[maleadt]

Since #3315, device_synchronize() always hands the synchronization to a worker thread, because a context can't be polled. On the benchmark dashboard, cuda/synchronization/context/auto went from ~1.0 µs to ~6.8 µs as a result. It reached ~7.8 µs after the #3318 stack landed: the worker handoff got ~10% slower, and gpuci1's clocks were locked the same day.

CUDA 12.5 added cuCtxRecordEvent, which records an event capturing all of a context's work (including its green contexts). Unlike the context, that event can be polled. On 12.5+ drivers, nonblocking device_synchronize now goes through the existing event path: poll, and if done, synchronize the event so that kernel output is flushed; otherwise wait on a worker as before. The event gets EVENT_BLOCKING_SYNC when the context is configured with CTX_SCHED_BLOCKING_SYNC, so long waits keep following the context's scheduling policy. Blocking synchronization and older drivers still use cuCtxSynchronize.

Median times on an RTX 5080 (3 interleaved runs):

	main	PR
context/auto	2.04–2.25 µs	0.66–0.77 µs
context/nonblocking	2.01–2.44 µs	1.94–1.99 µs
context/blocking	172–176 ns	171–176 ns

The printf-flush test now covers device_synchronize as well, and the non-blocking-stream gate test also runs with spin=false.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 282fa89	Previous: ea49380	Ratio
array/accumulate/Float32/1d	64294 ns	64036 ns	1.00
array/accumulate/Float32/dims=1	209193 ns	208812 ns	1.00
array/accumulate/Float32/dims=1L	231191 ns	230609 ns	1.00
array/accumulate/Float32/dims=2	93552 ns	92377 ns	1.01
array/accumulate/Float32/dims=2L	4414263 ns	4414351 ns	1.00
array/accumulate/Int64/1d	106455 ns	106105 ns	1.00
array/accumulate/Int64/dims=1	224501 ns	222970 ns	1.01
array/accumulate/Int64/dims=1L	404588 ns	403514 ns	1.00
array/accumulate/Int64/dims=2	112681 ns	111534 ns	1.01
array/accumulate/Int64/dims=2L	5302726 ns	5303201 ns	1.00
array/broadcast	10789 ns	10425 ns	1.03
array/broadcast launch	7303.75 ns	6876 ns	1.06
array/construct	951.0416666666666 ns	958.1 ns	0.99
array/copy	17160 ns	17084 ns	1.00
array/copyto!/cpu_to_gpu	210619 ns	206949 ns	1.02
array/copyto!/gpu_to_cpu	242513 ns	240012 ns	1.01
array/copyto!/gpu_to_gpu	9057 ns	8931 ns	1.01
array/iteration/findall/bool	47968 ns	48499 ns	0.99
array/iteration/findall/int	71072 ns	69835 ns	1.02
array/iteration/findfirst/bool	116565 ns	114590 ns	1.02
array/iteration/findfirst/int	113982 ns	112292 ns	1.02
array/iteration/findmin/1d	67618 ns	66134 ns	1.02
array/iteration/findmin/2d	68300 ns	67092 ns	1.02
array/iteration/logical	52954 ns	52718 ns	1.00
array/iteration/scalar	61768 ns	62269 ns	0.99
array/permutedims/2d	47814 ns	47213 ns	1.01
array/permutedims/3d	45014 ns	44662 ns	1.01
array/permutedims/4d	48297 ns	47947 ns	1.01
array/random/rand/Float32	11289 ns	11026 ns	1.02
array/random/rand/Int64	18093 ns	17878 ns	1.01
array/random/rand!/Float32	8452 ns	8338.333333333334 ns	1.01
array/random/rand!/Int64	15869 ns	15757 ns	1.01
array/random/randn/Float32	34339 ns	34043 ns	1.01
array/random/randn!/Float32	25136 ns	25361 ns	0.99
array/reductions/mapreduce/Float32/1d	50875 ns	49702 ns	1.02
array/reductions/mapreduce/Float32/dims=1	43934 ns	42350 ns	1.04
array/reductions/mapreduce/Float32/dims=1L	73017 ns	72621 ns	1.01
array/reductions/mapreduce/Float32/dims=2	45479 ns	44306 ns	1.03
array/reductions/mapreduce/Float32/dims=2L	74455 ns	73613 ns	1.01
array/reductions/mapreduce/Int64/1d	52227 ns	50450 ns	1.04
