perf: replace scalar for-loops with IEngine hardware-accelerated ops in 147 files

[ooples]

Summary

• Replaces scalar NumOps.Subtract/NumOps.Multiply/NumOps.Add for-loops with hardware-accelerated IEngine vector and tensor operations across 147 files
• Net reduction of ~1,218 lines of code (2,395 deletions, 1,177 insertions) by replacing verbose element-wise loops with single Engine method calls
• Adds Engine property to DeepReinforcementLearningAgentBase and ContinualLearningStrategyBase for subclass access

Categories of files converted:

• Video models (30+ files): frame interpolation, motion estimation, segmentation, generation, recognition, inpainting
• Finance models: LSTNet, TSMixer, neural forecasting, transformers, graph neural networks
• Document models (29 files): OCR, layout analysis, page segmentation, pixel-to-sequence
• Audio models (3 files): VITS, Tacotron2, Wav2Vec2
• SSL methods (9 files): SimCLR, MoCo, MoCoV3, BYOL, SimSiam, BarlowTwins, MAE, TeacherStudentSSL, SSLFineTuningPipeline
• RL agents (2 files): DQN, MuZero
• Regression (2 files): RegressionBase, NonLinearRegressionBase
• Neural networks (4 files): Gpt4Vision, Blip2, AudioVisual networks
• Tabular models (10+ files): layers, base classes, utilities
• SSM layers (10+ files): fill operations, reduce-sum, concatenation
• Other: NeRF, TimeSeries, AiModelBuilder, diffusion layers

Patterns replaced:

1. SGD parameter updates: for(i) params[i] = NumOps.Subtract(params[i], NumOps.Multiply(lr, grads[i])) → Engine.Subtract(params, Engine.Multiply(grads, lr))
2. EMA updates: for(i) target[i] = NumOps.Add(NumOps.Multiply(m, target[i]), NumOps.Multiply(1-m, online[i])) → Engine.Add(Engine.Multiply(target, m), Engine.Multiply(online, 1-m))
3. Tensor element-wise ops: scalar loops → Engine.TensorAdd, TensorSubtract, TensorMultiplyScalar, TensorFill
4. Vector subtract/add: manual loops → Engine.Subtract(vec1, vec2), Engine.Add(vec1, vec2)
5. Reduce-sum: manual accumulation loops → Engine.ReduceSum

Test plan

• Build succeeds with dotnet build src/AiDotNet.csproj --framework net10.0 -c Release (0 errors, 0 warnings)
• Run existing unit tests to verify behavioral equivalence
• Verify GPU acceleration path works when available

🤖 Generated with Claude Code

Closes #910

Summary by CodeRabbit

• Chores

  • Bumped core tensor and optional native acceleration packages to v0.8.0.

• Refactor

  • Replaced many element-wise loops with centralized engine-backed tensor/vector operations and added shared vector math helpers (norms, normalization, distances, cosine similarity) for consistency and performance.

• New Features

  • Many layers/models now expose standardized training/export interfaces and engine integrations for improved interoperability and exportability.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
aidotnet-playground-api	Ready	Preview, Comment	Mar 1, 2026 4:33am

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

Walkthrough

Widespread refactor replacing element-wise numeric loops with AiDotNetEngine tensor ops and new VectorHelper utilities, adding Engine accessors, migrating many tabular components to LayerBase, and bumping AiDotNet.Tensors and native packages to 0.8.0. Public APIs largely preserved.

