Update ideep submodule (#97430)

### Description

This PR is to update ideep submodule for the following two aspects:

1. At inductor side, we are supporting dynamic shape path for packed linear, which we hopes the packed weight of linear doesn't depend on the input shapes and still can get a better a performance using a packed weight got from a dummy input shapes. However the current ideep has a accuracy issue for this case. This updating will fix the issue. 
2. Add an extra arg is_channels_last for deconv to notify ideep whether to go channels last or not because the memory format checks of ideep (e.g. is_nhwc(), is_ndhwc()) is not 100% identical to suggest_memory_format() from pytorch.

### Performance Benchmark

Use TorchBench test in ICX with 40 cores
Intel OpenMP & tcmalloc were preloaded
![image](https://user-images.githubusercontent.com/61222868/229072474-193513ba-6727-4451-91ff-0d57e016736f.png)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/97430
Approved by: https://github.com/jgong5
This commit is contained in:
yanbing-j 2023-04-02 06:42:09 +00:00 committed by PyTorch MergeBot
parent 347c67d4a2
commit 2ab18a23e1

2
third_party/ideep vendored

@ -1 +1 @@
Subproject commit 7bc3e12f7c0cad7fb24f8d4ab63dcd467ffa60c7
Subproject commit fe8378249600442043b98f333b8b605bedca5a25