1. Thermal-Aware Data Flow Analysis José L. Ayala – Complutense University (Spain) David Atienza – EPFL (Switzerland) Philip Brisk – EPFL (Switzerland)

2. Problem formulation  Thermal dissipation in semiconductor devices has a strong correlation with power consumption (leakage dependency) and reliability (absolute values and thermal gradients).  Register file (RF) is one of the hot spots in processor architectures.  Thermal-aware compilation has been proposed but requires post-compiled temperature estimation. Compiler can predict thermal state of the processor AT EVERY POINT IN THE CODE

3. Motivating example  Register assignment algorithms assign a register to a variable from those not assigned to interfering variables.  Traditional algorithms maintain an ordered list (FIFO) of registers  small set of registers is repeatedly selected.  Thermal profile of the RF is determined by a random assignment of registers, an ordered assignment or a “chessboard” assignment.

4. Data flow analysis (basics)  Useful to determine correctness and propose optimizations.  A single bit of information is propagated in liveness analysis.  An interval is propagated in bitwidth analysis.  The thermal state is a much more complex information to propagate: continuous function, floorplan dependent, accuracy depends on the granularity.

5. Data flow analysis (thermal)  Proposed forward analysis, in a single procedure, for the RF Do Boolean stop  True For each basic block B For each instruction I ε B, in forward order Estimate thermal state after I If change in I’s thermal state exceeds δ stop  False EndIf EndFor While (stop = False) Output thermal state of each instruction

6. Data flow analysis (thermal)  There is not an explicit way to guarantee convergence → very irregular data usage.  In case of non convergence, program code can be re-optimized or following compiler phases can be guided.

7. Data flow analysis (thermal)  Traditional approach  Thermal analysis is applied after register assignment.  Proposed approach  Application at earlier stages based on predictive analyses.  Development of a set of rules that relate compiler decisions with thermal state.  Compiler executes thermal-aware optimization based on previous analysis.

8. Data flow analysis (thermal)  Thermal-aware optimizations  Spreading in space  Spilling critical variables to memory  Splitting via copy insertion  Spreading in time  Via instruction scheduling  Using register promotion  Inserting NOP instructions

9. Preliminary Results  Leon3 Register File – MPEG-2 Decoder O R I G I N A L M O D I F I E D

10. Conclusions  Compilers can estimate thermal state in early stages of compilation;  Experimental results have shown thermal-aware optimization methods for the register file;  Research goal: to develop comprehensive data flow thermal analyses and rules for the processor modules;

11. Thank you! Questions?