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
{{ message }}
Repository navigation
[BUG] oneAPI: indexed af::max / af::min along dim 1 return wrong values and indices #3732
On the oneAPI backend, af::max(val, idx, in, 1) and af::min(val, idx, in, 1) (indexed reductions along dim 1) return wrong values and indices when the array has few rows: only part of each row is reduced.
Description
Official Windows installer, ArrayFire v3.10.0 (build 492718b).
Backend: oneAPI only. The same binary on the same GPU gives correct results with the OpenCL backend, and with the CPU backend. CUDA / OpenCL / CPU on an NVIDIA machine are also correct.
Only the indexed form is affected. af::max(in, 1) (no indices) is correct, and the indexed call along dim 0 is correct.
Reproducible every run.
Workaround: transpose (or af::reorder) so the reduced dimension is dim 0, call af::max(val, idx, in, 0), and transpose back.
Expected: the same values and indices as af::max(val, idx, in.T(), 0) transposed back, as on the other backends.
Related but different: #3656 (wrong location for a subarray input, on CUDA/OpenCL/CPU). Here the input is a plain, contiguous array, and only oneAPI is affected.
Activity
docs(probes): oneAPI indexed min / max filed as arrayfire/arrayfire#3732
docs: KNOWN_ISSUES.md, shipped in every release package