eve::tensor::kernels 命名空间参考

函数
float	applyUnary (OpType t, float x, float s0, float s1)
bool	isElementwiseOp (OpType t)
bool	broadcastShape (const int *aDims, int aRank, const int *bDims, int bRank, int *outDims, int &outRank)
void	binaryOp (OpType type, const float *a, const int *aDims, int aRank, const float *b, const int *bDims, int bRank, float *out, const int *outDims, int outRank)
void	unaryOp (OpType type, const float *in, int count, float *out, float s0, float s1)
void	softmax (const float *in, const int *dims, int rank, int axis, bool logMode, float *out)
void	layernorm (const float *in, int rows, int cols, const float *scale, const float *bias, float eps, float *out)
void	rmsnorm (const float *in, int rows, int cols, const float *scale, float eps, float *out)
void	conv1d (const float *x, const int *xDims, const float *w, const int *wDims, const float *bias, int stride, int pad, float *out)
void	conv2d (const float *x, const int *xDims, const float *w, const int *wDims, const float *bias, int stride, int pad, float *out)
void	maxpool2d (const float *in, const int *dims, int ksize, int stride, int pad, float *out)
void	avgpool2d (const float *in, const int *dims, int ksize, int stride, int pad, float *out)
void	embedding (const float *table, int vocab, int dim, const float *indices, int count, float *out)
void	concat (const float *const *ins, const int *const *inDims, const int *inRanks, int n, int axis, float *out, const int *outDims, int outRank)
void	sliceOp (const float *in, const int *inDims, int inRank, int axis, int begin, int end, float *out, const int *outDims, int outRank)
void	permute (const float *in, const int *inDims, int rank, const int *order, float *out, const int *outDims)
void	reduceAxis (OpType type, const float *in, const int *dims, int rank, int axis, float *out, const int *outDims, int outRank)
void	argmax (const float *in, const int *dims, int rank, int axis, float *out, const int *outDims, int outRank)
void	sdpa (const float *q, const float *k, const float *v, const float *mask, int B, int H, int T, int S, int D, float scale, float *out)
void	resize2d (const float *in, const int *inDims, int outW, int outH, int mode, float *out)

详细描述

CPU reference kernels shared by the eager path and the graph interpreter.

All shapes are static, row-major, rank <= Tensor::kMaxRank. Every entry point is a plain loop implementation: correctness / portability first, throughput is provided by the generated GPU kernels when a Vulkan device is available.

函数说明

applyUnary()

float eve::tensor::kernels::applyUnary	(	OpType	type,
		float	x,
		float	s0,
		float	s1
	)

Apply a scalar unary op (elementwise) to one value.

在文件 CpuKernels.cpp 第 55 行定义.

引用了 eve::tensor::Abs, eve::tensor::AddScalar, eve::tensor::Clamp, eve::tensor::Cos, eve::tensor::DivScalar, eve::tensor::Exp, eve::tensor::Gelu, eve::tensor::Log, eve::tensor::MaximumScalar, eve::tensor::MinimumScalar, eve::tensor::MulScalar, eve::tensor::Neg, eve::tensor::PowScalar, eve::tensor::Relu, eve::tensor::Sigmoid, eve::tensor::Silu, eve::tensor::Sin, eve::tensor::Sqrt, eve::tensor::SubScalar, eve::tensor::Tanh , 以及 x.

被这些函数引用 unaryOp().

argmax()

void eve::tensor::kernels::argmax	(	const float *	in,
		const int *	dims,
		int	rank,
		int	axis,
		float *	out,
		const int *	outDims,
		int	outRank
	)

在文件 CpuKernels.cpp 第 428 行定义.

引用了 reduce , 以及 v.

被这些函数引用 eve::tensor::TF::argmax().

avgpool2d()

void eve::tensor::kernels::avgpool2d	(	const float *	in,
		const int *	dims,
		int	ksize,
		int	stride,
		int	pad,
		float *	out
	)

在文件 CpuKernels.cpp 第 302 行定义.

引用了 c, n , 以及 valid.

