命名空间
namespace	affine

namespace	glsl_detail

namespace	kernels

namespace	onnx_detail

namespace	q

namespace	wgsl_detail

类
class	CompiledFunction
	Optimized / scheduled graph ready to run with feeds. 更多...

class	Func
	Trace builder — TF2 `tf.function` analogue (`tf.func` in scripts). While active, TF ops record into this graph. 更多...

struct	FusedGroup
	FusedGroup public API. 更多...

class	GpuProgram
	GPU execution of a compiled tensor Graph via generated compute shaders. 更多...

class	Graph
	EVENGINE_API_DOMAINS public API. 更多...

struct	GraphNode
	GraphNode public API. 更多...

struct	KernelSpec
	KernelSpec public API. 更多...

struct	OnnxBuffer
	Immutable buffer shared by graph aliases; exactly one storage is authoritative. 更多...

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

class	OnnxDeviceStorage
	Run-scoped device storage. Readback completes pending work on the device thread. 更多...

struct	OnnxGpuResult
	Owning GPU execution output and completed dispatch count. 更多...

struct	OnnxKernel
	Synchronous GPU kernel request; all inputs are borrowed only until dispatch returns. 更多...

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

struct	OnnxModelInfo
	Owning admission report; unsupported nodes remain inspectable but cannot execute. 更多...

struct	OnnxNamedTensor
	Owning named feed/output value. 更多...

struct	OnnxRunOptions
	Per-call deterministic RNG and optional strict finite-output diagnostic. 更多...

struct	OnnxTensor
	Owning ONNX boundary tensor: row-major little-endian bytes and exact integer shape. 更多...

struct	OnnxTransferStats
	Actual transfers and submissions made during one GPU run. 更多...

struct	OptimizedGraph
	OptimizedGraph public API. 更多...

class	Tensor
	float32 / int32 tensor (rank 1–6), row-major. Eager: owns a buffer. Symbolic: node in a Func graph (no buffer until run). 更多...

class	TF
	TF2-like namespace module. Script: `tf <- eve.TF();` Default eager; `tf.func()` traces a graph for compile/run. 更多...

枚举
enum class	OpType : uint8_t { Placeholder = 0 , Const , Add , Sub , Multiply , Divide , AddScalar , SubScalar , MulScalar , DivScalar , Neg , Abs , Sqrt , Exp , Log , Sin , Cos , Tanh , Relu , Sigmoid , Gelu , Silu , PowScalar , Clamp , MaximumScalar , MinimumScalar , Where , MatMul , Transpose , Permute , Reshape , Flatten , Softmax , LogSoftmax , LayerNorm , RMSNorm , Conv1d , Conv2d , MaxPool2d , AvgPool2d , Embedding , Concat , Slice , ReduceSum , ReduceMean , ReduceMin , ReduceMax , ArgMax , Cast , ScaledDotProductAttention , Resize2d }
	OpType public API. 更多...

enum class	KernelVariant { Default , TiledMatMul }
	Internal choice of generated matrix multiplication implementation. 更多...

enum class	OnnxElement : int { Float32 = 1 , UInt8 = 2 , Int8 = 3 , Int32 = 6 , Int64 = 7 , Bool = 9 }
	ONNX wire element types; distinct from block-quantized Tensor storage. 更多...

enum class	GroupKind : uint8_t { Elementwise , MatMul , Conv1d , Conv2d , MaxPool2d , AvgPool2d , Softmax , LayerNorm , RMSNorm , Reduce , ArgMax , Embedding , Concat , Slice , Permute , Sdpa , Resize2d , Alias }
	GroupKind public API. 更多...

enum class	DType : uint8_t { Float32 = 0 , Int32 = 1 , Fp16 = 2 , Fp8E4M3 = 3 , Fp4E2M1 = 4 , Int8 = 5 , Int4 = 6 }
	Tensor element types. 更多...

