命名空间
namespace	detail

类
struct	ByteView
	Borrowed packed 8-bit values, valid for the duration of a synchronous call. 更多...

struct	ConvShape
	Explicit 1D/2D NCHW convolution geometry; missing 1D height is one. 更多...

struct	QuantizedActivation
	ONNX unsigned activation quantization; owns bytes and scalar parameters. 更多...

函数
Result< std::vector< uint8_t > >	quantize (std::span< const float > input, float scale, int zeroPoint, bool signedValues)
	Affine quantization to int8/uint8 bytes, saturating and rounding ties to even.

Result< QuantizedActivation >	dynamicQuantize (std::span< const float > input)
	Quantize finite FP32 activations with ONNX DynamicQuantizeLinear semantics.

Result< std::vector< float > >	dequantize (ByteView input, std::span< const float > scales, std::span< const int32_t > zeros, size_t inner=1)
	Affine dequantization using scalar or per-axis scale/zero point.

Result< std::vector< int32_t > >	matmul (ByteView a, ByteView b, size_t m, size_t k, size_t n, int aZero=0, int bZero=0, OnnxCompute *compute=nullptr)
	Integer row-major [M,K] x [K,N], subtracting scalar zero points.

Result< std::vector< int32_t > >	conv (ByteView x, ByteView w, const ConvShape &shape, int xZero, std::span< const int32_t > wZeros, OnnxCompute *compute=nullptr)
	Integer Conv, supporting groups, asymmetric padding and dilation.

函数说明

◆ conv()

EVENGINE_API_DOMAINS Result< std::vector< int32_t > > eve::tensor::affine::conv	(	ByteView	x,
		ByteView	w,
		const ConvShape &	shape,
		int	xZero,
		std::span< const int32_t >	wZeros,
		OnnxCompute *	compute = `nullptr`
	)

Integer Conv, supporting groups, asymmetric padding and dilation.

返回: Owning NCHW int32 values; rejects bad geometry or accumulator overflow.

参数

compute Optional borrowed GPU provider; null selects CPU. No retry on GPU failure.

注解: Synchronous. CPU is thread-safe; GPU uses the provider thread. Padding represents real zero.

在文件 AffineQuant.cpp 第 159 行定义.

引用了 b, c, compute, eve::Diagnostic::error(), eve::Failed, eve::tensor::onnx_detail::gpuConv(), eve::InvalidArgument, s, eve::tensor::affine::detail::validateConv(), value, w, x, y , 以及 z.

被这些函数引用 eve::tensor::onnx_detail::executeQuant().

◆ dequantize()

Result< std::vector< float > > eve::tensor::affine::dequantize	(	ByteView	input,
		std::span< const float >	scales,
		std::span< const int32_t >	zeros,
		size_t	inner = `1`
	)

Affine dequantization using scalar or per-axis scale/zero point.

参数

inner Number of contiguous values following the channel axis.

返回: Owning FP32 values; InvalidArgument on inconsistent channel layout.

注解: CPU, thread-safe; all views are borrowed for the call, no callbacks.

在文件 AffineQuant.cpp 第 62 行定义.

引用了 eve::Diagnostic::error(), input, eve::InvalidArgument, scales, value , 以及 zeros.

被这些函数引用 eve::tensor::onnx_detail::executeQuant().

◆ dynamicQuantize()

EVENGINE_API_DOMAINS Result< QuantizedActivation > eve::tensor::affine::dynamicQuantize ( std::span< const float > input )

Quantize finite FP32 activations with ONNX DynamicQuantizeLinear semantics.

参数

input Borrowed flat input; not retained.

返回: Owning uint8 values, scale, zero point; InvalidArgument for nonfinite input.

注解: CPU, thread-safe, no callbacks; nearest-even rounding independent of fenv.

在文件 AffineQuant.cpp 第 41 行定义.

引用了 bytes, eve::Diagnostic::error(), eve::Result< T >::failure(), input, eve::InvalidArgument, quantize(), eve::tensor::affine::QuantizedActivation::scale, eve::Result< T >::success(), eve::tensor::affine::QuantizedActivation::values, x , 以及 eve::tensor::affine::QuantizedActivation::zeroPoint.

被这些函数引用 eve::tensor::onnx_detail::executeQuant() , 以及 eve::tensor::onnx_detail::executeQuantLstm().

◆ matmul()

EVENGINE_API_DOMAINS Result< std::vector< int32_t > > eve::tensor::affine::matmul	(	ByteView	a,
		ByteView	b,
		size_t	m,
		size_t	k,
		size_t	n,
		int	aZero = `0`,
		int	bZero = `0`,
		OnnxCompute *	compute = `nullptr`
	)

Integer row-major [M,K] x [K,N], subtracting scalar zero points.

返回: Exact owning int32 accumulators; rejects dimension and int32 overflow.

参数

compute Optional borrowed GPU provider; null selects CPU. No retry on GPU failure.

注解: Synchronous. CPU is thread-safe; GPU uses the provider thread. Weights remain packed.

在文件 AffineQuant.cpp 第 80 行定义.

引用了 a, b, c, compute, eve::Diagnostic::error(), eve::Failed, eve::tensor::onnx_detail::gpuMatmul(), eve::InvalidArgument, m, n, r, value , 以及 z.

被这些函数引用 eve::tensor::onnx_detail::executeQuant().

◆ quantize()

EVENGINE_API_DOMAINS Result< std::vector< uint8_t > > eve::tensor::affine::quantize	(	std::span< const float >	input,
		float	scale,
		int	zeroPoint,
		bool	signedValues
	)

Affine quantization to int8/uint8 bytes, saturating and rounding ties to even.

返回: Owning packed bytes; rejects invalid scale, zero point or nonfinite input.

注解: CPU, thread-safe; spans are borrowed only during the call.

在文件 AffineQuant.cpp 第 24 行定义.

引用了 eve::Diagnostic::error(), input, eve::InvalidArgument, q, scale , 以及 x.

被这些函数引用 dynamicQuantize() , 以及 eve::tensor::onnx_detail::executeQuant().

命名空间

类

函数

函数说明

◆ conv()

◆ dequantize()

◆ dynamicQuantize()

◆ matmul()

◆ quantize()