Refactor index manipulation kernels around position-indexed subblocks by lkdvos · Pull Request #526 · QuantumKitHub/TensorKit.jl

lkdvos · 2026-09-03T15:18:48Z

Structural fix for #516 (adjoint permute! up to 80× slower than the plain path), superseding #519 and #520 and building on #518/#521.

What changes

StridedSubblocks: sector-independent, integer-indexed views into the flat data of a TensorMap in canonical fusion-tree order, with an optional lazy conjugation (op = identity/conj as type parameter). TreeSubblocks is the generic counterpart for any AbstractTensorMap, going through subblock.
TreeTransformers store only the mapping between subblock positions and recoupling coefficients (plus the two subblock structures), and are cached for every tensor type; the closure-based fallback and TrivialTreeTransformer are gone.
One kernel serves abelian and generic transformers, TensorMaps and other tensor types.
permute!/braid!/transpose! and TO.tensoradd! fold AdjointTensorMap sources and destinations (and conjA) into a conjsrc::Bool, relabeled p/levels and conjugated α/β; the flag is resolved into the view type only at the kernel entry, so everything stays type-stable.
BraidingTensor sources are converted before the transformer is chosen (the old overload could reach an untyped kernel).

Numbers (issue reproducer, adjoint vs plain permute!): fℤ₂ 1.30× → 1.09×, fℤ₂⊠U₁ 3.57× → 1.05×, SU₂ 4.44× → 1.01×, U₁ 2.23× → 1.11×; plain path unchanged within noise (+32 B from the extra Bool in the cache key).

Tests: adjoint source/destination/both for permute!/transpose!/braid! with accumulation, @tensor conj, BraidingTensor source, and a dot-based isometry check that would catch a wrongly conjugated recoupling matrix for complex sector scalar types. Benchmark suite gained adjoint = true permute variants.

Follow-ups (not here): keying transformers on sector structure only; passing conj into trace_permute!.

🤖 Generated with Claude Code

codecov · 2026-09-03T21:36:49Z

Codecov Report

❌ Patch coverage is 85.79545% with 25 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/tensors/blockiterators.jl	45.83%	13 Missing ⚠️
src/tensors/tensor.jl	56.25%	7 Missing ⚠️
src/tensors/treetransformers.jl	96.66%	2 Missing ⚠️
ext/TensorKitEnzymeExt/utility.jl	0.00%	1 Missing ⚠️
src/tensors/braidingtensor.jl	50.00%	1 Missing ⚠️
src/tensors/indexmanipulations.jl	98.52%	1 Missing ⚠️

Files with missing lines	Coverage Δ
src/TensorKit.jl	`17.24% <ø> (ø)`
src/tensors/abstracttensor.jl	`54.85% <100.00%> (ø)`
src/tensors/tensoroperations.jl	`96.44% <100.00%> (-0.10%)`	⬇️
ext/TensorKitEnzymeExt/utility.jl	`20.68% <0.00%> (-0.25%)`	⬇️
src/tensors/braidingtensor.jl	`87.58% <50.00%> (+0.65%)`	⬆️
src/tensors/indexmanipulations.jl	`91.13% <98.52%> (+1.56%)`	⬆️
src/tensors/treetransformers.jl	`96.34% <96.66%> (+1.49%)`	⬆️
src/tensors/tensor.jl	`81.28% <56.25%> (-2.32%)`	⬇️
src/tensors/blockiterators.jl	`43.69% <45.83%> (ø)`

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

lkdvos · 2026-09-06T07:16:36Z

Benchmark: `main` vs `ld-adjoint` (rusty, dedicated `rome` node, `--threads=4`)

benchpkg TensorKit --rev=main,ld-adjoint --bench-on=ld-adjoint --filter=indexmanipulations, i.e. both revisions ran this branch's suite including the new adjoint = true variants (permute!(C, A', p)). Ratio is main / ld-adjoint, so > 1 means this branch is faster.