函数
bool	gpuReduce (const float *data, int size, int op, float &outResult)
	GPU-accelerated reduction for large eager tensors. op: 0 = sum, 1 = min, 2 = max. Returns false (caller should fall back to CPU) when Vulkan/gpgpu isn't available. This compatibility facade preserves the CPU alternative execution contract; new GPU APIs should return a named status enum or Result.

bool	generateKernel (const Graph &graph, const FusedGroup &group, KernelSpec &out)
	Generate kernel.

bool	generateMatMulVariant (const Graph &graph, const FusedGroup &group, bool tiled, KernelSpec &out)
	Generate mat mul variant.

void	generateKernelWgslImpl (const Graph &graph, const FusedGroup &group, KernelSpec &out)

Result< KernelSpec >	generateWgslKernel (const Graph &graph, const FusedGroup &group, KernelVariant variant=KernelVariant::Default)
	Lower an optimizer-produced group directly to owning WGSL source and dispatch metadata.

EVENGINE_API_DOMAINS Result< std::unique_ptr< OnnxCompute > >	createOnnxGpuCompute (uint32_t compilerWorkers=4)
	Create a reusable GPU session for the active engine Gpgpu Vulkan device, or an error.

OptimizedGraph	optimizeGraph (const Graph &graph, int outputNode)
	Optimize graph.

int	groupKernelCount (const OptimizedGraph &opt)
	Group kernel count.

const char *	dtypeName (DType dtype)
	Dtype name.

bool	parseDType (const std::string &name, DType &out)
	Parse d type.

	Module_IMPL (TF, new TF())

变量
constexpr int	kMaxKernelBindings = 8

枚举类型说明

◆ DType

enum class eve::tensor::DType : uint8_t

strong

Tensor element types.

Script-visible tensors are float32; int32 tensors are used for index data (argmax outputs, embedding lookups, cast("int32")). Int32 values are stored losslessly as floats for |v| < 2^24, which comfortably covers model vocabularies / sequence lengths / simulation ids used in games.

枚举值
Float32
Int32
Fp16
Fp8E4M3
Fp4E2M1
Int8
Int4

在文件 Tensor.h 第 24 行定义.

◆ GroupKind

enum class eve::tensor::GroupKind : uint8_t

strong

GroupKind public API.

Kernel group produced by the optimizer (AITemplate-style fusion).

A group is either:

an Elementwise chain: multiple graph nodes fused into ONE generated kernel; nodes inside the chain never materialize a buffer;
a MatMul / Conv group with an optional bias + elementwise epilogue fused into the same kernel;
a single specialized op kernel (softmax, layernorm, attention, ...);
an Alias (reshape / flatten / cast): no kernel, output aliases input.

枚举值
Elementwise
MatMul
Conv1d
Conv2d
MaxPool2d
AvgPool2d
Softmax
LayerNorm
RMSNorm
Reduce
ArgMax
Embedding
Concat
Slice
Permute
Sdpa
Resize2d
Alias

在文件 Optimizer.h 第 24 行定义.

◆ KernelVariant

enum class eve::tensor::KernelVariant

strong

Internal choice of generated matrix multiplication implementation.

枚举值
Default
TiledMatMul

在文件 KernelGenWgsl.h 第 8 行定义.

◆ OnnxElement

enum class eve::tensor::OnnxElement : int

strong

ONNX wire element types; distinct from block-quantized Tensor storage.

枚举值
Float32
UInt8
Int8
Int32
Int64
Bool

在文件 OnnxModel.h 第 18 行定义.

◆ OpType

enum class eve::tensor::OpType : uint8_t

strong

OpType public API.

枚举值
Placeholder
Const
Add
Sub
Multiply
Divide
AddScalar
SubScalar
MulScalar
DivScalar
Neg
Abs
Sqrt
Exp
Log
Sin
Cos
Tanh
Relu
Sigmoid
Gelu
Silu
PowScalar
Clamp
MaximumScalar
MinimumScalar
Where
MatMul
Transpose
Permute
Reshape
Flatten
Softmax
LogSoftmax
LayerNorm
RMSNorm
Conv1d
Conv2d
MaxPool2d
AvgPool2d
Embedding
Concat
Slice
ReduceSum
ReduceMean
ReduceMin
ReduceMax
ArgMax
Cast
ScaledDotProductAttention
Resize2d

在文件 Graph.h 第 22 行定义.

函数说明

◆ createOnnxGpuCompute()

Result< std::unique_ptr< OnnxCompute > > eve::tensor::createOnnxGpuCompute ( uint32_t compilerWorkers = 4 )

Create a reusable GPU session for the active engine Gpgpu Vulkan device, or an error.

