Add Base.min override for Float16 and extend LLVM version guard to v20.

[maleadt]

LLVM 20 lowers Base.min(::Float16, ::Float16) to min.NaN.f16, a PTX instruction requiring sm_80+, causing failures on Turing (sm_75) GPUs. Add a Julia-level override matching the existing Base.max workaround, and extend the version guard from LLVM 18 to 20 since the upstream fix (llvm/llvm-project@6f318d47) only landed in LLVM 21.

As observed in #3020

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.33%. Comparing base (1810b7a) to head (de9be6a).
⚠️ Report is 4 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3038      +/-   ##
==========================================
- Coverage   89.49%   89.33%   -0.17%     
==========================================
  Files         148      148              
  Lines       13047    13047              
==========================================
- Hits        11676    11655      -21     
- Misses       1371     1392      +21     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: de9be6a	Previous: 1810b7a	Ratio
latency/precompile	44675226940.5 ns	44300180944.5 ns	1.01
latency/ttfp	13291171512 ns	13138137112 ns	1.01
latency/import	3784128603 ns	3757487166.5 ns	1.01
integration/volumerhs	9440873.5 ns	9441754.5 ns	1.00
integration/byval/slices=1	145616 ns	145846 ns	1.00
integration/byval/slices=3	422814 ns	423265 ns	1.00
integration/byval/reference	143792 ns	143916 ns	1.00
integration/byval/slices=2	284069 ns	284641 ns	1.00
integration/cudadevrt	102357 ns	102633 ns	1.00
kernel/indexing	13245 ns	13466 ns	0.98
kernel/indexing_checked	14083 ns	13982 ns	1.01
kernel/occupancy	635.202380952381 ns	699.625850340136 ns	0.91
kernel/launch	2025.5 ns	2067.8 ns	0.98
kernel/rand	14585 ns	16244 ns	0.90
array/reverse/1d	18615 ns	18605 ns	1.00
array/reverse/2dL_inplace	66177 ns	66133 ns	1.00
array/reverse/1dL	68804 ns	68870 ns	1.00
array/reverse/2d	21266 ns	20781 ns	1.02
array/reverse/1d_inplace	10491 ns	10493.666666666666 ns	1.00
array/reverse/2d_inplace	11367 ns	10765 ns	1.06
array/reverse/2dL	73210 ns	72777.5 ns	1.01
array/reverse/1dL_inplace	66188 ns	66166 ns	1.00
array/copy	18366 ns	18321 ns	1.00
array/iteration/findall/int	145622.5 ns	145251 ns	1.00
array/iteration/findall/bool	130340 ns	130303 ns	1.00
array/iteration/findfirst/int	85134 ns	83996 ns	1.01
array/iteration/findfirst/bool	82631 ns	81209 ns	1.02
array/iteration/scalar	67040 ns	64953 ns	1.03
array/iteration/logical	197058.5 ns	197334 ns	1.00
array/iteration/findmin/1d	83432 ns	85667.5 ns	0.97
array/iteration/findmin/2d	117087 ns	117130 ns	1.00
array/reductions/reduce/Int64/1d	38905 ns	38913 ns	1.00
array/reductions/reduce/Int64/dims=1	41600 ns	41855 ns	0.99
array/reductions/reduce/Int64/dims=2	58808 ns	59043 ns	1.00
array/reductions/reduce/Int64/dims=1L	87117 ns	87102 ns	1.00
array/reductions/reduce/Int64/dims=2L	84669 ns	84295 ns	1.00
array/reductions/reduce/Float32/1d	34237 ns	33785 ns	1.01
array/reductions/reduce/Float32/dims=1	43934 ns	48986 ns	0.90
array/reductions/reduce/Float32/dims=2	56239 ns	56655 ns	0.99
array/reductions/reduce/Float32/dims=1L	51394 ns	51438 ns	1.00
array/reductions/reduce/Float32/dims=2L	69575 ns	69460.5 ns	1.00
array/reductions/mapreduce/Int64/1d	39210.5 ns	38699 ns	1.01
array/reductions/mapreduce/Int64/dims=1	46057 ns	41686 ns	1.10
array/reductions/mapreduce/Int64/dims=2	58993 ns	58974 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	87229 ns	87184 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	84397 ns	84571 ns	1.00
array/reductions/mapreduce/Float32/1d	34022 ns	33512 ns	1.02
array/reductions/mapreduce/Float32/dims=1	39843 ns	47745 ns	0.83
array/reductions/mapreduce/Float32/dims=2	55903 ns	56241 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	51260 ns	51435 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	69261 ns	69604 ns	1.00
array/broadcast	20628 ns	20361 ns	1.01
array/copyto!/gpu_to_gpu	10673.333333333334 ns	10601.666666666666 ns	1.01
array/copyto!/cpu_to_gpu	213909 ns	214964 ns	1.00
array/copyto!/gpu_to_cpu	283527 ns	282717 ns	1.00
array/accumulate/Int64/1d	118150.5 ns	118054 ns	1.00
array/accumulate/Int64/dims=1	79533 ns	78929 ns	1.01
array/accumulate/Int64/dims=2	155242 ns	155861 ns	1.00
array/accumulate/Int64/dims=1L	1697447 ns	1705368 ns	1.00
array/accumulate/Int64/dims=2L	960552 ns	960330.5 ns	1.00
array/accumulate/Float32/1d	100637.5 ns	100426 ns	1.00
array/accumulate/Float32/dims=1	76099 ns	75943 ns	1.00
array/accumulate/Float32/dims=2	144215 ns	143974 ns	1.00
array/accumulate/Float32/dims=1L	1584181 ns	1584300 ns	1.00
array/accumulate/Float32/dims=2L	656485 ns	658063 ns	1.00
array/construct	1291 ns	1252.6 ns	1.03
array/random/randn/Float32	36310 ns	35435 ns	1.02
array/random/randn!/Float32	30120 ns	29972 ns	1.00
array/random/rand!/Int64	34550 ns	28260 ns	1.22
array/random/rand!/Float32	8320.166666666668 ns	8310 ns	1.00
array/random/rand/Int64	36976 ns	29927 ns	1.24
array/random/rand/Float32	12342 ns	12324 ns	1.00
array/permutedims/4d	50805 ns	51686 ns	0.98
array/permutedims/2d	52400 ns	52279 ns	1.00
array/permutedims/3d	52639 ns	52911 ns	0.99
array/sorting/1d	2734832 ns	2735042.5 ns	1.00
array/sorting/by	3304279 ns	3304486.5 ns	1.00
array/sorting/2d	1067131 ns	1066581 ns	1.00
cuda/synchronization/stream/auto	1064.090909090909 ns	993.5882352941177 ns	1.07
cuda/synchronization/stream/nonblocking	7534.299999999999 ns	7392.700000000001 ns	1.02
cuda/synchronization/stream/blocking	821.4470588235295 ns	811.8282828282828 ns	1.01
cuda/synchronization/context/auto	1150.3 ns	1160.9 ns	0.99
cuda/synchronization/context/nonblocking	7125.9 ns	7875.6 ns	0.90
cuda/synchronization/context/blocking	894.469387755102 ns	899.7058823529412 ns	0.99

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