RFC: automatic therock artifact selection

[simeonschaub]

This is a proof of concept for automatically downloading ROCm libraries via the artifact system. Not sure this is the best approach, maybe we should wrap these tarballs as JLLs through Yggdrasil instead?

It's currently partially vibe coded and Linux only, since I couldn't find a reliable way of querying the gfx_target_version on Windows.

[gbaraldi]

Have you seen that multiarch tarballs are available. Though I think they are gigantic

[simeonschaub]

Yes, I saw them, but like you said I don't think it's great to ship such huge tarballs to users

[simeonschaub]

This is now using the clang shipped by rocm as also proposed by @vchuravy in #931 (comment). Locally I am getting miscompilations in the triangular matmul tests, which I will try to reduce and open an upstream issue.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 0d12ab5	Previous: c3e9e19	Ratio
amdgpu/synchronization/context/device	552.5 ns	562.5 ns	0.98
amdgpu/synchronization/stream/blocking	237.5 ns	235 ns	1.01
amdgpu/synchronization/stream/nonblocking	312.5 ns	320 ns	0.98
array/accumulate/Float32/1d	173142.5 ns	74431.25 ns	2.33
array/accumulate/Float32/dims=1	316667.5 ns	279051.75 ns	1.13
array/accumulate/Float32/dims=1L	452844.25 ns	80943.75 ns	5.59
array/accumulate/Float32/dims=2	228906 ns	72506 ns	3.16
array/accumulate/Float32/dims=2L	15564961.25 ns	2617537.5 ns	5.95
array/accumulate/Int64/1d	178197.5 ns	78246 ns	2.28
array/accumulate/Int64/dims=1	331245 ns	243966 ns	1.36
array/accumulate/Int64/dims=1L	484725 ns	83891.25 ns	5.78
array/accumulate/Int64/dims=2	207750.75 ns	81418.75 ns	2.55
array/accumulate/Int64/dims=2L	15867657 ns	2867614.25 ns	5.53
array/broadcast	143444.5 ns	52263.25 ns	2.74
array/construct	2395.25 ns	2372.5 ns	1.01
array/copy	32078 ns	36485.5 ns	0.88
array/copyto!/cpu_to_gpu	111386.75 ns	86611 ns	1.29
array/copyto!/gpu_to_cpu	111921.75 ns	112164.25 ns	1.00
array/copyto!/gpu_to_gpu	38973 ns	35980.5 ns	1.08
array/iteration/findall/bool	254501.5 ns	141492 ns	1.80
array/iteration/findall/int	265581.75 ns	158094.75 ns	1.68
array/iteration/findfirst/bool	481505 ns	127356.75 ns	3.78
array/iteration/findfirst/int	491822.75 ns	127839.25 ns	3.85
array/iteration/findmin/1d	811835 ns	113976.75 ns	7.12
array/iteration/findmin/2d	850786 ns	107709 ns	7.90
array/iteration/logical	550746.25 ns	260086.25 ns	2.12
array/iteration/scalar	277744.5 ns	295631.5 ns	0.94
array/permutedims/2d	117421.75 ns	71536 ns	1.64
array/permutedims/3d	131454.5 ns	71118.5 ns	1.85
array/permutedims/4d	154204.75 ns	74313.5 ns	2.08
array/random/rand/Float32	44510.5 ns	45328 ns	0.98
array/random/rand/Int64	114436.75 ns	55040.75 ns	2.08
array/random/rand!/Float32	64981 ns	65493.5 ns	0.99
array/random/rand!/Int64	154014.75 ns	53880.75 ns	2.86
array/random/randn/Float32	83673.75 ns	76038.5 ns	1.10
array/random/randn!/Float32	76958.75 ns	60518.5 ns	1.27
array/reductions/mapreduce/Float32/1d	340852.75 ns	96758.75 ns	3.52
