Godkända
Hardware-software model co-simulation for GPU IP development
Jaime Gancedo Rodrigo (2016)
Start
2017-09-11
Presentation
2018-04-06 11:15
Plats:
Avslutat:
2018-04-19
Examensrapport:
Sammanfattning
When developing hardware components, like the Mali GPU, software models are a key tool. In the case of the Mali GPU, the model is used to understand system on chip properties, including functionality and performance. We would like to investigate the possibility of a mixed environment where components in the GPU could be swapped between model (written in C++) and HW (written in Verilog). This would be beneficial to accelerate the development of a new GPU by allowing to run individual parts in a system context before the full system is available in HW and at faster emulation speed. It might also open up completely new use-cases. The objective of this thesis is to explore the feasibility of HW-model co-simulation and the potential benefits it could bring. As a starting point the proposal is to use the texture unit in the Mali GPU since the interface of this component is well aligned between the model and the HW implementations. It would be necessary to investigate the different options of communication between the C++ and the Verilog code and select the most appropriate one. The experiments should primarily cover having the HW component connected to a model of the rest of the GPU and collect detailed results comparing simulations on this setup with purely model runs. This would allow us to understand the effects on aspects such as simulation time and differences in GPU performance predictions (using performance counters).
Handledare: Reimar Döffinger (ARM) och Liang Liu (EIT)
Examinator: Erik Larsson (EIT)