Use Julia's min and max for floats

[vchuravy]

Base.min/Base.max on floats were overridden on the device with OCML's __ocml_min_*/__ocml_max_*. Like C's fmin/fmax, those return the other argument when one is NaN, so max(NaN32, 1f0) gave 1f0 on the GPU and NaN32 on the CPU.

This drops the overrides. Julia's own definitions compile for AMDGPU:

• on Julia 1.12+ they are llvm.minimum/llvm.maximum, which have exactly Julia's semantics;
• before 1.12 they are the branch-free arithmetic definitions.

Julia's semantics here are that NaN propagates and -0.0 < 0.0.

Tests:

• New test in test/device/math.jl: NaN, ±0.0 and ±Inf, for Float16, Float32 and Float64.
• On an RX 6800 XT (gfx1030) with Julia 1.12.7, it fails before this change and passes after. The other tests in test/device/math.jl, including "Fast min/max", still pass.
• I haven't tested Julia 1.10, 1.11 or 1.13 locally.

Found through KernelAbstractions' KernelInterface test suite (JuliaGPU/KernelAbstractions.jl#831), which tests sub-group reductions of max with a NaN against Julia's semantics.

🤖 Generated with Claude Code

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: a83b644	Previous: f0d0502	Ratio
amdgpu/synchronization/context/device	555 ns	565 ns	0.98
amdgpu/synchronization/stream/blocking	232.5 ns	235 ns	0.99
amdgpu/synchronization/stream/nonblocking	317.5 ns	315 ns	1.01
applications/bitonic_sort	1150281.5 ns	1116058.5 ns	1.03
applications/convolution	99799.25 ns	78898.75 ns	1.26
applications/floyd_warshall	9131606.5 ns	9110058.5 ns	1.00
applications/histogram	802578 ns	827149.25 ns	0.97
applications/prefix_sum	187920.5 ns	187240 ns	1.00
array/accumulate/Float32/1d	82061.75 ns	80941.25 ns	1.01
array/accumulate/Float32/dims=1	262242.25 ns	279954 ns	0.94
array/accumulate/Float32/dims=1L	110482 ns	105479 ns	1.05
array/accumulate/Float32/dims=2	111932.25 ns	106766.5 ns	1.05
array/accumulate/Float32/dims=2L	3025676 ns	3030052.75 ns	1.00
array/accumulate/Int64/1d	90739.25 ns	86061.25 ns	1.05
array/accumulate/Int64/dims=1	244897 ns	249656 ns	0.98
array/accumulate/Int64/dims=1L	119737.25 ns	94328.75 ns	1.27
array/accumulate/Int64/dims=2	96234.25 ns	103489 ns	0.93
array/accumulate/Int64/dims=2L	3387696.75 ns	3393328 ns	1.00
array/broadcast	29328 ns	68516 ns	0.43
array/construct	2075 ns	2180 ns	0.95
array/copy	38040.75 ns	38240.5 ns	0.99
array/copyto!/cpu_to_gpu	88816.5 ns	89121.25 ns	1.00
array/copyto!/gpu_to_cpu	89709 ns	88303.5 ns	1.02
array/copyto!/gpu_to_gpu	33770.75 ns	33785.5 ns	1.00
array/iteration/findall/bool	127222.25 ns	130914.5 ns	0.97
array/iteration/findall/int	118779.75 ns	140924.5 ns	0.84
array/iteration/findfirst/bool	151870 ns	154504.5 ns	0.98
array/iteration/findfirst/int	187723 ns	189135.25 ns	0.99
array/iteration/findmin/1d	114932 ns	103571.5 ns	1.11
array/iteration/findmin/2d	100394.25 ns	89636.25 ns	1.12
array/iteration/logical	191903.25 ns	225498.25 ns	0.85
array/iteration/scalar	298785 ns	296164.25 ns	1.01
array/permutedims/2d	60866.25 ns	69475.75 ns	0.88
