Performance of @views with reshape

[jkozdon]

When view is used with a reshape in some circumstances there can be a lot of allocation.

Consider the following minimal example:

const N = 1000
function reshapefill(A)
  A = reshape(A, N, N)
  for j = 1:N, i = 1:N
    A[i, j] = 0
  end
end

function noreshapefill(A)
  for j = 1:N, i = 1:N
    A[i, j] = 0
  end
end

let
  A = Array{Float64}(undef, N, N, N)

  B = @views A[1, :, :]
  noreshapefill(B)
  reshapefill(B)

  println("B = @views A[1, :, :]")
  println("fill without reshape")
  @time noreshapefill(B)
  println("fill with reshape")
  @time reshapefill(B)

  C = @views A[:, 1, :]
  noreshapefill(C)
  reshapefill(C)

  println()
  println("C = @views A[:, 1, :]")
  println("fill without reshape")
  @time noreshapefill(C)
  println("fill with reshape")
  @time reshapefill(C)

  D = @views A[:, :, 1]
  noreshapefill(D)
  reshapefill(D)

  println()
  println("D = @views A[:, :, 1]")
  println("fill without reshape")
  @time noreshapefill(D)
  println("fill with reshape")
  @time reshapefill(D)
end

which results in the output

B = @views A[1, :, :]
fill without reshape
  0.022676 seconds (4 allocations: 160 bytes)
fill with reshape
  0.793889 seconds (8.00 M allocations: 320.419 MiB, 3.36% gc time)

C = @views A[:, 1, :]
fill without reshape
  0.000762 seconds (4 allocations: 160 bytes)
fill with reshape
  0.003002 seconds (5 allocations: 224 bytes)

D = @views A[:, :, 1]
fill without reshape
  0.000811 seconds (4 allocations: 160 bytes)
fill with reshape
  0.730238 seconds (8.00 M allocations: 320.419 MiB, 3.70% gc time)

As can be seen when @views is used with a reshape there can be a lot of allocations in some situations but not always.

Here is my versioninfo()

Julia Version 1.0.1
Commit 0d713926f8 (2018-09-29 19:05 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin14.5.0)
  CPU: Intel(R) Xeon(R) CPU E5-1650 v2 @ 3.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, ivybridge)

Not surprisingly, I see similar performance issues when I replace the @views macros with the view function.

[mbauman]

Thanks for the report. Looks like we're hitting a splatting penalty here… why isn't SubArray's setindex! inlining here?

julia> @code_warntype reshapefill(B)
…
  │    %53 = (Base.add_int)(%49, %52)::Int64 # computation of linear index
  │    %54 = (Core.tuple)(A, 0)::Tuple{SubArray{Float64,2,Array{Float64,3},Tuple{Int64,Base.Slice{Base.OneTo{Int64}},Base.Slice{Base.OneTo{Int64}}},true},Int64}
  │          (Core._apply)(Base.setindex!, %54, %53)
…

[vtjnash]

dup #29114