Effective CUDA code development for an open-source GPU hardware model
Parole chiave CUDA, GPU, PARALLEL PROGRAMMING
Riferimenti MATTEO SONZA REORDA
Gruppi di ricerca GR-05 - ELECTRONIC CAD & RELIABILITY GROUP - CAD
Descrizione GPUs are increasingly used in safety-critical embedded systems, e.g., in automotive to support the driver in Advanced Driver Assistance Systems (ADAS) and in robotics. Companies such as ARM, NVIDIA, Imagination Techn., MobilEye are very active in this area and are extremely interested in how to assess and improve their reliability with respect to both permanent and transient faults. Unfortunately, very few information are available about the internal structure and implementation of commercial devices. For this reason, a VHDL model (named FlexGripPlus) of a GPU which is compatible with some of the NVIDIA ones has been developed by the CAD Group of Politecnico di Torino, starting from a preliminary version made available by the University of Massachusetts (www.ecs.umass.edu/ece/tessier/andryc-fpt13.pdf). FlexGripPlus is currently being used in several research activities, which already demonstrated its suitability as a reference model, and led to several publications.
A thesis is proposed, to develop application and test code suitable to run on such a model. In particular, the development of CUDA applications whose execution can be simulated resorting to the model can lead to a better understanding of the GPU behavior, and allows code optimization. Moreover, the developed code can be effectively used to identify possible bugs in the VHDL code of the model, and/or possible faults existing in hardware implementations of the corresponding device.
Conoscenze richieste VHDL, CUDA
Scadenza validita proposta 18/06/2021 PROPONI LA TUA CANDIDATURA