Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack > Struct Template Reference

Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack > Struct Template Reference#

Composable Kernel: ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack > Struct Template Reference
Public Types | Public Member Functions | Static Public Member Functions | Static Public Attributes | Protected Attributes | Static Protected Attributes | List of all members
ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack > Struct Template Reference

#include <blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp>

Inheritance diagram for ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >:
ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >

Public Types

using Base
using ComputeDataTypeBuf = typename Base::ComputeDataTypeBuf
Public Types inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
using ThisThreadBlock
using ComputeDataTypeBuf
using HotLoopInstList
using Tuple4

Public Member Functions

template<bool HasMainLoop, TailNumber TailNum, typename AGridDesc, typename ABlockDesc, typename ABlockTransfer, typename AGridBuffer, typename ABlockBuffer, typename ABlockTransferStep, typename BGridDesc, typename BBlockDesc, typename BBlockTransfer, typename BGridBuffer, typename BBlockBuffer, typename BBlockTransferStep, typename CThreadBuffer, typename AScaleGridBuffer, typename AScaleGridDesc, typename AScaleThreadDesc, typename AScaleThreadTransfer, typename AScaleThreadTransferStep, typename BScaleGridBuffer, typename BScaleGridDesc, typename BScaleThreadDesc, typename BScaleThreadTransfer, typename BScaleThreadTransferStep>
__device__ void Run (const AGridDesc &a_grid_desc, const ABlockDesc &a_block_desc, ABlockTransfer &a_blockwise_copy, const AGridBuffer &a_grid_buf, ABlockBuffer &a_block_buf, const ABlockTransferStep &a_block_copy_step, const BGridDesc &b_grid_desc, const BBlockDesc &b_block_desc, BBlockTransfer &b_blockwise_copy, const BGridBuffer &b_grid_buf, BBlockBuffer &b_block_buf, const BBlockTransferStep &b_block_copy_step, CThreadBuffer &c_thread_buf, const AScaleGridDesc &a_scale_grid_desc, const AScaleThreadDesc &a_scale_thread_desc, AScaleThreadTransfer &a_scale_thread_copy, const AScaleGridBuffer &a_scale_grid_buf, const AScaleThreadTransferStep &a_scale_thread_copy_step, const BScaleGridDesc &b_scale_grid_desc, const BScaleThreadDesc &b_scale_thread_desc, BScaleThreadTransfer &b_scale_thread_copy, const BScaleGridBuffer &b_scale_grid_buf, const BScaleThreadTransferStep &b_scale_thread_copy_step, index_t num_loop, index_t num_loop_per_scale) const
__host__ __device__ constexpr auto & GetCThreadBuffer ()
Public Member Functions inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
__host__ __device__ constexpr auto & GetCThreadBuffer ()
__host__ __device__ BlockwiseGemmXdlops_pipeline_base (Tuple4 a_origin=CalculateAThreadOriginDataIndex(), Tuple4 b_origin=CalculateBThreadOriginDataIndex())
 Constructor for BlockwiseGemmXdlops_pipeline_base.

Static Public Member Functions

static __host__ constexpr bool BlockHasHotloop (index_t num_loop)
static __host__ constexpr TailNumber BlockLoopTailNum (index_t num_loop)
static __device__ auto CalculateCThreadOriginDataIndex (Number< m0 >, Number< n0 >, Number< xdlops_i >, Number< blk_i >)
static __device__ auto CalculateCThreadOriginDataIndex8D (Number< m0 >, Number< n0 >, Number< xdlops_i >, Number< blk_i >)
__host__ static __device__ constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4 ()
__host__ static __device__ constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4 ()
__host__ static __device__ constexpr auto MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 (const CGridDesc_G_M_N &c_grid_desc_g_m_n)
__host__ static __device__ constexpr auto MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 (const CGridDesc_M_N &c_grid_desc_m_n)
Static Public Member Functions inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
static __device__ auto GetWaveIdx ()
static __device__ auto CalculateAThreadOriginDataIndex ()
static __device__ auto CalculateAThreadOriginDataIndex6D ()
static __device__ auto CalculateBThreadOriginDataIndex ()
static __device__ auto CalculateCThreadOriginDataIndex (Number< m0 >, Number< n0 >, Number< xdlops_i >, Number< blk_i >)
static __device__ auto CalculateCThreadOriginDataIndex8D (Number< m0 >, Number< n0 >, Number< xdlops_i >, Number< blk_i >)
__host__ static __device__ constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4 ()
__host__ static __device__ constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4 ()
__host__ static __device__ constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ()
__host__ static __device__ constexpr auto MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 (const CGridDesc_M_N &c_grid_desc_m_n)
__host__ static __device__ constexpr auto MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 (const CGridDesc_G_M_N &c_grid_desc_g_m_n)
__host__ static __device__ constexpr auto GetCThreadDesc ()

