Class CudaKernelCompiler

Namespace

DotCompute.Backends.CUDA.Compilation

Assembly

DotCompute.Backends.CUDA.dll

CUDA kernel compiler implementation using NVRTC. Now delegates to specialized compilation pipeline for improved maintainability.

public sealed class CudaKernelCompiler : IDisposable, IAsyncDisposable

Inheritance

object

CudaKernelCompiler

Implements

IDisposable

IAsyncDisposable

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)

Constructors

CudaKernelCompiler(CudaContext, ILogger)

Initializes a new instance of the CudaKernelCompiler class.

[RequiresUnreferencedCode("This type uses runtime code generation and reflection")]
[RequiresDynamicCode("This type uses runtime code generation for CUDA kernel compilation")]
public CudaKernelCompiler(CudaContext context, ILogger logger)

Parameters

context CudaContext

The context.

logger ILogger

The logger.

Methods

ClearCache()

Performs clear cache.

public void ClearCache()

CompileAsync(KernelDefinition, CompilationOptions?, CancellationToken)

Gets compile asynchronously.

public Task<ICompiledKernel> CompileAsync(KernelDefinition definition, CompilationOptions? options = null, CancellationToken cancellationToken = default)

Parameters

definition KernelDefinition

The definition.

options CompilationOptions

The options.

cancellationToken CancellationToken

The cancellation token.

Returns

Task<ICompiledKernel>

The result of the operation.

CompileBatchAsync(KernelDefinition[], CompilationOptions?, CancellationToken)

Gets compile batch asynchronously.

public Task<ICompiledKernel[]> CompileBatchAsync(KernelDefinition[] definitions, CompilationOptions? options = null, CancellationToken cancellationToken = default)

Parameters

definitions KernelDefinition[]

The definitions.

options CompilationOptions

The options.

cancellationToken CancellationToken

The cancellation token.

Returns

Task<ICompiledKernel[]>

The result of the operation.

CompileKernelAsync(KernelDefinition, CompilationOptions?, CancellationToken)

Gets compile kernel asynchronously.

public Task<ICompiledKernel> CompileKernelAsync(KernelDefinition definition, CompilationOptions? options = null, CancellationToken cancellationToken = default)

Parameters

definition KernelDefinition

The definition.

options CompilationOptions

The options.

cancellationToken CancellationToken

The cancellation token.

Returns

Task<ICompiledKernel>

The result of the operation.

Dispose()

Performs dispose.

public void Dispose()

DisposeAsync()

Gets dispose asynchronously.

public ValueTask DisposeAsync()

Returns

ValueTask

The result of the operation.

GetAllMangledNames(string)

Gets all mangled function names for a kernel.

public static Dictionary<string, string>? GetAllMangledNames(string kernelName)

Parameters

kernelName string

Returns

Dictionary<string, string>

GetCacheStatistics()

Gets cache statistics for monitoring and debugging

public CacheStatistics GetCacheStatistics()

Returns

CacheStatistics

GetMangledFunctionName(string, string)

Gets the mangled function name for a kernel and function name.

public static string? GetMangledFunctionName(string kernelName, string functionName)

Parameters

kernelName string

functionName string

Returns

string

GetNvrtcVersion()

Gets NVRTC version information

public static (int major, int minor) GetNvrtcVersion()

Returns

(int x, int y)

IsNvrtcAvailable()

Checks NVRTC availability and version

public static bool IsNvrtcAvailable()

Returns

bool

SetTimingProvider(CudaTimingProvider?)

Sets the timing provider for automatic timestamp injection support.

public void SetTimingProvider(CudaTimingProvider? timingProvider)

Parameters

timingProvider CudaTimingProvider

The timing provider instance, or null to disable injection.

Remarks

When a timing provider is set and timestamp injection is enabled via EnableTimestampInjection(bool), all subsequently compiled kernels will have timestamp recording code automatically injected at their entry point.

This allows for automatic kernel profiling without manually modifying kernel code. The injected code records a GPU timestamp (via %%globaltimer on CC 6.0+) in parameter slot 0, shifting all user parameters by one position.

TryGetCached(string, out ICompiledKernel?)

Returns true if able to get cached, otherwise false.

public bool TryGetCached(string kernelName, out ICompiledKernel? compiledKernel)

Parameters

kernelName string

The kernel name.

compiledKernel ICompiledKernel

The compiled kernel.

Returns

bool

true if the operation succeeded; otherwise, false.