被这些函数引用 eve::tensor::TF::avgpool2d().

binaryOp()

void eve::tensor::kernels::binaryOp	(	OpType	type,
		const float *	a,
		const int *	aDims,
		int	aRank,
		const float *	b,
		const int *	bDims,
		int	bRank,
		float *	out,
		const int *	outDims,
		int	outRank
	)

Elementwise binary op with broadcasting.

在文件 CpuKernels.cpp 第 136 行定义.

引用了 a, b, eve::tensor::Tensor::kMaxRank , 以及 type.

broadcastShape()

bool eve::tensor::kernels::broadcastShape	(	const int *	aDims,
		int	aRank,
		const int *	bDims,
		int	bRank,
		int *	outDims,
		int &	outRank
	)

Broadcast two shapes to a common shape. Returns false when incompatible.

在文件 CpuKernels.cpp 第 120 行定义.

引用了 eve::tensor::Tensor::kMaxRank.

被这些函数引用 eve::tensor::Func::emitBinary().

concat()

void eve::tensor::kernels::concat	(	const float *const *	ins,
		const int *const *	inDims,
		const int *	inRanks,
		int	n,
		int	axis,
		float *	out,
		const int *	outDims,
		int	outRank
	)

concat n (2..4) tensors along axis; all other dims must match out dims.

在文件 CpuKernels.cpp 第 336 行定义.

引用了 n , 以及 v.

conv1d()

void eve::tensor::kernels::conv1d	(	const float *	x,
		const int *	xDims,
		const float *	w,
		const int *	wDims,
		const float *	bias,
		int	stride,
		int	pad,
		float *	out
	)

在文件 CpuKernels.cpp 第 230 行定义.

引用了 c, f, n, w , 以及 x.

被这些函数引用 eve::tensor::TF::conv1dBias().

conv2d()

void eve::tensor::kernels::conv2d	(	const float *	x,
		const int *	xDims,
		const float *	w,
		const int *	wDims,
		const float *	bias,
		int	stride,
		int	pad,
		float *	out
	)

在文件 CpuKernels.cpp 第 252 行定义.

引用了 c, f, n, w , 以及 x.

被这些函数引用 eve::tensor::TF::conv2dBias().

embedding()

void eve::tensor::kernels::embedding	(	const float *	table,
		int	vocab,
		int	dim,
		const float *	indices,
		int	count,
		float *	out
	)

table [vocab x dim], indices [count] -> out [count x dim]. Indices are clamped.

在文件 CpuKernels.cpp 第 325 行定义.

引用了 d , 以及 idx.

被这些函数引用 eve::tensor::TF::embedding().

isElementwiseOp()

bool eve::tensor::kernels::isElementwiseOp

(

OpType

type

)

Whether the op is a fusible elementwise op (also used by the optimizer).

在文件 CpuKernels.cpp 第 87 行定义.

引用了 eve::tensor::Abs, eve::tensor::Add, eve::tensor::AddScalar, eve::tensor::Clamp, eve::tensor::Cos, eve::tensor::Divide, eve::tensor::DivScalar, eve::tensor::Exp, eve::tensor::Gelu, eve::tensor::Log, eve::tensor::MaximumScalar, eve::tensor::MinimumScalar, eve::tensor::MulScalar, eve::tensor::Multiply, eve::tensor::Neg, eve::tensor::PowScalar, eve::tensor::Relu, eve::tensor::Sigmoid, eve::tensor::Silu, eve::tensor::Sin, eve::tensor::Sqrt, eve::tensor::Sub, eve::tensor::SubScalar, eve::tensor::Tanh , 以及 eve::tensor::Where.

被这些函数引用 eve::tensor::optimizeGraph().

layernorm()

void eve::tensor::kernels::layernorm	(	const float *	in,
		int	rows,
		int	cols,
		const float *	scale,
		const float *	bias,
		float	eps,
		float *	out
	)

在文件 CpuKernels.cpp 第 187 行定义.

引用了 c, scale, v , 以及 y.

