[maleadt]

The driver only flushes the device's printf buffer when synchronizing (cu{Ctx,Stream,Event}Synchronize), launching a kernel, loading a module, etc., not when querying for completion. Since #3315, synchronize() returns without synchronizing when a query finds the stream done, so the output of short kernels only appeared after a later launch or blocking synchronization.

Synchronizing the stream after polling, as before #3315, would wait for work other tasks submitted in the meantime. Instead, this synchronizes an event that is never recorded, which flushes the buffer without waiting for anything. Each context keeps one such event; creating and destroying one on every call, as in #3363, more than tripled the cost of synchronizing an idle stream. Exception reporting already flushed this way and now shares the helper.

synchronize() on an idle stream, RTX 5080, interleaved in one process:

	time
main	66 ns
#3363 (event per call)	292 ns
this PR (cached event)	159 ns

The second commit makes @grab_output synchronize the task's stream instead of the whole context, so that tests capturing kernel output go through the same synchronization as users do. A dedicated test covers the case where polling finds the stream done.

Supersedes #3363.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.23%. Comparing base (853cbdf) to head (2780c9f).
⚠️ Report is 21 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3368      +/-   ##
==========================================
- Coverage   87.40%   85.23%   -2.17%     
==========================================
  Files         195      194       -1     
  Lines       20072    19992      -80     
==========================================
- Hits        17543    17040     -503     
- Misses       2529     2952     +423     