Adjoint sources (the #516 case): 1.03–1.69× faster than main, and at parity with the plain TensorMap path on this branch (e.g. SU₂ Float64 [[1,3],[2,4]]: 1.27 ms → 0.78 ms next to 0.82 ms plain; on main the adjoint variant was 2.2× slower than plain).

Plain path: unchanged within noise. Two entries show a < 1 ratio (SU₂ Float64 [[1,3],[2,4]] 0.72 ± 0.19 and ℤ₂ [7264, 7264] 0.82 ± 0.03); both were re-measured locally against main at 4 threads and are at parity (0.63–0.67 vs 0.67 ms, and 49–50 vs 50 ms). The ℤ₂ adjoint variant executes the identical kernel (real eltype ⇒ op = identity) and is at parity in the table as well, and the untouched Trivial path shows the same kind of scatter (0.83 ± 0.14), so these are between-process memory-placement effects on the 400 MB transposes rather than code differences.

Full table

	main	ld-adjoint	main / ld-adjoint
indexmanipulations/permute/permute/("ComplexF64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[1, 3], [2, 4]]")	1.37 ± 0.22 ms	1.19 ± 0.23 ms	1.15 ± 0.29
indexmanipulations/permute/permute/("ComplexF64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[1, 3], [2, 4]]", "adjoint")	1.59 ± 0.085 ms	1.55 ± 0.32 ms	1.03 ± 0.22
indexmanipulations/permute/permute/("ComplexF64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[4, 2, 3], [1]]")	1.34 ± 0.25 ms	1.22 ± 0.22 ms	1.1 ± 0.28
indexmanipulations/permute/permute/("ComplexF64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[4, 2, 3], [1]]", "adjoint")	1.32 ± 0.12 ms	0.939 ± 0.15 ms	1.41 ± 0.26
indexmanipulations/permute/permute/("Float64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[1, 3], [2, 4]]")	0.59 ± 0.04 ms	0.822 ± 0.21 ms	0.717 ± 0.19
indexmanipulations/permute/permute/("Float64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[1, 3], [2, 4]]", "adjoint")	1.27 ± 0.19 ms	0.775 ± 0.12 ms	1.64 ± 0.35
indexmanipulations/permute/permute/("Float64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[4, 2, 3], [1]]")	0.811 ± 0.089 ms	0.665 ± 0.059 ms	1.22 ± 0.17
indexmanipulations/permute/permute/("Float64", "SU2Irrep", "[48, 48, 48, 48]", "[1.0, 1.0, 1.0, 1.0]", "Any[[4, 2, 3], [1]]", "adjoint")	0.684 ± 0.038 ms	0.5 ± 0.023 ms	1.37 ± 0.098
indexmanipulations/permute/permute/("Float64", "SU2Irrep", "[512, 512]", "[1.0, 1.0]", "Any[[2, 1], Any[]]")	0.144 ± 0.014 ms	0.0849 ± 0.0089 ms	1.69 ± 0.24
indexmanipulations/permute/permute/("Float64", "Trivial", "[43408, 1216]", "nothing", "Any[[2, 1], Any[]]")	0.0453 ± 0.0017 s	0.0545 ± 0.0086 s	0.832 ± 0.14
indexmanipulations/permute/permute/("Float64", "Trivial", "[7264, 7264]", "nothing", "Any[[2, 1], Any[]]")	0.0557 ± 0.0012 s	0.0548 ± 0.0011 s	1.02 ± 0.029
indexmanipulations/permute/permute/("Float64", "Z2Irrep", "[43408, 1216]", "[0.5, 0.5]", "Any[[2, 1], Any[]]")	27.6 ± 1.6 ms	27.3 ± 0.8 ms	1.01 ± 0.066
indexmanipulations/permute/permute/("Float64", "Z2Irrep", "[43408, 1216]", "[0.5, 0.5]", "Any[[2, 1], Any[]]", "adjoint")	27.7 ± 2.1 ms	27.8 ± 2.7 ms	0.995 ± 0.12
indexmanipulations/permute/permute/("Float64", "Z2Irrep", "[7264, 7264]", "[0.5, 0.5]", "Any[[2, 1], Any[]]")	22.9 ± 0.47 ms	28 ± 0.76 ms	0.819 ± 0.028
indexmanipulations/permute/permute/("Float64", "Z2Irrep", "[7264, 7264]", "[0.5, 0.5]", "Any[[2, 1], Any[]]", "adjoint")	23 ± 0.41 ms	23.1 ± 0.45 ms	0.994 ± 0.026

