Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

`copy` and `resize!` of a `ROCArray{T,N,Mem.HostBuffer}` silently switch to `Mem.HIPBuffer` / throw a convert MethodError

Open
  • Is something wrong?
#1,179 0 comments 0 reactions 1 assignee View on GitHub

Maintainers usually reply within 1 day

@luraess is already working on this.

Since Oct 21, 2026.

  • #1180 by @luraess — open

Assessment

This issue has not been assessed yet.

Description

ROCArray{T,N,B} carries its buffer type B, and similar preserves it, but copy and resize! do not:

using AMDGPU
x = ROCArray{Float32,1,AMDGPU.Mem.HostBuffer}(undef, 4); fill!(x, 1f0)
typeof(x)          # ROCArray{Float32, 1, AMDGPU.Runtime.Mem.HostBuffer}
typeof(copy(x))    # ROCArray{Float32, 1, AMDGPU.Runtime.Mem.HIPBuffer}   <- buffer type changed
typeof(similar(x)) # ROCArray{Float32, 1, AMDGPU.Runtime.Mem.HostBuffer}  (preserved)
resize!(x, 8)
# ERROR: MethodError: Cannot `convert` an object of type
#   GPUArrays.DataRef{AMDGPU.Managed{AMDGPU.Runtime.Mem.HIPBuffer}} to an object of type
#   GPUArrays.DataRef{AMDGPU.Managed{AMDGPU.Runtime.Mem.HostBuffer}}
# Stacktrace: setproperty!(::ROCArray{…,HostBuffer}, :data, ::DataRef{Managed{HIPBuffer}}) in resize!
y = ROCArray{Float32,1,AMDGPU.Mem.HIPBuffer}(undef, 4); resize!(y, 8)   # fine

Cause (as of main 284e548): Base.copy(X::ROCArray{T}) allocates ROCArray{T}(undef, size(X)), i.e. the default
HIPBuffer (src/array.jl:278-282); Base.resize!(A::ROCVector{T}, n) allocates
new_buf = Mem.HIPBuffer(bufsize; stream=stream()) unconditionally and assigns the new DataRef to a field typed
with the original buffer (src/array.jl:555-566). The TODO comment above resize! already says "Specialize ROCArray
on storage type / Check that it is not HostBuffer".

Consequence downstream: a struct holding HostBuffer arrays cannot be deepcopy'd (typeassert in
deepcopy_internal fails on the changed type) and its vectors cannot be grown with resize!, so host-coherent
arrays are not a drop-in for code written against ROCArray (seen with a Pauli-propagation vector backend on an
MI300A APU, where host-coherent memory is the natural choice). Related: #840 (unified memory support).

Expected: copy returns the same buffer type (as similar does); resize! either keeps the buffer type or throws
an ArgumentError stating that HostBuffer arrays cannot be resized.

Environment: AMDGPU.jl v2.8.0 (LLVM.jl 9.13.2; the same code is on main 284e548), Julia 1.11.6 (official x86_64
tarball), ROCm 7.2.4, HIP 7.2.53211, rocBLAS 5.2.0, AMD Instinct MI300A (gfx942:sramecc+:xnack-, 4 devices,
128 GiB each), AlmaLinux 9.8, kernel 5.14. AMDGPU.functional() == true. Note: Julia 1.11's LLVM prints
"'gfx942' is not a recognized processor for this target" during compilation; unrelated to this issue (kernels run).

Dominant language
Julia
Stars
348
Forks
81
Avg merge
2d 7h
Merged PRs (30d)
54

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from JuliaGPU/AMDGPU.jl

All issues in JuliaGPU/AMDGPU.jl

Similar issues

More Julia issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.