduce a major cache-miss problem previously identified with the Perf (2) tool.
Our proposed talk would first report about the numerical validation of the C++ PSA code, thanks to the Discrete Stochastic Arithmetic implemented in the CADNA library (3). After the code being properly instrumented, CADNA performs each computation three times with a random rounding mode. This allows, for each operation, to evaluate the number of exact significant digits using a Student test with 95% confidence threshold.
In a second step, we will report our successes and challenges while refactoring the code so to mix different numerical formats, using high precision only when necessary, and taking benefit of hardware speedup elsewhere. Such mixed-precision appears as a promising option for high performance computation in the next years, provided we use tools such as CADNA so to keep control of the accuracy of the computed results.
(1) https://www.agata.org/about
(2) https://perf.wiki.kernel.org/index.php/Main_Page
(3) http://cadna.lip6.fr
Origine : Fichiers produits par l'(les) auteur(s)