[luraess]

The "Wrap Array" test in test/core/rocarray_base.jl waits for a wrapped host array to be collected once the device is done with it. It bounded that wait by 10 s of wall-clock time, which stalled MI250 CI workers exceeded twice while everything else was green (#1162 on 2026-10-16, #1185 on 2026-10-23). In both runs that file took ~240 s instead of ~35 s, a known pattern on that partition.

The release needs two collections with a device round trip in between: the first finalizes the wrapper, whose finalizer enqueues a host function on the stream; the device reaching it wakes the task that lets go of the array; the second collects the array itself. The test now synchronizes the stream each round, so the device step is explicit, and bounds the wait by rounds rather than by time. It can only fail if the array is genuinely never released.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 2005202	Previous: 74f1eda	Ratio
amdgpu/synchronization/context/device	1775 ns	1810 ns	0.98
amdgpu/synchronization/stream/blocking	240 ns	235 ns	1.02
amdgpu/synchronization/stream/nonblocking	342.5 ns	350 ns	0.98
applications/bitonic_sort	1119541 ns	1110311 ns	1.01
applications/convolution	77026 ns	76166.25 ns	1.01
applications/floyd_warshall	9078637 ns	9085184.75 ns	1.00
applications/histogram	807741.75 ns	790563.75 ns	1.02
applications/prefix_sum	190390.25 ns	189820.25 ns	1.00
array/accumulate/Float32/1d	90856.25 ns	89828.75 ns	1.01
array/accumulate/Float32/dims=1	70786 ns	69438.5 ns	1.02
array/accumulate/Float32/dims=1L	105206.5 ns	104426.5 ns	1.01
array/accumulate/Float32/dims=2	105504 ns	104076.5 ns	1.01
array/accumulate/Float32/dims=2L	126794.25 ns	126476.75 ns	1.00
array/accumulate/Int64/1d	115884.25 ns	114141.75 ns	1.02
array/accumulate/Int64/dims=1	70476 ns	82011.25 ns	0.86
array/accumulate/Int64/dims=1L	113461.5 ns	102219 ns	1.11
array/accumulate/Int64/dims=2	109409 ns	108246.5 ns	1.01
array/accumulate/Int64/dims=2L	185112.75 ns	174135 ns	1.06
array/broadcast	57203.25 ns	58063.25 ns	0.99
array/construct	2192.5 ns	2317.5 ns	0.95
array/copy	39513 ns	33980.5 ns	1.16
array/copyto!/cpu_to_gpu	86073.75 ns	86038.75 ns	1.00
array/copyto!/gpu_to_cpu	86198.75 ns	85926.25 ns	1.00
array/copyto!/gpu_to_gpu	37868 ns	58045.75 ns	0.65
array/iteration/findall/bool	90733.75 ns	94236.25 ns	0.96
array/iteration/findall/int	107244 ns	106511.75 ns	1.01
array/iteration/findfirst/bool	55985.75 ns	54890.75 ns	1.02
array/iteration/findfirst/int	53738.25 ns	50660.75 ns	1.06
array/iteration/findmin/1d	90161.25 ns	94756.5 ns	0.95
array/iteration/findmin/2d	83546.25 ns	83726.25 ns	1.00
array/iteration/logical	91198.75 ns	89338.75 ns	1.02
array/iteration/scalar	304929.5 ns	295404.25 ns	1.03
array/permutedims/2d	49373.25 ns	61028.5 ns	0.81
array/permutedims/3d	48573.25 ns	58453.5 ns	0.83
array/permutedims/4d	62916 ns	50335.75 ns	1.25
array/random/rand/Float32	45580.5 ns	45608.25 ns	1.00
array/random/rand/Int64	45720.75 ns	45390.75 ns	1.01
array/random/rand!/Float32	66683.5 ns	65476 ns	1.02
array/random/rand!/Int64	64398.25 ns	63423.5 ns	1.02
array/random/randn/Float32	69291 ns	70596 ns	0.98
array/random/randn!/Float32	59711 ns	82401 ns	0.72
array/reductions/mapreduce/Float32/1d	61716 ns	69948.5 ns	0.88
array/reductions/mapreduce/Float32/dims=1	47513.25 ns	55098.25 ns	0.86
array/reductions/mapreduce/Float32/dims=1L	52165.75 ns	59223.5 ns	0.88
array/reductions/mapreduce/Float32/dims=2	53073.5 ns	58411 ns	0.91
array/reductions/mapreduce/Float32/dims=2L	62433.5 ns	73278.75 ns	0.85
array/reductions/mapreduce/Int64/1d	61996 ns	71083.5 ns	0.87
array/reductions/mapreduce/Int64/dims=1	51560.75 ns	51328.25 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	70621 ns	70761 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59048.25 ns	52118.25 ns	1.13
array/reductions/mapreduce/Int64/dims=2L	76411 ns	78523.5 ns	0.97
array/reductions/reduce/Float32/1d	70263.5 ns	70158.5 ns	1.00
array/reductions/reduce/Float32/dims=1	54605.75 ns	55425.75 ns	0.99
array/reductions/reduce/Float32/dims=1L	59920.75 ns	59165.75 ns	1.01
array/reductions/reduce/Float32/dims=2	60665.75 ns	59930.75 ns	1.01
array/reductions/reduce/Float32/dims=2L	72688.5 ns	72888.75 ns	1.00
array/reductions/reduce/Int64/1d	61248.5 ns	70608.75 ns	0.87
array/reductions/reduce/Int64/dims=1	51570.75 ns	43910.5 ns	1.17
array/reductions/reduce/Int64/dims=1L	70403.5 ns	70306 ns	1.00
array/reductions/reduce/Int64/dims=2	51750.5 ns	60678.5 ns	0.85
array/reductions/reduce/Int64/dims=2L	76303.5 ns	78253.5 ns	0.98
array/reverse/1d	40340.5 ns	41730.5 ns	0.97
array/reverse/1dL	66036 ns	66368.5 ns	0.99
array/reverse/1dL_inplace	77716 ns	73683.5 ns	1.05
array/reverse/1d_inplace	40183 ns	42845.75 ns	0.94
array/reverse/2d	48105.75 ns	47903.25 ns	1.00
array/reverse/2dL	57550.75 ns	77728.5 ns	0.74
array/reverse/2dL_inplace	79503.5 ns	94111.25 ns	0.84
array/reverse/2d_inplace	46005.75 ns	43678 ns	1.05
array/sorting/1d	336180 ns	342995 ns	0.98
gemm/tiled	1928065 ns	1924275 ns	1.00
gemm/tiled_unbounded	1923572.5 ns	1941765.25 ns	0.99
integration/byval/reference	40381 ns	40450 ns	1.00
integration/byval/slices=1	40441 ns	41060 ns	0.98
integration/byval/slices=2	156602 ns	159632 ns	0.98
integration/byval/slices=3	232423 ns	235493 ns	0.99
integration/volumerhs	4852679 ns	4845429 ns	1.00
kernel/indexing	34090.5 ns	33105.5 ns	1.03
kernel/indexing_checked	49010.75 ns	42875.75 ns	1.14
kernel/launch	1235 ns	1137.5 ns	1.09
kernel/rand	72841 ns	49568 ns	1.47
latency/import	1401844621 ns	1406040209 ns	1.00
latency/precompile	29100854052 ns	29142357571 ns	1.00
latency/ttfp	1784694352 ns	1788216920 ns	1.00
stencil/diffusion3d	1594885 ns	1575350 ns	1.01
stencil/diffusion3d_checked	1639093.25 ns	1635898.25 ns	1.00

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