[luraess]

Fixes a regression from #1120 that made core/codegen "Launch bounds" fail with hipErrorInvalidHandle on CSCS MI250 (Julia 1.13, main 9d33b4b).

HIP.task_stream picks the pool from HIP.device(). That is hipGetDevice, the one wrapper without prepare_state(), so it returns the device the OS thread was last used with, not the task's device. A fresh task on device 2, running on a thread last used with device 1, claims an idle device-1 stream. Its module is loaded on device 2, and HIP rejects the launch because the stream belongs to another device (ihipModuleLaunchKernel: stream device != ihipGetDevice() → hipErrorInvalidResourceHandle).

In CI, the KernelInterface testsuite's cross-device events test leaves device 2 as the default device (#1105, item 3), and its main task leaves an idle device-1 stream in the pool. core/codegen ran next on the same worker, and "Launch bounds" is its only launch.

The fix picks the pool from AMDGPU.device(), the task's device. That matches how test/core/tls.jl already looks up the pool.

Validated on MI250X (beverin mi200, Julia 1.13):

• new multi_gpu test "Pooled streams match the task's device": fails on main (47/48), passes with the fix (48/48)
• kernelinterface_tests → core/codegen in the same process: every launch fails on main, all pass with the fix
• core/tls passes

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 291ec15	Previous: 353b621	Ratio
amdgpu/synchronization/context/device	1787.5 ns	1812.5 ns	0.99
amdgpu/synchronization/stream/blocking	235 ns	225 ns	1.04
amdgpu/synchronization/stream/nonblocking	350 ns	340 ns	1.03
applications/bitonic_sort	1112918 ns	1105353.75 ns	1.01
applications/convolution	80536.25 ns	73976 ns	1.09
applications/floyd_warshall	8999922.75 ns	9054960.25 ns	0.99
applications/histogram	812431.25 ns	793314 ns	1.02
applications/prefix_sum	190512.75 ns	190742.75 ns	1.00
array/accumulate/Float32/1d	91579 ns	92523.75 ns	0.99
array/accumulate/Float32/dims=1	68426 ns	60563.5 ns	1.13
array/accumulate/Float32/dims=1L	87313.75 ns	87806.25 ns	0.99
array/accumulate/Float32/dims=2	103901.25 ns	88358.75 ns	1.18
array/accumulate/Float32/dims=2L	130944.25 ns	175740.25 ns	0.75
array/accumulate/Int64/1d	108129 ns	90441.5 ns	1.20
array/accumulate/Int64/dims=1	82528.5 ns	70633.5 ns	1.17
array/accumulate/Int64/dims=1L	98301.25 ns	92791.5 ns	1.06
array/accumulate/Int64/dims=2	108511.5 ns	92341.25 ns	1.18
array/accumulate/Int64/dims=2L	196562.75 ns	198483 ns	0.99
array/broadcast	53788.25 ns	57918.5 ns	0.93
array/construct	2390.25 ns	2355 ns	1.01
array/copy	39098 ns	38483 ns	1.02
array/copyto!/cpu_to_gpu	86036.25 ns	85918.75 ns	1.00
array/copyto!/gpu_to_cpu	85791.25 ns	86248.75 ns	0.99
array/copyto!/gpu_to_gpu	58781 ns	59498.25 ns	0.99
array/iteration/findall/bool	115549.25 ns	93529 ns	1.24
array/iteration/findall/int	129084.25 ns	104056.5 ns	1.24
array/iteration/findfirst/bool	49550.75 ns	56560.75 ns	0.88
array/iteration/findfirst/int	51920.75 ns	56940.75 ns	0.91
array/iteration/findmin/1d	79881 ns	97464 ns	0.82
array/iteration/findmin/2d	67583.5 ns	80563.5 ns	0.84
array/iteration/logical	91838.75 ns	93809 ns	0.98
