One column of raw voltage values is probably okay, and sampling rate doesn't matter as long as it's high enough and fairly constant. As Rick pointed out, if your EPI data are, for example, 200 s long then your physio data should be 200 s long as well, and in temporal alignment.

You don't need to do any peak detection beforehand as 3dretroicor has a peak detector for pulse data built in. It's a very simplistic peak detection algorithm: find a maximum above a threshold. Basically, if you can plot your pulse data and draw a straight horizontal line through it that isolates all the peaks from each other and from the noise below, you're good. If not, some transformation may be necessary, e.g. detrending, or even your own peak detector. I usually use the scanner's pulse trigger output rather than raw pulseox data, so I don't have that problem--everything works with the defaults.

You are doing the right thing by comparing the cardiac phase output with the original cardiac/pulse data. From your description, it sounds like the pulse peaks are not being detected. Note that the threshold for peak detection defaults to 1. This is unlikely to work with raw pulse data in general (although it's often okay with pulse triggers, which usually alternate between 0 and some positive number). Again, plot your pulse data to find a good value for the threshold. I also suggest you normalize the raw data so that the values are more consistent across runs and subjects, and you won't have to change the threshold as much.

Adding to what Ziad wrote, RetroTS.m has the added advantage of calculating additional regressors beyond those used in 3dretroicor. However, 3dretroicor is quick and convenient to use, e.g. you don't even need to do regression, so it fits easily into any pipeline. You may even want to try using both RetroTS.m and 3dretroicor.

Fred