Flush kernel output when synchronizing a stream that polling finds done

[maleadt]

Since #3315, synchronize() returns without synchronizing when a query finds the stream done. The driver only flushes the device's printf buffer when synchronizing, not when querying, so the output of short kernels only shows up after a later kernel launch or blocking synchronization:

using CUDA
kernel() = (@cuprintln("hello"); nothing)
@cuda kernel(); synchronize()
out = Pipe()
redirect_stdout(out) do
    @cuda kernel(); synchronize()
end
close(out.in); read(out, String)   # v6.4.1: "hello\n", v6.4.2 and main: ""

This broke cuTile.jl's printing tests, which capture the output of @cuda ...; synchronize(). CUDA.jl's own tests didn't catch it because @grab_output synchronizes the context, which always flushes.

Synchronizing the stream after a successful query would flush, but would also wait for work submitted since, which #3315 avoided (684 ms vs. 0.03 ms when another kernel was queued after the query). Instead, synchronize an event that was never recorded, as check_exceptions already did: that flushes without waiting for any work. Events keep synchronizing themselves after a successful query (#3346), which also flushes.

The event isn't cached, because CUDA 11 drivers can't detect a destroyed context whose handle is reused. On an RTX 5080, synchronize() of an idle stream goes from 0.07 to 0.29 µs, and a launch followed by synchronize() from 4.97 to 5.27 µs. A cached event would save 0.1 µs of that.

Not covered: synchronize(::CuArray) still doesn't flush when the array's last use is known to have completed (in v6.4.1 it synchronized the stream).

[maleadt]

Eh, not sure about this.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.34%. Comparing base (2f34355) to head (e768d39).
⚠️ Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3363      +/-   ##
==========================================
- Coverage   87.31%   85.34%   -1.97%     
==========================================
  Files         195      194       -1     
  Lines       20066    19984      -82     
==========================================
- Hits        17521    17056     -465     
- Misses       2545     2928     +383     

