[maleadt]

Stacked on #1147 (merge that first).

On the device, Base.min/Base.max for Float16/32/64 were overridden with ocml's min/max, which ignore NaN operands and leave the result for −0/+0 unspecified. Julia's min/max propagate NaNs and treat −0 as smaller than +0, so e.g. max(NaN32, 1f0) returned 1f0 in a kernel. This also affected @atomic max/min on floats, which on Julia ≤ 1.13 is a compare-exchange loop around Base.max/min, while Julia 1.14 lowers it to atomicrmw fmaximum, so results depended on the Julia version.

With Julia's LLVM 19 or later (Julia 1.13+), the overrides now emit llvm.minimum/llvm.maximum, like Base; LLVM expands these for every AMDGPU target and has native instructions on gfx12. Older LLVM versions can't lower those intrinsics in general (hence the existing Julia 1.12 FastMath workaround), so there the overrides use a comparison-based implementation. The choice is made from Julia's LLVM version, so this also applies with the external back-end. The fallback costs a few extra instructions: on gfx1036 with Julia 1.12, a Float32 max becomes four compares and a select, versus v_max_f32 + v_cmp_o_f32 + select with llvm.maximum on Julia 1.13.

Tests compare device results against the host for NaN in either or both operands, ±0 in both orders, infinities and subnormals (Float16/32/64), check atomic min/max with NaNs, and check offline codegen for gfx90a/gfx942/gfx1200 with both back-ends. Tested locally on gfx1036 with Julia 1.12 and 1.13 (external back-end), and before rebasing onto #1147 also with Julia 1.14-DEV and the in-process back-end; not on Julia 1.10/1.11.

