Support Julia 1.13

[eschnett]

Closes #3019.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 4c8e4e2	Previous: e2eab84	Ratio
latency/precompile	44846190635.5 ns	55351775865 ns	0.81
latency/ttfp	14453150957 ns	7675959782 ns	1.88
latency/import	4520087896 ns	4025008043 ns	1.12
integration/volumerhs	9443988.5 ns	9625265.5 ns	0.98
integration/byval/slices=1	145635.5 ns	147037 ns	0.99
integration/byval/slices=3	422956.5 ns	426089 ns	0.99
integration/byval/reference	143817 ns	144983 ns	0.99
integration/byval/slices=2	284679 ns	286492 ns	0.99
integration/cudadevrt	102555 ns	103531 ns	0.99
kernel/indexing	13575 ns	14161 ns	0.96
kernel/indexing_checked	14160 ns	15039 ns	0.94
kernel/occupancy	683.5454545454545 ns	792.6237623762377 ns	0.86
kernel/launch	2105.1 ns	2304.777777777778 ns	0.91
kernel/rand	14507 ns	18605.5 ns	0.78
array/reverse/1d	19160 ns	19692 ns	0.97
array/reverse/2dL_inplace	66497 ns	66886 ns	0.99
array/reverse/1dL	69339 ns	69816 ns	0.99
array/reverse/2d	21029.5 ns	21785 ns	0.97
array/reverse/1d_inplace	10596.666666666666 ns	9783 ns	1.08
array/reverse/2d_inplace	11915 ns	13411 ns	0.89
array/reverse/2dL	73024.5 ns	73680 ns	0.99
array/reverse/1dL_inplace	66420 ns	66877 ns	0.99
array/copy	18483 ns	20494 ns	0.90
array/iteration/findall/int	146058 ns	159432 ns	0.92
array/iteration/findall/bool	130524 ns	141132 ns	0.92
array/iteration/findfirst/int	84706 ns	161404 ns	0.52
array/iteration/findfirst/bool	82020 ns	162399 ns	0.51
array/iteration/scalar	65977.5 ns	73008 ns	0.90
array/iteration/logical	198658 ns	220542 ns	0.90
array/iteration/findmin/1d	82559 ns	94388 ns	0.87
array/iteration/findmin/2d	116988 ns	121456 ns	0.96
array/reductions/reduce/Int64/1d	39212 ns	43817 ns	0.89
array/reductions/reduce/Int64/dims=1	42352 ns	44722 ns	0.95
array/reductions/reduce/Int64/dims=2	59082 ns	61524.5 ns	0.96
array/reductions/reduce/Int64/dims=1L	87201 ns	88932 ns	0.98
array/reductions/reduce/Int64/dims=2L	84775.5 ns	88232 ns	0.96
array/reductions/reduce/Float32/1d	34234.5 ns	37401.5 ns	0.92
array/reductions/reduce/Float32/dims=1	40281.5 ns	51945.5 ns	0.78
array/reductions/reduce/Float32/dims=2	56532.5 ns	59868 ns	0.94
array/reductions/reduce/Float32/dims=1L	51686 ns	52569 ns	0.98
array/reductions/reduce/Float32/dims=2L	70130.5 ns	72250 ns	0.97
array/reductions/mapreduce/Int64/1d	39390 ns	43757 ns	0.90
array/reductions/mapreduce/Int64/dims=1	49646.5 ns	51077 ns	0.97
array/reductions/mapreduce/Int64/dims=2	59329 ns	61716 ns	0.96
array/reductions/mapreduce/Int64/dims=1L	87259 ns	89026 ns	0.98
array/reductions/mapreduce/Int64/dims=2L	84900.5 ns	88266 ns	0.96
array/reductions/mapreduce/Float32/1d	34051 ns	37130 ns	0.92
array/reductions/mapreduce/Float32/dims=1	45578 ns	42011 ns	1.08
array/reductions/mapreduce/Float32/dims=2	56455 ns	60008 ns	0.94
array/reductions/mapreduce/Float32/dims=1L	51769 ns	52656 ns	0.98
array/reductions/mapreduce/Float32/dims=2L	69321 ns	72333.5 ns	0.96
array/broadcast	20561 ns	20154 ns	1.02
array/copyto!/gpu_to_gpu	10673.166666666668 ns	12828 ns	0.83
array/copyto!/cpu_to_gpu	218524 ns	216999 ns	1.01
array/copyto!/gpu_to_cpu	284623 ns	282452 ns	1.01
array/accumulate/Int64/1d	118384 ns	125219 ns	0.95
array/accumulate/Int64/dims=1	79371 ns	88138 ns	0.90
array/accumulate/Int64/dims=2	155465 ns	162165 ns	0.96
array/accumulate/Int64/dims=1L	1694474 ns	1714037 ns	0.99
array/accumulate/Int64/dims=2L	960497 ns	971078.5 ns	0.99
array/accumulate/Float32/1d	100226 ns	109882.5 ns	0.91
array/accumulate/Float32/dims=1	76031 ns	84424 ns	0.90
array/accumulate/Float32/dims=2	144406.5 ns	151744.5 ns	0.95
array/accumulate/Float32/dims=1L	1585870 ns	1622762 ns	0.98
array/accumulate/Float32/dims=2L	656780.5 ns	702590.5 ns	0.93
array/construct	1265.3 ns	1267.85 ns	1.00
array/random/randn/Float32	36574 ns	48207 ns	0.76
array/random/randn!/Float32	30409 ns	25000 ns	1.22
array/random/rand!/Int64	34484 ns	27295 ns	1.26
array/random/rand!/Float32	8136.75 ns	8737.333333333334 ns	0.93
array/random/rand/Int64	36946 ns	30000 ns	1.23
array/random/rand/Float32	12492 ns	13155 ns	0.95
array/permutedims/4d	51599 ns	55023 ns	0.94
array/permutedims/2d	52539.5 ns	53878 ns	0.98
array/permutedims/3d	52918 ns	54959 ns	0.96
array/sorting/1d	2736673 ns	2758315 ns	0.99
array/sorting/by	3306442 ns	3344753.5 ns	0.99
array/sorting/2d	1068592 ns	1081270 ns	0.99
cuda/synchronization/stream/auto	974.5 ns	1017.0833333333334 ns	0.96
cuda/synchronization/stream/nonblocking	6809.700000000001 ns	7295.6 ns	0.93
cuda/synchronization/stream/blocking	816.5108695652174 ns	798.6601941747573 ns	1.02
cuda/synchronization/context/auto	1159.2 ns	1158.8 ns	1.00
cuda/synchronization/context/nonblocking	7289.8 ns	7658.9 ns	0.95
cuda/synchronization/context/blocking	894.8541666666666 ns	902.5217391304348 ns	0.99

