Class JitCompilerOptions

Namespace

AiDotNet.JitCompiler

Assembly

AiDotNet.dll

Configuration options for the JIT compiler.

public class JitCompilerOptions

Inheritance

object

JitCompilerOptions

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: Settings to control how the JIT compiler works.

You can:

• Enable/disable specific optimizations
• Turn caching on/off
• Configure compilation behavior
• Control how unsupported operations are handled

For most users, the defaults work great!

Properties

EnableAdaptiveFusion

Gets or sets a value indicating whether to enable adaptive fusion strategies. Default: false (currently uses standard fusion when enabled).

public bool EnableAdaptiveFusion { get; set; }

Property Value

bool

Remarks

Status: Architecture implemented, delegates to standard fusion. Adaptive fusion will intelligently select which operations to fuse based on graph structure, tensor sizes, and hardware characteristics.

EnableAutoTuning

Gets or sets a value indicating whether to enable auto-tuning of optimizations. Default: true.

public bool EnableAutoTuning { get; set; }

Property Value

bool

Remarks

Auto-tuning automatically determines the best optimization configuration for each graph based on graph analysis, tensor sizes, and operation types. It selects the optimal combination of fusion, unrolling, and vectorization strategies.

EnableCaching

Gets or sets a value indicating whether to enable caching of compiled graphs. Default: true.

public bool EnableCaching { get; set; }

Property Value

bool

EnableConstantFolding

Gets or sets a value indicating whether to enable constant folding optimization. Default: true.

public bool EnableConstantFolding { get; set; }

Property Value

bool

EnableDeadCodeElimination

Gets or sets a value indicating whether to enable dead code elimination. Default: true.

public bool EnableDeadCodeElimination { get; set; }

Property Value

bool

EnableLoopUnrolling

Gets or sets a value indicating whether to enable loop unrolling optimization. Default: true.

public bool EnableLoopUnrolling { get; set; }

Property Value

bool

Remarks

Loop unrolling improves performance for small, fixed-size loops by eliminating loop overhead and enabling better instruction pipelining. The optimizer automatically determines which loops benefit from unrolling based on tensor size and operation type.

EnableMemoryPooling

Gets or sets a value indicating whether to enable memory pooling for tensors. Default: true.

public bool EnableMemoryPooling { get; set; }

Property Value

bool

Remarks

For Beginners: Reuses tensor memory to reduce allocations.

Memory pooling improves performance by:

• Reducing garbage collection pauses
• Avoiding repeated memory allocations
• Improving cache locality

This is especially beneficial for training loops that create many temporary tensors.

EnableOperationFusion

Gets or sets a value indicating whether to enable operation fusion. Default: true.

public bool EnableOperationFusion { get; set; }

Property Value

bool

EnableSIMDHints

Gets or sets a value indicating whether to enable SIMD vectorization hints. Default: false (not yet fully implemented).

public bool EnableSIMDHints { get; set; }

Property Value

bool

Remarks

Status: Architecture planned, implementation pending. SIMD hints guide the code generator to use vector instructions (AVX, AVX-512) for better performance on element-wise operations.

LogUnsupportedOperations

Gets or sets whether to log warnings for unsupported operations. Default: true.

public bool LogUnsupportedOperations { get; set; }

Property Value

bool

Remarks

For Beginners: When enabled, you'll see warnings in logs when operations can't be JIT compiled. This helps you: - Identify which operations need fallback - Understand performance implications - Know when to request JIT support for new operation types

MaxElementsToPool

Gets or sets the maximum total elements in a tensor to pool. Tensors larger than this will not be pooled. Default: 10,000,000 (about 40MB for float32).

public int MaxElementsToPool { get; set; }

Property Value

int

MaxPoolSizePerShape

Gets or sets the maximum number of tensor buffers to keep per shape. Default: 10.

public int MaxPoolSizePerShape { get; set; }

Property Value

int

UnsupportedLayerHandling

Gets or sets how the JIT compiler handles unsupported operations. Default: Fallback (use interpreted execution for entire graph if any op is unsupported).

public UnsupportedLayerHandling UnsupportedLayerHandling { get; set; }

Property Value

UnsupportedLayerHandling

Remarks

For Beginners: When your model has operations the JIT can't compile, this setting controls what happens:

• Throw: Stop with an error - use when you need all ops compiled
• Fallback: (Default) Run the whole graph interpreted - always works
• Hybrid: JIT the supported ops, interpret the rest - best performance
• Skip: Ignore unsupported ops - dangerous, may give wrong results

Hybrid mode is recommended for production when you have mixed-support graphs. It gives you JIT speed for supported operations while still handling all ops correctly.