Add precompilation workload and move SpecialFunctions to an extension

[gbaraldi]

Mirrors CUDA.jl's startup strategy to cut time-to-first-kernel and using cost.

Precompilation workload

Adds src/precompile.jl (included at the end of AMDGPU.jl) with a
PrecompileTools @compile_workload that runs the GPUCompiler → AMDGPU codegen
pipeline on a dummy kernel during precompilation, so the first real kernel
launch no longer JIT-compiles the entire compiler.

Measured on Julia 1.12.6 (gfx, single GPU):

metric	before	after
first kernel (cold)	~8.2 s	~1.0 s
first A * B	full first-use compile	0 methods compiled

Details:

• The workload builds the compiler config manually (baseline gfx900 target), so
  it needs neither a GPU nor ROCm discovery — it only uses LLVM. Guarded by
  :AMDGPU in LLVM.backends() and VERSION >= v"1.12-" (matching CUDA.jl's
  foreign-MI-during-precompile workaround).
• Because the workload compiles before __init__/ROCm discovery runs,
  libdevice_libs is empty, which would poison the DEVICE_LIBS /
  _global_hostcalls caches with empty entries and bake them into the image
  (breaking __ocml_* linking at runtime). They are empty!-ed after the
  workload so they repopulate correctly once discovery has run.
• Adds the same explicit launch/library precompile directives CUDA.jl uses:
  hipfunction/actual_compilation/hiplink, plus the hot entry points of the
  bundled libraries (rocBLAS gemm!/*/mul!, rocSOLVER factorizations, rocFFT
  plans, rand).

SpecialFunctions extension

Moves SpecialFunctions from a hard dependency to a weakdep +
AMDGPUSpecialFunctionsExt, extracting its device overrides out of
device/gcn/math.jl (the Base-math overrides stay in the package). using AMDGPU no longer loads SpecialFunctions and its dependency tree; the device
overrides are registered when a user loads SpecialFunctions, same as CUDA.jl.

Verified: Base-math kernels (sqrt/exp/sin/pow), matmul, reductions, and
the extension's erf/loggamma device overrides all produce correct results.

