Qualcomm AI Stack SDK - Patched Version

This SDK has been patched to fix critical memory corruption issues in the model converter tools, specifically when handling complex graph transformations and shape inference.

Key Fixes & Patches

1. Python-based Shape Re-inference

• Core Fix: Replaced buggy C++ shape inference logic with a robust Python implementation in qti.aisw.converters.qnn_backend.qnn_backend_base.reinfer_graph_shapes.
• Topological Pass: Added a comprehensive graph traversal that re-computes all buffer shapes using Python logic to ensure consistency across the entire network.

2. Memory Corruption Prevention

• Python Caching: Dimensons, shapes, and operation parameters are now cached in Python objects to avoid direct calls to unstable C++ library functions which were causing segmentation faults and data corruption.
• Safe Property Accessors: Overrode __setattr__ and added Python-backed property accessors for critical metadata like perm, transpose, and dims.

3. Op-Specific Robustness Fixes

• Conv2D: Improved shape inference with better NCHW/NHWC layout detection and support.
• MatMul: Added auto-detection and fixing of transpose flags when inner dimensions do not match, preventing conversion failures.
• BatchNorm & Element-wise Ops: Ensured output shapes strictly match input shapes to prevent stale dimensions from propagating.
• Reshape & Transpose: Fixed NCHW to NHWC transformation logic and perm parameter handling.

4. Stability Improvements

• Input/Output Buffers: Removed direct C++ calls to inputs() and outputs() on certain nodes to avoid memory corruption during graph optimization passes.

Contents

• bin/: Executables and setup scripts.
• docs/: Original documentation for QNN, SNPE, and Genie.
• examples/: Sample code and models.
• include/: C++ header files.
• lib/: Shared libraries and Python modules (Patched).
• share/: Additional resources and templates.

Setup

Refer to bin/envsetup.ps1 (Windows) or bin/envsetup.sh (Linux) to set up your environment variables. Individual READMEs for QNN and Genie are available in the root directory:

• QNN_README.txt
• GENIE_README.txt

Support the Project

If these patches saved you hours of debugging or prevented a system crash, consider supporting the continued maintenance of this repository.

PayPal Account: drifterself01@gmail.com