DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp > Struct Template Reference

DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp > Struct Template Reference#

Composable Kernel: ck::tensor_operation::device::DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp > Struct Template Reference

Convolution Tensor Rearrange. More...

#include <device_conv_tensor_rearrange.hpp>

Inheritance diagram for ck::tensor_operation::device::DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp >:

Public Member Functions
virtual std::unique_ptr< BaseArgument >	MakeArgumentPointer (const void p_in, void p_out, const ck::index_t G, const ck::index_t N, const ck::index_t C, const std::array< index_t, NDimSpatial > &input_spatial_lengths, const std::array< index_t, NDimSpatial > &filter_spatial_lengths, const std::array< index_t, NDimSpatial > &output_spatial_lengths, const std::array< index_t, NDimSpatial+3 > &image_g_n_c_wis_strides, const std::array< index_t, 3 > &gemm_g_m_k_strides, const std::array< index_t, NDimSpatial > &conv_filter_strides, const std::array< index_t, NDimSpatial > &conv_filter_dilations, const std::array< index_t, NDimSpatial > &input_left_pads, const std::array< index_t, NDimSpatial > &input_right_pads)=0
	Make argument pointer for image to column.
virtual std::unique_ptr< BaseInvoker >	MakeInvokerPointer ()=0
Public Member Functions inherited from ck::tensor_operation::device::BaseOperator
	BaseOperator ()=default
	BaseOperator (const BaseOperator &)=default
BaseOperator &	operator= (const BaseOperator &)=default
virtual bool	IsSupportedArgument (const BaseArgument *)
virtual std::string	GetTypeString () const
virtual std::string	GetInstanceString () const
virtual std::string	GetTypeIdName () const
virtual std::optional< std::string >	GetObjectName () const
virtual std::optional< std::string >	GetTemplateInfo () const
virtual std::string	GetTypeIdHashCode () const
virtual size_t	GetWorkSpaceSize (const BaseArgument *) const
virtual void	SetWorkSpacePointer (BaseArgument p_arg, void p_workspace, const StreamConfig &=StreamConfig{}) const
virtual	~BaseOperator ()

Detailed Description

template<index_t NDimSpatial, typename ImageLayout, typename InputDataType, typename OutputDataType, typename ConvTensorRearrangeOp>
struct ck::tensor_operation::device::DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp >

Convolution Tensor Rearrange.

This Device operator supports converting an image to the GEMM representation (Image to Column) and converting a GEMM form to the image (Column to Image). Supported layouts: [G, N, Di, Hi, Wi, C] <-> [G, N * Do * Ho * Wo, Z * Y * X * C] [N, Di, Hi, Wi, G, C] <-> [N * Do * Ho * Wo, G, Z * Y * X * C]

Template Parameters

NDimSpatial	Number of spatial dimensions.
ImageLayout	Input Layout.
InputDataType	Input Data Type.
OutputDataType	Output Data Type.
ConvTensorRearrangeOp	Operation type: ImageToColumn, ColumnToImage.

Member Function Documentation

◆ MakeArgumentPointer()

template<index_t NDimSpatial, typename ImageLayout, typename InputDataType, typename OutputDataType, typename ConvTensorRearrangeOp>

virtual std::unique_ptr< BaseArgument > ck::tensor_operation::device::DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp >::MakeArgumentPointer	(	const void *	p_in,
		void *	p_out,
		const ck::index_t	G,
		const ck::index_t	N,
		const ck::index_t	C,
		const std::array< index_t, NDimSpatial > &	input_spatial_lengths,
		const std::array< index_t, NDimSpatial > &	filter_spatial_lengths,
		const std::array< index_t, NDimSpatial > &	output_spatial_lengths,
		const std::array< index_t, NDimSpatial+3 > &	image_g_n_c_wis_strides,
		const std::array< index_t, 3 > &	gemm_g_m_k_strides,
		const std::array< index_t, NDimSpatial > &	conv_filter_strides,
		const std::array< index_t, NDimSpatial > &	conv_filter_dilations,
		const std::array< index_t, NDimSpatial > &	input_left_pads,
		const std::array< index_t, NDimSpatial > &	input_right_pads )

pure virtual

Make argument pointer for image to column.

Parameters

p_in	A pointer to the device memory of the input image.
p_out	A pointer to the device memory of the output.
G	Convolution number of groups.
N	Convolution batch size.
C	Convolution number of channels.
input_spatial_lengths	Input spatial lengths.
filter_spatial_lengths	Filter spatial lengths.
output_spatial_lengths	Output spatial lengths.
image_g_n_c_wis_strides	Image strides in order [G, N, C, D, H, W].
gemm_g_m_k_strides	Gemm form strides.
conv_filter_strides	Convolution filter strides.
conv_filter_dilations	Convolution filter dilations.
input_left_pads	Convolution left pads.
input_right_pads	Convolution right pads.

Returns: Pointer to the argument.

Implemented in ck::tensor_operation::device::DeviceColumnToImageImpl< NDimSpatial, ImageLayout, InputDataType, OutputDataType, BlockSize, MPerBlock, KPerBlock, ThreadClusterLengths, ScalarPerVector,, bool, type >, and ck::tensor_operation::device::DeviceImageToColumnImpl< NDimSpatial, ImageLayout, InputDataType, OutputDataType, BlockSize, MPerBlock, KPerBlock, ThreadClusterLengths, ScalarPerVector,, bool, type >.

◆ MakeInvokerPointer()

template<index_t NDimSpatial, typename ImageLayout, typename InputDataType, typename OutputDataType, typename ConvTensorRearrangeOp>

virtual std::unique_ptr< BaseInvoker > ck::tensor_operation::device::DeviceConvTensorRearrange< NDimSpatial, ImageLayout, InputDataType, OutputDataType, ConvTensorRearrangeOp >::MakeInvokerPointer ( )