🤖 Generated with Claude Code

* Intercept `TO.tensoradd!` instead of `TensorKit.add_transform!` The blockwise `tensoradd` implementations were only reachable through `TensorKit`'s internals: `TO.tensoradd!` delegated to `permute!`, and the `add_transform!` methods defined here caught the kernel. Neither is a contract `TensorKit` owes us, and both stopped holding on TensorKit's index-manipulation refactor (QuantumKitHub/TensorKit.jl#526), where `TO.tensoradd!` calls the braid kernel directly and `add_transform!` gained a `conjsrc` argument. Nothing errors in that case: dense block tensors survive via TensorKit's generic subblock fallback, while sparse ones silently produce zeros, because the fallback writes through blocks a sparse container never materialized. Implement `TO.tensoradd!` for the block tensor types instead. That is the public entry point `@tensor` lowers to, so it is reachable no matter how TensorKit arranges its kernels, and `conjA` is handled explicitly rather than relying on it having been unwrapped into an adjoint beforehand. The mixed block/plain methods take the concrete `TensorMap` so that `(BlockTensorMap, SparseBlockTensorMap)` and the reverse resolve to the general method rather than being ambiguous. The added testset calls `TO.tensoradd!` directly rather than through `@tensor`, so this stays covered independently of TensorKit's routing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * Pin the planar entry points in the tests `planaradd!`, `planartrace!` and `planarcontract!` are reached through `transpose!`, `trace_permute!` and `contract!`, so they currently land on methods defined here and are not affected by TensorKit#526. They rest on the same undocumented delegation that broke `tensoradd!` though, and there was no planar coverage here at all, so pin them directly. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

lkdvos · 2026-09-17T18:13:09Z

All comments should now be addressed again, I think I also managed to simplify slightly more because of @Jutho's comments, in particular the abelian case now no longer needs to refer to the "adjoint spaces" and simply loops over the source fusiontrees and exchanges them on the fly, which at least reads a lot easier.

For the non-abelian case I still am doing it that way because swapping splitting and fusiontree in the FusionBlock requires rebuilding the block array and then sorting it (in order to have a unique cached value for the unitary generated there), so there's a lot more subtle annoyances to get rid of this. I think in this case that is just complexity that is required... :(

Jutho · 2026-09-18T13:26:29Z

+        f = gettokenvalue(fusiontrees(iter.t), i)
+        return f => iter.structure[i], i + 1
+    end
+end


Reposting as a new comment, so that it does not get lost:

I guess I am wondering to what extent this specialization is necessary. As far as I can tell, subblocks is not really used in any performance critical code, they all go via StridedSubblocks or TreeSubblocks directly. The subblocks function and the associated SubblockIterator seem mostly user convenience functions, for getting and setting data in the tensor. So I don't know if we really need to make this more complicated for negligible performance gain.

To address this more precisely, this is almost true, in the sense that for example operations that mix diagonal tensors and regular tensors would still end up with subblocks being called, and I do see future work benefiting from the knowledge that subblocks is a performant primitive to build around, e.g. for twist or flip implementations.

Jutho · 2026-09-18T13:26:59Z

Ok, I finally managed to make my way through. Looks really great. I have four final questions or suggestions, but otherwise fully approve.

