PIF, Higher Order and Optimized Scatter/Gather, #501 Splitup

[aaadelmann]

PR 501 Split Map

Source PR: #501

Compared against: origin/master at da43fd66a18848b65dcd82af90001f5b64a798ab

Original PR size: 41 commits, 149 files changed, 23,231 insertions, 2,205 deletions.

The original PR is reviewable only as a source branch. The extracted split is now the working structure:

Order	Branch	Scope	Depends on
1	pr501-pcg	PCG / preconditioner allocation cleanup	master
2	pr501-communication-particle-update	communication buffers, particle update, particle serialization, particle send/recv tests	pr501-pcg
3	pr501-hosg	higher-order scatter/gather, binning, autotune cache	pr501-communication-particle-update
4	pr501-fft	FFT backend / transform refactor, pruned transforms	pr501-hosg
5	pr501-nufft	native NUFFT plus FINUFFT/cuFINUFFT integration	pr501-fft
6	pr501-pif	Alpine ElectrostaticPIF examples	pr501-nufft

Split Rationale

The review load in PR #501 is dominated by three independent-looking but technically dependent areas:

• Particle communication/update: direct archive serialization, shared send/recv buffers, layout migration, sorting, and nghost plumbing.
• Interpolation/autotune: new scatter/gather dispatch, binning, tiling, tuning cache, and legacy CIC routing.
• FFT/NUFFT/PIF: transform/backend refactor, native NUFFT, FINUFFT/cuFINUFFT support, and Alpine PIF examples.

The dependency order is:

PCG
  -> communication + particle update
    -> higher-order scatter/gather
      -> FFT backend / transform refactor
        -> native NUFFT + FINUFFT/cuFINUFFT
          -> PIF examples

Important dependency details:

• Native NUFFT uses the new interpolation scatter/gather layer.
• PIF uses FFT<NUFFTransform, Field_t>, scatterPIFNUFFT, and gatherPIFNUFFT.
• Particle update and communication are coupled through ParticleAttrib::serialize, ParticleAttrib::deserialize, Archive, and BufferHandler.
• nghost changes touch FieldLayout and HaloCells; they belong with particle update because the update path needs kernel-width-aware halos.

Extracted Branches

PR 1: PCG Allocation / Solver Performance

Branch: pr501-pcg

Scope:

• src/LinearSolvers/PCG.h
• src/LinearSolvers/Preconditioner.h
• src/PoissonSolvers/PoissonCG.h

Purpose:

• Hoist repeated per-iteration allocation work out of PCG/preconditioner hot paths.
• Pass Field by reference through OperatorF.
• Keep this independent from particle, interpolation, FFT, NUFFT, and PIF changes.

PR 2: Communication And Particle Update Infrastructure

Branch: pr501-communication-particle-update

Scope:

• src/Communicate/*
• src/Particle/ParticleSpatialLayout*
• src/Particle/ParticleAttrib*
• src/Particle/ParticleAttribBase.h
• src/Particle/ParticleBase*
• src/Particle/ParticleSort.h
• src/Particle/SortBuffer.h
• src/FieldLayout/FieldLayout*
• src/Field/HaloCells*
• particle update/send-recv tests

Purpose:

• Replace per-attribute staging buffers with direct serialization into shared communication archives.
• Rework particle migration into a packed locate/send/recv path with reusable scratch.
• Preserve this as the non-PIF communication and particle update PR.

Important fix added during split:

• The receive path now pre-reserves particle attribute storage once per update before deferred receive deserialization.
• This avoids repeated preserving Kokkos::resize(offset + nrecvs) calls per source rank and per attribute.
• On GPU this removed the observed multi-second particleDeserialize bottleneck.

Key regression tests:

• ParticleSendRecv with at least 2 MPI ranks; 1 rank does not exercise receive/deserialization.
• ParticleUpdate
• ParticleUpdateNonuniform

PR 3: Higher-Order Scatter/Gather And Autotune

Branch: pr501-hosg

Scope:

• src/Interpolation/*
• cmake/AutoTunePresets.cmake
• cmake/IpplAutoTunePresets.h.in
• cmake/auto_tune/*
• interpolation tests

Purpose:

• Add scatter/gather dispatch over atomic, atomic-sort, tiled, and output-focused paths.
• Add particle binning and hardware tile-size cache lookup.
• Provide the algorithmic layer required by native NUFFT.

