[maleadt]

Follow-up to #3440. Two more chains that made Julia discard large parts of CUDACore's precompiled code when packages were loaded before CUDA.jl:

• invoke_frozen(GPUCompiler.compile, :asm, job) isn't inferred, so inspecting the IR it returns (to determine whether the device runtime is needed) was compiled with edges to generic methods like setdiff of an abstract collection. This asserts the module's type.
• ntuple(f, 53) returns a tuple of unknown length (ntuple only unrolls up to 10 elements), so the reserved bytes of CUmemPoolProps were converted to NTuple{53, UInt8} with generic ntuple/comparison methods, which ForwardDiff and others invalidate. Memory pools are created on the first allocation, so with ~1500 method instances this was the biggest remaining chain in CUDACore. The same applies to the texture descriptors. This passes the counts as Vals.

Tested with the core codegen, execution, exceptions, KernelAbstractions, broadcast, array, pool, texture and driver tests. Together with JuliaLLVM/LLVM.jl#703 and the other follow-ups, the first @cuda kernel (plus a broadcast and sum) with DataFrames loaded first goes from 6.42 s on main to 2.76 s, and with ChainRulesCore from 4.02 s to 2.22 s (2.12 s with nothing loaded first), on Julia 1.13 and an RTX 5080.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.58%. Comparing base (6747004) to head (d05e89b).