array/permutedims/3d	47058.25 ns	58178.25 ns	0.81
array/permutedims/4d	62431 ns	73456 ns	0.85
array/random/rand/Float32	44595.75 ns	44738 ns	1.00
array/random/rand/Int64	43570.75 ns	54233.5 ns	0.80
array/random/rand!/Float32	38405.75 ns	40805.75 ns	0.94
array/random/rand!/Int64	43728.25 ns	49345.75 ns	0.89
array/random/randn/Float32	71331.25 ns	65160.75 ns	1.09
array/random/randn!/Float32	55183.5 ns	56443.25 ns	0.98
array/reductions/mapreduce/Float32/1d	115112 ns	81013.5 ns	1.42
array/reductions/mapreduce/Float32/dims=1	87624 ns	88001.25 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	836309.75 ns	839651.75 ns	1.00
array/reductions/mapreduce/Float32/dims=2	93459 ns	89701.25 ns	1.04
array/reductions/mapreduce/Float32/dims=2L	134957.25 ns	135854.5 ns	0.99
array/reductions/mapreduce/Int64/1d	114999.5 ns	80863.5 ns	1.42
array/reductions/mapreduce/Int64/dims=1	88031.5 ns	88458.75 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	848372 ns	843502 ns	1.01
array/reductions/mapreduce/Int64/dims=2	91494 ns	86963.75 ns	1.05
array/reductions/mapreduce/Int64/dims=2L	136214.75 ns	136949.5 ns	0.99
array/reductions/reduce/Float32/1d	113534.5 ns	80531.25 ns	1.41
array/reductions/reduce/Float32/dims=1	87209 ns	87283.75 ns	1.00
array/reductions/reduce/Float32/dims=1L	840324.25 ns	840039.25 ns	1.00
array/reductions/reduce/Float32/dims=2	93421.75 ns	85036.25 ns	1.10
array/reductions/reduce/Float32/dims=2L	135189.75 ns	135587 ns	1.00
array/reductions/reduce/Int64/1d	100361.5 ns	80693.75 ns	1.24
array/reductions/reduce/Int64/dims=1	87899 ns	88476.25 ns	0.99
array/reductions/reduce/Int64/dims=1L	843064.25 ns	841001.75 ns	1.00
array/reductions/reduce/Int64/dims=2	72571.25 ns	90366.25 ns	0.80
array/reductions/reduce/Int64/dims=2L	135714.75 ns	135834.5 ns	1.00
array/reverse/1d	40220.5 ns	41358 ns	0.97
array/reverse/1dL	52121 ns	70341 ns	0.74
array/reverse/1dL_inplace	53273.5 ns	54481 ns	0.98
array/reverse/1d_inplace	34648.25 ns	35505.5 ns	0.98
array/reverse/2d	45848.25 ns	39643 ns	1.16
array/reverse/2dL	86046.5 ns	90556.25 ns	0.95
array/reverse/2dL_inplace	64541 ns	65046 ns	0.99
array/reverse/2d_inplace	35475.5 ns	36200.5 ns	0.98
array/sorting/1d	331121 ns	326947 ns	1.01
gemm/tiled	1948859.75 ns	2884390.75 ns	0.68
gemm/tiled_unbounded	1954799.75 ns	2904108.5 ns	0.67
integration/byval/reference	38451 ns	39421 ns	0.98
integration/byval/slices=1	40511 ns	38151 ns	1.06
integration/byval/slices=2	158003 ns	158542 ns	1.00
integration/byval/slices=3	238284 ns	241053 ns	0.99
integration/volumerhs	4888099 ns	4873209 ns	1.00
kernel/indexing	29357.75 ns	29223 ns	1.00
kernel/indexing_checked	36025.5 ns	33385.5 ns	1.08
kernel/launch	1200 ns	1160 ns	1.03
kernel/rand	73641.25 ns	98406.25 ns	0.75
latency/import	1492585412 ns	1462428687 ns	1.02
latency/precompile	23029244636 ns	22794416389 ns	1.01
latency/ttfp	2224788030 ns	2230777820 ns	1.00
stencil/diffusion3d	1626614.25 ns	1620078 ns	1.00
stencil/diffusion3d_checked	1647839.5 ns	1660428.5 ns	0.99

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