[maleadt]

Device-only code in the global method table gets compiled for the host by sysimage builds: Julia 1.12+ compiles every method with a compileable signature, and 1.11 keeps native code instances (e.g. those left by Core.Compiler.return_type(initialize_rng_state) in finish_module!). The image then refers to device symbols like julia.gpu.state_getter, and since sysimages are loaded with RTLD_NOW, it fails to load. Same fix as JuliaGPU/KernelAbstractions.jl#930 and JuliaGPU/OpenCL.jl#598, using the shared definitions from GPUToolbox 3.5.

• @device_override and @device_function are one-liners around GPUToolbox.Overlays.device_override / device_function; ExprTools is no longer needed.
• What GPUToolbox.Overlays.host_leaks(AMDGPU) reported is now a device function: output_context, printf_output_context, malloc_hc, free_hc, signal_exception, malloc, free (julia.gpu.state_getter), and ROCKernels' hardware_lane and lane_id (llvm.amdgcn.mbcnt.lo). Also kernel_state, which compile-all skips as a generated function, but which is the source of these. The KernelInterface functions built on the lane helpers already were @device_overrides.
• malloc and free lose their return type annotations, which Julia 1.10's @overlay can't parse; malloc asserts its result type instead.
• A new core/host_leaks test checks that host_leaks is empty for AMDGPU and its extensions.

Tested locally on a gfx1036 with Julia 1.13 (core, device, KernelAbstractions and KernelInterface tests) and 1.10.

[github-actions]

Benchmark results

Compared with 6ecd201bb, the merge base.
Times are the minimum time of every benchmark. A change is unusual when it exceeds both three times the benchmark's typical change between consecutive runs on the default branch and a minimum per category (performance 3%, latency 5%, applications 3%, other 3%). A benchmark that starts or stops allocating is listed however few the allocations. A single run can still be a fluke, so rerun before drawing conclusions.

	Geometric mean	Unusual: slower	Unusual: faster	Compared
Performance	−0.6%	3	2	69
Latency	+1.1%	0	0	3
Applications	−1.0%	0	0	14

Benchmark	Baseline	This PR	Change
Performance
array/permutedims/4d	50.2 µs	41.1 µs	−18.3%	🟢 faster
array/reductions/reduce/Float32/1d	69.3 µs	58.0 µs	−16.4%	🟢 faster
array/reductions/mapreduce/Float32/1d	58.7 µs	67.7 µs	+15.4%	🔴 slower
array/copy	33.3 µs	38.1 µs	+14.3%	🔴 slower
array/reductions/reduce/Float32/dims=2	46.1 µs	52.6 µs	+14.1%	🔴 slower

