[onert] Introduce DepthwiseConvOp in cpu kernel

[jyoungyun]

This commit introduces DepthwiseConvOp in cpu kernel.
This kernel uses depthwise_conv and bias_op eigen functions.

ONE-DCO-1.0-Signed-off-by: Jiyoung Yun [email protected]

[hseok-oh]

In file included from /home/runner/work/ONE/ONE/runtime/onert/backend/cpu/ops/DepthwiseConvolutionLayer.cc:20:
In file included from /home/runner/work/ONE/ONE/compute/cker/include/cker/operation/DepthwiseConv.h:32:
/home/runner/work/ONE/ONE/compute/cker/include/cker/eigen/bias_op.h:100:7: error: unused variable 'channel_dim' [-Werror,-Wunused-variable]
  int channel_dim = input_shape.DimensionsCount() - 1;
      ^
1 error generated.
gmake[3]: *** [runtime/onert/backend/cpu/CMakeFiles/onert_backend_cpu.dir/build.make:272: runtime/onert/backend/cpu/CMakeFiles/onert_backend_cpu.dir/ops/DepthwiseConvolutionLayer.cc.o] Error 1

[jyoungyun]

Need #13582

[jyoungyun]

@ys44kim

there is no code that access bias_data, and it is passed as argument , there is no need to check for nullptr.
of course, even when there is no explicit nullptr input, sometimes the static analysis tool judges it to be an error when accessing the variable without checking for nullptr. of course, this does not apply in this case.
so, i think it is a good choice.

The code was changed to support optional bias data in DepthwiseConv2D cpu kernel. :)

i think it would be better to add a exception action for the case where bias_data is nullptr.
for example, add assert (bias_data) as follows refer to biasHelper,

In ONERT, the DepthwiseConv2D always has bias data. However, this behaviour can be changed because the DepthwiseConv2D does not necessarily require bias data.
So, I want to support optional bias in DepthwiseConv2D cpu kernel.
I added a test case for optional bias test case for verifying this cpu kernel.

Please take another look. :)

[ys44kim]

@jyoungyun
i checked the commit history and review comment again and there was something I didn't understand well.
i will check more about the cpu kernel and review whether additional improvements are needed.
thank you for your kind reply.

[jyoungyun]

@ys44kim

In onert the DepthwiseConv2D always has bias data (please tell me if this is wrong.)

Yes, you're right.

and, In fact, I don't understand why it is tested when bias is nullptr because onert does not support optional bias.

The codes under compute/cker directory is implemented using cpu for each operation.
ONERT uses various kinds of backends. Currently, ONERT takes cpu, acl_cl, neon, trix(npu) as backend. The behavior of each backend may or may not match ONERT.

In this case, ONERT does not support optional bias_data. However, CPU kernel that I imlemented on this PR supports optional bias_data because I think it is better to support more diverse functions in the cpu kernel. If ONERT supports optional bias in the future, we can support it without modifying this kernel.

Therefore, this cpu kernel has decided to support optional bias even though ONERT does not support it. :)

[ragmani]

LGTM

[ys44kim]

LGTM

[jyoungyun]

@Samsung/one_onert Please merge this PR. :)