Switch to ParallelTestRunner

[maleadt]

Needs JuliaTesting/ParallelTestRunner.jl#129

[maleadt]

51/41/37 mins here vs 42/34/33 on master for Julia 1.11/1.12/1.13. And this surprisingly looks like the actual tests slowing down, e.g. on Julia 1.12:

gpuarrays/linalg/core                          (9) │   152.89 │   0.02 │      28.29 │   148.00 │   4.48 │  2.9 │   13223.51 │  4035.78 │
gpuarrays/linalg/norm                         (13) │   281.17 │   0.02 │       0.03 │   130.00 │   7.46 │  2.7 │   17295.61 │  3769.99 │

vs

gpuarrays/linalg/core                         (3) |   126.39 |   0.02 |  0.0 |      28.29 |   226.00 |   3.43 |  2.7 |   11518.49 | 10666.16 |
gpuarrays/linalg/norm                         (4) |   250.23 |   0.02 |  0.0 |       0.03 |   146.00 |   6.00 |  2.4 |   14957.74 |  6904.70 |

[giordano]

Try with --verbose which shows also the init time?

[maleadt]

Good idea. It's probably related to my tuning of the memory pool heuristics though, and not because of PTR.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 63a68c4	Previous: 9702041	Ratio
array/accumulate/Float32/1d	101140 ns	101336 ns	1.00
array/accumulate/Float32/dims=1	76267 ns	76569 ns	1.00
array/accumulate/Float32/dims=1L	1585826 ns	1585630.5 ns	1.00
array/accumulate/Float32/dims=2	143428 ns	143872.5 ns	1.00
array/accumulate/Float32/dims=2L	657493.5 ns	657817.5 ns	1.00
array/accumulate/Int64/1d	118646 ns	118546 ns	1.00
array/accumulate/Int64/dims=1	80280 ns	80328 ns	1.00
array/accumulate/Int64/dims=1L	1706680.5 ns	1694780.5 ns	1.01
array/accumulate/Int64/dims=2	157247.5 ns	156646 ns	1.00
array/accumulate/Int64/dims=2L	962003 ns	962572 ns	1.00
array/broadcast	20670.5 ns	20507 ns	1.01
array/construct	1264.5 ns	1260.6 ns	1.00
array/copy	17872 ns	18042.5 ns	0.99
array/copyto!/cpu_to_gpu	214938 ns	216139.5 ns	0.99
array/copyto!/gpu_to_cpu	283434 ns	282550 ns	1.00
array/copyto!/gpu_to_gpu	10684 ns	10770 ns	0.99
array/iteration/findall/bool	134489 ns	134891 ns	1.00
array/iteration/findall/int	149743 ns	150607 ns	0.99
array/iteration/findfirst/bool	81215 ns	81621 ns	1.00
array/iteration/findfirst/int	83429.5 ns	83931 ns	0.99
array/iteration/findmin/1d	88136.5 ns	88319.5 ns	1.00
array/iteration/findmin/2d	117332.5 ns	116740 ns	1.01
array/iteration/logical	200219.5 ns	199127 ns	1.01
array/iteration/scalar	67096 ns	69801 ns	0.96
array/permutedims/2d	52173.5 ns	51913 ns	1.01
array/permutedims/3d	52747 ns	52967 ns	1.00
array/permutedims/4d	51373 ns	51865.5 ns	0.99
array/random/rand/Float32	12818 ns	12969 ns	0.99
array/random/rand/Int64	24941 ns	24834 ns	1.00
array/random/rand!/Float32	8318.666666666666 ns	8996.666666666666 ns	0.92
array/random/rand!/Int64	21893 ns	21694 ns	1.01
array/random/randn/Float32	37834 ns	37552.5 ns	1.01
array/random/randn!/Float32	30772 ns	30840 ns	1.00
array/reductions/mapreduce/Float32/1d	34319 ns	35074 ns	0.98
array/reductions/mapreduce/Float32/dims=1	40492 ns	40645 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	51236 ns	51243.5 ns	1.00
