ndaniels / mrfy-icfp Goto Github PK

Layout change for JFP.

Acccess using first-class functinos

Achieved in commit ceaa5c641ed36332170b2d3b02d9f825d9bcacf0 in the MRFY repo

It looks like the profiling story is really improved. I am happy to revise the section, but I need help figuring out what has happened and what to say.

want to reduce code duplication to see if things simplify

this should happen on a branch other than nr-jfp

Is it poss8ible to change the plan7 figure (currently Figure 2 in jfp.tex) so that it shows clearly which edges emit residues? I was thinking maybe a small black dot next to an edge if that edge emits using a match-state emission table, and a small black diamond if it uses an insert-state emission table.

A rule I'm hoping for is that the probability on any edge is stored in the node from which the edge originates. And the probability of any transition is the product of the edge probability with

• 1 if the edge bears no dot or diamond
• an emission-table lookup if the edge bears a dot or a diamond

The emission table should also live on the state from which the node originates.

I'd also like to use the greater space in the JFP paper to pull out those emission tables and connect them with nodes in a way that is not just another arrow.

I'd also like to pick one or two of those edge probabilities and label them with an expression, e.g., m_d (node j)

Then measure the performance gain vs the status quo ante ICFP

I'd like to have a simple way of benchmarking the performance of Viterbi that has two properties:

• It gives numbers that we can compare with C++ Viterbi
• It is simple enough that even tenured faculty can run it

What is the effect of the UNPACK pragmas

• On the ICFP branch on our current benchmarks?
• On Noah's experiments branch (and how was that measured)?
• On our current best combination?

If GHC can specialize the Viterbi code better than I can, that's something others might wish to know about.

Sorry about this.

The original excuse for C-ish types was that an array is easy to cut into pieces.
Let us use REPA-like projection tricks to chop a more sophisticated type into pieces.