注解: Retain and reuse this owning provider across runGpu calls. Pipelines, immutable initializers and buffer pools persist until provider destruction or Graphics resource retirement. Cache bounds: 1024 shader variants, 128 MiB initializer payload and 512 MiB device buffers. CPU shader compilation uses a bounded queue; pipeline creation and submission remain on the device thread.

参数

compilerWorkers Number of CPU compiler workers (1..8); default 4. Source is owned by each job. @lifetime Graphics retirement clears device resources before device destruction; either destruction order is supported. A retired session rejects further GPU work; create a new session for a new device. @thread Graphics thread only, no concurrent calls, reentrancy or device destruction during runGpu.

在文件 OnnxGpgpu.cpp 第 347 行定义.

引用了 compilerWorkers, eve::Diagnostic::error(), eve::Failed, gp, eve::InvalidArgument, s, state, token , 以及 eve::Unsupported.

◆ dtypeName()

const char * eve::tensor::dtypeName ( DType dtype )

Dtype name.

在文件 Tensor.cpp 第 26 行定义.

引用了 Float32, Fp16, Fp4E2M1, Fp8E4M3, Int32, Int4 , 以及 Int8.

被这些函数引用 eve::tensor::Tensor::getDtype().

◆ generateKernel()

EVENGINE_API_DOMAINS bool eve::tensor::generateKernel	(	const Graph &	graph,
		const FusedGroup &	group,
		KernelSpec &	out
	)

Generate kernel.

Generate the specialized GLSL kernel(s) for a fused group. Returns false when the group cannot be lowered (e.g. too many inputs for the fixed 8-binding descriptor layout) — callers use the CPU interpreter. This compatibility facade preserves the optimizer alternate-execution contract; new APIs should return a named status enum or Result.

在文件 KernelGen.cpp 第 921 行定义.

引用了 Alias, ArgMax, AvgPool2d, Concat, Conv1d, Conv2d, Elementwise, Embedding, graph, group, LayerNorm, MatMul, MaxPool2d, Permute, Reduce, Resize2d, RMSNorm, Sdpa, Slice , 以及 Softmax.

◆ generateKernelWgslImpl()

void eve::tensor::generateKernelWgslImpl	(	const Graph &	graph,
		const FusedGroup &	group,
		KernelSpec &	out
	)

在文件 KernelGenWgsl.cpp 第 502 行定义.

引用了 Alias, ArgMax, AvgPool2d, Concat, Conv1d, Conv2d, Elementwise, Embedding, graph, group, LayerNorm, MatMul, MaxPool2d, Permute, Reduce, Resize2d, RMSNorm, Sdpa, Slice , 以及 Softmax.

被这些函数引用 generateWgslKernel().

◆ generateMatMulVariant()

bool eve::tensor::generateMatMulVariant	(	const Graph &	graph,
		const FusedGroup &	group,
		bool	tiled,
		KernelSpec &	out
	)

Generate mat mul variant.

Generate a specific matmul variant for autotuning (tiled=false: thread-per- output naive; tiled=true: shared-memory 16x16 tiles). Only rank-2 matmuls support the tiled variant.

在文件 KernelGen.cpp 第 950 行定义.

引用了 graph, group , 以及 MatMul.

◆ generateWgslKernel()

Result< KernelSpec > eve::tensor::generateWgslKernel	(	const Graph &	graph,
		const FusedGroup &	group,
		KernelVariant	variant = `KernelVariant::Default`
	)

Lower an optimizer-produced group directly to owning WGSL source and dispatch metadata.

参数

graph	Borrowed, immutable graph; must outlive this synchronous call.
group	Valid group produced by optimizeGraph for graph.
variant	TiledMatMul is supported only for unquantized rank-2 matrix products.

返回: Owning kernel specification, or Unsupported with a code-generation diagnostic.

注解: Reentrant and CPU-only; retains no graph references and creates no GPU resources.

在文件 KernelGenWgsl.cpp 第 530 行定义.

引用了 error, eve::Diagnostic::error(), eve::Result< T >::failure(), generateKernelWgslImpl(), eve::tensor::wgsl_detail::genMatMul(), graph, group, MatMul, eve::tensor::wgsl_detail::specializeInputBindings(), eve::Result< T >::success(), TiledMatMul, eve::Unsupported , 以及 variant.