PR 4: FFT Backend / Transform Refactor

Branch: pr501-fft

Scope:

• src/FFT/Backend/*
• src/FFT/Transform/CC.h
• src/FFT/Transform/RC.h
• src/FFT/Transform/Trig.h
• src/FFT/Transform/PrunedCC.h
• src/FFT/Transform/PrunedRC.h
• src/FFT/Transform/Common.h
• src/FFT/Traits.h
• existing FFT test updates

Purpose:

• Split the old monolithic FFT layer into backend and transform frontends.
• Expose CC, RC, pruned CC/RC, and trigonometric transforms without native NUFFT/PIF in the same PR.

Intentionally left for PR 5:

• src/FFT/NUFFT/*
• src/FFT/Transform/NUFFT.*
• NUFFT-specific tests
• FINUFFT/cuFINUFFT dependency use beyond dormant traits/config hooks.

PR 5: Native NUFFT And FINUFFT/cuFINUFFT Integration

Branch: pr501-nufft

Scope:

• src/FFT/NUFFT/*
• src/FFT/Transform/NUFFT.*
• IPPL_ENABLE_FINUFFT
• IPPL_ENABLE_CUFFTMP
• NUFFT tests

Purpose:

• Add native IPPL NUFFT and optional FINUFFT/cuFINUFFT backend support.
• Keep all PIF examples out of this PR.

Validation with FINUFFT disabled:

cmake -S . -B build-pr501-nufft-debug -DCMAKE_BUILD_TYPE=Debug -DIPPL_PLATFORMS=SERIAL -DIPPL_ENABLE_FFT=ON -DIPPL_ENABLE_UNIT_TESTS=ON
cmake --build build-pr501-nufft-debug --target ippl NUFFT NUFFTAccuracy -j 8
ctest --test-dir build-pr501-nufft-debug -R '^NUFFT$|^NUFFTAccuracy$' --output-on-failure

2/2 NUFFT tests passed
Total test time: 19.19 sec

PR 6: PIF / Alpine Examples

Branch: pr501-pif

Scope:

• alpine/ElectrostaticPIF/*
• PIF-specific particle attribute convenience APIs if not already included in NUFFT
• Alpine PIF wiring

Purpose:

• Add the ElectrostaticPIF examples after the core particle, interpolation, FFT, and NUFFT APIs are already reviewable.

Current status:

• LandauDampingPIF supports the default upsampled mode and an optional positional pruned mode.
• LandauDampingPIFPruned is not kept as a separate executable in the current split.
• pr501-pif has the receive pre-reserve particle update fix applied and pushed.

Build note:

• LandauDampingPIF is added when IPPL_ENABLE_ALPINE=ON and IPPL_ENABLE_FFT=ON.
• With IPPL_ENABLE_FINUFFT=OFF, it can compile against the native NUFFT stack, but runtime behavior depends on ChargedParticlesPIF::initNUFFT() selecting the native backend.

LUMI Status

The split branch stack is running successfully on LUMI. Current benchmark data show no evidence that the decomposition itself introduced a broad regression; several workloads improve relative to master.

Benchmark	Problem size	Nodes	Ranks	master	pr501-pcg	pr501-com	pr501-fft	pr501-hosg	pr501-nufft
FEM	513_10	8	64	28.10	28.13 (0%)	27.55 (-2%)	26.78 (-5%)	27.29 (-3%)	26.49 (-6%)
FFT	512_10	4	32	4.43	4.38 (-1%)	3.48 (-21%)	3.70 (-16%)	3.47 (-22%)	3.53 (-20%)
FFT	512_10	16	128	1.66	1.66 (0%)	1.14 (-31%)	1.13 (-31%)	1.49 (-10%)	1.08 (-35%)
PCG	512_10	1	8	72.28	70.17 (-3%)		67.58 (-6%)	67.65 (-6%)	67.59 (-6%)
PCG	512_10	4	32	34.80	32.95 (-5%)	32.34 (-7%)	36.40 (+5%)	31.93 (-8%)	31.93 (-8%)
PCG	512_10	64	512	25.09	22.96 (-8%)	22.93 (-9%)	22.11 (-12%)	22.09 (-12%)	22.21 (-11%)

