tcontract-rs

tcontract is a CPU library for dense tensor contraction in Rust, with a C ABI. It computes D = alpha * op(A) * op(B) + beta * op(C) for strided operands of f32, f64, c32 and c64, including batch indices, repeated indices (diagonals), reductions and conjugation, without transposing or copying the operands: a packed, block-scatter driver (TBLIS-style) packs general strides straight into microkernel panels, with faer as a strategy for problems that fuse into one copy-free batched GEMM. Threads and scratch memory come from an explicit execution context the host passes in; there is no ambient pool and no environment variable that changes behavior.

The C ABI is the standard TAPP tensor product interface, declared by the pinned upstream headers, plus DLPack operands and a few named extensions. It builds into one shared library, libtcontract_capi.

AI-assisted contributions are welcome; see AGENTS.md and REPOSITORY_RULES.md.

Crates

Crate	Owns
tcontract-exec	Execution context: serial, or the host's borrowed/owned/shared-Arc Rayon pool (or created and joined through the C API); width chosen from work; reusable, wrapper-owned scratch.
tcontract-kernel	The kernel layer: packed formats, kernel-family descriptors and their registry, resolution with frozen blocking, packing and write-back, cache blocking, and every microkernel family (scalar, AVX2, AVX-512 and NEON register-tile kernels, portable reference kernels, native complex kernels).
tcontract	Binary contraction over one validated Problem (from DotGeneral or from labels): Plan<T>, PlanConfig, the packed, faer and elementwise strategies chosen by the planner, contract_batched, the labels-based unary update (unary), shared errors, and the backend trait.
tcontract-capi	libtcontract_capi (cdylib, staticlib, rlib): the TAPP C ABI over tcontract, DLPack operands, executors. The only crate that produces a library for C.
tcontract-testkit	Test support: an independent label oracle, seeded fixtures, a naive second backend, downstream-kernel fixtures. Not a production fallback.
tcontract-bench (benchmarks/)	Benchmark harness (tcbench, C and Rust ABI rows); not part of the library.

strided-rs is an external dependency pinned to a published release; Cargo.toml carries the exact requirement, so this document does not repeat the number. Consumers must use the same source revision to share StridedView types.

Arrows mean "depends on" and are drawn from cargo tree (normal and build dependencies; tcontract-testkit is a dev-dependency of tcontract and tcontract-capi, which creates no cycle because it is only linked into tests).

flowchart TB
    CAPI["<b>tcontract-capi</b><br/>libtcontract_capi: TAPP, DLPack, executors"]
    CT["<b>tcontract</b><br/>Problem, Plan, strategies"]
    GK["<b>tcontract-kernel</b><br/>families, registry, packing, microkernels"]
    EX["<b>tcontract-exec</b><br/>execution context"]
    FA["faer"]
    ST["strided-rs (external)"]
    TK["tcontract-testkit"]
    CAPI --> CT
    CAPI --> EX
    CT --> GK
    CT --> EX
    CT --> FA
    CT --> ST
    TK --> CT
    TK --> GK
    TK --> EX
    classDef own fill:#e8f2ff,stroke:#2563a6,color:#132f50
    classDef ext fill:#edf0f4,stroke:#536477,color:#233244
    class CAPI,CT,GK,EX,TK own
    class FA,ST ext

Loading

Not in scope: N-ary einsum and contraction ordering (they sit above, calling tcontract per binary step), dense linear algebra, BLAS entry points (GEMM is a contraction), AD and GPU backends.

A contraction from Rust

crates/tcontract/examples/readme.rs (compiled and run in CI; a test checks that this text is that file):

//! The minimal example of the repository README (a test checks that the README
//! shows this file verbatim): `D = A * B` through `DotGeneral`, `Plan`, execute.
use strided_view::{StridedView, StridedViewMut};
use tcontract::api::{DType, DotGeneral, LayoutSpec, OperandSpec, Problem};
use tcontract::{Plan, PlanConfig, Result};
use tcontract_exec::Exec;

fn main() -> Result<()> {
    // D[i, k] = sum_j A[i, j] * B[j, k], all 2 x 2 and column-major.
    let layout = |dims: &[usize], strides: &[isize]| -> Result<OperandSpec> {
        Ok(OperandSpec::new(LayoutSpec::new(dims, strides, 0)?))
    };
    let dot = DotGeneral::new(&[1], &[0], &[], &[]);
    let problem = Problem::from_dot_general(
        DType::F64,
        layout(&[2, 2], &[1, 2])?,
        layout(&[2, 2], &[1, 2])?,
        layout(&[2, 2], &[1, 2])?,
        &dot,
    )?;

