Skip to content

GitLab

Commit 48c1b923 authored by rocking's avatar rocking
Browse files

1. Remove E parameter, because E is inside of device op 

2. Check vector size
parent 2d3264ba
Hide whitespace changes
Inline Side-by-side
Showing with 12 additions and 11 deletions
example/21_gemm_layernorm/gemm_add_add_layernorm_xdl_fp16.cpp
View file @ 48c1b923
...@@ -60,11 +60,11 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa ...@@ -60,11 +60,11 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// clang-format off // clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle
//######| ALayout| BLayout| DsLayout| HLayout| AData| BData| AccData| CShuffle| DsData| GammaData| BetaData| HData| A| B| CDE| H| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| PostShuffle| PostShuffle| Layernorm| Layernorm| Layernorm| Layernorm| Layernorm| Layernorm| //######| ALayout| BLayout| DsLayout| HLayout| AData| BData| AccData| CShuffle| DsData| GammaData| BetaData| HData| A| B| CDE| H| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| PostShuffle| PostShuffle| Layernorm| Layernorm| Layernorm| Layernorm| Layernorm|
//######| | | | | Type| Type| Type| DataType| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| ThreadClusterSize| ThreadSliceSize| ESrc| HDst| GammaSrc| BetaSrc| //######| | | | | Type| Type| Type| DataType| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| ThreadClusterSize| ThreadSliceSize| HDst| GammaSrc| BetaSrc|
//######| | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _M_N| _M_N| _M_N| VectorSize| VectorSize| VectorSize| VectorSize| //######| | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _M_N| _M_N| _M_N| VectorSize| VectorSize| VectorSize|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, GammaDataType, BetaDataType, HDataType, AElementOp, BElementOp, CDEElementOp, HElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, S<8, 32>, S<1, 8>, 8, 8, 8, 8>; < ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, GammaDataType, BetaDataType, HDataType, AElementOp, BElementOp, CDEElementOp, HElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, S<8, 32>, S<1, 8>, 8, 8, 8>;
// clang-format on // clang-format on
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
......
include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View file @ 48c1b923
...@@ -223,7 +223,6 @@ template <typename ALayout, ...@@ -223,7 +223,6 @@ template <typename ALayout,
index_t PostShuffleScalarPerVector, index_t PostShuffleScalarPerVector,
typename LayernormThreadClusterSize_M_N, typename LayernormThreadClusterSize_M_N,
typename LayernormThreadSliceSize_M_N, typename LayernormThreadSliceSize_M_N,
index_t LayernormESrcVectorSize,
index_t LayernormHDstVectorSize, index_t LayernormHDstVectorSize,
index_t LayernormGammaSrcVectorSize, index_t LayernormGammaSrcVectorSize,
index_t LayernormBetaSrcVectorSize, index_t LayernormBetaSrcVectorSize,
...@@ -236,7 +235,8 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle : public BaseOperator ...@@ -236,7 +235,8 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle : public BaseOperator
using MeanDataType = CShuffleDataType; using MeanDataType = CShuffleDataType;
using VarDataType = CShuffleDataType; using VarDataType = CShuffleDataType;
static constexpr index_t NumDTensor = DsDataType::Size(); static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr index_t LayernormESrcVectorSize = LayernormHDstVectorSize;
using LayernormBlockTileSize_M_N = using LayernormBlockTileSize_M_N =
Sequence<LayernormThreadClusterSize_M_N::At(0) * LayernormThreadSliceSize_M_N::At(0), Sequence<LayernormThreadClusterSize_M_N::At(0) * LayernormThreadSliceSize_M_N::At(0),
...@@ -851,10 +851,13 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle : public BaseOperator ...@@ -851,10 +851,13 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle : public BaseOperator
} }
// check vector store of E // check vector store of E
// only support RowMajor for now // E and H only support RowMajor for now
if constexpr(is_same_v<ELayout, Row> && is_same_v<HLayout, Row>) if constexpr(is_same_v<ELayout, Row> && is_same_v<HLayout, Row>)
{ {
if(arg.NRaw_ % PostShuffleScalarPerVector != 0) if(arg.NRaw_ % PostShuffleScalarPerVector != 0 ||
arg.NRaw_ % LayernormGammaSrcVectorSize != 0 ||
arg.NRaw_ % LayernormBetaSrcVectorSize != 0 ||
arg.NRaw_ % LayernormHDstVectorSize != 0)
{ {
return false; return false;
} }
...@@ -865,8 +868,6 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle : public BaseOperator ...@@ -865,8 +868,6 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle : public BaseOperator
} }
} }
// TODO - layernorm
return true; return true;
} }
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment