First program

Create app.ch with a function that applies ReLU and softmax to a tensor, then calls that function with three values:

def relu_then_softmax[n](x: tensor[n, f32]) -> tensor[n, f32] = x |> relu |> softmax(0)
result = [-1.0, 0.0, 1.0] |> to_tensor(f32) |> relu_then_softmax

The dimension variable n lets the function accept a vector of any length. tensor[n, f32] says that its elements have f32 precision. The |> operator passes its left value as the first argument of the next call: x |> softmax(0) means softmax(x, 0).

to_tensor(f32) states the element dtype for the bracket literal. Numeric literals inside to_tensor need either a dtype argument or explicit suffixes; they do not take a default dtype.

Run these commands in the directory containing app.ch:

chelis fmt --inplace app.ch
chelis check app.ch
chelis eval --file app.ch

The evaluator prints result as a tensor with shape [3]. Its values are approximately 0.212, 0.212, and 0.576.

To inspect the compiler's Deep representation and build a native executable, run:

chelis deep app.ch
chelis build app.ch --output out/
./out/app

chelis build invokes the system C compiler and links its carried runtime. Use --emit-c when you only want generated sources and runtime artifacts. See CLI Workflow for the other commands and their output.