zero copy permute opt for conv

[AshutoshSinghIntel]

Summary:

• This change targets optimization for permute->convolution subgraph typical in vision transformers. By aligning the Permute output to the Convolution's Channel-Last format, we achieve two performance wins:

1. Elimination of Reorder Nodes: Removing expensive data copy operations.
2. Zero-Copy "Virtual Transpose": Converting physical data movement into a logical layout cast

Implementation Details:

• reorder is getting added for changing planar to blocked fmt for conv
• existing perm->conv optimization optimize_out the permute operation and leaving the physical data in byxf. Since Input Physical == Output Physical, the Permute becomes a Zero-Copy Layout Cast.
• enforcing byxf format for conv eliminates the reorder node and improves performance
• for more details, kindly refer CVS-177918

Graph:

Performance Improvement:

Test details:

• EBGAN TF FP32 showed ~50% degradation, need to rerun to make sure that there was no temporary issue with machine during the run. EBGAN TF FP16 did not show such significant degradation.
• Geomean for rest of the models is fine.
• Full validation benchmark report is here

Tickets:

• CVS-177918

[AshutoshSinghIntel]

build_jenkins

[p-durandin]

Please check that node has only one user to avoid breaking data for other nodes

[AshutoshSinghIntel]

Hi @p-durandin,
in this case, dependency node, pnode (permute), has two users: permute -> conv, permute -> permute
while node var which is conv node has one user: conv -> permute

For which node do you want to add the check?