array/reductions/mapreduce/Int64/dims=1	46088 ns	45556 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	113518 ns	113168 ns	1.00
array/reductions/mapreduce/Int64/dims=2	46770 ns	46043 ns	1.02
array/reductions/mapreduce/Int64/dims=2L	100916 ns	99328 ns	1.02
array/reductions/reduce/Float32/1d	50629 ns	48970 ns	1.03
array/reductions/reduce/Float32/dims=1	44848 ns	42528 ns	1.05
array/reductions/reduce/Float32/dims=1L	73899 ns	72521 ns	1.02
array/reductions/reduce/Float32/dims=2	46348 ns	44789 ns	1.03
array/reductions/reduce/Float32/dims=2L	74798 ns	73801 ns	1.01
array/reductions/reduce/Int64/1d	52012 ns	50745 ns	1.02
array/reductions/reduce/Int64/dims=1	46296 ns	45758 ns	1.01
array/reductions/reduce/Int64/dims=1L	113841 ns	112810 ns	1.01
array/reductions/reduce/Int64/dims=2	46848 ns	46313 ns	1.01
array/reductions/reduce/Int64/dims=2L	101315 ns	100440 ns	1.01
array/reverse/1d	24024 ns	23468 ns	1.02
array/reverse/1dL	129723 ns	129150 ns	1.00
array/reverse/1dL_inplace	71117 ns	70962 ns	1.00
array/reverse/1d_inplace	13010 ns	12558 ns	1.04
array/reverse/2d	23213 ns	22933 ns	1.01
array/reverse/2dL	81687 ns	81376 ns	1.00
array/reverse/2dL_inplace	71643 ns	71131 ns	1.01
array/reverse/2d_inplace	15545 ns	15114 ns	1.03
array/sorting/1d	750713 ns	751253 ns	1.00
array/sorting/2d	195436 ns	194701 ns	1.00
array/sorting/by	928926 ns	928545 ns	1.00
cuda/graph/capture	38199 ns	39420 ns	0.97
cuda/graph/eager	36085 ns	35936 ns	1.00
cuda/graph/launch	13919 ns	13848 ns	1.01
cuda/synchronization/context/auto	2321 ns	7882 ns	0.29
cuda/synchronization/context/blocking	898.3 ns	900.695652173913 ns	1.00
cuda/synchronization/context/nonblocking	8450 ns	7934 ns	1.07
cuda/synchronization/stream/auto	858.6666666666666 ns	827.6833333333333 ns	1.04
cuda/synchronization/stream/blocking	756.2641509433962 ns	766.9134615384615 ns	0.99
cuda/synchronization/stream/nonblocking	7826.333333333333 ns	7800.75 ns	1.00
integration/byval/reference	147781 ns	147547 ns	1.00
integration/byval/slices=1	148894 ns	148428 ns	1.00
integration/byval/slices=2	290159 ns	289979 ns	1.00
integration/byval/slices=3	432385 ns	432003 ns	1.00
integration/cudadevrt	105190 ns	104936 ns	1.00
integration/volumerhs	9611281 ns	9609908 ns	1.00
kernel/indexing	11531 ns	11544 ns	1.00
kernel/indexing_checked	12287 ns	11943 ns	1.03
kernel/launch	2280.5555555555557 ns	2340.1111111111113 ns	0.97
kernel/occupancy	690.8095238095239 ns	688.5234899328859 ns	1.00
kernel/rand	15143 ns	14871 ns	1.02
latency/import	4750753782 ns	4767312023 ns	1.00
latency/precompile	5379427551 ns	5382871355 ns	1.00
latency/ttfp	5216826222 ns	5251068778 ns	0.99

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

[codecov]

Codecov Report

❌ Patch coverage is 92.59259% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.08%. Comparing base (ea49380) to head (282fa89).
⚠️ Report is 5 commits behind head on main.

Files with missing lines	Patch %	Lines
CUDACore/lib/cudadrv/synchronization.jl	85.71%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3396      +/-   ##
==========================================
- Coverage   86.56%   85.08%   -1.49%     
==========================================
  Files         192      191       -1     
  Lines       19497    19236     -261     
==========================================
- Hits        16878    16366     -512     
- Misses       2619     2870     +251     

☔ 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.