SPRADS4 October   2025 TDA4VM

 

  1.   1
  2.   Abstract
  3.   Trademarks
  4. 1Introduction
  5. 2Instructions
    1. 2.1 Scope
    2. 2.2 HWA Instruction Definition
    3. 2.3 Fault Injection Procedure
      1. 2.3.1 Block Diagram
      2. 2.3.2 Fault Injection: Steps
        1. 2.3.2.1 Underflow Error
        2. 2.3.2.2 Overflow Error
        3. 2.3.2.3 Offset Parity Error
        4. 2.3.2.4 Config Parity Error
          1. 2.3.2.4.1 A FSM Config Parity Error
          2. 2.3.2.4.2 B FSM Config Parity Error
          3. 2.3.2.4.3 C FSM Config Parity Error
          4. 2.3.2.4.4 X FSM Config Parity Error
        5. 2.3.2.5 C READ Error
        6. 2.3.2.6 C Write Error
  6. 3Flow Diagram
    1. 3.1 Code Changes
      1. 3.1.1 Return Hook Definition
      2. 3.1.2 Clear MMA Function
      3. 3.1.3 Sequence Test
  7. 4Summary
  8. 5References

1 Introduction

The Matrix Multiply Accelerator (MMA) is a tightly coupled matrix multiplication coprocessor that extends the scalar and vector facilities of the C7x processor architecture. Fed through the streaming engines (SEs), the MMA provides a large number of multiply-accumulate (MAC) operations, matrix-optimized storage, and matrix-optimized data movement that can be efficiently used for the dense linear algebra operations that dominate the computations found in vision (CNNs, SfM, filtering, and so forth), speech (xNNs), audio (convolutions), radar (FFTs) and controls (reinforcement learning, state updates) based applications.

This document covers the sequential process of fault injection, specifically focusing on Matrix Multiplier Accelerator (MMA)diagnostics.

This document explains about how these tests can be conducted sequentially, verifying a detailed understanding of fault handling and recovery.

This document does not discuss the recovery of the exception to execute the regular flow of the code after the sequence of error injection is done. Please refer to SDK 11.2 for integration example.