☔ 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: e768d39	Previous: 2f34355	Ratio
array/accumulate/Float32/1d	105912 ns	105820 ns	1.00
array/accumulate/Float32/dims=1	78020 ns	77794 ns	1.00
array/accumulate/Float32/dims=1L	1684559 ns	1675264 ns	1.01
array/accumulate/Float32/dims=2	148817 ns	148368 ns	1.00
array/accumulate/Float32/dims=2L	685784 ns	684464 ns	1.00
array/accumulate/Int64/1d	123990 ns	123400 ns	1.00
array/accumulate/Int64/dims=1	82375 ns	81768 ns	1.01
array/accumulate/Int64/dims=1L	1805420 ns	1798675 ns	1.00
array/accumulate/Int64/dims=2	163670 ns	163560 ns	1.00
array/accumulate/Int64/dims=2L	988832 ns	988728 ns	1.00
array/broadcast	17743 ns	17596 ns	1.01
array/broadcast launch	8368.666666666666 ns	8338.666666666666 ns	1.00
array/construct	906.9459459459459 ns	967.36 ns	0.94
array/copy	17825 ns	17621 ns	1.01
array/copyto!/cpu_to_gpu	208022 ns	207798 ns	1.00
array/copyto!/gpu_to_cpu	241069 ns	239887 ns	1.00
array/copyto!/gpu_to_gpu	9612 ns	9569 ns	1.00
array/iteration/findall/bool	141942 ns	140093 ns	1.01
array/iteration/findall/int	152875 ns	151472 ns	1.01
array/iteration/findfirst/bool	76429 ns	74672 ns	1.02
array/iteration/findfirst/int	78721 ns	77410 ns	1.02
array/iteration/findmin/1d	64188 ns	67569 ns	0.95
array/iteration/findmin/2d	104353 ns	104838 ns	1.00
array/iteration/logical	203373 ns	198787 ns	1.02
array/iteration/scalar	65099 ns	55807 ns	1.17
array/permutedims/2d	50497 ns	50509 ns	1.00
array/permutedims/3d	51977 ns	51420 ns	1.01
array/permutedims/4d	52895 ns	52481 ns	1.01
array/random/rand/Float32	10935 ns	10776 ns	1.01
array/random/rand/Int64	19173 ns	19059 ns	1.01
array/random/rand!/Float32	8217.666666666666 ns	8164 ns	1.01
array/random/rand!/Int64	16686 ns	16660 ns	1.00
array/random/randn/Float32	34773 ns	34551 ns	1.01
array/random/randn!/Float32	24891 ns	24847 ns	1.00
array/reductions/mapreduce/Float32/1d	35406 ns	33391 ns	1.06
array/reductions/mapreduce/Float32/dims=1	41146 ns	40679 ns	1.01
array/reductions/mapreduce/Float32/dims=1L	54404 ns	54218 ns	1.00
array/reductions/mapreduce/Float32/dims=2	60129 ns	59491 ns	1.01
array/reductions/mapreduce/Float32/dims=2L	72131 ns	72017 ns	1.00
array/reductions/mapreduce/Int64/1d	41624 ns	39675 ns	1.05
array/reductions/mapreduce/Int64/dims=1	43520 ns	43359 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	90243 ns	89909 ns	1.00
array/reductions/mapreduce/Int64/dims=2	61195 ns	60860 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	88556 ns	87707 ns	1.01
array/reductions/reduce/Float32/1d	35374 ns	33503 ns	1.06
array/reductions/reduce/Float32/dims=1	40803 ns	40509 ns	1.01
array/reductions/reduce/Float32/dims=1L	53815 ns	53371 ns	1.01
array/reductions/reduce/Float32/dims=2	59896 ns	59505 ns	1.01
array/reductions/reduce/Float32/dims=2L	72328 ns	72117 ns	1.00
array/reductions/reduce/Int64/1d	41637 ns	39620 ns	1.05
array/reductions/reduce/Int64/dims=1	43984 ns	43689 ns	1.01
array/reductions/reduce/Int64/dims=1L	90279 ns	89931 ns	1.00
array/reductions/reduce/Int64/dims=2	61105 ns	60828 ns	1.00
array/reductions/reduce/Int64/dims=2L	88219 ns	87372 ns	1.01
array/reverse/1d	17221 ns	16830 ns	1.02
array/reverse/1dL	71009 ns	70757 ns	1.00
array/reverse/1dL_inplace	68318 ns	68190 ns	1.00
array/reverse/1d_inplace	10098 ns	9767 ns	1.03
array/reverse/2d	21528 ns	21534 ns	1.00
array/reverse/2dL	75746 ns	75376 ns	1.00
array/reverse/2dL_inplace	68468 ns	68399 ns	1.00
array/reverse/2d_inplace	11436 ns	11264 ns	1.02
array/sorting/1d	2845601 ns	2842268 ns	1.00
array/sorting/2d	1090112 ns	1088961 ns	1.00
array/sorting/by	3396672 ns	3394714 ns	1.00
cuda/graph/capture	48735 ns	46949 ns	1.04
cuda/graph/eager	47261 ns	46953 ns	1.01
cuda/graph/launch	15002 ns	14823 ns	1.01
cuda/synchronization/context/auto	8124.75 ns	7993.75 ns	1.02
cuda/synchronization/context/blocking	893.7234042553191 ns	909.9767441860465 ns	0.98
cuda/synchronization/context/nonblocking	8158 ns	7999 ns	1.02
cuda/synchronization/stream/auto	1206.3 ns	348.4562211981567 ns	3.46
cuda/synchronization/stream/blocking	760.0084033613446 ns	767.9469026548672 ns	0.99
cuda/synchronization/stream/nonblocking	7748.666666666667 ns	7882.333333333333 ns	0.98
integration/byval/reference	148894 ns	148584 ns	1.00
integration/byval/slices=1	149782 ns	149648 ns	1.00
integration/byval/slices=2	291999 ns	291810 ns	1.00
integration/byval/slices=3	434910 ns	434504 ns	1.00
integration/cudadevrt	106334 ns	106021 ns	1.00
integration/volumerhs	9207468 ns	9207188 ns	1.00
kernel/indexing	12393 ns	12133 ns	1.02
kernel/indexing_checked	12878 ns	12620 ns	1.02
kernel/launch	2498.5555555555557 ns	2497.1111111111113 ns	1.00
kernel/occupancy	963.3157894736842 ns	939.4166666666666 ns	1.03
kernel/rand	15601 ns	15434 ns	1.01
latency/import	4334723655 ns	4338023747 ns	1.00
latency/precompile	5118293576 ns	5130684551 ns	1.00
latency/ttfp	4845303819 ns	4803836711 ns	1.01

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