All of your weighted projections should follow a formula such as the CPI example, where the metric to be weighted must be defined in terms of samples per instruction.  With CPI, this is cycles per instruction, and with your example you can use Energy Per Instruction.  This is necessary, because the weighting is computed based on the SimPoint sample space, which is based on instructions:


However you say that you also want to compute projected metrics for a cumulative stat such as overall Energy.  To extend the CPI example, let's propose that you wish to compute the projected overall cycle time of a simulation.  From the projected CPI you can simply multiply by the number of dynamic instructions in the entire benchmark to get the number of cycles projected to execute.

To summarize for your case: you will need a means of measuring Energy-Per-Instruction, Energy-Delay-Product-Per-Instruction and Energy-Delay-squared-Product-Per-Instruction.  For some of these you may need to base your sample over a region of instructions and then divide by the number of instructions to get a normalized metric.  Then you use the weighting formula to get an overall metric that is weighted to project the behavior of your entire program.  Last, multiply by the number of dynamic instructions to get the projected metric desired for your entire benchmark.

I hope this clarifies things,
Robert