[ARM]: Implement CPU plugin just-in-time emitter for Erf operation #27499

[shivam5522]

Details:

• Added a jit_erf_emitter derived class in aarch64/jit_eltwise_emitters
• Created entry Algorithm::EltwiseErf in the get_supported_precisions in nodes/kernels/aarch64
• Add the EltwiseErf entry in the aarch64 executors supported algorithms

Tickets:

#27499

[a-sidorova]

I saw your comment with asking help about erf impl (but seems like it was deleted).

You can take a look on JIT Emitter for GeluErf for AArch64 platforms. There is Erf impl

Also, we have JIT Emitter for Erf for x64 platforms. Please take a look - it might be useful.

[shivam5522]

Yeah I checked out the ERF approximation and also the ERF documentation in general. I assume that we have to use the approximation formula to implement ERF. Also can I directly use the same idea from the JIT Emitter for ERF for x64 and then port accordingly for AArch64?
Also, I work on a windows laptop, are there any steps to test this on windows? Sorry for the dumb doubts

[a-sidorova]

Also can I directly use the same idea from the JIT Emitter for ERF for x64 and then port accordingly for AArch64?

For sure! It makes sense 😃

Also, I work on a windows laptop, are there any steps to test this on windows? Sorry for the dumb doubts

I'd assume that your laptop is on x64 arch (intel or amd). In this case you need emulator to launch programs for arm64. But to be honest, I don't know any good options to do it on windows.
I know only one option with WSL - there you can use cross-compilation and launch programs with QEMU.

Sorry for the dumb doubts

No worries, everything is ok! 👍🏼

[shivam5522]

@a-sidorova Could take a look at it once? Thanks

[a-sidorova]

Also could you please update test instances? Please take a look at the last paragraph in the section "What needs to be done?" in the ticket #27499.

Then please locally launch tests using the section "Tests".

[a-sidorova]

If the method impl of derived class is the same as of base class and this method is virtual in base class - no need to override this method.

However, you use exp_emitter in the impl of erf_emitter. So you need to emit data for exp_emitter in this method too:

void jit_erf_emitter::emit_data() const {
    jit_emitter::emit_data();
    exp_emitter->emit_data();
}

Otherwise, exp_emitter cannot find the constant values in the table.

I believe that if we add exp_emitter->emit_data();, failed GHA jobs with the following error messages will be fixed:

[ RUN      ] smoke_Activation_Basic/ActivationLayerTest.Inference/IS=([])_TS={(1.50)}_TS=()_Erf_constants_value=()_netPRC=f32_trgDev=CPU
bad err=4 in Xbyak::Error

MEM_USAGE=1972392KB
src/tests/functional/shared_test_classes/src/base/ov_subgraph.cpp:97: Failure
Exception from src/inference/src/cpp/core.cpp:109:
Exception from src/inference/src/dev/plugin.cpp:53:
label is not found


[  FAILED  ] smoke_Activation_Basic/ActivationLayerTest.Inference/IS=([])_TS={(1.50)}_TS=()_Erf_constants_value=()_netPRC=f32_trgDev=CPU, where GetParam() = ((26, {}), f32, ({ ({}, { { 1, 50 } }) }, {}), "CPU") (46 ms)

[shivam5522]

Sure adding the tests right now

[shivam5522]

@a-sidorova Just to be sure, I need to update the SetUp() to use the ERF operation instead of the current GELU ERF operation right?

[a-sidorova]

@a-sidorova Just to be sure, I need to update the SetUp() to use the ERF operation instead of the current GELU ERF operation right?

You need to add a Erf as a new operation for testing. GeluErf is another operation and should be validated too. So please add Erf in activationTypes() list.

[shivam5522]

Other than that, could you review the changes once and see if they are correct?

[a-sidorova]

Other than that, could you review the changes once and see if they are correct?

I suggest to locally validate the implementation launching tests for Erf.
And I think there will be accuracy problems. Because previously we discussed that you can port JIT emitter impl from x64 to aarch64. However, I don't see polynomial calculation as it's done in x64 impl. Please pay attention on this part. Probably, this article also will help you to implement the emitter algorithm 😊

[shivam5522]

Sure sorry about that, are there any docs that I can follow for testing, mainly I cannot understand the activation.cpp and how to test my impl? So mainly any docs that I can follow to facilitate testing?
Thanks.

[a-sidorova]

Sure sorry about that, are there any docs that I can follow for testing, mainly I cannot understand the activation.cpp and how to test my impl? So mainly any docs that I can follow to facilitate testing? Thanks.

To implement tests, you need to take a look at this part of the issue description.

This paragraph contains links to the code parts which should be updated

How to test:

These paragraphs should be enough to test the implementation. If you have specific questions, feel free to ask them 😊

Also there is PR which added tests for jit_ceiling_emitter : #27527. It also might be useful for you.

[a-sidorova]

@shivam5522 Hello! How are things going with implementation and testing? Could you tell us please if you need our help? 😊

[github-actions]

This PR will be closed in a week because of 2 weeks of no activity.