[maleadt]

Ports AMDGPU.jl to GPUArrays 12, which implements Base's reductions, sorting, scans, findall, logical indexing and reverse once for every GPU array, on AcceleratedKernels 0.5 (JuliaGPU/GPUArrays.jl#790), and removes the GPUArrays.mapreducedim! hook.

• Deletes AMDGPU's own reduction kernel, sort!/sortperm!, scans, findall, logical indexing and reverse (src/kernels/). The vector sort!/sortperm! were ambiguous with GPUArrays' methods, and the scans used AcceleratedKernels 0.4's API.
• Drops the direct AcceleratedKernels dependency; its AMDGPU extension still loads through GPUArrays. (AMDGPU's current compat, 0.3.1, 0.4, couldn't load AK 0.5 anyway.)
• GPUArrays.default_rng is gone in GPUArrays 12; the RNG lives in gpuarrays_rng().

Intended as a minor release.

Tested on gfx1036 (Julia 1.13): core/rocarray_base, hip_rocarray, the GPUArrays testsuite and the KernelAbstractions tests, 17,968 pass. The 3 cholesky(::Diagonal) errors are fixed in Adapt 4.7.3.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 70f2603	Previous: 1554014	Ratio
amdgpu/synchronization/context/device	1777.5 ns	1760 ns	1.01
amdgpu/synchronization/stream/blocking	225 ns	230 ns	0.98
amdgpu/synchronization/stream/nonblocking	350 ns	337.5 ns	1.04
applications/bitonic_sort	1073345.25 ns	1081902.5 ns	0.99
applications/convolution	55503.25 ns	91548.75 ns	0.61
applications/floyd_warshall	9030625 ns	9056864.25 ns	1.00
applications/histogram	796388.75 ns	762330.5 ns	1.04
applications/prefix_sum	189467.75 ns	188532.75 ns	1.00
array/accumulate/Float32/1d	99116.25 ns	82493.5 ns	1.20
array/accumulate/Float32/dims=1	269243.75 ns	260561.25 ns	1.03
array/accumulate/Float32/dims=1L	107301.75 ns	102981.5 ns	1.04
array/accumulate/Float32/dims=2	105019 ns	106454 ns	0.99
array/accumulate/Float32/dims=2L	2459177.5 ns	2936696.25 ns	0.84
array/accumulate/Int64/1d	96868.75 ns	87393.75 ns	1.11
array/accumulate/Int64/dims=1	254568.5 ns	244228.5 ns	1.04
array/accumulate/Int64/dims=1L	133529.5 ns	112424 ns	1.19
array/accumulate/Int64/dims=2	87161.25 ns	107009 ns	0.81
array/accumulate/Int64/dims=2L	2762979.5 ns	3363919.75 ns	0.82
array/broadcast	63143.5 ns	29065.5 ns	2.17
array/construct	2305 ns	2027.5 ns	1.14
array/copy	40495.5 ns	38108 ns	1.06
array/copyto!/cpu_to_gpu	86103.75 ns	89163.75 ns	0.97
array/copyto!/gpu_to_cpu	85953.75 ns	89608.75 ns	0.96
array/copyto!/gpu_to_gpu	37040.5 ns	33963 ns	1.09
array/iteration/findall/bool	94348.75 ns	121821.75 ns	0.77
array/iteration/findall/int	103793.75 ns	112161.5 ns	0.93
array/iteration/findfirst/bool	128786.75 ns	153429.5 ns	0.84
array/iteration/findfirst/int	132737 ns	166444.75 ns	0.80
array/iteration/findmin/1d	107331.5 ns	115776.5 ns	0.93
array/iteration/findmin/2d	101094 ns	83061.25 ns	1.22
array/iteration/logical	90586.25 ns	182632.5 ns	0.50
array/iteration/scalar	306969.5 ns	296154.25 ns	1.04
array/permutedims/2d	60321 ns	58833.25 ns	1.03
array/permutedims/3d	58843.25 ns	56508.25 ns	1.04
array/permutedims/4d	65126 ns	65433.25 ns	1.00
