array support

[bpowers]

roughly in priority order

• [x] lexer + parser support for subscripts
• [x] simulation support for single-dimensional arrays
• [x] show simulation results in diagram
• [ ] edit a2a equations
• [ ] edit subscripted equations
• [x] simulation support for multi-dimensional arrays
• [ ] lexer + parser support for transpose (') unary postfix operator
• [x] use dimension names in A2A equations
• [x] use Dimension.SubscriptName directly in equations to get the integer subscript value (1, 2, etc)
• [ ] Apply-to-all arrays with non-apply-to-all graphical functions
• [ ] Slicing -- A[1, *] and SUM(A[1:3])
• [ ] "Dimension positions"

[bpowers]

I had written a bunch of notes on sd.js a few years ago that are relevant (more or less)

• diagram changes (minimal)
• figure out if it makes sense to do single-dimensional first, or multi- dimensional from the start
• parser:
  • array reference/slicing
  • transpose operator
• semantic analysis/validation:
  • validate indexing
  • validate slicing
  • transpose using dimension names
  • transpose using positions
  • array slicing (A[1, *])
• optimization?
  • the simple thing to do is have a nest of for loops for each individual variable. This is simple, I worry it will be slow for large models (which are very common users of arrays). If we're doing operations on multiple variables with the same dimensions in a row, we can merge them into a single loop. This is straightforward logically, but there is no optimization framework in place yet, so that would need to be added.
• codegen:
  • apply-to-all equations
  • nested for loop
  • non-A2A equations
  • non-A2A graphical functions
  • array slicing (A[1, *])
  • transpose using dimension names
  • transpose using positions
• runtime:
  • know about defined dimensions + their subscripts
  • allocate correct amount of space for arrayed variables
  • array builtins: MIN, MEAN, MAX, RANK, SIZE, STDDEV, SUM.
  • be able to enumerate all subscripted values for CSV output
  • be able to return results for arrayed_variable[int_or_named_dimension]
  • right now, the runtime is pretty dead-simple. Every builtin function expects one or more numbers as input. With the array builtins, this is no longer the case.
  • index into arrays with non-constant offsets: constants[INT(RANDOM(1, SIZE(foods)))]
  • create slices of arrays: SUM(array[chosen_dim, *]), where chosen_dim is an auxiliary variable.
  • this means runtime type checking, and I think runtime memory allocation (right now memory is allocated once, in one chunk, when a simulation is created, which is fast and optimial).