refactor discovery, ship therock tarballs as artifacts

[simeonschaub]

Some of the AI comments still need to be cleaned up but the logic looks reasonable from a first review, so I thought I'd put this up as a draft for feedback.

It turns out MIOpen needs all of libllvm and libclang, so unlike discussed in #919 (comment), I went back to shipping therock artifacts directly as done in that PR, since we don't gain much by repackaging anymore.

Otherwise, do a split similar to CUDA.jl of having ROCm_Runtime be a pseudo-jll providing automatic downloads through artifacts and having manual discovery live in ROCm_Runtime_Discovery. Either one of those is then loaded depending on a preference

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 0162ceb	Previous: 39fe271	Ratio
amdgpu/synchronization/context/device	527.5 ns	560 ns	0.94
amdgpu/synchronization/stream/blocking	222.5 ns	230 ns	0.97
amdgpu/synchronization/stream/nonblocking	297.5 ns	315 ns	0.94
array/accumulate/Float32/1d	80553.5 ns	70043.5 ns	1.15
array/accumulate/Float32/dims=1	280381.5 ns	284709 ns	0.98
array/accumulate/Float32/dims=1L	95523.75 ns	80278.75 ns	1.19
array/accumulate/Float32/dims=2	92016.25 ns	72131 ns	1.28
array/accumulate/Float32/dims=2L	2754104 ns	2614542.5 ns	1.05
array/accumulate/Int64/1d	81188.5 ns	78386.25 ns	1.04
array/accumulate/Int64/dims=1	248158.5 ns	244018.5 ns	1.02
array/accumulate/Int64/dims=1L	101566.5 ns	83741.25 ns	1.21
array/accumulate/Int64/dims=2	93793.75 ns	84758.75 ns	1.11
array/accumulate/Int64/dims=2L	2894465.75 ns	2891796.75 ns	1.00
array/broadcast	56058.5 ns	72998.75 ns	0.77
array/construct	2517.5 ns	2182.5 ns	1.15
array/copy	36475.5 ns	36628 ns	1.00
array/copyto!/cpu_to_gpu	111141.5 ns	110909.25 ns	1.00
array/copyto!/gpu_to_cpu	111876.75 ns	110519.25 ns	1.01
array/copyto!/gpu_to_gpu	59271 ns	45368.25 ns	1.31
array/iteration/findall/bool	144777 ns	132337 ns	1.09
array/iteration/findall/int	151544.75 ns	151897.25 ns	1.00
array/iteration/findfirst/bool	151932 ns	144392.25 ns	1.05
array/iteration/findfirst/int	150492 ns	145337.25 ns	1.04
array/iteration/findmin/1d	132394.25 ns	113169 ns	1.17
array/iteration/findmin/2d	122166.75 ns	108861.75 ns	1.12
array/iteration/logical	243088.5 ns	251833.5 ns	0.97
array/iteration/scalar	306136.75 ns	295024.25 ns	1.04
array/permutedims/2d	71986 ns	71828.5 ns	1.00
array/permutedims/3d	71476 ns	69728.5 ns	1.03
array/permutedims/4d	74198.5 ns	74073.75 ns	1.00
array/random/rand/Float32	44495.5 ns	45320.75 ns	0.98
array/random/rand/Int64	54270.75 ns	54208.25 ns	1.00
array/random/rand!/Float32	65776 ns	64303.25 ns	1.02
array/random/rand!/Int64	55200.75 ns	71623.5 ns	0.77
array/random/randn/Float32	81006.25 ns	79463.75 ns	1.02
array/random/randn!/Float32	80341 ns	80086.25 ns	1.00
array/reductions/mapreduce/Float32/1d	108146.75 ns	97711.5 ns	1.11
array/reductions/mapreduce/Float32/dims=1	94866.25 ns	85676.25 ns	1.11
array/reductions/mapreduce/Float32/dims=1L	842147 ns	832249.5 ns	1.01
array/reductions/mapreduce/Float32/dims=2	99669 ns	84791.25 ns	1.18
