Module tensor 

TensorAxis declaration + parametric square-matmul impl + shape.

Per Wiki ADR-031 the tensor sub-crate exposes TensorAxis as the canonical Layer-3 surface for tensor compute. The reference impl CpuI8MatmulSquare is generic over the square dimension DIM, with i8 inputs and saturating-i16 outputs — the integer-arithmetic determinism contract ADR-030 names as the axis substitution-determinism baseline.

Variable-rank tensor compute composes through verbs over partition_product!-declared shapes per ADR-033/044; the axis’s role is the fixed-shape atomic primitive.

ADR-055 substrate-Term verb body discipline

Per Wiki ADR-055 every AxisExtension impl satisfies the substrate-Term verb body discipline; the hand-written kernel below uses the default empty body_arena() emitted by foundation-sdk 0.4.11’s axis! companion macro (the primitive-fast-path-equivalent realization).

Explicit substrate-Term decomposition of CpuI8MatmulSquare<DIM>::matmul — fold_n(DIM, ...) over rows × fold_n(DIM, ...) over columns × fold_n(DIM, ...) over reductions, with a sign_extend sub-verb (matching Ge(operand, Literal(0x80, W8)) to select between Concat(0x00, operand) and Concat(0xff, operand)) plus W16 Mul + W16 Add accumulation plus saturation via Match over Ge(acc, Literal(0x7fff, W16)) / Lt(acc, Literal(0x8000, W16)) per ADR-054 § Substrate-Term realization examples — is syntactically expressible in foundation-sdk 0.4.11’s verb-body grammar. ADR-056 admits le/lt/ge/gt and concat in verb/axis bodies (only the route body’s syntactic surface retains the ψ-residuals rejection); foundation-sdk 0.4.11’s depth-2 const-generic-leaf partition-product projection covers the fold-n composition over matrix shapes. The remaining work is operational composition: the architectural witness verbs in crate::verbs (saturating-xor + concat-bytes) demonstrate the per-element primitives; the unfolded fold-over-rows-and-columns matmul body is a published-roster follow-on.

The hand-written for-loop kernel below is the operational form; byte-output equivalence with BLAS reference outputs at integer precision is checked at tests/conformance.rs.

Structs

CpuI8MatmulSquare

Parametric square DIM × DIM i8 × i8 → i16 matmul.

MatrixShape

Parametric ConstrainedTypeShape for a row-major ROWS × COLS matrix of ELEM_BYTES-byte elements. Per ADR-031’s Tensor<Element, Shape> shape commitment, restricted to matrix rank-2 here; higher ranks compose through partition_product! per ADR-033/044.

VectorShape

Parametric ConstrainedTypeShape for a length-N vector of ELEM_BYTES-byte elements. Per ADR-031’s Tensor<Element, Shape> for rank-1.

Constants

KERNEL_MATMUL

MAX_TENSOR_DIM

Maximum square dimension any CpuI8MatmulSquare instantiation supports. Cap at 16: output buffer = 2 * 16 * 16 = 512 bytes, inputs = 256 bytes each, total kernel byte budget bounded.

Traits

TensorAxis

Wiki ADR-031 tensor-compute axis.

Type Aliases

CpuI8Tensor4x4Matmul

4×4 i8 matmul — the canonical small-tensor reference.

CpuI8Tensor8x8Matmul

8×8 i8 matmul.

CpuI8Tensor16x16Matmul

16×16 i8 matmul (the MAX_TENSOR_DIM ceiling).