I would be inclined to use a std::pair object that points to the two arrays. Pass the pair objects through the pipeline. The transform stage should take a std::pair as its input and return a pointer to the merged results as its output.

If you want the I/P filter to process the pairs of input arrays in parallel, the simplest way is to write a parallel loop that loops over i=0,1 and does one of the arrays for each index. TBB is good about supporting nested parallelism, so putting the parallel_for inside a pipeline stage shouldn't cause any surprises.

An alternative, if the I/P operation is the same for each array, is to pass arrays into the I/P stage, and then write a serial "pairing" stage that does pairing. The pairing stage would return either the paired arrays (say a std::pair) or a dummy object. Each time the pairing stage is run, it inspects whether it has previously accumulated an array. If so, it pairs it with the new incoming array and returns a std::pair. Otherwise it accumulates the array and returns a dummy object. Later stages of the pipeline would have to know to skip processing of dummy objects.