GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ > Struct Template Reference

GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ > Struct Template Reference#

Composable Kernel: ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ > Struct Template Reference

The Grouped Convolution Backward Weight kernel template. More...

#include <grouped_convolution_backward_weight_kernel.hpp>

Classes
struct	SplitKBatchOffset

Public Types
using	TilePartitioner = remove_cvref_t<TilePartitioner_>
using	GemmPipeline = remove_cvref_t<GemmPipeline_>
using	EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>
using	GemmALayout = remove_cvref_t<typename GemmPipeline::ALayout>
using	GemmBLayout = remove_cvref_t<typename GemmPipeline::BLayout>
using	GemmCLayout = remove_cvref_t<typename GemmPipeline::CLayout>
using	InLayout = remove_cvref_t<typename GroupedConvTraitsType_::InLayout>
using	WeiLayout = remove_cvref_t<typename GroupedConvTraitsType_::WeiLayout>
using	OutLayout = remove_cvref_t<typename GroupedConvTraitsType_::OutLayout>
using	DsLayout = remove_cvref_t<typename GroupedConvTraitsType_::DsLayout>
using	GemmDsLayout = remove_cvref_t<typename EpiloguePipeline::DsLayout>
using	OutDataType = remove_cvref_t<typename GemmPipeline::ADataType>
using	InDataType = remove_cvref_t<typename GemmPipeline::BDataType>
using	DsDataType = remove_cvref_t<typename EpiloguePipeline::DsDataType>
using	WeiDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>
using	GroupedConvBwdWeightKernelArgsSpecialized

Public Member Functions
CK_TILE_DEVICE void	operator() (GroupedConvBwdWeightKernelArgsSpecialized kargs) const

Static Public Member Functions
static CK_TILE_HOST const std::string	GetName ()
static CK_TILE_HOST constexpr auto	GridSize (const GroupedConvBwdWeightKernelArgsSpecialized &kargs)
static CK_TILE_HOST constexpr auto	BlockSize ()
static CK_TILE_HOST constexpr GroupedConvBwdWeightKernelArgsSpecialized	MakeKernelArgs (const GroupedConvBwdWeightHostArgs &hostArgs)
static CK_TILE_HOST_DEVICE constexpr index_t	GetSmemSize ()
static CK_TILE_HOST auto	Preprocess (const GroupedConvBwdWeightKernelArgsSpecialized &kargs, const stream_config &s)
static CK_TILE_HOST bool	IsSupportedArgument (const GroupedConvBwdWeightKernelArgsSpecialized &kargs)
template<memory_operation_enum DstInMemOp = memory_operation_enum::set>
static CK_TILE_DEVICE auto	MakeGemmTensorViews (const OutDataType a_ptr, const InDataType b_ptr, const std::array< const void , NumDTensor > &ds_ptr, WeiDataType c_ptr, const GroupedConvBwdWeightKernelArgsSpecialized &kargs)
template<typename TensorView>
static CK_TILE_DEVICE auto	MakeGemmPadViews (const TensorView &views, const index_t k_batch)
template<typename PadView>
static CK_TILE_DEVICE auto	MakeGemmTileWindows (const PadView &views, const index_t i_m, const index_t i_n, const index_t i_k)
	Create views to the data that each workgroup will process.
static CK_TILE_DEVICE void	RunGemm (const OutDataType a_ptr, const InDataType b_ptr, const std::array< const void , NumDTensor > &ds_ptr, WeiDataType c_ptr, void *smem_ptr_0, const GroupedConvBwdWeightKernelArgsSpecialized &kargs, const index_t num_loop, const index_t block_idx_m, const index_t block_idx_n, const index_t block_idx_k)
	Runs single GEMM problem cooperatively by whole workgroup.
static CK_TILE_DEVICE void	RunGemm2LDS (const OutDataType a_ptr, const InDataType b_ptr, const std::array< const void , NumDTensor > &ds_ptr, WeiDataType c_ptr, void __restrict__ smem_ptr_0, void __restrict__ smem_ptr_1, const GroupedConvBwdWeightKernelArgsSpecialized &kargs, const index_t num_loop, const index_t block_idx_m, const index_t block_idx_n, const index_t block_idx_k)
	Runs single GEMM problem cooperatively by whole workgroup.

