Consolidate benchmarks around interpolative Gaussian inputs

.github/workflows/ci.yml

@@ -7,56 +7,27 @@ on:
 jobs:
   rust:
     runs-on: ubuntu-latest
-    # The Rust build compiles all of tensor4all-rs from source on a cold cache.
     timeout-minutes: 60
     steps:
       - uses: actions/checkout@v4
       - run: sudo apt-get update && sudo apt-get install -y libhdf5-dev liblapack-dev
       - uses: dtolnay/rust-toolchain@stable
       - uses: Swatinem/rust-cache@v2
-      - run: cargo test --release
-      - name: smoke elementwise
+      - run: cargo fmt --all -- --check
+      - run: cargo clippy --locked --all-targets -- -D warnings
+      - run: cargo test --release --locked
+      - name: smoke Fourier elementwise
         run: >
           BENCH_KS=4 BENCH_R=10 BENCH_RUNS=1 BENCH_WARMUPS=0
-          OUT_DIR=/tmp/smoke EXPORT_HDF5=/tmp/smoke
-          cargo run --release --bin elementwise_fourier
-      - name: smoke mpo
+          OUT_DIR=/tmp/smoke/raw
+          cargo run --release --locked --bin elementwise_fourier
+      - name: smoke Gaussian elementwise
         run: >
-          BENCH_RS=8 BENCH_NGAUSS=3 BENCH_RUNS=1 BENCH_WARMUPS=0 BENCH_SANITY=1e-1
-          OUT_DIR=/tmp/smoke EXPORT_HDF5=/tmp/smoke
-          cargo run --release --bin mpo_mpo_quantics
-      - name: smoke elementwise gauss2d
+          BENCH_NS=2 BENCH_RUNS=1 BENCH_WARMUPS=0
+          BENCH_INPUT_CACHE_DIR=/tmp/smoke-cache OUT_DIR=/tmp/smoke/raw
+          cargo run --release --locked --bin elementwise_gauss2d_patched
+      - name: smoke Gaussian MPO contraction using the shared cache
         run: >
-          BENCH_RS=8 BENCH_NGAUSS=3 BENCH_RUNS=1 BENCH_WARMUPS=0 BENCH_SANITY=1e-1
-          OUT_DIR=/tmp/smoke EXPORT_HDF5=/tmp/smoke-gauss2d
-          cargo run --release --bin elementwise_gauss2d
-      - name: smoke elementwise gauss2d scaling
-        # No EXPORT_HDF5: case 4 exports no instances, so there is no Julia
-        # check for it, and a single N is enough to exercise the runner.
-        run: >
-          BENCH_NS=8 BENCH_R0=8 BENCH_RUNS=1 BENCH_WARMUPS=0 BENCH_SANITY=1e-1
-          OUT_DIR=/tmp/smoke
-          cargo run --release --bin elementwise_gauss2d_scaling
-      - name: smoke elementwise gauss2d patched
-        # Both instance families, since they share only the harness: the default
-        # aniso one, whose N = 8 point is a fraction of a second, and the smooth
-        # one with R0 lowered so the patched construction stays cheap there, since
-        # at the default R0 one N already builds hundreds of patches. No
-        # EXPORT_HDF5, as for case 4.
-        run: >
-          BENCH_NS=8 BENCH_RUNS=1 BENCH_WARMUPS=0 BENCH_SANITY=1e-1
-          OUT_DIR=/tmp/smoke
-          cargo run --release --bin elementwise_gauss2d_patched
-          &&
-          BENCH_FAMILY=smooth BENCH_NS=8 BENCH_R0=8 BENCH_RUNS=1 BENCH_WARMUPS=0
-          BENCH_SANITY=1e-1 OUT_DIR=/tmp/smoke
-          cargo run --release --bin elementwise_gauss2d_patched
-      - uses: julia-actions/setup-julia@v2
-        with:
-          version: "1"
-      - name: julia checks
-        run: |
-          julia --project=julia -e 'using Pkg; Pkg.instantiate()'
-          julia --project=julia julia/check_elementwise.jl /tmp/smoke 4
-          julia --project=julia julia/check_mpo_mpo.jl /tmp/smoke 8
-          julia --project=julia julia/check_elementwise_gauss2d.jl /tmp/smoke-gauss2d 8
+          BENCH_NS=2 BENCH_RUNS=1 BENCH_WARMUPS=0
+          BENCH_INPUT_CACHE_DIR=/tmp/smoke-cache OUT_DIR=/tmp/smoke/raw
+          cargo run --release --locked --bin mpo_mpo_aniso_patched

.gitignore

@@ -3,5 +3,4 @@
 Cargo.lock.orig
 __pycache__/
 .venv/
-julia/Manifest.toml
 result/**/raw/*.h5

