eve::gpgpu::Sequence类 参考

#include <Sequence.h>

类
struct	Impl

Public 成员函数
	Sequence ()
	~Sequence ()
	Sequence (const Sequence &)=delete
Sequence &	operator= (const Sequence &)=delete
bool	isAvailable () const
void	begin ()
void	recordUpload (GpuBuffer *dst, const void *src, uint64_t nbytes, uint64_t dstOffset=0)
void	recordDownload (GpuBuffer *src, GpuBuffer *staging, uint64_t nbytes, uint64_t srcOffset=0)
void	recordDispatch (ComputeShader *shader, int groupsX, int groupsY=1, int groupsZ=1)
void	submit ()

详细描述

Kompute-style GPU command sequence.

Records storage-buffer transfers and compute dispatches into a single command buffer, then submits everything once. This turns an inference graph (placeholder uploads -> N fused kernel dispatches -> output download) from N+2 separate record/submit/wait round trips into one GPU submission.

Typical usage:

seq->begin();
seq->recordUpload(inputBuffer, inputData, nbytes);
seq->recordDispatch(shaderA, groupsA);   // shader bindings/push
seq->recordDispatch(shaderB, groupsB);   // constants already set
seq->recordDownload(outputBuffer, staging, nbytes);
seq->submit();                            // one submit, waits once
staging->downloadBytes(dst, nbytes);      // host-visible memcpy

A Sequence is reusable: call begin() again after submit(). Shaders keep their bindings between records; only buffers they reference may be re-bound between dispatches.

在文件 Sequence.h 第 32 行定义.

构造及析构函数说明

Sequence() [1/2]

eve::gpgpu::Sequence::Sequence

(

)

在文件 Sequence.cpp 第 49 行定义.

~Sequence()

eve::gpgpu::Sequence::~Sequence

(

)

在文件 Sequence.cpp 第 50 行定义.

引用了 eve::gpgpu::Sequence::Impl::v , 以及 eve::gpgpu::vulkanSequenceDestroy().

Sequence() [2/2]

eve::gpgpu::Sequence::Sequence ( const Sequence &  )

delete

成员函数说明

begin()

void eve::gpgpu::Sequence::begin

(

)

Start recording. Safe to call again after submit() to reuse.

在文件 Sequence.cpp 第 64 行定义.

引用了 eve::gpgpu::Sequence::Impl::v , 以及 eve::gpgpu::vulkanSequenceBegin().

被这些函数引用 eve::tensor::GpuProgram::run().

isAvailable()

bool eve::gpgpu::Sequence::isAvailable

(

)

const

True when the active backend (Vulkan) supports command recording.

在文件 Sequence.cpp 第 55 行定义.

引用了 eve::gpgpu::VulkanSequence::ensureReady(), eve::gpgpu::VulkanSequence::ready() , 以及 eve::gpgpu::Sequence::Impl::v.

operator=()

Sequence & eve::gpgpu::Sequence::operator= ( const Sequence &  )

delete

recordDispatch()

void eve::gpgpu::Sequence::recordDispatch	(	ComputeShader *	shader,
		int	groupsX,
		int	groupsY = 1,
		int	groupsZ = 1
	)

Record a compute dispatch using the shader's current bindings/push constants.

在文件 Sequence.cpp 第 76 行定义.

引用了 shader, eve::gpgpu::Sequence::Impl::v , 以及 eve::gpgpu::vulkanSequenceRecordDispatch().

被这些函数引用 eve::tensor::GpuProgram::Impl::recordGroup().

recordDownload()

void eve::gpgpu::Sequence::recordDownload	(	GpuBuffer *	src,
		GpuBuffer *	staging,
		uint64_t	nbytes,
		uint64_t	srcOffset = 0
	)

Record a device->host copy into staging (a "staging" usage buffer). Read the staging buffer after submit().

在文件 Sequence.cpp 第 71 行定义.

引用了 eve::gpgpu::Sequence::Impl::v , 以及 eve::gpgpu::vulkanSequenceRecordDownload().

被这些函数引用 eve::tensor::GpuProgram::run().

recordUpload()

void eve::gpgpu::Sequence::recordUpload	(	GpuBuffer *	dst,
		const void *	src,
		uint64_t	nbytes,
		uint64_t	dstOffset = 0
	)

Record a host->device copy. src is memcpy'd into an internal host-visible staging buffer immediately, so it only needs to stay valid until this call returns.

在文件 Sequence.cpp 第 66 行定义.

引用了 eve::gpgpu::Sequence::Impl::v , 以及 eve::gpgpu::vulkanSequenceRecordUpload().

被这些函数引用 eve::tensor::GpuProgram::run().

submit()

void eve::gpgpu::Sequence::submit

(

)

End recording, submit the whole sequence once, and wait for completion.

在文件 Sequence.cpp 第 81 行定义.

引用了 eve::gpgpu::Sequence::Impl::v , 以及 eve::gpgpu::vulkanSequenceSubmit().

被这些函数引用 eve::tensor::GpuProgram::run().

---

该类的文档由以下文件生成:

• src/modules/gpgpu/Sequence.h
• src/modules/gpgpu/Sequence.cpp