3. Numba for CUDA GPUs

• 3.1. Overview
  • 3.1.1. Terminology
  • 3.1.2. Programming model
  • 3.1.3. Requirements
    • 3.1.3.1. Supported GPUs
    • 3.1.3.2. Software
  • 3.1.4. Missing CUDA Features
• 3.2. Writing CUDA Kernels
  • 3.2.1. Introduction
  • 3.2.2. Kernel declaration
  • 3.2.3. Kernel invocation
    • 3.2.3.1. Choosing the block size
    • 3.2.3.2. Multi-dimensional blocks and grids
  • 3.2.4. Thread positioning
    • 3.2.4.1. Absolute positions
    • 3.2.4.2. Further Reading
• 3.3. Memory management
  • 3.3.1. Data transfer
    • 3.3.1.1. Device arrays
  • 3.3.2. Pinned memory
  • 3.3.3. Streams
  • 3.3.4. Shared memory and thread synchronization
  • 3.3.5. Local memory
  • 3.3.6. SmartArrays (experimental)
  • 3.3.7. Deallocation Behavior
• 3.4. Writing Device Functions
• 3.5. Supported Python features in CUDA Python
  • 3.5.1. Language
    • 3.5.1.1. Execution Model
    • 3.5.1.2. Constructs
  • 3.5.2. Built-in types
  • 3.5.3. Built-in functions
  • 3.5.4. Standard library modules
    • 3.5.4.1. cmath
    • 3.5.4.2. math
    • 3.5.4.3. operator
• 3.6. Supported Atomic Operations
  • 3.6.1. Example
• 3.7. Random Number Generation
  • 3.7.1. Example
• 3.8. Device management
  • 3.8.1. Device Selection
• 3.9. The Device List
• 3.10. Examples
  • 3.10.1. Matrix multiplication
• 3.11. Debugging CUDA Python with the the CUDA Simulator
  • 3.11.1. Using the simulator
  • 3.11.2. Supported features
• 3.12. GPU Reduction
  • 3.12.1. @reduce
  • 3.12.2. class Reduce
• 3.13. CUDA Ufuncs and Generalized Ufuncs
  • 3.13.1. Example: Basic Example
  • 3.13.2. Example: Calling Device Functions
  • 3.13.3. Generalized CUDA ufuncs
• 3.14. Sharing CUDA Memory
  • 3.14.1. Sharing between process
    • 3.14.1.1. Export device array to another process
    • 3.14.1.2. Import IPC memory from another process
• 3.15. CUDA Frequently Asked Questions
  • 3.15.1. nvprof reports “No kernels were profiled”