Grids · Blocks · Threads
Three levels of organization: launch a grid, schedule blocks, execute threads.
THE ATLAS · SEVEN DOMAINS · EVERY CONCEPT
The whole machine, mapped.
Reference coverage of the full CUDA surface: hardware, threads, memory, performance, asynchrony, tooling and libraries — exhibits where seeing beats reading.
Grids, blocks, threads, indexing math, divergence, synchronization.
Indexing Formulas — 2D & 3D
Linearize coordinates: row-major math for images and volumes.
Warp Divergence
If lanes disagree at a branch, the warp executes both paths serially.
Synchronization
__syncthreads() barriers within a block; warp intrinsics need no barrier.
◈ entries are hands-on exhibits. Everything else expands in place. The full narrative version of core topics lives in the journey.