"Abelian" is ambiguous for sectors: it can refer either to the fusion of two sectors having a unique result, or to the commutativity of the fusion rules. The transformer is selected on `FusionStyle(I) == UniqueFusion()`, so name it after that. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`StridedSubblocks` and `TreeSubblocks` no longer apply `identity`/`conj` to every view. Instead `conjsrc` is threaded through `add_transform_kernel!` into `_add_transform_block!`, where it is handed to `TO.tensoradd!` as its `conjA` argument, at the single-tree call and when packing a multi-tree block. This drops a type parameter from both collections, so the kernel compiles to one instance per (storage, numind) rather than one per conjugation. The runtime flag is free: `flag2op` is union-split, and `conj` of a real-eltype `StridedView` is a type-level no-op, which also makes the previous `scalartype(t) <: Real` guard redundant. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

lkdvos · 2026-09-19T20:49:12Z

Thanks for the careful review, I think all of these comments were helpful to further simplify and improve the implementation, which I've hopefully managed to achieve here.

In particular, your comment about the SubblockIterator specialization made me realize that the TreeSubblocks were functionally the same structure as that, and so I managed to merge these implementations, just adding the "token-based" access to the original. This has further reduced some code at otherwise zero additional cost, which is always nice 😄.

As for whether the TensorMap specialization is still worth it, I would argue yes, mostly because there are still other parts of the code that iterate fusiontrees and look at the data, such as flip and twist, and while I don't want to delay this PR further, it would definitely be useful to rewrite these in terms of the subblocks as well, further bypassing some symmetry overhead.

I'm hoping all tests pass, and if so might merge this Monday morning (ET) if no further comments, or unless someone merges this first.

Jutho · 2026-09-19T21:24:34Z

+@noinline _throw_subblock_bounds(iter, i) = throw(BoundsError(iter, i))
+@noinline _throw_subblock_missing(f) = throw(SectorMismatch(lazy"fusion tree pair $f is not present"))
+
+@propagate_inbounds function Base.getindex(iter::SubblockIterator{<:TensorMap}, i::Int)


This is now relying on iter being constructed as SubblockIterator(t). What happens if someone constructs SubblockIterator(t::TensorMap, fusiontrees(t)). This method is still being called and will fail, no?

Yes, although that is also true for someone calling SubblockIterator(t::AbstractTensorMap, nothing), in the sense that you really should not be doing that? I can try and restrict the method even further, but since this should all be internals I don't know if that makes too much sense 😄

Ok, fair enough.

Jutho · 2026-09-19T21:42:46Z

By the way, did you rerun the benchmarks from the original issue with this latest iteration of the PR?

lkdvos · 2026-09-20T12:47:33Z

Issue #516 reproducer, rerun against the final state of this PR

main @ 7a187c4 vs ld-adjoint @ 6b2dd79, rusty rome node, julia 1.12.7, single-threaded, @benchmark samples=500 seconds=10. The ratio is in-place adjoint permute! over plain TensorMap permute! within the same revision, so 1.00 means an adjoint source costs nothing extra.

sector type	ratio on `main`	ratio here	adjoint allocs, `main` → here
`fℤ₂`	1.31×	1.09×	134 → 68
`fℤ₂ ⊠ U₁`	3.82×	1.01×	675 → 320
`U₁`	2.01×	1.11×	299 → 145
`SU₂`	3.92×	1.01×	7096 → 1926
`fℤ₂ ⊠ U₁ ⊠ SU₃`	19.06×	0.99×	7873 → 486

In every case the adjoint path now allocates exactly what the plain path allocates and runs at the same speed. The plain path itself did not move (SU₂ 0.183 → 0.187 ms, fℤ₂ ⊠ U₁ ⊠ SU₃ 0.0296 → 0.0301 ms), so this is the adjoint side coming down to the baseline rather than the baseline drifting up. In absolute terms the adjoint column goes SU₂ 0.718 → 0.189 ms and fℤ₂ ⊠ U₁ ⊠ SU₃ 0.564 → 0.030 ms.

Two notes on the baseline: main already carries #518 and #521, so the comparison is against the post-hash-fix numbers from the issue thread, not the original 82×. And the U₁ row is not in the issue script — I added it with a guessed space so the U₁ figure quoted in the PR description has a counterpart.