array/random/rand/Float32	40975.5 ns	37123 ns	1.10
array/random/rand/Int64	39088.25 ns	44115.5 ns	0.89
array/random/rand!/Float32	37783 ns	41313 ns	0.91
array/random/rand!/Int64	58300.75 ns	37768 ns	1.54
array/random/randn/Float32	68746 ns	71366 ns	0.96
array/random/randn!/Float32	76478.5 ns	53588.25 ns	1.43
array/reductions/mapreduce/Float32/1d	77531.25 ns	110251.75 ns	0.70
array/reductions/mapreduce/Float32/dims=1	58948.5 ns	88178.75 ns	0.67
array/reductions/mapreduce/Float32/dims=1L	66343.5 ns	817171.25 ns	0.0811867769454689
array/reductions/mapreduce/Float32/dims=2	69021 ns	74603.5 ns	0.93
array/reductions/mapreduce/Float32/dims=2L	69668.5 ns	136832 ns	0.51
array/reductions/mapreduce/Int64/1d	80393.75 ns	110986.5 ns	0.72
array/reductions/mapreduce/Int64/dims=1	59181 ns	88126.25 ns	0.67
array/reductions/mapreduce/Int64/dims=1L	78348.5 ns	820481.5 ns	0.09549087944091365
array/reductions/mapreduce/Int64/dims=2	67785.75 ns	92678.75 ns	0.73
array/reductions/mapreduce/Int64/dims=2L	85068.75 ns	136504.25 ns	0.62
array/reductions/reduce/Float32/1d	79106.25 ns	112186.5 ns	0.71
array/reductions/reduce/Float32/dims=1	59108.25 ns	70328.5 ns	0.84
array/reductions/reduce/Float32/dims=1L	59401 ns	814998.75 ns	0.07288477436315087
array/reductions/reduce/Float32/dims=2	69165.75 ns	95113.75 ns	0.73
array/reductions/reduce/Float32/dims=2L	81406 ns	137447 ns	0.59
array/reductions/reduce/Int64/1d	80153.75 ns	110429 ns	0.73
array/reductions/reduce/Int64/dims=1	59428.25 ns	88551.25 ns	0.67
array/reductions/reduce/Int64/dims=1L	79213.5 ns	821399 ns	0.09643729782967839
array/reductions/reduce/Int64/dims=2	67653.5 ns	92529 ns	0.73
array/reductions/reduce/Int64/dims=2L	85053.75 ns	136359.5 ns	0.62
array/reverse/1d	40718 ns	43333.25 ns	0.94
array/reverse/1dL	68648.5 ns	68421 ns	1.00
array/reverse/1dL_inplace	77951.25 ns	53655.75 ns	1.45
array/reverse/1d_inplace	37983 ns	37160.5 ns	1.02
array/reverse/2d	48020.75 ns	46538.25 ns	1.03
array/reverse/2dL	77591 ns	90628.75 ns	0.86
array/reverse/2dL_inplace	82726.25 ns	64456 ns	1.28
array/reverse/2d_inplace	54603.25 ns	36693 ns	1.49
array/sorting/1d	324219.5 ns	322582 ns	1.01
gemm/tiled	1917469.5 ns	2811927 ns	0.68
gemm/tiled_unbounded	1912107 ns	2830087 ns	0.68
integration/byval/reference	39241 ns	38111 ns	1.03
integration/byval/slices=1	41600 ns	40750 ns	1.02
integration/byval/slices=2	128632 ns	140552 ns	0.92
integration/byval/slices=3	228744 ns	213973 ns	1.07
integration/volumerhs	4851009 ns	4851138 ns	1.00
kernel/indexing	57620.75 ns	30203 ns	1.91
kernel/indexing_checked	39893 ns	35820.5 ns	1.11
kernel/launch	1225 ns	1272.5 ns	0.96
kernel/rand	72798.5 ns	50990.75 ns	1.43
latency/import	1801892031 ns	1628767420 ns	1.11
latency/precompile	32543029095 ns	29419546616 ns	1.11
latency/ttfp	2366949199 ns	2181231733 ns	1.09
stencil/diffusion3d	1567239.5 ns	1557121.75 ns	1.01
stencil/diffusion3d_checked	1620250.5 ns	1593372.25 ns	1.02

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