liveBook · Manning

Tl.min

As explained in GPU Programming with Triton

tl.min is a Triton operation that selects the minimum value along one or more tensor dimensions, optionally returning its index.

tl.min reduces a tensor dimension to its minimum value. Its axis selects a dimension of the tile held by one Triton program instance; it does not reduce values across program instances. section 4.1.2

For a one-dimensional tile, axis=0 reduces the entire dimension to a scalar. For a two-dimensional tile, axis selects which dimension to collapse. A grid-wide reduction instead requires coordinating partial results, commonly with another kernel or with atomic operations. section 4.1.2

How GPU Programming with Triton covers it

How deep it goes
Explains axis-based tile reduction and stops short of detailing tl.min’s implementation.
Examples
Uses one- and two-dimensional tensor shapes to illustrate axis-based reduction.
What you can do after
Understand that tl.min reduces a tile dimension but does not synchronize values across program instances.
How it connects
Connects tl.min with tl.sum, tl.max, tl.argmax, tl.argmin, tl.reduce, and Triton’s program-instance model.

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.