Avoid boxing compiler jobs during kernel resolution

[maleadt]

Kernel resolution constructs a CompilerJob immediately before passing it through an unspecialized compile_or_lookup boundary, which can force the job onto the heap. Allow specialization on the job's target and compiler-parameter types, and keep the method @noinline so its body is not duplicated into callers specialized per kernel. Kernel identity remains a field of the job, so this does not introduce per-kernel specialization or another cache.

Applies the approach from KernelAbstractions.jl#769.

Syntax and whitespace checks passed, and the package loaded on Julia 1.13. Device execution and allocation measurements were unavailable because the local HIP runtime was not functional.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 52023c9	Previous: 63c6c40	Ratio
amdgpu/synchronization/context/device	552.5 ns	582.5 ns	0.95
amdgpu/synchronization/stream/blocking	230.25 ns	235 ns	0.98
amdgpu/synchronization/stream/nonblocking	315 ns	342.5 ns	0.92
applications/bitonic_sort	1129812.25 ns	1135780.75 ns	0.99
applications/convolution	104334 ns	105389 ns	0.99
applications/floyd_warshall	8864840.25 ns	8837117.5 ns	1.00
applications/histogram	810207.5 ns	803238.75 ns	1.01
applications/prefix_sum	231968.5 ns	231228.25 ns	1.00
array/accumulate/Float32/1d	75573.75 ns	79176 ns	0.95
array/accumulate/Float32/dims=1	266539 ns	270856.25 ns	0.98
array/accumulate/Float32/dims=1L	67636 ns	82303.75 ns	0.82
array/accumulate/Float32/dims=2	70118.75 ns	70921 ns	0.99
array/accumulate/Float32/dims=2L	2609173.5 ns	2753465.25 ns	0.95
array/accumulate/Int64/1d	78238.75 ns	77441 ns	1.01
array/accumulate/Int64/dims=1	242869 ns	242275.75 ns	1.00
array/accumulate/Int64/dims=1L	81491.25 ns	83723.5 ns	0.97
array/accumulate/Int64/dims=2	85266.25 ns	85516.25 ns	1.00
array/accumulate/Int64/dims=2L	2889973.25 ns	2887214.75 ns	1.00
array/broadcast	71943.75 ns	72326 ns	0.99
array/construct	2212.5 ns	2327.5 ns	0.95
array/copy	37200.5 ns	33308 ns	1.12
array/copyto!/cpu_to_gpu	111219.25 ns	111544 ns	1.00
array/copyto!/gpu_to_cpu	118657 ns	112239 ns	1.06
array/copyto!/gpu_to_gpu	58703.5 ns	53915.75 ns	1.09
array/iteration/findall/bool	131467.25 ns	134521.75 ns	0.98
array/iteration/findall/int	147562.5 ns	148569.5 ns	0.99
array/iteration/findfirst/bool	181535.5 ns	182537.5 ns	0.99
array/iteration/findfirst/int	141354.75 ns	140624.25 ns	1.01
array/iteration/findmin/1d	105612 ns	118906.5 ns	0.89
array/iteration/findmin/2d	106744.5 ns	109151.5 ns	0.98
array/iteration/logical	234893.75 ns	238428.25 ns	0.99
array/iteration/scalar	288022.25 ns	292276.25 ns	0.99
array/permutedims/2d	58606 ns	70986 ns	0.83
array/permutedims/3d	69891.25 ns	70406 ns	0.99
array/permutedims/4d	72351 ns	60245.75 ns	1.20
array/random/rand/Float32	45158 ns	45383 ns	1.00
array/random/rand/Int64	53278.25 ns	53550.75 ns	0.99
array/random/rand!/Float32	64536 ns	64251 ns	1.00