Additional details and impacted files

@@                  Coverage Diff                  @@
##           tb/invalidations    #3445       +/-   ##
=====================================================
+ Coverage             28.28%   85.58%   +57.30%     
=====================================================
  Files                   190      191        +1     
  Lines                 19229    19468      +239     
=====================================================
+ Hits                   5438    16661    +11223     
+ Misses                13791     2807    -10984     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: d05e89b	Previous: 44ac4eb	Ratio
array/accumulate/Float32/1d	55226 ns	55308 ns	1.00
array/accumulate/Float32/dims=1	39830 ns	40571 ns	0.98
array/accumulate/Float32/dims=1L	82357 ns	83176 ns	0.99
array/accumulate/Float32/dims=2	50354 ns	50536 ns	1.00
array/accumulate/Float32/dims=2L	397829 ns	399321 ns	1.00
array/accumulate/Int64/1d	88250 ns	88873 ns	0.99
array/accumulate/Int64/dims=1	60221 ns	60537 ns	0.99
array/accumulate/Int64/dims=1L	140077 ns	140593 ns	1.00
array/accumulate/Int64/dims=2	69611 ns	68766 ns	1.01
array/accumulate/Int64/dims=2L	797724 ns	796187 ns	1.00
array/broadcast	10241 ns	10301 ns	0.99
array/broadcast launch	6625.4 ns	6884.4 ns	0.96
array/construct	926.0277777777778 ns	894.6666666666666 ns	1.04
array/copy	16630 ns	17076 ns	0.97
array/copyto!/cpu_to_gpu	207570 ns	208724 ns	0.99
array/copyto!/gpu_to_cpu	240347 ns	240881 ns	1.00
array/copyto!/gpu_to_gpu	8663 ns	8932 ns	0.97
array/iteration/findall/bool	48712 ns	50126 ns	0.97
array/iteration/findall/int	65637 ns	65404 ns	1.00
array/iteration/findfirst/bool	35645 ns	35653 ns	1.00
array/iteration/findfirst/int	35074 ns	35876 ns	0.98
array/iteration/findmin/1d	50977 ns	51420 ns	0.99
array/iteration/findmin/2d	52207 ns	53587 ns	0.97
array/iteration/logical	50089 ns	51149 ns	0.98
array/iteration/scalar	59396 ns	61238 ns	0.97
array/permutedims/2d	47109 ns	47224 ns	1.00
array/permutedims/3d	44688 ns	44784 ns	1.00
array/permutedims/4d	47830 ns	48464 ns	0.99
array/random/rand/Float32	11038 ns	11176 ns	0.99
array/random/rand/Int64	17741 ns	18087 ns	0.98
array/random/rand!/Float32	8182 ns	8279.666666666666 ns	0.99
array/random/rand!/Int64	15569 ns	15676 ns	0.99
array/random/randn/Float32	33964 ns	34429 ns	0.99
array/random/randn!/Float32	24832 ns	25461 ns	0.98
array/reductions/mapreduce/Float32/1d	35487 ns	36711 ns	0.97
array/reductions/mapreduce/Float32/dims=1	20294 ns	20536 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	57774 ns	58081 ns	0.99
array/reductions/mapreduce/Float32/dims=2	28333 ns	28592 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	58398 ns	58294 ns	1.00
array/reductions/mapreduce/Int64/1d	36109 ns	36620 ns	0.99
array/reductions/mapreduce/Int64/dims=1	22727 ns	22994 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	97210 ns	97464 ns	1.00
array/reductions/mapreduce/Int64/dims=2	29169 ns	29433 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	83472 ns	83514 ns	1.00
array/reductions/reduce/Float32/1d	35879 ns	36457 ns	0.98
array/reductions/reduce/Float32/dims=1	20293 ns	20403 ns	0.99
array/reductions/reduce/Float32/dims=1L	58149 ns	57761 ns	1.01
array/reductions/reduce/Float32/dims=2	28131 ns	28033 ns	1.00
array/reductions/reduce/Float32/dims=2L	58403 ns	58514 ns	1.00
array/reductions/reduce/Int64/1d	36438 ns	36852 ns	0.99
array/reductions/reduce/Int64/dims=1	22441 ns	23003 ns	0.98
array/reductions/reduce/Int64/dims=1L	97498 ns	97475 ns	1.00
array/reductions/reduce/Int64/dims=2	29597 ns	29524 ns	1.00
array/reductions/reduce/Int64/dims=2L	84062 ns	84151 ns	1.00
array/reverse/1d	19854 ns	19992 ns	0.99
array/reverse/1dL	73695 ns	73831 ns	1.00
array/reverse/1dL_inplace	70590 ns	70561 ns	1.00
array/reverse/1d_inplace	12362 ns	12362 ns	1
array/reverse/2d	22559 ns	23022 ns	0.98
array/reverse/2dL	80854 ns	81365 ns	0.99
array/reverse/2dL_inplace	71169 ns	71062 ns	1.00
array/reverse/2d_inplace	15089 ns	14773 ns	1.02
array/sorting/1d	750292 ns	750546 ns	1.00
array/sorting/2d	194506 ns	194226 ns	1.00
array/sorting/by	927234 ns	928901 ns	1.00
cuda/graph/capture	36175 ns	36564 ns	0.99
cuda/graph/eager	32692 ns	34234 ns	0.95
cuda/graph/launch	13857 ns	13955 ns	0.99
cuda/synchronization/context/auto	2302.222222222222 ns	2315.8888888888887 ns	0.99
cuda/synchronization/context/blocking	902.1190476190476 ns	910.175 ns	0.99
cuda/synchronization/context/nonblocking	10820 ns	10524 ns	1.03
cuda/synchronization/stream/auto	870.25 ns	892.1086956521739 ns	0.98
cuda/synchronization/stream/blocking	791.979381443299 ns	794.3020833333334 ns	1.00
cuda/synchronization/stream/nonblocking	8753.666666666666 ns	8287.666666666666 ns	1.06
integration/byval/reference	147671 ns	147653 ns	1.00
integration/byval/slices=1	148796 ns	148788 ns	1.00
integration/byval/slices=2	290356 ns	290449 ns	1.00
integration/byval/slices=3	432493 ns	432565 ns	1.00
integration/cudadevrt	105004 ns	105032 ns	1.00
integration/volumerhs	9609724 ns	9610762 ns	1.00
kernel/indexing	11591 ns	11452 ns	1.01
kernel/indexing_checked	11996 ns	12086 ns	0.99
kernel/launch	2332.222222222222 ns	2338.8888888888887 ns	1.00
kernel/occupancy	688.0457516339869 ns	668.6883116883117 ns	1.03
kernel/rand	14817 ns	15164 ns	0.98
latency/import	4095438826 ns	4065671956 ns	1.01
latency/precompile	5034377782 ns	5034183307 ns	1.00
latency/ttfp	4995870249 ns	4961406691 ns	1.01

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