Matmul Mastery

Boehm's worklog teaches the method, not just the kernel: change one thing, profile, explain the delta, repeat. Each rung has a name you'll reuse forever (coalescing, shared-memory tiling, vectorized loads, warp tiling), and the biggest single jump is coalescing. That's why memory-access patterns, not FLOPs, are the first thing to check in any slow kernel. The flash-attention quest asks you to produce a worklog of your own in this genre, and worklogs like it have opened real doors.

The 40%-of-cuBLAS bar is set where it is on purpose: reachable with tiling done right, unreachable by accident. Afterward, don't skip the reduction, scan, and histogram patterns. Reductions are the skeleton of softmax, layernorm, and RMSNorm, which is most of what an inference kernel engineer actually ships.