Adapt to the library-based AMDGPU_LLVM_Backend_jll 23 and LLVMDowngrader_jll 0.10

[maleadt]

JuliaPackaging/Yggdrasil#14727 replaced the llc, lld and llvm-downgrade executables in the GPU LLVM JLLs with shared libraries exposing a small C API, moving the back-ends to LLVM 23 at the same time. JuliaGPU/GPUCompiler.jl#930 makes GPUCompiler call those libraries in-process, and bumps its compat to the new JLL majors. Since this package pins the same JLL, it needs its compat bumped in lockstep, which is what this PR does.

AMDGPU.jl used two of the removed tools itself, so besides the compat bumps (AMDGPU_LLVM_Backend_jll → 23, LLVMDowngrader_jll → 0.10, GPUCompiler → 2.7):

• Linking the kernel object into an HSA code object goes through AMDGPULink from libamdgpu (ld.lld -flavor gnu -shared in-process, no temporary files) whenever the artifact is used. A system ld.lld from ROCm is still preferred when found, and still spawned as before. lld_artifact now means "link through the library" and no longer comes with a tool path; versioninfo shows the library instead.
• Downgrading the device libraries for the in-process LLVM goes through LLVMDGDowngrade, one buffer at a time, with the downgrader's error message in the warning when a library is skipped.

[simeonschaub]

What's the reason we need to use dlopen and dlsym here for calling these libraries? Since they're just regular jlls, can't we use e.g. libamdgpu in ccall directly?

[maleadt]

What's the reason we need to use dlopen and dlsym here for calling these libraries? Since they're just regular jlls, can't we use e.g. libamdgpu in ccall directly?

Ah, yes. This was modeled after how we did things in JuliaGPU/GPUCompiler.jl#930, where the names are constructed more dynamically. I'll make it a plain ccall here.

[maleadt]

