feat: API expansion and quantics Rust parity by shinaoka · Pull Request #22 · tensor4all/Tensor4all.jl

shinaoka · 2026-03-31T13:16:43Z

Summary

This PR extends the existing Julia API expansion with quantics Rust parity on top of the merged tensor4all-rs C API work.

In addition to the earlier API cleanup and ComplexF64 work already on this branch, this update adds:

top-level DiscretizedGrid, InherentDiscreteGrid, and localdimensions exports
unfolding=:grouped support in QuanticsGrids
Julia C bindings for multivariable, affine, and binary-op quantics transforms
high-level QuanticsTransform wrappers for those operators
set_input_space!, set_output_space!, and set_iospaces! so LinearOperator application works with Julia-created MPS objects
focused regression tests for the new quantics grid/transform surface
a fix for the stale ComplexF64 tensor ABI in the Julia wrapper so complex tensor creation/data round-trip now matches the current Rust C API

Existing branch scope

This branch also contains the broader API expansion that was already open on PR #22:

1-indexed everywhere: SimpleTT evaluate, sitetensor now 1-indexed. TreeTCI pivots/evaluate now 1-indexed.
Function renames: site_dims→sitedims, link_dims→linkdims, site_tensor→sitetensor, local_dimensions→localdimensions
QuanticsTCI returns tuple: quanticscrossinterpolate now returns (qtci, ranks, errors)
QuanticsTCI signature: first arg is ::Type{V}
SimpleTT additions: ComplexF64 support, arithmetic, scale!, reverse, fulltensor, compress!, site-tensor constructors, and TreeTN conversions
TreeTCI cleanup: 1-indexed user API and naming alignment
TCI extension: SimpleTensorTrain ↔ TensorCrossInterpolation.TensorTrain

Why

ReFrequenTT needs the quantics backend surface more than it needs a revived TensorCI compatibility layer. The missing migration blockers were multivariable transforms, affine embeddings, mixed boundary conditions, and a clean path from Julia MPS objects into the quantics operator apply path.

The ComplexF64 failures turned out to be a separate wrapper bug: Tensor4all.jl was still calling the old split real/imag ABI for dense complex tensors, while the current Rust C API uses a single interleaved buffer. This PR aligns the Julia wrapper with the Rust ABI for both construction and extraction.

Validation

Passed:

TENSOR4ALL_RS_PATH=/home/shinaoka/tensor4all/tensor4all-rs/.worktrees/feat-quanticstransform-capi julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl deps/build.jl
julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl -e 'using Test, Tensor4all; const T4AIndex=Tensor4all.Index; const T4ATensor=Tensor4all.Tensor; include("/home/shinaoka/tensor4all/Tensor4all.jl/test/test_tensor.jl")'
julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl -e 'using Test, Tensor4all, Tensor4all.SimpleTT, Tensor4all.TreeTN; include("/home/shinaoka/tensor4all/Tensor4all.jl/test/test_conversions.jl")'
julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl -e 'using Test, Tensor4all, Tensor4all.TreeTN; const T4AIndex=Tensor4all.Index; include("/home/shinaoka/tensor4all/Tensor4all.jl/test/test_treetn.jl")'
julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl -e 'include("/home/shinaoka/tensor4all/Tensor4all.jl/test/test_quanticsgrids.jl")'
julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl -e 'include("/home/shinaoka/tensor4all/Tensor4all.jl/test/test_quanticstransform.jl")'
T4A_SKIP_HDF5_TESTS=1 julia --startup-file=no --project=/home/shinaoka/tensor4all/Tensor4all.jl /home/shinaoka/tensor4all/Tensor4all.jl/test/runtests.jl

Current suite status in this environment:

443 passed
0 failed
0 broken

with T4A_SKIP_HDF5_TESTS=1.

…CI c64 C API functions Add ~55 new ccall wrappers to C_API.jl for the expanded tensor4all-rs C API: - SimpleTT f64: compress, partial_sum, from_site_tensors, add, scale, dot, reverse, fulltensor - SimpleTT c64: full set (lifecycle, constructors, accessors, operations) - QtciOptions: lifecycle (default, release, clone) and all setters - QuanticsTCI f64: clone, max_bond_error, max_rank - QuanticsTCI c64: full set (lifecycle, accessors, operations, interpolation) - Update crossinterpolate_f64/discrete_f64 signatures with options, initial_pivots, output params Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