◆ gpuReduce()

EVENGINE_API_DOMAINS bool eve::tensor::gpuReduce	(	const float *	data,
		int	size,
		int	op,
		float &	outResult
	)

GPU-accelerated reduction for large eager tensors. op: 0 = sum, 1 = min, 2 = max. Returns false (caller should fall back to CPU) when Vulkan/gpgpu isn't available. This compatibility facade preserves the CPU alternative execution contract; new GPU APIs should return a named status enum or Result.

在文件 GpuBackend.cpp 第 401 行定义.

引用了 groups, parts, size , 以及 v.

被这些函数引用 eve::tensor::TF::reduceMax(), eve::tensor::TF::reduceMin() , 以及 eve::tensor::TF::reduceSum().

◆ groupKernelCount()

int eve::tensor::groupKernelCount ( const OptimizedGraph & opt )

Group kernel count.

Number of groups that need GPU kernels (i.e. not Alias).

在文件 Optimizer.cpp 第 498 行定义.

引用了 Alias, count, g , 以及 eve::tensor::OptimizedGraph::groups.

◆ Module_IMPL()

eve::tensor::Module_IMPL	(	TF	,
		new	TF()
	)

◆ optimizeGraph()

EVENGINE_API_DOMAINS OptimizedGraph eve::tensor::optimizeGraph	(	const Graph &	graph,
		int	outputNode
	)

Optimize graph.

Optimize a traced graph:

dead-code elimination + topological sort;
constant folding of elementwise chains over Const nodes;
elementwise chain fusion into FusedGroups;
matmul/conv bias + activation epilogue fusion;
static memory planning with buffer reuse.

在文件 Optimizer.cpp 第 175 行定义.

引用了 a, Add, Alias, b, best, eve::tensor::FusedGroup::biasNode, c, capacity, Cast, Const, Conv1d, Conv2d, Elementwise, eve::tensor::FusedGroup::epilogue, Flatten, g, graph, eve::tensor::OptimizedGraph::groupOrder, groups, eve::tensor::OptimizedGraph::groups, id, idx, inputs, eve::tensor::FusedGroup::inputs, eve::tensor::kernels::isElementwiseOp(), eve::tensor::FusedGroup::kind, m, MatMul, n, nodeId, eve::tensor::FusedGroup::nodes, eve::tensor::OptimizedGraph::nodeSlot, ok, order, eve::tensor::OptimizedGraph::order, eve::tensor::FusedGroup::outputNode, eve::tensor::OptimizedGraph::outputNode, p, eve::tensor::OptimizedGraph::persistentSlots, Placeholder, Reshape, s, size, slots, eve::tensor::OptimizedGraph::slotSize , 以及 u.

被这些函数引用 eve::tensor::CompiledFunction::fromFunc().

◆ parseDType()

bool eve::tensor::parseDType	(	const std::string &	name,
		DType &	out
	)

Parse d type.

在文件 Tensor.cpp 第 39 行定义.

引用了 Float32, Fp16, Fp4E2M1, Fp8E4M3, Int32, Int4, Int8 , 以及 name.

被这些函数引用 eve::tensor::TF::cast() , 以及 eve::tensor::TF::quantizeWeight().

变量说明

◆ kMaxKernelBindings

constexpr int eve::tensor::kMaxKernelBindings = 8

constexpr

Max storage bindings available per generated kernel.

在文件 KernelGen.h 第 69 行定义.

被这些函数引用 eve::tensor::glsl_detail::genElementwise() , 以及 eve::tensor::wgsl_detail::genElementwise().

命名空间

类

枚举

函数

变量

枚举类型说明

◆ DType

◆ GroupKind

◆ KernelVariant

◆ OnnxElement

◆ OpType

函数说明

◆ createOnnxGpuCompute()

◆ dtypeName()

◆ generateKernel()

◆ generateKernelWgslImpl()

◆ generateMatMulVariant()

◆ generateWgslKernel()

◆ gpuReduce()

◆ groupKernelCount()

◆ Module_IMPL()

◆ optimizeGraph()

◆ parseDType()

变量说明

◆ kMaxKernelBindings