Fix rocFFT error with multiple tasks

[luraess]

Problem

Running FFT operations on a Threads.@spawn-ed task (or any task other than the one that created the plan) would intermittently produce rocfft_status_failure. This was observed in CI when running tests in parallel on MI300.

The root cause is a constraint in the rocFFT API: rocfft_execution_info_set_stream must be called on the same OS thread as the subsequent rocfft_execute. Julia tasks can migrate between OS threads at yield points, so storing a single rocfft_execution_info per plan and updating its stream before execution is not sufficient — the task may land on a different OS thread between the stream update and the actual execute call. rocFFT's OS-thread constraint applies to the entire create → set_stream → execute sequence.

Fix

rocfft_execution_info is now created fresh per execution inside each unsafe_execute! call, rather than being stored in the plan struct. The stream and work buffer are set immediately before rocfft_execute, with no yield points between setup and execution. The execution_info is destroyed after each call.

This matches the intended C++ usage pattern: rocfft_plan is the expensive, reusable, thread-safe object; rocfft_execution_info is a lightweight per-call envelope carrying runtime context (stream, work buffer). The plan.stream and plan.execution_info fields are removed from both plan structs accordingly.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 959beea	Previous: 122a77d	Ratio
amdgpu/synchronization/context/device	600 ns	610 ns	0.98
amdgpu/synchronization/stream/blocking	240 ns	250 ns	0.96
amdgpu/synchronization/stream/nonblocking	330 ns	350 ns	0.94
array/accumulate/Float32/1d	86201 ns	87801 ns	0.98
array/accumulate/Float32/dims=1	401046 ns	337664 ns	1.19
array/accumulate/Float32/dims=1L	136262 ns	136022 ns	1.00
array/accumulate/Float32/dims=2	129491 ns	138791 ns	0.93
array/accumulate/Float32/dims=2L	2808079 ns	2812325 ns	1.00
array/accumulate/Int64/1d	97231 ns	96001 ns	1.01
array/accumulate/Int64/dims=1	283804 ns	325024 ns	0.87
array/accumulate/Int64/dims=1L	167763 ns	167872 ns	1.00
array/accumulate/Int64/dims=2	126642 ns	128101 ns	0.99
array/accumulate/Int64/dims=2L	2985942 ns	2978066 ns	1.00
array/broadcast	118422 ns	139582 ns	0.85
array/construct	1650 ns	1680 ns	0.98
array/copy	39840 ns	36000 ns	1.11
array/copyto!/cpu_to_gpu	114431 ns	98171 ns	1.17
array/copyto!/gpu_to_cpu	184013 ns	182952 ns	1.01
array/copyto!/gpu_to_gpu	66831 ns	130452 ns	0.51
array/iteration/findall/bool	180112 ns	181773 ns	0.99
array/iteration/findall/int	189393 ns	201313 ns	0.94
array/iteration/findfirst/bool	117711 ns	121972 ns	0.97
array/iteration/findfirst/int	115002 ns	116632 ns	0.99
array/iteration/findmin/1d	169213 ns	171813 ns	0.98
array/iteration/findmin/2d	155022 ns	155922 ns	0.99
array/iteration/logical	355715 ns	354734 ns	1.00
array/iteration/scalar	288554 ns	289223 ns	1.00
array/permutedims/2d	75141 ns	74161 ns	1.01
array/permutedims/3d	74591 ns	75121 ns	0.99
array/permutedims/4d	76421 ns	76851 ns	0.99
array/random/rand/Float32	51081 ns	52441 ns	0.97
array/random/rand/Int64	57730 ns	57411 ns	1.01
array/random/rand!/Float32	129092 ns	146302 ns	0.88
array/random/rand!/Int64	107592 ns	146922 ns	0.73
array/random/randn/Float32	101801 ns	89532 ns	1.14
array/random/randn!/Float32	115762 ns	160592 ns	0.72
array/reductions/mapreduce/Float32/1d	132742 ns	133381 ns	1.00
array/reductions/mapreduce/Float32/dims=1	94791 ns	95152 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	775101 ns	776099 ns	1.00
array/reductions/mapreduce/Float32/dims=2	96821 ns	97391 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	297964 ns	299993 ns	0.99
array/reductions/mapreduce/Int64/1d	133041 ns	135062 ns	0.99
array/reductions/mapreduce/Int64/dims=1	95102 ns	95441 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	780611 ns	784259 ns	1.00
array/reductions/mapreduce/Int64/dims=2	96181 ns	97071 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	296445 ns	301494 ns	0.98
array/reductions/reduce/Float32/1d	132782 ns	133202 ns	1.00
array/reductions/reduce/Float32/dims=1	94611 ns	96031 ns	0.99
array/reductions/reduce/Float32/dims=1L	774341 ns	775290 ns	1.00
array/reductions/reduce/Float32/dims=2	97251 ns	97391 ns	1.00
array/reductions/reduce/Float32/dims=2L	294274 ns	298723 ns	0.99
array/reductions/reduce/Int64/1d	132702 ns	133912 ns	0.99
array/reductions/reduce/Int64/dims=1	94921 ns	95231 ns	1.00
array/reductions/reduce/Int64/dims=1L	779441 ns	782440 ns	1.00
array/reductions/reduce/Int64/dims=2	95951 ns	96852 ns	0.99
array/reductions/reduce/Int64/dims=2L	298234 ns	298143 ns	1.00
array/reverse/1d	44210 ns	43171 ns	1.02
array/reverse/1dL	75241 ns	75361 ns	1.00
array/reverse/1dL_inplace	99501 ns	166932 ns	0.60
array/reverse/1d_inplace	75621 ns	140412 ns	0.54
array/reverse/2d	51390 ns	50811 ns	1.01
array/reverse/2dL	101191 ns	101331 ns	1.00
array/reverse/2dL_inplace	123482 ns	178202 ns	0.69
array/reverse/2d_inplace	68571 ns	61290 ns	1.12
array/sorting/1d	381835 ns	343314 ns	1.11
integration/byval/reference	39150 ns	39221 ns	1.00
integration/byval/slices=1	40461 ns	40110 ns	1.01
integration/byval/slices=2	160393 ns	137081 ns	1.17
integration/byval/slices=3	239514 ns	247933 ns	0.97
integration/volumerhs	5031862 ns	5039162 ns	1.00
kernel/indexing	118662 ns	101861 ns	1.16
kernel/indexing_checked	87422 ns	63861 ns	1.37
kernel/launch	1290 ns	1390 ns	0.93
kernel/rand	108201 ns	195682 ns	0.55
latency/import	1502528312 ns	1593049884 ns	0.94
latency/precompile	11784100596 ns	36362655526 ns	0.32
latency/ttfp	10974586647 ns	2158618452 ns	5.08

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

[luraess]

cscs-ci run

[luraess]

cscs-ci run

[luraess]

EDIT: Improved

• per-execution rocfft_execution_info lifetime with try/finally
• pointer buffers allocated before the thread-sensitive execution-info setup
• async FFT regression test repeated with an explicit yield() inside the spawned task