If you need to demultiplex based on barcodes within primer, you can try Checkout util (see docs). You'll need to specify a tab-delimited table with sample id, 5' barcode and reverse-complement of 3' barcode. As for the running time, I believe that I/O speed is a limiting factor here, not much to optimize.

CASAVA and bcl2fastq (both free downloads from Illumina) can handle dual barcodes. If you post the contents of the runParameters.xml file and your sample sheet, I can tell you what command line you should use.

EDIT:

Now that I know your updated requirements, it seems that in order to create a new FASTQ file for each unique barcode in a single script, you would need a lot of simultaneous filehandles open (or lots and lots and lots of RAM), both of which are problems on a desktop-grade machine if your distinct indices are as numerous as you claim.

Given that, I recommend iterating through the FASTQ data and dumping the data into a document/JSON storage database like CouchDB, RethinkDB, or the like. Basically, each unique barcode would have an array of FASTQ reads. Then, iterate through the indexes and dump the contents into separate FASTQ files. That will help you get around the RAM and filehandle hurdles.

BBMap has a tool for demultiplexing reads based on their barcodes, if you have them all in a fastq file with the barcode at the end of the read name. It's extremely fast.

demuxbyname.sh in=reads.fq out=%.fq suffix names=ACGTAC+GACTTG,ATATAT+CGCGCG

...where % gets replaced by the barcode, and names should list every literal barcode exactly as they appear in the read name, comma-delimited. If you have dual input files you can use in1, in2, out1, and out2. One (or two, for dual files) output file is created per barcode.