array/iteration/scalar	295784.25 ns	296092 ns	1.00
array/permutedims/2d	60703.5 ns	61688.5 ns	0.98
array/permutedims/3d	53105.5 ns	59256 ns	0.90
array/permutedims/4d	61910.75 ns	63150.75 ns	0.98
array/random/rand/Float32	45160.5 ns	46740.75 ns	0.97
array/random/rand/Int64	37873 ns	46675.75 ns	0.81
array/random/rand!/Float32	65695.75 ns	63673.5 ns	1.03
array/random/rand!/Int64	47825.75 ns	47478.25 ns	1.01
array/random/randn/Float32	73983.5 ns	72703.5 ns	1.02
array/random/randn!/Float32	82866 ns	82786.25 ns	1.00
array/reductions/mapreduce/Float32/1d	68823.5 ns	60425.75 ns	1.14
array/reductions/mapreduce/Float32/dims=1	43063 ns	43863 ns	0.98
array/reductions/mapreduce/Float32/dims=1L	58333.25 ns	51120.75 ns	1.14
array/reductions/mapreduce/Float32/dims=2	59455.75 ns	51973.5 ns	1.14
array/reductions/mapreduce/Float32/dims=2L	71441 ns	60900.75 ns	1.17
array/reductions/mapreduce/Int64/1d	70656 ns	61768.25 ns	1.14
array/reductions/mapreduce/Int64/dims=1	50313.25 ns	44065.75 ns	1.14
array/reductions/mapreduce/Int64/dims=1L	69691 ns	71006 ns	0.98
array/reductions/mapreduce/Int64/dims=2	58010.75 ns	50543.25 ns	1.15
array/reductions/mapreduce/Int64/dims=2L	74773.5 ns	76458.75 ns	0.98
array/reductions/reduce/Float32/1d	68635.75 ns	70176.25 ns	0.98
array/reductions/reduce/Float32/dims=1	49548 ns	51108.25 ns	0.97
array/reductions/reduce/Float32/dims=1L	58450.75 ns	59610.75 ns	0.98
array/reductions/reduce/Float32/dims=2	59143.5 ns	60643.5 ns	0.98
array/reductions/reduce/Float32/dims=2L	59298.25 ns	71181 ns	0.83
array/reductions/reduce/Int64/1d	70185.75 ns	72013.5 ns	0.97
array/reductions/reduce/Int64/dims=1	50028.25 ns	51468.25 ns	0.97
array/reductions/reduce/Int64/dims=1L	56020.75 ns	62391 ns	0.90
array/reductions/reduce/Int64/dims=2	57773.25 ns	59108.5 ns	0.98
array/reductions/reduce/Int64/dims=2L	61478.25 ns	62808.5 ns	0.98
array/reverse/1d	39800.5 ns	41488.25 ns	0.96
array/reverse/1dL	59875.75 ns	61570.75 ns	0.97
array/reverse/1dL_inplace	78263.75 ns	56503.25 ns	1.39
array/reverse/1d_inplace	46490.75 ns	39880.5 ns	1.17
array/reverse/2d	47400.5 ns	48433.25 ns	0.98
array/reverse/2dL	78006 ns	79033.75 ns	0.99
array/reverse/2dL_inplace	99731.25 ns	94686.5 ns	1.05
array/reverse/2d_inplace	57540.75 ns	67056 ns	0.86
array/sorting/1d	334162 ns	336870 ns	0.99
gemm/tiled	1905571.25 ns	1936281 ns	0.98
gemm/tiled_unbounded	1920509.25 ns	1949061 ns	0.99
integration/byval/reference	39140 ns	40880 ns	0.96
integration/byval/slices=1	40201 ns	41501 ns	0.97
integration/byval/slices=2	138162 ns	139832 ns	0.99
integration/byval/slices=3	235713 ns	239204 ns	0.99
integration/volumerhs	4866018 ns	4832111 ns	1.01
kernel/indexing	54630.75 ns	45395.5 ns	1.20
kernel/indexing_checked	47450.75 ns	59176 ns	0.80
kernel/launch	1245 ns	1230 ns	1.01
kernel/rand	72033.5 ns	72708.75 ns	0.99
latency/import	1958672081 ns	1953052041 ns	1.00
latency/precompile	34529641898 ns	34369458538 ns	1.00
latency/ttfp	2501540707 ns	2491942642 ns	1.00
stencil/diffusion3d	1594964.5 ns	1595708.5 ns	1.00
stencil/diffusion3d_checked	1633037.75 ns	1633254 ns	1.00

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