use AMDGPU_LLVM_Backend_jll instead of in-tree backend

[simeonschaub]

We might eventually want to switch to using the clang shipped by ROCm, so that the compiler, ld.lld, device libs, and the other vendor libs always stay in sync with upstream, but let's try the jll approach for now.

fixes #1002
cc @vchuravy since you worked on JuliaGPU/GPUCompiler.jl#857

[vchuravy]

Oh did I never finish that part :)

[luraess]

Interestingly, we have documentation and benchmark CI jobs that are failing, while others pass (with exception of this failing for known reason)

[simeonschaub]

Ah, I think I see what's happening: LLD_jll is a Julia stdlib, so we're forced to use the version Julia ships with. That version is too old to handle the code produced by AMDGPU_LLVM_Backend_jll, which is based on LLVM 22, though.

I (hopefully) fixed the precompile issues by just skipping creating the executable altogether, but this means we are now relying on the lld shipped by ROCm and we can't fall back to LLD_jll anymore. Maybe that's ok though and we should just remove the artifact fallback path for lld?

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 131d6cd	Previous: 724d15f	Ratio
amdgpu/synchronization/context/device	592.5 ns	577.5 ns	1.03
amdgpu/synchronization/stream/blocking	242.5 ns	237.5 ns	1.02
amdgpu/synchronization/stream/nonblocking	337.5 ns	322.5 ns	1.05
array/accumulate/Float32/1d	73401 ns	72383.75 ns	1.01
array/accumulate/Float32/dims=1	269466.25 ns	278954.25 ns	0.97
array/accumulate/Float32/dims=1L	81178.75 ns	80043.75 ns	1.01
array/accumulate/Float32/dims=2	72686 ns	71466 ns	1.02
array/accumulate/Float32/dims=2L	2758000.5 ns	2620236.25 ns	1.05
array/accumulate/Int64/1d	77081.25 ns	76381 ns	1.01
array/accumulate/Int64/dims=1	243728.5 ns	243786 ns	1.00
array/accumulate/Int64/dims=1L	84011.25 ns	84088.75 ns	1.00
array/accumulate/Int64/dims=2	83861.25 ns	83453.75 ns	1.00
array/accumulate/Int64/dims=2L	2893604.75 ns	2864940.25 ns	1.01
array/broadcast	74278.5 ns	68581 ns	1.08
array/construct	2347.5 ns	2195 ns	1.07
array/copy	37568.25 ns	37048 ns	1.01
array/copyto!/cpu_to_gpu	91788.75 ns	104081.5 ns	0.88
array/copyto!/gpu_to_cpu	104156.25 ns	104479 ns	1.00
array/copyto!/gpu_to_gpu	59395.75 ns	37960.5 ns	1.56
array/iteration/findall/bool	137027 ns	134574.5 ns	1.02
array/iteration/findall/int	152282.25 ns	148672.25 ns	1.02
array/iteration/findfirst/bool	147847.25 ns	123486.75 ns	1.20
array/iteration/findfirst/int	144967 ns	124531.75 ns	1.16
array/iteration/findmin/1d	111424 ns	110069 ns	1.01
array/iteration/findmin/2d	108049 ns	96364 ns	1.12
array/iteration/logical	249688.75 ns	245316.25 ns	1.02
array/iteration/scalar	308662 ns	296437 ns	1.04
array/permutedims/2d	72766 ns	71361 ns	1.02
array/permutedims/3d	72088.5 ns	70816 ns	1.02
array/permutedims/4d	75013.5 ns	73823.5 ns	1.02
array/random/rand/Float32	45880.75 ns	45335.75 ns	1.01
array/random/rand/Int64	55385.75 ns	54068.25 ns	1.02
array/random/rand!/Float32	56638.25 ns	47823 ns	1.18
array/random/rand!/Int64	72736 ns	66071 ns	1.10
array/random/randn/Float32	76931 ns	77073.75 ns	1.00
array/random/randn!/Float32	81738.5 ns	81381 ns	1.00
array/reductions/mapreduce/Float32/1d	88676 ns	94514 ns	0.94
array/reductions/mapreduce/Float32/dims=1	82503.5 ns	84653.75 ns	0.97
array/reductions/mapreduce/Float32/dims=1L	833466.25 ns	776879 ns	1.07
array/reductions/mapreduce/Float32/dims=2	80638.75 ns	83853.75 ns	0.96
array/reductions/mapreduce/Float32/dims=2L	145414.5 ns	143094.75 ns	1.02
array/reductions/mapreduce/Int64/1d	88693.75 ns	94963.75 ns	0.93
array/reductions/mapreduce/Int64/dims=1	83961.25 ns	84003.75 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	831863.75 ns	789344.25 ns	1.05
array/reductions/mapreduce/Int64/dims=2	84501 ns	83641.25 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	143724.5 ns	144282.25 ns	1.00
array/reductions/reduce/Float32/1d	88423.75 ns	92824 ns	0.95
array/reductions/reduce/Float32/dims=1	83018.5 ns	84624 ns	0.98
array/reductions/reduce/Float32/dims=1L	838582 ns	782216.75 ns	1.07
array/reductions/reduce/Float32/dims=2	85068.75 ns	84249 ns	1.01
array/reductions/reduce/Float32/dims=2L	145022.25 ns	143379.75 ns	1.01
array/reductions/reduce/Int64/1d	95243.75 ns	87928.75 ns	1.08
array/reductions/reduce/Int64/dims=1	83736 ns	84403.75 ns	0.99
array/reductions/reduce/Int64/dims=1L	834269.5 ns	789071.75 ns	1.06
array/reductions/reduce/Int64/dims=2	84751.25 ns	83748.75 ns	1.01
array/reductions/reduce/Int64/dims=2L	145784.75 ns	142694.5 ns	1.02
array/reverse/1d	45005.5 ns	43903.25 ns	1.03
array/reverse/1dL	70013.5 ns	71013.5 ns	0.99
array/reverse/1dL_inplace	80188.75 ns	79038.5 ns	1.01
array/reverse/1d_inplace	40200.5 ns	59903.25 ns	0.67
array/reverse/2d	49770.75 ns	49135.75 ns	1.01
array/reverse/2dL	81181.25 ns	79608.5 ns	1.02
array/reverse/2dL_inplace	91601.5 ns	91884 ns	1.00
array/reverse/2d_inplace	63113.25 ns	55113.25 ns	1.15
array/sorting/1d	336832.25 ns	338222.5 ns	1.00
integration/byval/reference	39661 ns	39280 ns	1.01
integration/byval/slices=1	40550 ns	40360 ns	1.00
integration/byval/slices=2	158302 ns	150632 ns	1.05
integration/byval/slices=3	237524 ns	236914 ns	1.00
integration/volumerhs	5003920 ns	5050775 ns	0.99
kernel/indexing	58080.75 ns	57423.5 ns	1.01
kernel/indexing_checked	49203.25 ns	47713.25 ns	1.03
kernel/launch	1670 ns	1567.5 ns	1.07
kernel/rand	99656.5 ns	98961.5 ns	1.01
latency/import	1720972876 ns	1712872529 ns	1.00
latency/precompile	37716303741 ns	37407776408 ns	1.01
latency/ttfp	5784812270 ns	5749419423 ns	1.01

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

[vchuravy]

Or have LLD be coming from AMDGPU_LLVM_Backend_jll

[luraess]

Thanks!