Fix KeyError in hipcompile when fetching early hostcall metadata

[luraess]

The global Dict side channel between link_libraries! and hipcompile could throw a KeyError if the key was missing (hash mismatch or concurrent compilation). Seen in Trixi.jl CI after #907 .

Replace it with task-local storage: link_libraries! writes into task_local_storage() and hipcompile reads from it on the same task, which is guaranteed since link_libraries! is called synchronously inside GPUCompiler.compile with fallback to late IR-based detection if the early pass was skipped.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 2786537	Previous: 30db2cc	Ratio
amdgpu/synchronization/context/device	600 ns	600 ns	1
amdgpu/synchronization/stream/blocking	240 ns	250 ns	0.96
amdgpu/synchronization/stream/nonblocking	340 ns	340 ns	1
array/accumulate/Float32/1d	85291 ns	87912 ns	0.97
array/accumulate/Float32/dims=1	281074 ns	377806 ns	0.74
array/accumulate/Float32/dims=1L	136052 ns	132962 ns	1.02
array/accumulate/Float32/dims=2	132082 ns	130232 ns	1.01
array/accumulate/Float32/dims=2L	2831838 ns	2828730 ns	1.00
array/accumulate/Int64/1d	93182 ns	94792 ns	0.98
array/accumulate/Int64/dims=1	282853 ns	296295 ns	0.95
array/accumulate/Int64/dims=1L	167762 ns	168273 ns	1.00
array/accumulate/Int64/dims=2	126802 ns	124332 ns	1.02
array/accumulate/Int64/dims=2L	3011530 ns	3012253 ns	1.00
array/broadcast	134902 ns	91621 ns	1.47
array/construct	1720 ns	1740 ns	0.99
array/copy	37390 ns	38370 ns	0.97
array/copyto!/cpu_to_gpu	115101 ns	114781 ns	1.00
array/copyto!/gpu_to_cpu	114992 ns	115392 ns	1.00
array/copyto!/gpu_to_gpu	61280 ns	59901 ns	1.02
array/iteration/findall/bool	178903 ns	183663 ns	0.97
array/iteration/findall/int	195812 ns	191663 ns	1.02
array/iteration/findfirst/bool	119832 ns	116691 ns	1.03
array/iteration/findfirst/int	115711 ns	116192 ns	1.00
array/iteration/findmin/1d	168992 ns	170762 ns	0.99
array/iteration/findmin/2d	155072 ns	156362 ns	0.99
array/iteration/logical	354645 ns	358726 ns	0.99
array/iteration/scalar	289054 ns	297164 ns	0.97
array/permutedims/2d	73441 ns	74201 ns	0.99
array/permutedims/3d	73481 ns	74851 ns	0.98
array/permutedims/4d	75531 ns	77271 ns	0.98
array/random/rand/Float32	51451 ns	51681 ns	1.00
array/random/rand/Int64	57051 ns	57821 ns	0.99
array/random/rand!/Float32	145032 ns	74771 ns	1.94
array/random/rand!/Int64	131881 ns	93162 ns	1.42
array/random/randn/Float32	88551 ns	92041 ns	0.96
array/random/randn!/Float32	115981 ns	84951 ns	1.37
array/reductions/mapreduce/Float32/1d	133162 ns	134042 ns	0.99
array/reductions/mapreduce/Float32/dims=1	94921 ns	95651 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	775700 ns	778621 ns	1.00
array/reductions/mapreduce/Float32/dims=2	97151 ns	97701 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	300664 ns	299124 ns	1.01
array/reductions/mapreduce/Int64/1d	133461 ns	134972 ns	0.99
array/reductions/mapreduce/Int64/dims=1	95401 ns	95752 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	784751 ns	782442 ns	1.00
array/reductions/mapreduce/Int64/dims=2	96312 ns	96952 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	305104 ns	300554 ns	1.02
array/reductions/reduce/Float32/1d	133132 ns	134001 ns	0.99
array/reductions/reduce/Float32/dims=1	94591 ns	95701 ns	0.99
array/reductions/reduce/Float32/dims=1L	776690 ns	775212 ns	1.00
array/reductions/reduce/Float32/dims=2	96461 ns	97311 ns	0.99
array/reductions/reduce/Float32/dims=2L	296444 ns	303524 ns	0.98
array/reductions/reduce/Int64/1d	133292 ns	134762 ns	0.99
array/reductions/reduce/Int64/dims=1	94581 ns	95241 ns	0.99
array/reductions/reduce/Int64/dims=1L	778970 ns	782341 ns	1.00
array/reductions/reduce/Int64/dims=2	95802 ns	96752 ns	0.99
array/reductions/reduce/Int64/dims=2L	295445 ns	300534 ns	0.98
array/reverse/1d	43991 ns	44471 ns	0.99
array/reverse/1dL	74741 ns	75792 ns	0.99
array/reverse/1dL_inplace	108561 ns	139412 ns	0.78
array/reverse/1d_inplace	133172 ns	74861 ns	1.78
array/reverse/2d	51180 ns	52100 ns	0.98
array/reverse/2dL	101101 ns	102302 ns	0.99
array/reverse/2dL_inplace	129242 ns	179532 ns	0.72
array/reverse/2d_inplace	135592 ns	115342 ns	1.18
array/sorting/1d	344244 ns	343644 ns	1.00
integration/byval/reference	38881 ns	39161 ns	0.99
integration/byval/slices=1	39691 ns	40220 ns	0.99
integration/byval/slices=2	148002 ns	143882 ns	1.03
integration/byval/slices=3	237913 ns	237113 ns	1.00
integration/volumerhs	5027049 ns	5034652 ns	1.00
kernel/indexing	44150 ns	107302 ns	0.41
kernel/indexing_checked	131332 ns	46721 ns	2.81
kernel/launch	1290 ns	1290 ns	1
kernel/rand	169732 ns	123912 ns	1.37
latency/import	1476602726 ns	1470667415 ns	1.00
latency/precompile	11915889321 ns	11877284792 ns	1.00
latency/ttfp	10278643291 ns	10304855031 ns	1.00

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