FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ > Struct Template Reference

FlatmmKernel&lt; TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ &gt; Struct Template Reference#

Composable Kernel: ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ > Struct Template Reference
Classes | Public Types | Public Member Functions | Static Public Member Functions | Static Public Attributes | List of all members
ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ > Struct Template Reference

#include <flatmm_kernel.hpp>

Inheritance diagram for ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >:
ck_tile::F16xMXF4FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ > ck_tile::GroupedFlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >

Classes

struct  SplitKBatchOffset
 

Public Types

using TilePartitioner = remove_cvref_t< TilePartitioner_ >
 
using FlatmmPipeline = remove_cvref_t< FlatmmPipeline_ >
 
using BlockGemmShape = remove_cvref_t< typename FlatmmPipeline::BlockGemmShape >
 
using EpiloguePipeline = remove_cvref_t< EpiloguePipeline_ >
 
using ALayout = remove_cvref_t< typename FlatmmPipeline::ALayout >
 
using BLayout = remove_cvref_t< typename FlatmmPipeline::BLayout >
 
using ELayout = remove_cvref_t< typename FlatmmPipeline::CLayout >
 
using DsLayout = remove_cvref_t< typename EpiloguePipeline::DsLayout >
 
using DsDataType = remove_cvref_t< typename EpiloguePipeline::DsDataType >
 
using ADataType = remove_cvref_t< typename FlatmmPipeline::ADataType >
 
using BDataType = remove_cvref_t< typename FlatmmPipeline::BDataType >
 
using EDataType = remove_cvref_t< typename EpiloguePipeline::ODataType >
 

Public Member Functions

template<class ScaleM , class ScaleN >
CK_TILE_DEVICE void operator() (FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()> kargs, int partition_idx=blockIdx.x) const
 

Static Public Member Functions

static CK_TILE_HOST const std::string GetName ()
 
static constexpr CK_TILE_HOST auto GridSize (index_t M, index_t N, index_t KBatch)
 
template<class ScaleM , class ScaleN >
static constexpr CK_TILE_HOST auto GridSize (const FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()> &kargs)
 
static constexpr CK_TILE_HOST auto BlockSize ()
 
template<class ScaleM , class ScaleN >
static constexpr CK_TILE_HOST FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()> MakeKernelArgs (const ScaleFlatmmHostArgs< ScaleM, ScaleN, DsDataType::size()> &hostArgs)
 
static constexpr CK_TILE_HOST_DEVICE index_t GetSmemPingSize ()
 
static constexpr CK_TILE_HOST_DEVICE index_t GetSmemPongSize ()
 
template<class KernelArgs >
static CK_TILE_HOST bool IsSupportedArgument (const KernelArgs &kargs)
 
template<memory_operation_enum DstInMemOp = memory_operation_enum::set, class KernelArgs >
static CK_TILE_DEVICE auto MakeGemmTensorViews (const ADataType *a_ptr, const BDataType *b_flat_ptr, const std::array< const void *, NumDTensor > &ds_ptr, EDataType *e_ptr, const KernelArgs &kargs, const SplitKBatchOffset &splitk_batch_offset)
 
template<typename TensorView >
static CK_TILE_DEVICE auto MakeGemmPadViews (const TensorView &views)
 
template<typename PadView >
static CK_TILE_DEVICE auto MakeGemmTileWindows (const PadView &views, const index_t i_m, const index_t i_n)
 
template<class ScaleM , class ScaleN , bool UseDefaultScheduler = true>
static CK_TILE_DEVICE void RunFlatmm (const ADataType *a_ptr, const BDataType *b_flat_ptr, const std::array< const void *, NumDTensor > &ds_ptr, EDataType *e_ptr, void *smem_ptr_ping, void *smem_ptr_pong, const FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()> &kargs, const SplitKBatchOffset &splitk_batch_offset, const index_t block_idx_m, const index_t block_idx_n)
 

Static Public Attributes

static constexpr index_t kBlockSize = FlatmmPipeline::BlockSize
 
static constexpr bool UsePersistentKernel = FlatmmPipeline::UsePersistentKernel
 
static constexpr index_t NumDTensor = DsDataType::size()
 
static constexpr auto I0 = number<0>()
 
static constexpr auto I1 = number<1>()
 
static constexpr auto I2 = number<2>()
 
static constexpr auto I3 = number<3>()
 

Member Typedef Documentation

◆ ADataType

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::ADataType = remove_cvref_t<typename FlatmmPipeline::ADataType>

◆ ALayout

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::ALayout = remove_cvref_t<typename FlatmmPipeline::ALayout>

◆ BDataType

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::BDataType = remove_cvref_t<typename FlatmmPipeline::BDataType>

◆ BLayout

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::BLayout = remove_cvref_t<typename FlatmmPipeline::BLayout>

◆ BlockGemmShape

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::BlockGemmShape = remove_cvref_t<typename FlatmmPipeline::BlockGemmShape>

