TODO

(newra) -*- mode: org; -*-

Help: [C-uc .] insert timestamp [C-cc] flip checkbox [C-uucc] partial flip checkbox [C-ct] flip TODO.

bugs [9/10]

• [X] (call-with-input-string “#3()” read) vs (call-with-input-string “#%3()” read)
• [X] reader loses type when reading rank-0 arrays
• [X] ra-equal? errors out with mismatched shapes
• [X] zero lengths break ra-format
• [X] (ra-dimensions (ra-iota))
• [X] (ra-reshape (ra-iota) …) doesn’t work
• [X] (ra-from (list->ra 2 ‘((a b c) (x y z))) 1 1) => #%0(y) contradicts doc
  • Resolved in 792491e163794d2788b049b071d6cd0740d59c92 <2023-02-16 Thu 09:33>
• [X] ra-slice-for-each shouldn’t return anything
  • Fixed with explicit (values) <2023-02-16 Thu 09:42>
• [X] define-inlinable-case doesn’t inline case-lambda forms
  • Fixed with a patch to guile:module/language/tree-il/peval (3b47f87618047ebb8812788c64a44877a4f2e0dd)
• [ ] docstrings don’t work for define-inlinable-case

documentation [1/2]

• [X] actually explain rank extension / prefix matching in the manual
• [ ] manual lacks a hands on intro ‘first (import (newra)) then etc’

compatibility with existing Guile arrays [8/9]

• [X] ra-slice-for-each ra-map! ra-for-each ra-cell ra-ref ra-set!
• [X] ra-transpose ra-reverse
• [X] make-ra make-ra-shared ra->list list->ra list->typed-ra
• [X] ra-iota ra-i
• [X] ra-index-map!
• [X] ra-shape ra-dimensions
• [X] ra-root ra-offset
• [X] ra-equal?
• [ ] root vectors are equivalent to rank-1 newra objects
• [ ] return roots for trivial rank-1 views on roots

whole array functions [9/15]

• [X] ra-fold
• [X] ra-any ra-every
• [X] raw prefix matching
• [X] generalized prefix matching
• [-] high level slicing [3/4]
  • [X] ra-from
  • [X] ra-amend!
  • [X] (dots)
  • [ ] axis insertion (?)
• [X] ra-ravel
  • [X] (ra-ravel a [n])
  • [X] (ra-order-c? a [n [org]])
  • [X] (ra-ravel a [n [org]])
• [X] ra-tile [2/2]
  • [X] accept arbitrary axis
  • [X] can produce dead axes
• [X] ra-cat (ra-cat, ra-cats)
  • [X] implementation
  • [X] tests
  • [X] tests with non-zero base indices
• [X] ra-reshape [3/3]
  • [X] basic function
  • [X] accept arbitrary axis
  • [X] placeholder
• [X] ra-clip
• [ ] ra-grade
• [-] ra-rotate [1/2]
  • [X] ra-rotate!
  • [ ] ra-rotate that isn’t just (ra-rotate! n (ra-copy x))
• [ ] as-ra
• [ ] ra-filter, ra-select, etc.
• [ ] ra-affine map that explicitly takes an [M×(N+1)] matrix

misc array manipulation [0/1]

• [ ] mapping of indices (beyond ra-reshape/ra-from). Maybe just explain in manual?

fundamental features [3/7]

• [X] unsized index vectors
• [X] tensor indices
• [X] broadcasting
• [ ] reader handles quasiquote
• [ ] ways to control iteration (in order, dim loop order, direction, etc.)
• [ ] a solution for (ra-amend! dst (RA-MAP OP (ra-from dst i …)) i …). Lenses?
• [ ] separate somehow the rank matching cases that can be done in parallel from those that cannot, e.g. using lower rank dest arg to perform reductions.

array language [/]

• [ ] drag-along facility / lazy ops
  • [ ] verbs
  • [ ] rank conjunction
  • [ ] each (cf array-curry in srfi-179 or the old Guile array-enclose).
• [ ] output types

performance [2/4]

• [X] u8 ra-fill! vs native (see bench.scm) ok in 3.0.8 with partial bytevector-fill!
• [X] f64 ra-copy! vs native (see bench.scm)
• [ ] ra-ref / ra-set! with 1 or 2 args vs built-in array-ref / array-set!
• [ ] ra->list / list->ra vs built-in array->list / list->array. Maybe those can be reused?

friendlier undefined sizes [2/6]

• [X] printing of ‘d typed arrays
• [X] printing with dead axes
• [ ] reading of ‘d typed arrays
• [ ] printing with infinite axes
• [ ] reading of arrays with inf axes
• [ ] reading of arrays with dead axes

quality of life [2/15]

• [X] make (ra …) work with generalized slices just like (ra-from …).
• [ ] make (set! (ra …) o) work with slices (now) or generalized slices (eventually) just like (ra …).
• [X] rank-1 ops in the unrolled section (e.g. bytevector-copy!, bytevector-fill!).
• [ ] precompute fixed parameters of rank-1 ops (len, step).
  • These can be used to improve the rank-1 op choice heuristic, and also not to rely on iteration order assumption.
• [ ] rank-2 ops or higher in the unrolled section (e.g. blis-dcopym!, blis-dsetm!).
• [ ] rank-1 (or higher!) cases for generating functions (e.g. array->list).
• [ ] the reverse of ra-singletonize, find a good name (see ‘differences with numpy’ in the doc).
• [ ] automatic handling of scalars (e.g. (ra-map #f + (ra-iota 3) 99) works).
• [ ] refactor macros
• [ ] bench tracking
• [ ] move struct defs to (newra types)
• [ ] ra-unique-elements?
• [-] pretty printer [4/6]
  • [X] basic ra-format
  • [X] print sizes in prefix
  • [X] CL/SRFI-163 style rank-0 (e.g. ‘#0 x’ instead of ‘#0(x)’).
  • [X] compact mode
  • [ ] handle unbounded axes
  • [ ] auto truncation
• [-] an ffi to Fortran based on iso_c_binding, be able to pass arrays & scalars at least [3/6]
  • [X] fix rank 0
  • [X] fix non-zero ra-offset
  • [X] verify behavior of lbounds
  • [-] fortran-library-function [3/7]
    • [X] c32/c64
    • [X] fix arg-types format
    • [X] working example with intent(inout)
    • [ ] support intent(out) array args (need array-mutable?)
    • [ ] logical <-> bool (need libffi / Guile ffi support)

    [ ] value, arguments, like real(C_FLOAT), value

    x

    • [ ] support out args, e.g. with boxes (?)
  • [ ] fix alignment assumptions (seems there’s padding in some versions of ISO_Fortran_binding.h :-/)
  • [ ] modulize, test, doc, etc.
• [ ] connections to other libraries
  • [ ] BLIS
    • [ ] fix unbound variable compilation warnings when (import (ffi blis)) fails.
  • [ ] fftw

other [0/1]

• [ ] review definition of (ra-untranspose a i …). Does it make sense to allow any i < (rank a) ?

packaging [1/3]

• [X] Global CMakeLists.txt
• [ ] Autoconf (?)
• [ ] Installation

replace Guile arrays [0/5]

• [ ] truncated-print support
• [ ] equal? support
• [ ] hooks into C / libguile
• [ ] compilation of literals
• [ ] replace all names