Using serial dilutions of a representative sample mixture that is qualitatively similar and/or identical to your experimental samples is the most accurate way to examine amplification efficiencies (and is considered the gold standard approach by Vandesompele, Hellemans et. al). Purified PCR products and plasmids will give you different efficiencies for the same targets examined in your experimental samples since they don't have the additional "gunk" in them that your experimental sample isolates will. In addition, "context" in which the target is offered to the qPCR will have an impact on the onset of the PCR phase, thus also altering final Cq values.  The schools of thought on this differ.  But, I have always found that using a mixture made up of a portion of each of the experimental samples themselves, or linearized plasmids or PCR products carried through the same extraction process and/or spiked with the same background 'gunk" or "matrix" as the experimental samples, offers the best results every time when examining serial dilutions to obtain primer pair (or primer-probe set) amplification efficiencies. If purified [linearized] plasmid or purified PCR product is used for absolute quantification reasons, then a target standard curve efficiency determined from serial dilution of the purified template and the standard curve efficiency determined from the curve made of the biological samples can (and must) both be reconciled with one another mathematically using the formula in the attached publication (see equations (C), (D) and (E) on the 4th page of the attachment here).

Efficiency SHOULD be "per primer pair", i.e. you should only need to calculate it once (and since a popular method for doing so is serial dilution of purified PCR product or plasmid anyway, this is normally how it's done).

If you think you might have contaminants or inhibitors in some of your samples that make the efficiency for those reactions lower, then the answer is not to calculate separate efficiencies, it's: remove the contaminants.

Because of the exponential nature of PCR, efficiency differences can make a huge difference to apparent Cq value, so making sure all your samples behave comparably is definitely the way forward.

Using serial dilutions of a representative sample mixture that is qualitatively similar and/or identical to your experimental samples is the most accurate way to examine amplification efficiencies (and is considered the gold standard approach by Vandesompele, Hellemans et. al). Purified PCR products and plasmids will give you different efficiencies for the same targets examined in your experimental samples since they don't have the additional "gunk" in them that your experimental sample isolates will. In addition, "context" in which the target is offered to the qPCR will have an impact on the onset of the PCR phase, thus also altering final Cq values.  The schools of thought on this differ.  But, I have always found that using a mixture made up of a portion of each of the experimental samples themselves, or linearized plasmids or PCR products carried through the same extraction process and/or spiked with the same background 'gunk" or "matrix" as the experimental samples, offers the best results every time when examining serial dilutions to obtain primer pair (or primer-probe set) amplification efficiencies. If purified [linearized] plasmid or purified PCR product is used for absolute quantification reasons, then a target standard curve efficiency determined from serial dilution of the purified template and the standard curve efficiency determined from the curve made of the biological samples can (and must) both be reconciled with one another mathematically using the formula in the attached publication (see equations (C), (D) and (E) on the 4th page of the attachment here).

You can check the PCR efficiency of each curve, using software like LineRegPCR. It is good to know whether all curves have a " minimum acceptable" efficiency, before venturing into detailed interpretations based on Cq.