Stream: helpdesk (published)

Topic: CUDA Scalar Indexing Issue


view this post on Zulip Dale Black (Aug 13 2021 at 01:27):

Any idea why this would be causing a scalar indexing problem?

begin
    for n in 4:6
        x4 = CuArray(rand([0, 1], n, n))
        sz = size(x4)
        dt = CUDA.zeros(Float32, sz),
        v = CUDA.ones(Int64, sz),
        z = CUDA.zeros(Float32, sz .+ 1)
    end
end
Scalar indexing is disallowed.

Invocation of getindex resulted in scalar indexing of a GPU array.

This is typically caused by calling an iterating implementation of a method.

Such implementations *do not* execute on the GPU, but very slowly on the CPU,

and therefore are only permitted from the REPL for prototyping purposes.

If you did intend to index this array, annotate the caller with @allowscalar.

error(::String)@error.jl:33
assertscalar(::String)@indexing.jl:53
getindex(::CUDA.CuArray{Float32, 2}, ::Int64)@indexing.jl:86
iterate@abstractarray.jl:1096[inlined]
iterate@abstractarray.jl:1094[inlined]
indexed_iterate(::CUDA.CuArray{Float32, 2}, ::Int64)@tuple.jl:89
top-level scope@Local: 5

view this post on Zulip Dale Black (Aug 15 2021 at 00:18):

Not sure why, but removing the [0, 1] from rand fixes the problem

begin
    for n in 4:6
        x4 = CuArray(rand(n, n))
        sz = size(x4)
        dt = CUDA.zeros(Float32, sz),
        v = CUDA.ones(Int64, sz),
        z = CUDA.zeros(Float32, sz .+ 1)
    end
end

view this post on Zulip Kyle Daruwalla (Aug 16 2021 at 02:21):

Maybe has to do with CUDA support for replacing Base intrinsics automatically. Stuff like sin or rand used to require calling CUDA.sin/CUDA.rand. Now CUDA will automatically replace those calls with CUDA equivalents when you are using CuArrays. So here you actually ended up calling CUDA.rand and sampling from [0, 1] will scalar index that vector.

view this post on Zulip Kyle Daruwalla (Aug 16 2021 at 02:22):

Maybe cu(Base.rand([0, 1], n, n)) will work?

view this post on Zulip Kyle Daruwalla (Aug 16 2021 at 02:22):

https://juliagpu.org/post/2021-04-09-cuda_3.0/ for more info

view this post on Zulip Michael Abbott (Aug 16 2021 at 02:37):

I think your problem is the stray commas. dt = zeros(sz), v = ones(sz), z = zeros(sz) is parsed as... some kind of unpacking like this:

julia> a, b = CUDA.ones(3)
ERROR: Scalar indexing is disallowed.

julia> sz = (4,4)
(4, 4)

julia> dt = CUDA.zeros(Float32, sz), v = CUDA.ones(Int64, sz), z = CUDA.zeros(Float32, sz .+ 1)
ERROR: Scalar indexing is disallowed.

view this post on Zulip Dale Black (Aug 16 2021 at 05:03):

Oh, I didn't even notice the commas before. That could've been the problem too, hmm


Last updated: Nov 22 2024 at 04:41 UTC