Modern GPU applications are composed of several kernels. We introduce the concept of Kernel Vulnerability Factor (KVF), which indicates the probability of faults in a kernel to affect computation. We apply KVF to Histogram of Oriented Gradients (HOG) algorithm, which is the base of many pedestrian detection systems. We measure the KVF of HOG with both architectural-level and high-level fault-injection. We also qualify the corrupted outputs distinguishing between tolerable and critical errors. By identifying the HOG portions which are more prone to affect detection we propose an efficient hardening technique able to detect 85% of critical errors with an overhead as low as 11.8%.
Kernel vulnerability factor and efficient hardening for histogram of oriented gradients / Weigel, L.; Fernandes, F.; Navaux, P.; Rech, P.. - 2018-:(2017), pp. 1-6. (Intervento presentato al convegno 13th IEEE International Symposium on Defect and Fault Tolerance in VLSI and Nanotechnology Systems, DFT 2017 tenutosi a gbr nel 2017) [10.1109/DFT.2017.8244439].
Kernel vulnerability factor and efficient hardening for histogram of oriented gradients
Rech P.
2017-01-01
Abstract
Modern GPU applications are composed of several kernels. We introduce the concept of Kernel Vulnerability Factor (KVF), which indicates the probability of faults in a kernel to affect computation. We apply KVF to Histogram of Oriented Gradients (HOG) algorithm, which is the base of many pedestrian detection systems. We measure the KVF of HOG with both architectural-level and high-level fault-injection. We also qualify the corrupted outputs distinguishing between tolerable and critical errors. By identifying the HOG portions which are more prone to affect detection we propose an efficient hardening technique able to detect 85% of critical errors with an overhead as low as 11.8%.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione
