RFC: Refactor indexing to use Varargs and ReshapedArrays #16251

mbauman · 2016-05-07T16:25:20Z

This puts some of the amazing new infrastructure through its paces, on our way towards #14770. Using a combination of Vararg and ReshapedArrays, we can always ensure that scalar, non-scalar, and view indexing all get indexed at an optimal dimensionality. This is slightly more verbose than would otherwise be necessary — it explicitly defines separate methods to support the general linear indexing cases that are on their way towards deprecation.

TODO:

@nanosoldier runbenchmarks(ALL)
Bikeshed the reshape(A, Val{N}) method. The current ~~implementation~~ name is just a stop-gap… as I was completing this I was thinking about doing something like this:

typealias Ravel{N, T} ReshapedArray{T, N}
(::Ravel{N}){N}(::AbstractArray) = # do the reshaping to N dimensions

timholy · 2016-05-07T20:55:41Z

base/abstractarray.jl

+## LinearFast Scalar indexing: canonical method is one Int
+_getindex(::LinearFast, A::AbstractArray, ::Int) = error("indexing not defined for ", typeof(A))
+_getindex(::LinearFast, A::AbstractArray, i::Real) = (@_propagate_inbounds_meta; getindex(A, to_index(i)))
+function _getindex{T,N}(::LinearFast, A::AbstractArray{T,N}, I::Vararg{Real,N})


My impression is that this is still slow without @inline, because of the lack of call specialization (which will hopefully happen someday, see #16159).

Oh, nvm, next line 😦

timholy · 2016-05-07T21:07:33Z

An awe-inspiring simplification---I really like how you've used the new reshape to such tremendously good effect. 💯

mbauman · 2016-05-07T21:17:09Z

I'm still having difficulty getting SubArray to play along with this whole scheme. I think I finally got something that will pass tests now… it'll be very interesting to see what it's performance is like, since we end up with ridiculously complicated nested structures from taking linear views into SubArrays.

mbauman · 2016-05-07T21:39:21Z

Does Nanosoldier hear me from within a bulleted list? Let's try it on a line by itself:

@nanosoldier runbenchmarks(ALL, vs = "JuliaLang/julia:master")

jrevels · 2016-05-07T22:59:26Z

Sorry, Nanosoldier isn't back online yet. I'm bringing it back up tomorrow, I'll trigger the build here once it's available.

timholy · 2016-05-07T23:03:19Z

Looks like you're almost there. The BitArray tests are definitely hard-core when it comes to specifying the return types. The same issue came up in #15449.

…lity

mbauman · 2016-05-07T23:15:44Z

Yup, it's a really wonderful catch and a simple fix — we don't want to let the reshaped view escape! I've cleaned things up so tests should pass at every stage along the way (despite how GitHub wants to display the order of the commits).

Thanks @jrevels. In the meantime, I did a quick run of test/perf/array.jl… and it's looking quite good. The huge wins are linear indexing into linear slow arrays, and there's a big win for logical indexing into BitArray, too. No tests are more than 20% worse, and there's not an obvious pattern to the 17 that are more than 10% worse.

mbauman · 2016-05-08T03:58:29Z

Any thoughts on the spelling of the reshape(A, Val{N}) API? No matter what, we need to pass the number of dimensions as some sort of type parameter… but should we use Val for this API? We already have a whole slew of types that have dimension parameters — could we use the N in some array type?

convert(AbstractArray{T,N}, …) is already pretty overloaded, and it has the downside that you'd need to use a special typealias to just express AbstractDimsFirstArray{N} in any case.
Numpy uses the verb ravel for this operation (well, the flattening to 1-dimension part). We could define typealias Ravel{N,T} ReshapedArray{T,N}, and then either use convert(Ravel{N}, A) or Ravel{N}(A). It is a little strange, though, since it's not guaranteed to return a ReshapedArray.
Or maybe we just define and export an empty immutable Ravel{N} end type just for this task. But again it's odd to use construction or conversion here since you're doing neither.
Or just stick with reshape(A, Val{N}). It's simple, it works, and it has no extra exports.

timholy · 2016-05-08T10:10:58Z

I don't like the convert option because it's not entirely clear that this is non-copying. Either the Val{N} or the Ravel{N} seem the best choices to me, and I don't have strong feelings about which is better.

timholy · 2016-05-08T10:35:03Z

Another option is reshape(A, ntuple(d->Colon(), Val{N})), although that's a lot to force users to type just to reshape to N dimensions.

jrevels · 2016-05-09T12:45:19Z

The maiden voyage of the new Nanosoldier configuration:

@nanosoldier runbenchmarks(ALL, vs = ":master")

nanosoldier · 2016-05-09T15:19:23Z

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

jrevels · 2016-05-09T15:28:25Z

Seems like I need to raise tolerances for the Big* scalar benchmarks, but those indexing improvements are pretty awesome!

