Petrobras CUDA Training

DAY 1: Fundamental CUDA Optimization

1. Latency Hiding

   • Thread and warp scheduling
   • Launch configuration

2. Memory Hierarchy and Access Patterns

   • Local storage, shared memory, and global memory

3. Understanding Bottlenecks

   • Memory-bound vs compute-bound codes
   • Extracting bandwidth
   • Memory alignment

4. Assignment

DAY 2: Atomics, Reductions, Warp Shuffle

1. Atomics and Reductions

   • Atomic operations
   • Classical parallel reduction
   • Parallel reduction + atomics

2. Warp Shuffle Techniques

   • Warp-Level reduction
   • Reduction with warp shuffle

3. Assignment

DAY 3: CUDA Concurrency

1. Pinned Memory

   • Definition and benefits
   • Usage and functions
   • Implications for host memory

2. CUDA Streams

   • Overview, purpose, and semantics
   • Creation, Usage, and copy-compute overlap
   • Stream behavior examples and default stream considerations

3. Multi-GPU Concurrency

   • Device management
   • Streams across multiple GPUs
   • Device-to-device data copying
   • Peer-to-peer transfers

4. Assignment

DAY 4: CUDA Performance Optimization

An alternate perspective on all delivered materials, using the GTC presentation as a baseline for CUDA performance optimization CUDA Performance Optimization.

DAY 5: Practical Advice, Example Codes Diving (Wave Propagators)

Nsight Systems & Nsight Compute

To proper open ".nsys-rep" files, please download the tool Nsight Systems [link]

Similarly, use Nsight Compute tool to open ".ncu-rep" files [link].