载入中...
搜索中...
未找到
类 | 函数 | 变量
eve::tensor::glsl_detail 命名空间参考

类

struct  ChainContext
 

函数

std::string header (int localX, int localY=1)
 Header.
 
std::string bufferDecl (int binding, const char *name)
 Buffer decl.
 
std::string bufferDeclUint (int binding, const char *name)
 
std::string emitQuantizedBVal (DType dt, int group, const char *bufName="b")
 
std::string pushConstant ()
 Pushes constant.
 
int groupsFor (int count)
 Groups for.
 
std::string scalarStr (float value)
 Scalar str.
 
void emitInputIndexExprs (std::ostringstream &os, const Graph &g, const FusedGroup &grp, std::vector< std::string > &indexExprs)
 
bool genElementwise (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genMatMul (const Graph &g, const FusedGroup &grp, bool tiled, KernelSpec &out)
 
bool genConv (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genPool (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genSoftmax (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genNorm (const Graph &g, const FusedGroup &grp, bool rms, KernelSpec &out)
 
bool genReduceOrArgmax (const Graph &g, const FusedGroup &grp, bool argmax, KernelSpec &out)
 
bool genEmbedding (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genConcat (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genSlice (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genPermute (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
bool genSdpa (const Graph &g, const FusedGroup &grp, KernelSpec &out)
 
void genResize2d (const Graph &, const FusedGroup &, KernelSpec &)
 Gen resize 2 d.
 

变量

constexpr int kLocalSize = 256
 

函数说明

◆ bufferDecl()

std::string eve::tensor::glsl_detail::bufferDecl ( int  binding,
const char *  name 
)

Buffer decl.

在文件 KernelGen.cpp 第 22 行定义.

引用了 binding , 以及 name.

被这些函数引用 genConcat(), genConv(), genElementwise(), genEmbedding(), genMatMul(), genNorm(), genPermute(), genPool(), genReduceOrArgmax(), genResize2d(), genSdpa(), genSlice() , 以及 genSoftmax().

◆ bufferDeclUint()

std::string eve::tensor::glsl_detail::bufferDeclUint ( int  binding,
const char *  name 
)

在文件 KernelGen.cpp 第 29 行定义.

引用了 binding , 以及 name.

被这些函数引用 genEmbedding() , 以及 genMatMul().

◆ emitInputIndexExprs()

void eve::tensor::glsl_detail::emitInputIndexExprs ( std::ostringstream &  os,
const Graph &  g,
const FusedGroup &  grp,
std::vector< std::string > &  indexExprs 
)

Per-input flat-index expression for an elementwise group. When an input has the exact output shape, the identity index is used; otherwise a broadcast index is decomposed from the output flat index with baked strides.

在文件 KernelGen.cpp 第 199 行定义.

引用了 d, eve::tensor::GraphNode::dims, g, eve::tensor::FusedGroup::inputs, eve::tensor::FusedGroup::outputNode, pad, eve::tensor::GraphNode::rank, s , 以及 t.

被这些函数引用 genElementwise().

◆ emitQuantizedBVal()

std::string eve::tensor::glsl_detail::emitQuantizedBVal ( DType  dt,
int  group,
const char *  bufName = "b" 
)

GLSL dequantizer float bval(uint elemIdx) for a packed weight buffer bufName (byte-addressed inside a uint SSBO) plus optional per-group scales bs.

在文件 KernelGen.cpp 第 41 行定义.

引用了 eve::tensor::Fp16, eve::tensor::Fp4E2M1, eve::tensor::Fp8E4M3, g, group, eve::tensor::Int4 , 以及 eve::tensor::Int8.

◆ genConcat()

bool eve::tensor::glsl_detail::genConcat ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 749 行定义.

引用了 bufferDecl(), eve::tensor::GraphNode::dims, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::KernelSpec::inputCount, eve::tensor::FusedGroup::inputs, kLocalSize, n, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::GraphNode::rank, eve::tensor::GraphNode::size , 以及 sz.

◆ genConv()

bool eve::tensor::glsl_detail::genConv ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 387 行定义.

引用了 eve::tensor::FusedGroup::biasNode, bufferDecl(), eve::tensor::Conv1d, eve::tensor::GraphNode::dims, eve::tensor::FusedGroup::epilogue, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::GraphNode::i1, eve::tensor::GraphNode::in0, eve::tensor::GraphNode::in1, eve::tensor::GraphNode::in2, eve::tensor::KernelSpec::inputCount, eve::tensor::FusedGroup::inputs, kLocalSize, eve::tensor::FusedGroup::nodes, eve::tensor::FusedGroup::outputNode, pad, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::GraphNode::type , 以及 u.