… new operations Complete rewrite of SimpleTT.jl: - Type parameter now Union{Float64, ComplexF64} (was T<:Real, Float64 only) - Type dispatch via _api(T, :name) helper for f64/c64 C API calls - All user-facing indices are 1-based (internal C API calls subtract 1) - Renamed functions: site_dims->sitedims, link_dims->linkdims, site_tensor->sitetensor - New operations: +, -, scalar *, dot, scale!, reverse, fulltensor, from_site_tensors - New functions: compress!, partial_sum (both supporting f64 and c64) - ComplexF64 data handled via reinterpret of interleaved doubles Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…, overloads - QuanticsTensorCI2{V} is now type-parameterized (V = Float64 or ComplexF64) - quanticscrossinterpolate returns (qtci, ranks, errors) tuple - First argument is ::Type{V} to match Pure Julia API - Build ephemeral QtciOptions handle from kwargs, release after call - Support initial_pivots, out_ranks, out_errors, out_n_iters - Add maxbonderror(qtci) and maxrank(qtci) accessors - ComplexF64 callback trampolines write [re, im] to result buffer - evaluate, sum, integral for c64 return ComplexF64 - to_tensor_train for c64 returns SimpleTensorTrain{ComplexF64} - Rename link_dims to linkdims - Add overloads: size tuple (discrete), Array (from dense) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Cover Float64 and ComplexF64 variants: - continuous grid (DiscretizedGrid) - discrete domain (size tuple) - from Array - evaluate, sum, integral, to_tensor_train - maxbonderror, maxrank, linkdims - callable interface - kwargs/options passthrough Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- All user-facing APIs now use 1-indexed (Julia convention) - crossinterpolate2: initialpivots default is [ones(Int, n)]; pivots are converted to 0-indexed before passing to C API - evaluate(ttn, indices): subtracts 1 from all indices before C API call - evaluate(ttn, batch::Matrix): subtracts 1 from all values - Callback trampolines: add 1 to batch values from C API before calling user function, so user always sees 1-indexed values - Rename bond_dims -> bonddims, max_bond_error -> maxbonderror, max_rank -> maxrank, max_sample_value -> maxsamplevalue - Replace TensorCI module include with TreeTCI in Tensor4all.jl - Add comprehensive test/test_treetci.jl Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add MPS(tt::SimpleTensorTrain) and SimpleTensorTrain(mps::TreeTensorNetwork{Int}) constructors in TreeTN.jl for converting between the two representations. Includes tests for round-trip conversions (both directions) with Float64 and ComplexF64. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add weak dependency on TensorCrossInterpolation and create Tensor4allTCIExt extension providing bidirectional conversion between TensorCrossInterpolation.TensorTrain and SimpleTensorTrain. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…t convention Define linkdims and compress! as generic functions in Tensor4all namespace so submodules (SimpleTT, TreeTN, QuanticsTCI) extend the same function instead of creating separate ones that collide when multiple submodules are loaded together. Export rank and compress! from SimpleTT. Fix ComplexF64 dot product to follow Julia convention (sesquilinear, conjugating the first argument) by adding a conj implementation for SimpleTensorTrain and using it before calling the Rust bilinear dot. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Define evaluate, maxbonderror, maxrank as generic functions in Tensor4all module (alongside linkdims and compress!) so submodules share a single dispatch table instead of creating module-local functions that can't be found across module boundaries. - Change QuanticsTCI.sum to extend Base.sum (matching SimpleTT pattern) so that sum(qtci) dispatches correctly without explicit import. - Update SimpleTT to import and re-export evaluate from parent module. - Update test imports to use Tensor4all-level generics. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ailures QuanticsTCI.evaluate expects grid indices (1-indexed, one per original dimension), not quantics indices. Changed tests to use origcoord_to_grididx instead of origcoord_to_quantics. Skip ComplexF64 round-trip conversion test since the C API does not yet support complex tensor creation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

shinaoka · 2026-04-06T08:20:22Z

This PR is superseded by #25.

feat/api-expansion drifted too far from current main and now conflicts heavily with already-merged work, so I rebuilt the actionable fix on a clean branch from main instead.

Please review and merge #25:

[codex] Fix ComplexF64 SimpleTT site tensor regression #25

That replacement PR contains the narrow fix for the ComplexF64 SimpleTT site-tensor regression plus the corresponding round-trip regression test.

shinaoka · 2026-04-06T08:20:23Z

Superseded by #25.

shinaoka and others added 16 commits March 31, 2026 20:48

refactor: rename local_dimensions to localdimensions

680d380

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

fix: QuanticsTCI linkdims via TT, relax integral tolerance for R=8 grid

2a97bed

fix: import rank in QuanticsTCI, fix linkdims implementation

994b205

fix: add rank method for QuanticsTensorCI2

8931639

fix: import SimpleTT module in QuanticsTCI for linkdims

b37d4ba

feat: expose quantics rust parity in Julia

e65e2c7

shinaoka changed the title ~~feat: API expansion — ComplexF64, 1-indexed, naming alignment, new operations~~ feat: API expansion and quantics Rust parity Apr 6, 2026

fix: align complex tensor ABI with Rust C API

2229c40

shinaoka enabled auto-merge (squash) April 6, 2026 08:08

shinaoka mentioned this pull request Apr 6, 2026

[codex] Fix ComplexF64 SimpleTT site tensor regression #25

Merged

shinaoka closed this Apr 6, 2026

auto-merge was automatically disabled April 6, 2026 08:20
Pull request was closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: API expansion and quantics Rust parity#22

feat: API expansion and quantics Rust parity#22
shinaoka wants to merge 17 commits intomainfrom
feat/api-expansion

shinaoka commented Mar 31, 2026 •

edited

Loading

Uh oh!

shinaoka commented Apr 6, 2026

Uh oh!

shinaoka commented Apr 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

shinaoka commented Mar 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Existing branch scope

Why

Validation

Related

Uh oh!

shinaoka commented Apr 6, 2026

Uh oh!

shinaoka commented Apr 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

shinaoka commented Mar 31, 2026 •

edited

Loading