Clean up tensor contraction APIs

.gitignore

@@ -26,6 +26,7 @@ build/
 Manifest.toml
 LocalPreferences.toml
 .julia-tests/
+benchmarks/results/*.h5
 
 # Avoid accidentally committing ad-hoc analysis notes at repo root
 /*.md

Cargo.toml

@@ -41,6 +41,7 @@ bnum = "0.12"
 uint = "0.10"
 libc = "0.2"
 paste = "1.0"
+smallvec = "1.15"
 thiserror = "2.0"
 rand = "0.9"
 rand_chacha = "0.9"
@@ -53,7 +54,7 @@ hdf5-metno = { version = "0.12", default-features = false }
 ndarray = "0.17"
 quanticsgrids = { git = "https://github.com/tensor4all/quanticsgrids-rs", rev = "a76b8fb" }
 hdf5-rt = { git = "https://github.com/tensor4all/hdf5-rt", default-features = false }
-tenferro = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "925e2511bc6bd019432f66596de39389dfced754", default-features = false }
-tenferro-device = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "925e2511bc6bd019432f66596de39389dfced754" }
-tenferro-einsum = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "925e2511bc6bd019432f66596de39389dfced754", default-features = false }
-tenferro-tensor = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "925e2511bc6bd019432f66596de39389dfced754", default-features = false }
+tenferro = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "6283c6f1a8a56f920a2600f6c27e9e6ade28beed", default-features = false }
+tenferro-device = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "6283c6f1a8a56f920a2600f6c27e9e6ade28beed" }
+tenferro-einsum = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "6283c6f1a8a56f920a2600f6c27e9e6ade28beed", default-features = false }
+tenferro-tensor = { git = "https://github.com/tensor4all/tenferro-rs.git", rev = "6283c6f1a8a56f920a2600f6c27e9e6ade28beed", default-features = false }

benchmarks/README.md

@@ -6,6 +6,7 @@ This directory contains benchmark code for comparing Rust and Julia implementati
 
 - `rust/`: Rust benchmark code using `tensor4all-rs`
 - `julia/`: Julia benchmark code using `ITensors.jl` and `ITensorMPS.jl`
+- `results/`: saved benchmark commands and representative local outputs
 
 ## Running Benchmarks
 
@@ -16,13 +17,73 @@ cd crates/tensor4all-itensorlike
 cargo run --release --example benchmark_contract
 ```
 
+Projected local-operator apply:
+
+```bash
+RAYON_NUM_THREADS=1 cargo run -p tensor4all-treetn --example benchmark_projected_apply --release -- 38 32 32 3 0
+```
+
+Prepared local linsolve:
+
+```bash
+RAYON_NUM_THREADS=1 cargo run -p tensor4all-treetn --example benchmark_local_linsolve --release -- 38 32 32 1 10 30 0
+```
+
+Non-AD local tensor operations:
+
+```bash
+RAYON_NUM_THREADS=1 cargo run -p tensor4all-core --example benchmark_tensor_ops --release -- 20000 6 2 2 6
+```
+
+TensorTrain-level operations against ITensorMPS:
+
+```bash
+RAYON_NUM_THREADS=1 cargo run -p tensor4all-itensorlike --example benchmark_tt_ops --release -- --L 32 --zipup-L 10 --chis 4,8,16,32,64
+```
+
+Inspect Julia-dumped local linsolve inputs:
+
+```bash
+cargo run -p tensor4all-hdf5 --example inspect_mps_inputs --release -- benchmarks/results/local_linsolve_inputs_N38_b32_o32.h5
+```
+
 ### Julia
 
