Repository navigation
perf(tableau): optimize measurement for generalized tableau - #56
Merged
Merged
Conversation
|
Contributor
There was a problem hiding this comment.
Pull request overview
This PR optimizes GeneralizedTableau::measure() (the dominant runtime hotspot for generalized tableau simulations with multiple T gates) by reducing allocations and HashMap work, and by simplifying overlap computation to accumulate only the real component.
Changes:
- Reworks
GeneralizedTableau::measure()to avoid cloning coefficients, add a case-b fast path, and optimize case-a partitioning/drain behavior. - Removes
trim_coefficients_for_measurement()(now inlined/handled inmeasure()’s case-b path). - Adds a new Criterion benchmark (
measure-scaling) and an ad-hoc profiling example, plus autotune docs/logs.
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
crates/ppvm-tableau/src/measure.rs |
Main performance refactor: mem::replace-based ownership transfer, case-b fast path (no HashMap), real-only overlap accumulation, retain-based A/B partition for case-a. |
crates/ppvm-tableau/src/data.rs |
Removes now-unused trim_coefficients_for_measurement() helper. |
crates/ppvm-tableau/benches/measure-scaling.rs |
Adds Criterion benchmark to track measurement scaling at different T-gate counts. |
crates/ppvm-tableau/Cargo.toml |
Registers the new measure-scaling benchmark target. |
crates/ppvm-tableau/examples/profile_measure.rs |
Adds an ad-hoc profiling example for end-to-end measurement time. |
docs/autotune/2026-04-12-measure-parallel/metric.toml |
Records perf results across micro-optimization iterations. |
docs/autotune/2026-04-12-measure-parallel/log.md |
Logs profiling breakdown and optimization notes. |
docs/autotune/2026-04-12-measure-parallel/eliminate-clone/prompts.md |
Captures the prompt/plan used for the “eliminate clone” iteration. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| let mut t = tab.fork(Some(42)); | ||
| let start = Instant::now(); | ||
| for q in 0..n_qubits { | ||
| t.measure(q); |
| || tab.fork(Some(42)), | ||
| |t| { | ||
| for q in 0..n_qubits { | ||
| t.measure(q); |
Comment on lines
+194
to
+196
| // Partition into A (k-bit=0) and B (k-bit=1) via drain, then merge. | ||
| // This avoids allocating a separate b_keys Vec and eliminates | ||
| // individual HashMap removes (which require probe+shift). |
Comment on lines
+9
to
+13
| | Component | 256 coeffs | 1024 coeffs | 4096 coeffs | 16384 coeffs | | ||
| |-----------|-----------|------------|------------|-------------| | ||
| | decomp | 1.8% | 0.6% | 0.1% | <0.1% | | ||
| | clone (coeffs→HashMap) | 23.9% | 26.8% | 36.1% | 26.3% | | ||
| | overlap loop | 21.1% | 22.7% | 17.6% | 22.4% | |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR optimizes GeneralizedTableau::measure() in ppvm-tableau, targeting the dominant end-to-end runtime cost for circuits with higher T-gate counts by reducing allocation/HashMap overhead and simplifying overlap accumulation.
Changes:
- Adds a case-b fast path (
stab_anticomm_bits == 0) that avoids building aHashMapand computes overlap/filtering directly from coefficient entries. - Refactors overlap computation to accumulate only the real component (and skips odd phases in the case-b fast path).
- Improves case-a drain/merge by partitioning with
HashMap::retaininstead of collecting/removing keys individually; adds a new Criterion bench and profiling/example + autotune docs.
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.
