Fix rocFFT error with multiple tasks

[luraess]

Fix rocfft_status_failure when executing FFT on a spawned Julia task (Julia 1.12+)

In Julia 1.12, Threads.@spawn seems to inherits task-local storage (via Base.copy?), so the spawned task sees the same HIPStream as the parent. update_stream! was guarded by plan.stream != new_stream, causing it to skip rocfft_execution_info_set_stream, however rocFFT requires this call to be made on the same OS thread as rocfft_execute. With multiple Julia threads the spawned task can land on a different OS thread, causing rocfft_status_failure.

The fix is to call rocfft_execution_info_set_stream unconditionally before every execution, dropping the plan.stream != new_stream guard. This is unobservable in a (single-threaded) REPL (all tasks share one OS thread?) but it manifests sometimes in CI.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: bfc256a	Previous: 756602c	Ratio
amdgpu/synchronization/context/device	610 ns	600 ns	1.02
amdgpu/synchronization/stream/blocking	250 ns	240 ns	1.04
amdgpu/synchronization/stream/nonblocking	340 ns	340 ns	1
array/accumulate/Float32/1d	86361 ns	86251 ns	1.00
array/accumulate/Float32/dims=1	397256 ns	393845 ns	1.01
array/accumulate/Float32/dims=1L	135712 ns	131681 ns	1.03
array/accumulate/Float32/dims=2	133462 ns	103022 ns	1.30
array/accumulate/Float32/dims=2L	2805579 ns	2827930 ns	0.99
array/accumulate/Int64/1d	96171 ns	96412 ns	1.00
array/accumulate/Int64/dims=1	407356 ns	285244 ns	1.43
array/accumulate/Int64/dims=1L	167053 ns	160812 ns	1.04
array/accumulate/Int64/dims=2	126791 ns	120772 ns	1.05
array/accumulate/Int64/dims=2L	2987371 ns	3014433 ns	0.99
array/broadcast	93662 ns	128932 ns	0.73
array/construct	1590 ns	1680 ns	0.95
array/copy	37621 ns	39371 ns	0.96
array/copyto!/cpu_to_gpu	184263 ns	114832 ns	1.60
array/copyto!/gpu_to_cpu	183652 ns	152432 ns	1.20
array/copyto!/gpu_to_gpu	126892 ns	88321 ns	1.44
array/iteration/findall/bool	179892 ns	181912 ns	0.99
array/iteration/findall/int	187303 ns	190933 ns	0.98
array/iteration/findfirst/bool	123721 ns	114451 ns	1.08
array/iteration/findfirst/int	118372 ns	116331 ns	1.02
array/iteration/findmin/1d	166743 ns	166203 ns	1.00
array/iteration/findmin/2d	155752 ns	156173 ns	1.00
array/iteration/logical	348885 ns	346025 ns	1.01
array/iteration/scalar	288354 ns	289864 ns	0.99
array/permutedims/2d	74901 ns	64761 ns	1.16
array/permutedims/3d	74231 ns	73791 ns	1.01
array/permutedims/4d	76831 ns	76481 ns	1.00
array/random/rand/Float32	50981 ns	51540 ns	0.99
array/random/rand/Int64	57291 ns	56210 ns	1.02
array/random/rand!/Float32	92261 ns	142162 ns	0.65
array/random/rand!/Int64	116052 ns	141832 ns	0.82
array/random/randn/Float32	100182 ns	86921 ns	1.15
array/random/randn!/Float32	116242 ns	152202 ns	0.76
array/reductions/mapreduce/Float32/1d	132841 ns	132902 ns	1.00
array/reductions/mapreduce/Float32/dims=1	95592 ns	95052 ns	1.01
array/reductions/mapreduce/Float32/dims=1L	772881 ns	777081 ns	0.99
array/reductions/mapreduce/Float32/dims=2	96871 ns	96731 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	298774 ns	299584 ns	1.00
array/reductions/mapreduce/Int64/1d	134062 ns	133322 ns	1.01
array/reductions/mapreduce/Int64/dims=1	95462 ns	78081 ns	1.22
array/reductions/mapreduce/Int64/dims=1L	783781 ns	783471 ns	1.00
array/reductions/mapreduce/Int64/dims=2	94881 ns	96252 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	300434 ns	308254 ns	0.97
array/reductions/reduce/Float32/1d	132662 ns	132802 ns	1.00
array/reductions/reduce/Float32/dims=1	95001 ns	94832 ns	1.00
array/reductions/reduce/Float32/dims=1L	773371 ns	774621 ns	1.00
array/reductions/reduce/Float32/dims=2	97222 ns	96802 ns	1.00
array/reductions/reduce/Float32/dims=2L	297484 ns	307245 ns	0.97
array/reductions/reduce/Int64/1d	133152 ns	129672 ns	1.03
array/reductions/reduce/Int64/dims=1	94852 ns	78151 ns	1.21
array/reductions/reduce/Int64/dims=1L	780821 ns	781931 ns	1.00
array/reductions/reduce/Int64/dims=2	96111 ns	96192 ns	1.00
array/reductions/reduce/Int64/dims=2L	303334 ns	298414 ns	1.02
array/reverse/1d	43921 ns	44380 ns	0.99
array/reverse/1dL	75401 ns	74131 ns	1.02
array/reverse/1dL_inplace	112692 ns	108282 ns	1.04
array/reverse/1d_inplace	77891 ns	86471 ns	0.90
array/reverse/2d	52211 ns	50661 ns	1.03
array/reverse/2dL	101491 ns	100341 ns	1.01
array/reverse/2dL_inplace	130102 ns	117622 ns	1.11
array/reverse/2d_inplace	79911 ns	95391 ns	0.84
array/sorting/1d	341775 ns	341945 ns	1.00
integration/byval/reference	39130 ns	38830 ns	1.01
integration/byval/slices=1	40161 ns	40880 ns	0.98
integration/byval/slices=2	140522 ns	158462 ns	0.89
integration/byval/slices=3	237983 ns	238013 ns	1.00
integration/volumerhs	5050459 ns	4942659 ns	1.02
kernel/indexing	129511 ns	43630 ns	2.97
kernel/indexing_checked	124942 ns	128022 ns	0.98
kernel/launch	1310 ns	1290 ns	1.02
kernel/rand	123941 ns	106671 ns	1.16
latency/import	1493016620 ns	1501349912 ns	0.99
latency/precompile	11942427457 ns	12041117438 ns	0.99
latency/ttfp	10901786827 ns	10491950084 ns	1.04

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

[luraess]

cscs-ci run