array/reductions/mapreduce/Float32/dims=2	56197 ns	56526 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	69138.5 ns	69833 ns	0.99
array/reductions/mapreduce/Int64/1d	42165.5 ns	43336 ns	0.97
array/reductions/mapreduce/Int64/dims=1	42715 ns	42698 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	87109 ns	87131 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59317 ns	59771 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	84317 ns	84773.5 ns	0.99
array/reductions/reduce/Float32/1d	34432 ns	34857.5 ns	0.99
array/reductions/reduce/Float32/dims=1	40092 ns	39525 ns	1.01
array/reductions/reduce/Float32/dims=1L	51381 ns	51365 ns	1.00
array/reductions/reduce/Float32/dims=2	56802 ns	56840 ns	1.00
array/reductions/reduce/Float32/dims=2L	69698 ns	69873 ns	1.00
array/reductions/reduce/Int64/1d	42618 ns	43072 ns	0.99
array/reductions/reduce/Int64/dims=1	43391.5 ns	42870 ns	1.01
array/reductions/reduce/Int64/dims=1L	87117 ns	87150 ns	1.00
array/reductions/reduce/Int64/dims=2	59838 ns	59657.5 ns	1.00
array/reductions/reduce/Int64/dims=2L	84509 ns	84754 ns	1.00
array/reverse/1d	17975 ns	17920 ns	1.00
array/reverse/1dL	68561 ns	68474 ns	1.00
array/reverse/1dL_inplace	65693 ns	65769 ns	1.00
array/reverse/1d_inplace	10259 ns	10329.333333333334 ns	0.99
array/reverse/2d	20756 ns	20545 ns	1.01
array/reverse/2dL	72734 ns	72599 ns	1.00
array/reverse/2dL_inplace	65813 ns	65843 ns	1.00
array/reverse/2d_inplace	9934 ns	9925 ns	1.00
array/sorting/1d	2734738 ns	2735157 ns	1.00
array/sorting/2d	1068707 ns	1068027 ns	1.00
array/sorting/by	3304798 ns	3304860 ns	1.00
cuda/synchronization/context/auto	1158.2 ns	1186.5 ns	0.98
cuda/synchronization/context/blocking	941.1315789473684 ns	933.8571428571429 ns	1.01
cuda/synchronization/context/nonblocking	8496 ns	7202.1 ns	1.18
cuda/synchronization/stream/auto	1027.875 ns	1047.75 ns	0.98
cuda/synchronization/stream/blocking	835.5714285714286 ns	845.7979797979798 ns	0.99
cuda/synchronization/stream/nonblocking	7455 ns	7400.299999999999 ns	1.01
integration/byval/reference	143779 ns	143847 ns	1.00
integration/byval/slices=1	145974 ns	145799 ns	1.00
integration/byval/slices=2	284791 ns	284592 ns	1.00
integration/byval/slices=3	423503.5 ns	423121 ns	1.00
integration/cudadevrt	102569 ns	102350 ns	1.00
integration/volumerhs	23499620 ns	23414128.5 ns	1.00
kernel/indexing	13265 ns	13228 ns	1.00
kernel/indexing_checked	14047 ns	14089 ns	1.00
kernel/launch	2127.8888888888887 ns	2207.5555555555557 ns	0.96
kernel/occupancy	722.0289855072464 ns	663.5094339622641 ns	1.09
kernel/rand	16687 ns	14119 ns	1.18
latency/import	3837952002 ns	3828719288.5 ns	1.00
latency/precompile	4578362120 ns	4584717609 ns	1.00
latency/ttfp	4425030626 ns	4415845715.5 ns	1.00

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

[maleadt]

The only remaining issue is the multigpu one. I'm looking into this separately, but I think we can merge this already.

[giordano]

What's the issue specifically?

[maleadt]

What's the issue specifically?

Test failure due to driver bug (presumably), unrelated to PTR but it does seem to make it occur more often.