Contents

• 1. CUDA for Tegra
• 2. Overview
• 3. Memory Management
  • 3.1. I/O Coherency
  • 3.2. Sysmem Full Coherency
  • 3.3. Estimating Total Allocatable Device Memory on an Integrated GPU Device
• 4. Porting Considerations
  • 4.1. Memory Selection
  • 4.2. Pinned Memory Guidelines on Tegra Systems with no I/O coherency
  • 4.3. Effective Usage of Unified Memory on Tegra
    • 4.3.1. Unified Memory where cudaDeviceProp::concurrentManagedAccess is 0 and only 1 iGPU device
    • 4.3.2. Unified memory where cudaDeviceProp::concurrentManagedAccess is 1
  • 4.4. GPU Selection
  • 4.5. Synchronization Mechanism Selection
  • 4.6. Multi-Process Service (MPS)
  • 4.7. CUDA Features Not Supported on Tegra
  • 4.8. GPUDirect RDMA
  • 4.9. Additional Reads
• 5. EGL Interoperability
  • 5.1. EGLStream
    • 5.1.1. EGLStream Flow
    • 5.1.2. CUDA as Producer
    • 5.1.3. CUDA as Consumer
    • 5.1.4. Implicit Synchronization
    • 5.1.5. Data Transfer Between Producer and Consumer
    • 5.1.6. EGLStream Pipeline
  • 5.2. EGLImage
    • 5.2.1. CUDA interop with EGLImage
  • 5.3. EGLSync
    • 5.3.1. CUDA Interop with EGLSync
    • 5.3.2. Creating EGLSync from a CUDA Event
    • 5.3.3. Creating a CUDA Event from EGLSync
• 6. CUDA Upgradable Package for Jetson
  • 6.1. Installing the CUDA Upgrade Package
    • 6.1.1. Prerequisite
    • 6.1.2. From Network Repositories or Local Installers
  • 6.2. Deployment Considerations for CUDA Upgrade Package
    • 6.2.1. Use the Right Upgrade Package
    • 6.2.2. Feature Exceptions
    • 6.2.3. Check for Compatibility Support
• 7. CUDA Instrumentation Methods
  • 7.1. Prerequisites
  • 7.2. GPU Task Tracker
    • 7.2.1. Introduction
    • 7.2.2. When to use?
    • 7.2.3. Procedure
• 8. CUDA Upgrade Package for Jetson
  • 8.1. Installing the CUDA Upgrade Package
    • 8.1.1. Prerequisite
    • 8.1.2. From Network Repositories or Local Installers
    • 8.1.3. CUDA Upgrade Package Files
    • 8.1.4. Example
  • 8.2. Deployment Considerations for the CUDA Upgrade Package
    • 8.2.1. Use the Right Upgrade Package
    • 8.2.2. Feature Exceptions
    • 8.2.3. Check for Compatibility Support
• 9. cuDLA
  • 9.1. Developer Guide
    • 9.1.1. Device Model
    • 9.1.2. Loading and Querying Modules
    • 9.1.3. Memory Model
    • 9.1.4. Task Execution and Synchronization Model
      • 9.1.4.1. Task Execution
        • 9.1.4.1.1. Multithreaded User Submission
      • 9.1.4.2. Synchronization
        • 9.1.4.2.1. Registering an external semaphore:
        • 9.1.4.2.2. Events setup for cudlaSubmitTask()
        • 9.1.4.2.3. Waiting on the signal event
        • 9.1.4.2.4. Supported Synchronization Primitives in cuDLA
        • 9.1.4.2.5. Setting NvSciSyncAttrKey_RequireDeterministicFences key in NvSciSyncAttrList
        • 9.1.4.2.6. Timestamp Support for NvSciFence
        • 9.1.4.2.7. Requesting Timestamp Support for NvSciSync Object
        • 9.1.4.2.8. Extracting Timestamp Value from Fence
      • 9.1.4.3. Fault Diagnostics
      • 9.1.4.4. NOOP Submission
    • 9.1.5. Error Reporting Model
  • 9.2. Migrating from NvMediaDla to cuDLA
  • 9.3. Profiling a cuDLA App
  • 9.4. cuDLA Release Notes
• 10. CUDA integration with other SoC engines
• 11. CUDA C++ Coding Guidelines
  • 11.1. Version 3.0.1
    • 11.1.1. Guidelines for Effective Safety-Critical Programming
• 12. Notices
  • 12.1. Notice
  • 12.2. OpenCL
  • 12.3. Trademarks