Switch back after device!(f, dev) on a fresh task

[luraess]

On a task with no AMDGPU state yet, context!(ctx) created that state on ctx and returned nothing, so context!(f, ctx) and device!(f, dev) never switched back and left the task on dev. context! now starts from the default state, so there is always a context to switch back to. This also drops the ctx.valid check, which can no longer fail since #605.

pool_cleanup wrapped reclaim() in HIP.device!, but reclaim reads the task-local device, so it trimmed the wrong device's pool. It now uses AMDGPU.device!.

Items 2 and 4 of #1105. The new multi_gpu test fails on main and passes here (mi200, 2 GCDs).

🤖 Generated with Claude Code

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: ba0404e	Previous: 8a66547	Ratio
amdgpu/synchronization/context/device	547.5 ns	550 ns	1.00
amdgpu/synchronization/stream/blocking	225 ns	222.5 ns	1.01
amdgpu/synchronization/stream/nonblocking	305 ns	307.5 ns	0.99
applications/bitonic_sort	1127180.75 ns	1160103.75 ns	0.97
applications/convolution	72141 ns	72376 ns	1.00
applications/floyd_warshall	9023781.25 ns	9100856.25 ns	0.99
applications/histogram	806093.75 ns	798998.75 ns	1.01
applications/prefix_sum	187992.75 ns	190535 ns	0.99
array/accumulate/Float32/1d	69691 ns	79346 ns	0.88
array/accumulate/Float32/dims=1	285226.5 ns	271536.25 ns	1.05
array/accumulate/Float32/dims=1L	83868.75 ns	72536 ns	1.16
array/accumulate/Float32/dims=2	74618.5 ns	74041 ns	1.01
array/accumulate/Float32/dims=2L	3028477.5 ns	2834326.75 ns	1.07
array/accumulate/Int64/1d	79803.5 ns	79098.5 ns	1.01
array/accumulate/Int64/dims=1	248723.5 ns	248393.25 ns	1.00
array/accumulate/Int64/dims=1L	88681.25 ns	89128.75 ns	0.99
array/accumulate/Int64/dims=2	75586 ns	76041 ns	0.99
array/accumulate/Int64/dims=2L	3389059.75 ns	3139866 ns	1.08
array/broadcast	43978.25 ns	50443.25 ns	0.87
array/construct	2192.5 ns	2247.5 ns	0.98
array/copy	39413 ns	39278 ns	1.00
array/copyto!/cpu_to_gpu	85866.25 ns	85676.25 ns	1.00
array/copyto!/gpu_to_cpu	85896 ns	85591.25 ns	1.00
array/copyto!/gpu_to_gpu	58961 ns	58288.25 ns	1.01
array/iteration/findall/bool	128669.25 ns	125954.25 ns	1.02
array/iteration/findall/int	144622 ns	146617 ns	0.99
array/iteration/findfirst/bool	156462.25 ns	155292.25 ns	1.01
array/iteration/findfirst/int	191840.25 ns	189260 ns	1.01
array/iteration/findmin/1d	103671.5 ns	103258.75 ns	1.00
array/iteration/findmin/2d	90151.25 ns	90081.25 ns	1.00
array/iteration/logical	227883.25 ns	226565.5 ns	1.01
array/iteration/scalar	296136.75 ns	295931.5 ns	1.00
array/permutedims/2d	70831 ns	70736 ns	1.00
array/permutedims/3d	47465.75 ns	69253.5 ns	0.69
array/permutedims/4d	75038.75 ns	50278.25 ns	1.49
array/random/rand/Float32	46080.75 ns	40065.75 ns	1.15
array/random/rand/Int64	55490.75 ns	48863 ns	1.14
array/random/rand!/Float32	44710.5 ns	65541 ns	0.68
array/random/rand!/Int64	57668.25 ns	69623.25 ns	0.83
array/random/randn/Float32	69196 ns	71161 ns	0.97
array/random/randn!/Float32	81926 ns	58721 ns	1.40
array/reductions/mapreduce/Float32/1d	81693.75 ns	81621 ns	1.00
array/reductions/mapreduce/Float32/dims=1	74416.25 ns	56660.75 ns	1.31
array/reductions/mapreduce/Float32/dims=1L	840539.25 ns	844514.25 ns	1.00
array/reductions/mapreduce/Float32/dims=2	75808.75 ns	74073.5 ns	1.02
array/reductions/mapreduce/Float32/dims=2L	136992 ns	136524.5 ns	1.00
array/reductions/mapreduce/Int64/1d	81546 ns	81663.5 ns	1.00
array/reductions/mapreduce/Int64/dims=1	74981 ns	73933.5 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	842879.5 ns	839174.25 ns	1.00
array/reductions/mapreduce/Int64/dims=2	75191 ns	75486.25 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	137824.25 ns	137611.75 ns	1.00
array/reductions/reduce/Float32/1d	81348.5 ns	81541.25 ns	1.00
array/reductions/reduce/Float32/dims=1	74341.25 ns	70296 ns	1.06
array/reductions/reduce/Float32/dims=1L	839396.75 ns	836984.25 ns	1.00
array/reductions/reduce/Float32/dims=2	75016 ns	75123.5 ns	1.00
array/reductions/reduce/Float32/dims=2L	136919.5 ns	136774.25 ns	1.00
array/reductions/reduce/Int64/1d	81423.5 ns	81368.75 ns	1.00
array/reductions/reduce/Int64/dims=1	74696 ns	74831 ns	1.00
array/reductions/reduce/Int64/dims=1L	843806.75 ns	844554 ns	1.00
array/reductions/reduce/Int64/dims=2	74978.5 ns	75411.25 ns	0.99
array/reductions/reduce/Int64/dims=2L	137827 ns	137047 ns	1.01
array/reverse/1d	43200.5 ns	43458 ns	0.99
array/reverse/1dL	69886 ns	70606 ns	0.99
array/reverse/1dL_inplace	79388.75 ns	79671 ns	1.00
array/reverse/1d_inplace	33625.5 ns	60390.75 ns	0.56
array/reverse/2d	47708.25 ns	47625.75 ns	1.00
array/reverse/2dL	57253.25 ns	57993.5 ns	0.99
array/reverse/2dL_inplace	79476.25 ns	79096 ns	1.00
array/reverse/2d_inplace	60936 ns	61191 ns	1.00
array/sorting/1d	323484.5 ns	326449.5 ns	0.99
gemm/tiled	2888903 ns	1961659.75 ns	1.47
gemm/tiled_unbounded	1956694.75 ns	1960334.75 ns	1.00
integration/byval/reference	39611 ns	39821 ns	0.99
integration/byval/slices=1	40481 ns	40830 ns	0.99
integration/byval/slices=2	158163 ns	130842 ns	1.21
integration/byval/slices=3	240243 ns	241884 ns	0.99
integration/volumerhs	4877228 ns	4900148 ns	1.00
kernel/indexing	35395.5 ns	41880.5 ns	0.85
kernel/indexing_checked	57765.75 ns	37075.5 ns	1.56
kernel/launch	1155 ns	1130 ns	1.02
kernel/rand	70338.5 ns	70343.5 ns	1.00
latency/import	1457911761 ns	1456646355 ns	1.00
latency/precompile	22612886279 ns	22642954217 ns	1.00
latency/ttfp	2246180072 ns	2251455325 ns	1.00
stencil/diffusion3d	1616775.25 ns	1611230 ns	1.00
stencil/diffusion3d_checked	1658688.25 ns	1655470.5 ns	1.00

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