Repository navigation

Unified threading - #44

Open

hakkelt wants to merge 75 commits into

kul-optec:masterfrom

hakkelt:unified-threading

Open

Unified threading#44
hakkelt wants to merge 75 commits into
kul-optec:masterfrom
hakkelt:unified-threading

Commits on Nov 2, 2026
- Port test coverage expansion and refactors beyond upstream GPU-support merge
  hakkelt
  committed
  241432e
- Remove AcceleratedDCTs from test/Project.toml deps
  hakkelt
  committed
  cdbf7c4
- minor fixes
  hakkelt
  committed
  4361f81
- Replace hand-rolled thread budgeting with NestedThreading.jl
  
  hakkelt
  and
  claude
  committed
  4ff137c
- Budget the threaded OperatorBroadCast adjoint loop
  
  hakkelt
  and
  claude
  committed
  f237635
- Activate the GPU test environment only when a GPU testitem runs
  
  hakkelt
  and
  claude
  committed
  74cab04
- Add threading policy infrastructure and thread the elementwise operators
  
  hakkelt
  and
  claude
  committed
  af87114
- Thread the DCAT block loop, gated on a measured two-part policy
  
  hakkelt
  and
  claude
  committed
  f854bb6
- Declare threading for the subpackage operators and fix a batching regression
  
  hakkelt
  and
  claude
  committed
  6e0becd
- Make `threaded` a per-operator, measurement-driven permission across all operators
  
  hakkelt
  and
  claude
  committed
  8831e56
- Fix Variation adjoint BoundsError on length-2 dimensions
  
  hakkelt
  and
  claude
  committed
  7525c8f
- Remove THREADING_PLAN.md
  
  hakkelt
  and
  claude
  committed
  7eea378
- Benchmark threaded paths, skip them on CI, split suites into files
  
  hakkelt
  and
  claude
  committed
  e81491e
- Apply Runic formatting
  
  hakkelt
  and
  claude
  committed
  cdf003b
- Remove temporarily sourced NestedThreading from Project.toml
  hakkelt
  committed
  efcec21
- Fix Windows LTS JET dispatch and Documentation CI failures
  hakkelt
  committed
  449c1d5
- Cover the threading-migration diff and fix a DiagOp storage-copy data-loss bug
  
  hakkelt
  and
  claude
  committed
  d4a2f44
- Cover Scale/AdjointOperator threading traits and fix a systemic SpreadingBatchOp gap
  
  hakkelt
  and
  claude
  committed
  6355f8e
- Cover OperatorWrapper threading traits, HCAT's tuple-idxs branches, and BroadCast's threaded construction
  
  hakkelt
  and
  claude
  committed
  88742c8
- Cover Sum and VCAT threading traits
  
  hakkelt
  and
  claude
  committed
  fda4817
- Cover LMatrixOp and MyLinOp copy_operator
  
  hakkelt
  and
  claude
  committed
  f993846
- Give Conv/Xcorr real FFTW threading; fix a factually wrong comment
  
  hakkelt
  and
  claude
  committed
  0991274
- Support storage_type in Conv/Xcorr copy_operator; fix a Runic-style return-before-loop
  
  hakkelt
  and
  claude
  committed
  6eac5e2
- Support storage_type in every remaining _copy_operator_impl that refused it
  
  hakkelt
  and
  claude
  committed
  9986e10
- Fix DSPOperators/NFFTOperators docstrings; revert accidental version bumps
  
  hakkelt
  and
  claude
  committed
  afa0740
- Remove stale 'previous implementation' comments; fix a factually wrong one
  
  hakkelt
  and
  claude
  committed
  192602a
- Fix benchmark scripts; add Julia version to threading measurement context
  
  hakkelt
  and
  claude
  committed
  479d756
- Give threaded HCAT/VCAT block loops their own serial-only block copy
  
  hakkelt
  and
  claude
  committed
  749f4b8
- Fix LBFGS copy_operator throwing on an unsupported threaded=true request
  
  hakkelt
  and
  claude
  committed
  16814f0
- Give MatrixOp/LMatrixOp real BLAS threading support
  
  hakkelt
  and
  claude
  committed
  98f199c
- Give SoftMax reduction-based threading via Polyester; measure its crossover
  
  hakkelt
  and
  claude
  committed
  a39ef69
- Unify nonlinear operators' threaded/serial mul! via FastBroadcast's Th singleton
  
  hakkelt
  and
  claude
  committed
  8623b5f
- Fix two CI test regressions from the threading changes
  
  hakkelt
  and
  claude
  committed
  e9d66ef
- Fix JET @test_opt regression: stop kwarg-splatting threaded through MatrixOp/LMatrixOp
  
  hakkelt
  and
  claude
  committed
  44de691
- Fix another CI flake: BLAS threading isn't bit-identical on every platform
  
  hakkelt
  and
  claude
  committed
  7ee83e8
- Fix threading-migration copy_operator bugs found in code review
  
  hakkelt
  and
  claude
  committed
  8052fe7
- Simplify threading-migration copy_operator code from review
  