◆ genElementwise()

bool eve::tensor::glsl_detail::genElementwise ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 247 行定义.

引用了 bufferDecl(), emitInputIndexExprs(), g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::KernelSpec::inputCount, eve::tensor::FusedGroup::inputs, kLocalSize, eve::tensor::kMaxKernelBindings, eve::tensor::FusedGroup::nodes, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::GraphNode::size , 以及 u.

◆ genEmbedding()

bool eve::tensor::glsl_detail::genEmbedding ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 712 行定义.

引用了 bufferDecl(), bufferDeclUint(), eve::tensor::GraphNode::constBytes, eve::tensor::GraphNode::dims, eve::tensor::GraphNode::dtype, eve::tensor::Fp16, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, eve::tensor::q::isQuantDType(), kLocalSize, eve::tensor::KernelSpec::outputBinding, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::KernelSpec::qDtype, eve::tensor::GraphNode::qGroup, eve::tensor::KernelSpec::qGroup, eve::tensor::KernelSpec::scalesBinding , 以及 eve::tensor::GraphNode::size.

◆ genMatMul()

bool eve::tensor::glsl_detail::genMatMul ( const Graph &  g,
const FusedGroup &  grp,
bool  tiled,
KernelSpec &  out 
)

Emit matmul kernel; tiled selects the shared-memory 16x16 variant.

在文件 KernelGen.cpp 第 273 行定义.

引用了 eve::tensor::FusedGroup::biasNode, bufferDecl(), bufferDeclUint(), eve::tensor::FusedGroup::epilogue, eve::tensor::Fp16, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, eve::tensor::KernelSpec::groupsY2, header(), eve::tensor::GraphNode::in0, eve::tensor::GraphNode::in1, eve::tensor::KernelSpec::inputCount, eve::tensor::FusedGroup::inputs, eve::tensor::q::isQuantDType(), kLocalSize, m, n, eve::tensor::FusedGroup::nodes, eve::tensor::KernelSpec::outputBinding, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::KernelSpec::qDtype, eve::tensor::KernelSpec::qGroup, eve::tensor::GraphNode::rank, eve::tensor::KernelSpec::scalesBinding , 以及 u.

◆ genNorm()

bool eve::tensor::glsl_detail::genNorm ( const Graph &  g,
const FusedGroup &  grp,
bool  rms,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 578 行定义.

引用了 bufferDecl(), cols, g, groupsFor(), eve::tensor::KernelSpec::groupsX1, eve::tensor::KernelSpec::groupsX2, eve::tensor::FusedGroup::hasBias, eve::tensor::FusedGroup::hasScale, header(), eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, eve::tensor::KernelSpec::inputsReadPass1, kLocalSize, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), rows, eve::tensor::GraphNode::s0, scalarStr(), eve::tensor::GraphNode::size, eve::tensor::KernelSpec::statsCount, eve::tensor::KernelSpec::statsSize , 以及 eve::tensor::KernelSpec::twoPass.

◆ genPermute()

bool eve::tensor::glsl_detail::genPermute ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 820 行定义.

引用了 bufferDecl(), eve::tensor::GraphNode::dims, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, kLocalSize, eve::tensor::Tensor::kMaxRank, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, eve::tensor::GraphNode::perm, pushConstant(), eve::tensor::GraphNode::rank , 以及 eve::tensor::GraphNode::size.

◆ genPool()

bool eve::tensor::glsl_detail::genPool ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 464 行定义.

引用了 bufferDecl(), eve::tensor::GraphNode::dims, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::GraphNode::i1, eve::tensor::GraphNode::i2, eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, kLocalSize, eve::tensor::MaxPool2d, eve::tensor::FusedGroup::outputNode, pad, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::GraphNode::size , 以及 eve::tensor::GraphNode::type.

◆ genReduceOrArgmax()

bool eve::tensor::glsl_detail::genReduceOrArgmax ( const Graph &  g,
const FusedGroup &  grp,
bool  argmax,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 653 行定义.

引用了 bufferDecl(), g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, kLocalSize, eve::tensor::FusedGroup::op, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), reduce, eve::tensor::ReduceMax, eve::tensor::ReduceMean, eve::tensor::ReduceMin, eve::tensor::ReduceSum , 以及 scalarStr().

