Class CudaKernelLibrary

Namespace

JYPPX.CudaSharp

Assembly

JYPPX.CudaSharp.dll

Owns a CUDA runtime kernel library and exposes only copied metadata. 拥有 CUDA runtime kernel library，并且只公开复制型元数据。

public sealed class CudaKernelLibrary : IDisposable

Inheritance

object

CudaKernelLibrary

Implements

IDisposable

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Properties

Inventory

Gets a copied validation snapshot produced by count/enumerate calls. 获取由 count/enumerate 调用产生的复制型校验快照。

public CudaKernelLibraryInventorySnapshot Inventory { get; }

Property Value

CudaKernelLibraryInventorySnapshot

KernelCount

Gets the CUDA-reported kernel count. 获取 CUDA 报告的 kernel 数量。

public uint KernelCount { get; }

Property Value

uint

Methods

ContainsKernel(string)

Determines whether the library contains a named kernel without exposing its borrowed handle. 判断 library 是否包含指定 kernel，不暴露 borrowed handle。

public bool ContainsKernel(string name)

Parameters

name string

Returns

bool

ContainsUnifiedFunction(string)

Determines whether a named unified function exists without exposing its function pointer. 判断指定 unified function 是否存在，不暴露其 function pointer。

public bool ContainsUnifiedFunction(string name)

Parameters

name string

Returns

bool

Dispose()

Performs application-defined tasks associated with freeing, releasing, or resetting unmanaged resources.

public void Dispose()

Launch(string, CudaKernelLaunchConfiguration, CudaStream, params CudaKernelArgument[])

Launches a named kernel with copied scalar arguments and owner-bound device-memory arguments. 使用复制型标量参数和 owner-bound device-memory 参数启动 named kernel。

public CudaKernelLaunch Launch(string kernelName, CudaKernelLaunchConfiguration configuration, CudaStream stream, params CudaKernelArgument[] arguments)

Parameters

kernelName string

configuration CudaKernelLaunchConfiguration

stream CudaStream

arguments CudaKernelArgument[]

Returns

CudaKernelLaunch

Remarks

The returned owner retains this library, the stream, and every device-memory allocation until disposal. Dispose or synchronize the launch before reading results. 返回 owner 会租用 library、stream 与全部 device-memory allocation；读取结果前必须同步或释放 launch。

Load(byte[])

Loads CUDA cubin, fatbin, or PTX bytes while the native owner retains its own copy. 加载 CUDA cubin、fatbin 或 PTX 字节，native owner 会保留独立副本。

public static CudaKernelLibrary Load(byte[] code)

Parameters

code byte[]

Returns

CudaKernelLibrary

LoadFromFile(string)

Loads a CUDA cubin, fatbin, or null-terminated PTX file. 加载 CUDA cubin、fatbin 或以 null 结尾的 PTX 文件。

public static CudaKernelLibrary LoadFromFile(string path)

Parameters

path string

Returns

CudaKernelLibrary

SetAttributeForDevice(string, CudaKernelAttribute, int, int)

Sets a mutable attribute on a named kernel for one device without exposing the borrowed kernel handle. 为指定设备上的 named kernel 设置可变属性，不暴露 borrowed kernel handle。

public void SetAttributeForDevice(string kernelName, CudaKernelAttribute attribute, int value, int deviceOrdinal)

Parameters

kernelName string

attribute CudaKernelAttribute

value int

deviceOrdinal int

TryGetGlobalSymbolSize(string, out ulong)

Tries to copy the size of a named global symbol without retrieving its device pointer. 尝试复制指定 global symbol 的大小，不获取其 device pointer。

public bool TryGetGlobalSymbolSize(string name, out ulong sizeInBytes)

Parameters

name string

sizeInBytes ulong

Returns

bool

TryGetManagedSymbolSize(string, out ulong)

Tries to copy the size of a named managed symbol without retrieving its pointer. 尝试复制指定 managed symbol 的大小，不获取其 pointer。

public bool TryGetManagedSymbolSize(string name, out ulong sizeInBytes)

Parameters

name string

sizeInBytes ulong

Returns

bool