array/reductions/mapreduce/Float32/dims=2L	144384.5 ns	144512 ns	1.00
array/reductions/mapreduce/Int64/1d	107486.5 ns	97688.75 ns	1.10
array/reductions/mapreduce/Int64/dims=1	94903.75 ns	84121.25 ns	1.13
array/reductions/mapreduce/Int64/dims=1L	838259.25 ns	836152 ns	1.00
array/reductions/mapreduce/Int64/dims=2	96038.75 ns	84931.25 ns	1.13
array/reductions/mapreduce/Int64/dims=2L	144549.5 ns	142327 ns	1.02
array/reductions/reduce/Float32/1d	107706.5 ns	97861.25 ns	1.10
array/reductions/reduce/Float32/dims=1	94466.25 ns	84426.25 ns	1.12
array/reductions/reduce/Float32/dims=1L	842487 ns	835814.5 ns	1.01
array/reductions/reduce/Float32/dims=2	97311.5 ns	84691.25 ns	1.15
array/reductions/reduce/Float32/dims=2L	144412 ns	143922.25 ns	1.00
array/reductions/reduce/Int64/1d	107349 ns	97814 ns	1.10
array/reductions/reduce/Int64/dims=1	92998.75 ns	83881.25 ns	1.11
array/reductions/reduce/Int64/dims=1L	834579.25 ns	827999.25 ns	1.01
array/reductions/reduce/Int64/dims=2	97911.25 ns	84931 ns	1.15
array/reductions/reduce/Int64/dims=2L	144567 ns	144367 ns	1.00
array/reverse/1d	46175.5 ns	45648.25 ns	1.01
array/reverse/1dL	75903.5 ns	73036 ns	1.04
array/reverse/1dL_inplace	63353.25 ns	80503.75 ns	0.79
array/reverse/1d_inplace	61746 ns	52233.25 ns	1.18
array/reverse/2d	50685.75 ns	50390.75 ns	1.01
array/reverse/2dL	91208.75 ns	82971.25 ns	1.10
array/reverse/2dL_inplace	92803.75 ns	92171.25 ns	1.01
array/reverse/2d_inplace	59695.75 ns	55451 ns	1.08
array/sorting/1d	335577.25 ns	335057.25 ns	1.00
integration/byval/reference	39521 ns	39661 ns	1.00
integration/byval/slices=1	40360 ns	40411 ns	1.00
integration/byval/slices=2	158202 ns	147452 ns	1.07
integration/byval/slices=3	238793 ns	248133 ns	0.96
integration/volumerhs	4989270 ns	5013333 ns	1.00
kernel/indexing	36878 ns	57200.75 ns	0.64
kernel/indexing_checked	60208.25 ns	54958.25 ns	1.10
kernel/launch	1320 ns	1382.5 ns	0.95
kernel/rand	119299.25 ns	98641.5 ns	1.21
latency/import	1998362070 ns	1895221133 ns	1.05
latency/precompile	41803446362 ns	38320507704 ns	1.09
latency/ttfp	6051840349 ns	5930819959 ns	1.02

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

[luraess]

Cool!

Do we also handle Instinct GPUs? I see they seem omitted in explicit cases which though relate mostly to Windows OS from the comments.

[simeonschaub]

Yes, instinct should work just fine on Linux, it's just Windows that doesn't support them because upstream doesn't provide tarballs for that combination.

We have external CI on MI300, correct? Could you trigger a run, or can I even do it myself now?

[luraess]

I see, the marketing name is Windows only indeed, so all seems fine and Instinct GPUs are included. Just tested this locally and all works fine.

[luraess]

We have external CI on MI300, correct? Could you trigger a run, or can I even do it myself now?

This would only be possible if the PR comes from an AMDGPU branch (limitation from external CI). Is there a way we could switch to a AMDGPU branch now that you have write access?