Class CudaTensorCoreManagerProduction

Namespace

DotCompute.Backends.CUDA.Advanced

Assembly

DotCompute.Backends.CUDA.dll

Production-grade CUDA Tensor Core manager with WMMA operations, mixed precision support, and performance profiling.

public sealed class CudaTensorCoreManagerProduction : IDisposable

Inheritance

object

CudaTensorCoreManagerProduction

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)

Constructors

CudaTensorCoreManagerProduction(CudaContext, CudaDeviceManager, ILogger<CudaTensorCoreManagerProduction>)

Initializes a new instance of the CudaTensorCoreManagerProduction class.

public CudaTensorCoreManagerProduction(CudaContext context, CudaDeviceManager deviceManager, ILogger<CudaTensorCoreManagerProduction> logger)

Parameters

context CudaContext

The context.

deviceManager CudaDeviceManager

The device manager.

logger ILogger<CudaTensorCoreManagerProduction>

The logger.

Properties

Capabilities

Gets tensor core capabilities.

public TensorCoreCapabilities Capabilities { get; }

Property Value

TensorCoreCapabilities

Statistics

Gets performance statistics.

public TensorCoreStatistics Statistics { get; }

Property Value

TensorCoreStatistics

TensorCoresAvailable

Gets whether tensor cores are available.

public bool TensorCoresAvailable { get; }

Property Value

bool

Methods

ConvolutionAsync(nint, nint, nint, ConvolutionParams, DataType, nint, CancellationToken)

Performs convolution using tensor cores.

public Task<TensorCoreResult> ConvolutionAsync(nint input, nint filter, nint output, ConvolutionParams parameters, DataType dataType, nint stream, CancellationToken cancellationToken = default)

Parameters

input nint

filter nint

output nint

parameters ConvolutionParams

dataType DataType

stream nint

cancellationToken CancellationToken

Returns

Task<TensorCoreResult>

Dispose()

Performs dispose.

public void Dispose()

MatrixMultiplyAsync(nint, nint, nint, int, int, int, DataType, DataType, nint, MatrixLayout, MatrixLayout, float, float, CancellationToken)

Performs mixed-precision matrix multiplication using tensor cores.

public Task<TensorCoreResult> MatrixMultiplyAsync(nint a, nint b, nint c, int m, int n, int k, DataType inputType, DataType outputType, nint stream, MatrixLayout layoutA = MatrixLayout.RowMajor, MatrixLayout layoutB = MatrixLayout.RowMajor, float alpha = 1, float beta = 0, CancellationToken cancellationToken = default)

Parameters

a nint

b nint

c nint

m int

n int

k int

inputType DataType

outputType DataType

stream nint

layoutA MatrixLayout

layoutB MatrixLayout

alpha float

beta float

cancellationToken CancellationToken

Returns

Task<TensorCoreResult>