You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Right now only full matrices are supported, and to pass in a Diagonal you have to convert it to a full dense array with diagm. Some additional special cases here would be great.
Hi @kshyatt , I was trying to call kron(Diagonal(ROCArray), Diagonal(ROCArray)) with allowscalar(false) on my RX9060XT (gfx1200) GPU, it crashes with a HIPError and poisons the entire GPU context, so nothing GPU-related works until you restart Julia.
**System**
- GPU: AMD Radeon RX 9060 XT (gfx1100)
- VRAM: 16 GiB
- Julia: 1.12.1
- AMDGPU.jl: v2.2.1
- ROCm: 7.8.0
- Julia: 1.12.1
**Reproducer**
using AMDGPU
using LinearAlgebra
AMDGPU.allowscalar(false)
n = 512
d = Diagonal(ROCArray(rand(Float32, n))) # hangs here
Is this something which is specific to RDNA4 as even or is this actually a bug worth creating an issue?
Activity
shreyas-omkar commentedon Feb 22, 2026
Hi @kshyatt , I was trying to call
kron(Diagonal(ROCArray), Diagonal(ROCArray))withallowscalar(false)on my RX9060XT (gfx1200) GPU, it crashes with a HIPError and poisons the entire GPU context, so nothing GPU-related works until you restart Julia.Is this something which is specific to RDNA4 as even or is this actually a bug worth creating an issue?
maleadt commentedon Oct 19, 2026
Fixed by #690, which added
kron/kron!for GPU-backedDiagonalwithDiagonaland dense matrices.