Skip to content

GitLab

"...composable_kernel.git" did not exist on "19c5d6e651d00d15b3909bf1ba44bf59df7f29cf"
Commit 597155e8 authored by Anthony Chang's avatar Anthony Chang
Browse files

change cshuffle precision type to f16; clean up

parent bf44991f
Hide whitespace changes
Inline Side-by-side
Showing with 18 additions and 18 deletions
example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp
View file @ 597155e8
...@@ -31,6 +31,7 @@ using BDataType = F16; ...@@ -31,6 +31,7 @@ using BDataType = F16;
using CDataType = F16; using CDataType = F16;
using C0DataType = F16; using C0DataType = F16;
using AccDataType = F32; using AccDataType = F32;
using CShuffleDataType = F16;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = ck::tensor_layout::gemm::ColumnMajor;
...@@ -47,18 +48,20 @@ struct Relu ...@@ -47,18 +48,20 @@ struct Relu
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough;
// Elementwise operation that operates on the output of matrix multiplication Acc = A * B
using AccElementOp = Relu; using AccElementOp = Relu;
// Elementwise operation that operates on the output of layer normalization
using CElementOp = Relu; using CElementOp = Relu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off // clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmLayerNorm_Xdl_CShuffle using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmLayerNorm_Xdl_CShuffle
//######| ALayout| BLayout| CLayout| AData| BData| CData| C0Data| GemmAcc| CShuffle| ReduceAcc| A| B| Acc| C| 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| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadCopy| //######| ALayout| BLayout| CLayout| AData| BData| CData| C0Data| GemmAcc| CShuffle| ReduceAcc| A| B| Acc| C| 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| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadCopy|
//######| | | | Type| Type| Type| Type| DataType| DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| //######| | | | Type| Type| Type| Type| DataType| DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector|
//######| | | | | | | | | | | 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| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| //######| | | | | | | | | | | 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| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, ADataType, BDataType, CDataType, C0DataType, AccDataType, AccDataType, AccDataType, AElementOp, BElementOp, AccElementOp, CElementOp, 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, 2, S<1, 32, 1, 8>, 8, S<64, 4>, 4>; < Row, Col, Row, ADataType, BDataType, CDataType, C0DataType, AccDataType, CShuffleDataType, AccDataType, AElementOp, BElementOp, AccElementOp, CElementOp, 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, 2, S<1, 32, 1, 8>, 8, S<64, 4>, 4>;
// clang-format on // clang-format on
using ReferenceInstance = ck::tensor_operation::host::ReferenceGemmLayernorm<ADataType, using ReferenceInstance = ck::tensor_operation::host::ReferenceGemmLayernorm<ADataType,
...@@ -252,16 +255,16 @@ int main(int argc, char* argv[]) ...@@ -252,16 +255,16 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
pass &= ck::utils::check_err( if constexpr(std::is_same<CShuffleDataType, F32>::value)
c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c"); {
pass &= ck::utils::check_err(
// if (!pass) c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c");
// { }
// LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",") else if constexpr(std::is_same<CShuffleDataType, F16>::value)
// << std::endl; {
// LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",") pass &= ck::utils::check_err(
// << std::endl; c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c", 1e-2, 1e-2);
// } }
} }
return pass ? 0 : 1; return pass ? 0 : 1;
} }
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
View file @ 597155e8
...@@ -664,7 +664,6 @@ struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1 ...@@ -664,7 +664,6 @@ struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0), make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
c_element_op}; c_element_op};
// add bias: load bias to vgpr buffer, add to LDS
const auto NBlock = c0_grid_desc_nblock_nperblock.GetLength(I0); const auto NBlock = c0_grid_desc_nblock_nperblock.GetLength(I0);
// for broadcasting bias, beta, gamma // for broadcasting bias, beta, gamma
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
View file @ 597155e8
...@@ -65,8 +65,6 @@ struct ReferenceGemmLayernorm : public device::BaseOperator ...@@ -65,8 +65,6 @@ struct ReferenceGemmLayernorm : public device::BaseOperator
} }
avg_acc_sq(i) = sum_acc_sq / N; avg_acc_sq(i) = sum_acc_sq / N;
avg_acc(i) = sum_acc / N; avg_acc(i) = sum_acc / N;
// std::cout << "avg_acc_(" << i << ") =" << avg_acc(i) << std::endl;
// std::cout << "avg_acc_sq_(" << i << ") =" << avg_acc_sq(i) << std::endl;
} }
// normalize // normalize
......
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