[microNPU] Add hardware constraints for binary elementwise

[Aleksei-grovety]

Does not fuse min and max operations with requantize if there are different scales as it is not supported on NPU. Since there are hardware constraints, we cannot perform min or max operation fused with requantize (please look at NPU_SET_OFM_SCALE register description https://developer.arm.com/documentation/102420/0200/Programmers-model/Command-stream/cmd1-commands-) when we have different scales.
min/max operations with matching scales are offloaded to NPU as ethosu_binary_elementwise
min/max operations with different scales are offloaded to NPU as ethosu_binary_elementwise + ethosu_identity.

cc @leandron, @ekalda, @lhutton1

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @Mousius, @leandron, @lhutton1 _{See #10317 for details}

_{Generated by tvm-bot}

[ekalda]

Thanks @alexey-yazev, looks good in general, some comments about the legalization tests :)

[ekalda]

Suggested change

	@pytest.mark.parametrize("activation_function", [None, tf.nn.relu, tf.nn.relu6, relu_n1_to_1])
	@pytest.mark.parametrize("activation_function", [None, tf.nn.relu])

Since this test looks at the graph entirely at Relay level, where all the RELUs are just Relay clip operations, I don't think we benefit much from extra 70 tests with just different min and max attributes to clip.

[Aleksei-grovety]

I agree with this, I will add separate test for case with MAX operation and relu_n1_to_1 activation.

[ekalda]

Do both of these blocks get run? It looks like we are using the same method of generating representative dataset (which will determine the qnn params) for all the tests, so I suspect we will always create IFMs with differing qnn params and therefore test only one of the patterns here.

[Aleksei-grovety]

Yes, both of these blocks get run, the first block is run for cases with MAX operation and relu_n1_to_1 activation for example test_tflite_binary_elemwise_legalize[relu_n1_to_1-ifm_shape0-ifm2_shape0-False-MAX]

fn (%tvmgen_default_ethos_u_main_0_ifms: Tensor[(48), int8] /* ty=Tensor[(48), int8] */, Inline=1, Compiler="ethos-u", global_symbol="tvmgen_default_ethos_u_main_0", Primitive=1) -> Tensor[(1, 2, 3, 4), int8] {
  %0 = split(%tvmgen_default_ethos_u_main_0_ifms, indices_or_sections=[24]) /* ty=(Tensor[(24), int8], Tensor[(24), int8]) */;
  %1 = %0.0 /* ty=Tensor[(24), int8] */;
  %2 = %0.1 /* ty=Tensor[(24), int8] */;
  %3 = reshape(%1, newshape=[1, 2, 3, 4]) /* ty=Tensor[(1, 2, 3, 4), int8] */;
  %4 = reshape(%2, newshape=[1, 2, 3, 4]) /* ty=Tensor[(1, 2, 3, 4), int8] */;
  %5 = contrib.ethosu.binary_elementwise(%3, %4, meta[relay.Constant][0] /* ty=Tensor[(0), int8] */, operator_type="MAX", ifm_scale=1f, ifm_zero_point=0, ifm2_scale=1f, ifm2_zero_point=0, ofm_scale=1f, ofm_zero_point=0, ifm_channels=4, ifm2_channels=4, activation="CLIP", clip_min=-128, ofm_dtype="int8") /* ty=Tensor[(1, 2, 3, 4), int8] */;
  contrib.ethosu.identity(%5, meta[relay.Constant][1] /* ty=Tensor[(0), int8] */, ifm_scale=0.00783747f, ifm_zero_point=-128, ofm_scale=0.00392157f, ofm_zero_point=-128) /* ty=Tensor[(1, 2, 3, 4), int8] */
} /* ty=fn (Tensor[(48), int8]) -> Tensor[(1, 2, 3, 4), int8] */

in this cases the scales are different in others the same.

[ekalda]

Ok cool, thanks for clarifying :)

[ekalda]

Thanks @alexey-yazev, LGTM!