[maleadt]

Re-lands the part of #1148 that never reached main: #1148 was merged into its stacked base branch after that branch had already been merged.

#1117 already removed the NaN-ignoring ocml min/max overrides, so with Julia 1.13 and the external back-end, device min/max on floats match the host. But Julia 1.12 lowers them to llvm.minimum/llvm.maximum, which its in-process LLVM 18 can't select on most GPUs ("Cannot select: fminimum"). This adds an explicit comparison-based implementation for LLVM < 19, and #1148's tests: more special values and Atomix's atomic min/max on the device, and codegen checks for gfx90a/gfx942/gfx1200 with both back-ends.

Tested locally on gfx1036 (core/codegen and device/math): 194/194 on Julia 1.10, 206/206 on 1.12, 212/212 on 1.13. Without the fix, core/codegen crashes on 1.12.

Fixes #1142

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 00fdba7	Previous: 0b3cee4	Ratio
amdgpu/synchronization/context/device	1820 ns	1790 ns	1.02
amdgpu/synchronization/stream/blocking	240 ns	232.5 ns	1.03
amdgpu/synchronization/stream/nonblocking	342.5 ns	345 ns	0.99
applications/bitonic_sort	1090758 ns	1120496 ns	0.97
applications/convolution	73553.5 ns	74001 ns	0.99
applications/floyd_warshall	8979620.75 ns	9041692.25 ns	0.99
applications/histogram	798016.25 ns	774031.25 ns	1.03
applications/prefix_sum	189477.5 ns	190332.75 ns	1.00
array/accumulate/Float32/1d	91046.5 ns	90656.25 ns	1.00
array/accumulate/Float32/dims=1	70761 ns	70808.5 ns	1.00
array/accumulate/Float32/dims=1L	102464 ns	105546.5 ns	0.97
array/accumulate/Float32/dims=2	105054.25 ns	105271.5 ns	1.00
array/accumulate/Float32/dims=2L	118901.75 ns	127276.75 ns	0.93
array/accumulate/Int64/1d	113794.25 ns	115419.25 ns	0.99
array/accumulate/Int64/dims=1	83811.25 ns	83676 ns	1.00
array/accumulate/Int64/dims=1L	110144 ns	110961.5 ns	0.99
array/accumulate/Int64/dims=2	108661.5 ns	108946.5 ns	1.00
array/accumulate/Int64/dims=2L	174747.5 ns	186815.25 ns	0.94
array/broadcast	57373.25 ns	33388 ns	1.72
array/construct	2510 ns	2442.5 ns	1.03
array/copy	40068.25 ns	39128 ns	1.02
array/copyto!/cpu_to_gpu	85816.25 ns	86096.25 ns	1.00
array/copyto!/gpu_to_cpu	85853.75 ns	86191.25 ns	1.00
array/copyto!/gpu_to_gpu	57255.75 ns	37893 ns	1.51
array/iteration/findall/bool	94451.5 ns	93611.25 ns	1.01
array/iteration/findall/int	104824 ns	103016.75 ns	1.02
array/iteration/findfirst/bool	56570.75 ns	55878.5 ns	1.01
array/iteration/findfirst/int	56070.75 ns	56053.25 ns	1.00
array/iteration/findmin/1d	97528.75 ns	96021.25 ns	1.02
array/iteration/findmin/2d	81188.75 ns	82588.75 ns	0.98
array/iteration/logical	91436.25 ns	91116.25 ns	1.00
array/iteration/scalar	306449.25 ns	296909.25 ns	1.03
array/permutedims/2d	61708.5 ns	49090.75 ns	1.26
array/permutedims/3d	55703.25 ns	58643.5 ns	0.95
array/permutedims/4d	51088.25 ns	62151 ns	0.82
array/random/rand/Float32	46713 ns	45763 ns	1.02