Fixes #1142

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 1b16a35	Previous: 6f37fdf	Ratio
amdgpu/synchronization/context/device	565 ns	547.5 ns	1.03
amdgpu/synchronization/stream/blocking	232.5 ns	222.5 ns	1.04
amdgpu/synchronization/stream/nonblocking	322.5 ns	305 ns	1.06
applications/bitonic_sort	1116487.75 ns	1121018 ns	1.00
applications/convolution	91206 ns	91856.25 ns	0.99
applications/floyd_warshall	9094040.5 ns	9088943 ns	1.00
applications/histogram	803021.5 ns	795198.5 ns	1.01
applications/prefix_sum	188049.75 ns	188172.75 ns	1.00
array/accumulate/Float32/1d	80405.75 ns	82703.5 ns	0.97
array/accumulate/Float32/dims=1	263577.5 ns	259653.5 ns	1.02
array/accumulate/Float32/dims=1L	105038.5 ns	91916.25 ns	1.14
array/accumulate/Float32/dims=2	104743.5 ns	105999 ns	0.99
array/accumulate/Float32/dims=2L	3029516 ns	3029801.75 ns	1.00
array/accumulate/Int64/1d	86945.75 ns	93226.5 ns	0.93
array/accumulate/Int64/dims=1	244262.25 ns	242948.25 ns	1.01
array/accumulate/Int64/dims=1L	111588.5 ns	112674.25 ns	0.99
array/accumulate/Int64/dims=2	106546 ns	80158.75 ns	1.33
array/accumulate/Int64/dims=2L	3391547.75 ns	3392649.25 ns	1.00
array/broadcast	29655.25 ns	29063 ns	1.02
array/construct	2170 ns	2220 ns	0.98
array/copy	32037.75 ns	38870.5 ns	0.82
array/copyto!/cpu_to_gpu	88106 ns	88961.25 ns	0.99
array/copyto!/gpu_to_cpu	89626 ns	89613.75 ns	1.00
array/copyto!/gpu_to_gpu	32960.25 ns	34700.5 ns	0.95
array/iteration/findall/bool	109521.25 ns	119901.75 ns	0.91
array/iteration/findall/int	113091.25 ns	112736.5 ns	1.00
array/iteration/findfirst/bool	153931.75 ns	152529.75 ns	1.01
array/iteration/findfirst/int	187524.75 ns	153119.5 ns	1.22
array/iteration/findmin/1d	107588.75 ns	104196.25 ns	1.03
array/iteration/findmin/2d	91266 ns	93761.25 ns	0.97
array/iteration/logical	182804.5 ns	183702.5 ns	1.00
array/iteration/scalar	310778.25 ns	309796.75 ns	1.00
array/permutedims/2d	59368 ns	59476 ns	1.00
array/permutedims/3d	57163 ns	57221 ns	1.00
array/permutedims/4d	65668 ns	43595.5 ns	1.51
array/random/rand/Float32	45462.75 ns	45258 ns	1.00
array/random/rand/Int64	43375.5 ns	37798 ns	1.15
array/random/rand!/Float32	40678 ns	41948 ns	0.97
array/random/rand!/Int64	38190.5 ns	37215.5 ns	1.03
array/random/randn/Float32	57073.25 ns	72228.25 ns	0.79
array/random/randn!/Float32	54168.25 ns	54075.75 ns	1.00
array/reductions/mapreduce/Float32/1d	108956.25 ns	108671.5 ns	1.00
array/reductions/mapreduce/Float32/dims=1	88181 ns	88146.25 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	835141.25 ns	836306.5 ns	1.00
array/reductions/mapreduce/Float32/dims=2	93868.5 ns	94228.75 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	135981.5 ns	135486.75 ns	1.00
array/reductions/mapreduce/Int64/1d	108563.5 ns	108869 ns	1.00
array/reductions/mapreduce/Int64/dims=1	88541 ns	88541.25 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	839383.75 ns	841141.75 ns	1.00
array/reductions/mapreduce/Int64/dims=2	92748.5 ns	92663.75 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	136191.5 ns	136469.5 ns	1.00
array/reductions/reduce/Float32/1d	108891 ns	108931.5 ns	1.00
array/reductions/reduce/Float32/dims=1	88318.5 ns	88088.5 ns	1.00
array/reductions/reduce/Float32/dims=1L	836026.5 ns	835606.5 ns	1.00
array/reductions/reduce/Float32/dims=2	94333.25 ns	94116.25 ns	1.00
array/reductions/reduce/Float32/dims=2L	135816.5 ns	135959.5 ns	1.00
array/reductions/reduce/Int64/1d	108416.25 ns	107759 ns	1.01
array/reductions/reduce/Int64/dims=1	88691 ns	87166.25 ns	1.02
array/reductions/reduce/Int64/dims=1L	841434 ns	837371.5 ns	1.00
array/reductions/reduce/Int64/dims=2	92703.5 ns	92491.25 ns	1.00
array/reductions/reduce/Int64/dims=2L	136036.5 ns	135859.25 ns	1.00
array/reverse/1d	43543 ns	35203 ns	1.24
array/reverse/1dL	68715.75 ns	69038.25 ns	1.00
array/reverse/1dL_inplace	54328 ns	54128.25 ns	1.00
array/reverse/1d_inplace	35277.75 ns	37770.5 ns	0.93
array/reverse/2d	38850.5 ns	46583 ns	0.83
array/reverse/2dL	91240.75 ns	91263.75 ns	1.00
array/reverse/2dL_inplace	64428 ns	65028.5 ns	0.99
array/reverse/2d_inplace	37227.75 ns	37158 ns	1.00
array/sorting/1d	329413 ns	330742 ns	1.00
gemm/tiled	2866764.5 ns	1949946.75 ns	1.47
gemm/tiled_unbounded	2904937.5 ns	2914637.75 ns	1.00
integration/byval/reference	39430 ns	39901 ns	0.99
integration/byval/slices=1	40751 ns	39770 ns	1.02
integration/byval/slices=2	158332 ns	157742 ns	1.00
integration/byval/slices=3	239693 ns	223213 ns	1.07
integration/volumerhs	4876286 ns	4883697 ns	1.00
kernel/indexing	30295.25 ns	28847.75 ns	1.05
kernel/indexing_checked	36608 ns	32488 ns	1.13
kernel/launch	1225 ns	1220 ns	1.00
kernel/rand	48615.5 ns	51218.25 ns	0.95
latency/import	1628483156 ns	1628061194 ns	1.00
latency/precompile	29033979810 ns	29168346331 ns	1.00
latency/ttfp	2177501902 ns	2181953516 ns	1.00
stencil/diffusion3d	1583072.75 ns	1583662 ns	1.00
stencil/diffusion3d_checked	1627073.25 ns	1637802.75 ns	0.99

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