A Scheme replacement for Guile's array system.
lloda 1439f5b089 Relax FindGuile.cmake | 1 년 전 | |
---|---|---|
.github | 1 년 전 | |
bench | 1 년 전 | |
config | 1 년 전 | |
docs | 1 년 전 | |
examples | 1 년 전 | |
mod | 1 년 전 | |
sandbox | 1 년 전 | |
test | 1 년 전 | |
CMakeLists.txt | 1 년 전 | |
LICENSE | 7 년 전 | |
README.md | 1 년 전 | |
TODO | 1 년 전 |
guile-newra (newra
) wants to replace the current (3.0) Guile array system, which is almost entirely implemented in C.
The new implementation should be at least as fast. I think this is feasible once the Scheme compiler goes native, because for the most part the array functions are used to call back to Scheme, and a Scheme implementation could get rid of the back and forth, optimize the type dispatches, etc.
The C API shouldn't be affected. Once you get an array handle it makes no sense to use the Scheme array functions anyway, since the memory layout is transparent.
Except for the tests and for the pair of functions ra->array
/ array->ra
, newra
is independent of the old array system.
newra
requires Guile 3.0.8. Run the tests with
> cmake . && make test
or manually
> $GUILE -L mod test/everything.scm
> ...
The manual is at lloda.github.io/guile-newra, and you can find some larger examples in examples/
.
To install the library, put mod/newra
and mod/newra.scm
somewhere in your Guile load path, and use it with (import (newra))
.
newra
can use guile-ffi-blis
for some functions (ra-fill!
in the current version), to provide a considerable speed up for operations with arrays of types s32
, u32
, f32
, s64
, u64
, f64
, c32
, or c64
. It can also be slower — there isn't a good heuristic yet. Run the benchmarks with
> $GUILE -L mod bench/bench.scm
> $GUILE -L mod -L path-to-guile-ffi-blis bench/bench-blis.scm
The old array compatibility layer is mostly finished, with only a naming change (array-xxx
becomes ra-xxx
). The newra
versions of the map
and for-each
functions are significantly faster already, but the -ref
/ -set!
functions are a bit slower, and some of the functions that have a fast path in C, such as ra->list!
, can be several times slower in newra
, depending on the types of the arguments.
These issues seem fixable, and besides, the Scheme compiler is only improving.
Compared with the old arrays, newra
offers many features:
(define ra (list->ra 2 '((1 2) (3 4))))
, (ra 1 1)
returns 4
and (ra 0)
returns #%1(1 2)
.(set! ((make-ra #f 2 3) 1 1) 99)
returns #%2:2:3((#f #f #f) (#f 99 #f))
.ra-iota
, ra-i
). These may be infinite: ((ra-iota #f 1) (- #e1e20 1))
returns 100000000000000000000
.(ra-map! (make-ra 'x 2 3) + (ra-i 2 3) (ra-iota 2 0 10))
returns #%2:2:3((0 1 2) (13 14 15))
. Prefix matching supports undefined dimensions; the previous expression and (ra-map! (make-ra 'x 2 3) + (ra-i #f 3) (ra-iota #f 0 10))
are equivalent.(define I (ra-iota))
and (define J (ra-transpose (ra-iota) 1))
, then (ra-map! (make-ra 'x 10 10) * I J)
is a multiplication table.ra-from
, ra-amend!
: index arguments can have any rank, and use of lazy index vectors (of any rank!) results in a shared array. A stretch index object (dots)
is supported; e.g. (ra-from A (dots) 0)
(or just (A (dots) 0)
) will produce the slice A[..., 0]
for an array of any rank.ra-cat
, ra-cats
).ra-reverse
, ra-any
, ra-every
, ra-fold
, ra-ravel
, ra-reshape
, ra-tile
.ra-format
).newra
is written entirely in Scheme, if a newra
operation takes too long, you can actually interrupt it, which is not always the case in the old system.Originally I wanted newra
to be a drop-in replacement for the old array system, reusing the same function names and all. Now I think it's better to have a parallel system where some of the flaws of old system can be cleaned up. Still it's important that programs can be easily ported to the new system.
With that in mind, here is what you'd have to change. Note that the ra-
names are not final, and neither is the #%
read syntax. I'm not sure yet how the old array syntax will be absorbed — maybe old array objects will be converted transparently for a while. Some of these are bugs that will eventually be fixed.
The functions ra->array
and array->ra
are provided to convert between the old and the new array types. Neither of these functions copy the contents of the array, so (let ((o (make-array 3))) (eq? (shared-array-root o) (shared-array-root (ra->array (array->ra o)))))
returns #t
. Note that some of the new ra
types aren't convertible in this manner; for example (ra->array (ra-iota 3))
is an error.
The new system matches sizes strictly. For instance (array-map! (make-array #f 2) - (make-array #t 3))
succeeds, but (ra-map! (make-ra #f 2) - (make-ra #t 3))
fails with a shape mismatch error.
The new system still supports non-zero base indices, but I'd advise against using them.
For most of the old functions array-xxx
, the equivalent function in newra
is ra-xxx
. Exceptions:
shared-array-root
is ra-root
.shared-array-offset
is ra-offset
.make-shared-array
is make-ra-shared
.transpose-array
is ra-transpose
.(array-copy! src dst)
is (ra-copy! dst src)
. This follows array-map!
/ ra-map!
and array-fill!
/ ra-fill!
which both use the first argument as destination.Most ra-
functions try to return something useful even when the corresponding array-
functions do not. For example (array-fill! (make-array 0 3) 4)
returns *unspecified*
, but (ra-fill! (make-ra 0 3) 4)
returns #%1:3(4 4 4)
.
The default writer defaults to printing all the sizes of the array, so (ra-i 3 2)
prints as #%2:3:2((0 1) (2 3) (4 5))
. Note that #2:3:2((0 1) (2 3) (4 5))
is a valid read syntax in the old system, just not the default.
The read syntax is like that of the old system except for an extra %
, so #2f64((1 2) (3 4))
becomes #%2f64((1 2) (3 4))
. The compiler doesn't support the new literal type yet. You can work around this using the reader like (call-with-input-string "#%2f64((1 2) (3 4))" read)
or say (list->ra 'f64 2 '((1 2) (3 4)))
.
The default printer and reader don't handle undefined size arrays as well as they could. For example (ra-transpose (ra-i 2) 1)
prints as #%2:d:2((0 1))
, but this cannot be read back. (ra-iota #f)
prints as #%d1:f
.
truncated-print
doesn't support newra
types, so you'll get a lone #
if truncation is necessary at all.
equal?
doesn't support newra
types, so it does just eqv?
. Instead, you can use ra-equal?
.
See the manual for another set of references.