Changes

Cohort / File(s)	Summary
Dependency Updates src/AiDotNet.csproj	Bumped AiDotNet.Tensors, AiDotNet.Native.OpenBLAS, AiDotNet.Native.CLBlast, AiDotNet.Native.OneDNN 0.7.0 → 0.8.0.
Engine accessors added src/**/AutoML/*, src/**/ContinualLearning/*, src/**/ReinforcementLearning/*, src/**/Tabular/*, src/**/TextSafetyModuleBase.cs, ...	Added protected/protected static IEngine Engine => AiDotNetEngine.Current in many bases and modules to route tensor ops through AiDotNetEngine.Current.
VectorHelper: new vector ops src/Helpers/VectorHelper.cs, callers in src/**/*	Added L2Norm/Normalize/NormalizeInPlace/CosineSimilarity/DotProduct/EuclideanDistance/ManhattanDistance and replaced numerous local dot/norm/cosine implementations with VectorHelper calls.
Vectorized parameter updates (ubiquitous) src/AiModelBuilder.cs, src/*/UpdateParameters, src/*/ApplyGradients (many files)	Replaced per-element SGD/EMA/EMA-like updates with vector ops: Engine.Subtract(params, Engine.Multiply(grads, lr)) or Engine.Add/Multiply patterns across many UpdateParameters/ApplyGradients paths.
Tensor op replacements (Finance, Video, Diffusion, Vision, etc.) src/Finance/Forecasting/**, src/Video/**, src/Video/Generation/**, src/Diffusion/**, src/Video/Inpainting/**	Replaced manual loops/Transforms with Engine.TensorAdd/Subtract/Multiply/MultiplyScalar/Concatenate/MatMul/Sigmoid/Softmax and similar engine helpers.
Tabular layers → LayerBase migrations src/NeuralNetworks/Tabular/{AttentiveTransformer,BatchEnsembleLayer,FeatureTransformer,GatedFeatureLearningUnit,InteractingLayer,IntersampleAttention,ObliviousDecisionTree,PiecewiseLinearEncoding,SoftTree,...}	Many tabular components now inherit LayerBase<T> and expose overrides (SupportsTraining, Forward/Backward, Get/SetParameters, UpdateParameters, ResetState, ExportComputationGraph); internals converted to NumOps/Engine patterns—large public-surface changes requiring API/behavior review.
High-density internal rewrites src/NeuralNetworks/Layers/{DeformableConvolutionalLayer,TimeEmbeddingLayer,ConvolutionalLayer,Layer*}, src/NeuralNetworks/Layers/SSM/*	Significant changes: matmul, reduce, gradient accumulation, parameter updates moved to Engine ops. These files are dense and flagged for deep review (see notes).
Helper consolidations & removed local helpers many src/** (dot/norm/cosine/add/subtract helpers removed)	Removed many private helpers (DotProduct, Normalize, CosineSimilarity, EuclideanDistance, AddTensors, etc.) in favor of VectorHelper and Engine methods; impacts consistency and numeric behavior across callers.
Minor initialization and fill optimizations src/NeuralNetworks/Layers/SSM/*, various layers	Replaced manual loops filling ones/zeros with ones.Fill(NumOps.One) / Engine.TensorFill and strengthened null-checks on gradients.
Estimated hotspots needing attention See detailed files below	Files with high review priority: TimeEmbeddingLayer.cs, DeformableConvolutionalLayer.cs, ConvolutionalLayer.cs, InteractingLayer.cs, FeatureTransformer.cs, BatchEnsembleLayer.cs, tabular model migrations, and VectorHelper.cs (new, broad impact).

Sequence Diagram(s)

sequenceDiagram
  participant Model as Model (UpdateParameters / Forward)
  participant Engine as AiDotNetEngine (IEngine)
  participant Helper as VectorHelper
  Model->>Engine: Engine.Multiply(gradients, lr) / Tensor ops
  Model->>Engine: Engine.Subtract(parameters, scaledGrads)
  Model->>Helper: VectorHelper.{L2Norm,Normalize,CosineSimilarity,...}
  Engine-->>Model: Tensor/Vector result
  Helper-->>Model: scalar/distance/similarity

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

Possibly related PRs

• docs: improve CPU matmul and GPU buffer reuse #720 — engine/backends and buffer pooling that enable the tensor ops (direct runtime dependency of these changes).
• fix: fix GEMM correctness defaults and expand CPU SIMD ops #709 — changes to AiDotNetEngine and engine implementations relied upon by new Engine calls.
• refactor: migrate to AiDotNet.Tensors NuGet package #741 — aligns repo with AiDotNet.Tensors package/native backends referenced by the dependency bump.

Suggested labels

feature

"Loops turned to vectors, engines hum,
Helpers unified, the math now runs,
New LayerBase paths, big files to comb,
Flag dense layers and export hooks as BLOCKING,
Review VectorHelper and numeric-edge cases, then ship."

