Add a Learning Path for integrating a KleidiAI SME2 kernel into XNNPACK

[geesun]

Summary

This PR adds an advanced Learning Path that demonstrates how to integrate a KleidiAI SME2 matrix multiplication microkernel into XNNPACK's qd8_f16_qc4w fully connected operator.

What is included

• An overview of the XNNPACK operator lifecycle and its QD8, QC4W, and FP16 data formats
• Guidance for selecting a KleidiAI microkernel based on its complete quantization contract
• Instructions for packing static QC4W weights into the KleidiAI RHS layout
• Instructions for packing dynamic QD8 activations without requantization
• SME2 runtime dispatch with the existing XNNPACK microkernels retained as a fallback
• Support for normal and transposed QC4W weight layouts
• Android NDK build and on-device correctness validation steps
• A four-part patch series that readers can apply to the documented XNNPACK baseline revision

Key learning outcomes

After completing this Learning Path, readers will understand how to:

• Match framework tensor formats to a compatible KleidiAI microkernel
• Preserve per-row QD8 quantization parameters
• Handle K-dimension padding correctly
• Integrate KleidiAI packing and execution into XNNPACK
• Select the SME2 path at runtime without breaking non-SME2 configurations
• Build and test the integration on an SME2-capable Android device

Validation

The Learning Path documents validation using:

• XNNPACK's FULLY_CONNECTED_NC_QD8_F16_QC4W correctness tests on an SME2-capable Android device
• A fallback build with KleidiAI disabled

Submission checklist

Before submitting a pull request for a new Learning Path, please review Create a Learning Path.

• [ x] I have reviewed Create a Learning Path

Please do not include any confidential information in your contribution. This includes confidential microarchitecture details and unannounced product information.

• [ x] I have checked my contribution for confidential information

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution under the terms of the Creative Commons Attribution 4.0 International License.