☔ 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: 2780c9f	Previous: 853cbdf	Ratio
array/accumulate/Float32/1d	104414 ns	104545 ns	1.00
array/accumulate/Float32/dims=1	76806 ns	77216 ns	0.99
array/accumulate/Float32/dims=1L	1675140 ns	1681678 ns	1.00
array/accumulate/Float32/dims=2	147550 ns	147990 ns	1.00
array/accumulate/Float32/dims=2L	683700 ns	684823 ns	1.00
array/accumulate/Int64/1d	122420 ns	123482 ns	0.99
array/accumulate/Int64/dims=1	81088 ns	82067 ns	0.99
array/accumulate/Int64/dims=1L	1798181 ns	1804719 ns	1.00
array/accumulate/Int64/dims=2	162182 ns	163653 ns	0.99
array/accumulate/Int64/dims=2L	987247 ns	988348 ns	1.00
array/broadcast	16908 ns	16933 ns	1.00
array/broadcast launch	7542.25 ns	7531.25 ns	1.00
array/construct	896.3846153846154 ns	918.3055555555555 ns	0.98
array/copy	16829 ns	16994 ns	0.99
array/copyto!/cpu_to_gpu	207017 ns	208853 ns	0.99
array/copyto!/gpu_to_cpu	240522 ns	241692 ns	1.00
array/copyto!/gpu_to_gpu	8844 ns	8852 ns	1.00
array/iteration/findall/bool	138262 ns	137908 ns	1.00
array/iteration/findall/int	148144 ns	148516 ns	1.00
array/iteration/findfirst/bool	73705 ns	73524 ns	1.00
array/iteration/findfirst/int	75725 ns	75762 ns	1.00
array/iteration/findmin/1d	60264 ns	66141 ns	0.91
array/iteration/findmin/2d	102740 ns	103221 ns	1.00
array/iteration/logical	192288 ns	188598 ns	1.02
array/iteration/scalar	59579 ns	52780 ns	1.13
array/permutedims/2d	49185 ns	49695 ns	0.99
array/permutedims/3d	50327 ns	50504 ns	1.00
array/permutedims/4d	51485 ns	51914 ns	0.99
array/random/rand/Float32	11088 ns	11126 ns	1.00
array/random/rand/Int64	17986 ns	18386 ns	0.98
array/random/rand!/Float32	8167 ns	8346.666666666666 ns	0.98
array/random/rand!/Int64	15677 ns	16183 ns	0.97
array/random/randn/Float32	34288 ns	34716 ns	0.99
array/random/randn!/Float32	24694 ns	24925 ns	0.99
array/reductions/mapreduce/Float32/1d	31760 ns	31156 ns	1.02
array/reductions/mapreduce/Float32/dims=1	39821 ns	39658 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	53283 ns	53129 ns	1.00
array/reductions/mapreduce/Float32/dims=2	58592 ns	58976 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	70709 ns	71388 ns	0.99
array/reductions/mapreduce/Int64/1d	38989 ns	38448 ns	1.01
array/reductions/mapreduce/Int64/dims=1	42343 ns	42737 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	88870 ns	89229 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59867 ns	60306 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	86369 ns	86904 ns	0.99
array/reductions/reduce/Float32/1d	31850 ns	31707 ns	1.00
array/reductions/reduce/Float32/dims=1	39437 ns	39862 ns	0.99
array/reductions/reduce/Float32/dims=1L	52606 ns	52911 ns	0.99
array/reductions/reduce/Float32/dims=2	58670 ns	58836 ns	1.00
array/reductions/reduce/Float32/dims=2L	70977 ns	71759 ns	0.99
array/reductions/reduce/Int64/1d	38667 ns	38435 ns	1.01
array/reductions/reduce/Int64/dims=1	42491 ns	42816 ns	0.99
array/reductions/reduce/Int64/dims=1L	89070 ns	89319 ns	1.00
array/reductions/reduce/Int64/dims=2	60004 ns	60402 ns	0.99
array/reductions/reduce/Int64/dims=2L	87210 ns	87189 ns	1.00
array/reverse/1d	15883 ns	16005 ns	0.99
array/reverse/1dL	69739 ns	69941 ns	1.00
array/reverse/1dL_inplace	67283 ns	67498 ns	1.00
array/reverse/1d_inplace	9040 ns	9218 ns	0.98
array/reverse/2d	20320 ns	20751 ns	0.98
array/reverse/2dL	74214 ns	74844 ns	0.99
array/reverse/2dL_inplace	67474 ns	67610 ns	1.00
array/reverse/2d_inplace	10450 ns	10714 ns	0.98
array/sorting/1d	2841387 ns	2849188 ns	1.00
array/sorting/2d	1087259 ns	1090151 ns	1.00
array/sorting/by	3393430 ns	3401762 ns	1.00
cuda/graph/capture	40386 ns	40998 ns	0.99
cuda/graph/eager	37550 ns	38515 ns	0.97
cuda/graph/launch	14231 ns	14283 ns	1.00
cuda/synchronization/context/auto	7867.666666666667 ns	7914 ns	0.99
cuda/synchronization/context/blocking	894.8636363636364 ns	861.1481481481482 ns	1.04
cuda/synchronization/context/nonblocking	7935.333333333333 ns	7894.666666666667 ns	1.01
cuda/synchronization/stream/auto	825.7692307692307 ns	340.4063926940639 ns	2.43
cuda/synchronization/stream/blocking	756.2372881355932 ns	760.6017699115044 ns	0.99
cuda/synchronization/stream/nonblocking	7744.5 ns	7937 ns	0.98
integration/byval/reference	147547 ns	147753 ns	1.00
integration/byval/slices=1	148648 ns	148908 ns	1.00
integration/byval/slices=2	290179 ns	290256 ns	1.00
integration/byval/slices=3	432082 ns	432393 ns	1.00
integration/cudadevrt	105099 ns	105057 ns	1.00
integration/volumerhs	9200309 ns	9207429 ns	1.00
kernel/indexing	11514 ns	11777 ns	0.98
kernel/indexing_checked	11891 ns	12029 ns	0.99
kernel/launch	2363.777777777778 ns	2453 ns	0.96
kernel/occupancy	877.6326530612245 ns	885.2075471698113 ns	0.99
kernel/rand	14792 ns	15123 ns	0.98
latency/import	4336855968 ns	4349657474 ns	1.00
latency/precompile	5128596926 ns	5113951581 ns	1.00
latency/ttfp	4853603558 ns	4845722062 ns	1.00

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