Transfer buffer ownership at conversion time instead of a per-launch locked phase

[gbaraldi]

While comparing host-side overheads against CUDA.jl on an MI300A, the kernel-launch path stood out: every launch allocates a Vector{Managed}, rocconvert pushes each argument's buffer into it, and a with_managed phase then does unique + sort(by=objectid) + lock/unlock of every buffer's ReentrantLock around the launch. CUDA.jl performs the equivalent stream-ownership tracking inline in pointer conversion, with no per-launch collection phase.

This PR adopts that design:

• Runtime.Adaptor carries the launching stream instead of a Managed collector (nothing skips ownership handling, for bare rocconvert used in reflection/argconvert);
• the ROCArray adapt rule transfers ownership inline via a new take_ownership_fast!: an unlocked fast path for the common case (buffer already owned by this stream and already dirty), falling back to the locked take_ownership! otherwise. The unsynchronized read only races with usage patterns that were already racy under the locked protocol (concurrent use of one array from tasks on different streams);
• the HIPKernel call operator drops the vector and the with_managed wrapper. with_managed/lock_managed remain for external users.

Measured on MI300A (gfx942, ROCm 7.2.4, Julia 1.12.6), empty kernel / 3-array+4-scalar kernel:

metric	before	after
no-arg kernel-object launch	4,781 ns / 144 B	4,649 ns / 0 B
launch with 3 arrays + 4 scalars	7,354 ns / 1,232 B	6,596 ns / 416 B
small broadcast c .= a .+ b	15,437 ns / 3,056 B	14,685 ns / 2,320 B

Correctness validated on hardware including the cross-stream case (array owned by the default stream, used from a second stream via stream! — the locked slow path still synchronizes the previous stream).

