.take

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

Hi @a-sidorova, I ran the tests and as expected the tests failed. I'm failing to understand the tests completely.

Here is one of the tests for reference:

smoke_Snippets_FQDecomposition_Scalars/FakeQuantizeDecompositionTest.CompareWithRefImpl/IS=(1.3.16.16)_netPRC=f32_D=CPU_IN=f32_OP=Abs_opset1_ON1=Subgraph_ON1=Abs,fakeQuantize_LP=1SH1=[]SH2=[]SH3=[]SH4=[] src/tests/functional/shared_test_classes/src/base/snippets_test_utils.cpp:52: Failure Expected equality of these values: originalLayersNames Which is: "Abs,fakeQuantize" name Which is: "Abs" src/tests/functional/shared_test_classes/src/base/snippets_test_utils.cpp:37: Failure Expected equality of these values: ref_num_nodes Which is: 4 num_nodes Which is: 3 Compiled model contains invalid number of nodes.

• What I understand, from seeing the log, the fakeQuantize node is being decomposed but not fused with the subgraph. I am also confused why the num_nodes here are 3. While only the "Abs" is recognized and others isn't. So, I would think that one node is for "Abs", while about the other two, is not logged (I would guess they are Maximum and Minimum nodes, from the decomposition pass).
• As far as supporting missing operations is concerned, on inspecting the decomposition pass here, I found all the operations (Maximum, Minimum, Add, Subtract, Multiply, Round, ConvertSaturation and Divide) already have their corresponding jitters in cpu_generator. I'm not sure which operation is missing during code generation phase.
• So, if all the operations already have their corresponding jitters, if i include the ov::is_type<ov::op::v0::FakeQuantize>(n) in is_supported_op, in the transformation pipeline, I thought that it should work but it doesn't.

I'm sure I am missing something, please guide me through this. Thanks.

@srinjoydutta03 thank you for the questions!

As far as supporting missing operations is concerned, on inspecting the decomposition pass here, I found all the operations (Maximum, Minimum, Add, Subtract, Multiply, Round, ConvertSaturation and Divide) already have their corresponding jitters in cpu_generator. I'm not sure which operation is missing during code generation phase

You're absolutely right - all emitters are already implemented. If some emitter is missed, you will see the following exception from target machine/generator: Check 'jitter != jitters.end()' failed at src/common/snippets/src/lowered/target_machine.cpp:19: Supported precisions set is not available for X operation.. It means that CPU Generator knows nothing about X operation and doesn't know how to compile the code/which precisions are supported. If you don't see this exceptions, CPU Generator supports everything.

Also let me help you with logs of tests:

src/tests/functional/shared_test_classes/src/base/snippets_test_utils.cpp:52: Failure Expected equality of these values: originalLayersNames Which is: "Abs,fakeQuantize" name Which is: "Abs"

It means, the test expects that the execution model (state after ov::Model compilation) will contain Subgraph op which contains Abs and FakeQuantize ops (what was originally tokenized by Snippets). The test shows that currently Subgraph has only Abs and FakeQuantize has not been tokenized. This test part should be fixed by your adding ov::is_type<ov::op::v0::FakeQuantize>(n) in is_supported_op, in the transformation pipeline as you already did 😃

src/tests/functional/shared_test_classes/src/base/snippets_test_utils.cpp:37: Failure Expected equality of these values: ref_num_nodes Which is: 4 num_nodes Which is: 3 Compiled model contains invalid number of nodes.

I believe that this check is more for x64 where we support blocked layouts in CPU Plugin. Please see brief comment. Since MaxPool forces blocked shapes on x64 platforms, x64 platform expect reorders around (we use these reorder ops to change layouts of tensors) -> expected node count is 4 (reorder + MaxPool + reorder + Subgraph). AArch64 forces nothing so reorders ops are missed and expected node count is another. But I've just found that we inserted MaxPool on inputs of model due to old and outdated limitations in Snippets which was removed for a long time ago. So I suggest to set empty vector here instead of vector with MaxPool. Then the tests should expect that after model compilation there will be only one op - Subgraph. So after that you need to replace first numbers in pairs with 1. I believe that after that - these tests will be green on the both platforms: x64 and aarch64.

By the way, I've found the tests legacyFuse here. They are relevant only for x64 platforms: on x64 platforms Conv op can fuse FakeQuantize op on output for better performance. This is not supported on aarc64. So I suggest to hide this whole namespace (with test instance) into #ifdef OPENVINO_ARCH_X86_64 ... #endif since they're needed and valid only on x64 platforms.

If you have more questions, feel free to ask them! 😊

Thank you so much for the help :).

I would think for other tests too, the per_channel and per_channel_inputs as well I should set the first parameter to 1 since both these tests also use Reorder and MaxPool ops.

I have enclosed the INSTANTIATE_TEST_SUITE_P under legacyFuse namespace with #ifdef and #endif directives.

On doing so the tests run successfully now, with 6 tests skipped for 16bit floating point precisions.

@srinjoydutta03 thank you for the status sharing! Now we're waiting for the PR from you! 😊

@srinjoydutta03 as for our discussion about the next tasks which can be interesting for you.

At the moment, we have the following ARM-related tasks which already have assignee.. But she has a lot of taken issues without any activity in these tasks:

• [Good First Issue] [ARM]: Implement CPU plugin just-in-time emitter for SoftPlus operation #24109
• [Good First Issue] [ARM] [Debug Caps] : Implement register prints in just-in-time Kernels #27715

Just leave please the comment in the interesting for you issue. I will reassign to you! 😊