Refactor optimization pipeline.#520
Merged
Merged
Conversation
bimalgaudel
changed the title
…raction total mem.
…urable
- Add OptFor { Flops, Memsize } and ReorderSum { Reorder, NoReorder } in
a shared optimize/flags.hpp; thread them through single_term_opt and
the top-level optimize() API.
- Add index_to_extent_t alias for the type-erased provider used by the
public API; templated single_term_opt callers still pass callables
directly via the has_index_extent concept.
- Parallelize the outermost Sum's per-summand single-term optimization
with sequant::for_each; recurse sequentially on nested Sums.
- Have opt::reorder reuse pre-binarized eval nodes instead of rebuilding
them inside clusters().
- Harden opt_mixed_product placeholder labels (non-identifier prefix,
starts_with match, digit-walk suffix parse, asserted invariants).
- Fix std::forward misuse in flops_counter / memsize_counter that
prevented compilation under clang.
- Rename optimize/flags.hpp → optimize/options.hpp; move OptFor,
ReorderSum, and index_to_extent_t into it alongside a new
OptimizeOptions struct (opt_for / reorder / idx_to_extent with
sensible defaults).
- Collapse the six optimize() overloads into three (ExprPtr&,
ResultExpr&, ResultExpr&&), each taking OptimizeOptions{} by
default. Existing default-argument call sites are source-compatible.
- Stop including <SeQuant/core/optimize/single_term.hpp> from
optimize.hpp. The public optimize() API now needs only the
lightweight options header, so single_term.hpp (and its range-v3
transitives) no longer leak into consumers. This fixes the CI
unity-build failure in utilities/external-interface, where
range/v3/view/indirect.hpp's use of meta:: became ambiguous against
sequant::meta once a sibling .cpp in the same TU had issued
`using namespace sequant;`.
…nt optimize() API
bimalgaudel
force-pushed
the
gaudel/feature/refactor_optimization
branch
from
2dc2175 to
59f4b57
Compare
Krzmbrzl
reviewed
May 22, 2026
|
|
||
| // Overloads for backwards compatibility | ||
|
|
||
| /// \deprecated Use the \c OptimizeOptions overload instead. |
Collaborator
There was a problem hiding this comment.
We can (and maybe should) make use of the [[deprecated]] attribute: https://en.cppreference.com/cpp/language/attributes/deprecated
Member
Author
There was a problem hiding this comment.
I thought about that. I might add it.
…ummands Add minimal tests for the refactored optimize() API: - both OptFor::Flops and OptFor::Memsize binarize a product correctly - ReorderSum knob, and bare optimize() == explicit Reorder - parallel summand optimization (set_num_threads 1 vs 4) yields identical results, guarding the optimize_impl(..., parallel_outer=true) path Also document the thread-safety invariants of the parallel branch (work on per-task clones; binarize()'s lazy Index::label() cache write is sequenced after for_each joins) and explain the deliberate IndexSet-vs-vector container choice and the exact-equality scalar sentinel in the flop/memsize counters. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
evaleev
reviewed
May 23, 2026
Member
There was a problem hiding this comment.
Review: Refactor optimization pipeline
Solid, well-structured refactor. I did a focused thread-safety audit of the new parallel path and added minimal regression tests + documentation; details below. Pushed as a1d0264 on this branch.
Parallelism (optimize_impl(..., parallel_outer=true)) — SAFE as written, conditional
A four-angle audit (TensorCanonicalizer · TensorNetwork/bliss · Index/Expr/registry · for_each mechanics) found no data race, resting on two invariants that currently hold:
TensorNetworkclones every tensor on insertion, so the unsynchronizedmutablelazy caches (Expr::hash_value_,Index::label_/full_label_) are only ever written on thread-local clones.Index::operator==/<=>compare immutable members only, so building index sets over shared indices is safe.- The one step that lazily mutates shared
Index/Exprcaches —binarize()viaIndex::label()— runs in the sequential loop afterfor_eachjoins (optimize.cpp), never insidedo_term.
Canonicalizer/bliss global state is read-only during canonicalization; the swap counter is thread_local; the loop writes disjoint, pre-sized new_smands[i] slots.
I documented invariants (1)/(2) in a comment at the parallel branch so a future edit doesn't silently break them by moving a ->hash_value()/Index::label() call into do_term.
Caveats worth knowing:
- The default
Contextandcardinal_tensor_labelsmust be configured before enteringoptimize()(their writes are unsynchronized unlessSEQUANT_CONTEXT_MANIPULATION_THREADSAFE);optimize()only reads them. - On Linux/GCC builds
for_eachusesstd::execution::par_unseq, which ignoresnum_threads()— soset_num_threads(1)will not serialize this loop there (it does on the Apple-Clang/macOS build, which uses the manual-thread fallback).
Tests added (commit a1d0264)
OptimizeOptionsAPI: bothOptFor::FlopsandOptFor::Memsizebinarize a product correctly;ReorderSumknob; bareoptimize()== explicitReorder.- Parallel/sequential equivalence: a 3-summand sum optimized under
set_num_threads(1)vs(4)must produce identical results — the safety net for the invariants above.
Minor doc comments added
flops_countervsmemsize_counter: documented why flops needs<IndexSet>(concatenated operands repeat contracted indices) while memsize can use the default vector (per-operand, no dups).- Documented the exact-equality
== 1./!= 1.scalar sentinel convention in both counters.
Strengths (unchanged from the PR)
- Placeholder-label hardening (
@__opt_+ digit-by-digit parse with asserts) fixes the oldI_-prefixstd::stoi-on-raw-data()collision risk. reorderno longer re-binarizes (precomputed nodes threaded through, withSEQUANT_ASSERT(nodes.size() == expr.size())).- Clean extraction of
init_results/build_subnet_metadata(correctlyinline); deprecated bool overloads correctly drop their default args to avoid ambiguity with theOptimizeOptions = {}overload.
LGTM once CI is green.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.