Research On Compiler-directed Management Methods And Power Model For Scratchpad Memory

As the development of embedded technology, the SoC (System on Chip) has been widely applied to national defense electronics, automotive electronics, medical technology, consumer technology, wireless communication and power system which are highly important in national economy. The SoC requires smaller system size, longer working hours and higher stability, thus the energy consumption has been the main bottleneck that impedes its development. There comes a trend to replace Cache with SPM (Scratchpad Memory) to deal with the SoC's increasingly energy consumption. Centered on this theme, this thesis focused on studying compile-time scratchpad memory management method and related technology which are closely related to save storage subsystem's energy consumption and increase system's real-time performance.The main work and contributions are presented in the following aspects:(1) This thesis puts forward an architecture level SPM dynamic energy-consumption model according to actual circuit activity.All existing models can only provide SPM's average accessing energy-consumption and they can neither reflect program's actual input/output value nor SPM's content layout's influence on circuit's dynamic power-consumption, it limited the optimizing effect to decrease the SPM operation energy-consumption. Aiming at this problem, this thesis puts forward an architecture level SPM dynamic energy-consumption model according to actual circuit activity value which calculates the value by acquiring runtime circuit turnover status from performance simulator. This model effectively reflect the characteristic that energy-consumption changed as circuit activity value did. It may be helpful to guide the compiler to adopt proper optimizing technique to decrease energy-consumption by decreasing the circuit activity. Experiment shows that some runtime program can save 9% energy only by changing address encoding method to low down SPM address decoder circuit activity.(2) This thesis puts forward three methods for optimization of the low energy-consumption dynamic management:merging variable storage space to decrease runtime dynamic storage demand, calculating the optimizing-replacement-candidate-set to decrease the dynamic management cost, optimizing the SPM layout to reduce the circuit activity for increasing the energy-consumption optimizing effect.First, aiming at how to decrease the runtime storage demand, this thesis puts forward a local-variable storage space merging method based on function chain-correlation analysis. In those functions which are not chain-correlated, the local-variables are merged to share global storage space.This method can effectively decrease the total runtime storage space. Some optimized program's demand on storage space may cut 69% so that the unit average accessing frequency raise three times to original one.Second, current methods that avoid SPM management code failing may lead to high cost and limitation. Aiming at this problem, this thesis puts forward a new dynamic replacement candidate-set calculating method based on node execution sequence graph. Based on node-execution-sequence-graph generated that from extended flow chart, it chooses input/output position of the basic node as the replace operation's insert position between SPM and memory, then it calculates the replacement-candidate-set according to the characteristic that the node in the execution-sequence-graph. Example analyzing shows that, combined with object life cycle and node relationship in the execution-sequence-graph, the candidate-set calculation method can effectively decrease the replacement operations between SPM and memory, besides, it has no other limitation.Finally, considering that different storage object's position arrangement in SPM may affect circuit activity, furthermore, the SPM energy cost, this thesis puts forward a new method to optimize SPM layout which use genetic algorithm. It can effectively decrease the SPM circuit activity by optimizing storage object position in SPM, thus improving the energy consumption optimizing effect.Experiment shows that when SMP capacity is less than 8KB, the merging and replacement-candidate-set calculating method can effectively improve current method's poor optimizing performance. On the basis that existing optimization method can only save 30%-40% energy consumption, the new method can save 7%-20% more, some can do even up to 70%. Optimizing the storage object in SPM layout by genetic algorithm can furthermore save 6% energy consumption, and there comes a trend that the optimizing effect become more obvious as SPM capacity increased.(3) This thesis puts forward a SPM assignment algorithm based on tabu search and oriented program WCET value optimization.The embedded system improves its real-time performance by analyzing and optimizing the path's execution time under worst-case. Existing methods usually ignored that execution-path under worst-case may dynamic change so some paths may be over optimized and the global optimizing performance was affected. By analyzing program's WCET optimizing process and characteristic of SPM assignment result set, this thesis constructed an utility function which can reflect the expected value on the decreasing program's WCET value while variable assigned to SPM. Then the function was used to generate a screening method of candidate set in the assignment algorithm. Experiments show that the new allocation method which merges function-local-variable to share the storage space can achieve better optimization effect than the existing algorithm. Some program's WCET value optimization performance can improve about 5%.(4) This thesis constructs a compile-time SPM management prototype system according to SUIF system.By extending the SUIF (Stanford University Intermediate Format) system, this thesis constructed a compile-time SPM management prototype system, and the system can integrate the original one to share the achievements.