被这些函数引用 eve::tensor::TF::layernorm() , 以及 eve::tensor::TF::layernormWB().

maxpool2d()

void eve::tensor::kernels::maxpool2d	(	const float *	in,
		const int *	dims,
		int	ksize,
		int	stride,
		int	pad,
		float *	out
	)

在文件 CpuKernels.cpp 第 282 行定义.

引用了 c , 以及 n.

被这些函数引用 eve::tensor::TF::maxpool2d().

permute()

void eve::tensor::kernels::permute	(	const float *	in,
		const int *	inDims,
		int	rank,
		const int *	order,
		float *	out,
		const int *	outDims
	)

在文件 CpuKernels.cpp 第 382 行定义.

引用了 idx , 以及 eve::tensor::Tensor::kMaxRank.

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

reduceAxis()

void eve::tensor::kernels::reduceAxis	(	OpType	type,
		const float *	in,
		const int *	dims,
		int	rank,
		int	axis,
		float *	out,
		const int *	outDims,
		int	outRank
	)

sum/mean/min/max along axis (out dims must already be computed by caller).

在文件 CpuKernels.cpp 第 401 行定义.

引用了 reduce, eve::tensor::ReduceMax, eve::tensor::ReduceMean, eve::tensor::ReduceMin, eve::tensor::ReduceSum, type , 以及 v.

resize2d()

void eve::tensor::kernels::resize2d	(	const float *	in,
		const int *	inDims,
		int	outW,
		int	outH,
		int	mode,
		float *	out
	)

resize2d NCHW; mode: 0 = nearest, 1 = bilinear.

在文件 CpuKernels.cpp 第 488 行定义.

引用了 c, n, value, x , 以及 y.

被这些函数引用 eve::tensor::TF::resize2d().

rmsnorm()

void eve::tensor::kernels::rmsnorm	(	const float *	in,
		int	rows,
		int	cols,
		const float *	scale,
		float	eps,
		float *	out
	)

在文件 CpuKernels.cpp 第 211 行定义.

引用了 c , 以及 scale.

被这些函数引用 eve::tensor::TF::rmsnorm() , 以及 eve::tensor::TF::rmsnormW().

sdpa()

void eve::tensor::kernels::sdpa	(	const float *	q,
		const float *	k,
		const float *	v,
		const float *	mask,
		int	B,
		int	H,
		int	T,
		int	S,
		int	D,
		float	scale,
		float *	out
	)

Scaled dot-product attention over the last two dims: q [B,H,T,D], k [B,H,S,D], v [B,H,S,D], mask [B,H,T,S] (optional, may be null).

在文件 CpuKernels.cpp 第 454 行定义.

引用了 b, d, h, s, scale , 以及 v.

被这些函数引用 eve::tensor::TF::sdpaMasked().

sliceOp()

void eve::tensor::kernels::sliceOp	(	const float *	in,
		const int *	inDims,
		int	inRank,
		int	axis,
		int	begin,
		int	end,
		float *	out,
		const int *	outDims,
		int	outRank
	)

在文件 CpuKernels.cpp 第 365 行定义.

被这些函数引用 eve::tensor::TF::slice().

softmax()

void eve::tensor::kernels::softmax	(	const float *	in,
		const int *	dims,
		int	rank,
		int	axis,
		bool	logMode,
		float *	out
	)

在文件 CpuKernels.cpp 第 163 行定义.

引用了 reduce , 以及 v.

被这些函数引用 eve::tensor::TF::logSoftmax() , 以及 eve::tensor::TF::softmax().

unaryOp()

void eve::tensor::kernels::unaryOp	(	OpType	type,
		const float *	in,
		int	count,
		float *	out,
		float	s0,
		float	s1
	)

Elementwise unary / scalar op.

在文件 CpuKernels.cpp 第 159 行定义.

引用了 applyUnary() , 以及 type.

被这些函数引用 eve::tensor::Tensor::addScalar(), eve::tensor::Tensor::clamp(), eve::tensor::Tensor::divScalar(), eve::tensor::Tensor::maximumScalar(), eve::tensor::Tensor::minimumScalar(), eve::tensor::Tensor::mulScalar(), eve::tensor::Tensor::powScalar() , 以及 eve::tensor::Tensor::subScalar().