  hakkelt
  and
  claude
  committed
  8da06ff
- Close real coverage gaps in batch threading and nonlinear operator constructors
  
  hakkelt
  and
  claude
  committed
  8414226
- Add missing _copy_operator_impl for WaveletOp
  
  hakkelt
  and
  claude
  committed
  eb6736d
- Benchmark both single- and multithreaded execution, not an ambiguous default
  
  hakkelt
  and
  claude
  committed
  8a58611
- Test the unreleased GPUEnv fix (PR #6) in CI, pinned to 0.2.2
  
  hakkelt
  and
  claude
  committed
  748fb05
- Revert temporary GPUEnv 0.2.1 workarounds now that 0.2.2 is released
  
  hakkelt
  and
  claude
  committed
  08e67b2
- Drop remaining redundant -single/-threaded splits (DFT, NFFTOp, elementwise ops)
  
  hakkelt
  and
  claude
  committed
  b97713a
- Fix time and memory regressions in the unified threading benchmarks
  
  hakkelt
  and
  claude
  committed
  8c0adc4
- Normalise NFFTOp storage traits and close two coverage gaps
  
  hakkelt
  and
  claude
  committed
  c85d3e0
- Drop the redundant and the never-threaded BroadCast benchmark splits
  
  hakkelt
  and
  claude
  committed
  53fb290
- fix: preserve threading flag through DCAT/HCAT reconstruction and DFT thread restore
  
  hakkelt
  and
  claude
  committed
  d1fcc0f
- fix: eliminate JET runtime-dispatch regressions across constructors and DSP mul!
  
  hakkelt
  and
  claude
  committed
  2b2d9cd
- fix: bind all size() type parameters for VCAT/HCAT/DCAT
  
  hakkelt
  and
  claude
  committed
  7d5e678
- fix: construct FastBroadcast thread markers as instances
  
  hakkelt
  and
  claude
  committed
  6dd6b87
- fix: keep operator shape introspection concrete for JET
  
  hakkelt
  and
  claude
  committed
  69d5869
- fix: index Variation output linearly in mul!
  
  hakkelt
  and
  claude
  committed
  7463a8c
- Revert "fix: bind all size() type parameters for VCAT/HCAT/DCAT"
  
  hakkelt
  and
  claude
  committed
  6821ae5
- fix: reattach the ndims docstring
  
  hakkelt
  and
  claude
  committed
  2812516
- fix: drop the type-form _fbbool methods
  
  hakkelt
  and
  claude
  committed
  6f7df24
- test: tag threading testitems :threading
  
  hakkelt
  and
  claude
  committed
  330f299
- fix(calculus): make the threaded block loops of DCAT/VCAT/HCAT safe
  
  hakkelt
  and
  claude
  committed
  9371271
- fix(operators): copy_operator for a scalar DiagOp and for LBFGS
  
  hakkelt
  and
  claude
  committed
  15cffea
- fix(threading): no CPU-thread machinery for device-backed operators
  
  hakkelt
  and
  claude
  committed
  8bd0ad8
- fix(batching): == compares the batch dimensions
  
  hakkelt
  and
  claude
  committed
  335ea5c
- test: make the threading assertions hold in a single-threaded session
  
  hakkelt
  and
  claude
  committed
  da82d3a
- test(wavelets): activate the GPU environment for WaveletOp (GPU)
  
  hakkelt
  and
  claude
  committed
  7153be9
- refactor(DiagOp): take the copy's storage from domain_array_type
  
  hakkelt
  and
  claude
  committed
  f1857ee
- perf(DCAT): copy each block once when rebuilding a block-threaded DCAT
  
  hakkelt
  and
  claude
  committed
  7b555f5
- test: compare the device MatrixOp allocations with the bare gemm
  
  hakkelt
  and
  claude
  committed
  30a0e00
- docs: trim the threading-fix comments
  
  hakkelt
  and
  claude
  committed
  6113046
- test: move the device threading testitems to GPUEnv 0.3
  
  hakkelt
  and
  claude
  committed
  ea9bb22
Commits on Nov 3, 2026
- fix(batching): honour `threaded = false` in batch operators
  
  hakkelt
  and
  claude
  committed
  1954443
- fix(calculus): copy only shared blocks for a threaded block loop
  
  hakkelt
  and
  claude
  committed
  fab33b8
- style: drop the module prefix from exported names
  
  hakkelt
  and
  claude
  committed
  c2496d6
- refactor: derive threading traits from `_children` by default
  
  hakkelt
  and
  claude
  committed
  88fd54f
- docs: fix the threading-trait docstrings after the `_children` defaults
  
  hakkelt
  and
  claude
  committed
  0a340a4
- fix: use `Base.IdSet`, which Julia 1.10 does not export
  
  hakkelt
  and
  claude
  committed
  87c0f2e
- test: make the LBFGS block-identity check meaningful and platform-stable
  
  hakkelt
  and
  claude
  committed
  d4f0393
- perf(nonlinear): drop per-call allocations and the serial Sigmoid slowdown
  
  hakkelt
  and
  claude
  committed
  72ee810

Commits

Commits on Nov 2, 2026

Commits on Nov 3, 2026