Probe rocSPARSE with Libdl only

[luraess]

Relates to #1040. Follow-up to #1001.

#1001 moved the rocSPARSE version query into a subprocess so a segfault on a broken ROCm install (#920) degrades to err instead of killing the session. But the child ran using AMDGPU under the active project which is expensive.

The child now uses Base.Libc.Libdl only: it dlopens librocsparse by the absolute path the parent already knows and calls rocsparse_create_handle / rocsparse_get_version directly. The result is memoized for the session.

cc @simeonschaub

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 7d78c4e	Previous: 021540a	Ratio
amdgpu/synchronization/context/device	542.5 ns	547.5 ns	0.99
amdgpu/synchronization/stream/blocking	227.5 ns	235 ns	0.97
amdgpu/synchronization/stream/nonblocking	307.5 ns	312.5 ns	0.98
array/accumulate/Float32/1d	72718.5 ns	73158.5 ns	0.99
array/accumulate/Float32/dims=1	276651.5 ns	275026.25 ns	1.01
array/accumulate/Float32/dims=1L	78858.75 ns	78546 ns	1.00
array/accumulate/Float32/dims=2	73126 ns	72268.75 ns	1.01
array/accumulate/Float32/dims=2L	2751166.5 ns	2753974 ns	1.00
array/accumulate/Int64/1d	77896.25 ns	77868.75 ns	1.00
array/accumulate/Int64/dims=1	245016 ns	242633.5 ns	1.01
array/accumulate/Int64/dims=1L	83876.25 ns	84091 ns	1.00
array/accumulate/Int64/dims=2	83463.75 ns	71476.25 ns	1.17
array/accumulate/Int64/dims=2L	2893341.25 ns	2894525.75 ns	1.00
array/broadcast	73333.5 ns	73271 ns	1.00
array/construct	2262.5 ns	2295 ns	0.99
array/copy	36825.5 ns	36888 ns	1.00
array/copyto!/cpu_to_gpu	110471.75 ns	110634 ns	1.00
array/copyto!/gpu_to_cpu	120766.75 ns	110424 ns	1.09
array/copyto!/gpu_to_gpu	58846 ns	46220.75 ns	1.27
array/iteration/findall/bool	138384.5 ns	137422 ns	1.01
array/iteration/findall/int	147324.75 ns	147062 ns	1.00
array/iteration/findfirst/bool	146894.5 ns	147042 ns	1.00
array/iteration/findfirst/int	145784.75 ns	146377.25 ns	1.00
array/iteration/findmin/1d	113379.25 ns	111171.75 ns	1.02
array/iteration/findmin/2d	109674 ns	107931.5 ns	1.02
array/iteration/logical	251106 ns	248031.25 ns	1.01
array/iteration/scalar	295749.25 ns	295989.25 ns	1.00
array/permutedims/2d	71878.5 ns	71926 ns	1.00
array/permutedims/3d	71426 ns	71156 ns	1.00
array/permutedims/4d	74483.5 ns	73986 ns	1.01
array/random/rand/Float32	45185.75 ns	44998.25 ns	1.00
array/random/rand/Int64	54473.25 ns	54190.75 ns	1.01
array/random/rand!/Float32	65216 ns	65228.25 ns	1.00
array/random/rand!/Int64	52290.75 ns	73168.5 ns	0.71
array/random/randn/Float32	77221 ns	78318.75 ns	0.99
array/random/randn!/Float32	80846.25 ns	76536.25 ns	1.06
array/reductions/mapreduce/Float32/1d	95996.25 ns	95341.5 ns	1.01
array/reductions/mapreduce/Float32/dims=1	84291.25 ns	88586.25 ns	0.95
array/reductions/mapreduce/Float32/dims=1L	833392 ns	832311.75 ns	1.00
array/reductions/mapreduce/Float32/dims=2	85361.25 ns	85021.25 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	144572 ns	144267 ns	1.00
array/reductions/mapreduce/Int64/1d	96606.25 ns	97431.25 ns	0.99
array/reductions/mapreduce/Int64/dims=1	84488.5 ns	84368.5 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	831479.25 ns	832297 ns	1.00
array/reductions/mapreduce/Int64/dims=2	84721.25 ns	84728.75 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	145389.5 ns	144659.5 ns	1.01
array/reductions/reduce/Float32/1d	98248.75 ns	95276.25 ns	1.03
array/reductions/reduce/Float32/dims=1	84633.75 ns	84243.75 ns	1.00
array/reductions/reduce/Float32/dims=1L	832226.75 ns	833756.75 ns	1.00
array/reductions/reduce/Float32/dims=2	84661.25 ns	91626.5 ns	0.92
array/reductions/reduce/Float32/dims=2L	144654.75 ns	144562 ns	1.00
array/reductions/reduce/Int64/1d	97661.25 ns	97216.25 ns	1.00
array/reductions/reduce/Int64/dims=1	84461.25 ns	84073.75 ns	1.00
array/reductions/reduce/Int64/dims=1L	835541.75 ns	831101.75 ns	1.01
array/reductions/reduce/Int64/dims=2	84598.75 ns	84513.75 ns	1.00
array/reductions/reduce/Int64/dims=2L	144357 ns	144894.5 ns	1.00
array/reverse/1d	45765.5 ns	45428.25 ns	1.01
array/reverse/1dL	73478.5 ns	72771 ns	1.01
array/reverse/1dL_inplace	80428.75 ns	79873.75 ns	1.01
array/reverse/1d_inplace	40028 ns	60975.75 ns	0.66
array/reverse/2d	50268.25 ns	50025.75 ns	1.00
array/reverse/2dL	82666.25 ns	82088.75 ns	1.01
array/reverse/2dL_inplace	91928.75 ns	91983.75 ns	1.00
array/reverse/2d_inplace	63415.75 ns	63378.5 ns	1.00
array/sorting/1d	336734.5 ns	334864.75 ns	1.01
integration/byval/reference	39601 ns	39681 ns	1.00
integration/byval/slices=1	40040 ns	40541 ns	0.99
integration/byval/slices=2	152272 ns	153712 ns	0.99
integration/byval/slices=3	239263 ns	237193 ns	1.01
integration/volumerhs	5008007 ns	5005451 ns	1.00
kernel/indexing	57185.75 ns	47583.25 ns	1.20
kernel/indexing_checked	34028 ns	33818 ns	1.01
kernel/launch	1395 ns	1412.5 ns	0.99
kernel/rand	97571.25 ns	97204 ns	1.00
latency/import	1713310429 ns	1709687565 ns	1.00
latency/precompile	39614017578 ns	39720673731 ns	1.00
latency/ttfp	2316216409 ns	2310174550 ns	1.00

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

[simeonschaub]

Yes, seems pretty much what I had in mind, thanks! Left some comments below

[luraess]

cscs-ci run