AGENTS.md

@@ -1,66 +1,35 @@
 # Agent guidance
 
-Read `README.md` first for the cases, arm names and environment knobs. This file
-covers only what is easy to get wrong.
-
-## Julia is JIT compiled
-
-Any TIMED comparison against Julia must run warmup iterations before it measures.
-A first invocation includes compilation time and is not a measurement.
-
-The scripts in `julia/` are correctness checks only, not benchmarks. They read an
-exported HDF5 instance back through ITensors.jl and compare against the same
-analytic formula the Rust side uses. Their wall time is dominated by compilation
-and must never be quoted as a timing result. A real Julia timing arm would need
-its own warmup loop and a stated count, as `BENCH_WARMUPS` does in Rust.
-
-## Runners are self checking
-
-Every runner exits nonzero when any arm's error exceeds its sanity gate, so a
-green exit is part of the evidence that a change is sound. The gates screen order
-unity wrongness, not precision. The one exception is the `zipup_treetn` arm of
-cases 3 and 4, gated at a hardcoded `5.0` because its error at the fixed output
-budget is genuinely of order one (README known issue 8), so for that arm the
-gate catches only a non-finite result or a runaway magnitude. Never loosen one
-silently to make a run pass.
-
-## Results are committed deliberately
-
-Files under `result/<profile>/` are generated by `scripts/run_all.sh <profile>`
-and committed on purpose, as the record of a full sweep on a stated machine. Do
-not hand edit them, and do not regenerate them as a side effect of an unrelated
-change. Use `OUT_DIR=/tmp/...` for probe runs. One profile belongs to one
-physical machine, named after the machine, and a sweep from a different machine
-goes to its own profile instead of overwriting another machine's numbers: wall
-times of the memory heavy arms differ by factors between machines (README known
-issue 10). Never put a hostname in a result file, the machine label and the
-hardware fields in `run.yaml` carry the identity.
-
-## Expensive benchmark inputs are cached
-
-Every runner with nontrivial deterministic input construction must cache that
-input and reuse it across algorithm comparisons and reruns. Input generation,
-cache loading, format conversion and patch preparation stay outside timed
-regions.
-
-- Use one shared ignored cache directory, `.cache/inputs/` by default.
-- Include every input-defining parameter, seed and an explicit input schema
-  version in the cache key. Bump the schema when generation semantics change.
-- Write cache entries atomically and treat an existing entry as immutable.
-- Validate loaded shapes, ranks, dtypes and shared index identities before use.
-- Record the cache key, hit or miss, load time and build time in each result.
-- Provide explicit refresh and cache-directory environment knobs. Never silently
-  accept a corrupt or incompatible entry.
-- Cache inputs only, never algorithm outputs or timed intermediates.
-- Tests whose subject is input generation or serialization must force a rebuild;
-  ordinary unit tests remain isolated and must not depend on persistent cache
-  state.
-
-## Upstream is pinned
-
-`Cargo.toml` pins every `tensor4all-rs` crate to one `rev`. Bump it deliberately,
-move all the crates together, and never mix backend features across them. A bump
-can change measured numbers, so it belongs in its own commit with a rerun.
+Read `README.md` before changing case definitions, timing boundaries, input generation, or environment settings.
+
+## Maintained scope
+
+The repository has exactly three benchmark cases: Fourier elementwise, Gaussian elementwise global versus patched, and Gaussian MPO-MPO contraction global versus patched. Do not add another case when an arm or metric in one of these cases answers the question.
+
+Gaussian inputs use independent two-dimensional interpolative decomposition per term, deterministic balanced pairwise add-and-compress, final relative-L2/SVD tolerance `1e-6`, and fixed patch cap 128. Do not add TCI generation, whole-mixture interpolation, isotropic controls, independent `R` sweeps, or a runtime patch-cap knob without an explicit design change.
+
+ACI residual tolerance is not an L2 tolerance. Keep its internal metric explicit and compare it with other arms only through the common achieved external error.
+
+## Timing
+
+Time only the benchmarked elementwise or contraction operation. Input construction, cache I/O, compression, conversion, patch preparation, output conversion, and accuracy evaluation stay outside timed regions.
+
+Run release builds with one pinned CPU core and single-threaded Rayon and BLAS. Warm up any JIT-compiled external implementation before timing it.
+
+## Results
+
+Files under `result/<profile>/` are generated by `scripts/run_all.sh <profile>` and committed deliberately. Do not edit them manually. Probe runs belong under `/tmp`. One profile belongs to one physical machine and contains no hostname.
+
+## Input cache
+
+Every expensive deterministic input must use the shared ignored `.cache/inputs/` directory by default.
+
+- Include every content-defining parameter, seed, upstream revision, and schema version in the key.
+- Write atomically and treat existing entries as immutable.
+- Validate loaded length, site dimensions, scalar type, and index compatibility.
+- Record key, hit or miss, load time, and build time.
+- Cache inputs only, never timed outputs.
+- Input-generation and serialization tests must force a rebuild.
 
 ## Pushing
 
@@ -76,7 +45,10 @@ together with the hardware behind them. Keep hostnames out of the description
 as well as out of the files: on a public repository a hostname leaks the
 operator's institution and location over time.
 
-## Prose style
+## Upstream
+
+All tensor4all-rs crates are pinned to one revision in `Cargo.toml`. Move every pin together. A pin change requires fresh measurements.
+
+## Prose
 
-No em dashes or en dashes anywhere in this repository's documents, comments or
-generated reports. Use commas, colons or separate sentences.
+Do not use em dashes or en dashes in documentation, comments, or generated reports.

Cargo.toml

@@ -12,7 +12,6 @@ tensor4all-treetn = { git = "https://github.com/tensor4all/tensor4all-rs", rev =
 tensor4all-itensorlike = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-itensorlike" }
 tensor4all-partitionedtt = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-partitionedtt" }
 tensor4all-partitionedtreetn = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-partitionedtreetn" }
-tensor4all-quanticstci = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-quanticstci" }
 tensor4all-interpolativeqtt = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-interpolativeqtt" }
 tensor4all-hdf5 = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-hdf5" }
 tensor4all-core = { git = "https://github.com/tensor4all/tensor4all-rs", rev = "9e9aedaebe0d3918b34dd399ff0981e337f3835b", package = "tensor4all-core" }