    // Planning chooses the strategy and kernel once; the plan is reusable.
    let plan = Plan::<f64>::new(&problem, &PlanConfig::default())?;
    println!("strategy: {:?}", plan.report().algorithm);

    let a = [1.0, 2.0, 3.0, 4.0];
    let b = [0.0, 1.0, 1.0, 0.0]; // swaps the columns of A
    let mut d = [0.0; 4];
    let av = StridedView::new(&a, &[2, 2], &[1, 2], 0).map_err(tcontract::Error::backend)?;
    let bv = StridedView::new(&b, &[2, 2], &[1, 2], 0).map_err(tcontract::Error::backend)?;
    let mut dv =
        StridedViewMut::new(&mut d, &[2, 2], &[1, 2], 0).map_err(tcontract::Error::backend)?;
    plan.execute_into(&Exec::serial(), 1.0, &av, &bv, &mut dv)?;

    assert_eq!(d, [3.0, 4.0, 1.0, 2.0]);
    println!("D = {d:?}");
    Ok(())
}

A contraction from C

crates/tcontract-capi/tests/c/standard_consumer.c: includes only the pinned upstream header <tapp.h>, compiles as C and as C++, and links -ltcontract_capi (compiled and run in CI, with the same verbatim check):

/* A consumer that includes only the pinned upstream TAPP headers (no tcontract
   header) and performs a serial contraction D = A * B through libtcontract_capi:
   the standard ABI is enough for serial use. */
#include <stdio.h>
#include <tapp.h>

#define CHECK(cond, msg) do { if (!(cond)) { fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, msg); return 1; } } while (0)

int main(void) {
  TAPP_handle handle;
  TAPP_executor exec;
  CHECK(TAPP_check_success(TAPP_create_handle(&handle)), "create handle");
  CHECK(TAPP_check_success(TAPP_create_executor(&exec)), "create executor");

  /* A: 2x3, B: 3x2, D: 2x2, all column-major (strides in elements). */
  int64_t ea[2] = {2, 3}, sa[2] = {1, 2};
  int64_t eb[2] = {3, 2}, sb[2] = {1, 3};
  int64_t ed[2] = {2, 2}, sd[2] = {1, 2};
  TAPP_tensor_info ia, ib, id;
  CHECK(TAPP_check_success(TAPP_create_tensor_info(&ia, TAPP_F64, 2, ea, sa)), "info A");
  CHECK(TAPP_check_success(TAPP_create_tensor_info(&ib, TAPP_F64, 2, eb, sb)), "info B");
  CHECK(TAPP_check_success(TAPP_create_tensor_info(&id, TAPP_F64, 2, ed, sd)), "info D");

  int64_t la[2] = {'i', 'k'}, lb[2] = {'k', 'j'}, ld[2] = {'i', 'j'};
  TAPP_tensor_product plan;
  CHECK(TAPP_check_success(TAPP_create_tensor_product(
            &plan, handle, TAPP_IDENTITY, ia, la, TAPP_IDENTITY, ib, lb,
            TAPP_IDENTITY, id, ld, TAPP_IDENTITY, id, ld, TAPP_DEFAULT_PREC)),
        "create product");

  double a[6] = {1, 2, 3, 4, 5, 6};   /* A[i + 2k] */
  double b[6] = {1, 2, 3, 4, 5, 6};   /* B[k + 3j] */
  double d[4] = {-1, -1, -1, -1};
  double alpha = 1.0, beta = 0.0;
  TAPP_status status = -1;
  CHECK(TAPP_check_success(TAPP_execute_product(plan, exec, &status, &alpha, a, b, &beta, TAPP_IN_PLACE, d)),
        "execute");
  CHECK(status == 0, "status written");
  /* D[i,j] = sum_k A[i,k] B[k,j] */
  CHECK(d[0] == 22 && d[1] == 28 && d[2] == 49 && d[3] == 64, "values");

  /* The default executor 0 is the same serial executor. */
  d[0] = d[1] = d[2] = d[3] = -1;
  CHECK(TAPP_check_success(TAPP_execute_product(plan, 0, NULL, &alpha, a, b, &beta, TAPP_IN_PLACE, d)),
        "execute on executor 0");
  CHECK(d[0] == 22 && d[3] == 64, "values on executor 0");

