Native ONNX import and CPU/GPU execution using tensor kernels, without ONNX Runtime. 更多...

#include <OnnxModel.h>

类
struct	Impl

Public 成员函数
	~OnnxModel ()
	Destroy model-owned graph and packed initializers; outputs remain valid.

OnnxModelInfo	info () const
	Return owning graph diagnostics; importing does not imply full operator support.

Result< std::vector< OnnxNamedTensor > >	run (std::span< const OnnxNamedTensor > feeds, std::span< const std::string > requested={}, OnnxRunOptions options={}) const
	Execute selected named values, or graph outputs when requested is empty.

Result< OnnxGpuResult >	runGpu (std::span< const OnnxNamedTensor > feeds, OnnxCompute &compute, std::span< const std::string > requested={}, OnnxRunOptions options={}) const
	Execute with GPU matrix, convolution, normalization, activation and resampling kernels.

静态 Public 成员函数
static Result< std::unique_ptr< OnnxModel > >	load (std::span< const uint8_t > bytes)
	Import bounded in-memory ONNX ModelProto; copies all retained data.

详细描述

Native ONNX import and CPU/GPU execution using tensor kernels, without ONNX Runtime.

注解: Owns immutable model data. Concurrent CPU run calls are independent; no script callbacks, global RNG or background threads. runGpu uses a caller-provided device on its owning thread. FP32 results use tolerance comparison. Seeded random excitation is local to a run. Loading is transactional. Unknown protobuf metadata is ignored, unknown operators are reported; unsupported semantic features fail explicitly, never become identity.

在文件 OnnxModel.h 第 63 行定义.

构造及析构函数说明

◆ ~OnnxModel()

eve::tensor::OnnxModel::~OnnxModel ( )

default

Destroy model-owned graph and packed initializers; outputs remain valid.

成员函数说明

◆ info()

OnnxModelInfo eve::tensor::OnnxModel::info ( ) const

Return owning graph diagnostics; importing does not imply full operator support.

在文件 OnnxModel.cpp 第 12 行定义.

◆ load()

Result< std::unique_ptr< OnnxModel > > eve::tensor::OnnxModel::load ( std::span< const uint8_t > bytes )

static

Import bounded in-memory ONNX ModelProto; copies all retained data.

参数

bytes Borrowed input, not retained. Maximum 512 MiB, 100000 recursive nodes, graph depth 16, rank 6.

返回: Owning model or ParseError/Unsupported/UnknownVersion; no partial publication.

注解: Supports ONNX IR 3..10 and default-domain opset 13..17. External data is rejected.

在文件 OnnxModel.cpp 第 13 行定义.

引用了 bytes, eve::tensor::onnx_detail::Failure::code, eve::Diagnostic::error(), eve::Failed, impl , 以及 eve::tensor::onnx_detail::parse().

被这些函数引用 eve::animation::SmrOnnxRunner::load().

◆ run()

Result< std::vector< OnnxNamedTensor > > eve::tensor::OnnxModel::run	(	std::span< const OnnxNamedTensor >	feeds,
		std::span< const std::string >	requested = `{}`,
		OnnxRunOptions	options = `{}`
	)		const

Execute selected named values, or graph outputs when requested is empty.

参数

feeds	Borrowed inputs, valid only during this call; never mutated or retained.
requested	Borrowed output names; intermediate outputs may be selected for parity tests.

返回: Owning outputs or structured node diagnostic; failed calls publish no outputs.

注解: CPU only. Unsupported dependencies fail before executing the selected subgraph.

在文件 OnnxModel.cpp 第 26 行定义.

引用了 options.

◆ runGpu()

Result< OnnxGpuResult > eve::tensor::OnnxModel::runGpu	(	std::span< const OnnxNamedTensor >	feeds,
		OnnxCompute &	compute,
		std::span< const std::string >	requested = `{}`,
		OnnxRunOptions	options = `{}`
	)		const

Execute with GPU matrix, convolution, normalization, activation and resampling kernels.

注解: Shape/index/control operations, padding and LSTM gates execute on CPU. Dynamic quantization reads three validation scalars; activations remain on GPU. Integer GEMM/Conv, LSTM projections, FP32 GEMM/ConvTranspose, elementwise math, Softmax, sum/mean reductions, normalization, Resize and CumSum execute on GPU. The borrowed compute provider must outlive this synchronous call; use its device thread. No automatic CPU retry occurs after a GPU error. Outputs own their storage.

在文件 OnnxModel.cpp 第 31 行定义.

引用了 bytes, compute, count, device, eve::tensor::OnnxCompute::endRun(), eve::Result< T >::failure(), inputs, options, r, scope, source, eve::Result< T >::success(), t , 以及 target.

该类的文档由以下文件生成:

src/modules/tensor/OnnxModel.h
src/modules/tensor/OnnxModel.cpp

类

Public 成员函数

静态 Public 成员函数