All benchmarks

Benchmark	Baseline	This PR	Change
Performance
amdgpu/synchronization/context/device	1.72 µs	1.79 µs	+4.4%
amdgpu/synchronization/stream/blocking	233 ns	230 ns	−1.1%
amdgpu/synchronization/stream/nonblocking	343 ns	353 ns	+2.9%
array/accumulate/Float32/1d	79.3 µs	78.7 µs	−0.7%
array/accumulate/Float32/dims=1	63.2 µs	62.4 µs	−1.3%
array/accumulate/Float32/dims=1L	105 µs	104 µs	−0.3%
array/accumulate/Float32/dims=2	103 µs	101 µs	−1.2%
array/accumulate/Float32/dims=2L	116 µs	116 µs	−0.4%
array/accumulate/Int64/1d	101 µs	101 µs	−0.6%
array/accumulate/Int64/dims=1	74.0 µs	73.5 µs	−0.6%
array/accumulate/Int64/dims=1L	93.6 µs	108 µs	+15.4%
array/accumulate/Int64/dims=2	106 µs	105 µs	−0.7%
array/accumulate/Int64/dims=2L	146 µs	143 µs	−1.8%
array/broadcast	21.7 µs	21.6 µs	−0.9%
array/construct	2.06 µs	2.01 µs	−2.3%
array/copy	33.3 µs	38.1 µs	+14.3%	🔴 slower
array/copyto!/cpu_to_gpu	83.8 µs	83.9 µs	+0.1%
array/copyto!/gpu_to_cpu	83.7 µs	83.7 µs	−0.0%
array/copyto!/gpu_to_gpu	24.9 µs	24.5 µs	−1.7%
array/iteration/findall/bool	90.8 µs	90.2 µs	−0.6%
array/iteration/findall/int	102 µs	101 µs	−0.5%
array/iteration/findfirst/bool	56.1 µs	54.9 µs	−2.2%
array/iteration/findfirst/int	54.1 µs	55.9 µs	+3.4%
array/iteration/findmin/1d	91.0 µs	91.4 µs	+0.4%
array/iteration/findmin/2d	75.4 µs	76.1 µs	+1.0%
array/iteration/logical	86.1 µs	86.4 µs	+0.3%
array/iteration/scalar	304 µs	296 µs	−2.4%
array/permutedims/2d	57.7 µs	56.5 µs	−2.1%
array/permutedims/3d	55.7 µs	55.2 µs	−0.9%
array/permutedims/4d	50.2 µs	41.1 µs	−18.3%	🟢 faster
array/random/rand!/Float32	23.7 µs	23.3 µs	−1.5%
array/random/rand!/Int64	24.6 µs	27.0 µs	+9.8%
array/random/rand/Float32	42.3 µs	41.8 µs	−1.2%
array/random/rand/Int64	43.3 µs	42.7 µs	−1.3%
array/random/randn!/Float32	30.5 µs	30.4 µs	−0.6%
array/random/randn/Float32	68.9 µs	68.7 µs	−0.3%
array/reductions/mapreduce/Float32/1d	58.7 µs	67.7 µs	+15.4%	🔴 slower
array/reductions/mapreduce/Float32/dims=1	51.8 µs	51.3 µs	−1.0%
array/reductions/mapreduce/Float32/dims=1L	56.0 µs	55.0 µs	−1.7%
array/reductions/mapreduce/Float32/dims=2	54.7 µs	54.1 µs	−1.1%
array/reductions/mapreduce/Float32/dims=2L	67.7 µs	67.2 µs	−0.7%
array/reductions/mapreduce/Int64/1d	70.0 µs	68.4 µs	−2.3%
array/reductions/mapreduce/Int64/dims=1	48.1 µs	47.6 µs	−1.0%
array/reductions/mapreduce/Int64/dims=1L	67.0 µs	66.7 µs	−0.6%
array/reductions/mapreduce/Int64/dims=2	55.3 µs	55.0 µs	−0.6%
array/reductions/mapreduce/Int64/dims=2L	73.1 µs	71.3 µs	−2.5%
array/reductions/reduce/Float32/1d	69.3 µs	58.0 µs	−16.4%	🟢 faster
array/reductions/reduce/Float32/dims=1	52.2 µs	50.5 µs	−3.1%
array/reductions/reduce/Float32/dims=1L	55.7 µs	55.1 µs	−1.0%
array/reductions/reduce/Float32/dims=2	46.1 µs	52.6 µs	+14.1%	🔴 slower
array/reductions/reduce/Float32/dims=2L	67.5 µs	67.0 µs	−0.7%
array/reductions/reduce/Int64/1d	69.6 µs	68.0 µs	−2.3%
array/reductions/reduce/Int64/dims=1	48.1 µs	47.6 µs	−1.1%
array/reductions/reduce/Int64/dims=1L	66.8 µs	66.1 µs	−1.1%
array/reductions/reduce/Int64/dims=2	55.2 µs	54.7 µs	−0.9%
array/reductions/reduce/Int64/dims=2L	72.7 µs	72.3 µs	−0.5%
array/reverse/1d	40.6 µs	39.2 µs	−3.4%
array/reverse/1d_inplace	27.6 µs	23.3 µs	−15.6%
array/reverse/1dL	64.2 µs	65.0 µs	+1.2%
array/reverse/1dL_inplace	30.9 µs	30.4 µs	−1.4%
array/reverse/2d	41.0 µs	42.2 µs	+2.9%
array/reverse/2d_inplace	26.4 µs	26.1 µs	−1.0%
array/reverse/2dL	70.8 µs	69.7 µs	−1.6%
array/reverse/2dL_inplace	36.1 µs	34.5 µs	−4.3%
array/sorting/1d	294 µs	294 µs	+0.2%
kernel/indexing	20.5 µs	20.2 µs	−1.4%
kernel/indexing_checked	25.0 µs	24.5 µs	−2.3%
kernel/launch	788 ns	793 ns	+0.6%
kernel/rand	28.4 µs	27.8 µs	−2.1%
Latency
import	1.27 s	1.31 s	+3.2%
precompile	28.6 s	28.4 s	−0.6%
ttfp	1.69 s	1.71 s	+0.9%
Applications
bitonic_sort	1.05 ms	1.05 ms	−0.0%
byval/reference	39.9 µs	39.4 µs	−1.3%
byval/slices=1	39.9 µs	40.4 µs	+1.1%
byval/slices=2	158 µs	144 µs	−8.9%
byval/slices=3	228 µs	230 µs	+0.7%
convolution	52.4 µs	52.2 µs	−0.3%
floyd_warshall	8.94 ms	8.92 ms	−0.2%
gemm/tiled	2.61 ms	2.60 ms	−0.3%
gemm/tiled_unbounded	2.60 ms	2.60 ms	+0.2%
histogram	754 µs	740 µs	−1.8%
prefix_sum	194 µs	189 µs	−2.2%
stencil/diffusion3d	1.55 ms	1.56 ms	+0.1%
stencil/diffusion3d_checked	1.59 ms	1.60 ms	+0.2%
volumerhs	4.84 ms	4.83 ms	−0.2%

History: https://juliagpu.org/perf/?pkg=amdgpu