[maleadt]

CUDA_Compiler.ptxas(), nvlink() and nvdisasm() aren't inferred, so interpolating them into command literals was compiled with an edge to Base.arg_gen(::Any). FilePathsBase extends arg_gen, so with a package depending on it loaded before CUDA.jl, Julia discarded the precompiled code that queries the compiler version when loading CUDA.jl, and with it a large part of CUDACore's precompiled compilation and launch code (about 1500 method instances with a set of common packages, including DataFrames, CairoMakie and Dagger, loaded first). These functions return Cmds for both the JLL and the local toolkit, so this asserts that.

This is part of a series addressing JuliaGPU/GPUCompiler.jl#1048 (with JuliaGPU/GPUCompiler.jl#1052, JuliaLLVM/LLVM.jl#702 and maleadt/CompilerCaching.jl#33). With all of them, the first @cuda kernel (plus a broadcast and sum) with ChainRulesCore loaded first goes from 4.25 s to 2.76 s (2.52 s without it), and with DataFrames from 7.12 s to 5.72 s, on Julia 1.13 and an RTX 5080. Tested with the core codegen, execution, exceptions, KernelAbstractions, broadcast and array tests.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 337cfd2	Previous: a9e28c0	Ratio
array/accumulate/Float32/1d	55120 ns	53368 ns	1.03
array/accumulate/Float32/dims=1	39901 ns	39146 ns	1.02
array/accumulate/Float32/dims=1L	82731 ns	82278 ns	1.01
array/accumulate/Float32/dims=2	50151 ns	49464 ns	1.01
array/accumulate/Float32/dims=2L	399975 ns	399782 ns	1.00
array/accumulate/Int64/1d	89063 ns	87737 ns	1.02
array/accumulate/Int64/dims=1	60450 ns	59754 ns	1.01
array/accumulate/Int64/dims=1L	140443 ns	139769 ns	1.00
array/accumulate/Int64/dims=2	69095 ns	68524 ns	1.01
array/accumulate/Int64/dims=2L	796007 ns	797529 ns	1.00
array/broadcast	10394 ns	10132 ns	1.03
array/broadcast launch	6742.4 ns	6591 ns	1.02
array/construct	906.55 ns	934.15 ns	0.97
array/copy	17239 ns	16973 ns	1.02
array/copyto!/cpu_to_gpu	208686 ns	206527 ns	1.01
array/copyto!/gpu_to_cpu	241308 ns	240008 ns	1.01
array/copyto!/gpu_to_gpu	8785 ns	8794 ns	1.00
array/iteration/findall/bool	48834 ns	47090 ns	1.04
array/iteration/findall/int	65488 ns	65086 ns	1.01
array/iteration/findfirst/bool	35545 ns	35389 ns	1.00
array/iteration/findfirst/int	36050 ns	34804 ns	1.04
array/iteration/findmin/1d	51026 ns	50387 ns	1.01
array/iteration/findmin/2d	52887 ns	51323 ns	1.03
array/iteration/logical	50616 ns	49399 ns	1.02
array/iteration/scalar	59599 ns	60128 ns	0.99
array/permutedims/2d	47205 ns	46699 ns	1.01
array/permutedims/3d	44741 ns	44729 ns	1.00
array/permutedims/4d	47711 ns	47477 ns	1.00
array/random/rand/Float32	10997 ns	10864 ns	1.01
array/random/rand/Int64	17785 ns	17540 ns	1.01
array/random/rand!/Float32	8257.666666666666 ns	8222.333333333334 ns	1.00
array/random/rand!/Int64	15484 ns	15438 ns	1.00
array/random/randn/Float32	34078 ns	33917 ns	1.00
array/random/randn!/Float32	25264 ns	24950 ns	1.01
array/reductions/mapreduce/Float32/1d	35880 ns	35344 ns	1.02
array/reductions/mapreduce/Float32/dims=1	20340 ns	19440 ns	1.05
array/reductions/mapreduce/Float32/dims=1L	57901 ns	57177 ns	1.01
array/reductions/mapreduce/Float32/dims=2	28606 ns	26865 ns	1.06
array/reductions/mapreduce/Float32/dims=2L	57968 ns	57426 ns	1.01
array/reductions/mapreduce/Int64/1d	36284 ns	35963 ns	1.01
array/reductions/mapreduce/Int64/dims=1	23212 ns	22076 ns	1.05