Static Public Attributes
static constexpr index_t	NDimSpatial = GroupedConvTraitsType_::NDimSpatial_
static constexpr ConvolutionSpecialization	ConvSpecialization
static constexpr index_t	NumDTensor = GroupedConvTraitsType_::NumDTensor
static constexpr index_t	kBlockSize = GemmPipeline::BlockSize
static constexpr bool	IsSplitKSupported = true
static constexpr auto	I0 = number<0>()
static constexpr auto	I1 = number<1>()
static constexpr auto	I2 = number<2>()
static constexpr auto	I3 = number<3>()

Detailed Description

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
struct ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >

The Grouped Convolution Backward Weight kernel template.

Overview

This class provides the grouped convolution backward weight kernel template. By semantic division of Implicit GEMM algorithm into following parts we achieve flexible, versatile and robust kernel implementation.

Prolog - The start of GEMM kernel implementation in operator() function call operator" which determines the work scope of each workgroup. @li @b GemmPipeline - The core part @a "heart" of matrix multiplication algorithm. This is the place where each workgroup is loading data from global memory and carrying out dot products. @li @b Epilogue - The @a "final" part of matrix multiplication implementation responsible for storing results to global memory. This is also the place where any additional operator fusion may take place. Additionally both @ref GemmPipeline_ "GemmPipeline" and @ref EpiloguePipeline_ "EpiloguePipeline" are parameterized with so called @a Policy which determines all internal details of those functional parts. You can think of it like both gemm and epilogue pipelines provides the control-flow logic controlled by policies. Moreover the policy is responsible for definition of all necessary data layouts and thread's work distribution. @tparam GroupedConvTraitsType_ The type of class providing traits for grouped convolution. @tparam TilePartitioner_ The type of class providing mapping of workgroup index into the output data tile to be calculated. It determines the workgroup to data relationship (or in other words - which data would be processed and calculated by which workgroup). @tparam GemmPipeline_ The type of class which provides the core part of matrix multiplication. This class should provide implementation of data loading from global memory and performing block-wise matrix multiplication. You can think of it as a work done by single workgroup point of view. @tparam EpiloguePipeline_ The type of class providing the final part of matrix multiplication implementation. It is responsible for storing results calculated by @ref GemmPipeline_ "GemmPipeline" to the output C tensor in global memory.

Member Typedef Documentation

◆ DsDataType

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::DsDataType = remove_cvref_t<typename EpiloguePipeline::DsDataType>

◆ DsLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::DsLayout = remove_cvref_t<typename GroupedConvTraitsType_::DsLayout>

◆ EpiloguePipeline

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>

◆ GemmALayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GemmALayout = remove_cvref_t<typename GemmPipeline::ALayout>

◆ GemmBLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GemmBLayout = remove_cvref_t<typename GemmPipeline::BLayout>

◆ GemmCLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GemmCLayout = remove_cvref_t<typename GemmPipeline::CLayout>

◆ GemmDsLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GemmDsLayout = remove_cvref_t<typename EpiloguePipeline::DsLayout>

◆ GemmPipeline

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GemmPipeline = remove_cvref_t<GemmPipeline_>

◆ GroupedConvBwdWeightKernelArgsSpecialized

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GroupedConvBwdWeightKernelArgsSpecialized

Initial value:

GroupedConvBwdWeightKernelArgs<GroupedConvTraitsType_>

ck_tile::GroupedConvBwdWeightKernelArgs

The Grouped Convolution kernel device arguments.