◆ DsDataType

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::DsDataType = remove_cvref_t<typename EpiloguePipeline::DsDataType>

◆ DsLayout

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::DsLayout = remove_cvref_t<typename EpiloguePipeline::DsLayout>

◆ EDataType

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::EDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>

◆ ELayout

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::ELayout = remove_cvref_t<typename FlatmmPipeline::CLayout>

◆ EpiloguePipeline

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>

◆ FlatmmPipeline

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::FlatmmPipeline = remove_cvref_t<FlatmmPipeline_>

◆ TilePartitioner

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
using ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::TilePartitioner = remove_cvref_t<TilePartitioner_>

Member Function Documentation

◆ BlockSize()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
static constexpr CK_TILE_HOST auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::BlockSize ( )
inlinestaticconstexpr

◆ GetName()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
static CK_TILE_HOST const std::string ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::GetName ( )
inlinestatic

◆ GetSmemPingSize()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
static constexpr CK_TILE_HOST_DEVICE index_t ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::GetSmemPingSize ( )
inlinestaticconstexpr

◆ GetSmemPongSize()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
static constexpr CK_TILE_HOST_DEVICE index_t ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::GetSmemPongSize ( )
inlinestaticconstexpr

◆ GridSize() [1/2]

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<class ScaleM , class ScaleN >
static constexpr CK_TILE_HOST auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::GridSize ( const FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()> &  kargs)
inlinestaticconstexpr

◆ GridSize() [2/2]

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
static constexpr CK_TILE_HOST auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::GridSize ( index_t  M,
index_t  N,
index_t  KBatch 
)
inlinestaticconstexpr

◆ IsSupportedArgument()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<class KernelArgs >
static CK_TILE_HOST bool ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::IsSupportedArgument ( const KernelArgs &  kargs)
inlinestatic

◆ MakeGemmPadViews()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<typename TensorView >
static CK_TILE_DEVICE auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::MakeGemmPadViews ( const TensorView &  views)
inlinestatic

◆ MakeGemmTensorViews()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<memory_operation_enum DstInMemOp = memory_operation_enum::set, class KernelArgs >
static CK_TILE_DEVICE auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::MakeGemmTensorViews ( const ADataType *  a_ptr,
const BDataType *  b_flat_ptr,
const std::array< const void *, NumDTensor > &  ds_ptr,
EDataType *  e_ptr,
const KernelArgs &  kargs,
const SplitKBatchOffset &  splitk_batch_offset 
)
inlinestatic

◆ MakeGemmTileWindows()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<typename PadView >
static CK_TILE_DEVICE auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::MakeGemmTileWindows ( const PadView &  views,
const index_t  i_m,
const index_t  i_n 
)
inlinestatic

◆ MakeKernelArgs()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<class ScaleM , class ScaleN >
static constexpr CK_TILE_HOST FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()> ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::MakeKernelArgs ( const ScaleFlatmmHostArgs< ScaleM, ScaleN, DsDataType::size()> &  hostArgs)
inlinestaticconstexpr

◆ operator()()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<class ScaleM , class ScaleN >
CK_TILE_DEVICE void ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::operator() ( FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()>  kargs,
int  partition_idx = blockIdx.x 
) const
inline

◆ RunFlatmm()

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
template<class ScaleM , class ScaleN , bool UseDefaultScheduler = true>
static CK_TILE_DEVICE void ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::RunFlatmm ( const ADataType *  a_ptr,
const BDataType *  b_flat_ptr,
const std::array< const void *, NumDTensor > &  ds_ptr,
EDataType *  e_ptr,
void *  smem_ptr_ping,
void *  smem_ptr_pong,
const FlatmmKernelArgs< ScaleM, ScaleN, DsDataType::size()> &  kargs,
const SplitKBatchOffset &  splitk_batch_offset,
const index_t  block_idx_m,
const index_t  block_idx_n 
)
inlinestatic

Member Data Documentation

◆ I0

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::I0 = number<0>()
staticconstexpr

◆ I1

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::I1 = number<1>()
staticconstexpr

◆ I2

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::I2 = number<2>()
staticconstexpr

◆ I3

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr auto ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::I3 = number<3>()
staticconstexpr

◆ kBlockSize

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr index_t ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::kBlockSize = FlatmmPipeline::BlockSize
staticconstexpr

◆ NumDTensor

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr index_t ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::NumDTensor = DsDataType::size()
staticconstexpr

◆ UsePersistentKernel

template<typename TilePartitioner_ , typename FlatmmPipeline_ , typename EpiloguePipeline_ >
constexpr bool ck_tile::FlatmmKernel< TilePartitioner_, FlatmmPipeline_, EpiloguePipeline_ >::UsePersistentKernel = FlatmmPipeline::UsePersistentKernel
staticconstexpr
The documentation for this struct was generated from the following file:
  • /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp