Make versioninfo() robust to version-probe crashes

[luraess]

Summary

AMDGPU.versioninfo() could take down the entire Julia session with an uncatchable SIGSEGV on broken or mismatched ROCm installs. This isolates the one crash-prone probe in a subprocess, and fixes rocRAND to report the actually loaded version. Relates to #920.

Motivation

Of all the library version() calls in versioninfo(), only rocSPARSE requires creating a handle (rocsparse_get_version(handle, …)), which initializes a HIP context. On a broken/mismatched ROCm stack that can segfault (see #920), and a SIGSEGV cannot be caught with try/catch, so the crash propagated out and killed the caller. This is worse now that KA 0.10 forwards KA.versioninfo(io, ::ROCBackend) here (#868), exposing the crash through the backend-agnostic API.

Separately, rocRAND.version() reported the compile-time ROCRAND_VERSION constant (the version AMDGPU was built against) rather than the loaded library, the other libraries already probe at runtime.

Changes

• rocSPARSE: probe its version in a short-lived subprocess with a timeout. A crash, hang, or nonzero exit degrades to "err" in the table plus a single warning, instead of crashing the session. The other libraries expose handle-free version queries and stay in-process.
• rocRAND: query rocrand_get_version() at runtime (handle-free, safe) instead of the compile-time constant.

"err" is distinct from "-": "-" means the library was not found, "err" means it is present but its version query failed — itself a useful diagnostic signal.

