Abstract
The modeling of embedded systems attempts to estimate the performance and costs prior to the implementation. The early stage predictions for performance and power dissipation reduces the more costly late stage design modifications. Workload modeling is an approach where an abstract application is evaluated against an abstract architecture. The challenge in modeling is the balance between fidelity and simplicity, where fidelity refers to the correctness of the predictions and the simplicity relates to the simulation time of the model and its ease of comprehension for the developer. A model named GSLA for performance and power modeling is presented, which extends existing architecture modeling by including GPUs as parallel processing elements. The performance model showed an average fidelity of 93% and the power model demonstrated an average fidelity of 84% between the models and several application measurements. The GSLA model is very simple: only 2 parameters that can be obtained by automated scripts.
Besides the modeling, this thesis addresses lower level signal processing system improvements by proposing Instruction Set Architecture (ISA) extensions for RISC-V processors. A vehicle classifier neural network model was used as a case study, in which the benefit of Bit Manipulation Instructions (BMI) is shown. The result is a new PopCount instruction extension that is verified in ETISS simulator. The PopCount extension of RISC-V ISA showed a performance improvement of more than double for the vehicle classifier application. In addition, the design flow for adding a new instruction extension for a re-configurable platform is presented.
The GPU modeling and the RISC-V ISA extension added new features to the state of the art. They improve the modeling features as well as reduce the execution costs in signal processing platforms.
Besides the modeling, this thesis addresses lower level signal processing system improvements by proposing Instruction Set Architecture (ISA) extensions for RISC-V processors. A vehicle classifier neural network model was used as a case study, in which the benefit of Bit Manipulation Instructions (BMI) is shown. The result is a new PopCount instruction extension that is verified in ETISS simulator. The PopCount extension of RISC-V ISA showed a performance improvement of more than double for the vehicle classifier application. In addition, the design flow for adding a new instruction extension for a re-configurable platform is presented.
The GPU modeling and the RISC-V ISA extension added new features to the state of the art. They improve the modeling features as well as reduce the execution costs in signal processing platforms.
Original language | English |
---|---|
Place of Publication | Tampere |
Publisher | Tampere University |
ISBN (Electronic) | 978-952-03-2723-1 |
ISBN (Print) | 978-952-03-2722-4 |
Publication status | Published - 2023 |
Publication type | G5 Doctoral dissertation (articles) |
Publication series
Name | Tampere University Dissertations - Tampereen yliopiston väitöskirjat |
---|---|
Volume | 733 |
ISSN (Print) | 2489-9860 |
ISSN (Electronic) | 2490-0028 |