array/random/rand!/Int64	71883.75 ns	71186 ns	1.01
array/random/randn/Float32	78326.25 ns	79301 ns	0.99
array/random/randn!/Float32	80408.75 ns	80246.25 ns	1.00
array/reductions/mapreduce/Float32/1d	92771.5 ns	94648.75 ns	0.98
array/reductions/mapreduce/Float32/dims=1	85004 ns	86616.25 ns	0.98
array/reductions/mapreduce/Float32/dims=1L	840941 ns	835469 ns	1.01
array/reductions/mapreduce/Float32/dims=2	81206.25 ns	82718.75 ns	0.98
array/reductions/mapreduce/Float32/dims=2L	137979.75 ns	138384.5 ns	1.00
array/reductions/mapreduce/Int64/1d	93644 ns	94203.75 ns	0.99
array/reductions/mapreduce/Int64/dims=1	70518.5 ns	86458.75 ns	0.82
array/reductions/mapreduce/Int64/dims=1L	843898.75 ns	841119 ns	1.00
array/reductions/mapreduce/Int64/dims=2	82213.75 ns	83068.5 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	139117.25 ns	138761.75 ns	1.00
array/reductions/reduce/Float32/1d	92884 ns	94551.5 ns	0.98
array/reductions/reduce/Float32/dims=1	79568.75 ns	86223.75 ns	0.92
array/reductions/reduce/Float32/dims=1L	838031 ns	841294 ns	1.00
array/reductions/reduce/Float32/dims=2	81986.5 ns	82688.5 ns	0.99
array/reductions/reduce/Float32/dims=2L	137454.75 ns	137874.25 ns	1.00
array/reductions/reduce/Int64/1d	92369 ns	94288.75 ns	0.98
array/reductions/reduce/Int64/dims=1	70558.75 ns	85981.25 ns	0.82
array/reductions/reduce/Int64/dims=1L	850099 ns	840701.25 ns	1.01
array/reductions/reduce/Int64/dims=2	80853.75 ns	83251.25 ns	0.97
array/reductions/reduce/Int64/dims=2L	138654.75 ns	138074.5 ns	1.00
array/reverse/1d	43295.75 ns	43553 ns	0.99
array/reverse/1dL	73666.25 ns	73793.5 ns	1.00
array/reverse/1dL_inplace	77761.25 ns	78213.5 ns	0.99
array/reverse/1d_inplace	59073.5 ns	41343 ns	1.43
array/reverse/2d	48696 ns	49133 ns	0.99
array/reverse/2dL	81496.25 ns	84381 ns	0.97
array/reverse/2dL_inplace	88741.5 ns	89121 ns	1.00
array/reverse/2d_inplace	60903.5 ns	61473.25 ns	0.99
array/sorting/1d	333430.25 ns	335059.75 ns	1.00
gemm/tiled	1892401.5 ns	1903771.5 ns	0.99
gemm/tiled_unbounded	1895439 ns	1922259 ns	0.99
integration/byval/reference	36781 ns	39361 ns	0.93
integration/byval/slices=1	41660 ns	41280 ns	1.01
integration/byval/slices=2	158513 ns	158962 ns	1.00
integration/byval/slices=3	237694 ns	237943 ns	1.00
integration/volumerhs	4913295 ns	4900479 ns	1.00
kernel/indexing	34908 ns	55808.25 ns	0.63
kernel/indexing_checked	56728.25 ns	38333 ns	1.48
kernel/launch	1152.5 ns	1337.5 ns	0.86
kernel/rand	85994 ns	99514 ns	0.86
latency/import	1714206163 ns	1748092483 ns	0.98
latency/precompile	39858876511 ns	39846434991 ns	1.00
latency/ttfp	2325183817 ns	2322353398 ns	1.00
stencil/diffusion3d	1618706.25 ns	1620301.5 ns	1.00
stencil/diffusion3d_checked	1664557 ns	1658012.25 ns	1.00

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

[luraess]

Again Int128 failing test, on 1.12 only

AMDGPU.jl/test/core/rocarray_broadcast.jl

Line 59 in 63c6c40

@test Array(test_kernel.(ROCArray(M), Int128(10))) == test_kernel.(M, Int128(10))