  CHECK(TAPP_check_success(TAPP_destroy_status(status)), "destroy status");
  CHECK(TAPP_check_success(TAPP_destroy_tensor_product(plan)), "destroy product");
  CHECK(TAPP_check_success(TAPP_destroy_tensor_info(ia)), "destroy A");
  CHECK(TAPP_check_success(TAPP_destroy_tensor_info(ib)), "destroy B");
  CHECK(TAPP_check_success(TAPP_destroy_tensor_info(id)), "destroy D");
  CHECK(TAPP_check_success(TAPP_destroy_executor(exec)), "destroy executor");
  CHECK(TAPP_check_success(TAPP_destroy_handle(handle)), "destroy handle");
  printf("standard_consumer ok\n");
  return 0;
}

Build the library with cargo build --release -p tcontract-capi, install it with crates/tcontract-capi/install.sh --prefix=DIR (headers, library, tcontract_capi.pc), and see examples/c-consumer for a CMake project that consumes it three ways. The extension headers are <tcontract/tcontract.h> (DLPack operands, status codes, tcontract_abi_version) and <tcontract/tapp_ext.h> (Rayon executors: one tcontract creates and joins for a host without a pool, or a tcontract_pool the C host owns and lends to an executor). Executor handles are valid for every plan of the library; handles of another TAPP provider cannot be mixed in.

Kernel selection and ids

The planner picks the strategy (plan.report().algorithm says which) and, for the packed driver, the highest-priority available kernel family. A caller can override it through PlanConfig: kernel (KernelChoice::Id, exact, no silent fallback), isa (KernelForce), method (complex method), blocking, partition and the other fields. Family ids have the form {isa}.{dtype}.{scheme}.{MR}x{NR}, for example avx2.f64.real.8x6, avx512.c32.planar.32x6 or ref.f64.real.4x4, where isa is ref, avx2, avx512 or neon; tcontract_kernel::list_kernels lists what this CPU offers and the x86-64 manifest is pinned by a snapshot test. Downstream crates can supply their own kernels through a caller-scoped catalog (Plan::new_with_selector). The full old-to-new id table is in the migration guide.

Build and test

Rust 1.89 or newer. CI runs on Linux (x86-64) and macOS (arm64); the aarch64 kernels are also checked with cargo check --target aarch64-apple-darwin.

cargo build -p tcontract-capi            # libtcontract_capi; the C tests link it
cargo test --workspace
cargo run --release -p tcontract --example readme
python3 scripts/check-agent-skills.py

The full local gate is in AGENTS.md.

Status

Phase 1 is done: the contraction stack above, the TAPP C ABI, and the benchmark harness. faer is an internal strategy, not API. Phase 2 optimizes the packed driver until it can replace faer, leaving one route for contractions with a K role that differs only by kernel family (all-batch problems stay delegated to strided-rs); hadamard.json must not regress. There is no stable API or ABI yet and no crate has been published. Performance claims need recorded measurements (PERFORMANCE_TIPS.md, docs/experiments.md). The source integration that produced this layout is described in docs/migration-2026-10.md and issue #37.

Provenance and authorship

The packed driver, its planning and the microkernels originate in tensorprimitives-rs by Lukas Devos, imported with history and authorship; the faer strategy follows the permute-plus-batched-GEMM contraction of tenferro-rs. The old tree is archived in docs/archive/tensorprimitives. Licenses, imported notices (TAPP headers BSD-3-Clause, DLPack Apache-2.0) and the per-project record are in docs/provenance.md. The code is licensed MIT OR Apache-2.0 (LICENSE-MIT, LICENSE-APACHE).

Design documents: design principles, architecture, research map, decision log.

Acknowledgements and citation

tcontract builds on faer by Sarah Quiñones El Kazdadi. If you use tcontract in published work, please also cite the faer paper:

S. Q. El Kazdadi, "faer: A linear algebra library for the Rust programming language", Journal of Open Source Software 11(123), 6099 (2026), doi:10.21105/joss.06099.

@article{Kazdadi2026,
  doi = {10.21105/joss.06099},
  url = {https://doi.org/10.21105/joss.06099},
  year = {2026},
  publisher = {The Open Journal},
  volume = {11},
  number = {123},
  pages = {6099},
  author = {Kazdadi, Sarah Quiñones El},
  title = {faer: A linear algebra library for the Rust programming language},
  journal = {Journal of Open Source Software}
}