Unable to create PencilFFTPlan using ROCArray #74
Comments
|
Without the stack trace, I'm guessing the error comes from this line: in the PencilArrays.jl package. The easy solution would be to remove the |
|
I did a simple modification: function unsafe_as_array(::Type{T}, x::AbstractVector{UInt8}, dims) where {T}
p = typeof_ptr(x){T}(pointer(x))
unsafe_wrap(typeof_array(x), p, dims, lock=false)
endAt least the error message is now different, here's the stack trace: And this error comes from multiplication |
|
Thanks for testing. So the More precisely, for in-place transforms, we reuse a single One reason for this failure would be that AMDGPU.jl is missing a definition of You can also try the following script (replacing data = Array{Float64}(undef, 200)
u = reshape(view(data, 1:100), 20, 5)
v = reshape(view(data, 1:40), 10, 4)
Base.mightalias(u, v) # true
Base.dataids(u) == Base.dataids(v) # trueIf the last line gives You might also get away with removing the failing |
|
I used test script: using AMDGPU
function test(SomeArray)
data = SomeArray{Float64}(undef, 200)
u = reshape(view(data, 1:100), 20, 5)
v = reshape(view(data, 1:40), 10, 4)
@show Base.mightalias(u, v)
@show Base.dataids(u) == Base.dataids(v)
end
test(Array)
test(ROCArray)Results: Yes, it seems that AMDGPU.jl is missing the definition found from [CUDA.jl}(https://github.com/JuliaGPU/CUDA.jl/blob/e1e5be2b6bf17f03a367cebeb18c4645e593f80d/src/array.jl#L99): Base.dataids(A::CuArray) = (UInt(pointer(A)),)A modified test script using AMDGPU
Base.dataids(A::ROCArray) = (UInt(pointer(A)),)
function test(SomeArray)
data = SomeArray{Float64}(undef, 200)
u = reshape(view(data, 1:100), 20, 5)
v = reshape(view(data, 1:40), 10, 4)
@show Base.mightalias(u, v)
@show Base.dataids(u) == Base.dataids(v)
end
test(Array)
test(ROCArray)Will now show true, but unfortunately using plan is still not working: Again, line 26 being |
|
Thanks, it looks like the Could you try replacing your first modification to function unsafe_as_array(::Type{T}, x::ROCVector{UInt8}, dims::Tuple) where {T}
p = typeof_ptr(x){T}(pointer(x))
unsafe_wrap(typeof_array(x), p, dims, lock=false)
end
unsafe_as_array(::Type{T}, x::ROCVector{UInt8}, N::Integer) where {T} = unsafe_as_array(T, x, (N,)) |
# Reinterpret UInt8 vector as a different type of array.
# The input array should have enough space for the reinterpreted array with the
# given dimensions.
# This is a workaround to the performance issues when using `reinterpret`.
# See for instance:
# - https://discourse.julialang.org/t/big-overhead-with-the-new-lazy-reshape-reinterpret/7635
# - https://github.com/JuliaLang/julia/issues/28980
#=
function unsafe_as_array(::Type{T}, x::AbstractVector{UInt8}, dims) where {T}
p = typeof_ptr(x){T}(pointer(x))
unsafe_wrap(typeof_array(x), p, dims, lock=false)
end
=#
using AMDGPU
function unsafe_as_array(::Type{T}, x::ROCVector{UInt8}, dims::Tuple) where {T}
p = typeof_ptr(x){T}(pointer(x))
unsafe_wrap(typeof_array(x), p, dims, lock=false)
end
unsafe_as_array(::Type{T}, x::ROCVector{UInt8}, N::Integer) where {T} = unsafe_as_array(T, x, (N,))And the modification Base.dataids(A::ROCArray) = (UInt(pointer(A)),)Now: So these changes are enough to make things work. |
|
Awesome! I think it would make sense to contribute the As for the |
|
There's a version conflict preventing me trying this |
|
It's likely a compatibility issue between MPI.jl and AMDGPU.jl. It may be fixed by adding Besides I think AMDGPU.jl needs a version bump to include your |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Any tips on where should I start to fix this...?
The text was updated successfully, but these errors were encountered: