cpu: aarch64: matmul: Move allocation of temporary tensors to scratchpad in acl_matmul

[annop-w]

Description

Introduce 3 new scrathpad memory key names.

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

[mgouicem]

tagging @snadampal as this relates to #1470

[vpirogov]

+@snadampal, could you please help reviewing these changes?

@annop-w, please resolve merge conflict.

[annop-w]

@annop-w, please resolve merge conflict.

Done.

[snadampal]

glad to see this change finally coming.
Hi @annop-w to understand the change a bit in detail, the scratchpad buffers for src and weights are the same buffers allocated in the framework, right . for example, by ideep in PyTorch and by mkldnn wrapper in TensorFlow?

thanks @mgouicem and @vpirogov for tagging me here.

[annop-w]

@snadampal I am not sure how the scratchpads are currently managed in ideep (or Tensorflow), but the idea here is to allow for users (i.e. PyTorch or TF) a chance to decide that, isn't it ? For example, PyTorch can now choose to allocate the same buffer for both src and wei, if sensible, which was not possible before. Does this help ?

[snadampal]

i mean reusing the buffer allocated in PT or TF via oneDNN user mode scratchpad, you clarified it, thanks.
this change LGTM.

[annop-w]

@snadampal Ah, yes, in that case, you're absolutely right. Thanks for the review.

[vpirogov]

Awesome. Thanks, @snadampal.