Microbenchmark System.Numerics.Tests.Perf_BitOperations.PopCount_ulongis 464 % slower on Wasm

[naricc]

Microbenchmark System.Numerics.Tests.Perf_BitOperations.PopCount_ulongis 464 % slower on Wasm vs. desktop interpreter.

Repro Instructions for running benchmarks locally

Results page when reported.

Automated runs

[ghost]

Tagging subscribers to this area: @tannergooding, @pgovind
See info in area-owners.md if you want to be subscribed.

[EgorBo]

ah, can fix as part of #40069

[SamMonoRT]

@EgorBo - was this fixed as part of #40069 If so, please close this issue.

[naricc]

Looking at the benchmark results for today an the last several days, it seems like this benchmark is now around 280 % slower on Wasm vs. Desktop interpreter. That is probably an acceptable slowdown for wasm.

Latest results . I don't know how to link to just this benchmark, so that will load all of them, which will take a while.

[naricc]

@EgorBo I think we can close this out unless there is any follow up that need to be done?

[EgorBo]

@SamMonoRT @naricc #40069 wasn't merged so that must be something else I guess.

[SamMonoRT]

Let's keep an eye on the results of this benchmark in next 3-4 days, and then close if we determine it is improved. @EgorBo do you think you changes might help make this even more performant ?

[tannergooding]

@naricc, is there any update here. Is this something that is going to be worked on for 6.0.0?

If not, I'd like to either move this to future or close, as appropriate.

[naricc]

@tannergooding This is not being actively worked on for 6.0.0; I am having some issues verifying weather or not is resolved right now. I will move it to future, and then close it when I am able to get the data (if merited).

[ilonatommy]

@naricc is there anything new about this? Are we able to ship it with 7.0?

[naricc]

@ilonatommy t looks like this is still happening.

wasm-interpreter: https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04_CompliationMode%3dwasm_RunKind%3dmicro%2fSystem.Numerics.Tests.Perf_BitOperations.PopCount_ulong.html

mono desktop interpreter:
https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04_LLVM%3dfalse_MonoAOT%3dfalse_MonoInterpreter%3dtrue_RunKind%3dmicro_mono%2fSystem.Numerics.Tests.Perf_BitOperations.PopCount_ulong.html

This doesn't block anything from shipping; it's not clear there is actually anything that can be done about it, although I haven't investigated enough to confirm that.

[lewing]

@SamMonoRT lets triage this properly when 8 starts

[lewing]

@kotlarmilos is there a status update on this?

[kotlarmilos]

The BitOperations have been intrinsified on both platforms in #82827. The microbenchmark System.Numerics.Tests.Perf_BitOperations.PopCount_ulongis is about 60% slower on the Wasm than on Interpreter.

Is it expected to have decreased performance when running on Wasm?

[lewing]

WasmBase was being incorrectly disabled and that was fixed in #88523 lets make sure we have accurate numbers for this now

[radekdoulik]

@kotlarmilos what is the configuration with 464% perf. Is it aot or interpreter on wasm?

edit: or actually these 60% you measured.

[lewing]

@kg is there a jiterpreter intrinsic for i64.popcnt ?

[lewing]

@kg is there a jiterpreter intrinsic for i64.popcnt ?

I see there is

[kotlarmilos]

@kotlarmilos what is the configuration with 464% perf. Is it aot or interpreter on wasm?

@radekdoulik Sorry for the late reply, it is wasm interpreter.

[lewing]

We looked into this and consider it closed