Particle receive pre-reserve fix evidence from LUMI:

Run	Before	After
32 ranks particleDeserialize wall max	7.88320 s	0.0106297 s
32 ranks updateParticle wall max	8.08981 s	0.338958 s
128 ranks particleDeserialize wall max	1.40305 s	0.0132888 s
128 ranks updateParticle wall max	1.59376 s	0.214328 s

Interpretation:

• The slow path repeatedly resized particle attributes during deferred receive finalization.
• Each preserving resize copied existing particle storage on the GPU.
• Pre-reserving once to localAfterDestroy + totalRecvs removes the repeated copy pattern.

LUMI Final Results

Benchmark	Problem size	Nodes	Ranks	master	pr501-pcg	pr501-com	pr501-fft	pr501-hosg	pr501-nufft	pr501-original	pr501-pif
FEM	513_10	8	64	28.08	27.93 (-1%)	27.01 (-4%)	26.56 (-5%)	26.60 (-5%)	26.32 (-6%)	28.76 (+2%)	27.78 (-1%)
FFT	512_10	4	32	4.50	4.46 (-1%)	3.51 (-22%)	3.59 (-20%)	3.42 (-24%)	3.55 (-21%)	3.81 (-15%)	3.48 (-23%)
FFT	512_10	16	128	1.65	1.64 (0%)	1.10 (-33%)	1.15 (-30%)	1.26 (-24%)	1.08 (-35%)	1.15 (-30%)	1.14 (-31%)
PCG	512_10	1	8	72.21	70.20 (-3%)		67.58 (-6%)	67.37 (-7%)	67.69 (-6%)	66.42 (-8%)	67.81 (-6%)
PCG	512_10	4	32	34.70	32.97 (-5%)	32.30 (-7%)	31.99 (-8%)	31.94 (-8%)	32.02 (-8%)	32.60 (-6%)	32.63 (-6%)
PCG	512_10	64	512	24.63	22.86 (-7%)	23.06 (-6%)	22.12 (-10%)	22.16 (-10%)	22.21 (-10%)	22.99 (-7%)	22.82 (-7%)
PIF	1024_10	16	128								132.20
PIF	1024_10	32	256								69.34
PIF	1024_10	64	512								33.78
PIF	512_10	4	32							59.02	59.10
PIF	512_10	8	64							31.05	29.16
PIF	512_10	16	128							15.50	15.63

Contrast With PR #501 Description

The live PR #501 description is still useful as a source inventory, but it no longer matches the finite split branches.

Area	PR #501 Description	Split Status
Overall scope	One large PR for PIF, FFT, NUFFT, scatter/gather, particle update, communication, timings, Alpine, and tests	Six finite branches with explicit dependency order
PCG	Not a central part of the PR description	Extracted as standalone pr501-pcg
Particle update	Describes direct archive serialization and shared buffers	Also includes receive pre-reserve fix for GPU deserialization performance
Scatter/gather	Described together with all other changes	Isolated as pr501-hosg
FFT	Mixed with native NUFFT	Backend/transform refactor isolated as pr501-fft
NUFFT	Native NUFFT and FINUFFT/cuFINUFFT described together with FFT and PIF	Isolated as pr501-nufft, after scatter/gather and FFT
PIF	Mentions separate LandauDampingPIFPruned and ChargedParticlesPIFPruned files	Current split uses LandauDampingPIF with optional pruned mode; check file list before reusing original wording
Validation	Generic unit/integration test list	Split map records concrete local tests and LUMI benchmark results

[PaulFisch]

Looks very nice!

[srikrrish]

Looks great! thanks a lot for all the effort @aaadelmann