liveBook · Manning

Tile size

As explained in GPU Programming with Triton

Tile size is the configuration specifying the dimensions or element count of a tile processed by a kernel.

Tile size specifies how many elements a kernel processes in a tile. Triton kernel writers choose tile shapes; Helion can search candidate sizes with its autotuner. section 1.2.2

For a one-dimensional softmax, a single program instance must cover the full reduction axis. Choose a tile size of at least N and launch one instance per vector. This keeps the reduction local, but increases on-chip resource use and work per instance. section 4.2.2

How GPU Programming with Triton covers it

How deep it goes
Explains a softmax sizing constraint and its resource tradeoff, but gives no general manual tuning procedure.
Examples
Uses a one-dimensional vector softmax with tile size at least N and one program instance per vector.
What you can do after
Configure a Triton softmax tile to cover its full reduction axis, trading global-memory staging for on-chip resource use.
How it connects
Connects tile sizing to Triton blocked kernels and Helion autotuning.

Also explained in

Compare how 2 books explain tile size →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of GPU Programming with Triton. The section links point to the parts of the book it draws on.