Convolution FP32 in oneDNN Changed from gemm:acl to gemm:ref in TensorFlow 2.16

[rpushkarr]

In TensorFlow version 2.16 (and later), the convolution operation implementation for FP32 data type in oneDNN was changed from gemm:acl to gemm:ref. However, this change has resulted in performance degradation compared to TensorFlow version 2.15, where gemm:acl was used.

System Information:

• TensorFlow Version: 2.16
• Previous Working Version: 2.15
• oneDNN Version: 3.2.1
• Hardware: Aarch64
• Operating System: Ubuntu 22.04

Issue Summary:

• In TensorFlow 2.15, the convolution operation for FP32 data type was routed through gemm:acl in oneDNN, which provided better performance.
• In TensorFlow 2.16 (and later), the implementation was changed to use gemm:ref, leading to a noticeable performance drop.