The primitive pieces for genotype calling: A position, a base represented as four likelihoods, an inserted sequence, and the length of a deleted sequence. The logic is that we look at a base followed by some indel, and all those indels are combined into a single insertion and a single deletion.

Represents our knowledge about a certain base, which consists of the base itself (A,C,G,T, encoded as 0..3; no Ns), the quality score (anything that isn't A,C,G,T becomes A with quality 0), and a substitution matrix representing post-mortem but pre-sequencing substitutions.

Unfortunately, none of this can be rolled into something more simple, because damage and sequencing error behave so differently.

Decomposes a BAM record into chunks suitable for piling up. We pick apart the CIGAR and MD fields, and combine them with sequence and quality as appropriate. Clipped bases are removed/skipped as appropriate. We also do apply a substitution matrix to each base, which must be supplied along with the read.

Statistics about a genotype call. Probably only useful for fitlering (so not very useful), but we keep them because it's easy to track them.

Genotype likelihood values. A variant call consists of a position, some measure of qualities, genotype likelihood values, and a representation of variants. A note about the GL values: VCF would normalize them so that the smallest one becomes zero. We do not do that here, since we might want to compare raw values for a model test. We also store them in a Double to make arithmetics easier. Normalization is appropriate when converting to VCF.

If GL is given, we follow the same order used in VCF: "the ordering of genotypes for the likelihoods is given by: F(jk) = (k*(k+1)2)+j. In other words, for biallelic sites the ordering is: AA,AB,BB; for triallelic sites the ordering is: AA,AB,BB,AC,BC,CC, etc."

Map quality and a list of encountered bases, with damage information and reference base if known.

Map quality and a list of encountered indel variants. The deletion has the reference sequence, if known, an insertion has the inserted sequence with damage information.

Running pileup results in a series of piles. A Pile has the basic statistics of a VarCall, but no GL values and a pristine list of variants instead of a proper call. We emit one pile with two BasePiles (one for each strand) and one IndelPile (the one immediately following) at a time.

The pileup enumeratee takes BamRaws, decomposes them, interleaves the pieces appropriately, and generates Piles. The output will contain at most one BasePile and one IndelPile for each position, piles are sorted by position.

This top level driver receives BamRaws. Unaligned reads and duplicates are skipped (but not those merely failing quality checks). Processing stops when the first read with invalid br_rname is encountered or a t end of file.

The pileup logic keeps a current coordinate (just two integers) and two running queues: one of active PrimBases that contribute to current genotype calling and on of waiting PrimBases that will contribute at a later point.

Oppan continuation passing style! Not only is the CPS version of the state monad (we have five distinct pieces of state) somewhat faster, we also need CPS to interact with the mechanisms of Iteratee. It makes implementing yield, peek, and bump straight forward.

The things we drag along in PileM. Notes: * The active queue is a simple stack. We add at the front when we encounter reads, which reverses them. When traversing it, we traverse reads backwards, but since we accumulate the BasePile, it gets reversed back. The new active queue, however, is no longer reversed (as it should be). So after the traversal, we reverse it again. (Yes, it is harder to understand than using a proper deque type, but it is cheaper. There may not be much point in the reversing, though.)

Sends one piece of output downstream. You are not expected to understand how this works, but at last it doesn't leak memory.

Inspect next input element, if any. Returns Just b if b is the next input element, Nothing if no such element exists. Waits for more input if nothing is available immediately.

Discard next input element, if any. Does nothing if input has already ended. Waits for input to discard if nothing is available immediately.

The actual pileup algorithm. If active contains something, continue here. Else find the coordinate to continue from, which is the minimum of the next waiting coordinate and the next coordinate in input; if found, continue there, else we're all done.

Feeds input as long as it starts at the current position

Checks waiting queue. If there is anything waiting for the current position, moves it to active queue.

Scans active queue and makes a BasePile. Also sees what's next in the PrimChunks: Indels contribute to an IndelPile, Seeks and deletions are pushed back to the waiting queue, EndOfReads are removed, and everything else is added to the fresh active queue.

We need a simple priority queue. Here's a skew heap (specialized to strict Int priorities and PrimBase values).