These articles cover performance optimization, assuming familiarity with the fundamentals and the profiling tools.

The Poisson example is a guided, end-to-end optimization of a complete application. This extended example demonstrates how to use the Nsight tools to carry out profile-guided optimization.

1. Part 1: Background and reference implementation (WIP)

2. Part 2: Iterative solver algorithm (WIP)

3. Part 3: Memory Pools (WIP)

4. Part 4: Kernel Fusion (WIP)

5. Part 5: cuSPARSE and Arithmetic Intensity

6. Part 6: Block Size Tuning (WIP)

7. Part 7: Laplace Kernel Tuning (WIP)

Deeper dives into individual phenomena:

• Bytes In Flight (WIP) — latency hiding, Little's law, ILP, and vectorized loads

• Instruction Latencies (WIP) — compute-bound optimization: measuring op costs, the roofline, function approximation

• Load Balancing (WIP) — data-dependent divergence, CUB block sort, producer/consumer queues

Case studies:

• SoA vs AoS — Which data layout is best for my application?

• Mesh Connectivity — Using Thrust to accelerate a mesh preprocessing

• Black Hole Rendering (WIP)

• Granular Flow (written document WIP, sources available here, video demo)