Full suite (linalg + indexmanipulations + tensornetworks) on both revisions is still running; I'll post the table when it lands.

* Draft changelog for v0.17.2 Consolidates the Unreleased section (which already included the real entries added by #526/#532 on merge) with entries for the remaining PRs merged since v0.17.1 (#487-#535), and retitles it as 0.17.2. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Bump version to v0.17.2 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Fix confirmed small bugs from the pre-release audit - isunitspace: require dim(V) == 1 for GenericUnit sectors (#537) - GradedSpace ⊕/supremum: check unit homogeneity of the result (#538) - isconj(::ComplexSpace): return isdual(V) instead of always true (#539) - multi_associator: return a vector, not a scalar, on early-exit for GenericFusion (#540) - split(f, 0): use leftunit(f.coupled) instead of indexing an empty uncoupled tuple (#541) - repartition: return a Pair in the identity branch, matching every other branch (#542) - Mooncake scalar_pullback: accumulate into the tangent instead of overwriting it (#543) - rand/randn/randexp/randisometry(rng, T, space): fix one(domain) typo (#544) - pinv(::DiagonalTensorMap): fix inverted atol/rtol defaulting and empty-tensor throw (#545) - t1 / t2: promote to a float scalartype, matching t1 \ t2 (#546) Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Add changelog entry for the audit bugfixes Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Address fable review findings on the audit bugfixes - split(f, 0): also guard innerlines_extended construction, which still indexed the empty uncoupled tuple for a 0-leg tree - pinv(::DiagonalTensorMap): use eps (not sqrt(eps)) for the default rtol, matching _default_rtol's convention and dense LinearAlgebra.pinv Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Address tuicr review comments on the audit bugfixes - pinv(::DiagonalTensorMap): reuse _default_rtol instead of duplicating its formula - Add regression tests for split(f, 0) on a genuine 0-leg tree and for multi_associator's early-exit branch on a GenericFusion sector Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * fix planar issues after MPSKit test rerun * harden Mooncake scalar pullback --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

lkdvos force-pushed the ld-adjoint branch from 02a2fbe to 5ece9d1 Compare September 9, 2026 01:48

This was referenced Sep 9, 2026

Add dedicated cached treebraider for AdjointTensorMap #519

Closed

Don't materialize an AdjointTensorMap in TO.tensoradd! when conjA = true #520

Closed

Jutho reviewed Sep 9, 2026

View reviewed changes

Comment thread src/tensors/blockiterators.jl

Jutho reviewed Sep 9, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

lkdvos force-pushed the ld-adjoint branch from fd8d295 to 48a30c3 Compare September 9, 2026 20:27

lkdvos marked this pull request as ready for review September 10, 2026 13:52

lkdvos requested a review from Jutho September 10, 2026 13:52

Jutho reviewed Sep 12, 2026

View reviewed changes

Comment thread benchmark/TensorKitBenchmarks/indexmanipulations/IndexManipulationBenchmarks.jl

Jutho reviewed Sep 14, 2026

View reviewed changes

Comment thread src/tensors/blockiterators.jl Outdated

Jutho reviewed Sep 14, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl

Jutho reviewed Sep 14, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

Jutho reviewed Sep 14, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

Jutho reviewed Sep 14, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

lkdvos force-pushed the ld-adjoint branch from 48a30c3 to a568761 Compare September 15, 2026 15:45

lkdvos mentioned this pull request Sep 15, 2026

Intercept TO.tensoradd! instead of TensorKit.add_transform! QuantumKitHub/BlockTensorKit.jl#78

Merged

Jutho reviewed Sep 16, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl

Jutho reviewed Sep 16, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

Jutho reviewed Sep 16, 2026

View reviewed changes

Comment thread src/tensors/blockiterators.jl Outdated

Jutho reviewed Sep 16, 2026

View reviewed changes

Comment thread src/tensors/tensor.jl Outdated

Jutho reviewed Sep 16, 2026

View reviewed changes

Comment thread src/tensors/tensoroperations.jl Outdated

lkdvos force-pushed the ld-adjoint branch from 0a39450 to 3b4054d Compare September 16, 2026 18:51