EDIT: Seems like the status icon didn't update even though the job is finished - I'll have to look into that...

timholy · 2016-05-09T15:49:28Z

@jrevels, that also seems a lot faster than I remember it. If so, nice!

jrevels · 2016-05-09T16:03:18Z

@jrevels, that also seems a lot faster than I remember it. If so, nice!

Thanks! The new execution strategy we're using has cut down the runtime for a single pass of the full suite from ~4 hours to around ~30 minutes; expected turnaround time for full PR builds like this one is down from ~11 hours to ~3 hours.

mbauman · 2016-05-09T16:57:20Z

Thanks so much @jrevels, this is great.

I'm okay with the current reshape(A, Val{N}) API — it's un-opinionated and is easy to change in the future. This is ready to go as far as I'm concerned. Any other comments?

tkelman · 2016-05-09T18:47:47Z

base/subarray.jl

    else
-        :(@_propagate_inbounds_meta; (merge_indexes(V, idxs, subidxs),))
+        :(error("mismatched index dimensionalities"))


any way of making this more descriptive if it ever happens?

Would it be better to call this "rank" instead of "dimensionality"?

Yeah, I should probably enforce this in the constructor. The whole point here is to ensure that the number of indices stored in the SubArray matches the number of dimensions in the parent array. Even if you do manage to create a corrupted SubArray, though, you'd almost certainly hit a missing method in the other ~~reshape~~ reindex signatures long before you ever get here.

We've typically stayed away from using the word rank to mean ndims because of rank.

and add a better error message

vtjnash · 2016-05-12T19:15:32Z

I seem to be consistently getting a failure in the core.jl test after this merged. It sometimes goes away with inlining, and can be prevented with --inline=no but the code is all crazy & missing a bounds check:

julia> [][] # should throw a bounds error
ERROR: UndefRefError: access to undefined reference
 in getindex(::Array{Any,1}) at abstractarray.jl:949
 in eval(::Module, ::Any) at ./boot.jl:226

julia> @code_typed [][]
1-element Array{Any,1}:
 LambdaInfo for getindex
:(begin  # abstractarray.jl, line 424:
        $(Expr(:meta, :propagate_inbounds)) # abstractarray.jl, line 425:
        $(Expr(:inbounds, false)) # abstractarray.jl, line 447: # abstractarray.jl, line 448: # abstractarray.jl, line 449:
        $(Expr(:boundscheck, true))
        true
        $(Expr(:boundscheck, :pop)) # abstractarray.jl, line 450:
        $(Expr(:inbounds, true))
        $(Expr(:inbounds, false))
        (Base.arraysize)(A,1)::Int64 # abstractarray.jl, line 949:
        $(Expr(:inbounds, :pop))
        r = (Base.arrayref)(A,1)
        $(Expr(:inbounds, :pop)) # abstractarray.jl, line 451:
        $(Expr(:inbounds, :pop))
        return r
    end)

vtjnash · 2016-05-12T19:20:11Z

Ah, wait. Maybe it's not the direct fault of this PR. The zero-arg bounds check test is wrong:

base/abstractarray.jl
 138 _internal_checkbounds(A::AbstractArray) = true

timholy · 2016-05-12T19:32:31Z

Where does that go wrong?

(I ask because I'm in the middle of revamping checkbounds yet again.)

mbauman · 2016-05-12T19:35:51Z

A must have at least one element for A[] to work.

This is unrelated, but It's unfortunately another case of a probably-should-be-ambiguous dispatch (#11242 (comment)):

julia> @which Base._getindex(Base.LinearFast(), A)
_getindex(::Base.LinearFast, A::AbstractArray, I::Real...) at abstractarray.jl:447
# I'd expect:
# _getindex{T,N}(::LinearFast, A::AbstractArray{T,N}, I::Vararg{Real,N}) at abstractarray.jl:438

timholy · 2016-05-12T22:23:28Z

Oh, duh, right.

vtjnash · 2016-05-13T02:11:43Z

the "obvious" fix seems to work for, is there any reason to do something more complicated:

diff --git a/base/abstractarray.jl b/base/abstractarray.jl
index 0972cae..6522bc3 100644
--- a/base/abstractarray.jl
+++ b/base/abstractarray.jl
@@ -135,7 +135,7 @@ throw_boundserror(A, I) = (@_noinline_meta; throw(BoundsError(A, I)))
 checkbounds(A::AbstractArray, I...) = (@_inline_meta; _internal_checkbounds(A, I...))
 # The internal function is named _internal_checkbounds since there had been a
 # _checkbounds previously that meant something different.
-_internal_checkbounds(A::AbstractArray) = true
+_internal_checkbounds(A::AbstractArray) = _internal_checkbounds(A,1)
 _internal_checkbounds(A::AbstractArray, I::AbstractArray{Bool}) = size(A) == size(I) || throw_boundserror(A, I)
 _internal_checkbounds(A::AbstractArray, I::AbstractVector{Bool}) = length(A) == length(I) || throw_boundserror(A, I)
 function _internal_checkbounds(A::AbstractArray, I1, I...)