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

[eschnett]

The self-tests fail because the linear algebra functions (e.g. matrix exponential) as implemented in LinearAlgebra use scalar iteration. See e.g. exp! in https://github.com/JuliaLang/LinearAlgebra.jl/blob/f55e4736fb6dce08fee8a7ac7f0aba1f2b54838e/src/dense.jl#L784.

How should this be handled? Rewrite exp!? Find a respective CUDA library function to call and add a new method to exp? Fall back to the Julia 1.12 implementation? How does this work in Julia 1.12?

[eschnett]

I think it's JuliaGPU/GPUArrays.jl#679.

[eschnett]

The buildkite error is

  ptxas /tmp/jl_PALmvKnqta.ptx, line 226; error   : Modifier '.NaN' requires .target sm_80 or higher
  ptxas /tmp/jl_PALmvKnqta.ptx, line 226; error   : Feature 'min.f16 or min.f16x2' requires .target sm_80 or higher

This seems unrelated to my changes, except that I am now running CI tests on Julia 1.12 and Julia 1.13...

[maleadt]

I guess #3025 needs to be active for all LLVM versions.

[eschnett]

Good news: CUDA.jl now works for Julia 1.12.
Bad news: There's an LLVM segfault for Julia 1.13.

�_bk;t=1770145810814�      From worker 5:	[271397] signal 11 (1): Segmentation fault
�_bk;t=1770145810814�      From worker 5:	in expression starting at /var/lib/buildkite-agent/builds/gpuci-9/julialang/cuda-dot-jl/test/base/texture.jl:41
�_bk;t=1770145810920�      From worker 5:	_ZN12_GLOBAL__N_124NVPTXReplaceImageHandles18findIndexForHandleERN4llvm14MachineOperandERNS1_15MachineFunctionERj.isra.0 at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810920�      From worker 5:	_ZN12_GLOBAL__N_124NVPTXReplaceImageHandles18findIndexForHandleERN4llvm14MachineOperandERNS1_15MachineFunctionERj.isra.0 at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810920�      From worker 5:	_ZN12_GLOBAL__N_124NVPTXReplaceImageHandles20runOnMachineFunctionERN4llvm15MachineFunctionE at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810921�      From worker 5:	_ZN4llvm19MachineFunctionPass13runOnFunctionERNS_8FunctionE.part.0 at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810921�      From worker 5:	_ZN4llvm13FPPassManager13runOnFunctionERNS_8FunctionE at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810921�      From worker 5:	_ZN4llvm13FPPassManager11runOnModuleERNS_6ModuleE at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810922�      From worker 5:	_ZN4llvm6legacy15PassManagerImpl3runERNS_6ModuleE at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)
�_bk;t=1770145810922�      From worker 5:	_ZL21LLVMTargetMachineEmitP23LLVMOpaqueTargetMachineP16LLVMOpaqueModuleRN4llvm17raw_pwrite_streamE19LLVMCodeGenFileTypePPc at /root/.cache/julia-buildkite-plugin/julia_installs/bin/linux/x64/1.13/julia-1.13-latest-linux-x86_64/bin/../lib/julia/libLLVM.so.20.1jl (unknown line)

[eschnett]

I think it's texture interpolation that is broken on 1.13. This line segfaults LLVM:

dst[i] = texture[u]

in test/base/texture.jl (function kernel_texture_warp_native).

[eschnett]

We will need to update KernelAbstractions.jl as well JuliaGPU/KernelAbstractions.jl#679.