Integer tensors, static shapes, rank 1 or 2, no control flow. Statements end at a newline
(newlines inside (...) and [...] do not end a statement). # starts a comment.
input A : i8[4, 8] # declare an input
const W : i8[2, 2] = [[1, -2], # declare a constant (shape and range are checked)
[3, 4]]
C = relu(A @ B + b) # define a value (names are single-assignment)
output C, D # one or more outputs (at least one is required)Reserved words: input const output relu i8 i32.
Types
i8 or i32, with 1 or 2 positive dimensions: i8[4, 8], i32[6].
Expressions
@ binds tighter than +; both are left-associative; parentheses group.
| Expression | Rule | Result |
|---|---|---|
a @ b |
both rank 2, a is [M,K] and b is [K,N]; operand types may be i8 or i32 |
i32[M,N] |
a + b |
same shape, or [M,N] + [N] in either order (the row is broadcast); i8 operands are widened |
i32, the larger shape |
relu(a) |
any | same shape and type as a |
Semantics
All arithmetic is exact (widened); every op result must fit in i32, otherwise evaluation
fails with RangeError, because tiny-gpu raises on overflow rather than wrapping. Inputs must
fit their declared type.
Errors
The first problem stops compilation and is shown with the source line and a caret:
matmul_inner_mismatch.tiny:3:5: error: matmul inner dimensions differ: 8 vs 4 (i8[4,8] @ i8[4,4])
3 | C = A @ B
| ^^^^^Every message is snapshot-tested (tests/golden/errors/).
Target legality (checked later, Phase 2)
Whether a program can run on a chip is not a language rule: tiny-tpu needs i8 matmul operands,
and tiny-gpu has a memory budget. See docs/PLAN.md §3.3.