array/random/rand/Int64	39878 ns	45873 ns	0.87
array/random/rand!/Float32	49515.75 ns	65503.5 ns	0.76
array/random/rand!/Int64	48310.75 ns	48105.75 ns	1.00
array/random/randn/Float32	67421 ns	64408.25 ns	1.05
array/random/randn!/Float32	82873.75 ns	58868.5 ns	1.41
array/reductions/mapreduce/Float32/1d	70493.5 ns	69296 ns	1.02
array/reductions/mapreduce/Float32/dims=1	45783 ns	53448.25 ns	0.86
array/reductions/mapreduce/Float32/dims=1L	50380.75 ns	59493.25 ns	0.85
array/reductions/mapreduce/Float32/dims=2	51510.75 ns	60155.75 ns	0.86
array/reductions/mapreduce/Float32/dims=2L	60638.5 ns	71963.5 ns	0.84
array/reductions/mapreduce/Int64/1d	62001 ns	71308.5 ns	0.87
array/reductions/mapreduce/Int64/dims=1	43438 ns	43990.5 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	56578.25 ns	64335.75 ns	0.88
array/reductions/mapreduce/Int64/dims=2	49545.75 ns	57796 ns	0.86
array/reductions/mapreduce/Int64/dims=2L	62848.5 ns	75973.75 ns	0.83
array/reductions/reduce/Float32/1d	59948.5 ns	69478.5 ns	0.86
array/reductions/reduce/Float32/dims=1	46345.75 ns	54248.5 ns	0.85
array/reductions/reduce/Float32/dims=1L	50770.5 ns	58768.25 ns	0.86
array/reductions/reduce/Float32/dims=2	51280.75 ns	59593.5 ns	0.86
array/reductions/reduce/Float32/dims=2L	60045.75 ns	71803.5 ns	0.84
array/reductions/reduce/Int64/1d	61421 ns	68288.5 ns	0.90
array/reductions/reduce/Int64/dims=1	49238.25 ns	50843.25 ns	0.97
array/reductions/reduce/Int64/dims=1L	64275.75 ns	70143.5 ns	0.92
array/reductions/reduce/Int64/dims=2	49893.25 ns	58103.5 ns	0.86
array/reductions/reduce/Int64/dims=2L	76011.25 ns	62133.5 ns	1.22
array/reverse/1d	35783 ns	39465.5 ns	0.91
array/reverse/1dL	47955.75 ns	60373.5 ns	0.79
array/reverse/1dL_inplace	57078.25 ns	78531 ns	0.73
array/reverse/1d_inplace	39325.5 ns	62158.5 ns	0.63
array/reverse/2d	48423 ns	41040.75 ns	1.18
array/reverse/2dL	73533.75 ns	78448.75 ns	0.94
array/reverse/2dL_inplace	93283.75 ns	74931.25 ns	1.24
array/reverse/2d_inplace	37423 ns	58875.75 ns	0.64
array/sorting/1d	336194.75 ns	336670 ns	1.00
gemm/tiled	1922802.75 ns	1893882.25 ns	1.02
gemm/tiled_unbounded	1939687.75 ns	1905582.5 ns	1.02
integration/byval/reference	41091 ns	37511 ns	1.10
integration/byval/slices=1	40481 ns	39661 ns	1.02
integration/byval/slices=2	149672 ns	153042 ns	0.98
integration/byval/slices=3	238223 ns	240084 ns	0.99
integration/volumerhs	4866169 ns	4960111 ns	0.98
kernel/indexing	39010.5 ns	32263 ns	1.21
kernel/indexing_checked	44688 ns	59088.25 ns	0.76
kernel/launch	1242.5 ns	1232.5 ns	1.01
kernel/rand	64603.25 ns	71428.5 ns	0.90
latency/import	1878175299 ns	1872656502 ns	1.00
latency/precompile	34296924846 ns	34344569766 ns	1.00
latency/ttfp	2302692363 ns	2288835992 ns	1.01
stencil/diffusion3d	1583310 ns	1594345.5 ns	0.99
stencil/diffusion3d_checked	1611955.5 ns	1623818.25 ns	0.99

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