For context from the same comparison: with this change the Julia-side launch overhead is essentially at parity with CUDA.jl (burst submission 3.7 vs 3.1 µs, and raw hipModuleLaunchKernel costs as much as the full AMDGPU.jl path) — the remaining gaps vs CUDA (launch throughput, 17 µs launch+sync round trip, and 10 µs hipMallocFromPoolAsync+hipFreeAsync pairs vs 1.5 µs on CUDA) are in the HIP runtime, not in this package.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 56a4329	Previous: 16f5974	Ratio
amdgpu/synchronization/context/device	552.5 ns	755 ns	0.73
amdgpu/synchronization/stream/blocking	235 ns	317.5 ns	0.74
amdgpu/synchronization/stream/nonblocking	317.5 ns	432.25 ns	0.73
array/accumulate/Float32/1d	74421 ns	97092.25 ns	0.77
array/accumulate/Float32/dims=1	278646.25 ns	299718.5 ns	0.93
array/accumulate/Float32/dims=1L	79666.25 ns	121446.75 ns	0.66
array/accumulate/Float32/dims=2	70461 ns	130896.75 ns	0.54
array/accumulate/Float32/dims=2L	2757744.75 ns	2806321.5 ns	0.98
array/accumulate/Int64/1d	77241 ns	113217 ns	0.68
array/accumulate/Int64/dims=1	242483.25 ns	269376.25 ns	0.90
array/accumulate/Int64/dims=1L	84176.25 ns	144501.75 ns	0.58
array/accumulate/Int64/dims=2	84701.25 ns	127879.25 ns	0.66
array/accumulate/Int64/dims=2L	2897086.5 ns	2948585.75 ns	0.98
array/broadcast	73251 ns	64054.75 ns	1.14
array/construct	2065 ns	3220 ns	0.64
array/copy	37335.5 ns	39815 ns	0.94
array/copyto!/cpu_to_gpu	111581.5 ns	94539.5 ns	1.18
array/copyto!/gpu_to_cpu	120241.5 ns	95039.5 ns	1.27
array/copyto!/gpu_to_gpu	59270.75 ns	42115 ns	1.41
array/iteration/findall/bool	131351.75 ns	198036.5 ns	0.66
array/iteration/findall/int	148492 ns	207411.5 ns	0.72
array/iteration/findfirst/bool	145439.25 ns	167311.75 ns	0.87
array/iteration/findfirst/int	144769.25 ns	155156.75 ns	0.93
array/iteration/findmin/1d	109584 ns	179444.25 ns	0.61
array/iteration/findmin/2d	96596.25 ns	161089.25 ns	0.60
array/iteration/logical	234700.75 ns	317088.25 ns	0.74
array/iteration/scalar	302924 ns	343653.25 ns	0.88
array/permutedims/2d	71433.5 ns	82694.75 ns	0.86
array/permutedims/3d	71073.5 ns	82232.25 ns	0.86
array/permutedims/4d	73741 ns	84369.75 ns	0.87
array/random/rand/Float32	46008.25 ns	53677.5 ns	0.86
array/random/rand/Int64	54220.75 ns	66274.75 ns	0.82
array/random/rand!/Float32	66036 ns	49389.75 ns	1.34
array/random/rand!/Int64	73293.5 ns	59654.75 ns	1.23
array/random/randn/Float32	78516.25 ns	93112.25 ns	0.84
array/random/randn!/Float32	77758.5 ns	69269.75 ns	1.12
array/reductions/mapreduce/Float32/1d	86956.25 ns	141249.25 ns	0.62
array/reductions/mapreduce/Float32/dims=1	81748.5 ns	105462 ns	0.78
array/reductions/mapreduce/Float32/dims=1L	834063.5 ns	862533.5 ns	0.97
array/reductions/mapreduce/Float32/dims=2	82713.5 ns	108059.5 ns	0.77
array/reductions/mapreduce/Float32/dims=2L	143542 ns	155446.75 ns	0.92
array/reductions/mapreduce/Int64/1d	86626 ns	140932 ns	0.61
array/reductions/mapreduce/Int64/dims=1	81926.25 ns	104764.5 ns	0.78
array/reductions/mapreduce/Int64/dims=1L	830286 ns	860696 ns	0.96
array/reductions/mapreduce/Int64/dims=2	82653.75 ns	107757 ns	0.77
array/reductions/mapreduce/Int64/dims=2L	144024.5 ns	163496.75 ns	0.88
array/reductions/reduce/Float32/1d	86826 ns	145184.25 ns	0.60
array/reductions/reduce/Float32/dims=1	81471 ns	105042 ns	0.78
array/reductions/reduce/Float32/dims=1L	831651 ns	856898.25 ns	0.97
array/reductions/reduce/Float32/dims=2	82553.5 ns	106659.5 ns	0.77
array/reductions/reduce/Float32/dims=2L	143539.25 ns	161129.25 ns	0.89
array/reductions/reduce/Int64/1d	86686 ns	141302 ns	0.61
array/reductions/reduce/Int64/dims=1	81928.5 ns	105144.5 ns	0.78
array/reductions/reduce/Int64/dims=1L	832378.5 ns	851721 ns	0.98
array/reductions/reduce/Int64/dims=2	82113.75 ns	107839.5 ns	0.76
array/reductions/reduce/Int64/dims=2L	144034.5 ns	162989.25 ns	0.88
array/reverse/1d	45635.75 ns	53687.25 ns	0.85
array/reverse/1dL	72140.75 ns	78502.25 ns	0.92
array/reverse/1dL_inplace	60095.75 ns	62952.25 ns	0.95
array/reverse/1d_inplace	54553.25 ns	45130 ns	1.21
array/reverse/2d	49178.25 ns	55499.75 ns	0.89
array/reverse/2dL	81723.75 ns	111822 ns	0.73
array/reverse/2dL_inplace	92346.25 ns	74057 ns	1.25
array/reverse/2d_inplace	44235.5 ns	49312.25 ns	0.90
array/sorting/1d	332882.25 ns	359238.75 ns	0.93
integration/byval/reference	39380 ns	41860 ns	0.94
integration/byval/slices=1	40310 ns	42700 ns	0.94
integration/byval/slices=2	125802 ns	133939 ns	0.94
integration/byval/slices=3	234613 ns	239880 ns	0.98
integration/volumerhs	4999428 ns	4916200 ns	1.02
kernel/indexing	34753 ns	44594.75 ns	0.78
kernel/indexing_checked	49023 ns	47782.25 ns	1.03
kernel/launch	1345 ns	1785 ns	0.75
kernel/rand	97633.75 ns	110472 ns	0.88
latency/import	1702522741 ns	2242598207 ns	0.76
latency/precompile	39534644771 ns	52102215931 ns	0.76
latency/ttfp	2328242218 ns	3070234127 ns	0.76

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

[luraess]

LGTM. Thanks, maybe here as well, one could trim inline comments to the minimum before merging.

[luraess]

Thanks!