Testing

New versioninfo probe isolation testset asserts that a clean child returns its output while SIGABRT / SIGSEGV / nonzero-exit / no-output / hang all degrade to nothing without crashing the caller (the first cases are GPU-independent). Verified locally on gfx1100 (RX 7900 XTX): full versioninfo() output unchanged, rocSPARSE version now probed out-of-process, failure path shows err + warning.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 2d9ed4d	Previous: c3e9e19	Ratio
amdgpu/synchronization/context/device	570 ns	562.5 ns	1.01
amdgpu/synchronization/stream/blocking	240 ns	235 ns	1.02
amdgpu/synchronization/stream/nonblocking	327.5 ns	320 ns	1.02
array/accumulate/Float32/1d	77481 ns	74431.25 ns	1.04
array/accumulate/Float32/dims=1	263748.75 ns	279051.75 ns	0.95
array/accumulate/Float32/dims=1L	89566.25 ns	80943.75 ns	1.11
array/accumulate/Float32/dims=2	83301 ns	72506 ns	1.15
array/accumulate/Float32/dims=2L	2769649.5 ns	2617537.5 ns	1.06
array/accumulate/Int64/1d	80831 ns	78246 ns	1.03
array/accumulate/Int64/dims=1	242393.5 ns	243966 ns	0.99
array/accumulate/Int64/dims=1L	83743.75 ns	83891.25 ns	1.00
array/accumulate/Int64/dims=2	88883.75 ns	81418.75 ns	1.09
array/accumulate/Int64/dims=2L	2866545.75 ns	2867614.25 ns	1.00
array/broadcast	73851 ns	52263.25 ns	1.41
array/construct	2292.5 ns	2372.5 ns	0.97
array/copy	36850.5 ns	36485.5 ns	1.01
array/copyto!/cpu_to_gpu	112021.5 ns	86611 ns	1.29
array/copyto!/gpu_to_cpu	110771.5 ns	112164.25 ns	0.99
array/copyto!/gpu_to_gpu	58870.75 ns	35980.5 ns	1.64
array/iteration/findall/bool	141642 ns	141492 ns	1.00
array/iteration/findall/int	151397 ns	158094.75 ns	0.96
array/iteration/findfirst/bool	119881.75 ns	127356.75 ns	0.94
array/iteration/findfirst/int	123009.25 ns	127839.25 ns	0.96
array/iteration/findmin/1d	124126.75 ns	113976.75 ns	1.09
array/iteration/findmin/2d	113961.5 ns	107709 ns	1.06
array/iteration/logical	252981 ns	260086.25 ns	0.97
array/iteration/scalar	296701.75 ns	295631.5 ns	1.00
array/permutedims/2d	72038.5 ns	71536 ns	1.01
array/permutedims/3d	71783.5 ns	71118.5 ns	1.01
array/permutedims/4d	74058.5 ns	74313.5 ns	1.00
array/random/rand/Float32	43330.5 ns	45328 ns	0.96
array/random/rand/Int64	55025.75 ns	55040.75 ns	1.00
array/random/rand!/Float32	65228.5 ns	65493.5 ns	1.00
array/random/rand!/Int64	66540.75 ns	53880.75 ns	1.23
array/random/randn/Float32	80446 ns	76038.5 ns	1.06
array/random/randn!/Float32	80706 ns	60518.5 ns	1.33
array/reductions/mapreduce/Float32/1d	103841.25 ns	96758.75 ns	1.07
array/reductions/mapreduce/Float32/dims=1	95346.25 ns	85553.75 ns	1.11
array/reductions/mapreduce/Float32/dims=1L	780553.5 ns	781805.75 ns	1.00
array/reductions/mapreduce/Float32/dims=2	94276.25 ns	84621.25 ns	1.11
array/reductions/mapreduce/Float32/dims=2L	143419.5 ns	144722 ns	0.99
array/reductions/mapreduce/Int64/1d	100499 ns	94921.25 ns	1.06
array/reductions/mapreduce/Int64/dims=1	94746.5 ns	84463.75 ns	1.12
array/reductions/mapreduce/Int64/dims=1L	784633.75 ns	791916 ns	0.99
array/reductions/mapreduce/Int64/dims=2	94451.25 ns	83748.75 ns	1.13
array/reductions/mapreduce/Int64/dims=2L	141677 ns	141764.5 ns	1.00
array/reductions/reduce/Float32/1d	104479 ns	96491.5 ns	1.08
array/reductions/reduce/Float32/dims=1	94498.75 ns	85201.25 ns	1.11
array/reductions/reduce/Float32/dims=1L	780106.25 ns	782083.25 ns	1.00
array/reductions/reduce/Float32/dims=2	93883.75 ns	84891 ns	1.11
array/reductions/reduce/Float32/dims=2L	143179.75 ns	142992 ns	1.00
array/reductions/reduce/Int64/1d	92141.5 ns	88566.25 ns	1.04
array/reductions/reduce/Int64/dims=1	93456.5 ns	84548.5 ns	1.11
array/reductions/reduce/Int64/dims=1L	787721.25 ns	788023.5 ns	1.00
array/reductions/reduce/Int64/dims=2	94168.75 ns	83401.25 ns	1.13
array/reductions/reduce/Int64/dims=2L	143089.5 ns	143302 ns	1.00
array/reverse/1d	43258.25 ns	43483.25 ns	0.99
array/reverse/1dL	72776 ns	73061 ns	1.00
array/reverse/1dL_inplace	79201.25 ns	78998.5 ns	1.00
array/reverse/1d_inplace	59733.25 ns	59848.25 ns	1.00
array/reverse/2d	49335.75 ns	49138.25 ns	1.00
array/reverse/2dL	85381.25 ns	79461.25 ns	1.07
array/reverse/2dL_inplace	90711.25 ns	91798.75 ns	0.99
array/reverse/2d_inplace	62633.5 ns	62230.75 ns	1.01
array/sorting/1d	342852 ns	340847.5 ns	1.01
integration/byval/reference	39570 ns	39480 ns	1.00
integration/byval/slices=1	40511 ns	40561 ns	1.00
integration/byval/slices=2	154052 ns	121111 ns	1.27
integration/byval/slices=3	233633 ns	234933 ns	0.99
integration/volumerhs	5045009 ns	5042886 ns	1.00
kernel/indexing	57708.5 ns	57603.25 ns	1.00
kernel/indexing_checked	53245.75 ns	59520.75 ns	0.89
kernel/launch	1572.5 ns	1570 ns	1.00
kernel/rand	106384.25 ns	80076.25 ns	1.33
latency/import	1765363002 ns	1763912896 ns	1.00
latency/precompile	37705375366 ns	37604981379 ns	1.00
latency/ttfp	5781907160 ns	5783496043 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.