🤖 Generated with Claude Code

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: b39908f	Previous: 7c9aab0	Ratio
amdgpu/synchronization/context/device	600 ns	600 ns	1
amdgpu/synchronization/stream/blocking	250 ns	250 ns	1
amdgpu/synchronization/stream/nonblocking	330 ns	330 ns	1
array/accumulate/Float32/1d	87581 ns	85972 ns	1.02
array/accumulate/Float32/dims=1	340805 ns	412075 ns	0.83
array/accumulate/Float32/dims=1L	136932 ns	137091 ns	1.00
array/accumulate/Float32/dims=2	130092 ns	130332 ns	1.00
array/accumulate/Float32/dims=2L	2810929 ns	2810115 ns	1.00
array/accumulate/Int64/1d	98781 ns	102751 ns	0.96
array/accumulate/Int64/dims=1	342704 ns	442706 ns	0.77
array/accumulate/Int64/dims=1L	168402 ns	167432 ns	1.01
array/accumulate/Int64/dims=2	122861 ns	127031 ns	0.97
array/accumulate/Int64/dims=2L	2988292 ns	2984467 ns	1.00
array/broadcast	147262 ns	70231 ns	2.10
array/construct	1660 ns	1700 ns	0.98
array/copy	39831 ns	40561 ns	0.98
array/copyto!/cpu_to_gpu	182913 ns	121541 ns	1.50
array/copyto!/gpu_to_cpu	114891 ns	114461 ns	1.00
array/copyto!/gpu_to_gpu	131672 ns	66551 ns	1.98
array/iteration/findall/bool	183593 ns	181832 ns	1.01
array/iteration/findall/int	198563 ns	192932 ns	1.03
array/iteration/findfirst/bool	121652 ns	122251 ns	1.00
array/iteration/findfirst/int	117022 ns	116342 ns	1.01
array/iteration/findmin/1d	167612 ns	170152 ns	0.99
array/iteration/findmin/2d	156902 ns	153822 ns	1.02
array/iteration/logical	351255 ns	350744 ns	1.00
array/iteration/scalar	297134 ns	296083 ns	1.00
array/permutedims/2d	75101 ns	74481 ns	1.01
array/permutedims/3d	75361 ns	74251 ns	1.01
array/permutedims/4d	77521 ns	76951 ns	1.01
array/random/rand/Float32	54780 ns	52171 ns	1.05
array/random/rand/Int64	57970 ns	58731 ns	0.99
array/random/rand!/Float32	144282 ns	85101 ns	1.70
array/random/rand!/Int64	148042 ns	69261 ns	2.14
array/random/randn/Float32	88831 ns	98642 ns	0.90
array/random/randn!/Float32	85702 ns	101231 ns	0.85
array/reductions/mapreduce/Float32/1d	134732 ns	134242 ns	1.00
array/reductions/mapreduce/Float32/dims=1	96252 ns	95431 ns	1.01
array/reductions/mapreduce/Float32/dims=1L	776151 ns	774349 ns	1.00
array/reductions/mapreduce/Float32/dims=2	97541 ns	97531 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	300234 ns	297464 ns	1.01
array/reductions/mapreduce/Int64/1d	135342 ns	134951 ns	1.00
array/reductions/mapreduce/Int64/dims=1	96162 ns	95301 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	782481 ns	781800 ns	1.00
array/reductions/mapreduce/Int64/dims=2	97292 ns	96801 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	298984 ns	299524 ns	1.00
array/reductions/reduce/Float32/1d	134262 ns	133912 ns	1.00
array/reductions/reduce/Float32/dims=1	95851 ns	95711 ns	1.00
array/reductions/reduce/Float32/dims=1L	776211 ns	775219 ns	1.00
array/reductions/reduce/Float32/dims=2	98121 ns	97621 ns	1.01
array/reductions/reduce/Float32/dims=2L	299415 ns	297424 ns	1.01
array/reductions/reduce/Int64/1d	135342 ns	134602 ns	1.01
array/reductions/reduce/Int64/dims=1	96441 ns	95311 ns	1.01
array/reductions/reduce/Int64/dims=1L	785441 ns	780269 ns	1.01
array/reductions/reduce/Int64/dims=2	97792 ns	97121 ns	1.01
array/reductions/reduce/Int64/dims=2L	300154 ns	299264 ns	1.00
array/reverse/1d	43911 ns	44550 ns	0.99
array/reverse/1dL	75561 ns	76661 ns	0.99
array/reverse/1dL_inplace	166332 ns	173202 ns	0.96
array/reverse/1d_inplace	75971 ns	84571 ns	0.90
array/reverse/2d	51641 ns	52831 ns	0.98
array/reverse/2dL	102322 ns	102811 ns	1.00
array/reverse/2dL_inplace	159442 ns	178873 ns	0.89
array/reverse/2d_inplace	144072 ns	96051 ns	1.50
array/sorting/1d	341175 ns	379995 ns	0.90
integration/byval/reference	39811 ns	39540 ns	1.01
integration/byval/slices=1	40810 ns	40350 ns	1.01
integration/byval/slices=2	146032 ns	159152 ns	0.92
integration/byval/slices=3	237953 ns	238933 ns	1.00
integration/volumerhs	5047571 ns	5031334 ns	1.00
kernel/indexing	130511 ns	65521 ns	1.99
kernel/indexing_checked	65341 ns	72491 ns	0.90
kernel/launch	1330 ns	1280 ns	1.04
kernel/rand	123352 ns	124252 ns	0.99
latency/import	1516539925 ns	1491816057 ns	1.02
latency/precompile	33049701658 ns	11773992921 ns	2.81
latency/ttfp	2079499186 ns	10954774141 ns	0.19

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

[luraess]

Thanks! Feel free to merge when ready.

[luraess]

@gbaraldi is this ready?