Class RoutingTableBuilder

Namespace

DotCompute.Backends.CUDA.RingKernels

Assembly

DotCompute.Backends.CUDA.dll

Builds and manages GPU-resident routing tables for inter-kernel messaging.

public sealed class RoutingTableBuilder : IDisposable

Inheritance

object

RoutingTableBuilder

Implements

IDisposable

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Extension Methods

BufferExtensions.IsSafeToDispose(IDisposable)

BufferExtensions.SafeDispose(IDisposable)

Remarks

The routing table builder performs the following operations: 1. Validates kernel registrations and assigns queue indices 2. Calculates optimal hash table capacity (power of 2, ~50% load factor) 3. Allocates GPU memory for routing table structures 4. Populates hash table with kernel ID → queue index mappings 5. Returns initialized KernelRoutingTable for GPU usage

Memory Management: - Hash table: capacity * 4 bytes (32-bit entries) - Output queues array: kernel_count * 8 bytes (device pointers) - Control blocks array: kernel_count * 64 bytes (RingKernelControlBlock)

Thread Safety: Not thread-safe. Builder should be used sequentially during initialization.

Constructors

RoutingTableBuilder(CudaContext)

Initializes a new routing table builder for the specified CUDA context.

public RoutingTableBuilder(CudaContext context)

Parameters

context CudaContext

CUDA context for GPU memory allocation.

Exceptions

ArgumentNullException

Thrown if context is null.

Methods

Build()

Builds the routing table and allocates GPU memory.

public KernelRoutingTable Build()

Returns

KernelRoutingTable

Initialized KernelRoutingTable with GPU-resident data.

Remarks

GPU Memory Allocation: - Hash table: capacity * 4 bytes - Output queues: kernel_count * 8 bytes - Control blocks: kernel_count * 64 bytes

Hash Table Construction: Uses linear probing to insert kernel ID → queue index mappings. Entry format: (kernel_id << 16) | queue_index

Exceptions

InvalidOperationException

Thrown if no kernels registered or validation fails.

ObjectDisposedException

Thrown if builder has been disposed.

Dispose()

Disposes the routing table builder (does not free GPU memory allocated during Build).

public void Dispose()

Remarks

GPU memory allocated by Build() must be freed separately using CudaInterop.cuMemFree.

RegisterKernel(string, long, long)

Registers a kernel in the routing table.

public int RegisterKernel(string kernelName, long outputQueuePtr, long controlBlockPtr)

Parameters

kernelName string

Unique kernel name (will be hashed to 16-bit ID).

outputQueuePtr long

Device pointer to kernel's output queue.

controlBlockPtr long

Device pointer to kernel's control block.

Returns

int

Assigned queue index for this kernel.

Exceptions

ArgumentException

Thrown if kernel name is null/empty or already registered.

ObjectDisposedException

Thrown if builder has been disposed.