 ```bash
 cd external/ITensorMPS.jl
 julia benchmark/benchmark_contract.jl
 ```
 
+Projected local-operator apply:
+
+```bash
+BLAS_NUM_THREADS=1 julia --project=benchmarks/julia benchmarks/julia/benchmark_projected_apply.jl 38 32 32 3 0
+```
+
+Prepared local linsolve:
+
+```bash
+BLAS_NUM_THREADS=1 julia --project=benchmarks/julia benchmarks/julia/benchmark_local_linsolve.jl 38 32 32 1 1 10
+```
+
+Non-AD local tensor operations:
+
+```bash
+BLAS_NUM_THREADS=1 julia --project=benchmarks/julia benchmarks/julia/benchmark_tensor_ops.jl 20000 6 2 2 6
+```
+
+TensorTrain-level operations against tensor4all:
+
+```bash
+BLAS_NUM_THREADS=1 julia --project=benchmarks/julia benchmarks/julia/benchmark_tt_ops.jl --L 32 --zipup-L 10 --chis 4,8,16,32,64
+```
+
+Dump local linsolve inputs as ITensorMPS-compatible HDF5:
+
+```bash
+BLAS_NUM_THREADS=1 julia --project=benchmarks/julia benchmarks/julia/dump_local_linsolve_inputs.jl benchmarks/results/local_linsolve_inputs_N38_b32_o32.h5 38 32 32
+```
+
 ## Benchmark Details
 
 Both benchmarks perform MPO-MPO contraction using zip-up algorithm:
@@ -31,3 +92,36 @@ Both benchmarks perform MPO-MPO contraction using zip-up algorithm:
 - Bond dimension: 50
 - Max rank: 50
 - Includes orthogonalization time in measurements
+
+The projected local-operator apply benchmarks isolate the local matvec hot path
+used by two-site TreeTN/ITensor-style local solves. The Rust source of truth is
+`benchmarks/rust/benchmark_projected_apply.rs`, included by the cargo example
+target under `tensor4all-treetn`; the Julia counterpart is
+`benchmarks/julia/benchmark_projected_apply.jl`.
+
+The prepared local linsolve benchmarks construct the operator, right-hand side,
+and initial state once, then time the local solve body. They also report local
+GMRES/apply/RHS/factorization buckets. Use Julia `maxiter=1, krylovdim=10` as a
+rough match to Rust's `krylov_maxiter=10` total-iteration cap; KrylovKit's
+`maxiter=10, krylovdim=30` performs far more local operator applications.
+
+The non-AD local tensor operation benchmarks isolate `inner`, `norm`, affine
+addition, and explicit `conj`-then-contract on a small dense tensor. The default
+shape `[6, 2, 2, 6]` mirrors the small two-site local tensors observed in the
+QuanticsNEGF long-time local Krylov test, where dispatch/allocation overhead can
+dominate floating-point work.
+
+The TensorTrain-level operation benchmarks compare tensor4all's
+`TensorTrain::inner`, strict direct-sum MPS addition, and prepared MPO×MPO zipup
+contraction against ITensorMPS.jl. They use deterministic Complex64 fixtures and
+print CSV-style rows with sample counts, min/median/mean/max milliseconds,
+result max bond dimension, and a checksum. The Rust source of truth is
+`benchmarks/rust/benchmark_tt_ops.rs`, included by
+`tensor4all-itensorlike/examples/benchmark_tt_ops.rs`; the Julia counterpart is
+`benchmarks/julia/benchmark_tt_ops.jl`.
+
+`dump_local_linsolve_inputs.jl` writes the prepared local operator as
+`operator_as_mps`, plus `rhs` and `init`, in one HDF5 file. The operator is a
+Julia `MPO` stored through the `MPS` schema by saving its site tensors as
+`MPS([H[i] for i in 1:length(H)])`; Rust reads all three groups with
+`tensor4all_hdf5::load_mps`.

benchmarks/julia/Project.toml

@@ -1,5 +1,12 @@
 [deps]
 FastMPOContractions = "f6e391d2-8ffa-4d7a-98cd-7e70024481ca"
+HDF5 = "f67ccb44-e63f-5c2f-98bd-6dc0ccc4ba2f"
 ITensorMPS = "0d1a4710-d33b-49a5-8f18-73bdf49b47e2"
+ITensorTDVP = "25707e16-a4db-4a07-99d9-4d67b7af0342"
+ITensors = "9136182c-28ba-11e9-034c-db9fb085ebd5"
 KrylovKit = "0b1a1467-8014-51b9-945f-bf0ae24f4b77"
 VectorInterface = "409d34a3-91d5-4945-b6ec-7529ddf182d8"
+
+[compat]
+ITensorTDVP = "^0.4"
+ITensors = "^0.6"