Show a summary per file
| File | Description |
|---|---|
crates/ppvm-tableau/src/measure.rs |
Implements the measurement micro-optimizations (case-b fast path, real-only overlap, retain-based partitioning) and factors overlap/merge helpers. |
crates/ppvm-tableau/src/data.rs |
Exposes compute_phase_with_mask_static (and the rayon threshold) for reuse; removes now-unused measurement helper methods. |
crates/ppvm-tableau/benches/measure-scaling.rs |
Adds a Criterion benchmark to measure scaling across T-gate counts. |
crates/ppvm-tableau/Cargo.toml |
Registers the new measure-scaling benchmark target. |
crates/ppvm-tableau/examples/profile_measure.rs |
Adds an ad-hoc profiling example for measurement scaling. |
docs/autotune/2026-04-12-measure-parallel/* |
Captures autotuning metrics, logs, and prompts for the measurement optimization work. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Optimize GeneralizedTableau::measure(), which dominates end-to-end runtime at 85-90% for circuits with 8+ T gates, through five micro-optimizations: 1. Eliminate coefficient clone: replace clone().into_iter() with mem::replace() — self.coefficients is never read after the clone. 2. Case_b fast path: when Z is already a stabilizer (stab_anticomm_bits == 0), skip HashMap construction entirely. The overlap is self-pairing (conj(c)*c = |c|^2), so we work directly on the Vec with no HashMap needed. 3. Two-pass case_b: iterate by reference for overlap, then filter directly into coefficients — avoids the push-all-then-retain pattern. 4. Real-only overlap accumulation: compute only Re(z_overlap) since the imaginary part is always ~0. For case_b, skip odd-phase entries entirely (zero contribution to real part). 5. Retain-based drain: replace collect-b_keys + individual-remove with HashMap::retain to partition A/B entries in a single pass, avoiding expensive per-entry HashMap removes. Also: extract overlap and merge logic into helper methods, remove unused compute_phase_with_mask wrapper and trim_coefficients_for_measurement, make compute_phase_with_mask_static pub(crate). Benchmark results (85 qubits, full measurement of all qubits): | Benchmark | Before | After | Improvement | |------------------------|----------|----------|-------------| | MSD-fused (5 T gates) | 63.8 µs | 57.5 µs | -9.9% | | 8 T gates (256 coeffs) | 23.1 µs | 15.0 µs | -34.8% | | 10 T gates (1024) | 49.5 µs | 36.0 µs | -27.3% | | 12 T gates (4096) | 156.6 µs | 118.6 µs | -24.3% | Note: rayon parallelization was tested for measurement overlap and B→A merge but showed regressions at all sizes (16K-65K coefficients). The per-element work is too cheap (~5ns) for rayon overhead, and the sequential HashMap accumulation dominates. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Roger-luo
force-pushed
the
autotune/measure-parallel
branch
from
April 13, 2026 01:19
3801f4f to
6e4a193
Compare
Roger-luo
added a commit
that referenced
this pull request
Apr 13, 2026
Optimize GeneralizedTableau::measure(), which dominates end-to-end runtime at 85-90% for circuits with 8+ T gates, through five micro-optimizations: 1. Eliminate coefficient clone: replace clone().into_iter() with mem::replace() — self.coefficients is never read after the clone. 2. Case_b fast path: when Z is already a stabilizer (stab_anticomm_bits == 0), skip HashMap construction entirely. The overlap is self-pairing (conj(c)*c = |c|^2), so we work directly on the Vec with no HashMap needed. 3. Two-pass case_b: iterate by reference for overlap, then filter directly into coefficients — avoids the push-all-then-retain pattern. 4. Real-only overlap accumulation: compute only Re(z_overlap) since the imaginary part is always ~0. For case_b, skip odd-phase entries entirely (zero contribution to real part). 5. Retain-based drain: replace collect-b_keys + individual-remove with HashMap::retain to partition A/B entries in a single pass, avoiding expensive per-entry HashMap removes. Also: extract overlap and merge logic into helper methods, remove unused compute_phase_with_mask wrapper and trim_coefficients_for_measurement, make compute_phase_with_mask_static pub(crate). Benchmark results (85 qubits, full measurement of all qubits): | Benchmark | Before | After | Improvement | |------------------------|----------|----------|-------------| | MSD-fused (5 T gates) | 63.8 µs | 57.5 µs | -9.9% | | 8 T gates (256 coeffs) | 23.1 µs | 15.0 µs | -34.8% | | 10 T gates (1024) | 49.5 µs | 36.0 µs | -27.3% | | 12 T gates (4096) | 156.6 µs | 118.6 µs | -24.3% | Note: rayon parallelization was tested for measurement overlap and B→A merge but showed regressions at all sizes (16K-65K coefficients). The per-element work is too cheap (~5ns) for rayon overhead, and the sequential HashMap accumulation dominates. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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
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.
Summary
Optimize the
GeneralizedTableau::measure()method which dominates end-to-end runtime at 85-90% for circuits with 8+ T gates. Five micro-optimizations that compose for 24-35% total improvement on measurement:clone().into_iter()withmem::replace()sinceself.coefficientsis never read after the clonestab_anticomm_bits == 0), skip HashMap construction entirely — the overlap is self-pairing (conj(c)*c = |c|^2)Re(z_overlap)since the imaginary part is always ~0. For case_b, skip odd-phase entries entirely (zero contribution)collect-b_keys + individual-removewithHashMap::retainto partition A/B entries in a single passBenchmark results (85 qubits, full measurement)
Test plan
cargo test -p ppvm-runtime -p ppvm-tableau -p ppvm-sym)test_t_gate_measurement_statisticsverifies measurement correctness🤖 Generated with Claude Code