◆ genResize2d()

void eve::tensor::glsl_detail::genResize2d ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

Gen resize 2 d.

在文件 KernelGenResize.cpp 第 5 行定义.

引用了 bufferDecl(), eve::tensor::GraphNode::dims, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, kLocalSize, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), scalarStr() , 以及 eve::tensor::GraphNode::size.

◆ genSdpa()

bool eve::tensor::glsl_detail::genSdpa ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 853 行定义.

引用了 bufferDecl(), eve::tensor::GraphNode::dims, g, eve::tensor::KernelSpec::groupsX2, eve::tensor::KernelSpec::groupsY2, header(), eve::tensor::GraphNode::in0, eve::tensor::GraphNode::in1, eve::tensor::KernelSpec::inputCount, eve::tensor::FusedGroup::masked, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::GraphNode::s0, scalarStr() , 以及 scale.

◆ genSlice()

bool eve::tensor::glsl_detail::genSlice ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 792 行定义.

引用了 begin, bufferDecl(), eve::tensor::GraphNode::dims, end, g, groupsFor(), eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::GraphNode::i1, eve::tensor::GraphNode::i2, eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, kLocalSize, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), eve::tensor::GraphNode::rank , 以及 eve::tensor::GraphNode::size.

◆ genSoftmax()

bool eve::tensor::glsl_detail::genSoftmax ( const Graph &  g,
const FusedGroup &  grp,
KernelSpec &  out 
)

在文件 KernelGen.cpp 第 512 行定义.

引用了 bufferDecl(), g, groupsFor(), eve::tensor::KernelSpec::groupsX1, eve::tensor::KernelSpec::groupsX2, header(), eve::tensor::GraphNode::i0, eve::tensor::GraphNode::in0, eve::tensor::KernelSpec::inputCount, eve::tensor::KernelSpec::inputsReadPass1, kLocalSize, eve::tensor::FusedGroup::logMode, eve::tensor::FusedGroup::outputNode, eve::tensor::KernelSpec::pass1, eve::tensor::KernelSpec::pass2, pushConstant(), reduce, rows, eve::tensor::GraphNode::size, eve::tensor::KernelSpec::statsCount, eve::tensor::KernelSpec::statsSize , 以及 eve::tensor::KernelSpec::twoPass.

◆ groupsFor()

int eve::tensor::glsl_detail::groupsFor ( int  count)

Groups for.

在文件 KernelGen.cpp 第 105 行定义.

引用了 count , 以及 kLocalSize.

被这些函数引用 genConcat(), genConv(), genElementwise(), genEmbedding(), genMatMul(), genNorm(), genPermute(), genPool(), genReduceOrArgmax(), genResize2d(), genSlice() , 以及 genSoftmax().

◆ header()

std::string eve::tensor::glsl_detail::header ( int  localX,
int  localY 
)

Header.

在文件 KernelGen.cpp 第 13 行定义.

被这些函数引用 genConcat(), genConv(), genElementwise(), genEmbedding(), genMatMul(), genNorm(), genPermute(), genPool(), genReduceOrArgmax(), genResize2d(), genSdpa(), genSlice() , 以及 genSoftmax().

◆ pushConstant()

std::string eve::tensor::glsl_detail::pushConstant ( )

Pushes constant.

在文件 KernelGen.cpp 第 101 行定义.

被这些函数引用 genConcat(), genConv(), genElementwise(), genEmbedding(), genMatMul(), genNorm(), genPermute(), genPool(), genReduceOrArgmax(), genResize2d(), genSdpa(), genSlice() , 以及 genSoftmax().

◆ scalarStr()

std::string eve::tensor::glsl_detail::scalarStr ( float  v)

Scalar str.

在文件 KernelGen.cpp 第 107 行定义.

引用了 v.

被这些函数引用 eve::tensor::glsl_detail::ChainContext::emit(), genNorm(), genReduceOrArgmax(), genResize2d() , 以及 genSdpa().

变量说明

◆ kLocalSize

constexpr int eve::tensor::glsl_detail::kLocalSize = 256
inlineconstexpr

在文件 KernelGenInternal.h 第 6 行定义.

被这些函数引用 genConcat(), genConv(), genElementwise(), genEmbedding(), genMatMul(), genNorm(), genPermute(), genPool(), genReduceOrArgmax(), genResize2d(), genSlice(), genSoftmax() , 以及 groupsFor().