Static Public Attributes

static constexpr index_t WgpPerCU
static constexpr index_t FullMemBandPrefetchStages
static constexpr index_t PrefetchStages
static constexpr index_t PrefillStages = 1
static constexpr index_t GlobalBufferNum = PrefetchStages
static constexpr auto I0
static constexpr index_t KRepeat
static constexpr auto xdlops_gemm
static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k
static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k
static constexpr index_t AMmaKStride
static constexpr index_t BMmaKStride
static constexpr index_t WaveSize
Static Public Attributes inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
static constexpr auto I0
static constexpr auto I1
static constexpr auto I2
static constexpr auto I3
static constexpr index_t MWaves
static constexpr index_t NWaves
static constexpr index_t WaveSize
static constexpr index_t A_K0
static constexpr index_t B_K0
static constexpr index_t A_K1
static constexpr index_t B_K1
static constexpr auto xdlops_gemm
static constexpr index_t AMmaKStride
static constexpr index_t BMmaKStride
static constexpr index_t KPerThread
static constexpr index_t KRepeat
static constexpr index_t KPerInnerLoop
static constexpr index_t KGroup
static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k
static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k

Protected Attributes

AThreadCopy a_thread_copy_
BThreadCopy b_thread_copy_
Protected Attributes inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
AThreadCopy a_thread_copy_
BThreadCopy b_thread_copy_

Static Protected Attributes

static constexpr auto a_thread_desc_
static constexpr auto b_thread_desc_
static constexpr auto c_thread_desc_
Static Protected Attributes inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
static constexpr auto a_thread_desc_
static constexpr auto b_thread_desc_
static constexpr auto c_thread_desc_

Additional Inherited Members

Public Attributes inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
StaticBufferTupleOfVector< AddressSpaceEnum::Vgpr, AccDataType, MRepeat *NRepeat, xdlops_gemm.GetRegSizePerXdlops(), true > c_thread_buf_
Protected Types inherited from ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, true >
using AThreadCopy
using BThreadCopy

Member Typedef Documentation

◆ Base

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
using ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::Base
Initial value:
BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack,
true>
ck::BlockwiseGemmXdlops_pipeline_base::BlockwiseGemmXdlops_pipeline_base
__host__ __device__ BlockwiseGemmXdlops_pipeline_base(Tuple4 a_origin=CalculateAThreadOriginDataIndex(), Tuple4 b_origin=CalculateBThreadOriginDataIndex())
Constructor for BlockwiseGemmXdlops_pipeline_base.
Definition blockwise_gemm_pipeline_xdlops_base.hpp:222

◆ ComputeDataTypeBuf

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
using ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::ComputeDataTypeBuf = typename Base::ComputeDataTypeBuf

Member Function Documentation

◆ BlockHasHotloop()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ constexpr bool ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::BlockHasHotloop ( index_t num_loop)
inlinestaticconstexpr

◆ BlockLoopTailNum()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ constexpr TailNumber ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::BlockLoopTailNum ( index_t num_loop)
inlinestaticconstexpr

◆ CalculateCThreadOriginDataIndex()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__device__ auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::CalculateCThreadOriginDataIndex ( Number< m0 > ,
Number< n0 > ,
Number< xdlops_i > ,
Number< blk_i >  )
inlinestatic

◆ CalculateCThreadOriginDataIndex8D()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__device__ auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::CalculateCThreadOriginDataIndex8D ( Number< m0 > ,
Number< n0 > ,
Number< xdlops_i > ,
Number< blk_i >  )
inlinestatic

◆ GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ( )
inlinestaticconstexpr

◆ GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ( )
inlinestaticconstexpr

◆ GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4 ( )
inlinestaticconstexpr

◆ GetCThreadBuffer()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ __device__ constexpr auto & ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCThreadBuffer ( )
inlineconstexpr

◆ GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ( )
inlinestaticconstexpr

◆ GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ( )
inlinestaticconstexpr

◆ GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4 ( )
inlinestaticconstexpr

◆ MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2 ( const CGridDesc_G_M_N & c_grid_desc_g_m_n)
inlinestaticconstexpr

◆ MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
__host__ static __device__ constexpr auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2 ( const CGridDesc_M_N & c_grid_desc_m_n)
inlinestaticconstexpr

◆ Run()

