3D FFT gets `DivideError: integer division error`

[islent]

My MWE:

using Distributed
addprocs(8)
@everywhere using Dagger

using FFTW
N = 64
A = rand(ComplexF64,N,N,N);
DA = DArray(A);
fft(DA)

It throws error:

ERROR: DivideError: integer division error
Stacktrace:
  [1] div
    @ .\int.jl:295 [inlined]
  [2] _cumlength(len::Int64, step::Int64)
    @ Dagger E:\.julia\packages\Dagger\Jl5aj\src\array\alloc.jl:62
  [3] map
    @ .\tuple.jl:383 [inlined]
  [4] map
    @ .\tuple.jl:386 [inlined]
  [5] partition
    @ E:\.julia\packages\Dagger\Jl5aj\src\array\alloc.jl:69 [inlined]
  [6] zeros(p::Blocks{3}, eltype::Type, dims::Tuple{Int64, Int64, Int64}; assignment::Symbol)
    @ Dagger E:\.julia\packages\Dagger\Jl5aj\src\array\alloc.jl:153
  [7] zeros
    @ E:\.julia\packages\Dagger\Jl5aj\src\array\alloc.jl:151 [inlined]
  [8] _fft!(output::DArray{…}, input::DArray{…}, dims::Tuple{…}; decomp::AbstractFFTsExt.Pencil)
    @ AbstractFFTsExt E:\.julia\packages\Dagger\Jl5aj\ext\AbstractFFTsExt.jl:124
  [9] fft(DA::DArray{ComplexF64, 3, Blocks{3}, typeof(cat)}, dims::Tuple{Int64, Int64, Int64}; decomp::AbstractFFTsExt.Pencil)
    @ AbstractFFTsExt E:\.julia\packages\Dagger\Jl5aj\ext\AbstractFFTsExt.jl:62
 [10] fft
    @ E:\.julia\packages\Dagger\Jl5aj\ext\AbstractFFTsExt.jl:59 [inlined]
 [11] fft(DA::DArray{ComplexF64, 3, Blocks{3}, typeof(cat)})
    @ AbstractFFTsExt E:\.julia\packages\Dagger\Jl5aj\ext\AbstractFFTsExt.jl:59
 [12] top-level scope
    @ REPL[30]:1

I also tried to manually control the partitions, whereas the DivideError cannot be solved:

DA = DArray(A, Blocks(N,N,div(N,8)));
DA = DArray(A, Blocks(N,N,2));
DA = DArray(A, Blocks(N,N,1));
DA = DArray(A, Blocks(8,8,8));

# Or change N and rerun
N = 20

Couldn't figure out how does the error arise from:

Dagger.jl/src/array/alloc.jl

Lines 68 to 71 in 13ac291

	function partition(p::AbstractBlocks, dom::ArrayDomain)
	DomainBlocks(map(first, indexes(dom)),
	map(_cumlength, map(length, indexes(dom)), p.blocksize))
	end

Dagger.jl/src/array/alloc.jl

Lines 151 to 155 in 13ac291

	function Base.zeros(p::Blocks, eltype::Type, dims::Dims; assignment::AssignmentType = :arbitrary)
	d = ArrayDomain(map(x->1:x, dims))
	a = AllocateArray(eltype, zeros, false, d, partition(p, d), p, assignment)
	return _to_darray(a)
	end

[islent]

It seems p.blocksize is zero, but why?

[jpsamaroo]

Ahh this is a bug in the FFT logic, we use div here where we should instead use cld, otherwise we might get a zero dimension when np is much larger than N:

Dagger.jl/ext/AbstractFFTsExt.jl

Line 124 in 12c77cc

A = zeros(Blocks(N, div(N, np), div(N, np)), T, size(input))

. Would you mind submitting a PR to change all the div calls to cld?