Changes to MklConvFwdPrimitiveFactory to support Arm Compute Library backend

[alenik01]

System information

• TensorFlow version (you are using): 2.4.1
• Are you willing to contribute it (Yes/No): Yes

Describe the feature and the current behavior/state.
At the moment MklConvFwdPrimitiveFactory reuses already created oneDNN primitives, stored in the vector fwd_primitives, with the help of GetConvFwd function which creates a key from convolution parameters stored in MklConvFwdParams and compares this key with the keys of existing primitives. However, when oneDNN is built with Arm Compute Library (ACL, see also oneDNN build options) there is specific requirement to create a separate primitive per constant weights tensor, this is a restriction implied by ACL code design which does not allow to use primitive caching mechanism in MklConvFwdPrimitiveFactory as it is. The solution may be to add a new entry, void* filter_address, to MklConvFwdParams struct and to include the address of weights to the key in CreateKey function:

...
key_creator.AddAsKey(convFwdDims.filter_dims);
#ifdef DNNL_AARCH64_USE_ACL
key_creator.AddAsKey(convFwdDims.filter_address);
#endif
key_creator.AddAsKey(convFwdDims.bias_dims);
...

Will this change the current api? How?
The only API change is the addition of a new field to CreateKey and MklConvFwdParams in mkl_conv_ops.cc, this is related to the limitations of ACL which is designed to create a separate primitive per weights tensor.

Who will benefit with this feature?
The users who run inference regime with TensorFlow on AArch64-based machines.

Any Other info.
This Issue is raised mainly to get feedback from maintainers of the oneDNN to TensorFlow integration.

[alenik01]

Please have a look, @agramesh1

[agramesh1]

Please have a look, @agramesh1

Adding @gzmkl from Intel-TF team who can help.

[alenik01]

Closed due to the raised PR.