Jutho reviewed Sep 17, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

kshyatt reviewed Sep 17, 2026

View reviewed changes

Comment thread ext/TensorKitEnzymeExt/utility.jl Outdated

kshyatt reviewed Sep 17, 2026

View reviewed changes

Comment thread src/tensors/blockiterators.jl Outdated

kshyatt reviewed Sep 17, 2026

View reviewed changes

Comment thread src/tensors/treetransformers.jl Outdated

Jutho reviewed Sep 17, 2026

View reviewed changes

Comment thread src/tensors/tensor.jl Outdated

lkdvos added 4 commits September 17, 2026 11:13

fix (unrelated) docstring sentence

e8b083c

all has_array_view

1977293

clamp -> min

2366bb2

docstring improvements

e0fddbf

lkdvos force-pushed the ld-adjoint branch from cd40007 to e0fddbf Compare September 17, 2026 15:13

lkdvos commented Sep 17, 2026

View reviewed changes

Comment thread src/tensors/treetransformers.jl Outdated

lkdvos commented Sep 17, 2026

View reviewed changes

Comment thread src/tensors/treetransformers.jl Outdated

simplify treetransformer implementations

11998d1

lkdvos enabled auto-merge (squash) September 17, 2026 18:13

lkdvos linked an issue Sep 17, 2026 that may be closed by this pull request

Permuting an AdjointTensorMap seems much slower than it should be #516

Closed

Jutho reviewed Sep 18, 2026

View reviewed changes

Comment thread src/tensors/treetransformers.jl

Jutho reviewed Sep 18, 2026

View reviewed changes

Comment thread src/tensors/treetransformers.jl Outdated

Jutho reviewed Sep 18, 2026

View reviewed changes

Comment thread src/tensors/indexmanipulations.jl Outdated

Jutho reviewed Sep 18, 2026

View reviewed changes

lkdvos disabled auto-merge September 18, 2026 22:38

lkdvos and others added 3 commits September 19, 2026 15:20

remove TreeSubblocks and route through StridedSubblocks

6b2dd79

Jutho reviewed Sep 19, 2026

View reviewed changes

Jutho approved these changes Sep 19, 2026

View reviewed changes

lkdvos merged commit bfca561 into main Sep 20, 2026
65 of 67 checks passed

lkdvos deleted the ld-adjoint branch September 20, 2026 13:03

This was referenced Sep 20, 2026

Release v0.17.2 #550

Merged

planarcontract! rejects a basic zero-contracted-index planar contraction as non-cyclic #551

Closed

Conversation

lkdvos commented Sep 3, 2026

Uh oh!

codecov Bot commented Sep 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

lkdvos commented Sep 6, 2026

Benchmark: main vs ld-adjoint (rusty, dedicated rome node, --threads=4)

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lkdvos commented Sep 17, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Jutho Sep 18, 2026

Choose a reason for hiding this comment

Uh oh!

lkdvos Sep 19, 2026

Choose a reason for hiding this comment

Uh oh!

Jutho commented Sep 18, 2026

Uh oh!

lkdvos commented Sep 19, 2026

Uh oh!

Jutho Sep 19, 2026

Choose a reason for hiding this comment

Uh oh!

lkdvos Sep 19, 2026

Choose a reason for hiding this comment

Uh oh!

Jutho Sep 19, 2026

Choose a reason for hiding this comment

Uh oh!

Jutho commented Sep 19, 2026

Uh oh!

lkdvos commented Sep 20, 2026

Issue #516 reproducer, rerun against the final state of this PR

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov Bot commented Sep 3, 2026 •

edited

Loading

Benchmark: `main` vs `ld-adjoint` (rusty, dedicated `rome` node, `--threads=4`)