array/reductions/mapreduce/Float32/dims=1	405808.75 ns	85553.75 ns	4.74
array/reductions/mapreduce/Float32/dims=1L	36221579 ns	781805.75 ns	46.33
array/reductions/mapreduce/Float32/dims=2	411476.25 ns	84621.25 ns	4.86
array/reductions/mapreduce/Float32/dims=2L	1872676.25 ns	144722 ns	12.94
array/reductions/mapreduce/Int64/1d	355515.25 ns	94921.25 ns	3.75
array/reductions/mapreduce/Int64/dims=1	401326.25 ns	84463.75 ns	4.75
array/reductions/mapreduce/Int64/dims=1L	39148376.25 ns	791916 ns	49.44
array/reductions/mapreduce/Int64/dims=2	406641.25 ns	83748.75 ns	4.86
array/reductions/mapreduce/Int64/dims=2L	1860943.75 ns	141764.5 ns	13.13
array/reductions/reduce/Float32/1d	342337.75 ns	96491.5 ns	3.55
array/reductions/reduce/Float32/dims=1	403396.25 ns	85201.25 ns	4.73
array/reductions/reduce/Float32/dims=1L	37828141 ns	782083.25 ns	48.37
array/reductions/reduce/Float32/dims=2	410719 ns	84891 ns	4.84
array/reductions/reduce/Float32/dims=2L	1871578.75 ns	142992 ns	13.09
array/reductions/reduce/Int64/1d	354010.5 ns	88566.25 ns	4.00
array/reductions/reduce/Int64/dims=1	397306 ns	84548.5 ns	4.70
array/reductions/reduce/Int64/dims=1L	36559043.75 ns	788023.5 ns	46.39
array/reductions/reduce/Int64/dims=2	408866.5 ns	83401.25 ns	4.90
array/reductions/reduce/Int64/dims=2L	1868976.5 ns	143302 ns	13.04
array/reverse/1d	101664.25 ns	43483.25 ns	2.34
array/reverse/1dL	449649.75 ns	73061 ns	6.15
array/reverse/1dL_inplace	337150.25 ns	78998.5 ns	4.27
array/reverse/1d_inplace	132677 ns	59848.25 ns	2.22
array/reverse/2d	125659.5 ns	49138.25 ns	2.56
array/reverse/2dL	610169.75 ns	79461.25 ns	7.68
array/reverse/2dL_inplace	487237.75 ns	91798.75 ns	5.31
array/reverse/2d_inplace	132472 ns	62230.75 ns	2.13
array/sorting/1d	13472808.25 ns	340847.5 ns	39.53
integration/byval/reference	112422 ns	39480 ns	2.85
integration/byval/slices=1	112432 ns	40561 ns	2.77
integration/byval/slices=2	223974 ns	121111 ns	1.85
integration/byval/slices=3	1274849 ns	234933 ns	5.43
integration/volumerhs	5056826 ns	5042886 ns	1.00
kernel/indexing	67913.5 ns	57603.25 ns	1.18
kernel/indexing_checked	111301.75 ns	59520.75 ns	1.87
kernel/launch	1502.5 ns	1570 ns	0.96
kernel/rand	134472.25 ns	80076.25 ns	1.68
latency/import	1771562179 ns	1763912896 ns	1.00
latency/precompile	37845526629 ns	37604981379 ns	1.01
latency/ttfp	2373223982 ns	5783496043 ns	0.41

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

[gbaraldi]

Only thing is I wonder if you can split therock from using clang just to reduce the diff

[simeonschaub]

Yes, I agree that makes sense. We'll have to figure out how this should interact with JuliaGPU/GPUCompiler.jl#857 as well, since with those changes we'd have 3 possible backends - in tree, AMDGPU_LLVM_Backend_jll, and clang from the rocm path. What I like about the latter approach is that the compiler, ld.lld, device libs, and the other vendor libs always match and we don't have to worry about keeping all those different jlls up to date and in sync.