Definition grouped_convolution_backward_weight_kernel.hpp:22

◆ InDataType

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::InDataType = remove_cvref_t<typename GemmPipeline::BDataType>

◆ InLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::InLayout = remove_cvref_t<typename GroupedConvTraitsType_::InLayout>

◆ OutDataType

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::OutDataType = remove_cvref_t<typename GemmPipeline::ADataType>

◆ OutLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::OutLayout = remove_cvref_t<typename GroupedConvTraitsType_::OutLayout>

◆ TilePartitioner

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::TilePartitioner = remove_cvref_t<TilePartitioner_>

◆ WeiDataType

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::WeiDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>

◆ WeiLayout

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

using ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::WeiLayout = remove_cvref_t<typename GroupedConvTraitsType_::WeiLayout>

Member Function Documentation

◆ BlockSize()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST constexpr auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::BlockSize ( )

inlinestaticconstexpr

◆ GetName()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST const std::string ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GetName ( )

inlinestaticnodiscard

◆ GetSmemSize()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST_DEVICE constexpr index_t ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GetSmemSize ( )

inlinestaticconstexpr

◆ GridSize()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST constexpr auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::GridSize ( const GroupedConvBwdWeightKernelArgsSpecialized & kargs )

inlinestaticconstexpr

◆ IsSupportedArgument()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST bool ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::IsSupportedArgument ( const GroupedConvBwdWeightKernelArgsSpecialized & kargs )

inlinestatic

◆ MakeGemmPadViews()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

template<typename TensorView>

CK_TILE_DEVICE auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::MakeGemmPadViews	(	const TensorView &	views,
		const index_t	k_batch )

inlinestatic

◆ MakeGemmTensorViews()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

template<memory_operation_enum DstInMemOp = memory_operation_enum::set>

CK_TILE_DEVICE auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::MakeGemmTensorViews	(	const OutDataType *	a_ptr,
		const InDataType *	b_ptr,
		const std::array< const void *, NumDTensor > &	ds_ptr,
		WeiDataType *	c_ptr,
		const GroupedConvBwdWeightKernelArgsSpecialized &	kargs )

inlinestatic

◆ MakeGemmTileWindows()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

template<typename PadView>

CK_TILE_DEVICE auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::MakeGemmTileWindows	(	const PadView &	views,
		const index_t	i_m,
		const index_t	i_n,
		const index_t	i_k )

inlinestatic

Create views to the data that each workgroup will process.

Parameters

views	padded views of A, B, D and C tensors
i_m	block m-index
i_n	block n-index
i_k	block k-index

Returns: tuple of tile windows for A, B, D and C tensors

◆ MakeKernelArgs()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST constexpr GroupedConvBwdWeightKernelArgsSpecialized ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::MakeKernelArgs ( const GroupedConvBwdWeightHostArgs & hostArgs )

inlinestaticconstexpr

◆ operator()()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_DEVICE void ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::operator() ( GroupedConvBwdWeightKernelArgsSpecialized kargs ) const

inline

◆ Preprocess()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_HOST auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::Preprocess	(	const GroupedConvBwdWeightKernelArgsSpecialized &	kargs,
		const stream_config &	s )

inlinestatic

◆ RunGemm()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_DEVICE void ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::RunGemm	(	const OutDataType *	a_ptr,
		const InDataType *	b_ptr,
		const std::array< const void *, NumDTensor > &	ds_ptr,
		WeiDataType *	c_ptr,
		void *	smem_ptr_0,
		const GroupedConvBwdWeightKernelArgsSpecialized &	kargs,
		const index_t	num_loop,
		const index_t	block_idx_m,
		const index_t	block_idx_n,
		const index_t	block_idx_k )

inlinestatic

Runs single GEMM problem cooperatively by whole workgroup.

Parameters

a_ptr	input A pointer
b_ptr	input B pointer
c_ptr	output C pointer
smem_ptr_0	The start memory pointer of the shared memory block.
kargs	Grouped Convolution Backward Weight kernel arguments
block_idx_m	The GEMM's output M dimension tile index processed by this workgroup.
block_idx_n	The GEMM's output N dimension tile index processed by this workgroup.