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 61.61% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	Title clearly summarizes main change: systematic replacement of scalar loops with IEngine hardware-accelerated operations across 147 files.
Linked Issues check	✅ Passed	PR fully addresses issue #910 Phase 1 objectives: replaced scalar loops with Engine tensor ops (TensorAdd, TensorSubtract, TensorMultiply, TensorMultiplyScalar) across video, finance, document, audio, SSL, RL, and tabular models. Added Engine properties to base classes. Scope, patterns, and priority modules align with issue requirements.
Out of Scope Changes check	✅ Passed	All changes are in-scope replacements per #910 Phase 1. Changes consistently replace scalar loops with IEngine operations, add Engine properties to base classes, and introduce VectorHelper for vector operations. No unrelated refactoring or feature additions detected outside Engine-based optimization scope.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch perf/iengine-tensor-ops-910

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[Copilot]

Pull request overview

This PR modernizes the AiDotNet model implementations by replacing many hand-written element-wise scalar loops with IEngine tensor/vector operations (SIMD/GPU-accelerated), and bumps the AiDotNet.Tensors/native acceleration package versions to align with the new API usage.

Changes:

• Replaced numerous scalar NumOps.* loops with Engine.* / Engine.Tensor* operations (e.g., add/subtract/multiply, concatenate, clamp, softmax, reduce-sum).
• Introduced/standardized access to the global engine in additional base classes (e.g., RL + continual learning) to support subclass conversions.
• Updated NuGet references to AiDotNet.Tensors and native backends from 0.7.0 → 0.8.0.