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
template<bool HasMainLoop, TailNumber TailNum, typename AGridDesc, typename ABlockDesc, typename ABlockTransfer, typename AGridBuffer, typename ABlockBuffer, typename ABlockTransferStep, typename BGridDesc, typename BBlockDesc, typename BBlockTransfer, typename BGridBuffer, typename BBlockBuffer, typename BBlockTransferStep, typename CThreadBuffer, typename AScaleGridBuffer, typename AScaleGridDesc, typename AScaleThreadDesc, typename AScaleThreadTransfer, typename AScaleThreadTransferStep, typename BScaleGridBuffer, typename BScaleGridDesc, typename BScaleThreadDesc, typename BScaleThreadTransfer, typename BScaleThreadTransferStep>
__device__ void ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::Run ( const AGridDesc & a_grid_desc,
const ABlockDesc & a_block_desc,
ABlockTransfer & a_blockwise_copy,
const AGridBuffer & a_grid_buf,
ABlockBuffer & a_block_buf,
const ABlockTransferStep & a_block_copy_step,
const BGridDesc & b_grid_desc,
const BBlockDesc & b_block_desc,
BBlockTransfer & b_blockwise_copy,
const BGridBuffer & b_grid_buf,
BBlockBuffer & b_block_buf,
const BBlockTransferStep & b_block_copy_step,
CThreadBuffer & c_thread_buf,
const AScaleGridDesc & a_scale_grid_desc,
const AScaleThreadDesc & a_scale_thread_desc,
AScaleThreadTransfer & a_scale_thread_copy,
const AScaleGridBuffer & a_scale_grid_buf,
const AScaleThreadTransferStep & a_scale_thread_copy_step,
const BScaleGridDesc & b_scale_grid_desc,
const BScaleThreadDesc & b_scale_thread_desc,
BScaleThreadTransfer & b_scale_thread_copy,
const BScaleGridBuffer & b_scale_grid_buf,
const BScaleThreadTransferStep & b_scale_thread_copy_step,
index_t num_loop,
index_t num_loop_per_scale ) const
inline

Member Data Documentation

◆ a_block_desc_m0_m1_m2_k

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
AMmaTileDesc ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::a_block_desc_m0_m1_m2_k
staticconstexpr

◆ a_thread_copy_

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
AThreadCopy ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::a_thread_copy_
protected

◆ a_thread_desc_

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::a_thread_desc_
staticconstexprprotected

◆ AMmaKStride

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::AMmaKStride
staticconstexpr

◆ b_block_desc_n0_n1_n2_k

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
BMmaTileDesc ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::b_block_desc_n0_n1_n2_k
staticconstexpr

◆ b_thread_copy_

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
BThreadCopy ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::b_thread_copy_
protected

◆ b_thread_desc_

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::b_thread_desc_
staticconstexprprotected

◆ BMmaKStride

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::BMmaKStride
staticconstexpr

◆ c_thread_desc_

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::c_thread_desc_
staticconstexprprotected

◆ FullMemBandPrefetchStages

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::FullMemBandPrefetchStages
staticconstexpr
Initial value:
= math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock)
ck::math::integer_divide_ceil
__host__ __device__ constexpr auto integer_divide_ceil(X x, Y y)
Definition utility/math.hpp:72
ck::BlockwiseGemmXdlops_pipeline_v2< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::WgpPerCU
static constexpr index_t WgpPerCU
Definition blockwise_gemm_pipeline_xdlops_v2.hpp:146

◆ GlobalBufferNum

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::GlobalBufferNum = PrefetchStages
staticconstexpr

◆ I0

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::I0
staticconstexpr

◆ KRepeat

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::KRepeat
staticconstexpr

◆ PrefetchStages

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::PrefetchStages
staticconstexpr
Initial value:
=
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2
ck::BlockwiseGemmXdlops_pipeline_v2< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::FullMemBandPrefetchStages
static constexpr index_t FullMemBandPrefetchStages
Definition blockwise_gemm_pipeline_xdlops_v2.hpp:148

◆ PrefillStages

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::PrefillStages = 1
staticconstexpr

◆ WaveSize

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::WaveSize
staticconstexpr

◆ WgpPerCU

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
index_t ck::BlockwiseGemmXdlops_pipeline_v2_ab_scale< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::WgpPerCU
staticconstexpr
Initial value:
=
(4 * WaveSize / BlockSize) >= 1 ? 4 * WaveSize / BlockSize : 1
ck::BlockwiseGemmXdlops_pipeline_v2< BlockGemmPipelineScheduler::Intrawave, BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack >::WaveSize
static constexpr index_t WaveSize
Definition blockwise_gemm_pipeline_xdlops_base.hpp:46

◆ xdlops_gemm

template<index_t BlockSize, typename ADataType, typename BDataType, typename ComputeDataType, typename AccDataType, typename ATileDesc, typename BTileDesc, typename AMmaTileDesc, typename BMmaTileDesc, index_t ABlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t MPerXDL, index_t NPerXDL, index_t MRepeat, index_t NRepeat, index_t KPack>
auto ck::BlockwiseGemmXdlops_pipeline_base< BlockSize, ADataType, BDataType, ComputeDataType, AccDataType, ATileDesc, BTileDesc, AMmaTileDesc, BMmaTileDesc, ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXDL, NPerXDL, MRepeat, NRepeat, KPack, TransposeC >::xdlops_gemm
staticconstexpr
The documentation for this struct was generated from the following file: