Add a TBLIS.jl package extension

[lkdvos]

Adds TensorOperationsTBLISExt, providing a TBLISBackend on top of TBLIS.jl v0.3.
Supersedes the standalone TensorOperationsTBLIS.jl.

What it does

TBLISBackend() routes tensoradd!, tensortrace! and tensorcontract! through TBLIS, which contracts strided tensors in place rather than reshaping them into matrices, avoiding the permuted intermediates the BLAS path has to materialize.

It is opt-in: loading TBLIS.jl registers no select_backend method.

using TensorOperations, TBLIS
TBLIS.set_num_threads(16)
@tensor backend = TensorOperations.TBLISBackend() D[a,b,c,d] := A[a,e,c,f] * B[g,d,e] * conj(C[g,f,b])

Design notes

• Unsupported arguments throw. Mixed or unsupported element types, non-strided arrays and conjugated output views raise an ArgumentError instead of being delegated to another backend, so a contraction that never reaches TBLIS cannot pass for one that did.
• Conjugation is a runtime flag on the descriptor, not part of a StridedView's type, so the entry points stay type stable without branching over conj variants (@inferred covers all three). A conjugation already carried by the input, as for an Adjoint, combines with the requested one.

Library quirks worked around

• tblis_tensor_mult ignores the per-tensor conjugation flags, while tblis_tensor_add honours them (verified against tblis 1.3). Resolved by conjugating the output in place when both factors are conjugated, and by materializing the conjugated factor into an allocator temporary otherwise.
• TBLIS.jl's tblis_tensor constructor only accepts StridedArray, so it cannot take the StridedViews these kernels use, and it cannot set the conjugation flag. The descriptor is initialized through the low-level bindings and every referenced buffer is rooted with GC.@preserve. Widening that constructor upstream would let this drop back to the public API.

Tests

test/tblis.jl, wired into runtests.jl: compared against StridedNative over all four supported element types for add / trace / contract, every conjugation combination, Adjoint inputs, non-contiguous views, outer products, scalar outputs, argument checking, the rejection paths, @tensor/ncon integration, a GC-pressure loop, and threading.
All 20 testsets pass against the registered TBLIS.jl v0.3.0.

β == 0 is exercised with NaN-poisoned outputs, confirming TBLIS ignores C in that case for both add and mult.

Performance

Measured separately (the harness is not part of this PR) on 2× Xeon Gold 6244, with OpenBLAS, TBLIS and Strided all on 16 threads.

TBLIS is not a throughput win on this hardware. It is roughly on par with StridedBLAS for permuted real contractions (1.19×) and slower for other shapes: 0.64× on GEMM, 0.94× on a DMRG-style network, 0.56× on a partial trace.
What it buys is memory: for a permuted contraction StridedBLAS allocates 76.6 MiB of temporaries per call where TBLIS allocates zero.

Warning

As of tblis_jll v1.3, TBLIS has no competitive support for complex element types. Complex contractions run at 0.04–0.11× of BLAS; a bare tblis_tensor_mult on a 1000³ complex matrix product reaches ~25 GFLOP/s against ~500 GFLOP/s for mul!. tblis_jll v1.2.0 behaves identically, so this is not a regression from the v1.3 bump. This is called out in the TBLISBackend docstring and in the backends documentation.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.97959% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
ext/TensorOperationsTBLISExt.jl	98.96%	1 Missing ⚠️

Files with missing lines	Coverage Δ
src/backends.jl	77.77% <100.00%> (+6.34%)	⬆️
ext/TensorOperationsTBLISExt.jl	98.96% <98.96%> (ø)

... and 6 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.