Reviewed changes

Copilot reviewed 147 out of 147 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
src/Video/VideoNeuralNetworkBase.cs	Uses engine scalar-divide/multiply+clamp for (de)normalization; replaces manual feature concatenation with TensorConcatenate.
src/Video/Understanding/VideoCLIP.cs	Replaces per-element transform add/subtract with TensorAdd/TensorSubtract.
src/Video/Segmentation/XMem.cs	Replaces manual accumulation/scale and channel concatenation with engine ops.
src/Video/Segmentation/SAM2.cs	Replaces loss gradient, concatenation, add, sigmoid with engine ops.
src/Video/Motion/RAFT.cs	Replaces loss gradient, GRU update math, concatenation, add, sigmoid with engine ops.
src/Video/Motion/GMFlow.cs	Replaces loss gradient/diff/concat/add with engine ops.
src/Video/Inpainting/ProPainter.cs	Replaces loss gradient, concatenation, add with engine ops.
src/Video/Inpainting/E2FGVI.cs	Replaces scale/add/concat/sigmoid with engine ops.
src/Video/Generation/StableVideoDiffusion.cs	Refactors diffusion algebra to engine tensor ops; replaces add helpers.
src/Video/Generation/OpenSora.cs	Refactors diffusion algebra and guidance to engine tensor ops; replaces add/sigmoid helpers.
src/Video/FrameInterpolation/RIFE.cs	Replaces loss gradient, concatenation, adds, flow scaling with engine ops.
src/Video/FrameInterpolation/FILM.cs	Replaces loss gradient, concatenation, sigmoid, add, flow scaling with engine ops.
src/Video/FrameInterpolation/DRVI.cs	Replaces scalar blending loop with tensor multiply/add.
src/Video/FrameInterpolation/ABME.cs	Replaces scalar blending loop with tensor multiply/add.
src/Video/ActionRecognition/VideoMAE.cs	Replaces manual softmax implementation with Engine.Softmax.
src/TimeSeries/TimeSeriesModelBase.cs	Uses engine vector ops for SGD-style parameter updates.
src/SelfSupervisedLearning/TeacherStudentSSL.cs	Uses engine vector ops for EMA + SGD parameter updates.
src/SelfSupervisedLearning/SimSiam.cs	Uses engine vector ops for SGD parameter updates.
src/SelfSupervisedLearning/SimCLR.cs	Uses engine vector ops for SGD parameter updates.
src/SelfSupervisedLearning/SSLFineTuningPipeline.cs	Adds engine access and uses it for SGD parameter updates.
src/SelfSupervisedLearning/MoCoV3.cs	Uses engine vector ops for SGD parameter updates.
src/SelfSupervisedLearning/MoCo.cs	Uses engine vector ops for SGD parameter updates.
src/SelfSupervisedLearning/MAE.cs	Uses engine vector ops for SGD parameter updates.
src/SelfSupervisedLearning/BarlowTwins.cs	Uses engine vector ops for SGD parameter updates.
src/SelfSupervisedLearning/BYOL.cs	Uses engine ops for SGD + EMA parameter updates.
src/ReinforcementLearning/Agents/MuZeroAgent.cs	Uses engine vector ops for gradient application.
src/ReinforcementLearning/Agents/DeepReinforcementLearningAgentBase.cs	Adds protected Engine accessor for subclasses.
src/ReinforcementLearning/Agents/DQNAgent.cs	Uses engine vector ops for gradient application.
src/Regression/RegressionBase.cs	Uses engine vector ops for gradient application.
src/Regression/NonLinearRegressionBase.cs	Uses engine vector ops for gradient application.
src/NeuralRadianceFields/Models/NeRF.cs	Uses engine vector ops for gradient descent update.
src/NeuralNetworks/Tabular/TabTransformerBase.cs	Adds engine accessor; uses engine tensor ops for embedding updates.
src/NeuralNetworks/Tabular/TabPFNBase.cs	Uses engine tensor ops for residuals/matmul/fills.
src/NeuralNetworks/Tabular/TabNetBase.cs	Adds engine accessor; replaces ones/zeros/mask loops with engine ops.
src/NeuralNetworks/Tabular/TabMBase.cs	Adds engine accessor; refactors embedding updates.
src/NeuralNetworks/Tabular/TabDPTBase.cs	Uses engine tensor ops for residuals/matmul/fills.
src/NeuralNetworks/Tabular/SoftTree.cs	Converts to LayerBase<T> and uses engine ops for fills/updates; adds parameter export plumbing.
src/NeuralNetworks/Tabular/SAINTBase.cs	Adds engine accessor; replaces add helper with TensorAdd.
src/NeuralNetworks/Tabular/PiecewiseLinearEncoding.cs	Converts to LayerBase<T> and uses engine ops for fills/updates; adds parameter export plumbing.
src/NeuralNetworks/Tabular/ObliviousDecisionTree.cs	Converts to LayerBase<T> and uses engine ops for fills/updates; adds parameter export plumbing.
src/NeuralNetworks/Tabular/NODEBase.cs	Adds engine accessor; replaces ensemble accumulation/scaling loops with engine ops.
src/NeuralNetworks/Tabular/IntersampleAttention.cs	Converts to LayerBase<T>; replaces transpose/attention math with engine ops.
src/NeuralNetworks/Tabular/GatedFeatureLearningUnit.cs	Converts to LayerBase<T>; replaces sigmoid/multiply/add loops with engine ops.
src/NeuralNetworks/Tabular/GANDALFBase.cs	Adds engine accessor; replaces sigmoid/multiply/add/update loops with engine ops.
