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.

On Julia 1.13 / GPUCompiler 2.7 with an RTX 5080, warm cufunction resolution dropped from 304 to 32 bytes per call. Kernel launch and six cache checks passed, covering cache reuse, compiler options, reflection, and kernel/callee redefinition. Two kernels produced one compile_or_lookup specialization.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.65%. Comparing base (65f6ddd) to head (8a7eabe).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3273      +/-   ##
==========================================
+ Coverage   92.64%   92.65%   +0.01%     
==========================================
  Files         194      194              
  Lines       16988    16988              
==========================================
+ Hits        15738    15741       +3     
+ Misses       1250     1247       -3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 8a7eabe	Previous: 65f6ddd	Ratio
array/accumulate/Float32/1d	99648 ns	99072 ns	1.01
array/accumulate/Float32/dims=1	72332 ns	71446 ns	1.01
array/accumulate/Float32/dims=1L	1588784 ns	1588291 ns	1.00
array/accumulate/Float32/dims=2	138257 ns	137838 ns	1.00
array/accumulate/Float32/dims=2L	662363 ns	662440 ns	1.00
array/accumulate/Int64/1d	119677 ns	119538 ns	1.00
array/accumulate/Int64/dims=1	75373 ns	75442 ns	1.00
array/accumulate/Int64/dims=1L	1699644 ns	1700551 ns	1.00
array/accumulate/Int64/dims=2	149449 ns	149989 ns	1.00
array/accumulate/Int64/dims=2L	988289 ns	988238 ns	1.00
array/broadcast	18156 ns	18134 ns	1.00
array/broadcast launch	9092 ns	9573 ns	0.95
array/construct	924.060606060606 ns	883.3035714285714 ns	1.05
array/copy	16341 ns	16308 ns	1.00
array/copyto!/cpu_to_gpu	206617 ns	206834 ns	1.00
array/copyto!/gpu_to_cpu	239917 ns	239908 ns	1.00
array/copyto!/gpu_to_gpu	10175 ns	8787 ns	1.16
array/iteration/findall/bool	133884 ns	134476 ns	1.00
array/iteration/findall/int	146521 ns	147563 ns	0.99
array/iteration/findfirst/bool	68884 ns	69589 ns	0.99
array/iteration/findfirst/int	70649 ns	71492 ns	0.99
array/iteration/findmin/1d	68314 ns	65074 ns	1.05
array/iteration/findmin/2d	100393 ns	99665 ns	1.01
array/iteration/logical	192493 ns	193914 ns	0.99
array/iteration/scalar	63034 ns	64850 ns	0.97
array/permutedims/2d	48865 ns	48559 ns	1.01
array/permutedims/3d	50091 ns	49871 ns	1.00
array/permutedims/4d	49834 ns	50293 ns	0.99
array/random/rand/Float32	11128 ns	10986 ns	1.01
array/random/rand/Int64	21105 ns	20771 ns	1.02
array/random/rand!/Float32	7642.75 ns	7657.5 ns	1.00
array/random/rand!/Int64	18730 ns	18743 ns	1.00
array/random/randn/Float32	33357 ns	33687 ns	0.99
array/random/randn!/Float32	23236 ns	23290 ns	1.00
array/reductions/mapreduce/Float32/1d	32233 ns	32510 ns	0.99
array/reductions/mapreduce/Float32/dims=1	37761 ns	37611 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	50854 ns	51090 ns	1.00
array/reductions/mapreduce/Float32/dims=2	56330 ns	56419 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	67178 ns	67348 ns	1.00
array/reductions/mapreduce/Int64/1d	39519 ns	39881 ns	0.99
array/reductions/mapreduce/Int64/dims=1	40578 ns	40406 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	88500 ns	88686 ns	1.00
array/reductions/mapreduce/Int64/dims=2	57712 ns	58014 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	83652 ns	83765 ns	1.00
array/reductions/reduce/Float32/1d	32347 ns	32834 ns	0.99
array/reductions/reduce/Float32/dims=1	37681 ns	37975 ns	0.99
array/reductions/reduce/Float32/dims=1L	50610 ns	50935 ns	0.99
array/reductions/reduce/Float32/dims=2	55112 ns	55515 ns	0.99
array/reductions/reduce/Float32/dims=2L	67373 ns	67712 ns	0.99
array/reductions/reduce/Int64/1d	39300 ns	40661 ns	0.97
array/reductions/reduce/Int64/dims=1	40485 ns	40565 ns	1.00
array/reductions/reduce/Int64/dims=1L	88478 ns	88618 ns	1.00
array/reductions/reduce/Int64/dims=2	57814 ns	57805 ns	1.00
array/reductions/reduce/Int64/dims=2L	83488 ns	83628 ns	1.00
array/reverse/1d	15748 ns	16887 ns	0.93
array/reverse/1dL	69618 ns	69686 ns	1.00
array/reverse/1dL_inplace	67283 ns	67634 ns	0.99
array/reverse/1d_inplace	8387.666666666666 ns	8774.333333333334 ns	0.96
array/reverse/2d	20172 ns	20484 ns	0.98
array/reverse/2dL	73669 ns	73815 ns	1.00
array/reverse/2dL_inplace	67521 ns	67476 ns	1.00
array/reverse/2d_inplace	12151 ns	10081 ns	1.21
array/sorting/1d	2643503 ns	2636501 ns	1.00
array/sorting/2d	1017532 ns	1018106 ns	1.00
array/sorting/by	3173019 ns	3173453 ns	1.00
cuda/synchronization/context/auto	1003.7 ns	1036.3 ns	0.97
cuda/synchronization/context/blocking	789.09375 ns	806.7 ns	0.98
cuda/synchronization/context/nonblocking	5811.666666666667 ns	5925.833333333333 ns	0.98
cuda/synchronization/stream/auto	868.0555555555555 ns	875.42 ns	0.99
cuda/synchronization/stream/blocking	677.7615894039735 ns	668.2352941176471 ns	1.01
cuda/synchronization/stream/nonblocking	5598 ns	5835.428571428572 ns	0.96
integration/byval/reference	147827 ns	148006 ns	1.00
integration/byval/slices=1	148997 ns	149001 ns	1.00
integration/byval/slices=2	291830 ns	292091 ns	1.00
integration/byval/slices=3	434515 ns	435187 ns	1.00
integration/cudadevrt	104953 ns	105008 ns	1.00
integration/volumerhs	9149895 ns	9143150 ns	1.00
kernel/indexing	12960 ns	13076 ns	0.99
kernel/indexing_checked	13461 ns	13739 ns	0.98
kernel/launch	2066.222222222222 ns	2367.8888888888887 ns	0.87
kernel/occupancy	752.6324786324786 ns	956.5652173913044 ns	0.79
kernel/rand	15739 ns	14149 ns	1.11
latency/import	4221177420 ns	4204215460 ns	1.00
latency/precompile	4976875259 ns	4989180514 ns	1.00
latency/ttfp	4680902562 ns	4686057154 ns	1.00

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