Previous CI run was all green, an latest change was NFC, so merging this a little more quickly such that we get GPUCompiler reverse CI back.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 91edd35	Previous: b3e831e	Ratio
amdgpu/synchronization/context/device	557.5 ns	575 ns	0.97
amdgpu/synchronization/stream/blocking	235 ns	235 ns	1
amdgpu/synchronization/stream/nonblocking	320 ns	317.5 ns	1.01
applications/bitonic_sort	1143756.25 ns	1135751.5 ns	1.01
applications/convolution	113714 ns	104889 ns	1.08
applications/floyd_warshall	8947917.5 ns	8875884.5 ns	1.01
applications/histogram	817656.5 ns	802806.75 ns	1.02
applications/prefix_sum	232690.75 ns	230405.75 ns	1.01
array/accumulate/Float32/1d	80258.5 ns	79131.25 ns	1.01
array/accumulate/Float32/dims=1	272946.5 ns	265801.5 ns	1.03
array/accumulate/Float32/dims=1L	69086 ns	81151 ns	0.85
array/accumulate/Float32/dims=2	87616.25 ns	81428.75 ns	1.08
array/accumulate/Float32/dims=2L	2623047.75 ns	2753025.75 ns	0.95
array/accumulate/Int64/1d	83183.75 ns	77788.75 ns	1.07
array/accumulate/Int64/dims=1	243648.5 ns	245128.75 ns	0.99
array/accumulate/Int64/dims=1L	98031.25 ns	83273.75 ns	1.18
array/accumulate/Int64/dims=2	91528.75 ns	86051.25 ns	1.06
array/accumulate/Int64/dims=2L	2900544.25 ns	3132783.75 ns	0.93
array/broadcast	72818.5 ns	72161 ns	1.01
array/construct	2327.5 ns	2235 ns	1.04
array/copy	37090.75 ns	37170.5 ns	1.00
array/copyto!/cpu_to_gpu	90263.75 ns	111469 ns	0.81
array/copyto!/gpu_to_cpu	111844 ns	111984 ns	1.00
array/copyto!/gpu_to_gpu	58743.25 ns	58868.25 ns	1.00
array/iteration/findall/bool	142189.5 ns	133412 ns	1.07
array/iteration/findall/int	149474.75 ns	147627.25 ns	1.01
array/iteration/findfirst/bool	181745 ns	182052.75 ns	1.00
array/iteration/findfirst/int	161587.25 ns	142917.25 ns	1.13
array/iteration/findmin/1d	123824.5 ns	119061.75 ns	1.04
array/iteration/findmin/2d	112974.25 ns	108601.5 ns	1.04
array/iteration/logical	240981 ns	238773.75 ns	1.01
array/iteration/scalar	290179.25 ns	292314.5 ns	0.99
array/permutedims/2d	59383.5 ns	70388.5 ns	0.84
array/permutedims/3d	71118.5 ns	69858.5 ns	1.02
array/permutedims/4d	73231 ns	72816 ns	1.01
array/random/rand/Float32	44868 ns	45410.75 ns	0.99
array/random/rand/Int64	53795.75 ns	53175.75 ns	1.01
array/random/rand!/Float32	64888.5 ns	64423.5 ns	1.01
array/random/rand!/Int64	72663.75 ns	66538.5 ns	1.09
array/random/randn/Float32	80501 ns	76001 ns	1.06
array/random/randn!/Float32	80818.5 ns	80356.25 ns	1.01
array/reductions/mapreduce/Float32/1d	101614 ns	94806.5 ns	1.07
array/reductions/mapreduce/Float32/dims=1	89986.5 ns	86078.75 ns	1.05
array/reductions/mapreduce/Float32/dims=1L	844724.75 ns	802139.5 ns	1.05
array/reductions/mapreduce/Float32/dims=2	94139 ns	75523.75 ns	1.25
array/reductions/mapreduce/Float32/dims=2L	139169.5 ns	136907 ns	1.02
array/reductions/mapreduce/Int64/1d	101571.5 ns	93919 ns	1.08
array/reductions/mapreduce/Int64/dims=1	90671.25 ns	89003.75 ns	1.02
array/reductions/mapreduce/Int64/dims=1L	858675 ns	806172 ns	1.07
array/reductions/mapreduce/Int64/dims=2	74703.5 ns	83146.25 ns	0.90
array/reductions/mapreduce/Int64/dims=2L	140309.5 ns	138227 ns	1.02
array/reductions/reduce/Float32/1d	101721.5 ns	94399 ns	1.08
array/reductions/reduce/Float32/dims=1	90168.75 ns	84823.75 ns	1.06
array/reductions/reduce/Float32/dims=1L	843050 ns	800964.5 ns	1.05
array/reductions/reduce/Float32/dims=2	95879 ns	75003.75 ns	1.28
array/reductions/reduce/Float32/dims=2L	138504.5 ns	137049.75 ns	1.01
array/reductions/reduce/Int64/1d	101284 ns	94093.75 ns	1.08
array/reductions/reduce/Int64/dims=1	90696.5 ns	88986.25 ns	1.02
array/reductions/reduce/Int64/dims=1L	848904.75 ns	808594.5 ns	1.05
array/reductions/reduce/Int64/dims=2	74486 ns	83041.25 ns	0.90
array/reductions/reduce/Int64/dims=2L	140372 ns	139134.75 ns	1.01
array/reverse/1d	44203.25 ns	43285.5 ns	1.02
array/reverse/1dL	74316 ns	70681 ns	1.05
array/reverse/1dL_inplace	78261.25 ns	77993.75 ns	1.00
array/reverse/1d_inplace	59681 ns	37600.5 ns	1.59
array/reverse/2d	49600.75 ns	48985.75 ns	1.01
array/reverse/2dL	87661.25 ns	82579 ns	1.06
array/reverse/2dL_inplace	89984 ns	89033.75 ns	1.01
array/reverse/2d_inplace	61613.5 ns	61158.25 ns	1.01
array/sorting/1d	332814.75 ns	331257.25 ns	1.00
gemm/tiled	1886369.5 ns	1907183 ns	0.99
gemm/tiled_unbounded	1919549.75 ns	1909665.5 ns	1.01
integration/byval/reference	39571 ns	39340 ns	1.01
integration/byval/slices=1	41641 ns	41231 ns	1.01
integration/byval/slices=2	158833 ns	123942 ns	1.28
integration/byval/slices=3	238033 ns	229293 ns	1.04
integration/volumerhs	5006621 ns	4913802 ns	1.02
kernel/indexing	56016 ns	47750.75 ns	1.17
kernel/indexing_checked	57395.75 ns	57591 ns	1.00
kernel/launch	1377.5 ns	1260 ns	1.09
kernel/rand	83368.75 ns	97451.5 ns	0.86
latency/import	1748778764 ns	1755925272 ns	1.00
latency/precompile	39786339983 ns	39950566320 ns	1.00
latency/ttfp	2320769848 ns	2345510054 ns	0.99
stencil/diffusion3d	1621991.75 ns	1629262.5 ns	1.00
stencil/diffusion3d_checked	1665800 ns	1657793 ns	1.00

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