src/NeuralNetworks/Tabular/FeatureTransformer.cs	Converts to LayerBase<T>; refactors GLU/residual scaling to engine ops.
src/NeuralNetworks/Tabular/FeatureTokenizer.cs	Uses engine ops for parameter updates.
src/NeuralNetworks/Tabular/ContrastivePretraining.cs	Adds engine accessor; uses engine ops for parameter updates and fill helpers.
src/NeuralNetworks/Tabular/ColumnEmbedding.cs	Adds engine accessor; uses engine ops for updates + fill.
src/NeuralNetworks/Tabular/CLSToken.cs	Adds engine accessor; uses engine ops for updates + fill.
src/NeuralNetworks/Tabular/BatchEnsembleLayer.cs	Converts to LayerBase<T>; replaces some scalar ops with shared NumOps and adds parameter plumbing.
src/NeuralNetworks/Tabular/AutoIntBase.cs	Adds engine accessor; replaces gradient reset loops and some updates with engine ops.
src/NeuralNetworks/Tabular/AttentiveTransformer.cs	Converts to LayerBase<T>; replaces element-wise ops with engine ops.
src/NeuralNetworks/Layers/UNetDiscriminator.cs	Replaces AddTensors loop with Engine.TensorAdd.
src/NeuralNetworks/Layers/TimeEmbeddingLayer.cs	Refactors FC layers and gradient computations to use matmul/broadcast add/reduce-sum engine ops.
src/NeuralNetworks/Layers/SwinTransformerBlockLayer.cs	Replaces AddTensors loop with Engine.TensorAdd.
src/NeuralNetworks/Layers/SpyNetLayer.cs	Replaces add helper with TensorAdd and changes gradient accumulation to use engine add + copy-back.
src/NeuralNetworks/Layers/SSM/*	Replaces repeated “create ones” loops with Fill(NumOps.One) and tightens null-guard checks in parameter updates.
src/NeuralNetworks/Layers/ResidualDenseBlock.cs	Refactors residual add/scale helpers to engine ops.
src/NeuralNetworks/Layers/RRDBNetGenerator.cs	Replaces AddTensors loop with Engine.TensorAdd.
src/NeuralNetworks/Layers/RRDBLayer.cs	Refactors residual add/scale helpers to engine ops.
src/NeuralNetworks/Layers/HighwayLayer.cs	Replaces nested summation loops with Engine.ReduceSum across all axes.
src/NeuralNetworks/Layers/GroupedQueryAttentionLayer.cs	Tightens null-guard checks in parameter updates; uses engine ops consistently.
src/NeuralNetworks/Layers/DeformableConvolutionalLayer.cs	Uses engine ops for sigmoid derivative and gradient summation; refactors parameter updates to engine ops.
src/NeuralNetworks/Layers/CrossAttentionLayer.cs	Refactors weight update to use engine ops then copies into target tensor.
src/NeuralNetworks/Layers/ConvolutionalLayer.cs	Refactors CPU parameter update to engine tensor ops.
src/NeuralNetworks/Gpt4VisionNeuralNetwork.cs	Uses engine vector ops for gradient descent update.
src/NeuralNetworks/Blip2NeuralNetwork.cs	Uses engine vector ops for gradient descent update.
src/NeuralNetworks/AudioVisualEventLocalizationNetwork.cs	Uses engine vector ops for gradient descent update.
src/NeuralNetworks/AudioVisualCorrespondenceNetwork.cs	Uses engine vector ops for gradient descent update.
src/Finance/Graph/DCRNN.cs	Replaces scalar combine/accumulate/scale loops with engine ops.
src/Finance/Forecasting/Transformers/TimesNet.cs	Replaces residual add loop with Engine.TensorAdd.
src/Finance/Forecasting/Transformers/TSMixer.cs	Uses engine ops for RevIN normalize/denormalize steps.
src/Finance/Forecasting/Transformers/TFT.cs	Refactors gated connection to engine ops.
src/Finance/Forecasting/Transformers/Crossformer.cs	Replaces residual add loop with Engine.TensorAdd.
src/Finance/Forecasting/Transformers/Autoformer.cs	Replaces add/subtract helpers with engine ops.
src/Finance/Forecasting/Neural/WaveNet.cs	Replaces add/multiply helpers with engine ops.
src/Finance/Forecasting/Neural/TCN.cs	Replaces add helper with engine ops.
src/Finance/Forecasting/Neural/NHiTSFinance.cs	Replaces add/subtract helpers with engine ops.
src/Finance/Forecasting/Neural/NBEATSFinance.cs	Replaces add/subtract helpers with engine ops.
src/Finance/Forecasting/Neural/LSTNet.cs	Uses engine ops for highway residual add.
src/Finance/Forecasting/Neural/DeepState.cs	Replaces mismatched-shape handling with TensorBroadcastAdd.
src/Finance/Forecasting/Neural/DeepFactor.cs	Replaces manual “maxLen” broadcasting behavior with TensorBroadcastAdd.
src/Finance/Forecasting/Foundation/TimesFM.cs	Replaces add helper with engine ops.
src/ContinualLearning/Strategies/ContinualLearningStrategyBase.cs	Adds protected static Engine accessor for subclasses.
src/Audio/TextToSpeech/VITSModel.cs	Uses engine vector ops for gradient descent update.
src/Audio/TextToSpeech/Tacotron2Model.cs	Uses engine vector ops for gradient descent update.
src/Audio/SpeechRecognition/Wav2Vec2Model.cs	Uses engine vector ops for gradient descent update.
src/AiModelBuilder.cs	Refactors distillation SGD update to engine vector ops.
src/AiDotNet.csproj	Bumps AiDotNet.Tensors and native backend packages to 0.8.0.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.