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: 9ea5a7c	Previous: 7c9aab0	Ratio
amdgpu/synchronization/context/device	540 ns	600 ns	0.90
amdgpu/synchronization/stream/blocking	230 ns	250 ns	0.92
amdgpu/synchronization/stream/nonblocking	300 ns	330 ns	0.91
array/accumulate/Float32/1d	364975 ns	85972 ns	4.25
array/accumulate/Float32/dims=1	489717 ns	412075 ns	1.19
array/accumulate/Float32/dims=1L	563168 ns	137091 ns	4.11
array/accumulate/Float32/dims=2	459447 ns	130332 ns	3.53
array/accumulate/Float32/dims=2L	15277269 ns	2810115 ns	5.44
array/accumulate/Int64/1d	367166 ns	102751 ns	3.57
array/accumulate/Int64/dims=1	486067 ns	442706 ns	1.10
array/accumulate/Int64/dims=1L	462337 ns	167432 ns	2.76
array/accumulate/Int64/dims=2	480176 ns	127031 ns	3.78
array/accumulate/Int64/dims=2L	15885978 ns	2984467 ns	5.32
array/broadcast	207103 ns	70231 ns	2.95
array/construct	1630 ns	1700 ns	0.96
array/copy	36680 ns	40561 ns	0.90
array/copyto!/cpu_to_gpu	153202 ns	121541 ns	1.26
array/copyto!/gpu_to_cpu	109752 ns	114461 ns	0.96
array/copyto!/gpu_to_gpu	59051 ns	66551 ns	0.89
array/iteration/findall/bool	486877 ns	181832 ns	2.68
array/iteration/findall/int	513967 ns	192932 ns	2.66
array/iteration/findfirst/bool	472767 ns	122251 ns	3.87
array/iteration/findfirst/int	479817 ns	116342 ns	4.12
array/iteration/findmin/1d	793041 ns	170152 ns	4.66
array/iteration/findmin/2d	819362 ns	153822 ns	5.33
array/iteration/logical	515258 ns	350744 ns	1.47
array/iteration/scalar	275884 ns	296083 ns	0.93
array/permutedims/2d	115622 ns	74481 ns	1.55
array/permutedims/3d	198453 ns	74251 ns	2.67
array/permutedims/4d	296474 ns	76951 ns	3.85
array/random/rand/Float32	47471 ns	52171 ns	0.91
array/random/rand/Int64	90451 ns	58731 ns	1.54
array/random/rand!/Float32	66481 ns	85101 ns	0.78
array/random/rand!/Int64	189223 ns	69261 ns	2.73
array/random/randn/Float32	136992 ns	98642 ns	1.39
array/random/randn!/Float32	100062 ns	101231 ns	0.99
array/reductions/mapreduce/Float32/1d	472847 ns	134242 ns	3.52
array/reductions/mapreduce/Float32/dims=1	434207 ns	95431 ns	4.55
array/reductions/mapreduce/Float32/dims=1L	32851041 ns	774349 ns	42.42
array/reductions/mapreduce/Float32/dims=2	438946 ns	97531 ns	4.50
array/reductions/mapreduce/Float32/dims=2L	1760865 ns	297464 ns	5.92
array/reductions/mapreduce/Int64/1d	490807 ns	134951 ns	3.64
array/reductions/mapreduce/Int64/dims=1	550688 ns	95301 ns	5.78
array/reductions/mapreduce/Int64/dims=1L	36316261 ns	781800 ns	46.45
array/reductions/mapreduce/Int64/dims=2	562338 ns	96801 ns	5.81
array/reductions/mapreduce/Int64/dims=2L	1759095 ns	299524 ns	5.87
array/reductions/reduce/Float32/1d	506997 ns	133912 ns	3.79
array/reductions/reduce/Float32/dims=1	426166 ns	95711 ns	4.45
array/reductions/reduce/Float32/dims=1L	35624312 ns	775219 ns	45.95
array/reductions/reduce/Float32/dims=2	548737 ns	97621 ns	5.62
array/reductions/reduce/Float32/dims=2L	1796905 ns	297424 ns	6.04
array/reductions/reduce/Int64/1d	495207 ns	134602 ns	3.68
array/reductions/reduce/Int64/dims=1	540128 ns	95311 ns	5.67
array/reductions/reduce/Int64/dims=1L	36295651 ns	780269 ns	46.52
array/reductions/reduce/Int64/dims=2	560428 ns	97121 ns	5.77
array/reductions/reduce/Int64/dims=2L	1772895 ns	299264 ns	5.92
array/reverse/1d	207643 ns	44550 ns	4.66
array/reverse/1dL	528128 ns	76661 ns	6.89
array/reverse/1dL_inplace	647099 ns	173202 ns	3.74
array/reverse/1d_inplace	269914 ns	84571 ns	3.19
array/reverse/2d	314174 ns	52831 ns	5.95
array/reverse/2dL	589058 ns	102811 ns	5.73
array/reverse/2dL_inplace	654099 ns	178873 ns	3.66
array/reverse/2d_inplace	339214 ns	96051 ns	3.53
array/sorting/1d	13185389 ns	379995 ns	34.70
integration/byval/reference	120982 ns	39540 ns	3.06
integration/byval/slices=1	120211 ns	40350 ns	2.98
integration/byval/slices=2	258893 ns	159152 ns	1.63
integration/byval/slices=3	1266579 ns	238933 ns	5.30
integration/volumerhs	5292186 ns	5031334 ns	1.05
kernel/indexing	78391 ns	65521 ns	1.20
kernel/indexing_checked	214043 ns	72491 ns	2.95
kernel/launch	1360 ns	1280 ns	1.06
kernel/rand	256404 ns	124252 ns	2.06
latency/import	1626534717 ns	1491816057 ns	1.09
latency/precompile	11900817336 ns	11773992921 ns	1.01
latency/ttfp	10743222244 ns	10954774141 ns	0.98

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