◆ RunGemm2LDS()

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

CK_TILE_DEVICE void ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::RunGemm2LDS	(	const OutDataType *	a_ptr,
		const InDataType *	b_ptr,
		const std::array< const void *, NumDTensor > &	ds_ptr,
		WeiDataType *	c_ptr,
		void *__restrict__	smem_ptr_0,
		void *__restrict__	smem_ptr_1,
		const GroupedConvBwdWeightKernelArgsSpecialized &	kargs,
		const index_t	num_loop,
		const index_t	block_idx_m,
		const index_t	block_idx_n,
		const index_t	block_idx_k )

inlinestatic

Runs single GEMM problem cooperatively by whole workgroup.

Note: RunGEMM2LDS in with two shared memory buffers using the ping pong buffer mechanism.

Parameters

a_ptr	input A pointer
b_ptr	input B pointer
c_ptr	output C pointer
smem_ptr_0	The starting pointer of 1st shared memory block.
smem_ptr_1	The starting pointer of 2nd shared memory block.
kargs	Grouped Convolution Backward Weight kernel arguments
block_idx_m	The GEMM's output M dimension tile index processed by this workgroup.
block_idx_n	The GEMM's output N dimension tile index processed by this workgroup.

Member Data Documentation

◆ ConvSpecialization

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

ConvolutionSpecialization ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::ConvSpecialization

staticconstexpr

Initial value:

=

GroupedConvTraitsType_::ConvSpecialization

◆ I0

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::I0 = number<0>()

staticconstexpr

◆ I1

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::I1 = number<1>()

staticconstexpr

◆ I2

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::I2 = number<2>()

staticconstexpr

◆ I3

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

auto ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::I3 = number<3>()

staticconstexpr

◆ IsSplitKSupported

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

bool ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::IsSplitKSupported = true

staticconstexpr

◆ kBlockSize

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

index_t ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::kBlockSize = GemmPipeline::BlockSize

staticconstexpr

◆ NDimSpatial

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

index_t ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::NDimSpatial = GroupedConvTraitsType_::NDimSpatial_

staticconstexpr

◆ NumDTensor

template<typename GroupedConvTraitsType_, typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>

index_t ck_tile::GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ >::NumDTensor = GroupedConvTraitsType_::NumDTensor

staticconstexpr

The documentation for this struct was generated from the following file:

grouped_convolution_backward_weight_kernel.hpp

GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ > Struct Template Reference

GroupedConvolutionBackwardWeightKernel&lt; GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ &gt; Struct Template Reference#

Classes

Public Types

Public Member Functions

Static Public Member Functions

Static Public Attributes

Detailed Description

Overview

Member Typedef Documentation

◆ DsDataType

◆ DsLayout

◆ EpiloguePipeline

◆ GemmALayout

◆ GemmBLayout

◆ GemmCLayout

◆ GemmDsLayout

◆ GemmPipeline

◆ GroupedConvBwdWeightKernelArgsSpecialized

◆ InDataType

◆ InLayout

◆ OutDataType

◆ OutLayout

◆ TilePartitioner

◆ WeiDataType

◆ WeiLayout

Member Function Documentation

◆ BlockSize()

◆ GetName()

◆ GetSmemSize()

◆ GridSize()

◆ IsSupportedArgument()

◆ MakeGemmPadViews()

◆ MakeGemmTensorViews()

◆ MakeGemmTileWindows()

◆ MakeKernelArgs()

◆ operator()()

◆ Preprocess()

◆ RunGemm()

◆ RunGemm2LDS()

Member Data Documentation

◆ ConvSpecialization

◆ I0

◆ I1

◆ I2

◆ I3

◆ IsSplitKSupported

◆ kBlockSize

◆ NDimSpatial

◆ NumDTensor

GroupedConvolutionBackwardWeightKernel< GroupedConvTraitsType_, TilePartitioner_, GemmPipeline_, EpiloguePipeline_ > Struct Template Reference#