array/reductions/mapreduce/Int64/dims=1L	97277 ns	96669 ns	1.01
array/reductions/mapreduce/Int64/dims=2	29844 ns	28562 ns	1.04
array/reductions/mapreduce/Int64/dims=2L	84044 ns	82641 ns	1.02
array/reductions/reduce/Float32/1d	35472 ns	35477 ns	1.00
array/reductions/reduce/Float32/dims=1	20638 ns	19590 ns	1.05
array/reductions/reduce/Float32/dims=1L	58237 ns	56900 ns	1.02
array/reductions/reduce/Float32/dims=2	28890 ns	27235 ns	1.06
array/reductions/reduce/Float32/dims=2L	58503 ns	57560 ns	1.02
array/reductions/reduce/Int64/1d	36122 ns	36458 ns	0.99
array/reductions/reduce/Int64/dims=1	23131 ns	21785 ns	1.06
array/reductions/reduce/Int64/dims=1L	97923 ns	96838 ns	1.01
array/reductions/reduce/Int64/dims=2	29733 ns	28069 ns	1.06
array/reductions/reduce/Int64/dims=2L	84525 ns	82996 ns	1.02
array/reverse/1d	19685 ns	19163 ns	1.03
array/reverse/1dL	73557 ns	73140 ns	1.01
array/reverse/1dL_inplace	70143 ns	70001 ns	1.00
array/reverse/1d_inplace	12070 ns	11808 ns	1.02
array/reverse/2d	22795 ns	21957 ns	1.04
array/reverse/2dL	81178 ns	80616 ns	1.01
array/reverse/2dL_inplace	70927 ns	70530 ns	1.01
array/reverse/2d_inplace	14854 ns	14461 ns	1.03
array/sorting/1d	751038 ns	749551 ns	1.00
array/sorting/2d	194450 ns	193577 ns	1.00
array/sorting/by	927915 ns	925685 ns	1.00
cuda/graph/capture	38086 ns	35335 ns	1.08
cuda/graph/eager	33886 ns	33696 ns	1.01
cuda/graph/launch	14053 ns	13916 ns	1.01
cuda/synchronization/context/auto	2456.4444444444443 ns	2327.1111111111113 ns	1.06
cuda/synchronization/context/blocking	908.421052631579 ns	896.9761904761905 ns	1.01
cuda/synchronization/context/nonblocking	8919 ns	10394 ns	0.86
cuda/synchronization/stream/auto	857.9117647058823 ns	869.9464285714286 ns	0.99
cuda/synchronization/stream/blocking	789.2857142857143 ns	774.0733944954128 ns	1.02
cuda/synchronization/stream/nonblocking	8494.666666666666 ns	8489 ns	1.00
integration/byval/reference	147589 ns	147636 ns	1.00
integration/byval/slices=1	148703 ns	148857 ns	1.00
integration/byval/slices=2	290104 ns	290245 ns	1.00
integration/byval/slices=3	432309 ns	432133 ns	1.00
integration/cudadevrt	104994 ns	104872 ns	1.00
integration/volumerhs	9609632 ns	9610189 ns	1.00
kernel/indexing	11486 ns	11356 ns	1.01
kernel/indexing_checked	12048 ns	11875 ns	1.01
kernel/launch	2400.6666666666665 ns	2348.3333333333335 ns	1.02
kernel/occupancy	726.6969696969697 ns	704.8057553956835 ns	1.03
kernel/rand	14867 ns	14971 ns	0.99
latency/import	4087712198 ns	4087629357 ns	1.00
latency/precompile	5044437374 ns	5041187065 ns	1.00
latency/ttfp	4976136576 ns	4978622722 ns	1.00

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

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 85.58%. Comparing base (44ac4eb) to head (337cfd2).
⚠️ Report is 4 commits behind head on main.

Files with missing lines	Patch %	Lines
CUDACore/lib/cudadrv/version.jl	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3440      +/-   ##
==========================================
- Coverage   87.46%   85.58%   -1.89%     
==========================================
  Files         191      191              
  Lines       19646    19469     -177     
==========================================
- Hits        17184    16663     -521     
- Misses       2462     2806     +344     

☔ 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.