[Perf -47%] PerfLabTests.BlockCopyPerf.CallBlockCopy

[DrewScoggins]

Run Information

Architecture	x64
OS	Windows 10.0.18362
Changes	diff

Regressions in PerfLabTests.BlockCopyPerf

Benchmark	Baseline	Test	Test/Base	Modality	Baseline Outlier
CallBlockCopy	5.03 ns	7.39 ns	1.47		False

Historical Data in Reporting System

Repro

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f netcoreapp5.0 --filter 'PerfLabTests.BlockCopyPerf*';

Histogram

PerfLabTests.BlockCopyPerf.CallBlockCopy(numElements: 0)

[4.738 ; 5.406) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[5.406 ; 6.074) | 
[6.074 ; 6.607) | 
[6.607 ; 7.275) | @@@@@@@@@@@@
[7.275 ; 7.979) | @@@@@@@

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this issue. Please help me learn by adding exactly one area label.

[DrewScoggins]

From @adamsitnik

Looks like a very simple regression. I would expect the BlockCopy to return very quickly for 0 elements.

This needs an investigation. We should also consider removing this test case - I am not sure if testing BlockCopy for 0 elements adds value.

[EgorBo]

Both BlockCopy and InitBlockUnaligned should be intrinsics in case of constant (small) size argument - so we can optimize it to no-op or unroll.

[jkotas]

BlockCopy comes with fairly non-trivial argument validation. You would need to hardcode all that into the JIT...

[EgorBo]

BlockCopy comes with fairly non-trivial argument validation. You would need to hardcode all that into the JIT...

We can start from Buffer.MemoryCopy
e.g.

Buffer.MemoryCopy(srcPtr, dstPtr, 8, 8);

to

*(long*)dstPtr = *(long*)strPtr
// or several (unrolled) lines

[jkotas]

I would focus on Span.CopyTo/TryCopyTo and tell everybody to use that if they want this sort of optimizations. Buffer.MemoryCopy is rarely used and legacy, not worth spending target effort on.

Tagging subscribers to this area: @tannergooding
Notify danmosemsft if you want to be subscribed.

[GrabYourPitchforks]

I am not sure if testing BlockCopy for 0 elements adds value.

I agree with this. It's useful as a unit test to make sure we don't error out, but as a perf test I don't think it's all that useful. I'd expect the vast majority use case to pass non-0 arguments to this method. And per Jan's earlier comment, we should really be encouraging use of strongly-typed APIs for optimal throughput.

Closing as won't fix. If this ends up being a problem please reopen.

[DrewScoggins]

Run Information

Architecture	x64
OS	ubuntu 18.04
Changes	diff

Regressions in PerfLabTests.BlockCopyPerf

Benchmark	Baseline	Test	Test/Base	Modality	Baseline Outlier
CallBlockCopy	2.65 ns	6.75 ns	2.54	Bimodal	False
CallBlockCopy	13.28 ns	23.06 ns	1.74	Bimodal	False

Related Issue on x64 Windows

[Perf -43%] PerfLabTests.BlockCopyPerf.CallBlockCopy

Related Issue on x86 Windows

[Perf -74%] PerfLabTests.BlockCopyPerf.CallBlockCopy

Historical Data in Reporting System

Repro

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f netcoreapp5.0 --filter 'PerfLabTests.BlockCopyPerf*';

Histogram

PerfLabTests.BlockCopyPerf.CallBlockCopy(numElements: 0)

[2.111 ;  2.873) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[2.873 ;  3.636) | 
[3.636 ;  4.399) | 
[4.399 ;  4.801) | 
[4.801 ;  5.563) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[5.563 ;  5.865) | 
[5.865 ;  6.549) | @@@@@@@@@@@@@@@@@@@@@
[6.549 ;  7.312) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[7.312 ;  7.924) | @@@@@@@@@@@@@@@
[7.924 ;  8.686) | 
[8.686 ;  9.425) | 
[9.425 ; 10.188) | @

PerfLabTests.BlockCopyPerf.CallBlockCopy(numElements: 1000)

[13.195 ; 15.231) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[15.231 ; 17.393) | @@@@
[17.393 ; 19.429) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[19.429 ; 21.180) | @@@
[21.180 ; 21.995) | 
[21.995 ; 24.031) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[24.031 ; 24.844) | @
[24.844 ; 27.444) | @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
[27.444 ; 29.479) | @@
[29.479 ; 31.515) | 
[31.515 ; 33.732) | 
[33.732 ; 35.803) | @@

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository