tiny-compiler

Live explorer

One tiny tensor language, two real chips. The same program compiles to tiny-gpu or tiny-tpu.

Open compiler explorer View source on GitHub

A recording of the compile sketch below, with the inputs changed.

Explore the idea.

Edit the two 2×2 matrices and compile. The result is computed in JavaScript, and the two lowerings are described. Explore the real compiler and its execution traces on the project site.

Interactive illustration · Simplified model

What it shows.

The same tensor program targets the tiny-gpu simulator and the tiny-tpu RTL.

Each lowering is checked against the same expected result, so the two targets cannot quietly disagree.

Three runs of one program.

The same expression on three inputs. Negative sums are clipped to zero by relu, which the sketch shows directly.

Default inputs: The result of relu(A @ B) for the starting matrices. — **Default inputs.** The result of relu(A @ B) for the starting matrices.

One value changed: A single edit changes the first output and the rest of the row. — **One value changed.** A single edit changes the first output and the rest of the row.

Everything clipped: Every product is negative, so relu gives a zero matrix. — **Everything clipped.** Every product is negative, so relu gives a zero matrix.