timholy · 2016-05-13T03:13:13Z

LGTM. Presumably could also just check !isempty.

mbauman force-pushed the mb/varargs branch 2 times, most recently from 37233cd to d2ac8f6 Compare May 7, 2016 19:05

timholy reviewed May 7, 2016
View reviewed changes

mbauman force-pushed the mb/varargs branch from d2ac8f6 to c2499b1 Compare May 7, 2016 21:22

kshyatt added the arrays [a, r, r, a, y, s] label May 7, 2016

mbauman added 5 commits May 7, 2016 19:06

Fix testing setindex! -- make sure we are using the proper dimensiona…

e082468

…lity

Refactor abstract scalar indexing to use Varargs

3440bf5

Refactor abstract nonscalar indexing to use Varargs

51b4bc1

Refactor SubArray indexing to use Varargs

7f2d845

Rewrite array interface doc in terms of Vararg

a6954cd

mbauman force-pushed the mb/varargs branch from c2499b1 to a6954cd Compare May 7, 2016 23:06

Use a lispy recursive function for reshape(A, Val{N})

a071314

mbauman mentioned this pull request May 9, 2016

indices or indexes? #12902

Closed

tkelman reviewed May 9, 2016
View reviewed changes

mbauman added 2 commits May 10, 2016 18:23

Enforce index/parent consistency in constructor

bd38ee9

and add a better error message

Use dispatch to avoid a branch in first(::Tuple)

e5fec99

tkelman merged commit 2c16e98 into master May 12, 2016

tkelman deleted the mb/varargs branch May 12, 2016 16:16

tkelman mentioned this pull request May 12, 2016

Change llvmcall docs binding #16332

Merged

tkelman mentioned this pull request May 20, 2016

deprecate (then remove) generalized linear indexing #14770

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

RFC: Refactor indexing to use Varargs and ReshapedArrays #16251

RFC: Refactor indexing to use Varargs and ReshapedArrays #16251

mbauman commented May 7, 2016 •

edited

Loading

timholy May 7, 2016

timholy May 7, 2016

timholy commented May 7, 2016

mbauman commented May 7, 2016

mbauman commented May 7, 2016

jrevels commented May 7, 2016

timholy commented May 7, 2016

mbauman commented May 7, 2016 •

edited

Loading

mbauman commented May 8, 2016 •

edited

Loading

timholy commented May 8, 2016

timholy commented May 8, 2016

jrevels commented May 9, 2016

nanosoldier commented May 9, 2016

jrevels commented May 9, 2016 •

edited

Loading

timholy commented May 9, 2016

jrevels commented May 9, 2016 •

edited

Loading

mbauman commented May 9, 2016

tkelman May 9, 2016

eschnett May 9, 2016

mbauman May 9, 2016 •

edited

Loading

vtjnash commented May 12, 2016 •

edited

Loading

vtjnash commented May 12, 2016

timholy commented May 12, 2016

mbauman commented May 12, 2016

timholy commented May 12, 2016

vtjnash commented May 13, 2016

timholy commented May 13, 2016

RFC: Refactor indexing to use Varargs and ReshapedArrays #16251

RFC: Refactor indexing to use Varargs and ReshapedArrays #16251

Conversation

mbauman commented May 7, 2016 • edited Loading

timholy May 7, 2016

Choose a reason for hiding this comment

timholy May 7, 2016

Choose a reason for hiding this comment

timholy commented May 7, 2016

mbauman commented May 7, 2016

mbauman commented May 7, 2016

jrevels commented May 7, 2016

timholy commented May 7, 2016

mbauman commented May 7, 2016 • edited Loading

mbauman commented May 8, 2016 • edited Loading

timholy commented May 8, 2016

timholy commented May 8, 2016

jrevels commented May 9, 2016

nanosoldier commented May 9, 2016

jrevels commented May 9, 2016 • edited Loading

timholy commented May 9, 2016

jrevels commented May 9, 2016 • edited Loading

mbauman commented May 9, 2016

tkelman May 9, 2016

Choose a reason for hiding this comment

eschnett May 9, 2016

Choose a reason for hiding this comment

mbauman May 9, 2016 • edited Loading

Choose a reason for hiding this comment

vtjnash commented May 12, 2016 • edited Loading

vtjnash commented May 12, 2016

timholy commented May 12, 2016

mbauman commented May 12, 2016

timholy commented May 12, 2016

vtjnash commented May 13, 2016

timholy commented May 13, 2016

mbauman commented May 7, 2016 •

edited

Loading

mbauman commented May 7, 2016 •

edited

Loading

mbauman commented May 8, 2016 •

edited

Loading

jrevels commented May 9, 2016 •

edited

Loading

jrevels commented May 9, 2016 •

edited

Loading

mbauman May 9, 2016 •

edited

Loading

vtjnash commented May 12, 2016 •

edited

Loading