GPU execution boundary for native ONNX; retains no model and retains compiled resources for the lifetime of the provider. 更多...

#include <OnnxCompute.h>

Public 成员函数
virtual	~OnnxCompute ()=default
	Release device resources after all synchronous calls have returned.

virtual Result< std::vector< uint8_t > >	dispatch (const OnnxKernel &kernel)=0
	Compatibility-only synchronous host-byte entry point; engine dispatch delegates to enqueue.

virtual Result< OnnxBuffer >	enqueue (const std::string &source, std::span< const OnnxBuffer > inputs, size_t outputBytes, uint32_t workItems)
	Queue a kernel using immutable host/device buffers; no readback is required.

virtual OnnxTransferStats	transferStats () const
	Observable transfer counters for the current or most recent run.

Protected 成员函数
virtual void	beginRun () noexcept
	Start a synchronous run cache scope; invoked only by OnnxModel on the device thread.

virtual void	endRun () noexcept
	Retire temporary activations/recordings; compiled resources may survive for later calls.

友元
class	OnnxModel

详细描述

GPU execution boundary for native ONNX; retains no model and retains compiled resources for the lifetime of the provider.

注解: Use on the device's owning thread, without concurrent calls or reentrancy. dispatch completes GPU work before returning host bytes; enqueue may defer completion. Errors never trigger CPU execution. Integer kernels require exact int32 results.

在文件 OnnxCompute.h 第 25 行定义.

构造及析构函数说明

◆ ~OnnxCompute()

virtual eve::tensor::OnnxCompute::~OnnxCompute ( )

virtualdefault

Release device resources after all synchronous calls have returned.

成员函数说明

◆ beginRun()

virtual void eve::tensor::OnnxCompute::beginRun ( )

inlineprotectedvirtualnoexcept

Start a synchronous run cache scope; invoked only by OnnxModel on the device thread.

在文件 OnnxCompute.h 第 69 行定义.

◆ dispatch()

virtual Result< std::vector< uint8_t > > eve::tensor::OnnxCompute::dispatch ( const OnnxKernel & kernel )

pure virtual

Compatibility-only synchronous host-byte entry point; engine dispatch delegates to enqueue.

注解: GLSL 450, local_size_x=64; inputs occupy bindings 0..N-1, output N.

被这些函数引用 enqueue().

◆ endRun()

virtual void eve::tensor::OnnxCompute::endRun ( )

inlineprotectedvirtualnoexcept

Retire temporary activations/recordings; compiled resources may survive for later calls.

在文件 OnnxCompute.h 第 71 行定义.

被这些函数引用 eve::tensor::OnnxModel::runGpu().

◆ enqueue()

virtual Result< OnnxBuffer > eve::tensor::OnnxCompute::enqueue	(	const std::string &	source,
		std::span< const OnnxBuffer >	inputs,
		size_t	outputBytes,
		uint32_t	workItems
	)

inlinevirtual

Queue a kernel using immutable host/device buffers; no readback is required.

注解: Default provider executes synchronously through dispatch. Engine provider records a resident graph segment inside runGpu, completing it at a host-read or resource-budget boundary. Inputs are retained until completion; returned device storage expires when runGpu ends.

Failure.

Error.

Success.

在文件 OnnxCompute.h 第 37 行定义.

引用了 bytes, dispatch(), eve::Diagnostic::error(), eve::Result< T >::failure(), input, inputs, eve::tensor::OnnxKernel::inputs, eve::InvalidArgument, r, source , 以及 eve::Result< T >::success().

◆ transferStats()

virtual OnnxTransferStats eve::tensor::OnnxCompute::transferStats ( ) const

inlinevirtual

Observable transfer counters for the current or most recent run.

在文件 OnnxCompute.h 第 64 行定义.

友元及相关函数文档

◆ OnnxModel

friend class OnnxModel

friend

在文件 OnnxCompute.h 第 67 行定义.

该类的文档由以下文件生成:

src/modules/tensor/OnnxCompute.h

Public 成员函数

Protected 成员函数

友元